Arthroscopy - 2026-09-19 - Journal Article
An Artificial Intelligence-Based Preoperative Planning System for High Tibial Osteotomy Significantly Enhances Planning Speed With Comparable Accuracy to Surgeons.
Li S, Li Z, Zhang H, Chen Y, Miao Z, Li L, Qin B, Meng X, Qian W, Lu Q, Liu P
Topics
Key Takeaway
An AI-based planning system (OAPS) completed OWHTO preoperative alignment measurements in 9.27 seconds versus 4.30 minutes for manual planning, with mean absolute HKA error of 0.47° compared to attending surgeons.
Summary Depth
Choose how much analysis to show on this article page.
Summary
This study developed and validated a UNet CNN-based system (OAPS) detecting 28 anatomical landmarks on standing hip-to-ankle radiographs to automate OWHTO preoperative planning. On the test set, OAPS achieved mean absolute errors of 0.47° for HKA and 0.43° for correction angle with no significant difference from attending surgeons (P>.05). External validation across two institutions confirmed OAPS-assisted junior residents matched attending-level accuracy and were significantly faster than both manual and semiautomatic methods.
Key Limitation
The study measures planning accuracy against surgeon measurements rather than against postoperative radiographic alignment or clinical outcomes, leaving the downstream surgical impact of AI-generated plans unvalidated.
Original Abstract
PURPOSE
To develop an artificial intelligence-based system capable of measuring preoperative alignment parameters, calculating correction angles, and visualizing outcomes for open wedge high tibial osteotomy (OWHTO) and to validate the accuracy and efficiency of the system in clinical cases.
METHODS
Between January and December 2023, standing hip-to-ankle radiographs from patients with varus knee osteoarthritis were retrospectively collected as the development cohort. The OWHTO auxiliary planning system (OAPS) was developed using a UNet convolutional neural network to detect 28 anatomical landmarks. The system was validated against manual measurements by attending surgeons and further evaluated through an external validation cohort across 2 institutions to compare accuracy and efficiency among junior residents using manual, semiautomatic, and OAPS-assisted methods.
RESULTS
The study included 346 knees for development and an external validation cohort of 110 knees. On the test set, the OAPS achieved a mean radial error of 2.703 pixels and a successful detection rate at the 6-pixel threshold of 0.951. For hip-knee-ankle and correction angles, the mean absolute errors were 0.47° (95% CI: 0.39°-0.56°) and 0.43° (95% CI: 0.36°-0.51°), respectively. No significant differences were found between OAPS and attending surgeons across all alignment parameters (P > .05). OAPS processed each image in 9.27 seconds, which was significantly faster than the 4.30 minutes required for manual planning on the test set (P < .001). External validation confirmed that OAPS-assisted resident measurements matched attending-level accuracy and were significantly faster than manual and semiautomatic methods.
CONCLUSIONS
The artificial intelligence-based OAPS enables accurate, efficient preoperative planning for OWHTO, achieving measurement precision comparable to experienced surgeons while significantly reducing planning time.
CLINICAL RELEVANCE
By automating anatomical measurement and planning, the OAPS reduces interobserver variability and enhances preoperative plan efficiency within OWHTO workflows, streamlining the preparation phase for surgical teams.