Application of an improved Swin-Unet with wavelet convolution and attention mechanism for MRI segmentation in HIFU treatment of uterine fibroids
GUO Lei
LI Li-ming
LIN Chu-lan
OU Xiang-hong
Abstract：Objective To propose an improved Swin-Unet network model incorporating frequency-domain en-hancement and multi-scale attention for high-precision automatic segmentation of uterine fibroids in MRI,supporting personalized and precise HIFU treatment.Methods In the Swin-Unet architecture,wavelet convolution was introduced after Patch Merging and Patch Expanding to capture multi-frequency details.The encoder employed a multi-scale dilat-ed convolution pyramid to enhance contextual awareness.Additionally,spatial-channel combined attention was embedded in the skip connections to adaptively refine high-level semantic features.Results Using a dataset of 1 582 T2-weigh-ted MRI images,the proposed model outperformed other deep learning methods,including Attention UNet,demonstrating superior accuracy and stability.Performance metrics were excellent,with mean recall,mean Dice coefficient,and mean Jaccard coefficient reaching 91.17,91.33,and 86.23,respectively.Conclusion The improved Swin-Unet with wave-let convolution and self-attention mechanism enables automated segmentation of uterine fibroids in MRI for HIFU treat-ment.It achieves better performance across multiple evaluation metrics compared with existing methods,offering a reliable tool for individualized treatment planning.
Keywords：medical image segmentationdeep learninguterine fibroidsattention mechanisms
Publication Date：2025-11-15
Online Publishing Date：2026-01-06(First online date of this platform, not the publication date of the document)
Pages：8( 1606-1613 )