Current trends and future directions of artificial intelligence in lung cancer diagnosis

Abstract

Lung cancer is the most lethal malignancy worldwide, largely due to its late detection after its progression to advanced stages. Over the last decade, artificial intelligence (AI) applications have shown significant potential in transforming lung cancer diagnostics by improving the speed, accuracy, and personalization of early detection strategies. This review provides a comprehensive overview of current AI application landscape in early lung cancer diagnosis, encompassing medical imaging, histopathology, liquid biopsy, natural language processing of electronic health records, and genomic profiling. We explain how machine learning, deep learning, and transformer-based models are employed in lung cancer diagnosis, and summarize recent cutting-edge advances, including multimodal AI platforms and Food and Drug Administration (FDA)-approved computer-aided diagnosis/detection (CAD) systems. Furthermore, we evaluate the challenges that impede clinical translation, including data heterogeneity, interpretability, and privacy, and present prospective directions such as federated learning and multi-omics integration. Through a comprehensive analysis of the dynamic evolution of AI applications in oncology, we aim to inform researchers, clinicians, and policymakers about its diagnostic potential and translational relevance in clinical practice.

Introduction

Lung cancer accounts for a disproportionately large share of cancer-related incidence and mortality worldwide (1). In its early stages, patients with lung micro-invasive carcinoma can achieve a 5-year postoperative survival rate approaching 100%, similar to those diagnosed with carcinoma in situ. Regrettably, most patients are diagnosed only once the lung cancer has progressed to advanced stages with a corresponding 5-year survival rate dropping to as low as 2% (2). Therefore, early diagnosis holds a critical role in combating lung cancer.

To date, a variety of diagnostic strategies are implemented in clinical practice in addition to standard diagnostic methods, including cytology, chest X-ray, computed tomography (CT) scan, bronchoscopy, and tissue biopsies. Lung cancer screening based on low-dose CT (LDCT) has been under consideration or is widely implemented in many countries (3). This is because several studies, especially the National Lung Screening Trial (NLST) and Nederlands–Leuvens Longkanker Screenings Onderzoek (NELSON) trial, have suggested that population screening is cost-effective and significantly reduces lung cancer-related mortality (4). Also, biomolecular marker-based diagnosis and medical radiomics are expected to address the current challenges in lung cancer diagnosis (5,6). However, the survival outcome remains suboptimal due to several existing challenges. These include specificity and overdiagnosis in heterogeneous imaging features and histopathology, lack of comprehensive guidelines and sufficient external validation, clinician-induced anxiety and errors due to high workloads, and difficulties in filtering and interpreting large datasets (7).

Over the last decade, with the rapid expansion of biomedical datasets, the continuous development of computing hardware, and the advent of deep learning (DL) algorithms (8), the integration of artificial intelligence (AI) has emerged as both a necessary and promising development. In this review, we outline the key concepts and current landscape of AI in lung cancer diagnosis, including the evolution of data streams, DL, machine learning (ML), and natural language processing (NLP). In addition, we provide a structured summary of selected AI applications categorized by different prospective diagnostic modalities. Specifically, we describe overarching challenges, including data curation, diagnostic bias, and the clinical integration of AI, as well as emerging strategies underway to address these challenges. Therefore, we aim to demonstrate the recent state-of-the-art applications of AI in lung cancer diagnosis and discuss related concepts to inform future research, providing opportunities for integration into real-world clinical diagnostic workflows.

Rapid advances in data streams for clinical oncology

In the early stages of modern clinical oncology, data and computational limitations necessitated a shift toward simplified, human-interpretable representations of unstructured patient data (e.g. imaging features, gene expression data). One of the most prominent examples is the TNM classification system, introduced in 1977 by the American Joint Committee on Cancer (AJCC). In the earliest editions, the framework incorporated only limited inputs and cancer stages, such as the second edition TNM classification system (9) (Figure 1). In subsequent decades, with the advancements in data integration methodologies, increased emphasis on clinical research, and improvements in clinical data storage capabilities, the evolution of lung cancer patient data representation has reflected a shift toward increasingly complex and multidimensional models. As a result, more predictive nomograms that incorporate various nuanced descriptors can be generated to assist in lung cancer diagnosis and treatment selection for patients (Figure 1).

Figure 1.

Comparison of 2^nd, 7^th and 9^th edition stage groups for lung cancer. The increasing complexity of TNM lung cancer staging reflects the growing volume of data streams and higher data dimensionality in clinical oncology. T, Tumor; N, Node; M, Metastasis.

Recent advances in emerging algorithms, combined with the increasing global burden of cancer, have accelerated the digitization of lung cancer diagnosis (10). Since the 21^st century, new data streams, including electronic health records (EHRs), digital radiology, pathology imaging, and routine genomic profiling, have been introduced. Key examples of lung cancer genomic profiling databases include the American Association for Cancer Project GENIE, the Cancer Genome Atlas (TCGA), and the most recent whole-genome sequencing data from UK clinics, conducted by Kinnersley et al. (11). Alongside the different stages of the patient pathway, starting from prevention to follow-up, relevant oncology data streams are progressively accumulated (Figure 2).

Figure 2.

An example of a lung cancer patient pathway converging with data streams from each stage. CT, computed tomography; T, tumor; N, node; M, metastasis; ECOG, Eastern Cooperative Oncology Group.

With the growing biological insight into cancer and the accumulation of patient longitudinal information through novel data streams, ML and related algorithms can leverage these data to identify patterns and generate predictions for future “invisible” patients. Based on the real-time availability of these data, one upcoming and tantalizing era is precision oncology (12) , in which AI models can theoretically generate optimized, personalized cancer management strategies tailored to an individual’s genomic profile, although this remains unfulfilled.

AI subdisciplines and algorithms applied in lung cancer

AI, first proposed by emeritus Stanford Professor John McCarthy in 1955, aims to replicate patterns of human decision-making. The general process involves data collection, data preprocessing, algorithm selection for model construction, model training and optimization, and performance evaluation informed by prior experience. Given the aforementioned challenges of high mortality, delayed diagnosis, and tumor heterogeneity in lung cancer, AI thus emerged as a prospective and effective tool in this domain. In this section, we introduce key AI subdisciplines applied in lung cancer diagnosis, including NLP, ML and DL. The hierarchical relationships among these subfields, along with representative algorithms used in lung cancer diagnosis, are illustrated in Figure 3A.

Figure 3.

Overview of AI subfields, ML and NLP pipeline in lung cancer diagnosis; (A) Venn diagram of AI, including ML, NN, and DL, NLP, and recent representative algorithms applied in lung cancer diagnosis; (B) In conventional ML, also known as shallow learning, feature extraction is performed outside of the algorithmic stage, where hand-crafted features and optimal engineering are based on expert knowledge; (C) NLP pipeline for recognition in EHR-based lung cancer data. AI, artificial intelligence; ML, machine learning; NLP, natural language processing; NN, neural network; DL, deep learning; EHR, electronic health record.

ML

Unlike traditional hard-coded software, which relies on manually defined logic and task-specific instructions, ML is the science of developing data-driven algorithms (Figure 3B). These algorithms can solve problems without being explicitly programmed and are often viewed as statistical tools for performing data analysis, visualization, prediction, and forecasting. In theory, ML models can be categorized into unsupervised learning, semi-supervised learning, supervised learning, and reinforcement learning (Figure 4). Within lung cancer research, each basic type of model has found tailored applications and emerged as a pivotal tool for enhancing early diagnosis. In this section, we highlight key algorithms applied in lung cancer diagnosis, including Bayesian networks (BNs), support vector machines (SVM), logistic regression (LR), decision trees (DT), K-nearest neighbor (KNN), and Q-learning.

Figure 4.

Concept map of four basic types of ML. (A) Supervised learning uses input-output pairs (labeled data) for training; (C) However, unsupervised learning only uses input (unlabeled data), which are classified based on inherent patterns; (B) Semi-supervised learning utilizes only a small fraction of labeled data by incorporating unlabeled data from the same distribution to augment labels and enhance training performance; (D) Similar to unsupervised learning, reinforcement learning does not require labeled data; its policy is iteratively updated through a reward-based system. Decision-makers (agents) interact with the environment by taking action based on a policy, which is followed by a change in state. The agent receives feedback in the form of rewards and updates its policy through a learning algorithm to dictate future actions. ML, machine learning.

SVMs, for instance, are particularly well-suited to classifying pulmonary nodules as benign or malignant. By maximizing the margin between the closest points from different lung cancer sample classes, known as support vectors, the optimal hyperplane in an N-dimensional space can be recognized. This approach demonstrates strong robustness and high accuracy in recent lung cancer diagnosis research, especially for those who aim to generalize datasets that exhibit high-dimensional features (13).

