Artificial intelligence in cancer: applications, challenges, and future perspectives

Abstract

Artificial intelligence (AI) is rapidly revolutionizing the landscape of oncological research and the advancement of personalized clinical interventions. Progress in three interconnected areas, including the development of methods and algorithms for training AI models, the evolution of specialized computing hardware, and increased access to large volumes of cancer data such as imaging, genomics, and clinical information, has converged, leading to promising new applications of AI in cancer research. AI applications are systematically organized according to specific cancer types and clinical domains, encompassing the elucidation and prediction of biological mechanisms, the identification and utilization of patterns within clinical data to improve patient outcomes, and the unraveling of the complexities inherent in epidemiological, behavioral, and real-world datasets. When applied in an ethical and scientifically rigorous manner, these AI-driven approaches hold the promise of accelerating progress in cancer research and ultimately fostering improved health outcomes for all populations. We review examples demonstrating the integration of AI within oncology, highlighting cases where deep learning has adeptly addressed challenges once deemed insurmountable, while also discussing the barriers that must be surmounted to facilitate broader adoption of these technologies.

Introduction

Cancer remains a principal cause of mortality worldwide [1]. Projections estimate approximately 35 million cases by 2050 [2]. This alarming rise highlights the imperative to accelerate progress in cancer research and the development of therapeutic strategies.

Over the last decade, there has been a renewed and growing interest in the integration of artificial intelligence (AI) within the medical field, propelled by the advent of advanced deep-learning algorithms, significant advancements in computational hardware, and the rapid growth of data leveraged for clinical decision-making [3–5]. Furthermore, its application in oncology exhibits remarkable and expanding potential, encompassing foundational scientific pursuits such as protein folding predictions [6, 7], translational initiatives such as biomarker discovery [8, 9], and clinical progress in the organization and management of trials [10, 11].

In this review, we aim to provide a comprehensive overview of the present state and evolving landscape of AI in the realm of oncology. We initiate our discussion by summarizing the major types of AI models and input data modalities. Next, we review recent advancements in AI across six key domains, including cancer screening and diagnosis, precision treatment, cancer surveillance, drug discovery, health care delivery, and mechanisms of cancer. Finally, we highlight the principal obstacles impeding the widespread clinical integration of AI and propose strategic, actionable approaches to catalyze future innovations in this rapidly evolving field.

AI models and data modalities

Artificial intelligence enables systems to learn from data, recognize patterns, and make decisions [12]. In oncology, AI uses diverse data modalities, including medical imaging, genomics, and clinical records, to address complex challenges [13]. The selection of AI models depends on the data type and clinical objective [3]. Structured data such as genomic biomarkers and lab values are often analyzed using classical machine learning (ML) models including logistic regression and ensemble methods for tasks such as survival prediction or therapy response [14]. Imaging data including histopathology and radiology utilize deep learning (DL) architectures such as convolutional neural networks (CNNs) to extract spatial features, enabling tumor detection, segmentation, and grading [15]. Sequential or text data such as genomic sequences and clinical notes employ transformers or recurrent neural networks (RNNs) to model long-range dependencies, facilitating tasks such as biomarker discovery or electronic health record (EHR) mining [16]. Recent advances in large language models (LLMs) such as GPT-5 enhance knowledge extraction from scientific literature and clinical text, accelerating hypothesis generation in cancer research. For an overview of AI model evolution and technical specifications, see Appendix (Fig. 1).

Fig. 1.

Overview of AI models and corresponding data modalities in precision oncology. A Developmental history of AI and key models. B Clinical origin datasets leveraged in AI models. CT, computed tomography; MRI, magnetic resonance imaging; H&E, hematoxylin and eosin; IHC, immunohistochemistry; CNNs, convolutional neural networks; GNNs, graph neural networks

AI applications in cancer research and care

The integration of AI in cancer research and clinical practice encompasses advancing screening and diagnostic accuracy, improving precision cancer treatment, enhancing cancer surveillance, accelerating drug discovery, optimizing healthcare delivery, and elucidating cancer mechanisms. These applications collectively strive to improve patient outcomes and streamline clinical workflows, fostering more efficient and personalized cancer management (Fig. 2).

Fig. 2.

Key applications of artificial intelligence across the cancer care continuum. LLMs, large language models; FDA, food and drug administration; EHR, electronic health record

Expediting cancer screening, detection and diagnosis

AI is playing an increasingly important role in enhancing various aspects of cancer screening and detection methods by significantly improving their speed, accuracy, and overall reliability. The extensive and continually growing datasets generated by current screening programs offer a remarkable opportunity for the development and implementation of advanced AI applications. As AI continues to evolve, its application across diverse cancer detection modalities promises to revolutionize early diagnosis and improve clinical decision-making processes (Table 1).

Table 1.

