Abstract Background Prostate cancer (PCa) is the second most common malignancy diagnosed in men worldwide, with approximately 1.47 million new cases reported in 2022. Biochemical recurrence (BCR), defined as a rising prostate-specific antigen (PSA) after radical treatment, is the first clinical sign of disease relapse and a harbinger of metastasis and cancer-specific mortality. Accurate, non-invasive prediction of BCR is essential for guiding individualized treatment decisions and optimizing long-term outcomes. Main body This narrative review critically evaluates the current evidence on artificial intelligence (AI)-based approaches—encompassing radiomics, machine learning (ML), and deep learning (DL)—applied to multiparametric magnetic resonance imaging (mpMRI) for the prediction of BCR in PCa following radical prostatectomy (RP) or radiation therapy (RT). The review further examines multimodal AI approaches integrating mpMRI with prostate-specific membrane antigen positron emission tomography (PSMA-PET), digital pathology, and genomic data. This manuscript is a narrative review; no systematic protocol was registered. Among the reviewed studies, mpMRI-based radiomics models achieved area under the receiver operating characteristic curve (AUC) values ranging from 0.72 to 0.97 for BCR prediction, though this wide range reflects substantial methodological and population heterogeneity. Deep learning models, particularly those combining mpMRI features with clinical parameters, demonstrated C-index values up to 0.83. Because the area under the receiver operating characteristic curve (a discrimination metric for binary classification) and the C-index (for time-to-event survival analysis) are distinct statistical measures, radiomics AUC and deep-learning C-index values are reported separately here and are not directly comparable or interchangeable. Conclusion AI-powered mpMRI analysis holds substantial promise for non-invasive, accurate BCR prediction in PCa. Integration of radiomics and DL with clinical and multi-omics data within standardized, multi-center frameworks represents the most promising future direction. Regulatory-compliant, externally validated models with demonstrated calibration are required before routine clinical implementation.
Narimani et al. (Thu,) studied this question.