Automatic detection of translucency using a deep learning method from patches of clinical basal cell carcinoma images

Translucency, defined as a jelly-like appearance, is a common clinical feature of basal cell carcinoma, the most common skin cancer. The feature plays an important role in diagnosing basal cell carcinoma in an early stage because the feature can be observed readily in clinical examinations with a high specificity of 93%. Therefore, translucency detection is a critical component of computer aided systems which aim at early detection of basal cell carcinoma. To address this problem, we proposed an automated method for analyzing patches of clinical basal cell carcinoma images using stacked sparse autoencoder (SSAE). SSAE learns high-level features in unsupervised manner and all learned features are fed into a softmax classifier for translucency detection. Across the 4401 patches generated from 32 clinical images, the proposed method achieved a 93% detection accuracy from a five-fold cross-validation. The preliminary result suggested that the proposed method could detect translucency from skin images.