Table 4.
Recent AI-based studies in radiogenomics for various oncology care.
| CIT * | Motivation | Radiomics Information |
Genomics Information | AI-Based Models | Dataset | PM $ | Performance Measure |
Outcomes |
|---|---|---|---|---|---|---|---|---|
| (a) | ||||||||
| [133] | Risk assessment in breast cancer | Traditional Radiographic, Texture Analysis, Pretrained CNN for deep features | BRCA1/2 | SVM Model |
456 clinical FFDM patients | AUC | BRCA1/2 gene-mutation: AUC = 0.86 unilateral cancer patients: AUC = 0.84 |
Fusion classifiers performed significantly better. Deep features performed very well. |
| [134] | Association Assessment of imaging phenotype with molecular subtype. | 529 tumor and tissue imaging features. | Luminal A, ER, PR, EGFR, Ki67, HRE2 |
ML-based multivariate models | 922 patients (Proprietary data) | AUC | Luminal A subtype: AUC = 0.697, TNBC: AUC = 0.654, ER: AUC = 0.649%, PR: AUC = 0.622% |
Application in early diagnosis with association relation between the MRI-based imaging features and genomic features. |
| [135] | Prediction of molecular subtypes and prognostic biomarkers | CT perfusion features include lymph node status, tumor grading, tumor size | ER, PR, Luminal A, Ki67, HRE2 |
SVM, RF, Decision tree, Naïve Bayes | 723 patients (Proprietary data) | AUC, ACC. | Random Forest: AUC: 0.86, Tumor grade and size: AUC: 0.88 and 0.85 ER and PR status: AUC: 0.88 and 0.85 HER2 and Ki67: AUC: 0.88 and 0.85 Molecular subtypes: AUC: 0.82 |
Helps in non-invasive diagnosis by performing a depth analysis of the relation between molecular subtype and CT-based imaging features. |
| [123] | Classification of breast cancer molecular subtype | Deep features | Luminal A | Google Net, VGG, & CIFAR network | 272 patients (Proprietary data) | AUC | Deep features: AUC = 65% TL: AUC = 60% |
Provides a non-invasive way to detect Luminal A tumor subtype with the help of DL. |
| [136] | Diagnosis of breast cancer | Features: tumor shape, size, morphology, enhancement texture, enhancement-variance kinetics, and kinetic curve assessment. | RNA sequencing, KEGG, GSEA | Radiogenomics | TCGA/TCIA | - | - | Detailed analysis of the association between the gene pathways and imaging features provides a future direction for the non-invasive diagnosis of breast cancer. |
| (b) | ||||||||
| [137] | CAD system | Traditional features: morphological, intensity, and textural features. | IDH1 | Logistic regression |
32 (WT IDH) and 7 (mutant IDH) patients from TCIA | ACC, SENS, SPEC | Morphology: ACC = 51% (20/39), SPEC = 50% (16/32), SENS = 57% (4/7); Intensity: ACC = 59% (23/39), SPEC = 59% (19/32), SENS = 57% (4/7); Texture: ACC = 85% (33/39), SENS = 86% (6/7), SPEC = 84% (27/32). |
Non-invasive diagnosis of tumor CAD system. |
| [138] | Prediction of IDH1 for LGG tumor | Texture, intensity, shape, and wavelet features. | IDH1 | CNN | 151 patients from the Department of Neurosurgery, Huashan Hospital. |
AUC, ACC, SPEC, SENS, NPV, PPV, MCC | IDH1 estimation, in radiomics method: AUC = 86%, DLR: AUC = 92%, DLR based on multiple-modality MRI, AUC = 95% |
Provides a direction for early researchers to choose the models as it gives a comparative performance analysis of DL-based radiomics and normal radiomics methods. |
| [139] | Classification of MGMT promoter | Nine textures, histogram, gray level-based features | MGMT, IDH1 | XGBoost | 262 subjects from TCGA and TCIA | AUC, ACC, SENS, SPEC, F1 score | AUC = 89.6% | Yields better treatment planning for patients with IDH1 wildtype GBM in the primary diagnosis phase. |
| [140] | Characterization of genetic heterogeneity over enhancing and non-enhancing tumor. | MR imaging texture features | EGFR, PTEN, PDGFRA, CDKN2A, TP53 and RB1. | Predictive decision-tree models. | 18 GBM Patients (Proprietary data) |
ACC, LOOCV, | Accuracy for 6 driver genes: EGFR = 75%, PDGFRA = 77.1%, CDKN2A = 87.5%, TP53 = 37.5%, RB1 = 87.5%, |
In primary diagnosis and better treatment planning of patient with GBM. |
| (c) | ||||||||
