argos-dentsight-stage1-fdi-v1

Status: research preview — v1 has a known classification-head calibration failure. v2 retrain with cls-head LR-multiplier + focal_alpha=0.25 is in progress. The reported mAP@0.5 = 0.74 is a ranking-only number computed at threshold=0.0; at any usable inference threshold the model returns no detections (max sigmoid ≈ 0.07 on real OPGs). Box localization is correct; per-class confidence is not. Treat v1 outputs as positional skeleton only.

Stage-1 of the argos-dentsight two-stage transformer detector for panoramic dental X-rays (OPGs). This head localizes the 32 FDI tooth positions (11–18, 21–28, 31–38, 41–48); stage 2 (Mobe1/argos-dentsight-stage2-conditions-v1) detects 13 condition classes within the same OPG.

Model description

D-FINE-Medium fine-tuned from ustc-community/dfine-medium-coco for 18 epochs on Mobe1/argos-dentsight-opg-v1 (v8 OPG REPORTING corpus, 32-FDI subset). Input 1024×512 (aspect-preserving Albumentations pipeline + pad). Single-LR AdamW; focal loss α=0.75 (inherited from D-FINE checkpoint), γ=2.0.

Intended use

• Research / dissertation companion to the AI The Dentist MSc dissertation (Tun Ye Minn, University of Essex). Reproduces and extends the dissertation's 41-class ensemble pipeline.
• Per-tooth bounding-box ranking for downstream IoU-match-to-stage-2 pipeline. Use the argmax-per-class boxes (top-1 per FDI), NOT the absolute confidence scores.

Out of scope

• Any clinical decision-making. This model has not been validated against dentist agreement.
• Standalone use without v2 — v1's confidence scores are not calibrated and should not be surfaced to clinicians or end users.
• Billing, insurance, or treatment-recommendation pipelines.

Training data

Mobe1/argos-dentsight-opg-v1 (private). v8 of the Roboflow "OPG REPORTING" corpus normalized into a 32-FDI

• 13-condition unified label space. Stage-1 training filters each row to FDI annotations only; condition annotations are dropped at this stage.

Evaluation results

metric	value	note
validation mAP@0.5 (threshold=0.0)	0.7444	ranking-only — not a deployment number
validation mAP@0.5 (threshold=0.05)	0.367	reproduced post-hoc
validation mAP@0.5 (threshold=0.30)	0.000	the demo's old default of 0.6 was guaranteed empty
max sigmoid score on a held-out OPG	0.069	below the focal-loss prior 1/33≈0.030
quadrant accuracy on per-class top-1 boxes	16/32 (50%)	classification near random; localization correct

Acceptance-gate verdict: The dissertation spec called for mAP@0.5 ≥ 0.92 (docs/specs/2026-05-02-dental-detection-v1-design.md). v1 fails this gate even on the ranking-only number (0.74 vs 0.92), and the head-collapse pathology means the gap is not just margin error.

Limitations (v1)

• Classification head bias collapse. All 32 logits in decoder.class_embed[3].bias drifted < 0.005 from the focal-loss bias prior −log(32) ≈ −3.466 over 18 epochs. The reported per-class mAP differences reflect ordinal ranking + box-IoU quality, not class-specific learning. Inference on real OPGs caps at ~0.07 sigmoid score; using any threshold ≥ 0.10 returns zero detections.
• Box regressor did converge — the per-class argmax boxes localize teeth correctly. The cls head is the bottleneck.
• v2 retrain in progress: 10× LR multiplier on class_embed / enc_score_head / denoising_class_embed; focal_loss_alpha 0.75 → 0.25.

Training procedure

Hyperparameters (v1)

• learning_rate: 5e-05 (single AdamW group)
• train_batch_size: 8
• num_epochs: 18
• focal-loss α: 0.75 (inherited from D-FINE checkpoint), γ: 2.0
• input resolution: 1024×512 (Albumentations FitInBox + pad)
• HF Jobs flavor: a10g-large (~40 min wall-clock, ~$1.07)

Framework versions

• Transformers 4.57.6
• PyTorch 2.x with CUDA
• Datasets 4.8.5
• Tokenizers 0.22.2

Citation

Companion model to the AI The Dentist MSc dissertation (Tun Ye Minn, University of Essex, MSc AI). Predecessor codebase: Raghu2411/AI_The_Dentist.

License

Inherits apache-2.0 from base model ustc-community/dfine-medium-coco. Project-license decision is open per docs/specs/2026-05-02-dental-detection-v1-design.md (open item §10 row 8); confirm before any public release.