Measurement of synovial tissue volume in knee osteoarthritis using a semiautomated MRI-based quantitative approach

Abstract

Purpose: Synovitis is common in knee osteoarthritis and is associated with both knee pain and progression of disease. Semiautomated methods have been developed for quantitative assessment of structure in knee osteoarthritis. Our aims were to apply a novel semiautomated assessment method using 3D active appearance modeling for the quantification of synovial tissue volume (STV) and to compare its performance with conventional manual segmentation.

Methods: Thirty-two sagittal T₁ -weighted fat-suppressed contrast-enhanced MRIs were assessed for STV by a single observer using 1) manual segmentation and 2) a semiautomated approach. We compared the STV analysis using the semiautomated and manual segmentation methods, including the time taken to complete the assessments. We also examined the reliability of STV assessment using the semiautomated method in a subset of 12 patients who had participated in a clinical trial of vitamin D therapy in knee osteoarthritis.

Results: There was no significant difference in STV using the semiautomated quantitative method compared to manual segmentation, mean difference = 207.2 mm³ (95% confidence interval -895.2 to 1309.7). The semiautomated method was significantly quicker than manual segmentation (18 vs. 71 min). For the semiautomated method, intraobserver agreement was excellent (intraclass correlation coefficient (3,1) = 0.99) and interobserver agreement was very good (intraclass correlation coefficient (3,1) = 0.83).

Conclusion: We describe the application of a semiautomated method that is accurate, reliable, and quicker than manual segmentation for assessment of STV. The method may help increase efficiency of image assessment in large imaging studies and may also assist investigation of treatment efficacy in knee osteoarthritis.

Keywords: osteoarthritis; segmentation; semiautomated; synovial tissue volume.

Figure 1

Stages of semi‐automated quantitative assessment of STV using our approach. (A) Preprocessing and development: Mean of masks deformed into a reference frame created for each slice in the MRI sequence to loosely identify regions in which synovitis commonly occurs; a 3D mask (the white regions) for a given slice. (B) Application: Synovitis‐shape model was registered to the target image. Position of the mask was moved, and/or manual editing of the applied shape model was then completed on a slice‐by‐slice basis if failing to overlay the synovium. (C) Targeted thresholding, based on voxel signal intensity, was completed manually to identify STV (red) and efficiently remove the remaining tissue and fluid (blue) from measurement. (D) Editing, post‐thresholding, of the remaining voxels was completed where appropriate by adding and removing voxels manually. Automatic calculation of STV was completed by summation of voxels across all slices, generating an absolute total volume. STV, synovial tissue volume

Figure 2

Semi‐automated quantitative assessment of STV using our approach. Software‐computer interface displays a sagittal T₁‐w postcontrast fat‐suppressed MRI (right) with overlaid synovitis model (green) and a 3D‐rendered axial plane with overlaid synovitis model (left). (A) Correct positioning of synovitis mask to the target image. (B) Displacement in the posterior direction of the synovitis mask to the target image. The displacement of the synovitis‐shape model can also be seen on the 3D rendered axial plane

Figure 3

An example of the 2 probability density functions (for low‐ and high‐intensity data) that were used during the thresholding step in the software to determine whether a voxel contained synovitis

Figure 4

Bland‐Altman plot of differences between semi‐automated segmentation and manual segmentation for the measurement of STV (mm³) versus the averages of the 2 segmentation methods for the assessment of STV (mm³) (N = 29). Measurements reported using the 2 methods were completed by a single reader (t.a.p.). Data taken from the Brace study