| [141] | Prediction of EGFR and KRAS mutation | Texture and Non-texture features | EGFR and KRAS | Ensemble model based on ML and CNN. | 99 patients from the TCIA | AUC, ACC, SENS, SPEC | AUC for ML models: EGFR = 75%, KRAS = 72%, For DL models: EGFR = 82.8% KRAS = 72.2%. |
Enhancing the performance of non-invasive diagnosis of lung cancer by predicting EGFR and KRAS mutation in a small dataset |
| [142] | Prediction histology and tumor Recurrence. |
117 radiomic features based on GLM. | KRAS, TP53, EGFR | ML and Generalized linear model | 151 Institutional databases | ACC, F1-score | AUC = 87% | Compressive analysis of showing a correlation between genomic and tumor subtype. |
| [143] | Prediction of tumor Recurrence in Non-small cell lung cancer (NSCLC. |
Handcrafted: GLCM, histogram-based statistics, Laplace of Gaussian. Deep features |
The RNA-sequencing. | Genotype-guided radiomics method | 162 patients from the TCIA dataset | AUC, ACC, SENS, SPEC | AUC = 76.67% and ACC = 83.28% | Showing an effective prediction method with low cost and improved accuracy. |
| [144] | Risk prediction of lung cancer | Feature: patient’s current and prior CT volumes | - | 3D CNN | 6716 National Lung Cancer Screening Trial cases | AUC | AUC = 94.4% in risk prediction | Clinical validation proves its low-biased performance and allows enhancement of the screening process via CAD and automated screening to the radiologist. |
| [143] | Classification of histology subtype | 1695 quantitative radiomic features (LOG, GLCM) | Histological subtypes | Incremental Forward Search and SVM | 278 patients (181 NSCLC and 97 SCLC) | AUC | SCLC vs. NSCLC: 74.1%, SCLC vs. AD: 82.2%, SCLC vs. SCC 66.5% and AD vs. SCC: 66.5% | Detailed analysis of phenotypic variation exists among various lung cancer histological subtypes in CT images. |
| [145] | Classify somatic mutations | Radiomic signature including tumor volume and maximum diameter, intensity. | EGFR and KRAS | Random Forest | Four independent datasets (PROFILE, TIANJIN, MOFFITT, xHARVARD-RT) |
AUC | AUC: 80% EGFR+ and KRAS+, 69% with EGFR+ and EGFR−, 63% with KRAS+/KRAS− radiomic signatures | Relation between the imaging phenotype captured with a genotype and EGFR mutant tumors has a clinical impact in selecting patients for targeted therapies. |
| (d) | ||||||||
| [146] | Prediction of early recurrence of HCC | 21 CT image-based radiomic signature | - | Machine learning | Proprietary data (215 HCC patients) | AUC, SENS, SPEC | Radiomic features: AUC = 81.7%, clinical data AUC = 78.1%, combined model AUC = 83.6% | Shows a direction towards preoperative estimation in early prediction of recurrence less than 1 year and helps radiologists with better treatment planning. |
| [147] | Diagnosis in HCC | Features include texture features, first-order histogram, and GLCM. | TP53, TOP2A, CTNNB1, CDKN2A and AKT1 | Machine learning | 27 patients from TCGA, and TCIA. | AUC, SPEC, SENS. | TP53: AUC = 86.61%, TOP2A 78.0%, CTNNB1: 86.8% |
Ability to categorize HCC tumors on a genetic level which helps the radiologist for early diagnosis of HCC patient |
| [148] | Prediction of progression-free survival (PFS) and overall survival in uHCC | SUV statistics, co-occurrence matrix, neighborhood intensity, neighborhood gray level dependence | Alpha-fetoprotein | Machine learning | Proprietary data (371 patients) | - | For survival PFS: [PFS-pPET-RadScore < 0.09] vs. 4.0 mo [95% CI(Confidence Interval): 2.3–5.7 mo] in high-risk group. median of 11.4 mo [95% CI: 6.3–16.5 mo] in a low-risk group. [OS-pPET-RadScore < 0.11] vs. 7.7 mo [95% CI: 6.0–9.5 mo] in high-risk group.[PFS-pPET-RadScore > 0.09]; p = 0.0004) and OS(Overall Survival): median of 20.3 mo [95% CI: 5.7–35 mo] in low-risk group. [OS-pPET-RadScore > 0.11]; p = 0.007) |
Helps in better treatment planning for the patients undergoing transarterial radioembolization using Yttrium-90. |
| [149] | Prediction of overall survival in HCC | Features including maximum diameter, histogram-based texture features | AFP, DCP | Machine learning | 178 patients (Proprietary data | Kaplan-Meier analysis | Random survival forest model’s high and low predicted individual risks are p = 1.1 × 10−4 for DFS, 4.8 × 10−7 for OS respectively, and based on the multivariate Cox proportional hazards model, high predicted individual risk (hazard ratio = 1.06 per 1% increase, p = 8.4 × 10−8) |
