Radiology-based artificial intelligence for predicting targeted therapy response in pan-cancer: a comprehensive review.

backgroundTargeted therapy is central to precision oncology, but identifying patients who will benefit remains challenging. Conventional molecular testing, though the current standard, provides limited predictive value. With recent advances in artificial intelligence (AI) and the widespread availability of imaging data, radiology-based AI models have emerged as valuable non-invasive tools for treatment response assessment.

methodsWe conducted a comprehensive review of 112 studies that developed radiology-based AI models for predicting responses to targeted therapy across various cancer types. The reviewed models were classified into direct prediction approaches, which use end-to-end imaging-based modeling to estimate therapeutic response, and indirect prediction approaches, which infer molecular biomarkers from imaging features to indirectly assess therapeutic sensitivity.

resultsAcross the identified literature, computed tomography (CT) was the most frequently used imaging modality, followed by magnetic resonance imaging (MRI), positron emission tomography (PET), and ultrasound (US). Lung and breast cancers were the most commonly studied diseases, though work has also expanded into gastric, colorectal, liver, kidney, brain, and ovarian cancers. Both machine learning (ML) and deep learning (DL) frameworks have been applied, with ML remaining dominant but DL gaining increasing attention in recent years, likely because ML offers interpretability and suitability for smaller datasets, whereas DL excels in handling complex, high-dimensional data. Collectively, these studies demonstrate promising performance in predicting response to targeted therapy, while also highlighting the diversity of cancer contexts and methodological designs.

conclusionRadiology-based AI offers a non-invasive approach to guide treatment selection and monitoring in targeted therapy. This review summarizes current progress, highlights strengths and limitations of direct and indirect prediction strategies, and discusses future directions. To support accessibility, we also provide a continuously updated interactive website of included resources.