Machine learning based on automated 3D radiomics features to classify prostate cancer in patients with prostate-specific antigen levels of 4-10 ng/mL.

Background: It can be difficult to decide clinically whether males with prostate-specific antigen (PSA) levels between 4 and 10 ng/mL should be suggested for a biopsy. This study aimed to develop a fully-automated magnetic resonance imaging (MRI) based prediction model for patients with PSA levels of 4-10 ng/mL to predict prostate cancer (PCa) preoperatively and reduce unnecessary biopsies. Methods: A retrospective study of 574 patients with PSA of 4-10 ng/mL was conducted, split into training (n=434) and testing (n=108) groups. A no-new-Net (nnU-net) model was trained for three-dimensional (3D) prostate segmentation on T2-weighted fast spin echo (T2FSE) MRI sequences and 1,595 radiomics features were extracted with PyRadiomics. There were 113 machine learning approaches compared to construct a radiomics model after features selection. The diagnostic performance of the model was compared with PSA and PSA density (PSAD). Results: The nnU-net model achieved relatively higher accuracy of segmentation for the prostate region in various datasets. The average dice was 95.33%, the average relative volume error (RVE) was 1.57%, and the average 95% Hausdorff distance (HD95) value was 2.73 mm. The radiomics model [area under the curve (AUC): 0.938; 95% confidence interval (CI): 0.916-0.960] shows superior accuracy to PSA (AUC: 0.542; 95% CI: 0.474-0.611) and PSAD (AUC: 0.718; 95% CI: 0.659-0.777) in predicting PCa (P<0.05). Conclusions: The automated 3D radiomics model holds the potential to reduce unnecessary biopsies and aid urologists in managing patients with PSA levels of 4-10 ng/mL.