Model Methodology, Validation and Limits

This page documents how the system produces a result, what it was measured at, and where it should not be trusted. Gender is never requested from the user — it is inferred from the image.

1. Routing (user first, model as fallback)

The under/over 18 branch decides which specialist model runs and in which unit the result is reported, so it is asked to the user first: a declared branch is binding. If the user does not answer, a dedicated gate model classifies the radiograph as under or over 18 years (threshold: 216 months) and routing follows that. The gate is reliable but not infallible near the threshold — this is exactly why the declaration takes precedence.

1. Minor specialist: bone age in months. Two independent systems run in parallel; in each, a coarse router first makes a rough estimate, then the age-band expert covering that estimate produces the final value (divide and conquer). The two systems' outputs are combined.
2. Adult specialist: age in years, from a single hybrid model.

2. Architecture

Multi-task convolutional networks (EfficientNet-B3 and ConvNeXt V2-Tiny backbones, 512 px input; 384 px for the gate). Age is learned as a probability distribution rather than a single number, and the prediction is the expectation of that distribution. Horizontal-flip test-time augmentation is applied to stabilise the output. All models are image-only: no clinical data, patient identity or user-supplied gender enters the inference. The architecture was designed by Assoc. Prof. Cüneyt Özdemir (Siirt University) together with Dr. Mehmet Ali Gedik (Kütahya City Hospital); it builds on our earlier peer-reviewed studies on the same dataset, listed under Publications below.

3. Image pre-processing

The models expect a square 512 px input, whereas real uploads are often phone screenshots, portrait-oriented files or images with viewer interface text around the edges. Three steps therefore run before inference:

1. Screenshot UI clean-up: frames, menus and text areas belonging to the viewer application are detected and discarded, so that only the radiograph itself reaches the model.
2. Orientation normalisation: a portrait or sideways upload is rotated so that the fingers point upwards, matching the orientation of the training data.
3. Wrist preservation: cropping is constrained so that the carpal bones and the distal radius/ulna are never cut off — these are the most informative regions for bone age.

The image the model actually receives is shown in the 'Model input' tab on the result screen, and a quality warning is raised when a step could not be completed (screenshot detected, low resolution, wrist not visible). After these steps went live, the mean error on the voluntary feedback set fell from 119 to 55 months; that set is small and self-selected, so the figure indicates a trend rather than an accuracy claim.

4. Explainability and uncertainty

A number alone cannot be audited. Every result therefore comes with two additional outputs that make the model's behaviour visible — and with clear limits on what they do not mean.

• Heatmap (Grad-CAM). For each result a Grad-CAM map is computed from the final convolutional layer; it highlights the regions that contributed most to the output. It shows where the model looked, not why it is right. Activation concentrated outside the expected regions (carpal bones, epiphyses) — on a label, an edge or a cast — is a sign that the estimate may be unreliable.
• Likely range (10th–90th percentile). Because age is learned as a label distribution (DLDL), the 10th and 90th percentiles of that distribution are reported as the likely range. When the two minor sub-systems disagree, the range is widened to cover both estimates. A wide range means the model was unsure on that image.
• Limits. The range is the model's own uncertainty, not a statistical confidence interval and not the clinical margin of error. A narrow range does not guarantee a correct result: if the model is systematically wrong (atypical anatomy, an artifact it has never seen), the range can be narrow in the wrong place. For clinical purposes the band experts' MAE (7–11 months) remains the margin of error, and the heatmap is a plausibility check, not evidence.

5. Internal validation results

Each model's own validation score, read directly from the trained checkpoints. MAE = mean absolute error.

6. Why adult accuracy is lower

The adult model's error is measured in years, not months, and this is a physiological limit rather than a software shortcoming: hand and wrist growth plates close around 18-19 years, after which the radiograph carries very little age-discriminating signal. No model can recover information the image does not contain. Treat over-18 output as a coarse estimate, and do not use it where a legally decisive age determination is required.

7. Validity scope

Expected performance applies to well-positioned PA hand-wrist radiographs resembling the training and validation data. Atypical anatomy, skeletal dysplasia, surgical implants, casts, severe artifacts, burned-in text, cropped images or incorrect positioning may all degrade the result — sometimes without any visible sign that the estimate is wrong.

8. Intended use

This is a clinical decision-support tool, not a diagnostic device, and it has no regulatory clearance as one. Every result must be reviewed by a physician or radiology specialist together with clinical findings. We recommend that institutions run an acceptance test on their own sample before putting it into routine use.

9. Continuous measurement

Users can report the patient's real age after an analysis. These reports let us measure error against ground truth in actual use, and we intend to publish those figures on this page as enough data accumulates. Note they will be voluntary — and therefore biased — so the trend over time matters more than the absolute value.

References

The clinical concept of skeletal maturity assessment this tool automates rests on the following standard works. They describe the reference methods, not this system's training data.

1. Greulich WW, Pyle SI. Radiographic Atlas of Skeletal Development of the Hand and Wrist. 2nd ed. Stanford University Press, 1959.
2. Tanner JM, Healy MJR, Goldstein H, Cameron N. Assessment of Skeletal Maturity and Prediction of Adult Height (TW3 Method). 3rd ed. WB Saunders, 2001.
3. Halabi SS et al. The RSNA Pediatric Bone Age Machine Learning Challenge. Radiology. 2019;290(2):498-503.

Our publications on this dataset

The models on this site were developed on the Kütahya hand-wrist radiograph dataset. The peer-reviewed studies below were carried out by our team on that same dataset and form the basis of the modelling approach described above.

1. Özdemir C, Gedik MA, Kaya Y. Age Estimation from Left-Hand Radiographs with Deep Learning Methods. Traitement du Signal. 2021;38(6):1565-1574. doi:10.18280/ts.380601Introduced the Kütahya hand-wrist radiograph dataset (KCRD) collected from four hospitals; reported an MAE of 4.3 months for bone age with a modified InceptionV3.
2. Özdemir C, Gedik MA, Küçüker H, Kaya Y. Gender Identification from Left Hand-Wrist X-ray Images with a Hybrid Deep Learning Method. Konya Journal of Engineering Sciences. 2023;11(4):1118-1133. doi:10.36306/konjes.1294139Gender inference from the image on 13,935 radiographs aged 2-79 with a hybrid InceptionV3 + DenseNet201 model; the approach behind this site's gender output.

Questions about the methodology? Contact us · See also the FAQ.