Artificial Intelligence (AI) in the Identification of Pemphigus Vulgaris as Compared with Healthy Mucosa Using Clinical Images: A Retrospective Study

Background: Pemphigus is a chronic autoimmune inflammatory condition characterized by the formation of blisters and ulcers on mucous membranes and skin. Proper diagnosis is essential for managing the disease. The gold standard method for diagnosis is biopsy. Objective: To assess the diagnostic accuracy and precision of AI algorithms in distinguishing pemphigus lesions from healthy oral mucosa. Methods: 200 photographic images, including 100 buccal mucosa images of Pemphigus vulgaris and 100 healthy buccal mucosa images, were used. A machine learning approach was employed using logistic regression, k-nearest neighbors algorithm [kNN], and random forest algorithms in Orange software to detect pemphigus vulgaris. Performance evaluation included accuracy, sensitivity, specificity, precision, recall, F1 score, and AUC-ROC curve. Results: The logistic regression model and random forest accurately distinguished pemphigus lesions from healthy mucosa, respectively, at 88.5% and 81.4%. In contrast, the kNN model showed an accuracy of 79.9% with an AUC (area under the curve). Conclusion: The logistic regression and random forest algorithms outperformed the kNN algorithm in detecting pemphigus images. Their superior performance is closely aligned with manual evaluations, particularly in terms of AUC. This indicates the promising potential of AI for automatically detecting pemphigus vulgaris lesions.