DAUNet: A deformable aggregation UNet for multi-organ 3D medical image segmentation

Abstract

Medical image segmentation is a critical component of medical image analysis. However, due to the limitations in the size and shape of the receptive field, neural networks often struggle to adapt to segmentation targets of different sizes in 3D medical images. Furthermore, they may not be able to adequately model the intra-slice and inter-slice relationships in 3D medical images. To overcome these challenges, we propose a Deformable Aggregation UNet (DAUNet) for multi-organ segmentation in medical images. We introduce two specially designed modules into the UNet structure, which is composed of residual blocks. Specifically, the Deformable Aggregation Module (DAM) incorporates deformable receptive fields and gate fusion methods. This enhances the fusion of information from various hierarchical levels, enabling DAUNet to adapt to the segmentation of multiple organs or structures at different scales within the abdominal region. Simultaneously, the Fourier Attention Module (FAM) leverages Fourier convolution to enhance long-range dependencies and relationships within the entire 3D volume of medical images, accommodating the detailed structural variations in different directions. The network was assessed on three publicly available datasets (ACDC, BTCV, and BraTS-Africa2024), as well as a private clinical dataset for the TME-CT segmentation task. Compared to existing state-of-the-art (SOTA) methods, DAUNet achieves superior performance, with improvements of 0.96%, 2.52%, 1.60%, and 4.57% in average Dice scores across the four datasets, respectively.