皮肤肿瘤轮廓的结构不规则性对黑色素瘤临床早期诊断具有重要意义。针对皮肤肿瘤轮廓结构不规则性度量和特征分类器设计问题，提出了一种基于小波子带分析的轮廓结构分量获取及多尺度特征神经网络分类器构建方法。利用肿瘤轮廓小波子带能量的Hausdorff距离寻找显著性小波子带，进而重构肿瘤轮廓的结构分量；给出基于显著性小波子带的轮廓多尺度结构不规则特征描述——基于统计和几何的轮廓结构不规则性度量。对单尺度/多尺度、小样本/大样本,以及特征选择前后神经网络分类器性能进行了实验。结果表明，在小样本情况下,基于显著性小波子带多尺度特征描述扩展了样本的特征空间维数，使得特征选择后分类器的灵敏度和特异度指标分别等于和优于大样本分类器。

Boundary Structural irregularity of skin lesions has shown the most significance in the early diagnosis of melanomas in clinic. Aiming to measure structural irregularities of skin lesion contours and designing feature oriented classifiers, a method for extracting structural components of contours based on wavelet sub-band analysis and neural network classifiers using multi-scale features is proposed in this paper. A set of salient wavelet sub-bands are searched using Hausdorff distance evaluations on energy at wavelet subbands. Then structural components of skin lesion contours are reconstructed. Furthermore, multi-scale descriptions of the contours structural irregularity are given using statistical and geometrical measures based on the salient wavelet sub-bands. Finally, various experimental schemes are taken on single scale/multi-scale features, for small training sets and large training sets as well as and before and after feature selection for a comparison on performances of the neural network classifiers. It is shown from experiments that the classifiers sensitivity and specialty after feature selection using small sample sets equal and exceed that of large training sets respectively as dimensions of feature spaces are expanded by the proposed multi-scale independent features at salient wavelet sub-bands.