FSDB-Diff:3D Medical Image Segmentation Model Based on Diffusion Model and Dual-Branch Encoding of Frequency-Space Domains

3D medical image segmentation is crucial for precision medicine, yet it faces challenges such as noise, blurred boundaries, and the balance between local and global features. This paper proposes the FSDB-Diff model, which integrates diffusion models with a 3D U-Net architecture. It features a dual-branch structure in the frequency domain and spatial domain, and incorporates high-low frequency fusion with cross-domain attention mechanisms. Experiments on the BraTS 2020 and BTCV datasets demonstrate that this method significantly outperforms current state-of-the-art approaches. The mean Dice score improves by 1.12%, while the HD95 metric decreases by 0.98. Ablation studies validate the effectiveness of each module, providing an efficient solution for clinical diagnostic assistance.