---
license: cc-by-4.0
tags: [cephalometric, landmark-detection, hrnet, medical-imaging]
---

# CEPHA29 cephalometric landmark model (HRNet-W32)

29-landmark HRNet-W32 for automatic cephalometric landmark detection, used by
Cephalometry Studio for in-browser auto-tracing (ONNX Runtime Web, WASM).

## Files
- `landmarks-cepha29.int8.onnx` — static-QDQ INT8, 768×768×3 input, output [1,29,192,192] (29 heatmaps)
- `landmarks-cepha29.onnx` — FP32 (same MRE)

## Performance (valid split, 150 images)
MRE 1.24 mm · SDR@2mm 83.1% · SDR@3mm 93.7% · SDR@4mm 97.2%

## Channel order (must match output index)
A ANS Ar B Co Gn Go LIA LIT LMT LPM Li Ls Me N N` Or PNS Pn Po Pog Pog` R S Sn UIA UIT UMT UPM

## Training
CephaloHRNet (MIT) · HRNet-W32 ImageNet backbone · CEPHA29/Aariz averaged annotations.

## Dataset & attribution
Trained on the **Aariz Cephalometric Dataset (CEPHA29)** — Figshare, **CC BY 4.0**,
DOI 10.6084/m9.figshare.27986417. Training code: **CephaloHRNet** (MIT). Backbone:
HRNet-W32 ImageNet (`hrnet_w32.ms_in1k`, Apache-2.0).

## Disclaimer
Research and education only. **Not a medical device**; not for diagnosis or treatment.