Frequently Asked Questions

Bone age is the biological maturity of the skeleton, assessed from the ossification centers on a hand-wrist radiograph. It is used to evaluate growth disorders, endocrine follow-up, and forensic age estimation.

After you pick the age group (under / over 18), the uploaded hand-wrist radiograph is pre-processed (screenshot clean-up, orientation correction, wrist-preserving crop) and the matching specialist deep-learning model estimates bone age, a likely range and gender. For minors, two independent model systems run in parallel and their outputs are combined. The whole process takes seconds, and the result comes with a heatmap showing where the model looked.

Age of Bones (KemikYaşı in Turkish) is a product of Medikalai Yazılım A.Ş. (Kütahya, Turkey). The model architecture was designed by Doç. Dr. Cüneyt Özdemir (Siirt Üniversitesi) and the medical content is reviewed by Dr. Öğr. Üyesi Mehmet Ali Gedik (Radiologist, Kütahya Şehir Hastanesi). The approach builds on the team's peer-reviewed studies on the same dataset: 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.380601) · 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.1294139) Full citations and DOI links are in the References section of the Methodology page.

The models were developed on the hand-wrist radiograph dataset collected from four hospitals in Kütahya, Turkey (Kütahya Child Radiology Dataset, KCRD), introduced in Traitement du Signal in 2021. It contains radiographs from the Turkish population taken on different devices and exposure settings, so its fit to the local population is closer than models trained only on foreign sets such as RSNA. The dataset is not public; images are anonymised.

You can; the system automatically removes viewer interface elements (frames, menus, text areas) from screenshots, rotates a portrait upload so the fingers point upwards and keeps the wrist inside the crop. Screenshots still lose resolution and may be cropped at the edges, so upload the original PNG or DICOM downloaded from e-Nabız whenever possible. The 'Model input' tab on the result screen shows the image the model actually received.

Warnings appear when a screenshot is detected, the radiograph was uploaded in portrait (and auto-rotated), the resolution is low, the wrist is not visible or the age-group decision is uncertain. A warning does not invalidate the result; it tells you the model ran under less reliable conditions. Re-upload the original file if you can, make sure the whole hand and wrist are visible, and check that the heatmap focuses on the carpal bones and epiphyses.

From the result screen you can re-run the same radiograph in the other age group (under / over 18) without using your quota; each record allows 3 re-runs. Your choice is binding: if you do not choose, the gate model decides automatically at the 216-month threshold, but it can be wrong near the threshold. If the two results differ widely, physician assessment is decisive.

JPG, JPEG, PNG, BMP, HEIC, WEBP images and DICOM (.dcm) files are supported. For best results, upload a clear, correctly positioned PA hand-wrist radiograph, and prefer the original file over a screenshot.

Sign in to e-Nabız (Turkey's national health record) with your e-Devlet password, open the hand-wrist radiograph under Sağlık Geçmişim → Radyolojik Görüntülerim, and save it as PNG (or DICOM if offered) with the viewer's download button. Do not take a screenshot: resolution drops and interface text ends up in the image. The step-by-step walkthrough is in the e-Nabız section of the Guide.

No. This is a clinical decision-support tool. The result is not a standalone diagnosis and must be reviewed by a physician or radiology specialist.

We publish every model's internal validation score as a table on the Methodology page: the gate model that separates under/over 18 scores 96.3% accuracy (AUC 0.994), the under-18 band experts sit at 6.9-11.1 months mean absolute error, and the over-18 model errs at year scale (3.9 years). That last figure is a physiological limit, not a software flaw: growth plates close around 18-19 years, after which the radiograph carries almost no age-discriminating signal. Internal validation always flatters a model, so we recommend institutions run an acceptance test on their own samples.

The model learns age as a probability distribution rather than a single number; the likely range is the 10th–90th percentile of that distribution (e.g. 147 months, likely range 138–156). If the two sub-systems disagree, the range widens. It is the model's own uncertainty on that image; the clinical margin of error is 7–11 months depending on the age band, and the range does not replace it. A wide range means the model was unsure; a narrow range does not guarantee the result is correct.

A colour overlay produced with Grad-CAM that shows the regions the model looked at most when deciding. The expected regions are the carpal bones and the epiphyses; if the map concentrates on irrelevant areas such as labels, edges or a cast, treat the result with caution. The 'Model input' tab shows the cropped, correctly oriented image the model actually received. The heatmap shows where the model looked; it does not prove the result is right.

