Bone age assessment from hand radiographs using the ianpan/bone-age deep learning model — a ConvNeXtV2-tiny 3-model ensemble trained on the RSNA Pediatric Bone Age 2017 dataset (14,036 left-hand PA radiographs, MAE 4.16 months).
Experimental — not clinically validated. Built as an engineering exercise to explore medical imaging inference. Only a radiologist's report has diagnostic value. Do not use for medical decisions.
The original Deeplasia paper was the first choice, but the authors never published the pretrained checkpoints. ianpan/bone-age is a public alternative trained on the same RSNA dataset with comparable performance (MAE 4.16 vs. Deeplasia's 3.87 months) and includes native DICOM support.
- Left hand, PA view, fingers up — matches training distribution
- All 5 fingers visible, wrist included, forearm cropped
- Manual crop recommended over the model's auto-crop (which can clip the thumb on rotated images)
- Screen photos work but degrade accuracy vs. original DICOM (gamma, glare, JPEG artifacts)
pip install -r requirements.txtpython bone-age.py \
--patient "Patient Example" \
--dob 2023-07-17 \
--sex female \
--image example.tif| Flag | Default | Description |
|---|---|---|
--patient |
Patient Example |
Patient name or identifier |
--dob |
2023-07-17 |
Date of birth (YYYY-MM-DD) |
--sex |
female |
male or female |
--exam-date |
today | Exam date (YYYY-MM-DD) |
--image |
example.tif |
Path to hand X-ray image (PNG/TIFF) |
The script preprocesses the image via histogram matching against the model's reference, runs inference, saves {image}_result.md to disk, and prints the report to the terminal.
| Detail | Value |
|---|---|
| Architecture | ConvNeXtV2-tiny 3-model ensemble |
| Parameters | 84.1M |
| Training data | RSNA Pediatric Bone Age 2017 |
| Training samples | 14,036 left-hand PA radiographs |
| MAE | 4.16 months (RSNA test set) |
| Input | 1-channel grayscale + sex |
| Preprocessing | Histogram matching |
- Bone age ≠ chronological age is normal in healthy children (±12 months variation)
- The model is sensitive to input quality — bad crop or skipped histogram matching can shift predictions by 10+ months
- This pipeline uses CPU inference (sufficient for single images)
MIT