| 摘 要: 针对皮肤病图像病变区域的形状大小不同、形状不规则、受周围毛发干扰、边缘模糊等情况以及现有网络分割效果不佳的问题,提出一种改进的特征金字塔网络(FPN)的皮肤病分割算法CG-FPN。通过加入通道和位置注意力模块(CPAM)以及门控融合模块(GFM),旨在增强对皮肤病关键特征通道的响应和精准定位病灶区域,显著提升对不规则形状及微小皮肤病灶的分割能力。在公开皮肤病数据集ISIC2017上的实验结果表明,CG-FPN算法在Dice相关系数(Dice)、JaccarD相似指数、准确率、灵敏度、特异性以及F1分数指标上相较于基准模型FPN分别提升了4.89、7.82、1.06、2.80、1.20和2.14个百分点,并且均优于其他算法。这些结果表明,CG-FPN分割算法效果显著,能够为皮肤病分割辅助诊断提供可靠、准确的分割结果。 |
| 关键词: 图像处理 皮肤病分割 特征金字塔网络 双注意力机制模块 门控融合模块 |
|
中图分类号: TP399
文献标识码: A
|
|
| Skin Disease Segmentation Based on CG-FPN Network |
|
FENG Zicheng, HE Liwen
|
(School of Internet of Things, Nanjing University of Posts and Telecommunications, Nanjing 210003, China)
1023071804@njupt.edu.cn; helw@njupt.edu.cn
|
| Abstract: Skin disease images typically present lesions with varying shapes and sizes, irregular contours, interference from surrounding hair, and blurred edges, which pose significant challenges to existing segmentation networks. This paper proposes an improved skin lesion segmentation algorithm, namely CG-FPN, based on the Feature Pyramid Network (FPN). By introducing a Channe-l Position Attention Module (CPAM) and a Gated Fusion Module (GFM), the proposed method enhances the response to key feature channels of skin lesions and accurately localizes lesion regions, thereby significantly improving segmentation performance on irregularly shaped and small lesions. Experimental results on the publicly available skin disease dataset ISIC 2017 demonstrate that, compared with the baseline FPN model, the proposed CG-FPN achieves improvements of 4.89, 7.82, 1.06, 2.80, 1.20, and 2.14 percentage points in the Dice similarity coefficient, JaccarD similarity index, Accuracy, Sensitivity, Specificity, and F1-score, respectively, and outperforms other competing algorithms. These results indicate that the CG-FPN segmentation algorithm delivers superior performance and can provide reliable and accurate segmentation results to support the auxiliary diagnosis of skin diseases. |
| Keywords: image processing skin lesion segmentation FPN dual attention mechanism gated fusion module |
