POS0925 AUTOMATIC SCORING OF ULTRASOUND SYNOVIAL HYPERTROPHY IN RHEUMATOID ARTHRITIS THROUGH INTEGRATING MULTIPLE CONVOLUTIONAL NEURAL NETWORK MODELS

Background

The OMERACT-EULAR Synovitis Scoring (OESS) system is worldwide used to evaluate arthritis severity on ultrasound (US) images. Because of inter-observer and intra-observer variability, deep learning (DL) has been applied in high-quality image interpretation and analysis. Previous studies mostly focused on Doppler US (DUS) classification by convolutional neural network (CNN), which could provide objective assessment. However, the reports of DL intervention in grey scale (GS) US image automatic measurements are limited.

Objectives

The aim of this study was to develop an integrated multiple CNN model in precise scoring GS US images from rheumatoid arthritis (RA) patients.

Methods

The standard US images from patients of RA were retrospectively selected by three 10-years US experienced rheumatologist together and were graded according to the OESS system. Six different joints data were taken, including proximal interphalangeal, metacarpophalangeal, wrist, elbow, knee and ankle joints. We conducted the DL model integrating three binary CNNs to predict four-class GS US scoring (Figure 1). The accuracy of the trained model was tested by an independent test data.

Results

Total 678 images from 447 patients of RA were used in this study. These images were divided into training (n=611) and testing (n=67) sets. The integrated multiple CNNs model could achieve a four-class accuracy of 77.6%. The individual accuracy of grades 0, 1, 2 and 3 were 68.4%, 77.3%, 73.3% and 100%, respectively (Table 1). Furthermore, we found that adding on anatomic site parameters or labeling areas of interest would establish a better average area under curve (AUC) with 92.6% and 89.0%.

Conclusion

Our study suggests the possibility of using the integrated multiple CNNs model in grading synovial hypertrophy of RA, which is critical in RA healthcare. External validation would be required to confirm the predictive ability of this model.

References

[1]D’Agostino MA et al. RMD Open. 2017 Jul 11;3(1):e000428. [2]Andersen JKH et al. RMD Open. 2019 Mar 30;5(1):e000891. [3]Christensen ABH et al. Ann Rheum Dis. 2020 Sep;79(9):1189-1193. [4]Shin Y et al. Ultrasonography. 2021 Jan;40(1):30-44. [5]Zhou Z et al. Patterns (N Y). 2022 Sep 29;3(10):100592.

Acknowledgements:

NIL.

Disclosure of Interests

None Declared.