Background/Purpose

 There is increasing need for sensitive, objective outcome measures in rheumatoid arthritis (RA). MRI is more sensitive than clinical examination and X-rays, using a semi-quantitative score (OMERACT RAMRIS). The objective was to use statistical shape models (SSMs) to identify RA pathology, and compare rates of change with RAMRIS in a RA population treated with a biologic therapy.

Methods

 MR images of the hand were acquired from 45 established, active (mean (SD) DAS28 = 5.4(1.1), seropositive) RA patients who received a single cycle of rituximab at 0 months in an open label study.  DAS28 improved by mean (SD) 1.7(1.4). Subjects were imaged at 0,3,6,9 & 12 months.  Hand bones were segmented using SSMs; resulting bone surfaces were used to quantify erosion and oedema (inside bone), synovitis and 3D joint space width (3DJSW, outside bone). RAMRIS scoring was performed by a single experienced reader blinded to time point. Multilevel linear regression analysis was used to model changes over time. Quantitative values were adjusted for differences in scale using total bone area at baseline. Observations were nested within patients; random slopes for time were included. Non-linear changes were modelled with polynomial terms. Likelihood ratio (LR) tests were used to compare nested models. Accuracy of automated segmentation was examined using point-to-surface distance between automated and manually segmented bone surfaces.

Results

Quantitative synovitis decreased over time; the extent of the change varied between patients (Table 1). By contrast, there was no change over time in RAMRIS synovitis. There was no change in quantitative or RAMRIS erosion and no inter-patient variation. In some patients there was evidence of cubic change in quantitative oedema; with relatively rapid change between 0 and 3m. There was no evidence of changes in RAMRIS oedema. There was strong evidence of change over time for 3DJSW (LR test chi-square=22.0, p<0.001), though the result was complex, with the most responsive patient showing increase in joint space width, while others showed a decrease. Mean accuracy of automatically generated bone surfaces was 0.2 mm (Table 2); 90^th percentile 0.44 mm

Conclusion

Fully automatic quantitative analysis of RAMRIS measures is now practical, and showed change for synovitis and oedema where RAMRIS did not.  3DJSW is a novel measure that needs careful interpretation as it includes more than one type of change. Quantitative analysis depends upon accurate and automatic identification of all hand bones; these results show excellent accuracy of around 1/3 pixel.