Performance comparison of AI systems vs. Human experts in oncology applications

Cancer type	Modality	Task	AI system	Dataset size	Gold-standard labels	Sensitivity	Specificity	Area Under the Curve (AUC)	External validation cohorts	Evidence level	Ref
Colorectal cancer	Colonoscopy	Malignancy detection	CRCNet	Training: 464,105 images from 12,179 patients Testing: 2,263 patients across 3 cohorts	A group of five skilled endoscopists	AI vs. Human: 1ST test set: 91.3% vs. 83.8% (p < 0.001) 2nd test set: 82.9% vs 87.6% (p = 0.29) 3rd test set: 96.5% vs. 90.3% (p = 0.006)	AI vs. Human: 1ST test set: 85.3% (AI) 2nd test set: 95.0% (AI) 3rd test set: 99.2% (AI)	1ST test set: 0.882 (95% CI: 0.828–0.931) 2nd test set: 0.874 (0.820–0.926) 3rd test set: 0.867 (0.795–0.923)	Three independent cohorts: 1. Tianjin Cancer Hospital (internal) 2. Tianjin First Central Hospital (external) 3. Tianjin General Hospital (external)	Retrospective multicohort diagnostic study with external validation	[17]
	Colonoscopy/Histopathology	Histological classification of colorectal polyps (neoplastic vs. nonneoplastic)	Real-time image recognition system (bag-of-features with SIFT descriptors + SVM classifier)	118 lesions (from 41 patients)	Histopathology (by a single GI pathologist blinded to real-time image recognition system)	AI: 95.9% (detection of neoplastic lesions)	AI: 93.3% (identification of nonneoplastic lesions)	NR	None (single-center study at Hiroshima University Hospital)	Prospective diagnostic accuracy study with blinded gold standard	[18]
	Colonoscopy/Histopathology	Compare autonomous AI to AI-assisted human (AI-H) optical diagnosis	CADx	AI: 238 patients/158 polyps AI-H: 229 patients/179 polyps	Histopathology by board-certified pathologists	AI vs. AI-H: 84.8% (95% CI: 76.8–90.9) vs. 83.6 (75.4–90.0)	AI vs. AI-H: 64.4% (48.8–78.1) vs. 63.8% (51.3–75.0)	NR	None (single-center study at Montreal University Hospital Center)	Randomized controlled trial	[19]
Breast cancer	2D Mammography	Screening detection	Ensemble of three DL models (lesion-, breast-, and case-level analysis)	UK: 25,856 women US: 3,097 women Reader study: 500 cases (465 analyzed)	Biopsy-confirmed cancer within extended follow-up: UK: 39 months US: 27 months Reader study: 27 months	UK vs. 1 st reader:  + 2.7% (P = 0.004 for non-inferiority) US vs. radiologists:  + 9.4% (P < 0.001)	UK vs. 1 st reader:  + 1.2% (P = 0.0096) US vs. radiologists:  + 5.7% (P < 0.001)	UK: 0.889 (95% CI 0.871–0.907) US: 0.8107 (95% CI 0.791–0.831) Reader study (US): 0.740	Model trained only on UK data tested on US dataset	Diagnostic case–control study with comparison to radiologists	[20]
	2D/3D Mammography	1. Screening detection 2. Early cancer detection	Progressively trained RetinaNet with MSP for DBT	Reader Study (Site D):131 index cancers + 154 confirmed negatives 120 pre-index cancers + 154 negatives	Cancers: Biopsy-proven malignancy Negatives: BI-RADS 1/2 + subsequent confirmed negative screen (9–39 months) Pre-index: Prior "negative" exams (12–24 mo pre-diagnosis)	Index exams:  + 14.2% absolute increase (95% CI: 9.2–18.5%) at avg. reader specificity Pre-index exams:  + 17.5% absolute increase (95% CI: 6.0–26.2%) at avg. reader specificity	Index exams:  + 24.0% absolute increase (95% CI: 17.4–30.4%) at avg. reader sensitivity Pre-index exams:  + 16.2% absolute increase (95% CI: 7.3–24.6%) at avg. reader sensitivity	Reader Study (Site D): Index: 0.94 OMI-DB: 0.963 ± 0.003 Site A-DM: 0.927 ± 0.008 Site A-DBT: 0.957 ± 0.010 Site E (China): 0.971 ± 0.005 Site E (size-adjusted): 0.956 ± 0.020	1. OMI-DB (UK): Screening population 2. Site A-DM (US): GE equipment 3. Site A-DBT (US): Hologic equipment 4. Site E (China): Diagnostic population, Hologic equipment 5. Reader Study (Site D, US): Hologic equipment	Diagnostic case–control study with comparison to radiologists	[21]
	B-mode and color Doppler ultrasound	1. Malignancy detection 2. Localization (saliency maps for lesion detection)	Weakly supervised DL model (ResNet-18 backbone) Trained with breast-level labels (no per-image/pixel annotations)	Internal (NYU): 288,767 exams (5,442,907 images) from 143,203 patients External (BUSI): 780 images from 600 patients	Pathology reports (biopsy/surgery within 30–120 days of ultrasound) Test set filtering: Positive exams: Cancers confirmed visible in ultrasound Negative exams: ≥ 1 cancer-negative follow-up or biopsy	AI: 94.5% (at radiologists' avg. specificity) Radiologists (avg): 90.1% Improvement: + 4.4% (95% CI: − 0.3%, 7.5%; P = 0.0278)	AI: 85.6% (at radiologists' avg. sensitivity) Radiologists (avg): 80.7% Improvement: + 4.9% (95% CI: 3.0%, 7.1%; P < 0.001)	Internal test: 0.976 (95% CI: 0.972–0.980) Reader study (AI): 0.962 (95% CI: 0.943–0.979) Radiologists (avg): 0.924 ± 0.02 External validation (BUSI): 0.927 (95% CI: 0.907–0.959)	BUSI dataset (Egypt): 780 images (437 benign, 210 malignant, 133 normal)	Retrospective reader study	[22]
	B-mode and color Doppler ultrasound	Improve the detection rate of early breast cancer while reducing misdiagnosis	EDL-BC	Training: 7,076 lesions (14,152 images) Internal validation: 879 lesions External validation 1 (TS): 448 lesions (Tangshan People's Hospital) External validation 2 (DZ): 245 lesions (Dazu People's Hospital)	Pathology biopsy (malignant) or 3-year follow-up confirmation (benign)	AI alone (threshold 0.76): 94.4% (internal), 100% (TS), 80% (DZ) Radiologists alone: 12.1% (average) Radiologists with AI assistance: 54.5%	AI alone (threshold 0.76): 84.2% (internal) Radiologists alone: 99.1% Radiologists with AI assistance: 98.2%	Internal validation: 0.950 (95% CI: 0.909–0.969) External validation (TS): 0.956 (95% CI: 0.939–0.971) External validation (DZ): 0.907 (95% CI: 0.877–0.938) Radiologists with AI assistance: 0.899 (vs. 0.716 without AI; p < 0.0001)	Two independent cohorts: 1. Tangshan (TS): 448 lesions (391 patients), northern China 2. Dazu (DZ): 245 lesions (235 patients), southwestern China	Retrospective, multicenter diagnostic study	[23]
	Ultrasound	Binary classification of metastatic vs. non-metastatic lymph nodes	Custom CNN (3 × 3 convolutional kernels, 9 layers including 6 residual layers)	Total: 338 images (169 patients) Training: 248 images (124 patients) Testing: 90 images (45 patients)	Core needle biopsy pathology (ground truth) Metastatic: 64 patients Non-metastatic: 105 patients	AI: 65.5% (SD ± 28.6)	AI: 78.9% (SD ± 15.1)	0.72 (SD ± 0.08)	None (single-center retrospective study)	Retrospective diagnostic study without external validation	[24]
	Histopathology (H&E/HES-stained whole slide images of biopsies)	1. Malignancy detection 2. Subtype classification 3. Feature identification (e.g. TILs, ALI)	Galen Breast (Ibex Medical Analytics) Architecture: Ensemble of CNNs	Internal test: 2,252 slides (1,090 cases) External validation: 841 slides (436 cases) Clinical deployment: 12,031 slides (5,954 cases)	Pathologist consensus (blinded review by ≥ 2 senior breast pathologists) Discrepancies resolved by third pathologist	AI: Invasive carcinoma detection: 95.51% (95% CI: 91.03–97.81%) DCIS/ADH detection: 93.20% (95% CI: 86.63–96.67%)	AI: Invasive carcinoma detection: 93.57% (95% CI: 90.07–95.90%) DCIS/ADH detection: 93.79% (95% CI: 88.63–96.70%)	Invasive carcinoma detection: 0.990 (95% CI: 0.984–0.995) DCIS/ADH detection: 0.980 (95% CI: 0.967–0.993)	1. Institut Curie (France) 2. Maccabi Healthcare Services (Israel)	Prospective validation in clinical setting with real-time deployment	[25]
	Histopathology (H&E-stained whole-slide images)	1. Binary classification of NHG 1 vs. NHG 3 2. Risk stratification of NHG 2 tumors into DG2-low/DG2-high	DeepGrade: Ensemble of 20 CNNs	Internal development: 1,567 patients External validation: 1,262 patients	Nottingham Histological Grade (NHG) assigned by pathologists during routine clinical diagnostics	NR	NR	NHG 1 vs. NHG 3 classification: - Internal AUC: 0.919–0.937 (95% CI: 0.884–0.987) - External AUC: 0.907 (95% CI: 0.885–0.930)	the Sweden Cancerome Analysis Network—Breast (SCAN-B) Lund cohort (n = 1,262, fully independent)	Retrospective cohort study with external validation	[26]
	Histopathology (H&E-stained whole-slide images)	Detection of metastases in lymph nodes (LNs) (micrometastases, macrometastases, and negative cases)	LYmph Node Assistant (LYNA)	Total: 70 digitized slides	Established by 3 board-certified pathologists (≥ 7 years' experience) using H&E and IHC Discrepancies resolved via adjudicated review	For micrometastases: Assisted: 91.2% (95% CI: 86.0–96.5%) Unassisted: 83.3% (95% CI: 76.4–90.2%) P = 0.023	NR	LYNA algorithm alone: Overall: 99.0% Micrometastases + negative slides: 98.5%	None. Study used internal test data only	Multireader, multicase retrospective study with intra-dataset validation	[27]
	Histopathology (H&E-stained whole-slide images)	Detection of metastases in LNs (macrometastases, micrometastases, isolated tumor cells)	Visiopharm Integrator System (VIS) metastasis AI algorithm	Total: 594 LNs SLN validation cohort: 234 sentinel LNs SLN consensus cohort: 102 sentinel LNs NSLN cohort: 258 nonsentinel LNs	Consensus review by 3 subspecialized breast pathologists	100% across all cohorts	SLN validation cohort: 41.5% NSLN cohort: 78.5%	NR	None (single-center retrospective study)	Retrospective diagnostic cohort study	[28]
Lung cancer	Chest radiography (X-ray)	Detection of pulmonary metastases from mixed primary cancers	Lunit INSIGHT CXR 1 (v1.1.2.0), DL-based CAD	Main study: CAD-assisted: 2,916 CXRs Conventional: 5,681 CXRs (after propensity matching) Subanalysis (reader test): 215 CXRs with paired CT	Clinical/pathological diagnosis + CT confirmation	Subanalysis (per-examination): Standalone CAD: 80% (73/91) Radiologists without CAD: 71% (257/364) Radiologists with CAD: 77% (279/364)	Subanalysis (per-examination): Standalone CAD: 75% (93/124) Radiologists without CAD: 73% (360/496) Radiologists with CAD: 75% (374/496)	NR	None (single-center retrospective study)	Retrospective diagnostic cohort study	[29]
	Chest Radiography (X-ray)	Detection of actionable lung nodules (Lung-RADS category 4)	AI-based CAD software (Lunit INSIGHT CXR v2.0.2.0)	10,476 participants (AI group: 5,238; non-AI group: 5,238)	CT scans within 3 months post-X-ray (actionable nodules: solid > 8 mm or subsolid solid portion > 6 mm)	AI-assisted radiologists: 56.4% (31/55 actionable nodules detected) Non-AI radiologists: 23.2% (13/56 actionable nodules detected)	AI group: 97.6% Non-AI group: 97.7%	NR	None (single-center trial: Seoul National University Hospital, Korea)	Prospective randomized controlled trial	[30]
	Low-dose computed tomography (LDCT)	End-to-end malignancy risk prediction and localization	3D DL model	Training/Tuning/Test (NLST): 42,290 CT cases (14,851 patients; 578 biopsy-confirmed cancers within 1 year) External validation: 1,139 cases (907 patients; 27 biopsy-confirmed cancers) from a US academic medical center	Cancer-positive: Biopsy/surgical confirmation within screening year Cancer-negative: Cancer-free on 1-year follow-up LDCT	Without prior CT: AI: + 5.2% improvement (95% CI: 0.38–9.9) at LuMAS 3+ vs. Lung-RADS 3+  Early-stage cancers: + 24.4% sensitivity improvement	Without prior CT: AI: + 11% improvement (95% CI: 7.8–15.1) at LuMAS 3+ vs. Lung-RADS 3+  With prior CT: Significant improvement for Lung-RADS 4A + threshold	NLST test set: 94.4% (95% CI: 91.1–97.3) External validation set: 95.5% (95% CI: 88.0–98.4)	Independent cohort from a US academic medical center (n = 1,139 cases)	Retrospective cohort study with external validation	[31]
	LDCT	Malignancy risk estimation of pulmonary nodules	DL algorithm using ensemble of 2D (ResNet50) and 3D (Inception-v1) CNNs	Development: 16,077 nodules (1,249 malignant) from NLST External validation: 883 nodules (65 malignant) from DLCST	Histopathologic confirmation for malignancies;  ≥ 2 years CT follow-up for benign nodules	Full DLCST cohort (90% specificity): AI: 84% (54/65) PanCan model: 63% (41/65) Cancer-enriched subset A: AI: 91% (54/59) at 90% specificity Clinicians (avg): NR Cancer-enriched subset B: AI: 54% (32/59) at 90% specificity Clinicians: Higher than 8/11 clinicians	Fixed at 90% for sensitivity comparisons (as per study design)	Full DLCST cohort: AI: 0.93 (95% CI: 0.89–0.96) PanCan model: 0.90 (95% CI: 0.86–0.93) Cancer-enriched subset A: AI: 0.96 (95% CI: 0.93–0.99) Clinicians (avg): 0.90 (95% CI: 0.87–0.94) Cancer-enriched subset B: AI: 0.86 (95% CI: 0.80–0.91) Clinicians (avg): 0.82 (95% CI: 0.77–0.86)	DLCST cohort: n = 883 Cancer-enriched subset A: n = 175 Cancer-enriched subset B: n = 177	Retrospective cohort study with external validation	[32]
	LDCT	Malignancy risk stratification of pulmonary nodules	Chinese Lung Nodules Reporting and Data System (C-Lung-RADS)	Primary cohort (MCC): 45,064 participants Independent testing cohort (MSC): 14,437 participants	Pathologically confirmed malignancy (label 4) or clinician ratings (labels 1–3)	C-Lung-RADS vs. Lung-RADS v2022 (Human): Internal testing: 79.9% vs. 60.3% (P < 0.001) External testing: 87.1% vs. 63.3% (P < 0.001)	C-Lung-RADS vs. Lung-RADS v2022 (Human): Internal testing: 91.8% (derived from 8.2% FPR) External testing: 94.1% (derived from 5.9% FPR)	Overall performance: Internal testing: 0.918 (95% CI: 0.918–0.919) External testing: 0.927 (95% CI: 0.926–0.928)	Mobile screening cohort (MSC): 14,437 participants screened via mobile CT units across Western China (Sichuan Province)	Retrospective cohort study with external validation using real-world data	[33]
	Liquid biopsy (blood-based cfDNA fragmentome analysis)	Early detection of lung cancer in screening-eligible individuals	Machine learning classifier	Total: 958 individuals Training: 576 (181 cases, 395 controls) Validation: 382 (248 cases, 134 controls)	Pathologically confirmed lung cancer diagnosis (cases) vs. negative thoracic CT ± 12-month clinical follow-up (controls)	Overall (validation): 84% (95% CI: 79–88%)	Overall (validation): 53% (95% CI: 45–61%)	NR	Independent held-out cohort from the same prospective study (DELFI-L101, NCT04825834), but no truly external cohort used	Prospective case–control study with independent validation split	[34]
	Blood-based liquid biopsy (ctDNA methylation analysis via ELSA-seq)	Early detection of lung cancer	Soft-margin support vector machine classifier	Total: 569 individuals Cases: 308 lung cancer patients Controls: 261 age-/sex-matched non-cancer individuals	Cases: Pathologically confirmed lung cancer Controls: Negative thoracic CT ± 12-month clinical follow-up	Overall: 52–81% (stage-dependent) Subgroup (115 individuals): 100% sensitivity at 100% specificity	Overall: 96% (95% CI: 93–98%) Subgroup (115 individuals): 100% (95% CI: 91–100%)	Training/Validation: 0.93 Independent test set: 0.90	Independent cohorts from two hospitals (Peking Union Medical College Hospital and Shanghai Chest Hospital)	Prospective case–control study with independent validation cohorts	[35]
	Histopathology (H&E-stained whole-slide images)	1. Classification: LUAD vs. LUSC vs. normal 2. Mutation prediction: 6 oncogenes	Google Inception v3 (deep CNN)	TCGA: 1,634 slides (459 normal, 567 LuAD, 608 LUSC) Tiles: 987,931 (512 × 512 px)	TCGA consensus pathology diagnosis + NYU pathologist annotations (for external cohorts)	AI: 89%	AI: 93%	Classification (LUAD/LUSC/Normal): AUC: 0.97 (TCGA test set) Mutation prediction (LUAD): • AUC range: 0.733–0.856 (for 6/10 genes: STK11, EGFR, FAT1, SETBP1, KRAS, TP53)	Independent cohorts (NYU Langone Medical Center): Frozen sections (n = 98) FFPE sections (n = 140) Biopsies (n = 102) EGFR mutation prediction (n = 63)	Prospective validation in multiple external cohorts	[36]
Prostate cancer	Biparametric magnetic resonance imaging (bpMRI)	1. Prostate segmentation 2. Detection of clinically significant PCa (csPCa) lesions 3. Automated report generation	AutoProstate framework	Training: PROSTATEx (n = 204 patients, 299 Lesions: 76 csPCa, 223 low-grade or benign lesions) External validation: PICTURE (n = 247 patients, 210 Lesions: 147 csPCa, 63 low-grade or benign lesions)	PROSTATEx (train): MR-guided targeted biopsy + histopathology (csPCa: Gleason ≥ 3 + 4). Annotations by radiologists PICTURE (test): Transperineal Template Prostate-Mapping (TTPM) biopsy + MR-targeted biopsy (csPCa: Gleason ≥ 3 + 4). Annotations by radiologists using histopathology correlation	Lesion-Level (external validation): AI: 76% Radiologist: 78% (p > 0.05, not statistically significant)	Lesion-Level (external validation): AI: 57% Radiologist: 48% (p > 0.05, not statistically significant)	Lesion-Level (external validation): AI: 0.70 Radiologist: 0.64 (p > 0.05, not statistically significant)	PICTURE dataset (n = 247)	External validation using a prospective cohort from a different institution	[37]
	bpMRI	1. Detection of index lesions 2. Classification: PCa vs. benign; csPCa 3. PI-RADS feature-based explanations	Explainable AI (XAI) model	Internal: 1,224 patients (3,260 lesions) Training/validation: 1,108 patients (3,094 lesions) Internal test: 116 patients (166 lesions) External: PROSTATEx dataset (204 patients, 330 lesions)	Histopathologic analysis (biopsy/prostatectomy)	csPCa detection (patient-level): 93% (95% CI: 87–98%) Lesion-level: 77% (CI: 69–85%) for csPCa	Lesion-level: 89% (CI: 84–95%) for csPCa	Internal: 0.89 (CI: 0.85–0.93) for csPCa External (PROSTATEx): 0.87 (CI: 0.81–0.93) for csPCa	PROSTATEx dataset	Retrospective diagnostic study with external validation	[38]
	bpMRI and multiparametric MRI (mpMRI)	Detection and localization of csPCa (Gleason grade group ≥ 2)	Ensemble of 5 DL models	Total: 10,207 MRI exams (9,129 patients) Training: 9,207 exams Testing: 1,000 exams (including 197 external cases)	Histopathology (biopsy/prostatectomy) + ≥ 3-year follow-up (median 5 years)	vs. Radiologists (reader study): 89.4% (matched specificity) vs. Clinical practice: 96.1%	vs. Radiologists (reader study): 79.1% (50.4% fewer false positives) vs. Clinical practice: 68.9%	Reader study cohort: 0.91 (95% CI 0.87–0.94) Full testing cohort: 0.93 (0.91–0.94)	Testing cohort included 197 exams from St. Olav's Hospital, Norway (unseen centre)	Large retrospective diagnostic study with independent external validation	[39]
	Digital histopathology (whole-slide images of biopsies)	1. PCa detection 2. Gleason grading (ISUP Grade Groups 1–5)	15 DL algorithms (ensemble-based CNNs)	Development: 10,616 biopsies (2,113 patients) Internal validation: 545 biopsies External validation: US (741 biopsies), EU (330 biopsies)	Internal: Consensus of 3–4 uropathologists US external: majority vote of 6 uropathologists + immunohistochemistry EU external: single expert uropathologist	US external: 98.6% (95% CI: 97.6–99.3) Human pathologists (US): 91.9% (95% CI: 89.3–95.5)	US external: 75.0% (95% CI: 61.2–82.7) Human pathologists (US): 95.0% (95% CI: 87.4–98.1)	NR	US cohort: 741 biopsies from 3 sites EU cohort: 330 biopsies from Karolinska Hospital	Retrospective diagnostic study with independent external validation	[40]
	H&E-stained histopathology	Gleason grading of biopsies	Extended U-Net architecture	Internal: 5,759 biopsies (1,243 patients) External: 245 TMA cores	Consensus reference standard by three uropathology experts	95.4% (AI) vs. pathologists' median	95.2% (AI) vs. pathologists' median	0.990 (benign vs. malignant) 0.978 (Grade Group ≥ 2) 0.974 (Grade Group ≥ 3)	Tissue microarray set (public dataset from University Hospital Zurich): (n = 245 cores)	Retrospective diagnostic study with external validation	[41]
Brain cancer	mpMRI	Binary classification of glioma grades (HGG vs. LGG)	Three machine learning classifiers	285 patients (210 HGG, 75 LGG)	Histopathological grading confirmed by surgical specimens	95.08%	70.22%	0.903	None	Retrospective diagnostic study with cross-validation	[42]
	Multi-modal MRI (3D black-blood [BB] + 3D gradient echo [GRE])	Automatic detection and segmentation of brain metastases	3D U-Net with encoder-decoder architecture + reconstruction decoder for regularization	Training: 188 patients (917 lesions) Test: 45 patients (203 lesions) + 49 metastasis-free patients	Manual segmentation by radiologists, confirmed/modified by senior neuroradiologist	93.10%	69.4%	0.78 (0.85–0.91)	None	Retrospective diagnostic study without external validation	[43]
	mpMRI	Non-invasive prediction of 1p/19q co-deletion status	Deep CNN based on ResNet-34 architecture	Total: 555 patients Training: 330 Internal validation: 123 External testing (TCIA): 102	1p/19q status determined by fluorescence in situ hybridization (FISH) on surgical tissue samples	Testing cohort: 93.0% (95% CI: 89.43–96.03)	Testing cohort: 93.6% (95% CI: 90.45–96.46)	Training: 0.999 Internal validation: 0.986 Testing (external): 0.983 (95% CI: 0.9110–0.9549)	the cancer imaging archive (TCIA) dataset (n = 102)	Retrospective multi-center validation with external cohort	[44]
Pan-Cancer (BRCA, COAD, LUNG, LIHC, PRAD)	cfDNA methylation profiling	1. Early cancer detection 2. Tumor localization (tissue-of-origin)	SRFD-Bayes (Semi-reference-free deconvolution + Bayesian diagnostic model)	Training: 207 normal controls + 113 late-stage cancer patients Validation: 207 normal controls + 79 early-stage cancer patients + 191 pre-diagnosis samples	Core needle biopsy pathology	AI: 86.1% (early detection)	AI: 94.7% (early detection)	Close to 0.98 (SRFD-Bayes for cancer detection)	Three independent cohorts: 1. GSE129374 (21 cirrhosis vs. 22 HCC) 2. HCC cohort (n = 1,050, stages I–IV) 3. Pre-diagnosis cohort (n = 191)	Retrospective diagnostic study	[45]

