Measuring Knee Alignment Following Distal Femur Fracture Surgery: A Novel Deep Learning Approach.

Abstract

Objectives: To evaluate whether a machine learning-based system can accurately measure coronal knee alignment parameters from routine postoperative radiographs following surgical fixation of distal femur fractures.

Methods: Design: Retrospective cohort study.

Setting: Single-center academic orthopedic trauma practice.

Patient selection criteria: Adult patients 50 years of age or older who underwent surgical fixation of low-energy OTA/AO 33A and 33C distal femur fractures (including periprosthetic fractures) between November 2020 and December 2024 were retrospectively identified.

Outcome measures and comparisons: A machine learning-based automated alignment assessment system was trained to identify standard radiographic landmarks, including the femoral intercondylar notch, femoral condyles, tibial spines, and tibial plateaus, and automatically calculate commonly used coronal alignment parameters including joint line convergence angle (JLCA), anatomic femorotibial angle (aFTA), anatomic lateral distal femoral angle (aLDFA), and anatomic medial proximal tibial angle (aMPTA). Automated measurements were compared with manual measurements annotated by a trained observer and validated by an orthopedic surgeon.

Results: A total of 287 postoperative AP knee and distal femur radiographs were obtained as part of routine postoperative care and randomly divided into training and testing datasets for model development and validation. The cohort was 78.0% female with a mean age of 72.3 years (range: 50 to 97 years). Automated measurements demonstrated strong agreement with surgeon-validated measurements. Mean absolute error was 1.1° for JLCA (range: 0.0° to 3.5°), 1.5° for aFTA (range: 0.0° to 4.4°), 1.0° for aLDFA (range: 0.0° to 3.1°), and 2.3° for aMPTA (range: 0.1 to 6.4°). Pearson correlation coefficients ranged from 0.4 to 1.0, and Bland-Altman analysis demonstrated minimal systematic bias, with 95% of differences falling within a clinically acceptable range of ±5°.

Conclusions: Automated alignment assessment utilizing a machine learning-based system using routine postoperative radiographs provided accurate measurements in distal femur fracture patients and may serve as a practical adjunct for longitudinal surveillance in orthopedic trauma practice.

Level of evidence: Level III.