BNs, also referred to as “belief networks”, are a specific type of probabilistic supervised learning model. In this model, variables are represented as nodes, and the directed edges between them represent conditional dependencies in a directed acyclic graph (DAG). In the context of lung cancer, which often involves multifactorial risk elements and heterogeneous data types, BNs are particularly advantageous for integrating diverse information streams, such as imaging findings, genomic markers, and patient history, into a unified predictive framework.

First, BNs can quantify diagnostic uncertainty by producing multiple outcomes with corresponding probabilities rather than generating a single definitive result. This ability to quantify uncertainty and interference is important for clinical decision-making by lung cancer clinicians (14), particularly when working with incomplete or ambiguous datasets. A notable example overcoming the present limitation of sparse EHR data in lung cancer is a recently developed interpretable model (15), BOUND (BN for large-scale lung cancer Digital prescreening). It can predict lung cancer risk and identify high-risk factors within datasets containing up to 70% missing data.

Additionally, lung cancer diagnosis typically involves multiple variables such as imaging data, omics data, or lifestyle factors, resulting in complex causal relationships. By representing these causal relationships as probabilistic dependencies in a DAG and integrating heterogeneous data sources into a unified framework, BNs enable more comprehensive analysis to identify key risk factors underlying lung tumorigenesis and their interactions. Moreover, by extending BNs to incorporate time-series data, Petousis et al. (16) developed dynamic BNs using data from the NLST dataset. In this model, each time slice is a BN that produces coherent longitudinal patient data through inter-slice connections. In most cancer and non-cancer cases, this approach generates higher predictive accuracy in the experiment compared with naïve BN models.

In contrast to BN, DT provide a simpler and straightforward hierarchical structure by recursively partitioning lung cancer datasets and can discriminate different lung cancer subtypes in a graphically intuitive way. For example, a 2019 study developed a decision-tree-based classifier to classify lung squamous cell carcinoma (LUSC) and lung adenocarcinoma (LUAD) subtypes using miRNA expression data from TCGA (17). During ensemble learning, DT can be transformed into a random forest (RF) through bootstrap aggregation of multiple DT. Although interpretability is reduced, these models demonstrate significantly improved robustness and accuracy, particularly in lung cancer diagnosis applications such as biopsy-based feature classification and CT radiomic analysis (18).

LR, a classical linear classifier, is widely utilized in lung cancer diagnosis due to its specificity and sensitivity based on lung histopathologically distinctive characteristics. Tirzïte et al. (19) reported an inspiring artificial olfactory sensor utilizing LR for breath analysis. Given 223 patients with lung diseases other than cancer and 252 lung cancer patients, the LR model successfully distinguished between the two groups, with both overall specificity and sensitivity exceeding 90%.

KNN identifies the K most similar lung cancer profiles by calculating the distance (e.g. Euclidean distance, Cosine similarity) between the new lung samples to be classified and existing labeled lung samples in the dataset based on the categories of K labels. Unlike other ML algorithms that require a training phase to construct a generalized model, KNN is considered a form of “lazy learning”, which is particularly advantageous in lung cancer diagnosis, as it only stores lung cancer training data without performing any computation until a prediction is requested. This enables the efficient incorporation of new lung cancer patient data at any time, providing great flexibility and accuracy within the dynamic diagnostic data flow for lung cancer.

It has been shown that KNN performs well in breathomics for distinguishing between LUAD and LUSC (20), similar to LR. Additionally, a comparative study involving 310 instances evaluated several learning algorithms based on lung cancer-related characteristics, with a particular focus on 11 symptoms (e.g. yellow fingers, chest pain) and two patient habits (e.g. smoking) (21). KNN was identified as the most effective method for early lung cancer prediction, with an accuracy of 92.86%, which highlights its extraordinary adaptability when applied to small lung cancer datasets.

Reinforcement learning, inspired by reward prediction-error signaling adaptations for optimal decision-making, holds significant potential for real-time personalized treatment planning and biomarker diagnosis selection in the future treatment of lung cancer. Recently, a widely used reinforcement learning algorithm called Q-learning has been employed to support lung cancer tumor segmentation, localization, and classification tasks (22).

DL

DL, a specialized subfield within ML, has recently emerged as a particularly successful and promising method in lung cancer diagnosis. It draws inspiration from the principles of human biological neural networks. Traditional ML approaches typically depend on manually crafted lung nodule features and require careful engineering to achieve optimal performance, limiting their ability to process clinically obtained lung cancer data in its raw form. Representation learning is a set of techniques that enables the automatic extraction of hierarchical lung cancer features from raw inputs, allowing models to learn discriminative patterns for tasks such as lung nodule detection, tumor subtype classification, or treatment outcome prediction. DL methods, such as generative adversarial networks (GAN), recurrent neural networks (RNN), transformers and convolutional neural networks (CNN), are representation learning methods that leverage multi-layered neural architectures to analyze diverse data modalities in lung cancer, such as radiological scans, sequential clinical records, or histopathological images (Figure 5A). Relatively less manual intervention is required compared to traditional methods during lung cancer diagnosis. Because DL performs complex representation transformations through layered networks, and involves millions of adjustable weights, DL models usually require enormous lung cancer datasets and computational demand for their full potential. Once training is completed, DL models can learn extremely complex functions with more outstanding performance than other methods in lung cancer diagnosis (Figure 5B). Thus, these models are particularly effective in extracting hidden patterns from complex, high-dimensional, and unstructured data sources relevant to lung cancer (23) (e.g. CT image features, tumor genomics).

Figure 5.

DL and CNN. (A) DL involves artificial neural networks with many layers; (B) DL techniques, when supplied with sufficiently large training datasets, have achieved higher performance than simple learning algorithms (e.g. SVM, RF); (C) Typical process of CNN in lung cancer diagnosis. LeNet, one of the first CNN architectures, is used as an example in this context. Many architectures are possible here, but finding the optimal one is as much a creative and artistic endeavor as it is a scientific one. Chest radiographs are reproduced with permission from Ref. 24, Springer Nature. DL, deep learning; CNN, convolutional neural network; ANN, artificial neural network; SVM, support vector machine; RF, random forest.

Here, we focus on introducing CNN, which has become the most widely adopted architecture for analyzing medical images in lung cancer (25). Most CNNs have similar architectures (Figure 5C), including convolution, subsampling, and non-linear activation layers that alternate, followed by fully connected layers. The convolutional layers apply learnable filters to extract local spatial features—such as edge contours, shapes, and textures—from input images, which are critical for identifying pulmonary nodules, masses, or other suspicious structures in lung cancer screening. Pooling layers reduce the spatial resolution of lung feature maps, enabling the network to abstract higher-order representations while improving translational invariance, a property particularly useful when lung lesions vary in size or position across patients. Rectified Linear Units (ReLUs), a popular non-linear activation function, introduce complexity into the model by enabling it to learn non-linear associations between features, supporting nuanced decisions such as distinguishing between solid and subsolid nodules or subtle changes in tumor margins. Finally, fully connected layers serve to utilize and integrate the acquired high-level semantic features for lung cancer prediction or classification.

The main strength of CNN in lung cancer screening is their ability to learn hierarchical representations of local features, thereby reducing the complexity of parameters and computational load. Additionally, the preprocessing cost is low, and CNN possesses strong robustness and excellent generalization capabilities, which facilitate the handling of high-dimensional and irregular images in lung cancer diagnosis (26,27).

The integration of CNN architecture with traditional ML algorithms has also become increasingly popular in lung cancer screening. Notable examples include deep reinforcement learning (22) and CNN integrated with SVM (28). Furthermore, CNN are widely employed in NLP and lung image analysis. Further details will be addressed and discussed in subsequent sections.

NLP

In lung cancer screening, language underpins the communication of clinical findings, from radiology reports to pathology summaries and physician notes. Before data analysis occurs, it is equally critical for AI systems to comprehend the vast amount of data generated during patient-doctor and healthcare system interactions. Recent developments in EHRs are revolutionizing lung cancer screening by digitizing clinical information and providing unprecedented opportunities for accurate and early lung cancer diagnosis. However, relevant patient data are often stored in siloed datasets and unstructured text, which typically require manual processing to transform into AI-readable structured lung screening data. This process can be inefficient, error-prone, and limited in scalability, particularly affecting lung cancer screening models that require complex computations and extensive data processing capabilities (29).

The development of NLP has emerged as a transformative tool to address this challenge by enabling automated extraction, interpretation, and structuring of unstructured text from EHRs. It combines computational linguistics, statistical methods, and ML techniques to interpret unstructured textual or verbal data in EHRs of patients with lung cancer and annotate relevant information. Figure 3C illustrates a standard pipeline for applying NLP techniques to process and analyze medical text data.

