The updated AI model for dental diagnostics, based on the YOLOv8l architecture, was trained on a refined dataset named Clean_Dental_v3. The training process utilized a two-stage approach on an NVIDIA A100 80GB GPU, with the final training phase lasting 29.3 minutes.
The dataset was compiled from three sources: the DENTEX Challenge, the Tufts Dental Database, and the Loki_Dental repository. The final training set includes four specific pathology classes: Caries, Deep Caries, Impacted teeth, and Periapical Lesion. Data cleaning involved removing 33 duplicates and filtering out 1 571 images based on aspect ratios and color saturation.
The model achieved an mAP50 of 0.642 on the validation set. During testing, the researchers determined that using Test-Time Augmentation (TTA) decreased precision for the Deep Caries and Periapical Lesion classes, leading to the decision to disable this feature in the final pipeline.
The team evaluated a Model Soup approach, which involved averaging the weights of the first and second training stages. While this method increased average precision at the maximum F1 point to 75.6%, it resulted in a lower overall mAP50 compared to the standalone second-stage checkpoint, which was ultimately selected for deployment.
