EVALUATION OF DIFFERENT METHODS OF AUTOMATED 3-D JOINT SPACE MAPPING FROM WEIGHT BEARING CT SUGGESTS A TIBIAL MESH-TO-MESH APPROACH IS MOST SENSITIVE

Abstract

INTRODUCTION

Weight bearing CT (WBCT) has the distinct advantage over radiography of being able to provide 3-D imaging of the knee joint while standing. It is also more practicable and better at depicting mineralized joint structures than MRI. Several different approaches to 3-D JSW measurement have been developed, but their repeatability has not been directly compared.

OBJECTIVE

To compare the test-retest repeatability of three different methods of 3-D joint space mapping (JSM) of the tibiofemoral compartment from WBCT imaging data.

METHODS

14 individuals recruited and consented at the University of Kansas Medical Center had baseline and follow-up WBCT imaging suitable for analysis. Participant demographics were: mean ± SD age 61.3 ± 8.4 years, BMI 30.7 ± 4.3 kg/m² and male:female ratio 8:6. All scanning was performed on the same XFI WBCT scanner (Planmed Oy, Helsinki, Finland) with the mean ± SD interval between baseline and follow-up attendances 14.9 ± 8.1 days. A Synaflexer^TM device was used to standardize knee positioning during scanning. Imaging acquisition parameters were 96 kV tube voltage, 51.4 mA tube current, 3.5 s exposure time. A standard bone algorithm was applied for reconstruction with 0.3 mm isotropic voxels and a 21 cm vertical scan range. Both knees were included in all analyses with SD adjustments made for multiple observations from the same individual. Participant ID and scan sequence were anonymized prior to analyses. An algorithm based on U-net was implemented in C++ using LibTorch and integrated into ScanXM software for automatic segmentation of the femur and tibia from all knees. Three different JSM techniques were applied: (1) femur-to-tibia deconvolution in which the femur was the base (performed in Stradview); (2) tibia-to-femur deconvolution in which the same was done but from the tibia; and (3) tibia-to-femur mesh-to-mesh distance using a custom MATLAB script. Results from each technique were registered using wxRegSurf and displayed on their average halfway joint space mesh (i.e. the middle plane of the joint space) using custom MATLAB scripts. Bland Altman descriptive statistics were calculated as 3-D bias (follow-up minus baseline) and limit of agreement (LOA) maps for all knees. Summary statistics also included root mean square coefficient of variation (RMSCV) and LOA as a % of the mean.

RESULTS

3-D bias and LOA maps for all knees are displayed on the halfway joint space patches as if viewing the right knee from the inferior aspect (Figure 1). Both deconvolution techniques showed similar noise patterns of bias around a zero value, while the mesh-to-mesh technique suggested systematically wider anterior and narrower posterior JSW at follow-up, but this was of sub-millimeter magnitude. Both deconvolution techniques also showed a pattern of worsening LOA towards the joint space patch margins, recognized as where errant or null values can be exaggerated by data smoothing. Mesh-to-mesh LOA was more robust across the whole joint space. When comparing repeatability measures for KLG < 2 and KLG = 2 groups (Table 1), LOAs from the whole joint space were similar for all techniques, ranging from 1.29 to 1.46 mm across groups, while the best LOA value of 0.13 mm was seen in the mesh-to-mesh KLG = 2 group at the inner aspect of both compartments.

CONCLUSION

Although differences between the three approaches to JSM were subtle, a tibial-based mesh-to-mesh technique may be more robust, in particular at the margins of the joint space. This approach also appeared to have a greater potential sensitivity for detecting smaller changes in JSW from having the lowest LOA (thus smallest detectable difference) in individuals with KLG = 2, an important stratification in OA clinical trials before structural disease is too severe. However, the caveat for a mesh-to-mesh approach to JSM derived from segmentation is that it relies on the accuracy of the segmentation technique that may vary between approaches, whereas a deconvolution approach has been proven to be accurate and is only marginally less repeatable.