Malignant Melanoma Classification Using Cross-Platform Dataset with Deep Learning CNN Architecture

Melanoma, a malignant skin lesion, is the deadliest of all types of skin cancer. Deep learning has been shown to efficiently identify patterns from images and signals from various application domains. Use of deep learning in medical image analysis is, however, limited till date. In the present paper, two well-known malignant lesion image datasets, namely Dermofit and MEDNODE, are both separately and together used to analyze the performance of a proposed deep convolutional neural network (CNN) named as CNN malignant lesion detection (CMLD) architecture. When Dermofit and MEDNODE datasets are used separately with tenfold data augmentation, the CNN gives 90.58 and 90.14% classification accuracy. When the datasets are mixed together the CMLD gives only 83.07% accuracy. The classification accuracy of the MEDNODE dataset using deep CNN is considerably high in comparison with the results found in the related literature. The classification accuracy is also high in case of Dermofit dataset in comparison with the traditional feature-based classification.