Computer-Aided Diagnosis of Clinically Significant Prostate Cancer in Transrectal Ultrasound

Diagnosing clinically significant prostate cancer (csPCa) using transrectal ultrasound (TRUS) remains a challenging task due to limited image contrast and the subtle appearance of malignant lesions. In this study, we propose a novel hybrid framework that integrates multi-modality and multi-task learning for csPCa classification from TRUS data. The framework leverages complementary information from B-mode and strain elastography (SE) imaging via a dual-stream encoder, enhanced by cross-modal attention gate (CMAG) modules at each level. These CMAG modules first adaptively integrate cross-modal features and then employ a gating mechanism to regulate the flow of cross-modal context into each modality-specific stream. To mitigate the scarcity of segmentation annotations, we introduce a multi-scale prototype-guided learning (MSPL) strategy, which propagates supervision signals from a small set of labeled samples to unlabeled ones by leveraging learned prototypes across multiple feature scales. This design allows the task-shared encoder to benefit from structural cues captured by the segmentation branch, even with minimal voxel-level supervision. Additionally, a context-aware multi-scale classifier (CMSC) aggregates hierarchical decoder features to further exploit the synergy between classification and segmentation tasks. The proposed method is evaluated on an in-house dataset containing 612 subjects who underwent prostate biopsy and attains favorable performance in identifying csPCa with an area under the curve (AUC) of 0.86. Extensive comparative and ablation results further highlight the efficacy of incorporating cross-modal interaction, prototype-guided learning, and context-aware classification for accurate and annotation-efficient csPCa diagnosis. The code is publicly available at https://github.com/mock1ngbrd/ProstateMMTNet .