Multi-classification Network for Detecting Skin Diseases using Deep Learning and XAI

This study compares four deep learning models in order to find the best performing model to construct a skin disease detection system for diagnosing three dangerous but uncommon skin conditions: Stevens-Johnson Syndrome, Erythema Multiforme, and Bullous Pemphigoid. Due to the ambiguity of these models, this work also makes an effort to explain these models’ predictions using Explainable Artificial Intelligence that is based on LIME. For feature extraction, the four pre-trained models ResNet50V2, VGG16, InceptionV3, and InceptionResNetV2 have been employed whose top layer has been replaced with a custom-made 10-layered architecture made up of Convolution, BatchNormalization, Dropout, and Dense Layers. An unprecedented hybrid dataset made up of colorful images of the diseases gathered from various sources (to subject the models to a plethora of varying images) served as the training ground for these models. Additionally, other machine learning classifiers, including Support Vector Machine, Random Forest, and Logistic Regression have been employed to identify the images in order to see how well they perform in comparison to a neural network approach. To determine which of the tried deep learning methods performs best, their accuracies have also been compared to one another. According to the analysis, the InceptionResnetV2 model has a 99.06 percent accuracy rate, whereas InceptionV3, VGG16, and Resnet50V2 have accuracy rates of 90.27 percent, 95.92 percent, and 98.26 percent, respectively.