Deep Learning Approach for Skin Lesion Attributes Detection and Melanoma Diagnosis

Reliable identification of skin lesions is an important pre-requisite for diagnosing melanoma and other skin diseases. Established models of melanoma assessment suggest either methods of pattern analysis or seven-point checklist criteria for examining skin lesion. Automated and correct skin lesion identification and subsequent diagnosis of melanoma remain an unresolved challenge. In addition, the two evaluation methods have drawbacks and a trade-off. This paper proposes a method of pattern analysis combining 7-point checklist with convolution neural network for melanoma diagnosis, where the lesion features are automatically extracted. Both clinical and dermoscopic images were considered input into the established multi-input convolution neural networks (CNNs) where a separate feature extraction model was implemented from each image type. The performance of the developed algorithm is assessed using a 2000 dermoscopic image dataset. The results obtained from the proposed system show promising and competitive performance for detection of lesions and the automated diagnosis of melanoma from dermoscopic images.