关键词: 皮肤病灶分割, 注意力机制, 空洞卷积, 多尺度信息, 特征融合

Abstract: Abstract: Aiming to address issues such as blurred skin lesion boundaries, hair interference, and varying lesion sizes, a skin lesion segmentation network based on a dual-path attention mechanism and multi-scale information fusion was proposed. First, a residual gated attention module (DGConv) utilizing depthwise separable convolution was designed in the encoder to capture local lesion information. Second, a multi-scale context extraction module (MCEM) was introduced at the bottleneck, employing horizontal and vertical average pooling to model context and integrating multi-scale features from a residual atrous convolution pyramid to enhance global lesion understanding. A dual-path attention module was applied at skip connections to refine lesion details, while a multi-scale feature enhancement (MSFE) module was utilized to enrich feature details by fusing multi-stage information. Finally, a feature fusion module (FM) was designed in the decoder to address receptive field mismatches, progressively combining encoder outputs and skip connection features to achieve the final segmentation. Experiments on the ISIC2017 and ISIC2018 datasets demonstrated that the proposed model outperformed suboptimal methods, with Dice improvements of 0.09 and 1.09 percentage points, and IoU improvements of 0.14 and 1.76 percentage points, respectively. Compared to U-Net, Dice was improved by 5.13 and 3.85 percentage points, and IoU by 7.74 and 6.03 percentage points, confirming the effectiveness of the proposed network.

Key words: skin lesion segmentation, attention mechanism, dilated convolution, multi-scale information, feature fusion