Abbreviation: ADH Atypical ductal hyperplasia, AI Artificial intelligence, ALI Angiolymphatic invasion, BRCA Breast invasive carcinoma, BUSI Breast Ultrasound Images dataset, CADe computer-aided detection, CADx Computer-aided diagnosis, cfDNA cell-free DNA, CI Confidence interval, CNN Convolutional neural network, COAD Colon adenocarcinoma, ctDNA circulating tumor DNA, CXR Chest radiography, DBT Digital breast tomosynthesis, DCIS Ductal carcinoma in situ, DG DeepGrade, DL Deep learning, DLCST Danish Lung Cancer Screening Trial, DM Digital mammography, EDL-BC Ensemble deep learning for breast cancer, FPR False-positive rate, HCC Hepatocellular carcinoma, HES Hematoxylin–eosin-saffron, HGG High-grade gliomas, H&E Hematoxylin and eosin, ISUP International Society of Urological Pathology, LGG Low-grade gliomas, LIHC Liver hepatocellular carcinoma, LUAD lung adenocarcinoma, LUMAS Lung malignancy scores, LUNG Lung squamous cell carcinoma & lung adenocarcinoma, Lung-RADS Lung Imaging Reporting and Data System, LUSC squamous cell carcinoma, MCC Medical checkup cohort, MSP Maximum suspicion projection, NLST National Lung Cancer Screening Trial, NR Not reported, NYU New York University, OMI-DB Optimam mammography imaging database, PCa Prostate cancer, SIFT Scale-invariant feature transform, SLN Sentinel lymph node, SVM Support vector machine, TILs Tumor-infiltrating lymphocytes, TMA Tissue microarray, UK United Kingdom, US United States, 2D Two-dimensional

Colorectal cancer

Colonoscopy remains the gold-standard for colorectal cancer (CRC) screening, though its efficacy varies with the operator's expertise, resulting in variability in adenoma detection rates and missed lesions that can contribute to the incidence of interval cancers [46–48]. To address these challenges, AI has emerged as a transformative breakthrough in colonoscopy, using DL and ML techniques to improve real-time decision-making by extensive, annotated datasets of colonoscopic images to train deep neural networks (DNNs), thereby enabling automated polyp detection, classification, and quality assessment [49, 50]. For example, Zhou et al. [17] introduced CRCNet, a DL model designed for the detection of CRC within endoscopic images, demonstrating high performance across three independent datasets. Multiple companies have achieved the Food and Drug Administration (FDA) clearance or the European Union (EU) certification for their computer-aided detection (CADe) systems designed to identify polyps in colonoscopy images (e.g., K211951, K223473), and a series of randomized controlled trials have demonstrated compelling evidence of these technologies’ clinical efficacy and potential to enhance diagnostic accuracy [51–56]. However, Mangas-Sanjuan et al. [57] designed a randomized controlled trial primarily to assess whether computer-assisted colonoscopy increases the detection of advanced colorectal neoplasias in patients with positive fecal immunochemical test (FIT) results in organized CRC screening programs. They concluded that CADe did not improve the colonoscopic identification of advanced colorectal neoplasias. In a meta-analysis of data from 18,232 patients across 21 randomized trials, the investigators found that using CADe for polyp detection during colonoscopy increases the detection of adenomas but not advanced adenomas, and leads to higher rates of unnecessary removal of nonneoplastic polyps [58]. The importance of detecting and removing these small growths, also known as polyps, still sparks debate [59].

AI-driven computer-aided diagnosis (CADx) systems employ advanced imaging analytics to distinguish benign from malignant lesions, and the integration of AI-derived histopathological predictions may serve as a valuable tool for improving diagnostic accuracy [46, 60]. An initial study comparing real-time image recognition system analysis with narrow-band imaging diagnosis, along with assessing the correlation between image analysis and pathological results, showed promising results, achieving approximately 90% accuracy and a negative predictive value (NPV) over 90% [18]. Data from the past five years offer clues on the NPV of CADx for neoplastic histologic prediction in small rectosigmoid polyps [61–64]. In three studies, CADx achieved an NPV of 90% or higher, meeting the performance threshold proposed by the American Society for Gastrointestinal Endoscopy (ASGE) for allowing the 'do not resect' strategy [65]. Research published in 2024 suggested that autonomous AI-based diagnosis has noninferior accuracy to endoscopist-based diagnosis [19]. Both autonomous AI and AI-assisted human (AI-H) showed relatively low accuracy for optical diagnosis. However, autonomous AI achieved higher agreement with pathology-based surveillance intervals [19]. Impressively, Hassan et al. [66] conducted a meta-analysis involving 7400 diminutive polyps, 3769 patients, and 185 endoscopists from 11 studies. They found that CADx did not provide benefit or harm for the resect-and-discard strategy, questioning its value in clinical practice. Improving the accuracy and explainability of CADx is desired.

In terms of histological diagnosis, Graham et al. [67] designed a comprehensive CNN architecture that preserves maximum information during feature extraction, which is crucial for successful gland instance segmentation in colon histology images. They also emphasized the generalizability of their method by processing whole-slide images (WSIs) from a different center with high accuracy. Zhao et al. [68] developed a DL model to quantify the tumor-stroma ratio (TSR) based on histologic WSIs of CRC and demonstrated its prognostic validity for patient stratification of overall survival (OS) in two independent CRC patient cohorts. This fully automatic method allows for objective and standardized assessment while reducing pathologists' workload.

Breast cancer

The application of computer-assisted techniques in enhancing the accuracy of medical imaging for breast cancer (BC) screening and detection has a storied history. Notably, CADe received approval from the FDA in 1998 for X-ray mammography, subsequently leading to its widespread integration into clinical practice [69]. However, CADe has been unable to fully meet the increasing demands for better mammographic performance due to limitations such as a high incidence of false positives and elevated recall rates [70]. Meanwhile, in recent years, the growing use of AI has transformed the capabilities of automated BC detection through mammographic interpretation [71].

In 2020, McKinney et al. [20] elegantly introduced an AI system that outperforms radiologists in a clinically relevant task of BC identification. Their system was trained and tested on two-dimensional (2D) mammograms from the United Kingdom (UK) and the United States (US), demonstrating its ability to generalize from training on UK data to testing on data collected from a US clinical site. Following this important study, Lotter et al. [21] developed a DL approach that effectively uses both strongly and weakly labeled data by gradually training in stages while keeping localization-based interpretability. Their approach also extends to 3D mammography, which is especially important given its increasing use as a primary screening method and the additional time needed for interpreting it. There are now several FDA-cleared AI products designed to aid radiologists in the detection of BC from mammograms (K220105, K211541, K200905) and prospective study in Sweden is also under way to assess the clinical utility of these products in real-world healthcare settings [72–75]. Furthermore, Yala et al. [76] developed the Mirai system, capable of predicting future five-year BC risk directly from mammograms, and it has been retrospectively validated across multiple hospitals. Vachon et al. [77] demonstrated that AI imaging algorithms not only enhance the detection of BC on mammography but also possess the potential to inform long-term risk prediction of invasive BC.

In addition to X-ray mammography, the availability of unique features and comprehensive imaging datasets such as ultrasound, magnetic resonance imaging (MRI), and positron emission tomography/computed tomography (PET/CT) offers opportunities for developing clinically impactful AI applications. Shen et al. [22] introduced a radiologist-level AI system that can automatically identify malignant lesions in breast ultrasound images. By validating its performance on an external dataset, they also provided initial results supporting its ability to generalize across a patient cohort with different demographic composition and image acquisition protocols. Moreover, ensemble DL models can detect subtle features in BC lesion images, thereby improving both diagnostic accuracy and efficiency in ultrasound interpretation, while the integration of ultrasound and elastography through DL techniques holds promise for more precise prediction of axillary lymph node metastasis, potentially reducing false positives and unnecessary biopsies [23, 24]. Additionally, FDA-approved AI algorithms have been developed to facilitate the interpretation of MRI (DEN1700, RRID:SCR_012945) and breast ultrasound examinations (K190442, K210670, P150043; RRID:SCR_012945) [78].

