Convolutional neural network technology can be used as an unbiased method for the scoring of disease activity on Doppler ultrasound images in patients with rheumatoid arthritis (RA) in both daily clinical practice and in clinical trials, according to the results of a study published in RMD Open.
The investigators sought to explore whether convolutional neural network architecture could be used for the interpretation of disease activity on Doppler ultrasound images, using the full-scale OMERACT EULAR Synovitis Scoring (OESS) system. They used 2 state-of-the-art neural networks to extract data from 1342 Doppler ultrasound images from patients with RA.
One of the neural networks divided the ultrasound images as either healthy (Doppler OESS score of 0 or 1) or diseased (Doppler OESS score of 2 or 3). The other neural network was used to score images across all 4 of the OESS systems Doppler ultrasound scores (from 0 to 3). The neural networks were then tested on a new set of RA Doppler ultrasound images (total of 176 Doppler ultrasound scans). The kappa statistic was used to measure agreement between rheumatologists’ scores and network scores.
For the neural network that evaluated healthy and diseased score, the highest accuracies achieved were 86.4% and 86.9%, respectively, compared with an expert rheumatologist, with sensitivities of 0.864 and 0.875 and specificities of 0.864 and 0.864. In contrast, the other neural network designed to use the 4-class Doppler ultrasound OESS system attained an average per-class accuracy of 75.0%, along with a quadratically weighted kappa score of 0.84.
The investigators concluded that convolutional neural network technology potentially can be used as a more unbiased technique for the scoring of ultrasound RA disease activity in both daily clinical practice and future clinical studies.
Andersen JKH, Pedersen JS, Laursen MS, et al. Neural networks for automatic scoring of arthritis disease activity on ultrasound images. RMD Open. 2019;5(1):e000891.