Longitudinal 3D quantification of joint space from CT images: Semi-automatic segmentation and feasibility of nnU-Net automation.
Authors
Affiliations (2)
Affiliations (2)
- UMR 7371 Laboratoire d'Imagerie Biomédicale, M.D., Ph.D. Laboratoire d'imagerie biomedicale Paris, France.
- UMR 7371 Laboratoire d'Imagerie Biomédicale, M.D., Ph.D. Laboratoire d'imagerie biomedicale Paris, France. Electronic address: [email protected].
Abstract
The most commonly imaging techniques used in knee osteoarthritis (OA) of the are Xrays and MRI. Computed tomography (CT) however, has his own advantages, especially three-dimensional high-resolution representation of subchondral bone and epiphyseal bone structure. Addition of a method allowing 3D joint space (JS) quantification would improve the clinical value of CT in knee OA follow-up. We here propose a method for quantifying the 3D JS from CT images using a semi-automatic and automatic segmentation based on deep learning method. Fifty-four subjects (6 men, 48 women; mean age: 62.8 ± 8.4 years), from a multicenter longitudinal study, with medial compartment OA (Kellgren-Lawrence grade 2 -3) who underwent high-resolution non-weightbearing CT scans performed 36 months apart (M00 and M36). A semi-automatic segmentation that could be manually corrected was used to segment the JS and to train a nnUNet algorithm. JS thickness mapping was obtained using a 3D sphere method both for the medial (MED) and lateral (LAT) compartments at times M00 and M36. Parameters measured included mean thickness (JS_mTh), standard deviation of thickness (JS-SDTh), minimum (JS_min), maximum (JS_max) and maximum/minimum ratio (JS_max/min). For the MED and LAT compartments, JS_mTh presented the best root mean square coefficient of variation (RMSCV%) about 2.7% and 2.8%, standard deviation (RMSSDmm) about 0.137 mm and 0.158 mm. JS_mTh, JS_SDTh and JS_max were significantly different for the MED compartment between M00 and M36, there is no differences for the LAT compartment, with 22% of the subjects with JS_mTh reduction beyond 0.5 mm for the MED compartment and 4% for the LAT compartment. The mean biases between semi-automatic and nn-Unet JS_mTh measurements were - 0.16mm±0.38 and -0.04±0.59 for MED and LAT compartments, respectively. This study shows that JS_mTh measured on non-weight bearing CT scans in patients with medial compartment OA had a satisfactory reproducibility and is able to measure significant variations between two CT-scans taken 36 months apart. Automatic segmentation based on nnUnet has the potential to replace semi-automatic segmentation.