OS prediction shows a better direction towards the improving survival of the patient. |
| (e) | ||||||||
| [150] | Diagnosis of prostate cancer | Features: Gabor texture, Gleason grade, and gland lumen shape | Gleason score, QH | ML | 54 patients from UPenn and 17 patients from SV | AUC | Prediction of Gleason grade based on Gabor texture features AUC = 69%, prediction of QH based on gland lumen shape features AUC = 0.75 | Relation between in vivo T2w MRI phenotype predicting prostate cancer status. |
| [151] | Prediction of tumor aggressiveness in prostate | Multiparametric (mp) MRI and 68Ga-PSMA-PET/CT phenotypes. | CNAs | - | 5 patients of the University of Heidelberg | - | Highly significant CNAs (≥10 Mbp) were found in 22 of 46 biopsies. | Correlating the most aggressive lesion with imaging features helps in future prostate cancer diagnosis and prognosis. |
| [152] | Diagnosis of prostate cancer | Texture Based features, morphological features | - | LSTM and ResNet101 | 230 for MRI by the Health Insurance Portability. | AUC, SENS ACC, SPEC, NPV, PPV, MCC | LSTM: AUC = 0.9999, ResNet − 101AUC = 100% |
Detection of prostate cancer prediction is better on a DL-based model. |
| [153] | CAD for prostate cancer | 564 radiomic features of texture, intensity, shape, and orientation. | - | CNN DL, radiomic model. | 644 patients from healthcare centers in Netherland. | AUC, ACC, SENS, SPEC. | DL: AUC = 89%, Active Surveillance dataset using Radiomic model AUC = 83% |
Developed a tool for significant-PCA classification with radiomic model. |
| (f) | ||||||||
| [154] | predicting early recurrence in HGSOC | Radiomic nomogram | - | KNN, SVM, and LR | Proprietary data (256 patients) | AUC, Kaplan-Meier survival analysis and Decision curve analysis | C-index for clinical factors model = 82% [95% CI (0.75–0.88)] (training set) (validation set): 77% [95% CI (0.59–0.90)] Radiomics nomogram C-index = 0.91 [95% CI (0.85–0.95)] (training set),, the C-index = 0.85 [95% CI (0.69–0.95)] (validation set) |
Helps in early individualized recurrence prediction in patients with HGSOC |
| [155] | Classification of ovarian cancers (SOCs). | Features include Histogram, Formfactor, GLSZM, RLM. | CEA, CA125 | ML | Proprietary data (110 patients) | AUC, SPEC, SENS | AUC = 85.4% | The model using radiomic features of arterial phase of CT with clinical features is the first study to develop a useful tool for differentiating the POC and SOC. |
| [156] | Prediction of PM in ovarian cancer. | Radiomics features: T2WIs, T2WIs, multi-value DWIs | - | LR | 89 patients Shanxi Medical University | AUC. | AUC = 96.3% (training) AUC of 0.928 (validation) |
Treated as a biomarker for risk stratification. |
| [157] | Prediction of PFS in advanced HGSOC. | Imaging features | Pelvic fluid, and CA-125 | 261 patients (Proprietary data) | AUC | AUC = 96.9% | The quantitative solution to predict PM in OC patients. | |
| [158] | Assessments of CT imaging features of HGSOC | Ovarian mass, size of pleural effusions and ascites, mesenteric implants and infiltration, lymphadenopathy, and distant metastases. | - | ML | 92 patients (Proprietary data) | Estimates of Krippendorff α and coverage probabilities | Pleural effusion and Ascites: α = 0.78, Intraparenchymal splenic metastases: α = 0.08 |
Experimental results show evidence of the clinical and biological validity of these image features. |
| (g) | ||||||||
| [159] | Prediction of mutation status and prognostic values in colorectal cancer | - | PIK3CA exon 9 and 20, NRAS exon 2 and 3, KRAS exon 2, 3 and 4, and BRAF exon 15 | PCR and direct sequencing | 353 CRC patients at Zhongda Hospital | - | 13.9% (49 out of 353) CRC patients carried mutations at RAS exons outside the KRAS exon 2. | Provides the importance of these novel molecular features in CRCs |