No. The heatmap is a 16×16 sensitivity map from the last convolutional layer: hot regions influenced the final decision most; a cold region does not mean 'ignored'. The model is not a template-matching system but a deep-learning ensemble that processes the whole image; phalangeal and carpal information is used in intermediate layers. Between roughly 10 and 14 years the distal radius/ulna epiphyses genuinely are the most discriminative region (the TW3-RUS method also weights that area). The 'Attention distribution' box on the heatmap tab shows the approximate share of the map across phalanges / metacarpals / carpal-wrist regions.

No. Pre-processing standardises the image so the fingers point left; if they point right, the image is flipped horizontally and a badge says so. The model was trained with random horizontal flips and averages a flipped copy at inference, so it never learned left/right laterality and the flip does not change the result. A 180° rotation was deliberately not used because it would move the thumb to a side the model did not see in training.

Yes. In the review screen ('Zoom & measure') enable the 'Hand region' tool, drag corner to corner over the hand-wrist region on the original radiograph and choose 'Re-run with this region'. The analysis runs again in the same age group on that region only; it does not use your quota but counts against the per-record re-run allowance (3). The marked region is stored with the record and shown as a badge on the 'Model input' tab.

With the 'Request expert review' button on the result screen you can ask a specialist radiologist to review your radiograph. The review is done personally by Dr. Mehmet Ali Gedik, a radiology specialist at Kütahya City Hospital and a Hacettepe University Faculty of Medicine graduate: he examines the image, the model input and the heatmap, then writes a bone-age opinion and notes. You are notified by e-mail within 1 business day and the opinion appears on the result screen next to the model estimate, with the difference. This is a paid service: only accounts with an active (Pro) subscription can open a request; at most 5 per subscriber in 30 days and 3 open at a time. The expert opinion is also decision support; diagnosis and treatment rest with the treating physician.

On the My Analyses page, tick the 'Compare' box on two records. The panel that opens at the top shows the change in bone age, the growth rate between the two acquisitions (relative to calendar time) and a projected time to skeletal maturity. The projection assumes linear growth; the clinical decision belongs to the physician. Growth rate is more meaningful on records where chronological age was entered.

In the review screen ('Zoom & measure') you can measure the distance between two points. Measurements are in pixels by default; if the DICOM file carries pixel-spacing (PixelSpacing) metadata, the same measurement is also given in mm. For JPG/PNG the physical scale is unknown, so only pixels are shown. The mm conversion applies to the original radiograph only; measurements are added to the printed report.

The image, result and audit record are stored securely in the database and are accessible only to authorized administrators. Model files are never served publicly. See the Privacy page.

Use the 'Delete my account' button at the bottom of the My Account page; you will be asked for your password to confirm (leave it empty if you signed up with Google). Deletion is permanent under KVKK: your analyses, images and personal data are removed and cannot be restored. Before deleting, you can download all your records as JSON with 'Download my data' on the same page. You can also change your password there.

No. Membership is required. When you register for free, you receive one trial analysis.

Everyone gets 1 free analysis. After that there are two routes: buy a single analysis or a credit pack (from $5; credits never expire), or subscribe with a monthly quota (Physician: 50 analyses a month, Clinic: 250; yearly billing gives 2 months free). The expert opinion is an optional add-on charged only when you request it. Current amounts are on the Pricing page; all prices include tax.

Pick the Physician or Clinic plan and a monthly or yearly period on the Pricing page and complete the payment with your card. Your account is activated as soon as the payment is confirmed — no waiting. Single analyses and packs are bought on the same page without a subscription.

Yes. You are charged automatically every month on the monthly plan, or every year on the yearly plan. You can stop it yourself at any time with the 'Cancel subscription' button on the Subscription page — no charge after cancellation, and you keep access until the end of the period you already paid for.

Yes. Use 'Print Result' on the result screen to print or save as PDF.

Yes, and we value it highly. On the result screen you can tell us whether you were satisfied and, if you know them, enter two values separately: the bone age assessed by the physician and the patient's chronological (calendar) age in years + months. Accuracy is measured against the physician's bone age; chronological age is used only to compute the deviation. If you learn the real age later, you can add it from the 'My Analyses' page. This is what lets us measure and improve the model's accuracy.

Click 'Forgot password' on the login page and enter your e-mail; a time-limited reset link is sent to you. Set your new password from that link. If you signed up with Google there is no password — simply sign in with Google. While signed in, you can also change your password on the My Account page.