Can ViTPose Automatically Measure HVA and IMA on Preoperative and Postoperative Radiographs, Including Images with Metal Implants?

Internal MAEs were 1.09° for HVA and 0.76° for IMA; external MAEs were 1.33° and 0.81°, respectively.

Research structure

01
Subject
Development: 2,985 images—2,375 for training, 297 for validation and 313 for internal testing. External validation: 300 images; internal postoperative subgroup: 19 images.
02
Methods
A Transformer-based pose-estimation model was developed, evaluated with MAE, ICC(2) and Bland–Altman analysis, and validated at two external institutions.
03
Reported Results
Internal MAEs were 1.09° for HVA and 0.76° for IMA; external MAEs were 1.33° and 0.81°, respectively.

An explanatory schematic separating the study population, methods, and results. It is not an original paper figure.

Participants and Methods

Development: 2,985 images—2,375 training, 297 validation and 313 internal test images. External validation: 300 images; internal postoperative subgroup: 19 images. A Transformer-based pose-estimation model was assessed with MAE, ICC(2) and Bland–Altman analysis and validated at 2 external institutions.

Check points in the original.

Performance on the 300 external validation images should be distinguished from performance in the postoperative subgroup. The internal test set included 19 postoperative images, of which 12 were after hallux valgus surgery. The 18 postoperative images in external validation showed similar osteotomy and screw fixation patterns, so these results do not establish performance across all procedures and fixation devices.

Main Findings

Internal mean absolute errors were 1.09° for HVA and 0.76° for IMA; external errors were 1.33° and 0.81°, respectively. In the postoperative subgroup, errors were 1.09° for HVA and 0.54° for IMA, with ICC(2) values of 0.98 and 0.95. Automatic measurement took 0.49 seconds per image, about 30 times faster than manual measurement.

The point of reading on surgery

The model assists quantitative measurement for surgical planning and follow-up assessment. Improved patient function and fewer repeat operations were not evaluated.

Limitations

Postoperative validation involved 19 images and was preliminary. Further validation across surgical procedures and fixation devices is needed.

Let's put it briefly.

What did this study confirm?

Internal MAEs were 1.09° for HVA and 0.76° for IMA; external MAEs were 1.33° and 0.81°, respectively.

Can it be applied directly to individual patients' treatment decisions?

Postoperative validation involved 19 images and was preliminary. Further validation across surgical procedures and fixation devices is needed.

Keywords

Hallux valgus
Artificial intelligence
Deep learning
AI
hallux valgus
external validation
surgical planning

Abstract · Original text

Sang Gyo Seo, Donghyuk Kwak, Sehun Kim, Min Gyu Kyung, Dong Yeon Lee, Du Hyun Ro, Jahyung Kim. Externally Validated Automated Measurement of Hallux Valgus on Pre- and Postoperative Weight-Bearing Radiographs Using a Transformer-Based Model. Clinics in Orthopedic Surgery. 2026. doi:10.4055/cios26232.

Background

Accurate measurement of the hallux valgus angle (HVA) and intermetatarsal angle (IMA) is essential for surgical planning and postoperative assessment in hallux valgus. However, manual measurement is time-consuming and subject to interobserver variability. Existing deep learning models are limited by reliance on preoperative radiographs and lack of external validation. Therefore, this study aimed to develop and externally validate a fully automated artificial intelligence (AI)–based system for HVA and IMA measurement on weight-bearing radiographs and to preliminarily evaluate its feasibility in postoperative radiographs containing metallic implants.

Methods

A transformer-based pose estimation architecture (ViTPose) was developed for automated HVA and IMA measurement on weight-bearing anteroposterior foot radiographs. The model was developed using 2,985 radiographs, including 217 postoperative radiographs with metallic implants or hardware-related image artifacts. The dataset was divided into training, validation, and internal test sets of 2,375, 297, and 313 radiographs, respectively. Performance was evaluated using internal and external validation cohorts. Accuracy was assessed using mean absolute error (MAE), intraclass correlation coefficient type 2 (ICC2), and Bland- Altman analysis. Subgroup analyses were performed by deformity severity and postoperative status.

Results

In the internal test cohort, MAEs were 1.09° for HVA and 0.76° for IMA, with excellent agreement with expert measurements (ICC2 = 0.99 and 0.95). External validation on 300 independent radiographs from 2 institutions demonstrated comparable performance against a consensus-reviewed reference standard, with MAEs of 1.33° for HVA and 0.81° for IMA. In the postoperative subgroup of the internal test set (n = 19), the model showed MAEs of 1.09° for HVA and 0.54° for IMA, with ICC2 values of 0.98 and 0.95, respectively. Severity classification agreement between predicted and reference measurements was 89.5% for HVA and 78.6% for IMA. Automated measurement required 0.49 seconds per image, approximately 30-fold faster than manual measurement.

Conclusions

This externally validated transformer-based system provided accurate, fully automated measurement of HVA and IMA on weight-bearing radiographs and showed preliminary feasibility in a limited postoperative subgroup. The proposed approach may serve as a practical AI-based measurement-support tool for preoperative assessment. Broader postoperative application requires further validation in larger cohorts encompassing diverse hallux valgus procedures and fixation constructs.

Read abstract at source Terms of use

The original paper.

Externally Validated Automated Measurement of Hallux Valgus on Pre- and Postoperative Weight-Bearing Radiographs Using a Transformer-Based Model

Sang Gyo Seo, Donghyuk Kwak, Sehun Kim, Min Gyu Kyung, Dong Yeon Lee, Du Hyun Ro, Jahyung Kim
Clinics in Orthopedic Surgery · 2026

DOI · 10.4055/cios26232 Based data

Reviewed academic journal articles