Front Oncol. 2026 ;16
1733046
Predicting prostate cancer (PCa) recurrence after radical treatment is crucial for personalised adjuvant therapy. This study aimed to compare different algorithms in order to select the best model for predicting recurrence. Therefore, a retrospective cohort analysis was conducted on 72 patients with radical prostate cancer, including 39 patients with biochemical recurrence and 33 patients without recurrence. We extracted features from imaging data, construct and evaluate 10 machine learning models and 8 deep learning models. Model performance was assessed using the area under the curve (AUC), accuracy, sensitivity, specificity, precision, 10-fold cross-validation AUC, and F1-score. In addition, the feature importance was analysed. Among all models, the MLP-Mixed-Act model exhibited superior performance in all evaluation indicators (AUC = 0.910, accuracy =0.819, sensitivity =0.744, specificity =0.909, precision =0.912, F1 = 0.817), thereby indicating its strong predictive ability and clinical application potential. This study provides a theoretical basis for the development of preventive and non-invasive recurrence prediction tools. Especially in the context of valuing the tumor microenvironment, accurate recurrence prediction can effectively help select immunotherapy strategies, improve treatment efficacy and prognosis, and support for personalized treatment of PCa.
Keywords: deep learning; positron emission tomography/computed tomography; prostate cancer; radiomics; recurrence prediction