OPMD Classification using a Mixture of Experts Framework

Oral Potentially Malignant Disorders (OPMDs) pose significant diagnostic challenges due to subtle visual differences and severe class imbalance, particularly in resource-limited healthcare settings. To address these issues, we developed a multi-expert classification framework that utilizes a mixture of expert models, each leveraging the MobileNetV2 architecture with distinct loss functions tailored for class imbalance. The framework includes three specialized experts: the Backward Expert (Inverse Softmax Loss), the Forward Expert (Weighted Cross-Entropy Loss), and the Uniform Expert (Balanced Softmax Loss). An aggregation module combines these experts' outputs, enhancing the model's ability to generalize across rare and common classes. In test evaluations, the aggregated Mixture of Experts model achieved an accuracy of 0.636 with a weighted F1 score of 0.422, outperforming individual models and showing promise for early OPMD screening in underserved settings. The codes are available in GitHub repository.