Automatic measurement of knee parameters in three-dimensional computed tomography models: a deep learning-based approach

This study evaluates a deep learning algorithm for automated measurement of knee parameters from three-dimensional (3D) computed tomography (CT) reconstructions.

We analyzed 323 CT scans from patients with normal knee morphology (mean age 39.10 ± 11.34 years; 200 male), randomly split into a model development dataset (Dataset 1) and a hold-out test set (Dataset 2). Initial 3D reconstructions of the femur, distal femoral epiphyseal line, tibia, and proximal tibial epiphyseal line were generated using Materialise Mimics software, followed by manual annotation of 24 anatomical landmarks. A 3D convolutional neural network (CNN) based on the GU2-Net architecture was trained on Dataset 1 to predict landmark probability maps, from which 22 knee joint morphological parameters were calculated. Model performance was assessed on two datasets using intraclass correlation coefficients (ICC), mean absolute differences (MAD), root mean square (RMS) values, paired t-tests, Benjamini-Hochberg false discovery rate (FDR) correction for multiple testing, and two one-sided tests (TOST) for clinical equivalence verification.

The mean absolute difference between predicted and manually annotated landmarks was 1.95 mm in Dataset 1 and 2.04 mm in Dataset 2. The percentage of correct keypoints (PCK) at a 4 mm threshold was 95.83% in Dataset 1 and 94.34% in Dataset 2. Automated measurements showed strong agreement with manual results (ICC = 0.880-0.993, MAD = 0.39-1.46, RMS = 0.46-1.97), with no statistically significant differences after FDR correction (all adjusted p > 0.05). TOST with clinical margins (± 2.0 mm for linear parameters, ± 1.0° for angular parameters) confirmed clinical equivalence for most parameters, with only 5 parameters failing to reach equivalence due to strict margins.

The GU2-Net-based automated measurement system achieved high accuracy, comparable to manual measurements, demonstrating its potential for clinical application.

Not applicable.