Automatic Diagnosis of melanoma using Hyperspectral data

Abstract

Melanoma is a type of superficial tumor and should be treated in an early stage. Early-stage melanoma is difficult to diagnose because it looks like a benign lesion. However, melanoma is still subjectively diagnosed by a dermatologist. Therefore, there is strong need for development of a quantitative diagnostic method. We are developing an automatic melanoma diagnostic system using convolutional neural network and hyperspectral imager. Hyperspectral imager acquires position and wavelength information simultaneously. 3D patches (16x16x201) extracted from hyperspectral data (560x680x210) is input to our newly proposed network. The total number of hyperspectral data was 202, including 88 melanomas and 114 non-melanomas. Our dataset was evaluated by 5-fold cross-validation and calculated sensitivity, specificity and accuracy. In this paper, the result of training and validation using patch data is reported.