Artificial Intelligence and Radiomics for Non-Invasive Prediction of Genomic Alterations in Solid Tumors: A Scoping Review

Abstract

This scoping review aims to map the current extent, range, and nature of evidence on the use of artificial intelligence (AI), machine learning, deep learning, radiomics, and radiogenomics applied to medical imaging for the non-invasive prediction or characterization of genomic alterations and molecular phenotypes in patients with solid tumors. Genomic alterations play an important role in the diagnosis, prognosis, treatment selection, and therapeutic response of patients with cancer. Conventional identification of genomic alterations commonly requires tissue acquisition and molecular testing, which may be invasive and may not fully capture tumor heterogeneity. Imaging-derived biomarkers, radiomics, radiogenomics, and AI-based approaches have emerged as potential non-invasive methods for inferring molecular and genomic characteristics from routinely acquired clinical imaging. The review will follow the Joanna Briggs Institute (JBI) methodology for scoping reviews and will be reported in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews (PRISMA-ScR). Eligibility criteria will be structured according to the Population, Concept, and Context (PCC) framework. The population will include human participants with histologically, cytologically, clinically, or radiologically established solid tumors. The review will encompass a broad range of solid malignancies, including lung/NSCLC, glioma/CNS tumors, breast, colorectal, liver, prostate, gastric, renal, head and neck, ovarian, pancreatic, and other solid tumors. Studies involving mixed cancer populations will be eligible when relevant solid-tumor results are separately reported or can be clearly extracted. The concept of interest is the application of AI, machine learning, deep learning, radiomics, radiogenomics, or computer-aided analysis to medical imaging for the non-invasive prediction, classification, characterization, or inference of genomic or genetic alterations and molecular phenotypes. Eligible genomic and molecular targets will include gene mutations, genomic alterations, molecular biomarkers, molecular subtypes, receptor status, gene-expression signatures, and related molecular classifications. Imaging modalities will include computed tomography (CT), low-dose CT, magnetic resonance imaging (MRI), dynamic contrast-enhanced MRI, multi-sequence MRI, positron-emission tomography/PET-CT, mammography, digital breast tomosynthesis, ultrasound, and other clinically relevant imaging modalities. A comprehensive literature search will be conducted from database inception to the final search date using PubMed/MEDLINE, Scopus, Web of Science Core Collection, and Embase, where available. Google Scholar will be used as a supplementary source for grey literature and citation discovery. Backward reference-list screening, forward citation searching, and hand-searching will also be performed where appropriate. No date restriction will be applied at the database level. English-language publications will be included, while non-English studies with an accessible English abstract will be considered during screening, with translation sought where feasible. Original human research will be eligible, including prospective and retrospective cohort studies, case-control and cross-sectional studies, diagnostic accuracy and prediction-model studies, retrospective imaging-genomic correlation studies with an AI/radiomics component, and case series with sufficient methodological information. Relevant theses and dissertations may also be considered as grey literature. Reviews, editorials, commentaries, protocols without original study data, animal-only, phantom-only, and purely in-vitro studies will be excluded. Studies using imaging solely for cancer detection or diagnosis without attempting to characterize or predict a genomic, genetic, or molecular alteration will also be excluded. Two reviewers will independently perform title/abstract and full-text screening. Disagreements will be resolved through discussion and, when necessary, adjudication by a third reviewer. A standardized and piloted data-charting form will be used by two reviewers to independently extract information from included studies. Data will be charted on bibliographic characteristics, study design, cancer type and population, genomic target, reference standard, imaging modality and acquisition characteristics, radiomics and AI methodology, model development, validation strategy, reported performance measures, comparators, limitations, reproducibility, explainability, and potential clinical implementation. Performance measures such as area under the receiver operating characteristic curve (AUC), sensitivity, specificity, accuracy, precision, recall, F1-score, calibration, predictive values, decision-curve or clinical utility measures, and other reported metrics will be extracted where available. The evidence will be synthesized descriptively and narratively rather than quantitatively pooled, given the anticipated heterogeneity across cancer types, genomic targets, imaging modalities, algorithms, datasets, and outcome measures. The review will map the distribution of evidence across cancer types and genomic alterations, summarize imaging and AI/radiomics methodologies, characterize model-development and validation approaches, and describe reported model performance where sufficiently comparable. Expected outcomes include an evidence map demonstrating the current landscape of AI- and radiomics-based non-invasive prediction of genomic alterations in solid tumors; identification of commonly studied cancer–genomic alteration combinations, imaging modalities, and AI methodologies; characterization of dataset and validation practices; and identification of under-studied areas and methodological gaps. Particular attention will be given to issues affecting reproducibility and clinical translation, including small sample sizes, class imbalance, feature-selection concerns, overfitting, inconsistent imaging protocols, lack of external or multicenter validation, limited prospective evaluation, and inadequate clinical validation. The review is expected to provide a structured overview of the field and identify priorities for future research, including standardized reporting, robust external validation, prospective evaluation, improved dataset diversity, imaging standardization, reproducibility, and development of clinically deployable AI and radiomics models.