基于边缘关键点和边缘注意力的黑色素瘤图像分割方法

Melanoma image segmentation method based on edge key points and edge attention

黑色素瘤图像的高精度分割对早期诊断和提高患者的生存率至关重要。然而,由于黑色素瘤的边缘区域模糊,呈现出不规则的形状,使得现有分割方法难以获取边缘特征信息,影响了黑色素瘤图像分割的准确率。为解决该问题,提出了一种基于边缘关键点和边缘注意力的黑色素瘤图像分割方法。首先,在编码器中设计了点渲染的边缘关键点选择模块和组合卷积变压器块,通过边缘关键点的选择,引导获取边缘的局部特征和全局特征。然后,在编码器中设计边缘细化模块用于细化深层网络的边缘特征,最后,在跳连接中设计了多尺度边缘注意力模块,能够更好地捕获图像多尺度的边缘形状特征。将所提方法在ISIC 2018和PH2两个数据集上进行实验,实验结果表明,与现有分割方法相比,所提方法具有较好的分割性能。

High-precision segmentation of melanoma images is crucial for early diagnosis and improving patient survival.However,the blurring of the edge region of melanoma,which presents irregular shapes,makes it difficult for existing segmentation methods to obtain edge feature information,affecting the accuracy of melanoma image segmentation.To solve this problem,a melanoma image segmentation method based on edge key points and edge attention is proposed.An edge key point selection module for point rendering and a combined convolution transformer block are designed in the encoder to guide the acquisition of local and global features of the edge by selecting edge key points.Then,the edge refinement module is designed in the encoder to refine the edge features of the deep network,and finally,the multi-scale edge attention module is designed in the skip connection,which enables the capture of the edge shape features at multiple scales.The tests on two datasets(ISIC 2018 and PH2)demonstrate that the proposed method has better segmentation performance than the existing segmentation methods.