Skip to main content
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2026 Feb 12.
Published in final edited form as: AJNR Am J Neuroradiol. 2026 Feb 3;47(2):289–292. doi: 10.3174/ajnr.A8946

AI-Assisted Detection of Amyloid-related Imaging Abnormalities (ARIA): Promise and Pitfalls

Jeffrey R Petrella 1, Andrew J Liu 1, Laura A Wang 1, P Murali Doraiswamy 1
PMCID: PMC12867066  NIHMSID: NIHMS2130372  PMID: 40738658

Abstract

The advent of anti-amyloid therapies (AATs) for Alzheimer’s disease (AD) has elevated the importance of MRI surveillance for amyloid- related imaging abnormalities (ARIA) such as microhemorrhages and siderosis (ARIA-H) and edema (ARIA-E). We report a literature review and early quality assurance experience with an FDA-cleared assistive AI tool intended for detection of ARIA in MRI clinical workflows. The AI system improved sensitivity for detection of subtle ARIA-E and ARIA-H lesions but at the cost of a reduction in specificity. We propose a tiered workflow combining protocol harmonization and expert interpretation with AI overlay review. AI- assisted ARIA detection is a paradigm shift that offers great promise to enhance patient safety as disease-modifying therapies for AD gain broader clinical use; however, some pitfalls need to be considered.

The Rise of Anti-Amyloid Therapies and the Need for Vigilant ARIA Monitoring

The recent full U.S. Food and Drug Administration approval of lecanemab and donanemab represent a shift in the treatment of Alzheimer disease (AD) and mild cognitive impairment (MCI) (1, 2). While these beta-amyloid targeted monoclonal antibodies, now commonly referred to as anti-amyloid therapies (AATs), were approved based on their efficacy in slowing cognitive decline, both carry risks of amyloid-related imaging abnormalities (ARIA), including vasogenic edema (ARIA-E) and cerebral microhemorrhages or superficial siderosis (ARIA-H) (1, 2). These complications, while often relatively asymptomatic or mild, can sometimes be clinically significant or even life-threatening, necessitating regular MRI surveillance throughout treatment.

Challenges in Real-World ARIA Detection and Reader Variability

To address these risks, recent guidance from the American Society of Neuroradiology (ASNR) and other expert groups has outlined ARIA surveillance protocols, including recommended imaging sequences and timing (3). Yet, despite standardized approaches, real-world implementation remains imperfect. One key concern is reader variability—particularly among radiologists with limited real- world ARIA-specific experience. This issue is exacerbated when pre-treatment baseline scans are not done using the same sequences or on the same scanners as follow up scans. Even at large academic centers, including our own at Duke University, prior work and clinical experience has demonstrated significant inter-reader variability in ARIA detection. For example, a prior study documented that up to 84% of ARIA cases may be missed by local radiologists (versus a central expert reader) and that this error rate dropped to 18% with additional experience (4).

AI as a Potential Solution: Regulatory Approval and Technical Capabilities

In this context, artificial intelligence (AI) has emerged as a promising tool to improve efficiency, detection sensitivity and reproducibility and reduce interpretive bias. On November 7, 2024, the U.S. FDA granted 510(k) clearance to a Class II AI software device, termed icobrain aria, for detecting, measuring, and grading ARIA (5). A Class II device is one that poses moderate to high risk to the user. A 510(k) clearance does not indicate the FDA has approved the device, just that it has determined it is substantially equivalent to a predicate in terms of safety and efficacy, and has robust technical performance. Though there are other 510(k) Class II tools for quantitating microhemorrhages and white matter hyperintensities on the market, to date, this remains the first and only software product marketed in the US specifically for ARIA monitoring. The technical details of the software algorithm can be found elsewhere (6). The tool generates measurements on the number of brain regions and the longest axis impacted by ARIA-E, the number of new microhemorrhages and number of new areas of superficial siderosis as well as the severity (6). icobrain aria provides a report as well as the annotated segmentation images for review by the radiologist. As with any assistive software, the radiologist must make the final diagnosis. To date, icobrain aria has been used across 8 large radiology centers in several hundred patients.

Evidence from Reader Studies: Performance Metrics and Limitations

Using this tool, Sima et al (6) conducted a multi-case, multi-reader (MRMC) study of 199 retrospective cases that included 16 ARIA-reading trained board-certified radiologists (including 2 neuroradiologists). Among the 199 cases, 40 subjects had no ARIA, 84 cases had both ARIA-E and ARIA-H and the rest had either ARIA-E or ARIA-H. Interrater reliability for unassisted reads was suboptimal at kappa coefficients of 0.72 and 0.66 for ARIA-E and ARIA-H, respectively. Compared to unassisted reads, the AI- assistive reads led to an increase in sensitivity of 16% for ARIA-E and 10% for ARIA-H, but specificity decreased by 9% for ARIA-E and 3% for ARIA-H (6). Unassisted reads missed approximately 29% of ARIA-E and 31% of ARIA-H cases (including more than 53% of mild ARIA-E and 40% of mild ARIA-H cases) (6). Indeed, had this AI-software been utilized in the pivotal clinical trials of amyloid- targeted antibodies, the package inserts of these drugs would likely have listed substantially higher ARIA rates.

