Uso de la inteligencia artificial como modelo de predicción de riesgos en pacientes con hallazgos equívocos en...

Introduction: PI-RADS 3 lesions represent a diagnostic grey zone on multiparametric MRI, with clinically significant prostate cancer (csPCa) detected in only 10-30%. Their equivocal nature leads to both unnecessary biopsies and missed cancers. Artificial intelligence (AI) has emerged as a potential tool to provide objective, reproducible risk prediction. This review summarises current evidence on AI for risk stratification in patients with indeterminate mpMRI findings, including clarification of key multicentre initiatives such as the PI-CAI (Prostate Imaging-Artificial Intelligence) study-a global benchmarking effort comparing AI systems against expert radiologists. Methods: A narrative review of PubMed and Embase (search updated to August 2025) was conducted using terms including "PI-RADS 3", "radiomics", "machine learning", "deep learning", and "artificial intelligence." Eligible studies included those evaluating AI-based prediction of csPCa in PI-RADS 3 lesions using biopsy or long-term follow-up as reference standards. Both single-centre and multicentre studies were included, with emphasis on externally validated models. Results: Radiomics studies demonstrate that handcrafted features extracted from T2-weighted and diffusion-weighted imaging can distinguish benign tissue from csPCa, particularly in the transition zone, with area-under-the-ROC curves typically 0.75-0.82. Deep learning approaches-including convolutional neural networks and large-scale representation-learning frameworks-achieve higher performance and can reduce benign biopsy rates by 30-40%. Models that integrate imaging-based AI with clinical predictors such as PSA density further improve discrimination. The PI-CAI study, the largest international benchmark to date (>10,000 MRI exams), shows that state-of-the-art AI systems can match or exceed expert radiologists for csPCa detection across diverse scanners, centres, and populations, though prospective validation remains limited. Conclusions: AI shows strong potential to refine management of PI-RADS 3 lesions by reducing unnecessary biopsies, improving csPCa detection, and mitigating inter-reader variability. Translation into routine practice will require prospective multicentre validation, harmonised imaging protocols, and integration of AI outputs into clinical workflows with clear thresholds, decision support, and safety-net recommendations.