Conventional NLP methods built on rule-based and statistical techniques require significant time and effort to model and exhibit limited generalizability beyond narrowly defined lung cancer documentation tasks, such as keyword-based tumor stage classification or radiology term extraction. In contrast, modern NLP techniques, particularly those integrated with DL, are capable of dynamically processing free-text data and hold broad applicability across medical records of patients with lung cancer. This highlights the feasibility of automatically collecting clinically relevant, real-world lung cancer outcomes from free-text EHRs. Furthermore, these advanced NLP models can learn from lung cancer screening data and improve over time, thereby reducing the need for manual annotation of terms and providing greater flexibility in understanding and interpreting lung medical texts. In a subsequent study conducted by Kenneth’s research team (30), they developed the PRISSMM data model. They combined limited manual clinical annotation with deep neural networks to successfully train an interpretable deep NLP model on medical oncologist notes and imaging reports for patients with non-small cell lung cancer (NSCLC). The model demonstrates strong generalizability across cancer types outside the training dataset. Some researchers have been dedicated to training NLP using a wide range of cancer patient EHRs and applying the model for lung cancer diagnosis. Nobel et al. (31) applied NLP to radiology reports for automated classification of lung tumors, providing a standardized quality assurance tool to support lung cancer screening workflows, clinical management guidance and report generation. Another study (32) demonstrated the use of NLP in analyzing chest CT reports, illustrating the consistency of NLP in detecting pulmonary nodules.

In lung cancer diagnosis, breakthroughs in DL are based on the transformer architecture with attention mechanisms, thereby removing the reliance on recurrent structures and convolutional layers. Since then, attention-based algorithms, such as Google’s Bidirectional Encoder Representations (BERT), have emerged as one of the major algorithms in lung cancer NLP. Fei et al. (33) developed one of the earliest NLP systems tailored specifically for Chinese radiology reports. Their approach combined a bi-directional long short-term memory network with a conditional random field (Bi-LSTM + CRF) to identify entities related to follow-up recommendations for pulmonary nodules. From reports, the system generated automated follow-up suggestions using a knowledge graph in conjunction with rule-based templates, thereby supporting quality management in clinical practice. Agnikula Kshatriya et al. (34) implemented a hybrid DL model (BERT and Bio-BERT) to extract text features from EHRs. To evaluate doctors’ documents, they utilized semi-supervised training methods. Additionally, the Wang research group (35) from the UK developed a transformer-based DL model, MedAlbert, for the early diagnosis of lung cancer using EHRs. This model achieved an AUC of 0.924, outperforming classic regression approaches and offering valuable insights into patient care pathways.

More intriguingly, most state-of-the-art large language models (LLMs) are also based on the transformer architecture, including GPT-4, Llama, and Mistral. In lung cancer diagnosis, LLMs have been applied to transform clinical notes into structured data, emerging as a prospective approach to enhance clinical decision-making and support research. A recent study employed ChatGPT-3.5 (36) to analyze 1,026 lung cancer pathology reports, achieving 89% accuracy in classifying lung tumor stages and histological types. This outperforms the selected traditional NLP methods, such as a DL-based named entity recognition method and a keyword-based search algorithm. In addition, an oncology-specific LLM, Woollie, trained on real-world clinical oncology data from Memorial Sloan Kettering Cancer Center (37), demonstrated promising performance in radiology-based lung cancer detection with an AUC of 0.95, suggesting that LLMs may extend beyond information extraction to broader clinical decision support.

AI applications in lung cancer diagnosis

Recently, the use of AI to advance imaging technology has achieved notable successes in field of lung cancer diagnosis, including radiomics and computer-aided diagnosis/detection (CAD) systems. Valuable sources of data provided by multiple imaging modalities, including CT, X-ray, magnetic resonance imaging (MRI) and positron emission tomography (PET), have also been analyzed through these diverse approaches. In this section, we discuss both the current successes and the future potential of AI applications in lung cancer diagnosis.

Screening and detection

Early detection plans for cancer rely on screening as a first point of entry, whether in symptomatic patients or through incidental detection. Currently, many scientists and enterprises are increasingly investing resources into developing AI systems to assist physicians in screening lung nodules (38). These developments will enable AI to identify suspected nodules for both screening and follow-up treatment recommendations, and better utilize imaging features to create individualized risk models that incorporate EHRs, thereby reducing missed diagnoses and misdiagnoses. Recent AI applications for detecting and screening pulmonary nodules are outlined in Table 1. However, most cited studies in Table 1 are retrospective and conducted in single-center or multi-center retrospective settings. Prospective validation and real-world clinical evidence are needed for clinical utility in the future.

Table 1. Recent AI applications in lung nodule detection and screening.

Model proposed by studies	Finding	Reference	Year
AI, artificial intelligence; SVM, support vector machine; LASSO, least absolute shrinkage and selection operator; CT, computed tomography; ML, machine learning; LDA, linear discriminant analysis; ANN, artificial neural network; PET, positron emission tomography; CADx, computer-aided diagnosis; CNN, convolutional neural network; LUNA16, Lung Nodule Analysis 2016; DL, deep learning; Res-Net, residual neural network; CMixNet, comprehensive mixed network; DI2IM, deep image-to-image network; RCNN, region-based convolutional neural network; GLCM, Gray Level Co-occurrence Matrices; LCP, lung cancer prediction; NLST, National Lung Screening Test; 18F-FDG, 18F-2’-deoxy-2-fluoro-D-glucose; CXR-LC, lung cancer incidence risk model; EHR, electronic health record; RADS, reporting and data system; LDCT, low-dose CT; AdaBoost, Adaptive Boosting; SNMV, Self-Normalized Multi-View; CAD, computer-aided diagnosis/detection; B-RGS, bates distribution coati optimization; MET, mesenchymal epithelial transformation; 3D-UTE, three-dimensional ultrashort echo time; MRI, magnetic resonance imaging; NSCLC, non-small cell lung cancer; LGBM, light gradient boosting machine; XGBoost, extreme gradient boosting; LR, logistic regression; AUC, area under the receiver operating characteristic curve; 95% CI, 95% confidence interval; CMS, Centers for Medicare & Medicalaid Services; LIDC-IDRI, Lung Image Database Consortium and Image Database Resource Initiative; FP, false positives; MGH, Massachusetts General Hospital; CGMH, Chang Gung Memorial Hospital; LCRAT, Lung Cancer Risk Assessment Tool.
SVM-LASSO model utilizes two CT radiomic features.	Predicted malignancy of pulmonary nodules with 84.6% accuracy, outperforming that of Lung-RADS.	(39)	2018
Combination of ML techniques SVM and sensor array technique LDA	Produced AUCs of 0.91 (95% CI: 0.79−1.00) with LDA and 0.90 (95% CI: 0.80−0.99) with SVM, reflecting strong diagnostic accuracy.	(40)	2018
Feed-forward autoencoder-based deep neural network classifier	Classified pulmonary nodules with 91.2% accuracy, demonstrating it is a prospective tool for lung cancer detection with prospective validation.	(41)	2018
Multi-parameterized ANN based on personal health information	In validation set for predicting lung cancer risk, sensitivity was 75.3% (68.9%−81.6%), specificity was 80.6% (80.3%−80.8%), and AUC was 0.86 (0.84−0.89).	(42)	2018
Deep neural networks applied to ultralow dose PET scans	In automated detection of lung cancer, sensitivity was 91.5% and specificity is 94.2%.	(43)	2018
Collaborative CADx system unifying eye-tracking systems and CADx	Achieved a classification accuracy of 97% and a 91% average Dice Similarity Coefficient for distinguishing between nodules and non-nodules.	(44)	2019
Amalgamated-CNN analyzed with LUNA16 and Ali Tianchi data	Sensitivity was 81.7% and 85.1%, with average false positives per scan of 0.125 and 0.25, respectively.	(45)	2019
Fusion algorithm combining features learned by 3D deep CNN and handcrafted features	Achieved the highest accuracy, AUC, specificity, and sensitivity within all competitive classification models.	(46)	2019
DL model utilizing CMixNet architectures in conjunction with clinical factors	Achieved 94% sensitivity and 91% specificity, which is better than existing methods.	(47)	2019
In-house-deployed AI algorithms complexes consist of DI2IN, region growing, and faster RCNN.	Facilitated reliable 3D segmentation and detection of T1/T2 lung tumors on FDG-PET/CTs	(48)	2019
GLCM technique	GLCM features are accurate in predicting lung tumors, albeit at a slower rate than histogram features.	(49)	2019
LCP-CNN	Achieved an AUC of 89.6% and yielded only one false negative with 234 nodules in 229 patients (19.3%) when using the predefined thresholds.	(50)	2019
DL-based AI algorithm for detecting lung cancer and pulmonary nodules on chest radiographs using data from NLST	Outperformed NLST radiologists in detecting pulmonary nodules on digital radiographs.	(51)	2020
A DL algorithm using pre-trained Res-Net to automatically identify small 18F-FDG-avid pulmonary nodules in PET scans	Accurately identified small 18F-FDG-avid pulmonary nodules, supporting automated clinical PET/CT interpretation.	(52)	2020
A fusion CNN (CXR-LC) using EHRs	Predicted incident lung cancer with high discrimination than CMS eligibility (AUC: 0.755 vs. 0.634).	(53)	2020
A radiomics model applied LASSO feature selection and score calculation, followed by multivariate logistic regression to construct a classification model and nomogram.	Achieved AUCs of 0.836 and 0.809 in training cohort and validation cohorts, respectively to classify benign and malignant pulmonary nodules.	(54)	2020
An AI-CNN prototype (Siemens Healthiness, AI-RAD Companion) to detect pulmonary nodules on LDCT	The agreement between the AI findings and experts was excellent (lung nodules Cohen’s kappa=0.846), with 0.99 sensitivity and 0.708 specificity.	(55)	2021
Fusion CNN integrating four advanced object detectors	Malignancy risk prediction accuracy was slightly lower than the performance of the observers, but the model is useful for assisting less experienced radiologists.	(56)	2021
Innovative model called AdaBoost-SNMV-CNN	Achieved 0.93 sensitivity, 0.92 accuracy, and 0.92 specificity for lung nodules detection and outperformed the majority of the models based on LIDC-IDRI dataset.	(57)	2022
Open-source DL tool (CXR-LC) based on X-ray images and common EHRs	Identified individuals at elevated risk who are potential candidates for lung cancer screening CT.	(58)	2022
A DL-based CAD system developed for a Chinese low-dose CT lung cancer screening program	For lung nodule detection independent of size or type, sensitivity was 90.1% with 1.0 FP/scan, compared with 76.0% and 0.04 FP/scan for double reading (P=0.001).	(59)	2022
An optimized early warning model for lung cancer risk (OWL) based on XGBoost algorithm	Demonstrated robust accuracy and clinical utility for identifying individuals at high risk of lung cancer.	(60)	2023
The Sybil model, trained on NLST LDCTs, requires only one scan and no annotations, enabling real-time use on radiology stations.	AUCs were 0.92 (NLST), 0.86 (MGH), and 0.94 (CGMH) for 1-year lung cancer prediction.	(61)	2023
Recalibrated LCP-CNN	Predicted 1-year lung cancer risk with an AUC of 0.87, exceeding that of LCRAT+CT (0.79) or Lung-RADS (0.69). It showed the highest accuracy for 1-year risk prediction and the lowest likelihood of delayed diagnosis under biennial screening.	(62)	2023
An advanced lung cancer risk prediction model using transfer learning (PResNet classifier) and a modified B-RGS algorithm	Achieved 98.2% accuracy, 98.7% precision, and 97.5% recall rates, which is comparable to other state-of-the-art methods.	(63)	2024
A CT-based deep learning model, METnet, employs grouped convolution blocks to predict MET dysregulation.	Predicted MET dysregulation with 0.746 accuracy and an AUC of 0.793. The method could potentially aid precise diagnosis and therapy at the molecular scale with prospective validation.	(64)	2024
A random forest model combing 18F FDG-PET and 3D-UTE MRI to assess lymph node status in NSCLC	Achieved AUCs of 0.912 and 0.791 in the training and test sets, respectively, demonstrating different degrees of improvement over individual models.	(65)	2024
An ML model based on a dynamic ensemble model for lung cancer detection, which combines four classification models: LGBM, XGBoost, LR, and SVM.	Achieved 76.2% sensitivity, an AUC of 0.77, and 63.8% specificity in 9,940 samples, outperforming five pulmonologists with a sensitivity of 67.4% and a specificity of 70.3%.	(66)	2024
A model for early detection of lung cancer in primary care using ML with deep ‘transformer’ models on EHR data	Achieved an AUC of 0.924 with a positive predictive value of 3.6% and 96.6% sensitivity based on three years of data.	(35)	2024

