PanorAI-FDI: Automated Tooth Enumeration in Panoramic Radiographs Using YOLOv8 with Progressive Training

Background/Objectives: Automated tooth enumeration in panoramic radiographs is a clinically relevant yet underexplored task, requiring the classification of up to 32 tooth positions following the FDI numbering system (ISO 3950). Manual interpretation is time-consuming and subject to inter-observer variability, motivating the development of reliable automated systems. Methods: We propose PanorAI-FDI, a YOLOv8n-based detection system trained on the DENTEX 2023 benchmark (634 images, 18,095 annotations) using a progressive training strategy with checkpoint reuse across incremental rounds of 10, 20, and 30 epochs. Progressive training is compared against a direct 30-epoch baseline initialized from the same COCO pre-trained weights. A spatial interpolation mechanism for estimating missing tooth positions within each quadrant is also introduced. Results: Progressive training substantially outperforms direct training, achieving mAP@0.5 of 0.9128 vs. 0.632 (+44%) and mAP@0.5:0.95 of 0.4974 vs. 0.284 (+75%), with precision of 0.8241 and recall of 0.8670. Consistent metric improvement was observed across all three training rounds with no signs of overfitting. Conclusions: Progressive training with checkpoint reuse is an effective strategy for 32-class FDI tooth enumeration under limited data conditions, enabling competitive performance with a lightweight model trained on only 634 images. The approach shows potential for extension to other medical imaging tasks with constrained datasets.