MRI-based radiomics for prediction of biochemical recurrence in prostate cancer: a systematic review and meta-analysis.

Abstract

Background and purpose: Biochemical recurrence (BCR) following prostate cancer (PCa) treatment is a significant indicator of metastasis and mortality. Early prediction of BCR can guide treatment decisions, and optimize patient management strategies. MRI is essential for the diagnosis and surveillance of PCa. This study aimed to assess the accuracy and quality of MRI radiomics-based machine learning (ML) models for predicting post-treatment BCR in PCa.

Methods: A systematic literature search was conducted across five electronic databases (PubMed, Scopus, Embase, Web of Science, and IEEE) up to December 23, 2024, to identify studies developing ML models based on MRI-derived radiomics features for the prediction of BCR in PCa. Studies were assessed for quality using the QUADAS-2 and METRICS tools. A meta-analysis of radiomics, clinical, and clinical-radiomics models in validation cohorts was performed to pool sensitivity, specificity, and area under the curve (AUC) using a bivariate random-effects model.

Results: A total of 24 studies were incorporated into the systematic review, with 14 included in the meta-analysis. The pooled AUC, sensitivity, and specificity for radiomics-based ML models were 0.75, 72%, and 78%, respectively. Clinical-radiomics models showed the highest performance with a pooled AUC of 0.88, sensitivity of 85%, and specificity of 79%. QUADAS-2 revealed significant methodological biases, particularly in the index test and flow and timing domains. The mean METRICS score across studies was 65.68%, ranging from 43.8 to 82.2%, showing overall good quality but highlighting methodological gaps in some domains.

Conclusion: MRI-based radiomics demonstrates potential for predicting BCR in PCa, especially when integrated with clinical variables. However, it is still far from widespread clinical use, necessitating further standardization and key methodological improvements for better generalizability and robustness. Future studies should adopt multi-center designs and conduct thorough external validation to enhance applicability across diverse patient populations.