The study
A systematic review published in the Journal of Orthopaedic Surgery and Research on 4 November 2026 by Yin and colleagues from Chengdu University of Traditional Chinese Medicine, China. It was registered in advance (PROSPERO CRD420261417898) and searched four databases up to 17 June 2026.
It included 14 reports, published from 2022 to 2026, testing artificial intelligence (AI) that automatically measures the hallux valgus angle (HVA) and intermetatarsal angle (IMA) on plain foot radiographs. The authors didn't pool the results in a meta-analysis, because studies defined and reported their measurements differently and some shared patients.
The software
Commercial software: four studies independently evaluated commercial tools. Three tested BoneMetrics by Gleamer, and one tested Milvue's TechCare Bones.
Smartphone app: one study tested an AI smartphone app.
Research models: the rest were models built in-house. Only two of these were tested on data from outside the hospital that developed them.
Some reports overlap. Three come from one hospital and may share patients, and two use the same 92 patients. The 14 reports aren't 14 independent samples.
Reading the numbers
Mean absolute error (MAE): the average gap, in degrees, between the AI's angle and the human reference measurement, ignoring whether the AI was over or under.
Intraclass correlation coefficient (ICC): agreement between AI and human raters, from 0 to 1. Above 0.9 is usually called excellent.
What they found
Average HVA error ranged from 0.69° to 4.14° across 11 reports, and average IMA error from 0.57° to 3.22° across 10. Those ranges run from the most accurate single study to the least accurate, and the review didn't combine them into one figure.
Agreement with human raters was generally high. ICCs for HVA ran from 0.82 to above 0.99 in the studies reporting them.
Some individual measurements were well off. In one study of 386 radiographs, HVA was more than 5° out in 8 and IMA more than 3° out in 18. On an external dataset of 121 feet, 8 HVA measurements were more than 5° out. Software sometimes failed to produce a measurement (3 of 178 in one study) or needed manual correction (7% and 3% of cases in another).
Measurement time dropped by about 72%, from 288 to 80 seconds, in one simulated workflow with a single junior radiologist.
In one study, AI measurements led to the same treatment category as one surgeon in 67 of 84 cases. No study looked at patient outcomes.
Other angles, including the distal metatarsal articular angle and sesamoid position, weren't analysed.
Quality
Risk of bias was rated with an adapted QUADAS-2 tool, which the authors note isn't validated for this type of study. Patient selection was high risk in 4 reports and unclear in 5, and the reference standard was unclear in 10. Only two reports were low risk in every domain, both commercial-software evaluations.
Funding and conflicts
Funded by the Sichuan Provincial Administration of Traditional Chinese Medicine and a Chinese National Health Commission education centre. The review authors declared no competing interests. The review doesn't report whether the included studies had vendor funding or vendor-employed authors, so we couldn't check that for the commercial software evaluations.
What it means for your practice
An error of a few degrees can matter near a cut-off. Commonly used upper limits of normal sit around 15° for HVA and 9° for IMA. The review doesn't discuss thresholds, and an error of 3 to 4 degrees could move a foot between severity grades.
The authors state that low average errors don't mean every individual measurement was acceptable.
External testing was limited and severe deformity was under-represented. Most data were retrospective and from single centres, and how well the tools handle postoperative feet varied by system.
Plain radiograph angles can't capture metatarsal pronation or 3D sesamoid position, whether a human or AI measures them.
Bottom line
Across 14 reports, AI tools measured HVA and IMA on plain films with average errors from under 1° to about 4° and generally high agreement with human raters, mostly in single-centre retrospective data. It seems like these tools could save time on routine measurement. The individual outliers and the limited data on severe and postoperative feet mean their outputs aren't yet reliable enough to go unchecked.