Development and evaluation of a deep learning model for dental image classification
TL;DR
OdontoAI, a minimum viable product designed for the automatic classification of dental images into six categories, demonstrates that rigorous curation protocols directly affect reported metrics, making direct comparisons with studies that do not control for data leakage across partitions invalid.
Abstract
Dental image analysis can be affected by factors such as lighting, framing, image quality, and visual similarity among different oral conditions. This study aims to present the development and evaluation of OdontoAI, a minimum viable product designed for the automatic classification of dental images into six categories: calculus, caries, gingivitis, hypodontia, mouth ulcer, and tooth discoloration. The project used the public dataset Oral Diseases (Kaggle), submitted to a rigorous curation process including duplicate removal by MD5 hash, grouping by perceptual hash (pHash), and controlled partitioning at a 70/15/15% ratio. The final dataset contained 3,507 images. Classification was performed using a ResNet-50 architecture adapted through two-phase transfer learning. On the test set (524 images), the model achieved an accuracy of 96.18% and a macro F1-score of 90.78%. The best results were observed for gingivitis and hypodontia; caries presented a recall of 64.29%, reflecting difficulties associated with class imbalance and visual similarity between categories. Qualitative Grad-CAM analysis indicated activations in relevant dental regions, treated as a complementary resource rather than a diagnostic tool. The study demonstrates that rigorous curation protocols directly affect reported metrics, making direct comparisons with studies that do not control for data leakage across partitions invalid.