Skin melanoma detection based on hyperspectral imaging and deep-learning techniques

Skin cancer is one of the most common cancers. Most skin cancers are not life threatening, but malignant melanoma is fatal. Currently, it still remains a challenge to discriminate malignant melanoma from benign melanoma by using conventional diagnostic techniques, such as ultrasonography, computed tomography, magnetic resonance imaging and positron emission tomography. As a new type of bio-optical imaging technology, hyperspectral imaging (HSI) has become the focus of research. It can provide information about hemoglobin and melanin content for the differentiation of various skin diseases. In this study, we propose a hyperspectral imaging system based on push-broom imaging spectrometer to image skin-pigmented nevus, and then segment the nevus out from surrounding normal skin through pixel-wised spectrum classification with deep learning techniques. The HIS system can produce hyperspectral image over the spectral range of 465-630nm and with a spectral resolution of 2.1 nm. Meanwhile, we evaluated the performance of K-means, Gaussian Mixture Model (GMM) and Hierarchical Clustering (HAC) in detecting the nevus with manual segmentation as the gold standard. The results show that these three techniques all have a good accuracy in differentiating the nevus from normal skin, which proves that the hyperspectral system combined with classification techniques has a good potential to detect the pigmented nevus on the skin.