Prostate cancer (PCa) is a common malignant tumor in men, and multiparametric magnetic resonance imaging (mpMRI) plays a crucial role in the early diagnosis and staging of PCa. However, the current diagnostic system has low specificity and is highly influenced by the experience of radiologists. To address this issue, this study combines semi-supervised learning to achieve high-precision prostate region segmentation, utilizing limited labeled data and a large amount of unlabeled data to improve segmentation accuracy. Based on the anatomical structure prior information extracted from the segmentation results, an auxiliary branch and main trunk network are designed to work in tandem, enhancing the model's ability to distinguish clinically significant prostate cancer (csPCa). Specifically, the U-Net architecture is adopted as the backbone network, combined with interpolation consistency training (ICT) to improve segmentation performance, achieving Dice coefficients of 76.38% and 87.72% in the peripheral zone (PZ) and transition zone (TZ), respectively. In the classification stage, a dynamic weight fusion mechanism based on structural prior information is introduced, utilizing an improved SPD-ResNet50 network. Experimental results show that the structure-a priori-assisted branch improves the accuracy of the SPD-ResNet model to 87.10%, with a sensitivity of 100%, an F1 score of 87.50%, and an AUC of 0.84, significantly outperforming models without structure-a priori fusion. This method reduces the cost of anatomical structure annotation while enhancing the diagnostic efficacy of csPCa, offering new insights for prostate cancer diagnosis.