Diagnosis with Deep Learning and a Novel Pre-processing Strategy on a Large-Scale Dermoscopic Image Dataset

Abstract

Skin cancer, in particular melanoma, has become a major health problem worldwide. Early diagnosis is the most important factor to consider for successful treatment. The latest advances in diagnosis have increased melanoma survival rates significantly, include early detection techniques such as imaging technologies that can detect melanoma in its earliest stages when treatment is most effective. AI has also been a game changer in the field of diagnosis, providing automated analysis data with a high level of accuracy. In this article, a novel computer-assisted diagnosis is presented, which consists of a new preprocessing technique to improve the quality of images. Data augmentation is used to increase data size by applying transformations to improve model generalization. The transfer learning efficiency is proved using the MobileNetV2 model. Improving and fine-tuning this architecture for the skin lesion classification task. The trained model can achieve performance, with an accuracy of 95.1% on 7 classes, and a very high AUC score of 94% for the precision-recall curve on the HAM10000benchmark dataset. These results show how advanced deep learning techniques can be used in dermatological practice, thus creating a promising alley for improved skin cancer diagnosis.