CephTrace/cephtrace-v4

CephTrace v4 — Anatomy-Guided Cephalometric Landmark Detection

1.050 mm MRE across 25 landmarks on a 151-image held-out test set, using image-adaptive spatial priors generated by anatomical analysis of each radiograph.

Model Description

CephTrace v4 is a two-stage pipeline for automatic cephalometric landmark detection from lateral skull radiographs:

• Stage 0 (Anatomical Initialization): A multi-phase module that detects the soft-tissue profile, partitions the image into anatomical zones, extracts bony contours, derives anchor landmarks via geometric rules, and generates 25 per-landmark Gaussian attention maps — all adapted to each patient's individual anatomy.
• Stage 1 (Heatmap Regression): An HRNet-W32 backbone (32M params) that accepts the 28-channel input (3 RGB + 25 attention maps) and outputs 25 landmark heatmaps at 256×256 resolution.

The key innovation is that the attention priors are image-adaptive: each patient receives maps centered at their estimated anatomy, not fixed population-average positions. Controlled experiments show this reduces MRE by 30.9% compared to the same architecture without priors.

ONNX Models

All models are exported as ONNX (opset 14) for cross-platform inference.

File	Stage	Purpose	Size	Input	Output
v4_stage0_profile.onnx	0A	Soft-tissue profile segmentation	26.8 MB	(1,1,512,512) float32	(1,1,512,512) sigmoid mask
z1_cranial_base_contours.onnx	0C	Cranial base contour segmentation	26.8 MB	(1,1,256,256) float32	(1,1,256,256) logits
z2_midface_contours.onnx	0C	Midface contour segmentation (palatal + upper incisor)	26.8 MB	(1,1,256,256) float32	(1,2,256,256) logits
z3_mandible_contours.onnx	0C	Mandible contour segmentation (border + symphysis + lower incisor)	26.8 MB	(1,1,256,256) float32	(1,3,256,256) logits
z4_posterior_contours.onnx	0C	Posterior contour segmentation (mandible + cranial base)	26.8 MB	(1,1,256,256) float32	(1,2,256,256) logits
phase0e_model.onnx	0E	Anchor → derived landmark MLP	455 KB	(1,14) float32	(1,36) float32
v4_stage1.onnx	1	HRNet-W32 heatmap regression	130 MB	(1,28,512,512) float32	(1,25,256,256) float32

Total: 264 MB

Pipeline Flow

Lateral Cephalogram (any resolution)
    │
    ▼ resize to 512×512
Phase 0A ──► Soft-tissue profile mask (Dice 0.80)
    │
    ▼
Phase 0B ──► 5 anatomical zones + 6 soft-tissue landmarks (geometric rules)
    │
    ▼ per-zone CLAHE enhancement
Phase 0C ──► Bony contour masks (4 zone-specific U-Nets)
    │
    ▼ Douglas-Peucker simplification
Phase 0D ──► 7 anchor landmarks (0.11 mm MRE, topological rules)
    │
    ▼
Phase 0E ──► 18 derived landmarks (MLP, 114K params)
             + 25 Gaussian attention maps (256×256, 3-tier σ)
    │
    ▼ bilinear upsample to 512, concat with RGB → 28 channels
Stage 1  ──► 25 heatmaps (256×256) → peak decode → 25 landmarks

Inference time: ~410 ms total (Stage 0: ~40 ms, Stage 1: ~350 ms) on A100 GPU.

Landmark Set (25 landmarks, CANONICAL_25 order)

 0: S (Sella)           1: N (Nasion)          2: Or (Orbitale)
 3: Po (Porion)         4: ANS                 5: PNS
 6: A (Subspinale)      7: B (Supramentale)    8: Pog (Pogonion)
 9: Gn (Gnathion)      10: Me (Menton)        11: Go (Gonion)
12: Ar (Articulare)    13: Co (Condylion)     14: U1_tip
15: U1_root            16: L1_tip             17: L1_root
18: UL (Upper Lip)     19: LL (Lower Lip)     20: Pm (Pterygomaxillare)
21: Ba (Basion)        22: Pog_soft           23: Sn (Subnasale)
24: Prn (Pronasale)

Performance

Controlled Ablation (151-image held-out test set)

Configuration	Input	MRE (mm)	SDR@2mm
HRNet backbone (no priors)	3-ch	1.520	86.6%
HRNet + Phase 0E priors	28-ch	1.050	87.8%
Improvement		0.470 (30.9%)	+1.2%

Same 1,201 training images, architecture, and recipe. Only variable: prior channels.

