Acral melanocytic lesion segmentation with a convolution neural network (U-Net)

Melanocytic lesions of acral sites (ALM) are common, with an estimated prevalence of 28 - 36% in the USA. While the majority of these lesions are benign, differentiation from acral melanoma (AM) is often challenging. Much research has been done in segmenting and classifying skin moles located in acral volar areas. However, methods published to date cannot be easily extended to new skin regions because of different appearance and properties. In this paper, we propose a deep learning (U-Net) architecture to segment acral melonacytic lesions which is a necessary initial step for skin lesion pattern recognition, furthermore it is a prerequisite step to provide an accurate classification and diagnosis. The U-Net is one of the most promising deep learning solution for image segmentation and is built upon fully convolutional network. On the independent validation dataset including 210 dermoscopy images our implemented method showed high segmentation accuracy. For the U-Net convolutional neural network, an average DSC of 0.92, accuracy 0.94, sensitivity 0.91, and specificity 0.92 has been achieved. ALM due to small size and similarity to other local structures create enormous difficulties during the segmentation and assessment process. The use of advanced segmentation methods like deep learning models especially convolutional neural networks have the potential to improve the accuracy of advanced medical area segmentation.