To date, more advanced CAD systems have been developed and designed to augment radiologists’ classification abilities. CAD systems are broadly categorized into two types: computer-aided diagnosis (CADx) and computer-aided detection (CADe), each serving distinct but complementary roles. By analyzing the morphological features of nodules, CADe is primarily used to identify potential lung nodules or lesions in chest images, helping radiologists locate suspicious areas, whereas CADx is designed to characterize lesions, such as identifying them as malignant or benign. These systems apply different DL and ML algorithms to allow automatic segmentation and identification of lung nodules in chest CT and X-ray, offering radiologists both a “second opinion” and quantitative metrics useful for patient monitoring, while showing comparable sensitivity and negative predictive value to those demonstrated by radiologists (67).

To support the high input demand for testing and training these AI algorithms for lung cancer detection and classification, several open databases are available. For X-rays, the larger dataset is chest X-ray14, which comprises 112, 120 frontal-view X-ray images for fourteen common diseases, including lung mass and lung nodules. For CT, a large database with annotated examinations datasets includes the NLST mentioned previously (10), the Early Lung Cancer Action Program (ELCAP) datasets, and the Lung Image Database Consortium (LIDC)/Image Database Resource Initiative (IDRI) (68). Although CAD systems have not yet been systematically adopted across imaging providers or cancer hospitals, some of these systems have been approved by the Food and Drug Administration (FDA), as shown in Table 2.

Table 2. Recent FDA approvals for AI applications in lung cancer diagnosis.

Device name	Data type	FDA summary	Task	Year
FDA, Food and Drug Administration; AI, artificial intelligence; CT, computed tomography; DL, deep learning; MRI, magnetic resonance imaging; CADe, computer-aided detection.
Riverain ClearRead CT	CT	https://www.accessdata.fda.gov/ cdrh_docs/pdf16/k161201.pdf	The detection of pulmonary nodules in the asymptomatic population.	2016
Arterys Oncology DL	CT MRI	https://www.accessdata.fda.gov/ cdrh_docs/pdf17/K173542.pdf	Segmentation of lung nodules, 3D visualization and cross-comparison of medical images across modalities and time points.	2017
Philips Lung Nodule Assessment and Comparison Option	CT	https://www.accessdata.fda.gov/ cdrh_docs/pdf16/K162484.pdf	Characterization of nodule type, lobe location, quantification of nodule parameters, and comparison over time.	2017
Siemens AI-Rad Companion (Pulmonary)	CT	https://www.accessdata.fda.gov/ cdrh_docs/pdf18/K183271.pdf	Segmentation and volume measurements of identified lung lesions.	2019
Siemens syngo CT Lung CADe	CT	https://www.accessdata.fda.gov/ cdrh_docs/pdf19/K193216.pdf	Detection of solid nodules alerts the radiologists to potentially overlooked regions of interest.	2020
Coreline AVIEW LCS	CT	https://www.accessdata.fda.gov/ cdrh_docs/pdf20/K201710.pdf	Automatic nodules detection by integration with 3rd party CADe, characterization of nodule location, type and related measurements.	2020
Infervision InferRead Lung CT.AI	CT	https://www.accessdata.fda.gov/ cdrh_docs/pdf19/K192880.pdf	Pulmonary nodule detection in asymptomatic patients.	2020
Optellum Virtual Nodule Clinic	CT	https://www.accessdata.fda.gov/ cdrh_docs/pdf20/K202300.pdf	Provide lung cancer prediction CNN score and help with tracking, assessment, and characterization of pulmonary nodules.	2020
MeVis Veolity	CT	https://www.accessdata.fda.gov/ cdrh_docs/pdf20/K201501.pdf	Detection of solid nodules alerts the radiologists to potentially overlooked regions of interest.	2021
IMIDEX VisiRad XR	X-ray	https://www.accessdata.fda.gov/ cdrh_docs/pdf22/K223133.pdf	Identify and mark regions of suspect interest in suspicious lung nodules and masses on chest radiographs.	2023
Qure.ai qCT LN Quant	CT	https://www.accessdata.fda.gov/ cdrh_docs/pdf24/K240740.pdf	Provide advanced quantitative characterization of lung nodules, track volumetric growth, and offer detailed 2D and 3D reconstructions.	2024
Thirona LungQ V3.0.0	CT	https://www.accessdata.fda.gov/ cdrh_docs/pdf23/K232412.pdf	3D segmentation and isolation of sub-compartments, density evaluations, volumetric analysis, fissure evaluation and reporting.	2024

Among current imaging modalities, low-dose CT and X-rays are the most routine screening methods. Low-dose CT is more sensitive and accurate than X-rays, making it the most prominent imaging technique for AI applications in lung oncology currently (69,70). However, the potential of X-ray-based lung cancer diagnosis should not be overlooked due to its widespread availability in healthcare settings, low cost [$21 in Centers for Medicare & Medical Services (71) and £25 in the U.K. National Health Service (72)], and minimal radiation exposure. DL has recently emerged as the primary area of interest in the medical imaging community, resulting in significant progress in enhancing X-ray screening for lung cancer detection. Yoo et al. (51) proposed a deep CNN (DCNN) algorithm based on a residual neural network architecture (13). Similarly, Sim et al. (73) evaluated the clinical efficacy of another DCNN. Lee et al. (74) validated a commercially available DL algorithm from Lunit Company (Seoul, Korea). Schultheissi et al. (75) trained a CNN-based one-stage detector, Retina Net, which was robust to foreign bodies that may have caused misclassification in earlier CNN-based nodule detection systems. Collectively, these algorithms demonstrated diagnostic performance comparable to that of expert radiologists in identifying malignant pulmonary nodules on chest radiographs.

