Dermatoscopy Image Classification Using Convolutional Neural Networks

This paper examines the problem of diagnosing skin melanoma based on dermatoscopy images using modern computer technologies. In particular, it considers the problem of classifying dermatoscopy images using convolutional neural networks to obtain preliminary melanoma diagnoses. The main open sources that can be used to create image datasets for network training and testing are described. The challenges encountered in creating a dataset for neural network training and solutions to problems associated with the imbalance in existing datasets are discussed. Two neural network architectures are proposed for solving the problem of dermatoscopy image classification: a simple convolutional network and a network built on the Inception architecture. The key metrics used to assess diagnostic quality in medical applications are described. The results obtained using the proposed networks are compared, and the choice of Inception is justified as providing greater sensitivity in melanoma detection. The developed networks were tested using our own database of dermatoscopy images. Possible causes of discrepancies between the test and validation image sets were analyzed, and the need for further data augmentation during training was substantiated. Tools for additional data augmentation were proposed and integrated into the network architecture during the training stage. A comparison of the results obtained through augmentation was performed, demonstrating a significant improvement in the network sensitivity when working with the test set of images.