# Subtalar Arthroereisis with Calcaneus Stop Screws—Can the Angles on Pre- and Post-Surgical X-Ray Images Be Reliably Measured by Artificial Intelligence?

## TL;DR

Artificial intelligence can reliably measure certain foot angles in X-rays but struggles with others, especially in lateral views, suggesting manual measurements are still needed for accurate diagnosis.

## Contribution

The study evaluates AI's ability to measure radiographic angles in children with flat feet, revealing its strengths and limitations in clinical settings.

## Key findings

- AI accurately measured angles in the dorsoplantar plane but showed significant deviations in lateral view measurements.
- AI performed similarly on pre- and post-operative X-rays and was not affected by calcaneus stop screws.
- AI measured angles on incorrectly taken radiographs as often as on correctly taken ones.

## Abstract

What are the main findings?
AI was proven to deliver excellent results for angles in dorsoplantar plane, while in lateral view, AI either showed significant deviations from manual measurements or could only measure angles on a small proportion of X-rays.AI neither measured angles on preoperative X-rays significantly more often nor significantly better than on postoperative X-rays.AI measured angles on incorrectly taken radiographs at least as often as on correctly taken radiographs.

AI was proven to deliver excellent results for angles in dorsoplantar plane, while in lateral view, AI either showed significant deviations from manual measurements or could only measure angles on a small proportion of X-rays.

AI neither measured angles on preoperative X-rays significantly more often nor significantly better than on postoperative X-rays.

AI measured angles on incorrectly taken radiographs at least as often as on correctly taken radiographs.

What are the implications of the main findings?
AI still requires training in measuring angles on radiographs of children with severe flat feet on the lateral plane.AI was not disturbed by calcaneus stop screws.AI did not exclude incorrectly taken radiographs from angle measurements.

AI still requires training in measuring angles on radiographs of children with severe flat feet on the lateral plane.

AI was not disturbed by calcaneus stop screws.

AI did not exclude incorrectly taken radiographs from angle measurements.

Background/Objectives: Flexible symptomatic flat foot in children can be surgically treated with calcaneus stop screws. This raises the question of whether pre- and postoperative radiographs (X-ray) can be analyzed in two planes using AI. Methods: In this monocentric retrospective study, angle measurements generated by Bone Metrics AI (Gleamer) were compared with manual measurements using Centricity™ (GE Healthcare). A total of 659 X-rays from 124 operated feet (2014–2024) were available, of which 422 were analyzable by AI and 299 met defined quality criteria. Bland–Altman plots were used to assess agreement. Linear and logistic regression analysis examined the influence of age, gender, accessory navicular bone, additional foot pathologies, and flat foot severity on comparability of the measurement methods and measurability by the AI. Finally, radiographs meeting and missing quality criteria were compared. Results: AI measurements were comparable to manual measurements for calcaneus inclination, hallux valgus, 1st–2nd and 1st–5th metatarsal angle both pre- and post-operatively. For the talus-1st metatarsal and medial arch angles, AI results differed significantly (p < 0.001 and p ≤ 0.013) from manual measurement. AI generated talus-1st metatarsal angle was measured larger by 6.14°, 95% [−7.14; −5.14] pre-operatively and 2.80°, 95% [−3.79; −1.81] post-operatively. Medial arch angle was smaller by 1.63° pre-operatively, 95% [1.03; 2.23] and 0.52° post-operatively, 95% CI [0.11; 0.93] with AI. Post-operative measurability was not significantly lower than pre-operative. AI measured angles on incorrectly taken radiographs as often or more often than on correctly taken ones. Discussion: Screw implantation did not negatively impair measurability or AI accuracy. However, age, gender, and flat foot severity influenced AI performance. Bad radiograph quality did not affect AI measurability negatively, indicating that AI cannot yet distinguish between X-rays suitable and unsuitable for angle measurements. Conclusions: Manual measurements are still indispensable in the diagnosis of children’s flat feet. In the future, continuous training of the AI is expected to bring it into line with manually measured radiological values.