Although CT-based screening has demonstrated a 20%−24% reduction in lung cancer mortality, it does not address a critical issue: the low participation rate in lung cancer screening compared to other established screening programs. For instance, fewer than 5% of the 8 million eligible Americans undergo screening (58). One reason is that Centers for Medicare & Medicaid Services (CMS) eligibility guidelines for CT lung cancer screening, which require detailed smoking history, miss many incident lung cancer cases that should be considered eligible. Recently, a more pragmatic open-source CNN (CXR-LC) using EHR data and patterns from chest radiographs was developed to complement CMS guidelines (53). Its performance was assessed and validated in two large-scale lung cancer screening trials (PLCO and NLST), showing its ability to identify high-risk smokers who would probably benefit from additional CT screening.

For CT scans, a recent meta-analysis of 115 studies evaluated the performance of DL in CT-based respiratory imaging and reported an average AUC of 0.937 for diagnosing lung cancer (76). In one of the earliest studies employing CNNs, from 2016, Setio et al. proposed a multi-view CNN comprising multiple streams of 2D ConvNets (77). This model extracts 2D patches from different CT-oriented planes and reaches high detection sensitivity of up to 90.1% at 4 false positives per scan. However, conventional 2D CNNs lose spatial nodule context information because lung nodule CT data are inherently multi-dimensional. To address this limitation, employing a one-time multi-view strategy, such as a 3D CNN, can further improve the classification and sensitivity of the model, while simultaneously reducing the false positive rate. Huang’s group developed the first DL prediction algorithm without using CAD tools (78). By accounting for both pertinent nodules and non-nodule features in screening chest CT scans, this algorithm can accurately and sensitively assess a person’s risk of developing lung cancer and associated mortality over three years. Ardila et al. (79) developed a promising end-to-end modeling framework with three key components, including a 3D CNN model using LDCT volume, a CNN-based cancer risk prediction model, and a CNN-based region of interest (ROI) detection model. In the NLST dataset, their model identified early stages of lung cancer in 94.4 % of cases, exceeding six expert radiologists. Moreover, 3D scans can capture additional details, such as vascular structures surrounding primary tumor mass. As Etemadi commented in a 2020 Nature interview (80), “The 3D volume starts highlighting areas far away from the tumor. It has shown us some things we wouldn’t expect. We’re opening a whole new area of scientific enquiry.”

In 2023, researchers from the Massachusetts Institute of Technology developed and validated a state-of-the-art AI tool known as Sybil (61), which is independent of clinical data or radiologist-annotated ROI and requires only a single LDCT scan. The model can predict lung cancer development up to 6 years before diagnosis, supporting more personalized screening strategies. For example, it could reduce unnecessary follow-up imaging scans or invasive biopsies for patients with low-risk pulmonary nodules. With the increasing complexity of tasks in the era of big data and the growing demand for computational power, scientists have introduced several innovations concurrently. Although the goal of every developer is to build robust models that generalize across tasks, the resulting models are sometimes weak classifiers that excel only in specific domains. Ensemble learning addresses this limitation by aggregating multiple trained weak classifiers into a stronger, more generalizable classifier. Using an AdaBoost classifier, Huang’s group (45) combined results from three CNN frameworks into an amalgamated model, reducing false positives per scan to 0.125 while achieving a sensitivity of 81.7% for nodules ≥5 mm. Similarly, Venkadesh et al. (81) proposed an ensemble DL algorithm of 2D and 3D CNN to predict nodule malignancy risk directly from voxels of LDCT screening images only. During validation, the model significantly outperformed the PanCan model and demonstrated performance comparable to that of thoracic radiologists (AUC=0.93). Given the higher dimensionality of data, training a 3D CNN requires exponentially more data and computational resources. To address this need, Hussein’s team (82) was among the first to validate the effectiveness of transfer learning in a 3D network of lung nodule characterization. Transfer learning is a type of ML characterized by the sharing of representations and joint optimization. This helps to leverage a large, well-trained dataset and transfer the information to the target task only with fine-tuning. The team proposed a CNN multi-task learning framework trained on the Sports-1M video dataset, and it illustrates state-of-the-art results in risk stratification on LDCT scans from LIDC-IDRI datasets.

To overcome some critical limitations of current DL models, such as the weakness of multi-label lung nodule classification and data imbalance of CNN, Yi et al. established MLSL-Net, which employs multi-label softmax loss (MLSL) as a performance index (83). MLSL can optimize ranking performance both between and within labels simultaneously. In addition, DL models require predefined anchor parameters (size, number, etc.) and often lack robustness when dealing with lung lesions or nodules with various sizes. Luo et al. introduced an anchor-free method called 3D sphere representation-based center-points matching detection network (SCPM-Net) (84). The SCPM-Net achieved a lung nodule detection sensitivity of 89.2% given seven preset false positives per scan on the LUNA16 dataset.

Pathology

Pathology is the study of the nature and effects of disease, primarily through laboratory examinations. It is considered the gold standard for the diagnosis and treatment of lung cancer, playing a critical role in ensuring favorable patient outcomes. Beyond traditional practices such as histological grading and subtyping, modern lung cancer pathology has expanded to include molecular and cellular profiling techniques, such as tumor microenvironment analysis, biomarker detection, and multi-omics approaches (e.g. genomics, proteomics). In parallel, digital pathology has evolved over the past few decades, coupled with advances in ML and big data acquisition, thereby creating numerous opportunities for improving clinical lung cancer diagnosis. A particularly transformative technology is whole slide imaging, which digitizes histopathology glass slides into whole slide images (WSIs). Like glass slides on a microscope, these digital slides can be viewed and navigated on devices with extremely high resolution. However, the lack of fine-grained annotation, the large scale of a standard whole slide image (e.g. 150,000×150,000), and multiple-instance learning problems associated with WSIs present unique computational and practical challenges. Recently, several novel WSI foundation models with state-of-the-art performance (85) have been proposed to fill these gaps. Wang et al. (86) were the first to explore the potential of a small number of coarse annotations in image-level labels for weakly supervised learning. Their model achieves an accuracy of 97.3% with high performance on the TCGA WSIs lung cancer dataset. In a similar way, Kanavati et al. (87) developed an EfficientNet-B3-based CNN incorporating weakly-supervised learning and transfer learning to predict lung cancer in WSIs. This model reported an AUC of 0.98 across four independent test sets for carcinoma detection in WSIs. Researchers at Harvard Medical School (88) unveiled another general-purpose weakly supervised ML framework called Clinical Histopathology Imaging Evaluation Foundation (CHIEF ) in WSI-level tasks in different cancer types. Using over 60,000 WSIs, the model achieved up to 36% higher accuracy in lung cancer detection compared to other state-of-the-art methods, with an AUC greater than 0.9 for cancer cell detection in LUSC.

Histological classification

In this section, we explore how AI may improve the efficiency and accuracy of interpreting these complex pathological features. Lung cancer is a heterogeneous disease with a wide range of clinicopathological features (90) (Figure 6). It is primarily classified into small cell lung cancer (SCLC) and NSCLC, accounting for about 15% and 85% of total diagnoses, respectively. NSCLC can be further subdivided into three major subtypes, including LUAD, LUSC, and large cell carcinoma (89). Therefore, accurate histological classification of these subtypes is essential to improve diagnostic precision and guide targeted treatment strategies.

Figure 6.

Lung cancer histology. Lung cancer is categorized as SCLC or NSCLC (A), with NSCLC further divided into squamous and non-squamous forms (B); Common proportions of oncogenic driver mutations in NSCLC (C). Reproduced with permission from Ref. 89, Elsevier. SCLC, small cell lung cancer; NSCLC, non-small cell lung cancer.

Substantial efforts have been made to distinguish the most common subtypes observed in lung biopsies. CNN is among the most widely applied models as it provides significant assistance to pathologists in determining lung cancer subtypes. Le Page et al. (91) and Coudray et al. (92) trained and employed deep CNN predominantly with the Inception version 3 architecture to classify squamous and non-squamous subtypes in WSIs, achieving an AUC of 0.78 and 0.97, respectively. Nonetheless, the performance of the model in Coudray’s group declined on poorly differentiated biopsy samples (n=34), with AUCs of 0.809 for LUAD and 0.822 for LUSC. To address the diagnostic challenges associated with such poorly differentiated tumors, Kanavati et al. (93) trained a DL model consisting of RNN and CNN to classify indeterminate WSIs, achieving an AUC of 0.99 on a test set of 83 samples. Khosravi et al. (94) developed an innovative fine-tuned DL-based pipeline, termed CNN-Smoothie, to distinguish LUSC from LUAD in 100% high-resolution, locally magnified digital pathology images, reaching classification accuracies between 75% and 90%. In parallel, Janßen et al. (95) developed a fully automated, dual-modality AI-based classification framework. It consists of two parts: a U-net architecture, which is a fully convolutional network for segmenting WSIs, and a second neural network called IsotpeNet for feature extraction and classification, yielding a test accuracy of 94.7 % to classify LUAD and LUSC regions across 16 test sections.

