3D medical image segmentation based on group mixing dual attention
Year: 2025
Type: article
Abstract: In recent years, vision transformers have become a significant focus in 3D medical image segmentation research due to their advantage in capturing long-range dependencies. In this paper, we propose a new method for 3D medical image segmentation, called GMDA UNETR. This method combines the comprehensive group mixing mechanism with the advantages of a novel dual attention (spatial and channel attention) module to provide high-quality segmentation masks. Specifically, we introduce a new type of group mixing dual attention module that can simultaneously capture correlations among token-token, token-token group, and token group-token group, and enhance discriminative features in spatial and channel aspects through the dual attention (spatial and channel attention) module. Additionally, we introduce a channel compression-spatial excitation module to alleviate the spatial information limitations caused by channel attention. Extensive evaluations were conducted on two benchmark datasets, ACDC and INSTANCE 2022, and experiments demonstrate the effectiveness of the proposed method in terms of efficiency and accuracy. The model achieved a DSC score of 82% and an HD95 of 9.92mm on the INSTANCE 2022 dataset, outperforming the state-of-the-art nnUNet method. On the ACDC dataset, the DSC score was 92.54%, surpassing the state-of-the-art UNETR++.
Institution: North University of China
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $2,590
OpenAlex ID: W4410739010