Unsupervised Learning Reduces Local Image Redundancy and Improves Supervised Segmentation of Dermoscopy Lesions

Melanoma is the most common form of skin cancer, and skin disease image segmentation plays a vital role in automated diagnosis of skin cancer. A primary challenge of image segmentation and other automated object recognition techniques is the large amount of redundant input information which often obfuscates critical input features. In the context of dermatoscopy lesion segmentation we show that unsupervised clustering algorithms applied to input images can reduce local image redundancy and result in dramatic improvements in segmentation performance. Our work proposes a skin disease image segmentation algorithm combining an unsupervised simple linear iterative cluster algorithm (SLIC), and the supervised deep learning U-Net model. The unsupervised SLIC method can detect the fine structure of skin damage highlighting critical features that improve segmentation performance of the supervised U-Net model. Both the superpixel dermoscope image and original image are used as input information for the U-Net training deep learning model. Finally, a fully-connected conditional random field (CRF) is used for image post-processing. This algorithm achieves an Intersection Over Unit(IOU) coefficient reaching 83%, dice coefficient 90%, sensitivity 90%, improved by 10%, 7% and 4% respectively in comparison with the results of the classic U-Net, showing that this approach improves the performance of network image segmentation.