Hierarchically enhanced feature fusion and loss prevention for prostate segmentation on micro-ultrasound images

Micro-ultrasound (micro-US) provides superior resolution compared to conventional ultrasound, enabling improved visualization of anatomical structures and supporting more accurate prostate cancer detection. However, prostate segmentation in micro-US remains challenging due to imaging artifacts and indistinct tissue boundaries. While recent Transformer-based UNet models have achieved promising results, they still suffer from inefficient multi-scale feature fusion, limited contextual modeling, and inadequate integration of local and global information. To address these issues, we propose HEFFLPNet, a Transformer-based UNet architecture that introduces hierarchically enhanced feature fusion and loss prevention mechanisms. HEFFLPNet integrates three novel modules: the Tri-Cross Attention with Feature Enhancement (TriCAFE) module for strengthening semantic consistency across skip connections, the Multi-Scale Prediction Map Attention (MSPMA) module for scale-specific feature refinement, and the Upsample Fusion Attention (UpFA) module for attention-guided fusion of spatial details. The model adopts annotation-guided binary cross entropy (AG-BCE) loss and deep supervision across multiple scales, building on the MicroSegNet framework. We evaluate HEFFLPNet on two datasets: the micro-US dataset (55 training and 20 testing cases) and the CCH-TRUSPS dataset. Our model achieves a Dice coefficient of 0.938 and a Hausdorff distance of 2.12 mm on the micro-US dataset, and 0.914 Dice and 3.63 mm HD95 on CCH-TRUSPS, outperforming existing state-of-the-art methods. These results demonstrate the effectiveness and generalization ability of HEFFLPNet in segmenting prostate structures under challenging ultrasound conditions.