Explainable Deep Learning for Oral Disease Detection with LLM Decision Support.

Abstract

This thesis develops a comprehensive pipeline that automate detection and classification of oral diseases from RGB intraoral photographs, which addresses the three main challenges of accuracy, interpretability, and clinical relevance. Using an advanced convolutional neural network InceptionResNetV2, fine-tuned on the Mouth and Oral Disease dataset, which includes images across seven classes which includes Canker Sores, Gingivostomatitis, Oral Cancer, Oral Lichen Planus, Cold Sores, Oral thrush and Mouth Cancer. The model achieved an overall accuracy of 99.60%, with F1-scores, precision, and recall that were near perfect for most classes. To support transparency, we use Grad-CAM++ to produce heatmaps that highlight clinically relevant lesion regions, enabling dentists to interpret and validate the model’s predictions. Furthermore, to generate human readable narrative a large language model DeepSeek R1 was used: one prompt generates clinically detailed explanations for dentists, including disease descriptions, potential causes, and management guidelines; while another prompt generates simplified, patient-friendly explanations to improve understanding and engagement. This dual-perspective approach ensures that both healthcare professionals and patients benefit from the diagnostic pipeline, enhancing confidence, interpretability, and usability.