Melanoma Segmentation and Classification in Clinical Images Using Deep Learning

In this paper, a deep learning computer aided diagnosis system (CADs) is proposed for automatic segmentation and classification of melanoma lesions, containing a fully convolutional neural network (FCN) and a specific convolutional neural network (CNN). FCN, which consists of a 28-layer neural structure, is designed for segmentation and with a mask for region of interest (ROI) as its output. Later, the CNN only uses the segmented ROI of raw image to extract features, while the DLCM features, statistical and contrast location features extracted from same ROI are merged into CNN features. Finally, the combined features are utilized by the fully connected layers in CNN to obtain the final classification of melanoma, malignant or benign. The training of FCN and CNN are separated with different loss functions. Publicly available database ISBI 2016 is used for evaluating the effectiveness, efficiency, and generalization capability with evaluating indicator, such as accuracy, precision, and recall. Preprocessing methods, such as data argumentation and balancing are utilized to make further improvements to performance. Experiments on a batch size of 100 images yielded an accuracy of 92%, a specificity of 93% and a sensitivity of 94%, revealing that the proposed system is superior in terms of diagnostic accuracy in comparison with the state-of-the-art methods.