About Bone Ager
Updated October 12, 2026
Bone Ager is a free educational bone age calculator run by Joseph Willis of Willis Worxs. It estimates skeletal age from one left-hand radiograph, shows a Greulich and Pyle Brush mean for the chronological age, and does not save the uploaded film.
The result is educational research. Clinical performance and screenshot accuracy have not been established. A clinician must review medical interpretations.
Measured Error
What error has Bone Ager measured?
On the official RSNA validation set of 1,425 images, the mean absolute error is 6.44 months. The median absolute error is 5.00 months, the root mean square error is 8.62 months, and the 90th-percentile absolute error is 13.68 months. Estimates fall within 6 months for 57.9% of those images and within 12 months for 85.9%. A bootstrap 95% confidence interval for the mean absolute error is 6.16–6.76 months. A 90% split-conformal interval, calibrated on 1,260 reserved images, has a radius of 14.03 months, and observed coverage on the official validation set is 90.46%. This is a population-level coverage measurement, not a 90% probability claim for each individual case. External clinical validation, real screenshot validation, automatic laterality and anatomy suitability checks, and superiority over published models remain unestablished. Official case 1401 was inspected during interface testing and was not used to select the checkpoint. Excluding it leaves 1,424 cases with the same 6.44 month mean absolute error.
| Group | Images | Mean Absolute Error |
|---|---|---|
| Female | 652 | 6.65 months |
| Male | 773 | 6.26 months |
| Under 5 | 94 | 7.65 months |
| 5 to 10 | 394 | 7.29 months |
| 10 to 15 | 809 | 5.97 months |
| 15 and Over | 128 | 5.88 months |
The checkpoint was selected on development error. Training, development, and calibration used 10,091, 1,260, and 1,260 images. Partitions are image-disjoint. Patient identifiers are unavailable, so patient-level independence is unverified.
What the Calculator Uses
The image model is a sex-conditioned ConvNeXt V2 Tiny network trained on expert-labeled RSNA radiographs. Chronological age is shown beside the estimate and is excluded from the image model. Reference films shown after an estimate are RSNA training examples. They are separate from Greulich and Pyle atlas plates.
The radiograph is processed in memory for the estimate and is not saved. Do not upload identifying information.
The starting ConvNeXt V2 Tiny weights are used under CC BY-NC 4.0. The RSNA radiographs used for training carry noncommercial terms.