Multi-class Skin Disease Classification Using Transfer Learning Model

Abstract

The human body’s major organ is the skin, and it protects human beings from the outside environment. Detecting skin disease at an earlier stage is a big challenge because of the similar appearance of skin disease. Although skilled dermatologists find it challenging to forecast skin lesions due to lack of contrast between adjoining tissues. Therefore, there is a need for an automated system that can detect skin lesions timely and precisely. Recently Deep Learning (DL) has attained outstanding success in the diagnosis of various diseases. Thus, in this paper, a transfer learning-based model has been proposed with help of pre-trained Xception model. The Xception model was modified by adding layers such as one pooling layer, two dense layers and one dropout layer. A new Fully Connected (FC) layer changed the original Fully Connected (FC) layer with seven skin disease classes. The proposed model has been evaluated on a HAM10000 dataset with large class imbalances. The data augmentation techniques were applied to overcome the unbalancing in the dataset. The new results showed that the model has attained an accuracy of 96.40% for classifying skin diseases. The proposed model is working best on Benign Keratosis and the values of precision, sensitivity and F1 score are 99%, 97% and 0.98 respectively. This method can provide patients and doctors with a good notion of whether or not medical assistance is required, thus, avoiding undue stress and false alarms.