AI to Spot ARIA? FDA Says Yes

To keep patients on amyloid immunotherapy safe, doctors need to be able to spot subtle signs of ARIA on MRI scans. For memory clinics that have limited experience with these therapies, this may prove challenging. Now those clinics can consider a new tool to help, i.e. a software package called icobrain aria that uses artificial intelligence to detect and grade this side effect.

In a diagnostic study, radiologists using the software made more accurate calls for ARIA-E and ARIA-H than did radiologists reading scans without this assistance. Based on these findings, the U.S. Food and Drug Administration approved the technology for clinical use November 7.

Steve Salloway at Butler Hospital in Providence, Rhode Island, has consulted for icometrix, the company that developed icobrain aria. He is enthusiastic about the software’s potential, noting that it offers a standardized read of every scan. “It simulates the central reader we had in clinical trials. It doesn’t replace radiologists or the clinical read, but it augments them,” he told Alzforum before its FDA approval. He would like to see it widely adopted.

Other Alzheimer’s clinicians contacted by Alzforum are still unfamiliar with the technology, but interested in it. “Our health system’s neuroradiologists could consider taking it for a test run to get a feel for its utility,” Russell Swerdlow at Kansas University Medical Center, Kansas City, wrote to Alzforum. The approach makes sense to Vijay Ramanan at the Mayo Clinic in Rochester, Minnesota, but he reserved judgment about its application. “It will be interesting to see how this approach functions in real-world practice, where there can be variation in scanner type, sequences, and reader expertise,” he wrote (comments below).

Based in Leuven, Belgium, and in Boston, icometrix’s software measures factors such as the number, location, and size of areas of ARIA-E, and integrates them to produce a numerical severity score. For ARIA-H, it adds up the number of new microhemorrhages and regions of superficial siderosis.

In a clinical study, 16 radiologists read 199 retrospective MRI scans from the aducanumab Phase 2 and 3 trials either with or without the assistance of icobrain aria. The radiologists had an average of 10 years of clinical experience. The software nudged upward their ability to detect ARIA, though it was still not perfect. Sensitivity went from 71 to 87 percent for ARIA-E, and from 69 to 79 percent for ARIA-H. Specificity decreased, i.e., there were more false positives. Even so, overall accuracy went up from 0.82 to 0.87 for ARIA-E, and from 0.79 to 0.83 for ARIA-H (Sima et al., 2024). 

“Particularly given that some cases of ARIA can include mild or subtle findings which can have implications for management, progress in this space is a positive,” Ramanan noted.—Madolyn Bowman Rogers

Comments

  1. Using AI to help detect ARIA is an interesting concept, and one might imagine this could represent a logical application of AI technology. As to whether the benefits of augmenting a human read are tangible, I guess that question is hard to answer. This is something maybe our health system’s neuroradiologists could consider taking for a test-run, to get a feel for its utility.

    View all comments by Russell Swerdlow

  2. Accurate detection and characterization of ARIA is critical for optimizing delivery of anti-amyloid therapies. The technology behind this new tool is intriguing, and the concept of an assistive tool for image interpretation makes sense as one approach toward improved ARIA recognition.

    It will be interesting to see how this approach functions in real-world practice, where there can be variation in scanner type, sequences, and reader expertise. It will also be important that this and similar tools remain adjuncts for MRI interpretation rather than substitutes. 

    However, particularly given that some cases of ARIA can include mild or subtle findings which can have implications for management, progress in this space is a positive.

    View all comments by Vijay Ramanan

  3. AI-driven quantitative measures have been steadily making their way into radiological screening and decision-making. The ability to objectively quantify ARIA E and ARIA H will help physicians in managing mABs more precisely.

    View all comments by Marwan Sabbagh

Make a Comment

To make a comment you must login or register.

References

Therapeutics Citations

  1. Aduhelm

Paper Citations

  1. Sima DM, Phan TV, Van Eyndhoven S, Vercruyssen S, Magalhães R, Liseune A, Brys A, Frenyo P, Terzopoulos V, Maes C, Guo J, Hughes R, Gabr RE, Huijbers W, Saha-Chaudhuri P, Curiale GG, Becker A, Belachew S, Van Hecke W, Ribbens A, Smeets D. Artificial Intelligence Assistive Software Tool for Automated Detection and Quantification of Amyloid-Related Imaging Abnormalities. JAMA Netw Open. 2024 Feb 5;7(2):e2355800. PubMed.

External Citations

  1. approved

Further Reading

Annotate

To make an annotation you must Login or Register.