This is a logistic regression classifier trained on 88,064-dimensional embeddings extracted by Google's CXR Foundation model. It converts CXR Foundation — a chest X-ray foundation model — into a musculoskeletal fracture detector, demonstrating that CXR Foundation's learned representations transfer beyond its original thoracic training domain.

Architecture: sklearn.linear_model.LogisticRegression with StandardScaler preprocessing and temperature calibration (T=3.49)
Input: 88,064-dim CXR Foundation embedding vector
Output: Fracture probability (0–1)
File format: joblib dict with keys model, scaler, temperature
Size: ~700 KB

Training Data

Trained on FracAtlas — 4,083 musculoskeletal X-rays (717 fractures, 3,366 normal) across four body regions: hand, leg, hip, and shoulder.

Performance

5-fold stratified cross-validation with bootstrap confidence intervals (1,000 iterations):

Region	n_train	AUC	95% CI
Hand	1,510	0.850	[0.829, 0.873]
Leg	2,237	0.888	[0.859, 0.914]
Hip	179	0.864	[0.764, 0.953]*
Shoulder	98	0.848	[0.664, 0.972]*
Overall	4,024	0.882	--

* Wide CI due to small sample size — statistically unreliable.

Data Efficiency

The probe achieves strong performance with remarkably few labeled examples, demonstrating that CXR Foundation embeddings carry transferable signal even for out-of-domain musculoskeletal anatomy:

Training examples	AUC
10	0.555
25	0.578
50	0.607
100	0.683
250	0.785
500	0.820
4,024 (full)	0.882

0.820 AUC at just 500 examples suggests viable deployment in data-scarce clinical settings.

Usage

python

import joblib
import numpy as np

# Load the probe
probe = joblib.load("fracture_probe.joblib")
model = probe["model"]
scaler = probe["scaler"]
temperature = probe["temperature"]

# Given a CXR Foundation embedding vector
# embedding = extract_embedding(image)  # shape: (88064,)

embedding_scaled = scaler.transform(embedding.reshape(1, -1))
logit = model.decision_function(embedding_scaled)
probability = 1 / (1 + np.exp(-logit / temperature))

Intended Use

Primary: Fracture screening triage in resource-limited settings without radiologist coverage
Deployment target: NVIDIA Jetson Orin Nano (8 GB) or similar edge devices
NOT intended for: Standalone clinical diagnosis. This is a screening aid — all positive findings require radiologist confirmation.

Limitations

Trained only on FracAtlas (hand, leg, hip, shoulder). Performance on other body regions is unknown.
Hip and shoulder results have wide confidence intervals due to small sample sizes (179 and 98 training examples respectively).
Requires CXR Foundation for embedding extraction, which expects chest X-ray-format input images.
Not validated in a prospective clinical trial.

Integrity

SHA-256 checksum for fracture_probe.joblib:

5ab04c187564f84cd26dfac5467f2949094f1045efc9b59d2d97260ec3f0dad6

A .sha256 sidecar file is included in this repository for automated verification.

Citation

If you use this model, please cite:

@misc{edgefracture2026,
  title={EdgeFracture: CXR Foundation Fracture Probe for Edge Deployment},
  author={Ibrahim Dagib},
  year={2026},
  url={https://huggingface.co/ibdagib/edgefracture-cxr-fracture-probe}
}

Part of EdgeFracture

This probe is one component of the EdgeFracture pipeline — a two-model fracture triage system combining CXR Foundation embeddings with MedGemma clinical reasoning, built for the Google MedGemma Impact Challenge.