Background: Oral cancer is one of the ten most common cancers worldwide, with most cases detected at late stages leading to poor survival. Accurate diagnosis of oral lesions, some of which may be malignant, is essential to lessen the disease burden. Deep learning (DL) methods have recently succeeded in image recognition, prompting its use to reduce diagnostic delays and uncertainty. However, most DL models do not produce any estimate of their predictive uncertainty and often perform poorly in out-of-distribution settings.  

Objective: We develop an evidential deep learning (EDL) model that accurately classifies oral lesion types from intra-oral images while providing uncertainty estimates. We also compare the EDL model with a Bayesian DL approach in uncertainty quantification. 

Methods: A retrospective cohort of electronic health record gathered by an oral pathologist during routine examinations, was used. Ground truth labels were obtained via biopsy or expert diagnosis. The dataset includes 12,865 images of 66 oral lesions from 6000 patients. Model: An EfficientNet-B5 pretrained on ImageNet served as the backbone. Two approaches were compared: (i) Evidential Deep Learning (EDL), which uses the Dempster-Shafer Theory to estimate Dirichlet distribution parameters, and (ii) MC-DropConnect (MCDC), an approximate Bayesian method that produces samples from the posterior predictive distribution. 

Results: The EDL model achieved an average F1-score of 71.4% versus 69.7% for MCDC. Both models had similar discrimination (AUC = 0.94 and 0.92, respectively), though the EDL produced better-calibrated probability estimates. 

Conclusion: Our study shows that both uncertainty-aware DL approaches perform comparably for oral lesion diagnosis. However, the EDL model offers the benefit of reliable uncertainty quantification without requiring multiple forward passes. These findings highlight the potential of uncertainty-aware DL models to improve decision-making in oral cancer diagnosis.