Optimization of a mobile imaging system to aid in evaluating patients with oral lesions in a dental care setting.

Significance: Oral cancer is often diagnosed at an advanced stage even though the oral cavity is easily accessible for examination. There is a need for tools to support early detection and help non-expert clinicians make appropriate referral recommendations. Aim: We aim to evaluate a mobile detection of oral cancer (mDOC) smartphone-based autofluorescence imaging system by collecting data in a representative, low-prevalence population and optimizing a referral algorithm suitable for deployment in community dental settings. Approach: Data were collected from patients at a community dental clinic using the mDOC system, which captures white light and autofluorescence images. A multi-input deep learning algorithm was developed using both previously acquired and newly collected data in a rehearsal training method. Performance was evaluated on training, validation, and test sets using expert referral decision as the ground truth. Metrics included the Brier Skill Score, area under the ROC curve (AUC-ROC), sensitivity, and specificity. Results: Data from 252 anatomic sites across 50 patients were collected, with five sites in four patients referred to an oral cancer specialist based on expert review. The best-performing algorithm achieved a Brier Skill Score of 0.126. It reached an AUC-ROC of 0.986 on the validation set and an AUC-ROC of 0.778 on the test set. Sensitivity and specificity on the test set were 60.0% and 88.0%, respectively. Conclusions: These findings demonstrate the potential of the mDOC system and referral algorithm to assist in early detection and triage of oral mucosal lesions in low-prevalence, community dental clinic settings.