DynaSynergic-UNet:面向医学图像分割的动态协同UNetOA

The accuracy of medical image segmentation is crucial for disease diagnosis and treatment planning,yet existing methods still have limitations in multi-scale feature perception optimization,cross-level feature fusion alignment,and synergy between edge extraction and overall segmentation tasks.To address the above challenges,a dynamic cooperative segmentation network,DynaSynergic-UNet,is proposed.DynaSynergic-UNet consists of the deformable attention trans-former++(DAT++),the multi-scale large kernel channel-spatial attention(MLCSA)module,the bidirectional deformable alignment cooperative attention fusion(BiDCAF)module,and the edge guided multi-scale refinement(EGMR)module in concert.Among them,the DAT++encoder is used to encode the spatial structural feature representation of the input image;the MLCSA module captures the local details through the multi-scale parallel large kernel depth wise convolution with the ability of global sensing at the same time,and it enhances the focusing and discriminative perception of important target regions through channel attention and spatial attention;the BiDCAF module performs dynamic spatial feature alignment through the deformable mechanism and combines gated fusion,cross-layer cooperative attention calibration,and feature cross-complementarity to solve the encoder-decoder spatial feature representation difference and feature semantic gap problem;and the EGMR module performs edge fine-grained extraction through the multi-scale target region cooperative Sobel operator,which promotes the consistency improvement between edge accuracy and target region,and introduces boundary difference over union loss(Boundary DoU Loss)to optimize the effect of edge-guided segmentation.The pro-posed method is experimentally compared with recent advanced models such as EMCAD,AgileFormer,D-LKA Net,Swin-UNet,etc.On the public medical image segmentation datasets—the Synapse abdominal multi-organ CT dataset,the ACDC cardiac MRI dataset,and the ISIC2018 dermatology dataset,the small version of the model achieves average Dice similarity coefficients of 85.29%,92.33%,and 89.44%with 6.064×107 parameters,the tiny version achieves average Dice similarity coefficients of 84.49%,92.11%,and 89.19%with 2.724×107 parameters.The proposed model has a better trade-off among segmentation performance,resource utilization,and inference speed.The proposed model has also conducted cross-domain fine-tuning experiments using the ISIC2018 dermatology dataset and the LiTS2017 liver tumor dataset to further explore its generalization capabilities.DynaSynergic-UNet provides an efficient and robust dynamic synergistic solution for medical image segmentation,which is valuable for precise clinical diagnosis and treatment.