autoscoRA: Deep Learning to Automate Sharp/van der Heijde Scoring of Radiographic Damage in Rheumatoid Arthritis

Abstract

Objective: Regular imaging by conventional radiography to assess for joint damage is a cornerstone in the management of rheumatoid arthritis (RA). Scoring systems to quantify such damage, such as the widely used Sharp/van der Heijde (SvdH) score, are limited by the requirement of time and experienced staff as well as intra- and inter-rater variability. To alleviate these problems, autoscoRA, a fully automated scoring system to assign SvdH scores to radiographs of the hands and feet was developed. Methods: Using the hitherto largest dataset of adult rheumatoid arthritis patients, autoscoRA, a deep learning-based system, was trained to automatically perform joint extraction and scoring of joint space narrowing and bone erosion. Results: The dataset included 769 patients (155 of which in the test set) with 3437 visits (707) and 12144 radiographs (2507). The model reached excellent agreement with a human scorer for joint space narrowing, erosion, and combined scores both on the joint level and for summed total SvdH scores (ICC 0.9). On a subset of data scored by a second human reader, the model outperformed the former in terms of agreement with the first human reader. In addition, autoscoRA demonstrated good agreement with a human reader for detecting longitudinal progression of joint damage across different SvdH score cut-offs defining the presence of progression (average agreement of 70 %). Conclusion: Automated systems like autoscoRA could be used to facilitate scoring of radiographic joint damage in clinical trials, registries and observational studies, and, eventually, routine clinical care.

Links & Resources

Authors

Cite This Paper