Real-World Experience: A Quality Control Initiative

To assess the performance of this tool in a real-world setting, a retrospective quality control (QC) initiative was done at our institution earlier this year. We applied the icobrain aria software to 73 de-identified MRI surveillance scans from 23 patients receiving lecanemab therapy as part of their clinical care. Of these, 66 scans were able to be analyzed by the software. By and large the reports were easy to interpret. In 3 of 7 patients with ARIA, ARIA was flagged only in AI reads. In two patients, microhemorrhages and siderosis were missed by the local radiologist. One case was asymptomatic but in the other case the patient reported multiple transient neurological symptoms (headache, confusion, dizzyness) that did not result in treatment discontinuation. In a third case, the software highlighted an isolated MH (Figure 1) that preceded ARIA-E, followed on later scans by a petechial pattern of hemorrhage. In this case, the subject also had some transient dizzyness and treatment was temporarily discontinued for 2 months and then resumed. Thus, overall, we found the software helpful in flagging small areas of abnormality, particularly subtle cortical microhemorrhages (ARIA-H) and superficial siderosis, which might otherwise have gone unrecognized.

FIG 1.

FIG 1.

Figure illustrates SWI (TR/TE 27/20) (a,b) and FLAIR (TR/TE 5000/390) (c,d) imaging sequences from surveillance MRIs in a patient following treatment with lecanemab. The first surveillance MRI scan (a,c) was read as not having any ARIA by the local neuroradiologist but the AI-tool (b) showed a new MH in the left occipital lobe (white arrows). The patient subsequently developed mild ARIA-E (measuring up to 4.5 cm) (d) and several additional microhemorrhages (not shown) in the left occipital lobe. Treatment was paused or 2 months. ARIA-H remained stable, ARIA-E resolved and treatment was resumed.

Common Pitfalls and Technical Limitations

We also observed a number of pitfalls. These included false positives such as: 1) stable hypertensive microhemorrhages in the brainstem misclassified as ARIA-H, 2) old infarcts flagged erroneously as ARIA-E lesions; 3) subdural hematomas labeled as extensive ARIA-E and >120 MHs; and 4) volume averaging of cortex and motion artifact leading to false positive ARIA-E calls. While the performance of the tool in tracking ARIA progression (severity, location) over time was reasonable, occasionally there were discrepancies in the quantitative data (number or size of ARIA) over sequential scans. For, example, there was a case where the same pontine microhemorrhage was called “new” on 3 consecutive scans. There were also cases where some microhemorrhages were originally missed then later picked up, and others where microhemorrhages were originally picked up then later missed on subsequent scans. In addition, some scans could not be processed due to pulse sequence differences between baseline and surveillance scans. Until such limitations are overcome, ARIA AI-assistive tools should be used as an adjunct, not a replacement, for expert review.

Correlation with AI-Assisted Study Findings

Our observations are broadly consistent with those of Sima et al (6), who demonstrated that radiologists reviewing AI-flagged regions detected more ARIA lesions with increased confidence and diagnostic efficiency. Like their study, our real-world experience showed improved sensitivity for subtle findings, with the caveat that specificity remains a challenge due to frequent false positives. Importantly, in our QC sample, the AI software classified approximately two-thirds of scans as negative—findings that were later confirmed by radiologist review. This suggests future potential for a triage model, in which AI filters out definitively negative scans, allowing radiologists to devote more attention to positive or ambiguous cases.

Proposed Tiered Integration into Clinical Workflow

Based on this cumulative experience, we propose a tiered approach for introducing these tools into the clinical workflow:

  1. Education of radiology and clinician team on strengths and weaknesses; Inter-departmental communication to ensure protocol consistency while ordering scans.

  2. An initial interpretation, performed independently by a radiologist, without reference to the AI results.

  3. Review of the AI findings, prompting reassessment of any areas flagged by the software, followed by a final interpretation made by the radiologist.

  4. Treatment decision made by the clinician; Follow up MRIs as indicated.

The above approach may slow workflow slightly but minimizes false negatives resulting from satisfaction-of-search bias (i.e., prematurely terminating the search if the AI detects nothing) while preserving the radiologist’s independent judgment, thereby minimizing false positives.

Implications for Patient Care and Future Research

The implications for AD and MCI patient care are significant. Increased sensitivity and earlier identification of subtle microhemorrhages by AI prompts would allow for closer monitoring and, in some cases, changes in treatment decisions which might prevent serious harm or even death. Further, it would reduce variability and eventually, through appropriate triage, make workflow more efficient. The reports are easy to convey to referring-clinicians, patients and caregivers. Finally, the quantitative ARIA data generated by the AI tool will facilitate objective real-world studies of the effectiveness and safety of disease modifying therapies.

