基于多尺度的肿瘤轮廓结构不规则性特征分析

Feature Analysis of Tumors Boundary Structure Irregularity Based on Multi-scale

皮肤肿瘤轮廓的结构不规则性特征提取在计算机辅助诊断黑色素瘤中具有重要意义。在前期轮廓不规则性的局部分形维基础上,提出基于多尺度曲率的轮廓不规则特征提取方法,采用相邻尺度间特征差异度量来增强大尺度下的甄别良恶性皮肤肿瘤的能力。通过局部分形维和曲率分析比较表明,相邻尺度间特征差异度量方法具有类间Hausdorff距离随尺度增大的特性,但局部分形维较曲率分析具有较大的Hausdorff距离值。实验结果表明,上述方法不仅具有较强的结构不规则性的分类能力,并且有助于削弱纹理不规则性对分类结果的影响。

Feature extraction of structure irregularity for skin lesion boundaries has a great significance in computer-aided diagnosis for melanomas. Based on previous work using local Fractal Dimension（local FD） for contour irregularity descriptions, this paper proposes the feature extraction method of boundary irregularity using multi-scaled curvature analysis. Feature differences from adjacent scales are used to enhance the discrimination power between malignant and benign shin tumors. Comparative experiments of both local FD and curvature analysis show that the features difference of adjacent scales has characteristics of larger inter-class Hausdorff distance with the scale increase, while local FD performs better than curvature analysis with larger Hausdorff distance value. Experimental results show that the two methods not only have strong classification capacity of structure irregularity, but also are helpful to weaken texture irregularity for the classification.