Xiaoyang Ge, Yi Wang, Jia Cheng, Tong Luo
To address insufficient texture recovery and limited global-structure modeling in image-domain super-resolution (SR) of prostate T2-weighted magnetic resonance imaging (T2W MRI), we propose HDBU-Net, a hybrid dual-branch U-Net. HDBU-Net uses U-Net as its backbone and comprises shallow feature extraction, dual-branch feature extraction, and high-frequency enhancement at the end of reconstruction, forming an end-to-end SR model that improves structural fidelity and detail recovery in low-resolution prostate T2W MRI. The dual-branch feature extraction module uses a Local Feature Enhancement Branch (LFEB) and a Global Structure Modeling Branch (GSMB) to capture high-frequency details and long-range dependencies, respectively. We further design a Cross-Branch Adaptive Interaction Module (CrossAIM), which uses channel and spatial gating to provide bidirectional guidance and adaptive fusion between the two branches. In addition, a HighPassGate module is introduced at the reconstruction end to strengthen the high-frequency representation of edges and fine-grained textures. The model was evaluated on the public ProstateX and PI-CAI datasets, using Gaussian blurring and bicubic downsampling to generate synthetic low-resolution images. The results show that the proposed method provides stable reconstruction performance across different upscaling factors, achieves favorable results for most evaluation metrics, and exhibits an ability to preserve structures and recover local textures.