What's New

A release shaped by user feedback: expert radiologist review (human in the loop), manual hand-region marking, attention distribution and pre-processing badges on the heatmap, and a pilot analysis quota for physicians. The scrolling bug in the review screen is fixed.

• Expert radiologist review

  Open a request from the result screen; a specialist radiologist reviews the radiograph, model input and heatmap, then writes a bone-age opinion and notes. You are notified by e-mail; the opinion is shown next to the model estimate with the difference and printed in the report. A paid service: available to active subscribers only, at most 5 requests in 30 days.

  FAQ: expert review →
• Manual hand-region marking

  If the automatic crop is wrong, draw the hand-wrist region with the 'Hand region' tool in the review screen; the analysis re-runs on that region only. No quota is used; it counts against the per-record re-run allowance and the region is stored with the record.

• Attention distribution and pre-processing badges

  The heatmap tab now shows the approximate percentage split of the map across phalanges / metacarpals / carpal-wrist regions and explains why a cold region does not mean 'ignored'. The 'Model input' tab labels steps such as horizontal flip, rotation and left-hand isolation, and explains why the flip does not affect the result.

  FAQ: heatmap and flipping →
• Pilot analysis quota for physicians

  Paediatric endocrinologists and radiologists can be granted 5–10 free analyses per account (the remaining allowance is shown on the Account page). The goal is to test the system in outpatient practice and collect feedback; use the contact page to request access.

  My account →
• Review screen scrolling fix

  In the full-screen review window the measurement list and footer note could not be reached on small screens; the window now accounts for the mobile address bar and scrolls vertically when content overflows.

The result is no longer a single number: a likely range, a heatmap showing where the model looked, a review screen with measurements, and a temporal comparison between two analyses. Uploading and account management were also reworked.

• Likely age range

  Next to the estimate, the result now shows a range derived from the model's age distribution (e.g. 147 months, likely range 138–156). The range reflects the model's 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.

  How is it computed? →
• Heatmap (Grad-CAM)

  The 'Heatmap' tab on the result screen shows the regions the model looked at most when deciding. The 'Model input' tab shows the cropped, correctly oriented input — the image the model actually received.

• Review screen: zoom and measure

  Zoom the radiograph up to 6×, pan, and measure the distance between two points. Measurements are in pixels; for DICOM files they are also given in mm using pixel-spacing metadata. Measurements are added to the printed report.

• Temporal comparison

  Select two analyses in My Analyses to get the change in bone age, growth rate and a projected time to skeletal maturity. The projection assumes linear growth; physician judgement prevails.

  My Analyses →
• Image quality warnings

  The result screen now warns explicitly about screenshots, portrait uploads (auto-rotated), low resolution, a wrist that is not visible, and an uncertain age-group decision.

• Re-run in the other age group

  If the age-group decision looks doubtful, re-run the same radiograph in the other group (under / over 18) without using your quota. Three re-runs are allowed per record.

• Upload improvements

  HEIC, WEBP and DICOM support; a progress bar and a cancel button. The result survives a page refresh, every past record can be printed as a report, and chronological age is entered as years + months.

• My Account

  Password change and a 'Forgot password' flow were added. You can download your data as JSON and permanently delete your account under KVKK (Turkish data-protection law).

  My Account →
• e-Nabız guide

  Step-by-step instructions for downloading a radiograph from e-Nabız (Turkey's national health record) as an original file were added to the guide. An original file gives a more reliable result than a screenshot.

  Open the guide →
• Scientific basis and team

  The Methodology page now lists the two peer-reviewed studies published on the Kütahya dataset the models were developed on (with DOI links) and the researcher responsible for the model architecture. The FAQ and Guide were updated with everything in this release; the site now publishes llms.txt for AI assistants.

  Publications →
• Two separate ages in feedback

  The feedback form now asks separately for the physician's bone-age assessment and the patient's chronological age. Accuracy is measured against the physician's bone-age value; chronological age is used only for the deviation.

Image pre-processing: support for phone screenshots and portrait uploads.

• Screenshots and portrait uploads

  Phone screenshots are stripped of interface text, portrait uploads are rotated to the correct orientation, and the wrist is preserved when cropping. On the voluntary feedback set the mean error dropped from 119 to 55 months; the set is small, so this indicates a trend rather than a definitive accuracy figure.

  Pre-processing steps →