Applications of a Convolutional Neural Network to Aid Patients in Performing Machine-Learning Aided Dermatological Diagnosis
Abstract – Over a quarter of the world suffers from over 1500 skin conditions. With the shortage of physicians and user propensity to self-diagnose through a Google search instead of visiting a doctor, misdiagnoses or neglect of skin conditions is becoming widespread. Data scientists have started stepping in to help by offering machine learning models that can help identify skin conditions. These efforts are in the nascent stages with few solutions that have high-confidence outcomes. Machine learning and deep learning algorithms have been applied to tasks of medical image segmentation in areas such as spotting tumors, etc. before. New research applies these algorithms to dermatological disease recognition and aiding in their medical diagnosis. This paper explores Convolutional Neural Network (CNN) models and refinement algorithms to optimize machine learning aided diagnosis, specifically in the context of dermatological condition recognition. 22 skin diseases were reviewed for the purposes of this study, featuring 10 different body segments. 7 pre-trained CNN models were tested to determine optimal architectures for a given task, determining the ideal candidate for refinement. Backpropagation and support vector machine algorithms were utilized to refine the ideal CNN model post-experimentation. The final proposed model of Xception with backpropagation and support vector machine algorithms yielded an average accuracy of 0.9969 across 10 body parts. By allowing for the processing of nonlinear data and adjusting kernel weights, we saw significant improvements in pattern recognition amongst similar appearing symptoms, achieving another objective of this study, which was detecting dermatological anomalies with high accuracy.