Medical Imaging in Immunotherapy Response Evaluation: from RECIST to AI-driven Longitudinal Models, and the Distinction Between Risk Prediction and Kinetic Diagnosis

Immune checkpoint inhibitors have reshaped cancer treatment, but the response patterns they produce—pseudoprogression, dissociated responses, and hyperprogressive disease (HPD)—fit awkwardly into the size-based logic of RECIST 1.1. This narrative review, prepared with reference to the SANRA criteria, follows the evolution of response assessment from RECIST 1.1 through immune-adapted criteria such as iRECIST to functional and molecular imaging and, most recently, artificial intelligence (AI) models trained on longitudinal imaging. One distinction runs through the entire argument: current AI and radiomics models are risk-stratification tools, not diagnostic classifiers of HPD. Because HPD is defined as acceleration of tumor growth relative to a pretreatment trajectory, its diagnosis requires at least three imaging time points (pre-baseline, baseline, and first on-treatment assessment); no single-time-point model, however sophisticated, can reconstruct that trajectory. We appraise Image Biomarker Standardisation Initiative-compliant radiomics, convolutional neural networks, vision transformers, and early longitudinal modeling, and we compare early, joint, and late strategies for fusing imaging with clinical and molecular data. The clinical meaning of reported performance gains is weighed against standard assessment practice. Most evidence remains retrospective, and persistent obstacles include heterogeneous acquisition protocols, small cohorts, missing pre-baseline scans, endpoint misclassification, weak external validation, and limited biological interpretability. Prospective multicenter data collection, transparent model reporting, calibration and domain-shift testing, and biologically anchored validation are prerequisites before imaging-based AI biomarkers can enter routine immuno-oncology care.