A Deep Learning Model Integrating FrCN and Residual Convolutional Networks for Skin Lesion Segmentation and Classification

Abstract

Automated diagnosis of various skin lesion diseases through medical dermoscopy images is still a very challenging task. In this work, an integrated model for segmentation of skin lesion boundaries and classification of skin lesions is proposed by cascading novel deep learning networks. In the first stage, a novel full resolution convolutional networks (FrCN) is utilized to segment the boundaries of skin lesions from dermoscopy images. Then, the segmented lesions are passed into a deep residual networks (i.e., ResNet-50) for classification. The pre-segmentation process enables ResNet-50 to extract more specific and representative features from skin lesions and use them for improved classification. We have tested and evaluated our diagnostic deep model for skin lesions using the publicly available International Skin Imaging Collaboration (ISIC) 2017 challenge dataset which contains three different skin diseases: benign, seborrheic keratosis, and melanoma. The integrated model exhibits its capability to segment the skin lesions with an overall accuracy of 94.03% and an average Jaccard similarity index of 77.11% via FrCN. Meanwhile, the overall prediction accuracy and F1-score of multiple skin lesions classification task via ResNet-50 achieved 81.57% and 75.75%, respectively. The integrated model could be utilized as a computer-aided diagnosis (CAD) system for dermatology.