Wavelet-based logistic discriminator of dermoscopy images

Proper diagnosis of cutaneous melanoma is a life-saving factor. The most important limitation is the early and sensitive recognition of melanoma relative to dysplastic nevi. We have studied wavelet-based features extracted from dermoscopic images as efficient signals of neoplastic changes. We recursively treat the dermoscopic images and all their transformation channels (wavelet packets) through the Mallat transform. All the four decomposition filters from each decomposition level are the source of features based on three functions of the pixel values. We train the logistic classifier regularized by either the Lasso or the Ridge penalty and test its AUC metric for a set of different wavelet bases, and as a function of image resolution. A total of three different data sets with respectively 185, 117, and 413 images, and 52 wavelet bases are tested. Classification performance as a function of the wavelet number strongly depends on the image resolution and image compression. There is also a large variation in the classification performance within the self same wavelet family. Degradation of image resolution makes the overall classification performance lower and more dispersed between the regularizes. Some wavelets do not lower, but increase the learning performance at the reduced image resolutions, which is consistent with the melanoma feature-extraction studies based on other learning paradigms. The logistic classifier can extract high-performance, resolution-invariant wavelet features of melanoma.