Deep Learning Model for Skin Disease Classification

Abstract

Skin diseases are very common health disorders, from common disorders to life-threatening cancers like melanoma. Their early and correct diagnosis is very important but typically challenging owing to similar appearances among diseases, different skin color, and a lack of experts. This thesis proposes a deep learning solution for the automatic diagnosis of skin diseases using dermatoscopic images, relying on CNNs and transfer learning. ResNet50, EfficientNetV2-B1, and Xception were trained on an improved version of the HAM10000 dataset. The results showed that Xception was the best performing at 98% accuracy, followed by EfficientNetV2-B1 at 96%, and then ResNet50 at 93%. The results show that CNN models, specifically Xception, can help dermatologists with quick and accurate diagnosis.