Prior Ablation

Configuration	MRE (mm)	vs. No Priors
Random priors (shuffled channels)	2.240	+15.6% worse
No priors (baseline)	1.938	—
Fixed textbook priors	1.869	−3.6% (marginal)
Image-adaptive priors (Phase 0E)	1.043	−46.2%

Attention Map Confidence Tiers

Tier	σ (at 256×256)	Landmarks	Mean Improvement
High	5–7	S, N, Me, ANS, Prn, Sn	−0.74 mm
Medium	8–13	Go, Gn, Pog, Or, UL, LL, Pog', A	−0.44 mm
Low	18–22	Po, Co, B, PNS, U1r, L1r, Ba, Pm	−0.17 mm

Clinical Reliability

• Vertical skeletal classification (FMA): Cohen's κ = 0.78 (substantial agreement)
• 20/25 landmarks improve with priors; 1 degrades (Basion, lowest confidence tier)

Usage

import onnxruntime as ort
import numpy as np
import cv2

# Load Stage 1 model
sess = ort.InferenceSession("v4_stage1.onnx")

# Prepare input (28 channels: 3 RGB + 25 attention maps from Stage 0)
image = cv2.imread("cephalogram.jpg")
image_512 = cv2.resize(image, (512, 512))
rgb = image_512.astype(np.float32) / 255.0  # (512, 512, 3)
rgb = np.transpose(rgb, (2, 0, 1))  # (3, 512, 512)

# attention_maps shape: (25, 512, 512) from Stage 0 pipeline
# (See Stage 0 inference code for generating these)
input_28ch = np.concatenate([rgb, attention_maps], axis=0)  # (28, 512, 512)
input_tensor = input_28ch[np.newaxis]  # (1, 28, 512, 512)

# Run inference
input_name = sess.get_inputs()[0].name
heatmaps = sess.run(None, {input_name: input_tensor})[0]  # (1, 25, 256, 256)

# Decode landmarks from heatmap peaks
landmarks = []
for i in range(25):
    hm = heatmaps[0, i]
    y, x = np.unravel_index(np.argmax(hm), hm.shape)
    # Scale from heatmap (256) to image (512) coordinates
    landmarks.append((x * 2, y * 2))

Training Data

Aggregated from three public sources (1,502 total images):

Source	Images	Landmarks	Scanner(s)
ISBI 2015	400	19	Soredex CRANEX
Aariz/CEPHA29	1,000	29	7+ device types
DentalCepha	102	19	Mixed

Split: 1,201 train / 150 validation / 151 test (stratified by source, seed=42).

Citation

@article{mohapatra2025cephtrace,
  title={CephTrace: Anatomy-Guided Spatial Attention Priors for
         Sub-Millimeter Cephalometric Landmark Detection},
  author={Mohapatra, Sidhartha and Mohanty, Pallavi},
  journal={arXiv preprint arXiv:2605.03358},
  year={2025},
  url={https://arxiv.org/abs/2605.03358}
}

Links

Resource	URL
Paper	arXiv:2605.03358
Code	github.com/sidwiz/cephtrace-research
Data & Weights	Zenodo DOI 10.5281/zenodo.20032162
Website	cephtrace.com

License

This work is licensed under CC BY-NC-SA 4.0. Commercial use requires a separate license — contact research@cephtrace.com.

Three U.S. provisional patent applications are pending (#64/037,246; #64/037,252; #64/039,042).

Limitations

• Trained on 2D lateral cephalograms only; not validated on 3D CBCT or PA cephalograms.
• Phase 0A requires visible soft-tissue profile; severely overexposed or cropped images may degrade.
• Basion (Ba) accuracy degrades slightly with priors due to low Phase 0E confidence (σ=22).
• Cross-source generalization without priors is poor (22–37 mm MRE in LOSO experiments); Phase 0's anatomical analysis provides scanner-invariant features.