基于双路编码器融合与注意力门控的脑卒中图像分割网络OA

Ischemic stroke lesions in 3D magnetic resonance imaging(MRI)are characterized by small volume,heterogeneous morphology,and indistinct boundaries,making automatic segmentation highly challenging.To enhance global semantic consistency and local detail representation,a 3D medical image segmentation network,DFAG-Net,is developed based on dual-encoder fusion and attentiongating mechanisms.The network integrates a Swin Transformer encoder with a lightweight 3D convolutional encoder.Cross-branch and multi-scale semantic interaction are achieved through an Encoder Fusion Module(EFM),while selective spatial filtering across hierarchical features is performed using an Attention-Gated Skip Module(AGSM).Experiments conducted on the ISLES 2022 and ATLAS v2.0 stroke datasets demonstrate that DFAG-Net surpasses multiple mainstream segmentation models in Dice coefficient,HD95,precision,and recall.Specifically,the Dice on ISLES 2022 and ATLAS v2.0 reached 74.50%and 56.80%,respectively.The results indicate that the network exhibits strong stability and adaptability in 3D stroke lesion segmentation,providing effective technical support for quantitative analysis and intelligent assisted diagnosis of ischemic stroke lesions.