基于ESS-Net模型的皮肤病变分割算法研究OA

The current skin lesion segmentation algorithms face challenges in accurately segmenting multi-dimensional and multi-scale skin lesions.To address this issue,we propose ESS-Net,an improved skin lesion segmentation model based on EGE-UNet.Our al-gorithm combines the advantages of both multi-scale attention and multi-dimensional attention.Specifically,we introduce an efficient multi-scale attention module in the encoder to capture lesions at different scales,and incorporate a parameter-free multi-dimensional at-tention module after the original bridge layer to comprehensively extract features across channel and spatial dimensions.Furthermore,we add a spatial-channel reconstruction convolution module after the convolutional layers in the decoder to enable multi-dimensional recon-struction and reduce redundancy.Experimental results on the ISIC 2017 dataset demonstrate that ESS-Net achieves superior perfor-mance with mean Intersection over Union(mIoU)of 80.04%and Dice Similarity Coefficient(DSC)of 88.89%.Comparative experiments with other models show that ESS-Net outperforms existing approaches,while ablation studies confirm the effectiveness of the improve-ments made to the baseline model.