In recent years, liquid biopsy of body fluid samples, which include a diverse range of tumor-derived components, has gained considerable attention and momentum for the detection and characterization of BC. Zhou and colleagues [45] evaluated the application of circulating cell‐free DNA (cfDNA) methylation profiling for the characterization of tumor constituents and the detection of nascent neoplasms, utilizing a semi-reference deconvolution algorithm that integrates tumor scores and ML models, ultimately achieving high sensitivity and specificity in early BC detection.

The advent of innovative pathological techniques has yielded more comprehensive and intricate large-scale datasets encompassing breast fine-needle aspiration samples and tissue specimens [79]. For example, Sandbank et al. [25] introduced a DL model capable of classifying invasive and noninvasive BC subtypes and generating predicted clinical and morphological features, which was validated using external datasets from the Institut Curie and subsequently piloted as a second-reader diagnostic system at Maccabi Healthcare Services in Israel. Wang et al. [26] developed and validated a novel method, DeepGrade, for histological grading of BC, focused on re-stratification of the Nottingham histological grade (NHG) 2 cases. The approach provides a cost-effective alternative to molecular profiling to extract information relevant for clinical decisions. Another key focus in this field has been the detection of metastatic lesions in sentinel lymph nodes, given that treatment strategies in BC are frequently contingent upon the identification of such lesions [80]. For instance, Google [27] developed the LYmph Node Assistant (LYNA), and reported that pathologists employing this system achieved improved accuracy and efficiency in detecting micrometastases while reducing their review time. Challa et al. [28] reported that the Visiopharm Integrator System (VIS) metastasis AI algorithm demonstrated 100% sensitivity and NPV in detecting lymph node metastasis with readily recognized false negative causes and less time consumed during pathologists' review compared with immunohistochemistry (IHC) slides. This suggests its potential as a useful screening tool in routine clinical digital pathology workflows to improve efficiency. Recently, the Paige Lymph Node, an AI-assisted diagnostic tool developed by the company Paige, has received FDA Breakthrough Device Designation, recognizing its potential to aid pathologists in the detection of BC metastases within lymph node tissue [81]. In addition, quantitative assessment of tissue biomarkers, including Ki-67 and human epidermal growth factor receptor 2 (HER2), is essential for BC evaluation. However, the interpretation of these markers remains influenced by inherent subjectivity [82, 83]. Data from several studies demonstrated that AI-assisted assessment of Ki-67 could achieve a lower mean error [84] and a lower standard error deviation [85]. As another example, Dy et al. reported AI's potential to standardize Ki-67 scoring, especially between 5 and 30% proliferation index [86]. In 2025, DAN Albuquerque et al. [87] conducted a diagnostic meta-analysis to assess AI’s ability to classify HER2 IHC scores. They found that AI holds promising potential in accurately identifying HER2-low patients and excels in distinguishing 2 + and 3 + scores.

Lung cancer

AI-assisted interpretation for chest radiography has been one of the earliest applications of AI in medical imaging, as detecting abnormalities, especially small pulmonary nodules, remains a challenging task even for experienced thoracic radiologists [88]. Early studies demonstrated that AI achieved radiologist-level or superior performance in detecting various abnormalities on chest radiographs, and radiologists' detection accuracy was further improved when assisted by AI [89, 90]. In recent years, the real-world impact of AI CADe tools on interpretation accuracy has started to be documented. For example, a retrospective diagnostic cohort study [29] documented that a DL–based CADe system improved the diagnostic yield for newly visible metastasis on chest radiographs in patients with cancer with a similar false-referral rate. Nam et al. [30] conducted a randomized controlled trial involving patients undergoing chest radiography for health check-ups, demonstrating that AI CADe-assisted reading improved the detection rate of actionable lung nodules.

Currently, low-dose CT (LDCT) screening is the only exam currently proven to reduce mortality from lung cancer (LC) [91, 92]. A variety of software devices have been approved by the FDA to improve workflow efficiency and performance through enhanced detection of lung nodules [92–94]. In addition, assessing the risk of malignancy in detected lung nodules is another crucial aspect of comprehensive lung nodule evaluation. Impressively, Ardila et al. [31] employed advanced DL techniques to develop models with cutting-edge performance by utilizing full 3D LDCT volumes, pathology-confirmed case results, and previous volumes. If clinically validated, these models could assist clinicians in both localization and lung cancer risk assessment tasks. Venkadesh et al. [32] illustrated that DL–based models can predict the malignancy risk of pulmonary nodules with greater accuracy than radiologists' interpretations or existing risk prediction models. Wang et al. [33] presented the Chinese Lung Nodules Reporting and Data System (C-Lung-RADS), a multiphase approach to evaluate the malignancy risk of pulmonary nodules, improving early lung cancer detection while optimizing healthcare resources.

For blood-based liquid biopsy, Mazzone et al. [34] described the development and validation of a new blood-based LC screening test that uses a highly affordable, low-coverage genome-wide sequencing platform to analyze cfDNA fragmentation patterns. The test could increase LC screening rates leading to substantial public health benefits. Furthermore, Liang et al. [35] demonstrated that deep methylation sequencing, combined with a ML classifier analyzing methylation patterns, facilitates the detection of circulating tumour DNA (ctDNA) at dilution ratios as low as 1 in 10,000, thereby providing advantages in cancer screening and the assessment of treatment efficacy.

In histologic diagnosis of non–small cell lung cancer (NSCLC), such as distinguishing between lung adenocarcinoma (LUAD) and lung squamous carcinoma (LUSC), the process is relatively straightforward, and AI has the potential to support pathologists in the near future by handling these routine tasks. In line with this notion, there are already documented instances of AI-based approaches capable of effectively distinguishing LUAD from LUSC using hematoxylin and eosin (H&E) slides [95, 96]. Impressively, Coudray et al. [36] developed a DL model trained on The Cancer Genome Atlas (TCGA) H&E images to classify NSCLC subtypes, and their study was also among the first to predict mutation status of key driver genes directly from H&E images. Lu et al. [97] introduced a method named clustering-constrained-attention multiple-instance learning (CLAM) that can localize well-known morphological features on WSIs without needing spatial labels. It outperforms standard weakly supervised classification algorithms and is adaptable to independent test cohorts, smartphone microscopy, and varying tissue content. As another example, Wang et al. [98] developed an automated cell type classification pipeline, ConvPath, which includes nuclei segmentation, CNN-based classification of tumor cells, stromal cells, and lymphocytes, as well as extraction of tumor microenvironment-related features for LC pathology images. However, rare cases such as various neuroendocrine carcinoma subtypes frequently require additional time and specialized expertise from pathologists to achieve accurate diagnoses, and currently, the utility of clinical AI tools in these complex and atypical diagnostic scenarios remains questionable [99].

Prostate cancer

According to international guidelines, multiparametric MRI (mpMRI) of the prostate is recommended for radiologists to effectively identify lesions harboring clinically significant prostate cancer (csPCa) prior to conducting confirmatory pathological biopsies [100]. Various AI-powered methodologies have been developed to facilitate the analysis of MRI images for the detection, staging, and segmentation of PCa [101, 102]. For example, Mehta et al. [37] introduced AutoProstate, which uses patient data and biparametric MRI (bpMRI) to generate an automatic web-based report that includes segmentations of the entire prostate, prostatic zones, and potential csPCa lesions. It also presents several derived characteristics with clinical value. Hamm et al. [38] developed an explainable AI (XAI) model for detecting and classifying csPCa. The model improved confidence and reduced reading time for nonexperts while providing visual and textual explanations using well-established imaging features. Additionally, there are several prostate segmentation AI algorithms available on commercial platforms for users. All of these tools have received approval for use in the US by the FDA, including Prostate MR® (Siemens) [103], Quantib Prostate® (Quantib) [104], OnQ Prostate® (Cortechs.ai) [105], PROView® (GE Medical Systems) [106], and qp-Prostate® (Quibim) [107].

Various concurrent AI data challenges are currently ongoing. Among them, the Prostate Imaging–Cancer AI (PI-CAI) challenge stands out as a new grand challenge, featuring over 10,000 carefully curated prostate MRI exams to validate modern AI algorithms and assess radiologists’ performance at csPCa detection and diagnosis [39]. The prostate cancer grade assessment (PANDA) framework exemplifies this approach through its two-phase study design (development and validation) [40]. In the validation phase, submitted algorithms underwent independent evaluation by expert pathologists under strictly blinded conditions to ensure unbiased assessment. These initiatives establish a transparent and rigorous framework for the evaluation and benchmarking of AI algorithms, thereby fostering the progression of future algorithmic innovations.

The prevailing standard for PCa diagnosis entails the histopathological examination of biopsy specimens. Currently, numerous studies have demonstrated encouraging outcomes in the application of DL for autonomous cancer detection within digital WSIs of prostate biopsy specimens [108–110]. In 2019, Campanella et al. [108] introduced a multiple instance learning-based DL system that uses only the reported diagnoses as labels for training, thereby avoiding costly and time-consuming pixel-wise manual annotations. This model was further examined by other researchers and was later named Paige Prostate, which has received FDA approval for clinical use in the automated detection of PCa in core needle biopsies [111–113]. Beyond the initial diagnosis of PCa, DL–based approaches have been extensively investigated to improve the precision of Gleason grading, a method for determining the aggressiveness of PCa based on microscopic examination of tissue samples. For instance, Bulten et al. [41] demonstrated that an automated DL system achieved performance similar to that of pathologists in Gleason grading and could potentially aid in PCa diagnosis. The system might assist pathologists by screening biopsies, offering second opinions on grade groups, and providing quantitative measurements of volume percentages. Similarly, Nagpal et al. [114] developed a DL model that could improve the accuracy of Gleason scoring and subsequent therapy decisions, particularly where specialist expertise is unavailable.

Brain cancer

Brain tumors are frequently initially characterized by MRI before further diagnosed through histopathological examination and AI-driven tools have the potential to assist neuroradiologists in lesion detection and differential diagnosis [115, 116]. For example, Cho et al. [42] showed that glioma grades could be accurately determined by combining high-dimensional imaging features, an advanced feature selection method, and ML classifiers. Park et al. [43] demonstrated that DL-based models markedly improve both the detection and segmentation of brain metastases, with a particular emphasis on accurately identifying small metastases. In a retrospective multi-center study, researchers [44] found that a deep CNN model built from preoperative mpMRI could predict the 1p/19q status in lower-grade gliomas (LGG) patients with high accuracy, sensitivity, and specificity. The imaging-based DL has the potential to serve as a noninvasive tool for predicting key molecular markers in adult diffuse gliomas.

In terms of pathological diagnosis, CNNs trained on WSIs of gliomas have been utilized to deliver unbiased diagnoses of gliomas. For instance, Ertosun et al. [117] trained two CNNs on publicly available H&E-stained images of gliomas from TCGA. One CNN was designed to differentiate glioblastoma (GBM) from LGG, while the other aimed to distinguish between grade 2 and grade 3 gliomas. Li et al. [118] used DL on H&E-stained histopathology images to classify central nervous system (CNS) tumors and to assess biomarkers including IDH1 mutation and p53 mutation. While conventional methods that include imaging and tissue biopsies reliably identify many brain tumors, there are exceptions such as high-grade astrocytoma with piloid features, which was introduced in the 2021 WHO classification [119]. This condition requires methylome profiling for accurate diagnosis [120]. To address this, Vermeulen et al. [121] evaluated the use of rapid nanopore sequencing combined with ML to improve intraoperative diagnosis of CNS tumors, developing 'Sturgeon', a neural network trained to subclassify CNS tumors during surgery using sparse methylation profiles obtained through nanopore sequencing. Similarly, Hoang et al. [122] developed Deep lEarning from histoPathoLOgy and methYlation (DEPLOY), a DL model that classifies CNS tumors into ten major categories from histopathology within a clinically relevant short time frame.

Facilitating precision cancer treatment

While cancer diagnosis and classification remain crucial for informing patient treatment, emerging AI algorithms are increasingly being developed to directly enhance therapeutic interventions by assisting in treatment selection, designing personalized treatments, and guiding during treatment delivery. This form of cancer care often involves analyzing a large amount of data with advanced computational approaches to help clinicians in decision-making and to facilitate the evaluation of biomarkers for prognostic purposes.

Colorectal cancer

Surgical intervention remains the principal and most efficacious approach for the management of patients with CRC. Advancements in AI have ushered in a new era in CRC surgery, exemplified by the substantial progress achieved with the da Vinci surgical system, which has the potential to enhance surgical precision, visualization, and surgeon ergonomics, thereby reducing tissue trauma, accelerating recovery, and decreasing complication rates [123, 124].

