Early Detection of Oral Ulcers Using Photographic Evidence: A Novel Approach using Ensemble of Convolution Neural Network

Oral lesions are difficult to diagnose and often predispose as a challenge to clinicians due to the intersection of their properties and physical appearances. Early detection of mouth ulcers is critical as they indicate an underlying condition such as Autoimmune disorders (e.g., SLE), inflammatory conditions (e.g., reactive arthritis), bacterial or fungal infections (e.g., secondary syphilis or candida) or deficiencies like Vitamin B12 or Iron deficiency. In this paper, we proposed novel approach that combines convolutional neural network (CNN) and ensemble learning (EL), called CNN-EL, to identify the type of oral ulcer using photographic evidence: i.e., classification between (1) traumatic ulcer and infectious ulcer, (2) infectious ulcer and malignant ulcer, and (3) malignant and traumatic ulcer. For each task, we trained a large number of CNN models on photographic images obtained from 89 patients with oral ulcers. The CNN models with the highest predictive performance were then ensemble. We evaluated performance using five-fold cross-validation method, revealing an accuracy rate of 85.7 ± 0.07, 88.2 ± 0.05, 91.1 ± 0.03 for classifying traumatic ulcers vs. infectious ulcers, infectious ulcers vs. malignant ulcers, and traumatic ulcers vs. malignant ulcers, respectively. These results suggest that our CNN-EL approach can discriminate among the oral ulcers.