Related Experiment Video
Updated: Aug 21, 2026
Four-Dimensional CT Analysis Using Sequential 3D-3D Registration
Published on: November 23, 2019
3D QUANTITATIVE MEASUREMENTS OF JOINT SPACING BASED ON SEMI-AUTOMATIC AND AUTOMATIC SEGMENTATION, CONDUCTED 36 MONTHS
J B Deloges1, S Uk2, J Cohen2
1Laboratoire d'Imagerie Biomédicale CNRS, Inserm, Sorbonne Université, Paris, France.
Introduction:
Computed tomography (CT) is less frequently used than X-rays or MRI in knee osteoarthritis (OA), but remains very useful for studying changes in the subchondral bone. Furthermore, CT offers good spatial and isotropic resolution, good bone-soft tissue contrast, and as such is a good tool for quantitative imaging in OA. Cartilage degeneration results in variations in joint space (JS) width. quantifying JS width in 3D would enhance the clinical value for monitoring disease progression using CT.
Objective:
We propose a simple method for quantifying the 3D JS variations from CT images, using semi-automatic segmentation as ground truth and automatic segmentation based on deep learning.
Methods:
A subset of fifty-four subjects (6 men, 48 women; mean age: 62.8 ± 8.4 years) with OA of the medial compartment (grade 2-3 according to Kellgren-Lawrence) originally recruited for a multicenter longitudinal study (Sequoia - Russia, Belgium and France). High-resolution CT Siemens scans (0.25 × 0.25 × 0.30 mm) in the supine position with knees extended were performed at two timepoints 36 months apart (M00 and M36). Semi-automatic segmentation using ItkSNAP (version 3.8.0), which allowed for manual correction was performed by an expert. Based on the expert's segmentation, a deep learning model, nnUNet, was trained with Russian data (n=65) and validated based on Belgium and France data (n= 37). The mapping of JS thickness was obtained using the 3D sphere method, applied to the medial (MED) and lateral (LAT) compartments at M00 and M36. Measured parameters were JS volume (JS_vol), mean thickness (JS_mTh), standard deviation of thickness (JS-SDTh), and minimum (JS_min) and maximum (JS_max) values.
Results:
Individual results of JS_mTh derived from the expert are presented in Figure 1. Comparing the semi-automatic JS_mTh measurements from expert and those obtained by the nn-UNet, the mean DICE coefficients and the Hausdorff distances were 0.86, and 5.1 mm, and 0.83, 7.3 mm for the MED and LAT compartments, respectively. The mean biases are presented in Figure 2 and were -0.15 mm ± 0.38 for the MED and -0.04 mm ± 0.59 for the LAT compartments.
Conclusion:
This study demonstrates that measuring JS_mTh from CT scans in patients with OA of the MED compartment will be a viable option for monitoring the progression of knee OA. Furthermore, automatic segmentation based on nnUNet demonstrates its ability to replace semi-automatic segmentation performed by an expert.