However, the current World Health Organization classification is much more complex (96), especially within LUAD, which encompasses subtypes such as acinar, lepidic, solid, papillary, and micropapillary patterns. Such pronounced heterogeneity within the samples introduces additional bias and complexity in quantifying each subtype, thereby hindering the accuracy of diagnosis. Wei et al. (97) developed a CNN model to identify regions of neoplastic cells and aggregate those classifications to infer predominant and minor histological patterns of LUAD for any given WSI. After evaluation in 143 WSIs, the model demonstrated performance comparable to that of three experienced pathologists. Expanding on this, Pan et al. (98) developed an AI method using a multi-order attention mechanism within CNN, named pyrAmid pooliNg crOss stReam Attention network (ANORAK), to map complex growth patterns in LUAD spatially. The growth pattern proportions can be automatically converted to the predominant pattern in all LUAD subtypes, with a moderate overall agreement between pathologists and AI across four cohorts (n=1,372).

While morphological assessment under light microscopy remains the cornerstone of lung cancer subtyping, poorly differentiated tumors with overlapping morphological features often require adjunctive immunohistochemistry (IHC) for a definitive diagnosis. Currently, decisions about performing IHC are largely subjective and vary among pathologists, often requiring multiple stains, consuming valuable tissue samples, and delaying downstream molecular testing. These limitations underscore the need for AI tools that can assist with IHC interpretation. Koh et al. (99) applied decision-tree and SVM-based ML methods to subtype NSCLC cases using a prospective three-marker IHC panel (e.g. TTF-1, Napsin A, and p40) with equivocal IHC staining patterns. The classification accuracy ranged from 72.2% (e.g. p63, CK5/6) to 91.7% (e.g. TTF-1), depending on the marker patterns. In a similar study, Wang et al. (100) developed a CNN-based cell type classifier, ConvPath, to generate spatial maps of various cell types within the tumor microenvironment. ConvPath achieved an average 90.1% classification accuracy and helped elucidate tumor progression, metastasis, and biomarker discovery (101).

However, many algorithms for IHC face challenges with domain shifts, where models fail to generalize across novel immunostain-cancer type combinations that are not present in the training datasets. These limitations, coupled with their substantial time and resource demands, become particularly significant when evaluating multiple antibody candidates. Recently, Brattoli et al. (102) developed a DL-based Universal IHC (UIHC) analysis method characterized by innovative multi-cohort training-derived models. This system excels in interpreting protein expression results across different cancer and immunostain types without requiring matching with specific training datasets. In a WSI-level lung test set immunostained for programmed cell death ligand 1 (PD-L1) using the 22C3 antibody (n=479), the model achieved a Cohen’s kappa score of 0.652 when compared to the tumor proportion score categories from pathologists and an accuracy of 0.793.

Although many AI-based CAD systems demonstrate excellent performance in detecting lung nodules, they face challenges in handling domain shifts where specific patterns or features do not appear in the training datasets. These issues are critical when evaluating multiple antibody candidates. To address this limitation, Jian et al. (103) developed an open-source CT dataset containing expert-level cancer type annotations and achieved over 50% accuracy across ten models (e.g. ResNet and EfficientNet). Additionally, Diosdado et al. (104) segmented high-resolution images from their LungHist700 dataset into three differentiation levels for both LUAD and LUSC. Through the application of DL and multiple instance learning techniques, they achieved classification accuracy ranging from 81% to 92%.

Analysis of liquid biopsy

Apart from tissue biomarker analysis, liquid biopsy has gained increasing attention and use in lung cancer diagnosis. A liquid biopsy is a less invasive sampling approach that analyzes various types of tumor-derived biomarkers in blood or bodily secretions. Given the low patient adherence to repeated invasive tissue biopsy procedures, common liquid biopsy specimens include urine and blood. Therefore, liquid biopsy is easier to perform and is less burdensome for patients. More importantly, due to its capability for continuous sampling, it enables real-time monitoring of tumors by capturing temporal heterogeneity (105). This diagnostic modality is applicable in various clinical contexts. It offers substantial advantages in lung cancer management (106), including treatment strategies, resistance prediction, therapeutic response monitoring, and prognosis assessment, with a particular emphasis on early diagnosis in this section. Furthermore, the identification and validation of the distinct clinical utility of unique candidate biomarkers remain costly and labor-intensive, and most serum biomarkers are present in extremely low concentrations. Additionally, current techniques, such as plasma next-generation sequencing (NGS) with large panels, will generate complex, high-dimensional, multi-modal data that exceed the scope of manual analysis. Hence, research on the use of AI in analyzing liquid biopsies is needed to enhance diagnostic specificity and sensitivity for lung cancer.

Tumor-educated platelets (TEPs) are blood cells that exhibit altered RNA splicing patterns due to the uptake and enrichment of tumor-derived RNA and proteins within the tumor microenvironment (Figure 7A). Single-platelet RNA-based biomarkers and associated RNA signatures have emerged as robust tools for early diagnosis and treatment monitoring in NSCLC (109). Antunes-Ferreira et al. (110) developed a novel ML algorithm based on particle swarm optimization to analyze TEPs collected from NSCLC patients. Specifically, this algorithm employs multiple candidate solutions inspired by biological swarms, such as bird flocks and fish schools, to iteratively search for optimal results across numerous variables (111). The study identified an 881-RNA biomarker panel from the 4,082 spliced RNAs detected in TEPs and achieved an AUC of 0.88 in a validation cohort comprising 558 participants. Besides, mature circulating miRNAs (Figure 7B) also play an indispensable role in the clinical diagnosis and monitoring of lung cancer, as they are highly stable in body fluids and demonstrate high specificity for lung cancer early diagnosis (112). Zhang et al. (113) used synthetic minority oversampling (SMOTE) to correct subset class imbalance and applied RF to miRNA profiles, yielding an AUC of 0.99 in lung cancer detection.

Figure 7.

Clinical AI application of liquid biopsy in lung cancer diagnosis. TEP, miRNA, exosome, ctDNA, epigenetic information, oncRNA, lipidomics and metabolomics information in the venous blood from lung cancer patients offer potential clinical biomarkers to support early lung cancer diagnosis. Images G and H are reproduced with permission from Ref. 107, American Association for the Advancement of Science and Ref. 108, Springer Science and Business Media LLC; Nature Publishing Group, respectively. AI, artificial intelligence; TEP, tumor-educated platelet; miRNA, microRNA; ctDNA, circulating tumor DNA; oncRNA, orphan non-coding RNA.

Currently, biomarkers such as exosomes (Figure 7C) and circulating tumor DNA (ctDNA) (Figure 7D) have been more extensively studied and have emerged as attractive detection targets for lung cancer (106). Exosomes are nano-sized extracellular vesicles actively secreted into the extracellular space as messengers from various living cells, including lung cancer cells. Shin et al. (114) proposed a method for isolating exosomes, capturing their surface-enhanced Raman spectroscopy (SERS) signals, and applying a CNN-based multiple instance learning model to achieve accurate diagnosis of lung cancer. Notably, the model demonstrated an AUC of 0.93 and 88% accuracy on 100 test samples. Similarly, Lu et al. (115) constructed another AI-assisted SERS strategy using CNN to extract features from exosomes in forty samples, combined with an SVM algorithm to classify health controls and patients with LUAD. This model also demonstrated significant performance, with an 83.3% sensitivity, 83.3% specificity, and an AUC of 0.84. ctDNA, a class of circulating free DNA (cfDNA), provides a real-time snapshot of tumor genomes. Bruhm’s team (116) pioneered genome-wide mutational incidence for non-invasive (GEMINI) detection of cancer, an ML-based screening tool utilizing low-coverage cfDNA whole-genome sequencing to detect lung cancer. GEMINI profiled various somatic alterations in cfDNA and distinguished the regional differences between non-cancer and cancer mutation profiles. It demonstrated an AUC of 0.85 in the LUCAS cohort (n=74 without cancer, n=89 with lung cancer) and detected over 90% of lung cancer patients when followed by CT imaging.

