Magnetic Resonance Imaging (MRI)-based radiomic models in prostate cancer diagnosis, staging and response prediction: systematic review and meta-analysis.

Purpose: To assess quality of published machine learning models using MRI-based radiomic features in diagnosis, staging and response assessment of prostate cancer.

Methods/Materials: A PubMed database query with no date or language restrictions using the keywords (prostate AND MRI AND radiomics) revealed 264 studies. After exclusion of non-relevant publications, 70 studies between 2017-2023 with a total of 14453 subjects (range: 16–644) were included in meta-analysis.

Results: There was a significant increase in the number of publications after 2020, majority of which originated in China and Italy. The outcome of interest involved diagnosis, staging and treatment response in 41 (62%), 21 (31%) and 5 (7%) of the publications. All studies conducted internal validation using a total of 4360 subjects (range: 13–380). The number of radiomic features included in the highest performing models ranged between 3–540 (median: 13). In addition, 25% of the studies utilized clinical variables (most frequently age [n=13]) and 29% utilized laboratory values (most frequently prostate-specific antigen (PSA) [n=14] and PSA density [n=10]). Pooled area under receiver operating characteristic (AUROC) curve (95% confidence interval) for the best models for diagnosis, staging and treatment response prediction were 0.889 (0.863–0.915), 0.872 (0.845–0.941) and 0.765 (0.616–0.915), respectively. On meta-regression, there was a significant positive association between publication year and AUROC in treatment response prediction studies but not other two categories (β: 0.074, P< 0.0001). There was no association between the number of the radiomic features and model AUC (β: -0.0002, P=0.850).

Conclusions: MRI-based radiomics machine learning models show high validity in diagnosis, staging and treatment response prediction of prostate cancer. Increase in the number of radiomic features does not improve predictive value of the models. Post-treatment radiomic models have improved since their conception in predicting treatment response but not diagnosis and staging. Wide heterogeneity in patient selection and applied statistical methods can limit generalizability of the models.