BounDA-SAM:复杂边界医学图像分割方法OA

Although the Segment Anything Model(SAM)performs excellently in general image segmentation tasks,its perfor-mance in medical image segmentation is limited due to its inability to effectively capture complex boundary details in medical im-ages.To address this issue,this paper proposes a SAM-based medical image segmentation method called BounDA-SAM.First,a dual-backbone feature enhancement mechanism is constructed in the hybrid encoder:the front convolution(Conv-I)extracts low-level texture features of medical images,and the rear convolution(Conv-V)strengthens the high-level semantic expression output by the Vision Transformer(ViT).Second,the Boundary-Awareness Module(BAM)fuses the features of the dual back-bones,uses the Sobel operator to generate multi-scale boundary pseudo-prompts,which are then input together with image em-beddings into the decoder guided by the boundary-aware and residual local adapter.Finally,on the premise of freezing the back-bone parameters of SAM,a lightweight residual local adapter is embedded to achieve local feature enhancement through"three-stage residual convolution+depth-wise separable convolution".Experimental results on the COVID-19 dataset and the private Left Atrial Appendage(LAA)dataset show that the Dice coefficients are 87.76%and 85.78%,respectively,and the 95%Haus-dorff distances(HD95)are 25.49 and 28.91,respectively.Comparative experiments with mainstream methods such as U-Net and SAM show that,compared with SAM,the proposed method achieves a Dice coefficient improvement of 5.27%and 4.52%and an HD95 reduction of 56.90 and 29.03 on the COVID-19 and LAA datasets,respectively;compared with U-Net,the Dice coefficient increases by 18.40%and 5.51%,with the HD95 decreasing by 8.11 and 5.07 correspondingly,indicating the effec-tiveness of the proposed method in medical image segmentation.