Skin Disease Classification Using CNN with Transfer Learning

Abstract

Skin diseases are one of the most common public health issues and some of these conditions like melanoma are very dangerous to human lives unless detected at an early age. The latest development in deep-learning algorithms has shown a lot of potential in analyzing medical images, especially in the field of disease diagnosis based on visual information. This paper examines the process of automatic detection of dermatological disorders using dermoscopic visualization and convolutional neural networks (CNNs). This research is based on the HAM10000 dataset which consists of dermoscopic images of seven different categories of skin diseases. Both the standard CNN architecture and the pre-trained models of deep-learning are used: AlexNet is trained as a standard reference, and ResNet-18 and DenseNet are fine-tuned on the basis of the transfer learning. Strict data preprocessing guidelines and class-balancing strategies are followed to improve the performance of the model. Performance indicators such as accuracy, precision, recall, and F1-score are embraced to determine the effectiveness of the model. The results of the experiments prove that transfer learning significantly increases the performance of the classification, and the overall accuracy of DenseNet was the highest of the considered models. The results highlight the possibility of deep-learned diagnostic systems to be used successfully to help dermatologists diagnose skin diseases correctly.