Compared to research on rare somatic mutations in blood, epigenetic profiling (Figure 7E), such as methylome analysis of ctDNA, has gained significant attention (117) due to several notable advantages. First, the cell-type-specific nature of methylation signatures enables the determination of tumor tissue origin. Second, aberrant methylation arises early in tumorigenesis, serving as a sensitive diagnostic marker. Third, because these methylation alterations are often concentrated in specific genomic regions, for example, CpG islands, this allows for efficient analysis via targeted sequencing (118). For instance, Liang et al. (119) developed a soft-margin SVM-based classifier of methylation patterns to assist their high-resolution enhanced linear-splinter amplification sequencing (ELSA-seq) tool in detecting tumor-derived signals. The assay demonstrated its effectiveness by achieving a specificity of 96% while identifying 52%−81% of patients across lung cancer stages IA−III (n=567, including 308 patients and 261 non-cancer controls). Recently, orphan non-coding RNA (oncRNAs) (Figure 7F), which is a class of small RNA, has been identified as arising from cancer-specific genomic reprogramming (120), characterized by the abundant, active, and stable secretion of these RNAs into blood serum from living cancer cells. Karimzadeh et al. (121) developed a semi-supervised multi-input variational autoencoder named Orion to analyze serum oncRNAs in 1,050 untreated NSCLC subjects and matched controls. This approach achieved an overall 87% specificity and 94% sensitivity across all NSCLC stages for the detection of disease.

With advances in sample preparation, mass spectrometry (MS) and chromatographic separation, plasma lipidomics (Figure 7G) and metabolomics (Figure 7H) also represent promising diagnostic avenues. Wang et al. (122) used scRNA-seq to reveal aberrant lipid metabolism in early-stage lung cancer. They identified a targeted panel of nine lipids as features most relevant to early-stage cancer detection. Combining SVM with an MS technique termed analytical power multiple reaction monitoring, the tool called Lung Cancer Artificial Intelligence Detector (LCAID) version 2.0 achieved an AUC of over 95% in both screening (n=1,036) and prospective (n=109) cohorts. For metabolomics analysis, a sparse regression ML model (108) was constructed to classify serum metabolic patterns and detect early-stage LUAD, with specificities ranging from 90% to 93% and sensitivities ranging from 70% to 90%.

However, the wide variety of liquid biopsy biomarkers introduces computational challenges, particularly in integrating heterogeneous data from different sources. This underscores the need for multimodal AI frameworks to build a more complete diagnostic pathway and precision intervention strategies in lung cancer. Wang et al. (123) developed an integrative serum metabolic fingerprints (SMFs) based multi-modal platforms integrating several clinical indices [carcinoembryonic antigen (CEA), image features, and protein tumor markers] via RF for LUAD early detection, demonstrating higher performance (AUC=0.91) than approved clinical models (Mayo Clinic and Veterans Affairs). Zhang et al. (124) investigated the potential of multimodal fusion of radiomics features from CT, extracellular vesicle long RNA (eVLRNA) features, and physician-observed features by employing the XGBoost ML framework for the differential diagnosis of early-stage LUAD. Notably, the fusion model achieved a benign-malignant AUC of 0.94 in a cohort of 146 participants. In recent work published in January, a novel multiomics model named clinic-RadmC (125), which integrates the radiomic, clinical, and circulating cfDNA fragmentomics features in 5-methylcytosine (5mC)-enriched regions, is capable of predicting the malignancy risk of indeterminate pulmonary nodules. In a validation test involving 344 participants, clinic-RadmC yielded an AUC of 0.92 and outperformed single-omics models.

Genetic mutations and gene expression

While immunohistochemical assessment remains a primary focus in the current applications of AI in lung cancer, ML and DL techniques are increasingly applied to the analysis of gene expression profiles. To date, commercial platforms such as NGS have revolutionized genomic sequencing by generating vast volumes of data. Genomic testing for driver mutations has been recommended as standard practice since 2018 by the National Comprehensive Cancer Network (NCCN) guidelines for lung cancer patients (126). However, genomic testing still faces significant barriers to widespread adoption. One major challenge is the long turnaround time for results, which can delay critical treatment decisions, especially for patients with rapidly progressing disease. Additionally, the high costs hinder accessibility, particularly in resource-constrained healthcare systems. Another vital issue is the requirement for high-quality tissue samples, as small or degraded biopsies often yield insufficient material for comprehensive genomic profiling, leading to incomplete or inconclusive results. Thus, developing well-trained AI models to understand key (e.g. EGFR, ALK, KRAS) or novel genetic alterations could be a promising solution for future treatment selection and the assessment of metastatic risk (127).

Several pioneering patch-level DL approaches have been proposed recently. For instance, in another experiment by Courdray et al. (92), their CNN inception model successfully predicted and distinguished the mutational status of six commonly mutated genes (FAT1, KRAS, SETBP1, STK11, EGFR, and TP53) in LUAD from WSI data alone, achieving AUCs from 0.73 to 0.85 in a held-out population. The model’s generalizability was validated on an independent dataset with an AUC of 0.75. Different from the aforementioned approaches, which extract only tumor information, the novel fully automated artificial intelligence system (FAIS) by Wang et al. (128) represents significant progress in predicting EGFR genotypes by incorporating both tumor and whole-lung information from datasets derived from patients in China and the United States, demonstrating superior performance compared to conventional DL-based tumor-focused methods (129). Beyond DL, unsupervised learning also holds its unique potential in predicting genetic mutations. Cook et al. (130) used a novel unsupervised learning model based on RF, which is not only able to classify LUAD and LUSC in small gene expression datasets, but also facilitates the identification of novel mutations, such as PIGX, an oncogenic driver associated with breast cancer progression, meriting further exploration.

The most significant advancement came from Zhao and his colleagues (131), who proposed DeepGEM, an annotation-free bag-level and instance-level co-supervised multiple-instance learning model in 2025. Validated on the largest multicenter dataset to date, DeepGEM accurately predicts gene mutations (KRAS, EGFR, ROS1, ALK, TP53, LRP1B) from routine histopathological slides, achieving AUCs of 0.90−0.97 and accuracy of 0.91−0.97 on excisional biopsy samples, and generating additional spatial gene mutation maps to visualize intratumor heterogeneity.

Challenges and future directions

The integration of AI into lung cancer diagnostics marks a transformative shift in the field of oncology. Although significant progress has been achieved in algorithm development and early-stage clinical applications, numerous barriers that must be addressed before AI can be widely and equitably implemented in routine clinical practice remain. Here, we summarize core challenges related to AI in lung cancer and outline potential future directions to address these challenges (Figure 8).

Figure 8.

Challenges in lung cancer diagnosis. Like other domains in oncology, AI applications in lung cancer diagnosis also face significant challenges, including curating and sharing data across institutions, the existence of bias from training through to deployment, and integrating into real-world lung cancer clinical workflows. Reproduced with permission from Ref. 132, American Association for Cancer Research.

Data curating and sharing

The development of robust, generalizable, and clinically insightful AI models for lung cancer diagnosis critically depends on the availability of large-scale, high-quality training datasets that are ideally obtained from multisite, independent resources (133). This requirement is further highlighted by the inherent heterogeneity of lung cancer, reflecting variability in tumor biological characteristics, clinical manifestations, and treatment protocols across institutions. Individual healthcare institutions are often constrained by small patient cohorts, limited demographic diversity, and site-specific clinical practices, with rare lung cancer subtypes facing additional challenges. Despite the widely acknowledged significance of data sharing, substantial barriers persist, such as a lack of data interoperability, concerns over patient privacy, and intellectual property protection (132). These obstacles are governed by regulatory frameworks such as the EU General Data Protection Regulation (GDPR), the Federal Policy for the Protection of Human Subjects, and the Health Insurance Portability and Accountability Act (HIPAA) (134).

To address these challenges in data privacy and sharing, two primary options have been explored by the lung cancer research community. Current AI assessments for lung cancer diagnosis are typically one-off snapshots based on only one data modality, such as CT, and therefore do not evaluate health as a continuous process. In contrast, to deliver the best care, radiologists rely on the totality of information, such as family history, genomic data, prior clinical notes, lab results, pathology images, and radiology scans (Figure 9).

Figure 9.

Opportunities in data curating and sharing in lung cancer diagnosis. (A) Multimodal data integration from lung cancer patients; (B) Workflow of federated learning from federated server to local clients; (C) One specific example of federated learning: flexible federated learning. Multiple medical centers collaborate to train a shared backbone network and individual lung cancer classification heads using their datasets (e.g. chest radiographs). Images A, B and C are reproduced with permission from Ref. 135, Elsevier, Ref. 136, MDPI AG and Ref. 137, Nature Publishing Group UK; Nature Publishing Group, respectively. CT, computed tomography.

One approach, involving the development of multimodal lung cancer datasets and multimodal AI, partially bridges this gap (Figure 9A). For instance, Niu and his colleagues (138) developed a large-scale, multimodal question-answering model called M3FM to process various combinations of clinical data types, CT series, and other parameters involved in lung cancer screening. This study illustrates that M3FM could improve lung cancer risk prediction by up to 20%. Oncu and his colleagues (139) integrated CNN and artificial neural network (ANN) to develop a novel multimodal framework for classifying lung tissue histological categories, achieving a weighted accuracy of 92%. The second strategy is federated learning (FL) (Figure 9B), where data are stored locally at each institution, and model training is collaboratively performed through iterative updates exchanged among sites. This decentralized paradigm protects patient privacy and institutional autonomy while addressing key concerns related to data sharing (Figure 9C). Although FL introduces non-trivial challenges such as communication overhead, system heterogeneity, and privacy-preserving optimization, recent advancements have significantly narrowed the performance gap between FL methods (distributed approaches) and centralized learning methods. Importantly, FL has successfully demonstrated its growing potential in clinical applications of AI through multiple oncology domains (140).