AI tools are being increasingly integrated into clinical workflows to optimize the radiotherapy treatment process. For instance, AI integration in magnetic resonance-guided radiation therapy (MRgRT) represents a significant leap forward in oncologic treatment by utilizing real-time MRI for highly precise visualization of tumors and adjacent structures. It is especially advantageous in the treatment of complex soft tissue neoplasms, delivering enhanced image resolution and precise localization that significantly exceed the standards set by conventional imaging-guided radiation therapy [125].

Prognostic AI models for CRC have been extensively developed utilizing various data modalities such as histopathological analysis [126, 127] and multiplex imaging approaches [128–130]. For example, DoMore Diagnostics company [131, 132] has a conformité européenne (CE)-marked product that predicts CRC prognosis from H&E slides, and several validation studies of histopathology-based prognostic models have been conducted, including research demonstrating that a prognostic feature initially identified through AI analysis can be effectively learned and applied by pathologists [133, 134]. In addition to directly predicting clinical outcomes, several studies have employed AI to forecast previously characterized prognostic and predictive biomarkers, with emphasis on the development and validation of models for predicting microsatellite instability (MSI), a key biomarker associated with treatment response and clinical prognosis in immunotherapy [135, 136]. This application is increasingly approaching commercialization, exemplified by Owkin’s CE-marked product that predicts MSI directly from H&E images [137].

Breast cancer

AI is gradually revolutionizing the surgical landscape in BC management, offering advancements in oncology aesthetics, preoperative planning, intraoperative guidance, and postoperative assessment. Pfob et al. [138] utilized ML algorithms to predict breast satisfaction during follow-up in women contemplating mastectomy and reconstruction as part of their BC treatment strategy, which provided a personalized reference. In addition, ensuring clear margins is crucial to prevent the recurrence of BC in breast-conserving surgery. Kothari et al. [139] reported that combining Laser Raman spectroscopy (LRS) with two ML algorithms offers rapid, quantitative, and probabilistic tumor assessment with real-time error analysis. This approach can detect cellular changes characteristic of cancer tissue in vivo during surgery, enabling real-time margin evaluation.

Advancements in AI are accelerating the evolution of drug therapies for BC treatment. Park et al. [140] developed an interpretable DL model to predict responses to palbociclib, a cyclin-dependent kinase 4/6 inhibitor (CDK4/6i) used in BC therapy, based on a reference map of multiprotein assemblies in cancer. This study provides an integrated assessment of how a tumor’s genetic profile influences resistance to CDK4/6 inhibitors. Moreover, Sammut et al. [141] demonstrated that ML models integrating clinical, genomic, and transcriptomic data from patients undergoing chemotherapy or targeted therapy substantially outperform those relying solely on clinical variables in predicting BC outcomes, with the high accuracy observed in external validation indicating their robustness and potential to inform therapeutic decision-making in future clinical trials.

In addition to surgical intervention and drug therapy, numerous AI-driven predictive models grounded in histological analysis have been developed to augment clinical decision-making processes. For example, Ogier du Terrail et al. [142] demonstrated that federated learning enables collaborative multi-center training of ML models on WSIs to predict histological response to neoadjuvant chemotherapy in triple-negative breast cancer (TNBC), outperforming local models and clinical baselines while identifying predictive features such as tumor-infiltrating lymphocytes (TILs), apocrine tumor cells, and fibrosis through interpretability techniques. In another study, Amgad et al. [143] introduced the Histomic Prognostic Signature (HiPS), an interpretable and comprehensive scoring system that assesses survival risk based on the morphological characteristics of the BC microenvironment. HiPS leverages DL techniques to precisely map cellular and tissue architectures, enabling the quantification of features related to epithelial, stromal, immune components, and their spatial interactions. Additional investigations have employed AI techniques on H&E images to directly predict the presence of TILs, a prognostic and predictive biomarker in BC, as well as the expression status of programmed death ligand-1 (PD-L1), a key biomarker for immunotherapy response [144–146]. Furthermore, DL can directly determine molecular biomarker status from routine histology. In line with notion, Shamai et al. [147] developed a DL system termed morphological-based molecular profiling (MBMP) that predicts BC molecular biomarker expression, such as estrogen receptor (ER), progesterone receptor (PR), and HER2, directly from standard H&E-stained tissue images. Results suggest tissue morphology encodes molecular information, providing a potentially faster, cheaper alternative to IHC for biomarker profiling in many patients.

Lung cancer

The application of AI in non-invasive liquid biopsy for personalized treatment of LC has seen remarkable advancements. Assaf et al. [148] demonstrated that alterations in ctDNA, modeled within a ML framework and validated across both a hold-out test set and an external cohort of NSCLC patients, can enhance patient risk stratification and enable the sensitive detection of treatment arm differences at early stages within clinical trial contexts. Widman et al. [149] presented a ML-guided whole-genome sequencing platform for ctDNA single-nucleotide variant detection that enables plasma-only (non-tumor-informed) disease monitoring in advanced LC, providing clinically valuable tumor fraction assessments for patients undergoing immune checkpoint inhibition. Heeke et al. [150] introduced minimal residual disease (MRD)-EDGE, a ML-enhanced plasma whole-genome sequencing platform that improves ctDNA detection sensitivity by approximately 300-fold for single-nucleotide variants (SNVs) and reduces the required aneuploidy for copy-number variant (CNV) detection from 1 Gb to 200 Mb. This enables ultrasensitive monitoring of tumor burden, MRD, immunotherapy response, and even ctDNA shedding from precancerous lesions.

Additionally, numerous studies have highlighted the potential of AI to predict immunotherapy biomarkers from pathological data in patients with NSCLC, including PD-L1 expression, and TILs. For example, Park et al. [151] developed an AI-powered spatial analyzer of TILs in H&E images, which correlates with tumor response and progression-free survival in patients with advanced NSCLC undergoing ICI therapy, potentially serving as a supplementary biomarker to the tumor proportion score evaluated by a pathologist. This product is currently CE-marked and approved for quantifying PD-L1 expression by the AI startup Lunit [152]. Vanguri et al. [153] developed a multimodal ML model (DyAM) integrating radiology (CT scans), pathology (digitized PD-L1 slides), and genomics to predict response to PD-(L)1 blockade immunotherapy in advanced NSCLC patients.

Furthermore, AI has revolutionized the identification and analysis of genomic profiles, playing a vital role in forecasting disease prognosis, treatment responses, and survival rates. For instance, Wang et al. [154] proposed a fully automated artificial intelligence system (FAIS), which offers a non-invasive method to detect epidermal growth factor receptor (EGFR) genotype and identify patients with an EGFR mutation at high risk of tyrosine kinase inhibitor (TKI) resistance. The superior performance of FAIS over tumour-based DL methods indicates that genotype and prognostic information can be obtained from the whole lung instead of only tumour tissues. Apart from the driver mutation EGFR, Rakaee et al. [155] proposed that ML-based immune phenotyping, which analyzes the spatial distribution of T cells in resected NSCLC, can identify patients at greater risk of disease recurrence after surgical resection. Specifically, LUADs with concurrent KEAP1 and STK11 mutations are enriched for altered and desert immune phenotypes. Besides, Ricciuti et al. [156] highlighted the genomic and immunophenotypic heterogeneity of immune checkpoint inhibitor (ICI) resistance in patients with NSCLC, utilizing comprehensive tumor genomic profiling and ML-based assessment of TILs.

Prostate cancer

AI systems can also enhance situational awareness, optimize surgical approaches, and improve patient outcomes during robotic-assisted radical prostatectomy (RARP) [157]. For instance, automated performance metrics collected by the Da Vinci surgical robot have been utilized to predict postoperative length of stay following RARP [158]. AI also helps integrate interactive 3D imaging systems for preoperative and intraoperative planning. In a prospective study, a real-time 3D augmented reality system was developed to identify prostate lesions at the neurovascular bundle level, thereby enabling enhanced nerve-sparing techniques during RARP [159].

AI has also been extensively investigated for non-surgical treatment planning in PCa. Data from a prospective study demonstrated that fully automated, ML-generated therapeutics are realizable in a clinical environment. ML delivers reproducible, high-performance radiation therapy treatments for patients with PCa and provides time savings that allow for better reallocation of human resources [160]. Nouranian et al. [161] developed an advanced ML-based multi-label segmentation algorithm aimed at providing rapid and clinically relevant segmentations for seed implantation planning in low-dose-rate prostate brachytherapy, a treatment modality involving the placement of small radioactive seeds within or adjacent to the prostate gland.

Choosing the best therapy for a PCa patient is challenging, as oncologists must find a treatment with the highest chance of success and the lowest risk of side effects. International guidelines for predicting outcomes rely on non-specific and semi-quantitative tools, often resulting in over- and under-treatment [162]. To address this, Esteva et al. [163] developed a multimodal deep learning model (MMAI) using clinical data and digital histopathology images from randomized trials to predict long-term PCa outcomes (e.g., metastasis, survival). The MMAI outperformed standard risk stratification from the National Comprehensive Cancer Network (NCCN) across all endpoints, offering a globally accessible tool for therapy personalization through digital pathology. Furthermore, Parker and colleagues [164] aimed to evaluate whether the MMAI algorithm could predict outcomes in very advanced PCa using data from four phase 3 trials of the STAMPEDE platform protocol. They found that diagnostic prostate biopsy samples contain prognostic information in patients with, or at high-risk of, radiologically overt metastatic PCa. The MMAI algorithm combined with disease burden improves prognostication of advanced PCa.

Efforts are also ongoing to leverage AI for enhancing PCa prognostication. For example, Elmarakeby et al. [165] elegantly showed P-NET is a sparse DL model that analyzes molecular profiling data within a biologically informed, pathway-driven framework to predict PCa states such as metastasis, with the resulting prediction scores demonstrated to independently correlate with patient prognosis. Kartasalo et al. [166] developed an AI model utilizing H&E images to detect perineural invasion, a critical prognostic marker associated with adverse outcomes in PCa. Additionally, ArteraAI, an AI-focused company, has developed and validated an AI-based predictive model that can identify patients with a predominantly intermediate risk for prostate cancer who are likely to benefit from short-term androgen deprivation therapy (ADT) [167]. This innovative technology has been commercialized as the ArteraAI Prostate Test and is now available for clinical use through a Clinical Laboratory Improvement Amendments (CLIA)-certified laboratory.

Brain cancer

Approximately 96% of patients with GBM have isocitrate dehydrogenase-1 (IDH1) wildtype mutations, and the treatment success rate for these patients (i.e., concomitant adjuvant temozolomide (TMZ) therapy) can be predicted via the O6-methylguanine-DNA methyltransferase (MGMT) gene promoter [168]. To address this, Le et al. [168] developed a radiomics model using XGBoost and F-score feature selection to non-invasively classify MGMT promoter methylation status in IDH1 wildtype GBM patients. Using nine key MRI-derived radiomics features, the model achieved high accuracy outperforming other methods and potentially supporting treatment planning. Furthermore, Do et al. [169] proposed a hybrid ML feature selection model to identify the most informative radiomics feature set and meticulously evaluate its capability of accurately classifying MRI images into methylated and unmethylated ones.

AI-based methods excel at predicting therapy responses, facilitating more effective treatment planning. Kawahara et al. [170] reported that AI techniques can predict responses to gamma knife radiosurgery for metastatic brain tumors by utilizing radiomic features extracted from contrast-enhanced MRI scans. There are also several studies in which radiomics features were used to improve the accuracy of distinguishing necrosis from tumor progression and to enable early detection of adverse radiation events after radiotherapy of brain tumors [171–173].

In brain tumor care, AI plays a crucial role in enhancing prognostic capabilities, enabling more accurate predictions of disease progression and patient outcomes. For instance, Macyszyn et al. [174] integrated diverse imaging markers to accurately predict patient survival and classify GBM molecular subtypes using ML on preoperative MRI, revealing distinctive radiographic phenotypes that can improve diagnosis and guide treatment without additional invasive testing. Zheng et al. [175] presented a scalable approach to elucidate the transcriptional heterogeneity of GBM and establish a vital link between spatial cellular architecture and clinical outcomes.

