Using Deep Learning Models Based on WGAN: A Solution to Improve Melanoma Diagnosis in Dermoscopy Images

Melanoma is one of the types of cancerous skin lesions, where early detection is crucial to prevent patient mortality. One method for early detection of melanoma involves using dermoscopic images of skin lesions to train deep learning models, which can then be used to classify skin lesions in patients, including the diagnosis of melanoma. A significant limitation of deep learning models is their need for substantial amounts of labeled data. This article discusses data augmentation using the Wasserstein GAN (WGAN) network to overcome the issue of limited diversity in images generated by GAN networks (a problem known as Mode Collapse). By generating ۵,۰۰۰ high-quality synthetic images of the melanoma class and adding these images to the unbalanced HAM۱۰۰۰۰ dataset, an improved accuracy in diagnosing this disease was achieved using the pre-trained deep ResNet۵۰ model. The proposed model improved melanoma classification accuracy by ۱۰% without significantly altering the overall model accuracy. These results suggest that using the WGAN network for data augmentation can enhance the classification accuracy of melanoma.

Keywords:

Melanoma , Data Augmentation , GAN , WGAN , Pre-trained Deep Learning Model , Dermoscopy