AI model predicts bone marrow edema, may reduce diagnostic delay in spondyloarthritis
Click Here to Manage Email Alerts
Key takeaways:
- A fully automatic machine learning algorithm was able to predict bone marrow edema in sacroiliac joint MRI.
- During testing, the algorithm demonstrated a balanced accuracy of 72.1%.
A fully automated, machine learning model that predicts bone marrow edema in sacroiliac joint MRI could improve screening for spondyloarthritis and decrease diagnostic delay, according to data published in Arthritis & Rheumatology.
“Standardized methods to evaluate [sacroiliac (SI)] joint MRI ... lack consistency,” Joris Roels, PhD, of Ghent University, in Belgium, and colleagues wrote. “In a diagnostic context, detection and especially interpretation of [bone marrow oedema (BMO)] on MRI of the SI joints require expertise given the broad range of sacroiliitis mimickers such as mechanical stress-induced BMO or degenerative disease.”
To develop, train and test a machine learning model for spondyloarthritis screening, Roels and colleagues used data from patients with SpA, as well as postpartum patients and healthy control individuals, seen at Ghent University Hospital. Scans of postpartum patients and healthy controls were used to train the learning model regarding lesions that are not related to SpA. Meanwhile, scans from patients in the SpA group were split — half were used to train the model, and the other half to test the model.
The researchers defined the basis for these lesions through manual interpretation. The test dataset consisted entirely of patients with SpA. In all, 279 scans of patients with SpA, 71 scans from postpartum patients and 114 healthy control scans were used to train the model, while 243 scans were used to validate the algorithm.
The algorithm was able to detect sacroiliac joints at a precision rate of 98.4%. During cross-validation, it demonstrated an area under the curve (AUC) of 94.5%, a balanced accuracy of 80.5% and produced an F1 score of 64.1%. The algorithm also produced an AUC of 88.2% during testing, a balanced accuracy rate of 72.1% and an F1 score of 50.8%, according to the researchers.
“This study is promising as a first step towards efficient, objective and reliable prediction and quantification of BMO of the SI joints on MRI,” Roles and colleagues wrote. “This could be a potential solution for reducing the time to a correct diagnosis and improve objective assessment of MR images in SpA patients, especially in high-throughput or non-expert settings.”
Perspective
Back to Top
Everyone practicing in the field of rheumatology has likely ordered an MRI to evaluate the sacroiliac joints of a patient suspected to have spondyloarthropathy. However, most do not perform follow-up scans to monitor response.
Prior to the familiarity of MRI for this purpose, we relied on X-rays — and rheumatologists are very proficient at interpreting sacroiliac (SI) joint X-rays. That said, patients with active inflammation without more advanced changes are missed, and X-ray changes cannot be used to predict or monitor response to therapy.
We recognize that MRI is a useful tool but the lack of standardization in performing and interpreting the changes currently limits its utility. Artificial Intelligence (AI) and machine learning (ML) could be employed in the future not only to screen the EHR of patients with back pain, but also to image at-risk populations.
Roles and colleagues describe a fully automated model that predicts bone marrow oedema, (BMO), representing SI joint synovitis, the hallmark of spondyloarthropathies. The investigators’ approach is novel as it is potentially fully automated, different than previously described algorithms. Their algorithm identifies the SI joint and segregates the sacrum from the ilium. Overall, it performed well at SI joint detection and segmentation, as well as prediction of inflammation. However, the prediction of SPARCC scores did not show “acceptable accuracy.”
Although promising, the authors did recognize the limitations of their model.
Personally, I find the promise of AI and ML exciting. This will certainly enhance our ability to accurately diagnose earlier and to intervene more meaningfully. However, for the near future, I don’t see the clinical rheumatologist being replaced. Even when these tools are ready for prime time, human interpretation still seems necessary to put the results into proper context.
Collapse
Click Here to Manage Email Alerts