Deep auto-context fully convolutional neural network for skin lesion segmentation

Automatic segmentation of skin lesions in dermoscopy and clinical images is a common initial step in computer aided diagnosis. However, the low contrast of lesion boundaries and the existence of misleading image artifacts (e.g., hair), make segmenting skin lesions a challenging task. We propose a deep auto-context architecture that incorporates image appearance information as well as contextual information to predict the pixel-wise probability of a skin lesion. A sequence of fully convolutional networks is trained in a consecutive manner, where the input of each classifier is the original image concatenated with a degraded a posteriori probability estimated by the previous classifier. In contrast to common approaches that use morphological operations or thresholds to correct irregularities in the predicted lesion segmentation mask, our auto-context architecture efficiently refines the skin segmentation without any post-processing. Our experiments demonstrate that using our deep auto-context framework improves the segmentation performance of U-Net by 3.5% in terms of Dice similarity coefficient.