In line with trends in lung cancer applications, Zhu et al. (141) developed a three-stage FL framework for pulmonary nodule segmentation and detection based on CT images. This framework integrates U-Net++ architecture for segmentation, an EHRs-guided classifier discriminator, and a RF detector. The model exchanges parameters between client devices and a central server to improve segmentation fidelity while enhancing robustness against non-independent and identically distributed (IID) data through alignment of predicted image attributes with EHRs-derived descriptors. This adaptive knowledge-sharing framework achieved a mean dice similarity score of 76% and an 89% average competition metric performance on decentralized datasets, outperforming multiple baseline methods in centralized ML. Simultaneously, a study conducted by Rajendran et al. (142) implemented FL using EHRs across two institutions to classify lung cancer risk based on tobacco and radon exposure. They employed two federated strategies: single-round FL and iterative weight transfer between ANN and LR models. Their results demonstrated that only ANN-based models benefited from federated training, improving prediction accuracy from 68% in centralized training to 74% in FL, while LR performance remained essentially unchanged. This suggests that the benefits of FL are model-architecture-dependent and more pronounced for non-linear architectures like artificial neural networks. To further promote cross-institutional collaboration, Tayebi Arasteh et al. (137) proposed a flexible federated learning (FFL) architecture for chest X-ray classification across five global centers with varying labeling protocols. By decoupling the feature extraction backbone from task-specific classification heads, their FFL framework enables joint training on semantically related but label-diverse datasets. This model achieved superior AUC performance compared to local and traditional FL methods in large-scale clinical applications of lung cancer imaging.

Bias and fairness

Creating fairness in performance and use across populations is a critical challenge in AI in medicine. This issue particularly applies in lung cancer diagnosis, where disparities in demographic background, socioeconomic status, or geographical location perpetuate inequities in early detection and treatment outcomes. For example, one primary source of bias lies in the biological heterogeneity of cancer across populations (143). Genetic ancestry has been shown to influence tumorigenesis and cancer susceptibility, with evidence demonstrating notable differences in lung cancer incidence among ethnic groups, even after controlling environmental risk factors such as smoking. For instance, African American and Native Hawaiian populations exhibit disproportionately higher susceptibility to lung cancer compared to other demographic groups (144). By adopting population-agnostic assumptions or being trained predominantly on data from a narrow demographic base, AI models that disregard such inter-population biological variability have risk producing skewed outputs in underrepresented groups, since genetic ancestry can influence certain traits and disease susceptibility. This can lead to biased patient classification or equitable subgroup allocation, prolonged diagnostic times, and ultimately, inferior clinical outcomes for specific ethnic populations. For example, Rakaee et al. (145) evaluated two open-source AI pathology models for predicting EGFR mutation status from hematoxylin and eosin (H&E) slides in LUAD. Although the best-performing model achieved strong AUCs of 0.83 and 0.81 in US and European cohorts, respectively, ancestry-stratified analysis revealed a substantially lower AUC of 0.68 among patients of Asian ancestry. In lung cancer diagnosis, such limitations lie in the lack of diversity within widely utilized public genomic repositories, including TCGA and the Genomic Initiative for NCI Ecosystems GENIE (146). Moreover, socioeconomic disparities further compound these challenges. Most current AI diagnostic tools in oncology rely heavily on high-resolution imaging modalities such as MRI, PET and CT, which remain inaccessible to lung cancer patients in most low- and middle-income countries (LMICs) (147). In LMICs, there is only one CT scanner available for every 1.7 million people, whereas in high-income countries one scanner is available for every 25,000 people (148). Despite the theoretical global reach of AI tools, particularly those enabled by internet-accessible LLMs, practical deployment in resource-limited settings remains constrained by systemic limitations in digital infrastructure and hardware availability. Alarmingly, most AI models have been developed and validated using datasets that exclude these contexts, rendering them unsuitable for clinical translation in LMICs. Consequently, failure to consider the socio-technical realities of deployment environments perpetuates a form of structural inequity, whereby populations in great need of diagnostic innovation are systematically excluded from its benefits. This disconnect raises serious concerns about the global applicability of such models and emphasizes the development of tools that function with more readily available diagnostics, such as chest X-rays, portable ultrasounds, and cytological smears (149) for low-resolution data or images. For example, Qure.ai provided an AI algorithm for the detection of lung nodules using X-rays as a more accessible alternative to high-cost imaging tools (150).

Integration in clinical workflows

The successful application of AI in lung cancer diagnosis depends on its smooth integration into clinical workflows, a goal that is currently hampered by practical and technical barriers (146). In practice, the integration of AI into routine clinical workflows remains uncommon due to limitations related to user interface design, processing speed, program complexity, internet connectivity, and system resource demands. More infrastructure needs to be constructed before AI can transition from supportive tools to frontline diagnostic agents in lung cancer care.

Among these limitations, determining the optimal format for AI outputs requires a clear understanding of the clinical objectives and the intended use of the system. In this context, the lack of transparency commonly associated with “black box” AI models remains a frequently noted challenge in lung cancer diagnosis. Despite achieving diagnostic accuracy on par with seasoned clinicians in the context of lung cancer, as shown in the studies mentioned above, current AI systems fall short in delivering clinically interpretable rationales or offering mechanistic and pathological explanations as physicians do. Encouragingly, advances in imaging technologies and interface design are progressively improving the interpretability and trustworthiness of AI models. One notable example is a collaborative-CAD (C-CAD) system with eye-tracking by Bagci and his colleagues (44), which enables AI to interact with radiologists by incorporating their visual attention patterns into diagnostic analysis (Figure 10). By converting gaze data into graph-based representations and applying attention-based stratification, the system identifies clinically relevant ROI. A 3D multi-task CNN then performs simultaneous segmentation and false positive reduction within these regions. This synergistic approach between human expertise and AI not only enhances diagnostic accuracy but also improves transparency, achieving a 97% classification accuracy and a Dice score of 91% in lung cancer screening tasks.

Figure 10.

C-CAD system workflow. Reproduced with permission from Ref. 44, Elsevier. C-CAD, collaborative computer aided diagnosis. ROI, region of interest; CNN, convolutional neural network.

Conclusions

AI is beginning to redefine the diagnostic landscape of lung cancer. Across radiology, digital pathology, liquid biopsy, EHRs and molecular profiling, AI has shown considerable potential to improve early detection, refine histological and molecular classification, and support more personalized diagnostic decision-making. Despite these advances, most current models remain confined to retrospective, narrowly curated settings, and their real-world clinical value has not been fully established.

The next phase of progress will depend on a shift from modality-specific models towards integrative and clinically deployable frameworks. In particular, multimodal foundation models capable of jointly learning from imaging, pathology, genomic, liquid-biopsy and longitudinal clinical data may offer a more complete representation of tumor biology and patient context than any single modality alone. At the same time, federated learning networks provide a promising strategy to overcome institutional data silos and privacy barriers, enabling collaborative model development across diverse populations and healthcare systems without requiring centralized data sharing. Equally important, prospective AI trials and real-world implementation studies are now urgently needed to determine whether these technologies can improve diagnostic accuracy, reduce unnecessary procedures, shorten time to diagnosis, and ultimately translate into better patient outcomes.

Beyond technical performance, the future success of AI in lung cancer diagnosis will depend on whether these systems are interpretable, equitable and compatible with routine clinical workflows. Models that perform well only in highly selected datasets are unlikely to achieve meaningful clinical impact unless they remain robust across populations, practice settings and resource levels. Thus, the central challenge is no longer simply to build more accurate algorithms, but to develop trustworthy AI systems that can be validated prospectively, integrated responsibly and used effectively in everyday care. With continued collaboration among clinicians, data scientists, engineers, regulators and industry, AI has the potential not merely to augment existing diagnostic pathways, but to reshape lung cancer diagnosis into a more precise, scalable and patient-centered discipline.

Funding Statement

This work was supported by the National Science and Technology Major Project (No. 2025ZD0551200), the Innovation Research Group Program of the Sichuan Provincial Natural Science Foundation for High-Altitude Medicine (No. 2026NSFSCZY0080), and the Sichuan Science and Technology Program (No. 2025ZDZX0097).

Author contributions

Conceptualization and methodology: WL Hu, YC Yao; Writing – original draft, visualization: WL Hu, G Wang, L Ren, XP Wu; Writing – review & editing: J Hu, WH Zhuang, YC Yao, CD Wang, F Ye, WJ Mao, WC Hu; Supervision: WC Hu.