Skin Diseases Diagnosis System Based on Machine Learning

Abstract

Skin diseases are very common across the globe and especially in underdeveloped countries. Increasing climate changes and global warming are among the major causes of cancerous skin diseases. Skin diseases can be deadly if not diagnosed and treated at their initial stages. Clinical skin disease diagnosis methods are very rarely available for a common person and these methods are very expensive as well. The advancement in the field of image processing and machine learning made it possible to use machine learning methods for feature extraction and diagnosis of medical illnesses. Deep learning on the emerging subfields of machine learning methods that are more advance and intelligent in the classification of images based on different features. In this work, different deep learning models are used which are based on Convolution Neural Networks (CNN) architecture like ResNet-50, the DenseNet-121, and sequential CNN model for feature extraction and the classification of skin disease categories using HAM10000 dataset from ISIC archive 2018. The HAM10000 dataset consists of 10015 images from seven different skin disease categories. The dataset is further subdivided into testing and training data with a ratio of 20% and 80% correspondingly. The skin disease diagnosis is a three-step process that includes dataset images preprocessing in the first stage, feature extraction using CNN models in the second stage, and classification of skin disease category based on the extracted features using different classifiers in the third stage. Image augmentation techniques were used to reduce the imbalance in different skin disease classes. The sequential CNN-based model using 7 convolution layers gives an accuracy of 98% and 99% Area Under the Curve (AUC) for all seven skin disease categories. ResNet-50 and DenseNet-121 give an accuracy of 84% and 89% respectively. The performance of different CNN models on the given dataset is compared and evaluated using different performance matrixes.