Morphological, Texture and Auto-encoder based Feature Extraction Techniques for Skin Disease Classification

Development of computer aided diagnostic system for skin disease identification has a great impact on early and accurate diagnosis of the ailment. It is a challenging task for the experts to make diagnosis only by visual inspection of the skin lesion area or of the related dermoscopic images, owing to the closely similar appearance and complex structural property. Here, the morphological and wavelet based fractal texture features have been used along with the stacked auto-encoder based features for the identification of four disease classes from dermoscopic images. The introduction of auto-encoder based features in combination with handcrafted features has classified the melanoma, nevus, basal cell carcinoma (BCC) and seborrheic keratosis (SK) diseases with the correct identification accuracy of 96.71%, 97.77%, 98.03% and 98.11% respectively.