CANCER LESION CLASSIFICATION WITH GAN-BASED IMAGE AUGMENTATION METHOD FROM SKIN IMAGES

Abstract

Skin cancer is one of the most common types of cancer in the world and an important public health problem. Rapid diagnosis and correct treatment are important in the treatment of skin cancer. However, the in-depth imaging and analysis required to diagnose skin cancer is a time-consuming process that requires expertise. Unfortunately, the lack of labeled data in the classification and the limited size of training datasets affect the performance of deep learning architectures and the correct diagnosis of the disease. In this study, cancer lesion classification was performed by using deep convolution networks with a GAN (Generative Adversarial Network) based data augmentation method for skin lesion classification. This model synthesizes new skin lesion images by analyzing skin lesion images in training data. The images synthesized by GAN are processed by the classification model in the same way as the real images and added to the training set for better classification performance. In all classification tasks, 3 different convolutional neural network architectures were used for classification and feature extraction. The classification accuracy of the architectures consisting of images obtained with the ISIC (International Skin Imaging Cooperation) Archive dataset is 86.00% for CNN, 90.81% for MobileNet, and 81.66% for ResNet-18, respectively. The performance of the models in the hybrid dataset consisting of the combination of synthetic images and the original primary dataset increased to 86.34% for CNN, 91.67% for MobileNet and 91.78% for ResNet-18. This study shows that GAN-based data enhancement models in skin lesion classification also make an important contribution to medical image analysis.