Remaining Caveats and Areas for Development

Caveats remain. AI ARIA detection tools currently require substantial oversight, monitoring and strategies to reduce false-positive rates, including model improvements, more training cases with a rigorous gold standard, automated motion correction, and longitudinal imaging protocol harmonization. To strengthen clinical utility, multi-center validation and integration of clinical data (e.g., APOE4 status, cortical amyloid beta levels) into AI algorithms would improve robustness. While we focused only on the icobrain tool in our report, it would also be important for future studies to perform direct comparison studies of icobrain versus other tools (such as NeuroQuant and Neurophet AquaAD) that can also be utilized for microhemorrhage or ARIA detection.

Conclusion: Moving Toward Scalable ARIA Monitoring

Assistive AI tools are not meant to serve as standalone diagnostics and should not replace the joint judgment of radiologist and referring clinician. Relying too much on the AI (“cognitive offloading”) could lead to errors in diagnosis - as noted earlier, the tool prioritizes increased sensitivity at the cost of slightly reduced specificity so could flag findings that may not be ARIA. Icobrain aria was trained on 2D FLAIR and T2* sequences and hence is not ideal for use with 3D FLAIR or SWI, though we applied it exclusively to these latter sequences in our project. Last, but not least, greater interdepartmental communication between the referring clinicians and radiologist is needed to optimize its utility so referring clinicians are clear that the radiology report, not the AI-generated report, represents the final interpretation.

From a public health standpoint, the value of AI-assisted ARIA detection software tools will only increase as more patients seek AATs. As treatment expands beyond infusion centers to community clinics and as new administration routes (e.g., subcutaneous delivery) and better predictors of risk are developed, accurate, scalable MRI monitoring becomes critical. If used judiciously, AI tools that increase sensitivity and consistency of ARIA surveillance across readers and institutions can help ensure safer access to these disease-modifying therapies.

ACKNOWLEDGMENTS

We wish to thank icometrix for providing icobrain aria software analysis at no cost. icometrix provided no funding and had no role in the preparation of this manuscript.

JRP is supported by grants from the NSF (DMS-2052676), NIH (P30AG072958) and a gift from Gates Ventures, and has served on the advisory boards of cortechs.ai, icometrix and Biogen. AJL is a consultant for Lucent Technologies and receives support from the NIH (P30AG072958). LAW has no disclosures. PMD is supported by grants from NIH (P30AG072958) and Duke UNC ADRC Seed Grant, Cure Alzheimer Fund, as well as gifts from Gates Ventures, Karen L Wrenn Trust, lululemon and Steve Aoki Foundation. PMD has served as an advisor to several pharmaceutical and technology companies, and serves on the board of AHEL Ltd. PMD owns stock in several biotech companies and is a co-inventor (through Duke) on patents related to dementia diagnosis or therapy.

ABBREVIATIONS:

AAT

anti-amyloid therapy

ARIA

amyloid-related imaging abnormalities

ARIA-H

amyloid-related imaging abnormality-hemorrhage

ARIA-E

amyloid-related imaging abnormality-edema

REFERENCES

  • 1.van Dyck CH, Swanson CJ, Aisen P, et al. Lecanemab in early Alzheimer’s disease. N Engl J Med 2023;388:9–21. [DOI] [PubMed] [Google Scholar]
  • 2.Mintun MA, Lo AC, Duggan Evans C, et al. Donanemab in early Alzheimer’s disease. N Engl J Med 2021;384:1691–1704. [DOI] [PubMed] [Google Scholar]
  • 3.Cogswell PM, Madhavan A, Kanekar S, et al. Alzheimer’s disease anti-amyloid immunotherapies: imaging recommendations and practice considerations for ARIA monitoring. AJNR Am J Neuroradiol 2024;45. doi: 10.3174/ajnr.A8469 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 4.Gutt S; Brocco A; Sharma S P1-045: Variability of Aria Detection in Patients Receiving Monoclonal Antibodies Against Amyloid-β Plaques. Alzheimers Dement. 2019, 15, P250. DOI: 10.1016/j.jalz.2019.06.070 [DOI] [Google Scholar]
  • 5.U.S. Food and Drug Administration. 510(k) premarket notification: K240712. https://www.accessdata.fda.gov/cdrh_docs/pdf24/K240712.pdf. Accessed June 14, 2025.
  • 6.Sima DM, Phan TV, Van Eyndhoven S, et al. Artificial Intelligence Assistive Software Tool for Automated Detection and Quantification of Amyloid-Related Imaging Abnormalities. JAMA Netw Open. 2024;7(2):e2355800. doi: 10.1001/jamanetworkopen.2023.55800 [DOI] [PMC free article] [PubMed] [Google Scholar]

RESOURCES