Improving cancer surveillance

Beyond individual patient care, AI plays a crucial role in population-level cancer surveillance, enabling more efficient data collection, analysis, and risk prediction to inform public health strategies. Cancer surveillance involves the continuous collection and analysis of patient data and epidemiological statistics. AI techniques are increasingly used to speed up the extraction of relevant information for surveillance reports and to discern meaningful patterns within population-level cancer datasets. For example, a collaboration between the National Cancer Institute (NCI) and the department of energy, known as Modeling Outcomes using Surveillance data and Scalable Artificial Intelligence for Cancer (MOSSAIC), employs advanced AI methodologies to expedite the submission of data to NCI’s Surveillance, Epidemiology, and End Results (SEER) program, thereby enhancing the efficiency and timeliness of cancer data reporting [176]. As part of this initiative, Alawad et al. [177] developed sophisticated AI algorithms capable of automatically extracting tumor characteristics from unstructured clinical narratives, thereby saving thousands of hours of manual effort. Chandrashekar et al. [178] developed a long-sequence AI transformer called Path-BigBird, which extracts data from six SEER cancer registries to provide cancer researchers and clinicians with more accurate information on cancer diagnosis and management, supported by the MOSSAIC initiative. Moreover, NCI-supported researchers are harnessing DL algorithms trained on population-scale disease data to predict individual risk for pancreatic cancer, thereby laying the groundwork for earlier detection and improved patient outcomes [179].

Additionally, LLMs used for EHR surveillance are assisting researchers in gaining deeper insights into social determinants of health, which are potentially vital for the prevention, early detection, and effective treatment of cancer [180]. Pan et al. [181] proposed that an LLM-based pipeline can facilitate the interpretation of EHR notes without needing manually curated labels, thereby enabling comprehensive and real-time disease surveillance.

Revolutionizing cancer drug discovery

AI is also transforming the upstream process of cancer drug development, accelerating the identification of novel targets, therapeutic candidates, and optimization strategies.

Enhancing target identification and validation

AI-driven computational predictions of protein structures can substantially advance structure-based drug design, facilitating the development of novel therapeutic targets. Recent advancements in DL-based AI models have revolutionized the prediction of protein three-dimensional structures, exemplified by AlphaFold [6], the modeling of molecular interactions such as those achieved by AlphaFold 3 [182], and the simulation of dynamic processes including protein folding and unfolding, as demonstrated by AI²BMD [183]. In addition, target validation through the utilization of cellular and animal models constitutes a critical phase in the target discovery process, and an expanding array of AI-identified candidates are moving toward successful validation. For example, Ren et al. [184] reported that a highly potent small-molecule inhibitor, designed using generative AI, exhibited selective antiproliferative activity in a hepatocellular carcinoma cell line.

Accelerating drug discovery

AI holds the potential to markedly compress drug discovery timelines and reduce costs relative to conventional methodologies. Impressively, Zhavoronkov et al. [185] developed a deep generative model called generative tensorial reinforcement Learning model for de novo small-molecule design, which optimizes synthetic feasibility, novelty, and biological activity, and used it to discover a potent TKI within Just 21 days. Ren et al. [184] combined AlphaFold-predicted structures with generative AI (Chemistry42) to identify a novel CDK20 inhibitor for hepatocellular carcinoma. Within 30 days of target selection, their first hit molecule (ISM042-2–001) was validated after synthesizing only seven compounds. A second AI-driven optimization round yielded ISM042-2–048 within 60 total days. The entire process synthesized only 13 synthesized compounds, highlighting a significant improvement over traditional drug discovery, which typically takes 10–15 years and costs up to $2.8 billion [186].

Beyond de novo design, AI-enabled drug repurposing has emerged as a complementary strategy to further compress timelines and reduce risk. For example, Tran et al. [187] used AI (Standigm Insight™) to rapidly repurpose antiviral drug Z29077885 as a novel serine/threonine kinase 33 (STK33)-targeting anticancer agent. By leveraging its existing safety profile, this approach bypassed early-stage development costs, with efficacy validated through in vitro/in vivo studies. Recently, Abdel-Rehim et al. [188] utilized GPT-4 to identify unexpected combinations of everyday drugs that could help treat BC. One example is the combination of simvastatin (commonly used to lower cholesterol) and disulfiram (used in alcohol dependence), which showed an inhibitory effect on BC cells. By focusing on affordable, FDA-approved drugs not typically associated with cancer, these treatments have a higher potential to be fast-tracked for real-world application.

Improving drug design and optimization

AI-driven approaches are transforming cancer drug development by improving precision, interpretability, and treatment optimization. For example, Vries et al. [189] introduced a revolutionary AI 'fingerprint' technology that can accurately show how cancer cells respond to new drugs, by simply observing changes in their shape. Zhao et al. [190] developed an ensemble of predictive models that reveal the impact of cancer mutations on responses to common DNA replication stress-inducing agents, enabling both multidrug response prediction and mechanistic interpretation.

Improving access to cancer care

Finally, the advent of AI-powered patient engagement tools (e.g., chatbots, virtual assistants) shows promise in supporting cancer care delivery. Chatbots emulating human conversation can enhance medication adherence and self-management. Chaix et al. [191] demonstrated that the use of chatbots enhanced medication adherence among patients with BC. Similarly, Tawfik et al. [192] developed ChemoFreeBot, a chatbot on the Microsoft Azure platform, to educate women with BC, aiming to enhance self-care behaviors and reduce chemotherapy-related side effects through personalized information and improved access to real-time, high-quality data. However, limitations include potential reinforcement of health inequities if training data lacks diversity, and challenges in handling complex patient queries [127, 193, 194]. Future integration should focus on complementing (not replacing) clinician-led care.

Advancing fundamental knowledge of cancer biology

AI methods are increasingly utilized to deepen our understanding of the mechanisms underlying cancer initiation, progression, and drug resistance. The extensive scientific literature offers a wealth of information and insights on cancer. Experts in AI are leveraging LLMs to develop innovative computational tools aimed at enhancing the extraction of knowledge from research publications. For example, complex molecular regulatory pathways (MRPs) are essential for deciphering the mechanisms that govern cancer biology, and knowledge graphs (KGs) have emerged as indispensable tools for organizing and analyzing MRPs, providing structured frameworks to represent intricate biological interactions. Wu et al. [195] illustrated that reguloGPT, an innovative variant of GPT-4 within the LLMs framework, is designed for the end-to-end joint Name entity recognition, N-ary relationship extraction, and context predictions from a sentence describing regulatory interactions with MRPs. Using reguloGPT predictions on 400 annotated PubMed titles centered on N⁶-methyladenosine (m⁶A) regulation, they constructed the m⁶A knowledge graph (m⁶A-KG) and showed its effectiveness in elucidating the regulatory mechanisms of m⁶A in cancer phenotypes across multiple cancer types, thereby underscoring the transformative potential of reguloGPT in advancing the extraction of biological insights from scientific literature [195].

Additionally, researchers are harnessing AI to model the atomic dynamics of the RAS protein, one of the most frequently mutated proteins involved in cancer [196]. The AI-driven multiscale investigation of the RAS/RAF activation lifecycle (ADMIRRAL) project employs a novel integration of molecular dynamics, coarse-grain modeling, and DL to comprehensively investigate the RAS/RAF interaction across progressively biologically relevant time scales. A more comprehensive understanding of the interactions between RAS and its associated proteins could unveil new therapeutic opportunities for targeting oncogenic mutations within the RAS gene. Moreover, transposons are integral to processes such as evolution, gene regulation, and cancer development, yet existing methods for their identification often lack standardized frameworks, including a unified taxonomy scheme and consistent output file formats [197]. Riehl et al. [198] developed TransposonUltimate, an AI-powered software platform designed for the classification, detection, and annotation of transposon-related events. Wang et al. [199] introduced DeepBIO, a comprehensive platform integrating 42 DL algorithms aimed at enhancing the accuracy of functional annotation, visualization analysis, and automated interpretation of high-throughput biological sequence predictions. It offers streamlined biological sequence analysis with minimal programming effort, thereby providing detailed functional insights at both the nucleotide and sequence levels.

Challenges and opportunities for AI in oncology

Data privacy

In healthcare, concerns about data privacy and security are more urgent than in nearly any other industry, particularly in data-intensive domains such as oncology where AI relies on sensitive multimodal data (imaging, genomics, EHRs) [200–202]. It is well recognized that the average individual remains largely unaware of the extensive scope of their data that is collected, stored, sold, and shared across various entities [203]. This opacity is intensified by re-identification risks even with pseudonymized data and by the loss of individual control as commercial holders govern health information [204, 205]. These persistent challenges necessitate the development and implementation of concrete strategies such as privacy-enhancing technologies (PETs) to mitigate privacy risks while enabling critical oncology AI advancements.

Federated learning (FL) is widely used in healthcare, enabling multi-institutional model training without exposing raw data beyond institutional boundaries and thereby preserving local control while supporting collaborative analytics [206, 207]. For example, Elbachir et al. [208] proposed a federated 3D U-Net on the BraTS 2020 dataset for brain tumor segmentation, training across institutions without sharing patient data and demonstrating the feasibility of privacy-preserving, multi-institutional learning in medical imaging. Despite its promise, FL faces challenges such as the statistical diversity of data across clients and communication constraints. Techniques such as Agnostic FL and q-Fair FL have been developed to address data heterogeneity [209, 210]. Meanwhile, client selection protocols, model compression, and update reduction strategies focus on boosting communication efficiency [200, 211, 212]. Additionally, incorporating secure multi-party computation (SMPC) and differential privacy in FL frameworks enhances data privacy and security, making it a valuable tool for healthcare applications [211, 213]. Data from 2025 indicate that using FL combined with DP effectively balances data privacy and model accuracy in BC diagnosis [214]. The findings show that FL outperforms traditional centralized models, demonstrating its capability to generalize across decentralized data without compromising predictive performance.

Synthetic data augmentation technology is increasingly adopted in healthcare to support privacy-preserving research, algorithm training, and patient profiling. By mimicking the statistical characteristics of real data without revealing identifiable information, it helps balance innovation and data protection [215]. For instance, Zhou et al. [216] introduced a novel technique named DiffGuard that generates diverse synthetic medical images with annotations, even indistinguishable for experts, to replace real data for DL model training, which cuts off their direct connection and enhances privacy safety. Walonoski et al. [217] developed Synthea, an open-source software that simulates the Lifespans of synthetic patients, modeling the 10 most common for primary care encounters and the 10 chronic conditions with the highest morbidity in the US. These innovative approaches are crucial for preserving patient privacy and maintaining AI model integrity, enabling the development and testing of healthcare AI applications without risking the exposure of sensitive patient information.

Source-free domain adaptation (SFDA) has recently gained attention in the medical field. It aims to adapt a model well trained on source domain to target domains without accessing source domain data nor requiring target domain labels, thus promoting privacy protection and annotation efficiency [218, 219]. For example, Wang et al. [220] developed a dual reference strategy to select domain-invariant and domain-specific representative samples from a specific target domain for annotation and model fine-tuning without relying on source-domain data. This method ensures data privacy and reduces the workload for oncologists as it just requires annotating a few representative samples from the target domain and does not need to access the source data. However, Guichemerre et al. [221] analyzed the effectiveness of four representative SFDA methods for weakly supervised object localization (WSOL) in histology images and results indicated that these SFDA methods typically perform poorly for localization after adaptation when optimized for classification. Therefore, while SFDA offers promising advantages in terms of privacy and efficiency, its application, particularly in large datasets, remains challenging and requires further research to improve effectiveness.

Blockchain technology offers a secure, decentralized method for storing and managing health data, employing a distributed ledger system to ensure that patient records remain immutable and verifiable, thereby preventing unauthorized access and potential data breaches. It also facilitates secure data sharing among verified entities [222, 223]. In parallel, homomorphic encryption provides a solution for conducting computations on encrypted data without decryption. This enables AI to process and learn from the data, while it remains encrypted, thereby upholding privacy even when third-party analysis is involved [202, 224]. Lastly, differential privacy (DP) adds controlled noise to data sets to obscure individual identities, permitting the training of AI models on population-representative data without compromising individual privacy [61, 200]. Together, these technologies create a comprehensive framework for safeguarding patient privacy in AI-driven healthcare. Their implementation signifies a commitment to upholding ethical standards in healthcare AI and demonstrating a dedication to respecting and protecting patient rights, especially concerning privacy and autonomy. In addition, by adopting these technologies, healthcare providers can comply with laws such as the Health Insurance Portability and Accountability Act (HIPAA) in the US and the General Data Protection Regulation (GDPR) in Europe, which enforce strict data privacy and security standards [200, 225].

