NEW YORK — On the basis of MRI images of the hand, a neural network has been trained to distinguish seronegative and seropositive rheumatoid arthritis (RA) from psoriatic arthritis (PsA) as well as from each other, according to a study that was presented at the annual meeting of the Group for Research and Assessment of Psoriasis and Psoriatic Arthritis.
In the work so far, the neural network was correct about 70% of the time in the absence of any further clinical analyses, according to David Simon, MD, a rheumatologist in the department of internal medicine at Friedrich-Alexander University, Erlangen, Germany.
Previous to this work, “there has been no study that has exclusively used hand MRI data and deep learning without requiring further expert input for the classification of arthritides,” Simon said.
In fact, when demographic and clinical data were added, there was no improvement in the performance of patient classification relative to the deep learning classification alone, according to the data presented by Simon.
The images were evaluated with residual neural networks (ResNet), which represents a sophisticated form of deep learning to facilitate the flow of information across the network layers as they form to improve accuracy in their ability to distinguish one form of disease from the other. The training was performed on images from the T1 coronal, T2 corona1, T1 coronal fat suppressed with contrast, T1 axial fat suppressed with contrast, and T2 fat suppressed axial sequences.
The study included hand MRI scans from 135 patients with seronegative RA, 190 with seropositive RA, 177 with PsA, and 147 with psoriasis. The performance was judged on the basis of area under the receiver operating characteristics curve (AUROC) with and without input of clinical characteristics. Patients who had psoriasis without clinical arthritis were included as a control population.
The AUROC for accuracy was 75% for seropositive RA relative to PsA, 74% for seronegative RA relative to PsA, and 67% for seropositive relative to seronegative RA. Of the patients who had psoriasis without arthritis, 98% were classified as PsA and 2% as RA.
Subsequent to the classification of the patients with psoriasis, 14 of the 147 (9.5%) have developed PsA so far over a relatively short follow-up. All of these were among those identified as PsA by neural network evaluation of the hand MRIs.
This suggests that “a PsA-like pattern may be present early in the course of psoriatic disease,” Simon said.
In the groups with joint disease, who had mean ages ranging from 56 to 65, the mean disease durations were 2.6 years for those with seropositive RA, 1.3 years for those with seronegative RA, and 0.8 years for those with PsA. The patients with psoriasis were younger (mean age, 40.5 years) but had a longer disease duration (mean 4.2 years).
All of the MRI sequences were relevant for classification, but contrast did not appear to help with accuracy.
“If the images with contrast enhancement were deleted, the loss of performance was only marginal,” Simon reported.
The accuracy of neural networks increases with data, making it likely that further refinements in methodology will lead to a greater degree of accuracy, according to Simon. While the methodology is not yet ready for routine use in the clinic, the study demonstrates that neural network analysis of hand MRI to distinguish forms of arthritis “is possible.” Further studies are planned toward the goal of creating a viable clinical tool.
“Of course, if we could create an accurate tool with ultrasound, this would be even more practical,” said Simon, recognizing the value of an office tool, but he cautioned that this would be far more challenging.
“The precision of MRI is an important factor for effective neural network training,” he said.
Utility: “In Challenging Cases if the Accuracy Improves?”
A viable method for objectively and rapidly distinguishing inflammatory joint diseases, particularly in patients with an ambiguous clinical presentation, is an unmet need, according to Philip J. Mease, MD, director of rheumatology research at Swedish Medical Center, Seattle.
Although the data presented are promising, Mease said in an interview that he believes there is a fair amount of work to be done before imaging analysis based on deep learning makes its way into routine clinical care. He is also hoping for methods to distinguish RA from PsA that are easier and less expensive, such as serum biomarkers. However, he agreed that a MRI-based tool could be useful when differentiating disease that is challenging.
“MRI is an expensive way for routine classification of disease, but this approach could be useful in challenging cases if the accuracy improves,” he said.
Meanwhile, other clinical researchers might want to test the principle. “You can try it,” said Simon, who reported that his team has made the methodology publicly available.
Simon reported no conflicts of interest. Mease reported financial relationships with more than 10 pharmaceutical companies, most of which make products used for the treatment of inflammatory joint diseases.
This article originally appeared on MDedge.com, part of the Medscape Professional Network.