Chest radiography (CXR) interpretation is highly accurate when using artificial intelligence (AI)-assisted diagnosis; nevertheless, it is unknown whether this technology can help non-radiologist physicians identify lung abnormalities on CXR.
The intervention (with AI-assisted interpretation) and control (without AI-assisted interpretation) groups were assigned to eligible patients who attended an outpatient clinic by the researchers. Seven non-radiologists recorded lung lesions on CXR, either with or without AI support. For lung lesions, the reference standard was set by three radiologists. The doctors’ clinical judgment and diagnostic precision were the main and secondary goals.
Data presented at ACR Convergence 2023 showed that an automated deep learning model, based on hand radiographs of patients with rheumatoid arthritis, effectively recognized and forecasted joint space narrowing and erosion.
A total of 162 patients were randomized into the intervention and control groups between October 2020 and May 2021. For the CXR and lung lesion levels, the intervention group’s area under the receiver operating characteristic curve was noticeably larger than the control group’s. The intervention group exhibited reduced erroneous referral rates for the lung lesion level and higher sensitivity for both CXR and lung lesion levels. Clinical judgments made by the doctors were unaffected by AI-assisted CXR interpretation.
The researchers discovered that AI-assisted CXR interpretation improved the diagnostic ability of non-radiologist physicians in recognizing aberrant lung results.
An associate professor at the University of Manitoba named Carol Hitchon, MD, FRCPC, MSc, stated at a press conference that radiographic damage is one of the primary outcome measures for rheumatoid arthritis. As clinicians, we typically perform serial plain X-rays in order to monitor and identify joint damage in a patient over time, which informs our management choices.
The Sharp-Van der Heijde score, one of the more widely used metrics for assessing joint degradation, indicates how much collaboration space is contracting and how many joint erosions there are, however it is impractical in clinical practice.
She added that it takes a lot of time to grade these joints using these scores. Since there is a great deal of intra- and inter-observer variability, one needs a fair level of knowledge. In many centers, the knowledge necessary to provide these scoring and assess radiographs is frequently lacking.
Constructing the system
In order to train a convolutional neural network-based algorithm to identify joints in 240 training and 89 test pediatric hand X-rays from the Radiologic Society of North America database, Hitchon and colleagues first developed a deep learning system that would automatically detect joints and predict Sharp-Van der Heijde scores in patients with RA based on hand X-rays.
According to Hitchon, their goal was to create a model that would allow the computer to search, X-ray, and locate joints. Its joint type will be identified by the model. Next, it determines whether the joint is injured and provides the score for erosions and joint space narrowing.
ocating the outcomes
According to their findings, the deep learning model—which had an adult data F1 score of 0.812 and a pediatric F1 score of 0.991—could correctly identify the target joints. The vision transformer model produced very accurate predictions, with root main squared errors of 0.93 and 0.91 for erosion and joint space narrowing, respectively.
According to Hitchon, this model might be useful in bigger investigations where joint damage is the end result, like randomized clinical trials or medication trials involving sizable cohorts.