Regulatory frameworks and compliance

Most of the described AI applications are subject to oversight by regulatory agencies, which varies based on regional regulations and the intended use of the technology. In the US, the FDA (https://www.fda.gov/) handles authorizations and regulates AI applications under the umbrella of Software as a Medical Device (SaMD) with similar regulatory processes as other SaMD devices and non-AI algorithm [226]. The requirements for FDA clearance primarily depend on the device's class and intended use, with classes ranging from I (lowest risk) to III (highest risk). Typically, the FDA evaluates medical devices through an appropriate premarket pathway, such as premarket clearance (510(k)), De Novo classification, or premarket approval [227]. For example, with the FDA's De Novo pathway authorization of DermaSensor, a novel moderate-risk AI device, manufacturers needed to submit clinical evidence from three key studies [228]. First, the pivotal DERM-SUCCESS trial, involving 1,005 patients and 1,579 Lesions, demonstrated 95.5% sensitivity and a 96.6% negative result by device confirmed by biopsy, though it had low specificity at 20.7%. Additionally, a supplemental melanoma validation study, DERM-ASSESS, was conducted, along with a clinical utility study showing that the use of the device halved missed cancers by primary care physicians from 18 to 9%. Importantly, post-market surveillance requires performance testing in underrepresented populations, such as those with Fitzpatrick skin types IV–VI, to address trial disparities, given that 97.1% of participants were White. This approach contrasts with predecessors such as MelaFind, which followed the premarket approval pathway but was discontinued due to 10% specificity. In addition, the FDA has sought to distinguish CADe systems from the more stringently regulated computer-assisted diagnosis devices. The former are designed to 'identify, mark, highlight, or in any other manner direct attention' to imaging features, rather than autonomously diagnose, stage, or triage pathology. To further reduce the regulatory burden on SaMD developers, the FDA is considering reclassifying CADe systems used for visualizing breast lesions and Lung nodules into a lower-risk category, requiring a 510(k) submission instead of premarket approval [226].

In contrast to the US FDA's single SaMD guidance, the EU employs two comprehensive regulations for medical device safety and efficacy, including AI applications. These are the Medical Device Regulation (MDR), which governs devices implanted within the body, and the In Vitro Diagnostic Devices Regulation (IVDR), for devices testing specimens outside the human body. In the EU, device assessments are not conducted by a single central agency but by accredited organizations authorized to issue CE marks. Oncology AI systems also fall under the EU AI Act, which is the first comprehensive AI regulation by a major regulator globally. According to the EU AI Act, AI systems used in health care are generally 'high-risk'. This classification necessitates implementing a risk-management system, data governance controls (including documentation of the quality and representativeness of training, validation, and test data), technical documentation, transparency and human oversight, assurance of accuracy, robustness, and cybersecurity, as well as a post-market monitoring plan [229]. In Europe, some mammography AI readers and digital pathology triage tools have obtained CE marking under IVDR [230–232]. For example, based on the retrospective study evaluating the CE-marked AI system (Vara version 2.8) in BreastScreen Norway, the evidence required for regulatory pathways included large-scale real-world validation across diverse screening populations. The study analyzed 1,017,208 screening examinations from ten centers and achieved an area under the receiver operating curve (AUROC) of 0.921–0.927 for detecting screen-detected and interval cancers. The system showed potential to reduce workload for the radiologists and potentially increase the sensitivity of mammography [230]. Post-market surveillance could be maintained through ongoing retrospective studies and registry linkages. Firstly, subgroup validations based on mammographic density and equipment vendors can ensure consistent performance across clinical settings. Secondly, integration with other national cancer registries could enable long-term tracking of false negatives and cancer outcomes.

Reliability and scalability

While AI represents the frontier of technological innovation, similar to how dial-up internet once was, widespread adoption remains distant, and ongoing challenges related to interoperability, data quality refinement, and other issues are unavoidable.

Code sharing for AI models is a crucial step toward ensuring transparency and reproducibility, thereby enhancing their suitability for clinical application [233]. While most published studies validate their models on external datasets, true clinical relevance and translatability require that these models be independently reproducible by other researchers, just as with any other robust scientific discovery. In addition to code sharing, disparities exist between the ease of acquiring data from various platforms and the accessibility of that data for independent use by external institutions, particularly regarding private or controlled-access datasets [234]. Moreover, these advanced models are currently niche, complex, and costly, rendering widespread consumer-level adoption a considerable distance into the future. Beyond these barriers to adoption, the path towards reliable, equitable, and clinically impactful AI faces significant challenges rooted in data representativeness, algorithmic fairness, and methodological heterogeneity.

Data representativeness

In the clinical context, access to data that comprehensively represents the diverse human population is essential for developing robust AI models [235]. It is increasingly evident that disparities related to race, gender, and socio-economic status collectively influence disease risk and recurrence among individuals [236]. However, many of the datasets routinely used to train and evaluate AI models in cancer research remain fundamentally biased toward specific racial and ethnic groups [237]. For example, TCGA, the largest repository of diverse cancer datasets, has a median of 83% European ancestry individuals (range 49–100%) [238]. A 2016 genomic analysis of TCGA demonstrated insufficient representation among all ethnic minority populations, thereby restricting the capacity to reliably identify potentially clinically significant genetic alterations [239]. The under-representation of non-European ancestries in widely used resources such as TCGA poses a critical challenge to equitable AI development. AI models trained on biased datasets can perpetuate and amplify existing health disparities. If a model is trained primarily on data from one population, it may not perform well on individuals from other populations due to differences in genetic architecture and disease susceptibility [239, 240]. Another example is the genome-wide association study (GWAS) catalog, the largest genomic database with detailed ancestry classification, which currently consists of approximately 95% European data [241]. A study on GWAS data showed that model effectiveness is linked to the size of population samples [242]. Populations with limited or no representation have greater disparities in disease model performance and gain minimal benefit from benchmark models. In addition, a survey of cell‑line data suggested that Just about 5% of Transcriptomes and 2% of analyzed genomes come from people of African descent, even though Africa contains greater human genetic diversity than all other continents together [243–245]. Consequently, models largely based on European ancestry data might inaccurately assess risk, overlook biology specific to certain ancestries, or detect false correlations that reflect population structure instead of true causal signals. To address this under-representation of non-European ancestries, Smith et al. [246] recently introduced PhyloFrame, a ML method aimed at equitable genomic precision medicine. PhyloFrame corrects for ancestral bias by integrating functional interaction networks and population genomics data with transcriptomic training data.

Besides the under-representation of non-European ancestries in TCGA and other public large databases, commercial medical imaging datasets used to train diagnostic AI models also often lack diversity in race, geography, and socioeconomic status. Models trained on urban, high-resource settings may fail in rural or low-income populations due to differences in imaging equipment, acquisition protocols, and disease. These repositories often contain data primarily from specific demographics or clinical settings, potentially excluding individuals from other backgrounds and the process of selecting images for inclusion in a repository can introduce bias, favoring certain types of cases or patients over others [247, 248]. Efforts to address bias include incorporating data from various demographic groups, using rigorous testing and validation protocols, and conducting ongoing monitoring of model performance. For example, Pinaya et al. [249] utilized generative adversarial networks and latent diffusion models to create synthetic datasets, such as brain MRI data based on age, sex, and brain structure volumes. They found that signals identifying race are also present in these synthetic datasets. Theoretically, it could be easier to validate models using these controlled datasets because specific subgroup sizes can be predetermined, but further research is needed for proper validation.

Algorithmic fairness

Moving beyond data representativeness, achieving algorithmic fairness presents a distinct and critical future challenge. While biased data is a primary cause, unfair outcomes arise from complex interactions within model development and deployment, manifesting as systematic disparities in model performance (e.g., accuracy, false negative rates, calibration) across protected subgroups such as those defined by race, ethnicity, gender, or socioeconomic status [250–252]. This constitutes algorithmic disparate impact, where models unintentionally cause disproportionate harm, even if disparate treatment (intentional discrimination) is absent [253, 254].

Addressing this requires algorithmic bias mitigation strategies that span the entire modeling pipeline and work in a complementary manner. In pre-processing, practitioners reweight samples or transform features to weaken correlations between protected attributes or their proxies and the outcome before any model is trained [255–257]. During in-processing, fairness constraints are incorporated directly into the learning objective; for instance, adversarial debiasing penalizes models for learning representations tied to protected attributes in order to promote subgroup-invariant features, and other approaches enforce statistical parity metrics within the optimization procedure to guide the learner toward fairer solutions [258–261]. After training, post-processing methods modify model outputs by setting subgroup-specific decision thresholds to meet criteria such as equalized odds, which targets equal false positive and false negative rates, or predictive parity, which targets equal positive predictive value [262–265]. Even with these techniques in place, significant challenges remain, as accuracy and fairness often entail unavoidable trade-offs, and when prevalences differ across groups it is mathematically impossible to satisfy all fairness criteria at once [266–268].

Implementing these approaches in clinical AI faces unique hurdles because performance and fairness can degrade under dataset shift when models encounter distributions not represented during training, including genetic variation, differences in imaging protocols, changes in disease prevalence, and evolving taxonomies, which can disproportionately affect underrepresented groups [269–272]. This challenge is compounded by the frequent unavailability of protected attributes due to strict regulations, since sensitive data such as self-reported race are often needed both to audit fairness and to apply many mitigation techniques [273–275]. In addition, models must contend with concept drift as clinical definitions evolve over time, with revisions to coding systems and disease classifications, such as updates to the international classification of diseases (ICD) and the Banff classification, changing the meaning of labels and thereby undermining model validity and fairness, which in turn affects equity assessments and real-world performance monitoring [276–278].

Promising paths forward include the use of FL, which enables training on diverse, distributed datasets without centralizing raw data and can inherently improve representation while helping to mitigate population and acquisition shifts [279–282]. Building on this, FL frameworks that integrate adversarial learning or disentanglement are designed to learn fairer, site-invariant representations across participating institutions [283, 284]. To ensure that such technical advances translate into equitable real-world performance, rigorous fairness auditing should be required, with stratified performance reporting across relevant subgroups during both validation and ongoing monitoring, complemented by attribution techniques such as Shapley additive explanations (SHAP) values to identify and quantify sources of disparity [285, 286]. In parallel, explainability is critical for trust, so developing interpretable methods that help clinicians understand model behavior and diagnose why disparities arise is essential for regulatory approval and for sustainable clinical adoption [287, 288].

Heterogeneity in AI study designs

Heterogeneity in study design across AI oncology studies hinders the interpretation and comparison of reported results and limits assessment of clinical utility. Major sources of variability and potential bias include differences in dataset size, gold-standard labels, and external validation cohorts.

Various studies use datasets ranging from small, single-institution cohorts (n < 100) [27] to large, multi-center collections (n > 10,000) [17, 20, 31, 33, 39, 40]. Smaller datasets are inherently more susceptible to overfitting, potentially leading to inflated performance metrics (e.g., accuracy, AUC) during internal validation that fail to hold in broader populations. While larger datasets generally offer greater robustness, their quality and representativeness remain crucial.

The quality and consistency of the gold-standard used for training and validation are important. Heterogeneity in medical AI studies primarily stems from several key sources. Firstly, pathologist variability is a well-documented challenge, encompassing both inter- and intra-observer disagreement in areas such as histopathological diagnosis, grading (e.g., Gleason, Nottingham), and biomarker assessment (e.g., PD-L1, Ki-67). Artificial intelligence models trained on labels from a single pathologist or a small group inheriting specific biases will reflect those limitations. Inconsistent application of diagnostic criteria (e.g., for sessile serrated lesions in colonoscopy or rare neuroendocrine subtypes further complicates labeling. Furthermore, differences in how regions of interest are annotated on images (e.g., WSIs, radiology scans) or how specific features are defined and labeled introduce noise. Lack of standardized annotation guidelines across studies makes aggregating data or comparing models difficult. Finally, variations in the clinical definitions used as labels (e.g., endpoints such as overall survival vs. progression-free survival) directly impact what the model learns and reports.

One of the most critical factors affecting real-world applicability is external validation. Many studies rely solely on internal validation, which optimistically biases performance estimates. Studies employing external validation often use cohorts from similar institutions or healthcare systems, limiting generalizability to truly diverse settings. The scarcity of robust, prospective, multi-center external validation studies represents a major barrier to clinical trust and adoption. Performance frequently degrades when models encounter data from different scanners, acquisition protocols, patient demographics, or clinical practices not represented in the training set.

Addressing these challenges requires concerted efforts towards standardization. Key priorities include defining minimum dataset size recommendations for different tasks, establishing clear and consistent gold-standard diagnostic and outcome definitions (leveraging expert consensus panels), mandating rigorous independent external validation using prospectively collected cohorts from diverse settings as a prerequisite for claims of clinical utility, and ensuring transparent reporting of all methodologies and data characteristics (e.g., adhering to guidelines such as CONSORT-AI and STARD-AI) [289, 290]. Ultimately, future research must prioritize robustness and generalizability alongside technical accuracy to ensure AI models deliver meaningful benefits in real-world clinical application.

