Enhanced Super-Resolution Generative Adversarial Networks (ESRGAN) on Melanoma Skin Cancer Detection

Abstract

Neural networks and image enhancement have become significant in oncology by aiding in the early diagnosis of various cancer types, including Melanoma, the most lethal type of skin cancer. The conventional approach to diagnosing melanoma includes using dermoscopic tools to capture skin lesions. However, these captured skin lesions have limitations, such as low resolution, artifacts on the skin, and variations in lighting conditions. One promising method for improving the resolution of these dermoscopic images is the Enhanced Super-Resolution Generative Adversarial Network (ESRGAN). To evaluate the effectiveness of ESRGAN in melanoma skin cancer detection, researchers trained CNN models with ESRGAN-enhanced images and not enhanced images and compared their performance. They used the ISIC 2020 dataset and balanced it with random undersampling and data augmentation. The study utilized two deep learning models, VGG16 and ResNet50, to compare their performance with and without ESRGAN enhancement. The results showed that the enhanced dataset outperformed the unprocessed dataset, with ResNet50 achieving an impressive accuracy of 98.2% and VGG16 achieving 94.74%. Additionally, training with the enhanced dataset took 5 minutes longer in VGG16 and 18 minutes longer in ResNet50 which led to significantly better results. In conclusion, the study shows that ESRGAN can improve the performance of deep learning models in melanoma skin cancer detection.