Background <p>Radiomics shows promise for non-invasive detection of clinically significant prostate cancer (csPCa) using prostate-specific membrane antigen positron emission tomography/computed tomography (PSMA-PET/CT). However, the impact of segmentation method—manual versus automated—on feature reproducibility and diagnostic performance remains unclear.</p> Objectives <p>To compare the diagnostic value of radiomics features extracted from manual versus convolutional neural network (CNN)-based segmentations in PSMA-PET/CT for csPCa detection, and to evaluate their performance against maximum standardized uptake value (SUVmax).</p> Methods <p>Radiomics features (<i>n</i> = 1155) were extracted from both manual and CNN-based (3D U-Net) segmentations in 110 patients. The cohort was split into training and test sets (7:3). Key features were selected using least absolute shrinkage and selection operator (LASSO) regression. Feature consistency between methods was assessed using intraclass correlation coefficient (ICC) and Cohen’s kappa. Diagnostic models were built on the training set and validated in the independent test set using area under the receiver operating characteristic curve (AUC) and accuracy (ACC).</p> Results <p>Of the 1155 features, 258 (22.24%) showed strong agreement (kappa ≥ 0.7) and 484 (41.90%) moderate agreement (0.4 &lt; kappa &lt; 0.7) between segmentation methods. Among the 10 LASSO-selected features, 4 showed strong and 3 moderate agreement. The CNN-based radiomics model achieved the highest diagnostic performance in the test set (AUC: 0.895, ACC: 0.873), outperforming both the manual model and SUVmax (AUC: 0.822).</p> Conclusion <p>Radiomics features from PSMA-PET/CT offer superior diagnostic performance for csPCa compared to SUVmax. The use of CNN-based segmentation facilitates the development of more efficient, objective, and stable radiomics models, with reduced inter-observer variability and improved reproducibility.</p>