Skin Disease Classification using Pre-Trained Convolution Neural Network with Transfer Learning

Skin disease diagnosis is critical for prompt intervention, effective treatment, public health management, and general well-being of people affected by dermatological disorders. Skin disease detection is a challenging task while classifying those diseases based on the images of common symptoms such as blisters- like rashes, skin lesions, pimples like papules & reddened skin is further more complex. The objective of this paper is to illustrate the effectiveness of transfer learning and application of the pre-trained models originally designed for the Imagenet Dataset in the context of detecting and categorizing various skin diseases. The aim of the model is to differentiate among monkeypox, measles, chickenpox and normal skin. The proposed framework utilizes the convolutional and pooling layers of ResNet50, DenseNet121, EfficientNetB5 and Xception which are pretrained models used for feature extraction. It incorporates a newly designed fully connected (FC) layer along with an output layer for the purpose of image classification. The proposed model is based on Monkeypox Skin Images Dataset (MSID) created at Islamic University in Bangladesh. The proposed framework exhibits that Xception achieved the best result among all the four pre-trained networks with 98.39% as training accuracy, 96.46% as validation accuracy and 95.80% as testing accuracy.