beaunix/dentex-ai

DentexAI — Dental Vision Models

This repository hosts the two vision models that power DentexAI, a dental assistant agent (FDI tooth numbering + lesion detection), feeding a PubMedBERT RAG and multi-tool backend.

Both models are built with Ultralytics YOLOv11. Full credit to Ultralytics for the open YOLO architecture and training tooling.

Models

1. FDI Tooth Numbering — dentex-fdi-numbering.{pt,onnx}

• Task: Instance segmentation, FDI tooth numbering scheme (35 classes: T11–T48, Bridge, Crown, Implant)
• Performance: mAP50 = 0.99
• Training data: FDI_numbering Dataset by Md Anas, Roboflow Universe, CC BY 4.0

2. Lesion Detection — dentex-lesion-detection.{pt,onnx}

• Task: Object detection of dental lesions
• Performance: mAP50 = 0.617
• Training data: sourced as a dataset archive from Mendeley Data. The original Mendeley record could not be relocated after the fact; the only surviving reference is a now-deleted Roboflow Universe demo project ("dental-disease-detection-hpn1d/demo-wjml3", CC BY 4.0, provided by a Roboflow user) noted at download time (2025-03-13). I do not remeber where does this second dataset come from.

Usage

ONNX Runtime (inference only)

```python import onnxruntime as ort

session = ort.InferenceSession("dentex-fdi-numbering.onnx") outputs = session.run(None, {"images": input_tensor}) ```

Ultralytics (.pt — inference or further training)

```python from ultralytics import YOLO

model = YOLO("dentex-fdi-numbering.pt") results = model.predict("panoramic_xray.jpg") ```

License

Released under AGPL-3.0, in line with the Ultralytics YOLO license these models were trained with (no Enterprise license was purchased).

Related

• Project repo: https://github.com/BeauBryanDev/dentex-ai/tree/master