Social impact

AI remains a human creation, crafted by individuals who may possess malicious intent or be inherently fallible. Consequently, issues of bias and fairness can emerge, stemming from the biases of their creators or even from ethically benign sources such as inadequate sampling methods [291].

An additional critical consideration is the issue of interpersonal responsibility when human oversight is minimized or removed [292]. For example, if AI systems make errors, questions arise regarding who should be held liable, such as in malpractice suits, and how this might influence insurance coverage and liability scope [293]. These are long-term, unresolved questions, but it is essential that industries address them thoughtfully and ethically to ensure that accountability is not simply transferred onto consumers.

Clinical integration and validation

Currently, nearly all AI models developed for cancer diagnosis rely on clinical data collected at the time of development, which may include patient reports or sequencing results. This raises the question of whether there are AI systems capable of recommending additional diagnostic tests or treatment options, or even assisting in prescribing anticancer medications, without dependence on traditional clinical data. As multiomics technologies continue to advance, incorporating diverse data types such as methylation profiles and fragmentomics, it is conceivable that once an AI model's dataset reaches sufficient scale and diversity, it could potentially predict the likelihood of cancer development solely based on data from healthy individuals [294]. Furthermore, by comparing sequencing results from cancer patients against an extensive database, it may become possible to recommend personalized chemotherapy regimens.

Looking ahead, prevention rather than treatment may emerge as the most compelling application of AI in cancer care. Pioneering research has already enabled the scientific community to compile a comprehensive portfolio of cancer risk factors, paving the way for more effective early intervention strategies. Advances in technology have facilitated multiple methods of collecting data at the individual patient level. In addition to genetic testing and EHRs, wearable biosensors have revolutionized healthcare, especially in cancer detection and monitoring [295]. They are embedded into smartwatches, patches, and clothing, continuously collecting extensive physiological and biochemical data streams to enhance diagnosis and treatment [296]. Traditional methods such as biopsies and imaging are invasive and costly, limiting their use. Advances in microfluidics and surface engineering have broadened the scope of wearable biosensors, utilizing body fluids such as sweat, saliva, tears, and interstitial fluid for non-invasive tumor biomarker detection [297]. Moreover, AI paired with smartphone-based imaging for in vivo cancer detection is an emerging field that Leverages advancements in 5th generation (5G) and 6G sensing technology. The high-speed data Transmission capabilities of 5G networks enable the efficient processing of large data volumes, allowing AI algorithms to quickly analyze smartphone-captured images and provide cancer diagnoses. For example, the integration of advanced imaging sensors, such as high-resolution cameras, thermal, and multispectral sensors in smartphones, enhances image quality and aids in the accurate identification of potential tumors [298, 299]. These diverse, real-time data streams have the potential to feed into integrated AI platforms, creating dynamic, personalized risk profiles by synthesizing information from genomics, EHRs, lifestyle factors (via sensors), physiological fluctuations (from wearables), and visual changes (from smartphone imaging). This offers real-time management guidance for modifiable risk factors, such as setting personalized activity goals based on wearable data insights. It also enables personalized early intervention recommendations, such as prompting high-risk individuals for specific screening tests based on smartphone-based imaging. Furthermore, these integrated AI systems could enable proactive remote monitoring of cancer survivors and alert clinicians to early physiological or behavioral changes that may indicate recurrence or complications long before traditional follow-up schedules would.

Clinical validation is the pivotal step that determines whether AI systems meaningfully improve cancer care beyond technical performance in development datasets. Unlike technical validation, which focuses on accuracy metrics within held-out or cross-validated data, clinical validation evaluates generalizability, calibration, safety, and impact in real patients and workflows. Robust external validation across institutions, geographies, devices, and time is essential to assess resilience to dataset shift and to quantify whether models maintain discrimination and calibration in populations that differ from the training data. Appropriate endpoints should extend from diagnostic accuracy to decision impact, downstream care processes, patient-centered outcomes, and harms, with transparent, pre-specified performance thresholds and failure analyses. For tools that learn or are periodically updated, adaptive or platform trial structures and learning health system approaches help align evaluation with model change. Equity considerations require planned subgroup analyses by demographics, tumor subtype, site, scanner or device, and socioeconomic context, coupled with bias mitigation and reporting of heterogeneous effects. Regulatory pathways (Sect. " Regulatory frameworks and compliance") emphasize intended use, benefit–risk assessment, Good Machine Learning Practice (GMLP), change control for adaptive algorithms, and post-market surveillance. After deployment, continuous monitoring for performance drift, re-calibration when needed, incident reporting, and human-in-the-loop safeguards are necessary to maintain safety and effectiveness.

In oncology, much of the current literature remains retrospective and single-center [24, 28, 29, 43], with fewer prospective impact evaluations [25, 35, 36] and only a limited number of randomized studies [19, 30]. Where rigorous prospective validation has been conducted, results have sometimes confirmed utility and at other times revealed diminished performance, underscoring the need for careful and context-specific evaluation. Ultimately, AI systems for cancer care should be considered ready for routine use only after they demonstrate reliable, equitable clinical benefit in well-designed studies and are supported by ongoing governance and surveillance in real-world practice.

However, it must be emphasized that, despite the rapid advancements and promising prospects of AI, it can never fully replace clinicians and will ultimately serve as a crucial tool to assist healthcare professionals in their practice (Fig. 3).

Fig. 3.

Challenges and future pathways for AI integration in oncology

Limitations

This review has several limitations that warrant acknowledgment. First, our synthesis of evidence relies predominantly on retrospective studies and proof-of-concept trials (e.g., CADe validation for colonoscopy in Sect. " Expediting cancer screening, detection and diagnosis"). While these demonstrate technical feasibility, they may overstate real-world performance due to idealized datasets and limited external validation. Prospective randomized trials assessing patient survival or cost-effectiveness remain underrepresented in our analysis, including trials of AI-driven surveillance interventions. Second, our assessment of algorithmic fairness is constrained by inconsistent reporting in primary literature. Although we highlight biases in training data (for example, TCGA’s under-representation of non-European ancestries), we do not provide a quantitative synthesis of performance disparities across demographic groups, including for prognostic models in breast and prostate cancer (Sect. " Improving cancer surveillance"). This limits our ability to assess equitable generalizability. Third, the regulatory analysis in Sect. " Clinical integration and validation" focuses on FDA/CE pathways but omits emerging frameworks in Asia (e.g., China’s National Medical Products Administration guidelines for oncology AI). This geographical gap reflects our inclusion bias toward English-language publications and may reduce relevance for global audiences. Finally, our evaluation of clinical integration barriers prioritizes technological hurdles over socioeconomic determinants. We provide limited analysis of how factors such as hospital resources and payer policies shape adoption, a critical gap in our review. These limitations highlight constraints specific to our methodology but do not undermine the challenges discussed in Sect. " Challenges and opportunities for AI in oncology". Future studies in this field would benefit from standardized bias reporting, prospective registry data, and deliberate inclusion of global regulatory cases.

Conclusion

Fueled by the exponential increase in data, advancements in AI algorithms, and enhanced computational capabilities, AI possesses the transformative potential to revolutionize precision oncology across the entire continuum of cancer care, encompassing prevention, diagnosis, treatment, and drug development. Achieving this potential relies on coordinated efforts focused on translational research, which encompasses expanding access to extensive and diverse datasets, developing robust and interpretable AI models, and conducting rigorous validation through unbiased, prospective clinical trials. Furthermore, the development of robust regulatory frameworks is crucial to ensure the safe, equitable, and effective deployment of AI technologies. By prioritizing these initiatives and fostering interdisciplinary collaboration among key stakeholders, we can expedite the integration of AI into routine clinical practice, thereby maximizing its impact on patient outcomes and ultimately alleviating the global burden of cancer.

Looking ahead, key research priorities must be addressed to unlock the full potential of AI in oncology: (a) Development of multi-institutional, multi-modal reference datasets: Establishing large-scale, curated datasets that integrate diverse data modalities (imaging, genomics, transcriptomics, proteomics, pathology, EHRs, liquid biopsy) across multiple institutions and diverse patient populations is paramount. Initiatives such as MOSSAIC demonstrate the value of such efforts, but broader collaboration and standardized data sharing frameworks are essential to overcome biases, enhance model generalizability, and fuel discovery across rare cancer types and underrepresented groups. (b) Creation of XAI methods that clinicians find trustworthy: Moving beyond 'black box' predictions is critical for clinical adoption. Research must focus on developing intuitive and reliable XAI techniques (e.g., visual saliency maps, feature attribution, natural language explanations) that provide clinicians with actionable insights into why an AI model arrived at a specific diagnosis, risk prediction, or treatment recommendation. Building trust requires transparency that aligns with clinical reasoning pathways, enabling effective human-AI collaboration, as highlighted by the need for interpretability in models predicting treatment response or biomarker status [300, 301]. (c) Integration of socio-behavioral and real-world data into risk-prediction and care algorithms: To achieve truly personalized prevention and care, AI models must evolve to incorporate factors beyond traditional clinical and molecular data. This includes integrating data on social determinants of health, behavioral patterns (e.g., from wearables or patient-reported outcomes), environmental exposures, and longitudinal real-world evidence. Developing AI capable of synthesizing this complex tapestry of information, while carefully addressing privacy and equity concerns, will be crucial for predicting individual cancer risk with higher fidelity, tailoring screening strategies, optimizing treatment adherence, and understanding the impact of non-biological factors on outcomes, as envisioned in future AI-driven prevention models. Addressing these priorities, alongside ongoing advancements in algorithms, computing, and ethical frameworks, will be instrumental in translating the promise of AI into tangible improvements in cancer prevention, early detection, therapeutic efficacy, and ultimately, patient survival and quality of life.

At the time of writing, shortly after the release of GPT-5, Sam Altman, the chief executive officer (CEO) of OpenAI, the company behind ChatGPT, said in an interview that by 2035 AI will cure or at least treat many diseases that currently plague humanity [302, 303]. For example, in the era of GPT-8, we might ask it to 'cure a certain type of cancer.' It will first read all existing research and data and come up with some treatment ideas. Then it will tell us, 'I need you to find an experimenter to conduct these nine experiments and tell me the results.' After two months of cell culture, when the experimenter sends the results back to GPT-8, it may say, 'Well, there's an unexpected discovery. I need to conduct one more experiment.' Then it will tell you, 'Synthesize this molecule and test it on mice.' If it works, then conduct human trials. Finally, it will say, 'Okay, here's the process for submitting it to the FDA.' These scenarios underscore AI's transformative potential in biomedical discovery and care, but realizing them will require rigorous validation, ethical safeguards, robust oversight, and integration with clinical expertise to ensure safety, equity, and reproducibility.

Appendix: evolution of AI models and data modalities

At a workshop held at Dartmouth College in the summer of 1956, McCarthy et al. introduced the term 'artificial intelligence', also known as 'thinking machines' [304]. The first ML algorithms were developed in the 1950 s to the 1970s. Frank Rosenblatt's perceptron was one of the neural network models, primarily employed for binary classification tasks [305]. From the 1980 s to the 2010 s, ML research has led to development and application of a number of 'shallow' learning algorithms, including earlier generalized classic linear models such as logistic regression, Bayesian algorithms, decision trees, and ensemble methods [306, 307]. The early 2010 s signified a pivotal turning point with the advent of DL, which exhibited superiority over traditional ML methods in applications such as image and speech recognition [308]. In computer science, ML constitutes a specialized branch of AI, while DL represents a particular subset of ML that emphasizes the development and application of deep artificial neural networks [309]. The 2020 s marked the introduction of the transformer architecture, a revolutionary DL framework distinguished by its attention mechanism, which proficiently models long-range dependencies within sequential data such as text and video [310]. Transformers have further significantly enhanced the capabilities of AI, particularly through their pivotal role in the development of LLMs. Through training on extensive datasets sourced from the internet using advanced supercomputing infrastructure, LLMs such as ChatGPT, DeepSeek, LLaMA and Grok demonstrate extraordinary capabilities in comprehending human input and producing responses that closely emulate human communication [311].

Authors’ contributions

CHC researched data for this manuscript. All authors discussed the content of the manuscript, and wrote, reviewed and/or edited the manuscript prior to submission.

Funding

This study is supported by funds from Research Grants Council-General Research Fund (14101321; 24100520), RGC-Collaborative Research Fund (C4008-23WF, C4042-24GF); Heath and Medical Research Fund (06170686; 08190706); The National Key Research and Development Program of China (2021YFF1201300)

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

Not applicable.

Competing interests

The authors declare no competing interests.