| [160] | Prediction of KRAS/NRAS/BRAF mutations in colorectal cancer (CRC). | Features include shape features, GLCM features, and GLRLM features. | KRAS/NRAS/BRAF | RELIEFF and SVM | 117 patients (Proprietary data) | AUC, SENS, SPEC. | Prediction of KRAS/NRAS/BRAF mutations, AUC = 86.9% | The predicted association is useful for the analysis of tumor genotype in CRC and hence helps in therapeutic strategies. |
| [161] | Prediction of KRAS mutations using MRI | polypoid pattern, axial tumor length | KRAS | - | 275 patients (Proprietary data) | - | The frequency of KRAS mutations was higher in the N2 stage (53.70%), and polypoid tumors (59.09%). | Helps in finding the imaging predictor of KRAS which helps the radiologist to make a better therapy strategy. |
| [162] | Prediction of the mutation status molecular subtype in colorectal cancer. | Features: tumor size, degree of the tumor, C-reactive protein level, differentiation, and TNM stage | KRAS | Machine Learning | 58 patients (Proprietary data) | AUC | AUC on predicting the KRAS mutant = was 86.5% | Provides a higher performance for the prediction of the KRAS mutation status in CRC. |
| [163] | Classification of imaging predictors. | - | KRAS | Naive Bayes classifier | 457 patients (Proprietary data) | - | - | Ability to identify disease course relation with mutated oncogenes and provides a cheaper, quicker substitute for genome sequencing. |
| (h) | ||||||||
| [164] | Predicting of lymph node metastasis. | Features include intensity features, shape, GLZLM, GLRLM, GLCM. | - | SVM | 490 patients (Proprietary data). | AUC | LN+, AUC = 82.4% (training and validation), AUC = 76.4% (test data) |
Shows a promising tool for the preoperative prediction of LN status in patients with GC. |
| [115] | Prediction of PD-L1 status in gastric cancer (GC). | - | PD-L1 | SVM and RF | 358 patients of Nanjing Drum Tower Hospital | AUC | Using SVM AUC = 70.4%, 79.9% in primary and validation cohort. | A promising tool to predict PD-L1 status and helps to improve clinical decision-making about immunotherapy. |
| [165] | PET based radiomic model for prediction of PM of gastric cancer. | Features including GLCM, GLZLM, NGLDM, and GLRLM | CA 125, PM, SUVmax. | Multivariate LR | 355 patients (Proprietary data). | AUC | Radiomics model: AUC = 86%, 87%, Clinical prediction model: AUC = 76% and 69% |
Provides a novel tool for predicting peritoneal metastasis of gastric cancer. |
| [166] | Prediction and investigation of the efficiency of neoadjuvant chemotherapy in survival stratification. | Texture, filter transformed, and wavelet features. | - | Randomized tree | 106 patients (Proprietary data) | AUC | Rad_score: AUC = 82%, clinical score: AUC = 62% | Effective prediction treatment for neoadjuvant chemotherapy and stratifying patients into various survival groups. |
| [167] | Predict the status of lymph node metastasis (LNM). | Shape-based features, first-order based, texture-based features. | Genome stable, Epstein–Barr virus-positive, chromosomal and microsatellite instability. | Multivariate LR | 768 patients (Proprietary data) | AUC | AUC = 92% (training cohort), AUC = 86% (validation cohort) AUC = 85% (EGC patients) |
Serves as a non-invasive tool for preoperative evaluation of LNM in EGC. |
Note: ACC—accuracy; SPEC—specificity; SENS—sensitivity; HCC—hepatocellular carcinoma; NSCLC—non-small cell lung cancer; small-cell lung cancers (SCLC); LOG—laplacian of Gaussian; HGSOC—high-grade serous ovarian cancer; CAN—chromosomal copy number alterations; CA 125—carbohydrate antigen 125; QH—quantitative histomorphometry; UPenn—University of Pennsylvania; SV—St. Vincent’s Hospital; GLZLM—gray-level zone length matrix; NGLDM—neighborhood gray-level dependence matrix; GLRLM—gray-level run-length matrix; GLCM—Gray-level co-occurrence matrix; MCC—Mathews correlation coefficient; GLSZM—gray-level size zone matrix; RLM—run-length matrix; PM—peritoneal metastasis; PES—progression-free survival; LNM—lymph node metastasis; T2-weighted images—T2WIs; fat suppressed—T2WIs; diffusion-weighted images—DWIs; Logistic Regression—LR; Machine Learning—ML; CIT *—citations; PM $—Performance metrics.