基于分组注意力并行编码的医学图像分割网络OA

For the task of medical image segmentation,traditional single-branch convolutional neural network(CNN)ar-chitectures,limited by their receptive fields,struggle to effectively integrate local details with global semantics.This res-ults in insufficient modeling of multi-scale structures and weak generalization across modalities.To address these limita-tions,we propose a novel segmentation network with group attention parallel fusion encoding.It employs a parallel dual-branch encoder combining Transformer and CNN to extract global semantic information and local details from im-ages,respectively.The group attention parallel fusion module following the encoder resolves the disconnection between local and global semantics as well as conflicts in multi-scale structures.Additionally,the spatial-channel dual-attention gating module in the decoder effectively tackles feature interference in medical image segmentation,thereby enhancing segmentation accuracy.Extensive experiments were conducted on the Synapse,ACDC and AVT datasets,yielding dice similarity coefficient(DSC)of 84.86％,91.66％,and 88.79％,and 95％hausdorff distance(HD95)of 13.54,1.20,and 4.02 mm,respectively.The results demonstrate that compared with existing major medical image segmentation models,our approach achieves higher segmentation accuracy and robustness,especially when dealing with complex anatomical structures of organ data.