We saw this LCD topic as Final in February from CGS* and in Draft from Wellpoint Federal, here it comes from Palmetto GBA in draft in October..
On October 1, Palmetto released for comment LCD DL40464, "Automated Detection and Quantification of Brain MRIs." The take-home lesson is simple:
- This is a non-coverage policy for artificial intelligence (AI) assistive software tools for automated detection and quantification of the brain.
Examples include ARIA detection and quantification (vascular and tissue abnormatlies in dementia therapy. Petrella 2026).
Obviously, the same principle could be applied to detection and quantification of features and properties of glass slides intead of MRI images.
75 footnotes. Links to an NCD 200.3for monoclonal treatment of Alzheimer's. Draft billing article DA60550. [boilerplate re ABN etc] Comment to 11/14/2026.
#######
#######
The MAC writes,
INTRODUCTION
There is interest in AI algorithms, including machine learning and deep learning, that automate neuroimaging in an effort to improve accuracy, reduce bias, and aid clinical decision-making. Machine learning is used to evaluate candidate biomarkers for assessing multiple sclerosis and disease severity, with random forest models commonly performing well, while deep learning remains less widely adopted. Translating these methods into clinically meaningful decision-support tools is limited by insufficient data quality and quantity, as well as a lack of shared standards. Advancing clinical impact will require robust, shared data resources and integrated analyses across multiple data modalities rather than isolated, single-purpose solutions.
ANALYSIS
Investigations have explored the potential of automated quantification technology for the evaluation of ARIA, MS, TBI, epilepsy, brain tumors, and other neurologic conditions. However, progress has been limited by the lack of established measurement standards and insufficient access to large, diverse datasets for training these tools. While expert radiologists interpret imaging based on visual patterns, these programs quantify brain volumes. Although this approach is promising, there are currently no standardized reference values defining normal brain volumes, and each program relies on proprietary data that is not interchangeable across platforms. In addition, the datasets used to train these models lack sufficient diversity to account for variations related to age, gender, or ethnicity. This limitation is particularly relevant in the Medicare population, where age related changes in brain volume are common. Without standardized data, it is difficult to determine whether subtle changes reflect true pathology or normal variation. At this time, there is insufficient evidence of clinical utility or validity, and use of this technology is considered investigational and therefore not covered. Palmetto GBA will continue to monitor the evolving research related to these devices.
##
##
It's important to note this is not a fixed or statutory position, not a benefit category exclusion. The conclusion seems to be based on today's analysis of utility data.
##
This is a concensus paper, Petrella 2026 (not cited in LCD), abstract in full:
Amyloid-targeted monoclonal antibody therapies have introduced a new era in the treatment of early Alzheimer disease. However, their use has increased the importance of detecting and monitoring amyloid-related imaging abnormalities (ARIA) on MRI, as such findings may influence treatment continuation, dose modification, and patient safety assessment. As anti-amyloid therapies expand into routine practice, increasing surveillance MRI volumes, interreader variability, and the potential for missed subtle abnormalities have generated interest in artificial intelligence (AI)–based clinical decision support tools.
A multidisciplinary panel of neuroradiologists and Alzheimer disease clinicians examined the extent of evidence supporting clinical implementation of AI-assisted ARIA detection tools, these tools' safe integration into practice, and remaining evidence gaps. The panel concluded that AI-assisted ARIA detection is likely to enhance patient safety when used as clinical decision support within a radiologist-in-the-loop framework.
Panelists also noted substantial variation among commercially available tools in regulatory status, technical capabilities, and validation evidence. Moreover, they emphasized the need for further studies to assess the impact of improved detection on clinical outcomes. Overall, the panel supported conditional implementation with radiologist oversight, ongoing quality assurance, and prospective monitoring of clinical performance.
###
Here is an AI response to the LCD, in 350 words:
Palmetto’s proposed LCD carefully attempts to appraise a large, varied literature, including distinct clinical applications and individual products. The concern is whether its categorical noncoverage conclusion adequately reflects the evidence, particularly for AI-assisted monitoring of amyloid-related imaging abnormalities (ARIA).
The LCD reports that AI assistance improved radiologists’ sensitivity from 71% to 87% for ARIA-E and from 69% to 79% for ARIA-H. It questions whether these gains improve outcomes, noting that additional findings often involve mild abnormalities that may not require stopping treatment. But continued treatment does not make detection irrelevant: identifying abnormalities can affect surveillance intensity and subsequent management. These potential benefits require assessment alongside false positives and unnecessary treatment interruptions.
Petrella and colleagues’ 2026 multidisciplinary consensus, absent from the bibliography, supports conditional clinical implementation with radiologist oversight, quality assurance, and prospective monitoring. Its position is guarded: products differ in regulatory status, capabilities, and validation, and improved patient outcomes remain unproven. This consensus does not establish a reimbursement entitlement, but it offers a different judgment about how to act under uncertainty.
That judgment is the central issue. The LCD treats uncertainty primarily as a reason to withhold coverage. It gives less attention to whether withholding validated assistance perpetuates avoidable detection errors in care Medicare already purchases. Neither possibility should simply be assumed; both belong in the assessment.
There is also a policy-design problem. As written, this is a noncoverage determination, not a foundational framework for evaluating future coverage. Even an application that subsequently develops persuasive evidence would face an explicit exclusion requiring LCD revision.
A better approach would establish potential coverage categories and the evidentiary requirements for each, even if no application qualifies today. The MRD LCD offers a useful precedent: it established categories of potential coverage before every category had a qualifying test. Defining a pathway does not require lowering the threshold.
For brain MRI software, that framework could specify requirements for clinical validation, demonstrated management relevance or patient benefit, generalizability, and physician oversight. Individual applications could then undergo the MAC’s technology assessment against those published standards.
The recommendation is to make the LCD foundational and durable: retain noncoverage for applications that fail its requirements, while allowing those that subsequently satisfy them to qualify within the existing framework. Revisions would remain appropriate when the framework itself needs changing, rather than becoming necessary whenever a product’s evidence improves.
###
###
What do you think of the Bibliography?
###
The bibliography’s distribution shows two dominant centers of attention: Alzheimer’s disease/ARIA and multiple sclerosis, with a substantial additional emphasis on the reliability of automated brain volumetry.
The Alzheimer’s/ARIA references cover treatment recommendations, imaging surveillance, severity scales, reader performance, and dementia-related volume measurements. Multiple sclerosis receives another extensive cluster addressing atrophy, disability correlations, automated diagnosis, and possible effects on treatment decisions. Smaller clusters address brain tumors, epilepsy, and traumatic brain injury. FDA clearance records establish the regulatory context for individual products.
Across these disease categories, numerous references examine agreement between software packages, reproducibility, normative populations, and scanner or acquisition effects. Thus, the bibliography is heavily concerned with whether measurements are dependable and clinically interpretable; comparatively little directly tests whether using the software changes care and improves outcomes.
That distribution helps explain the MAC’s conclusions, but also creates an interpretive challenge: the extensive literature on brain-volume measurement should not disproportionately determine judgments about a different task, such as detecting treatment-emergent ARIA. The important question is how much relevant evidence supports each application, rather than how large the combined bibliography appears.
###
###
Mindset of Author
###
My reading is a skeptical evidence reviewer who sees their primary responsibility as preventing premature coverage. The document supports that interpretation more strongly than either “hatchet job” or “enthusiastic but carefully balanced assessment.”
The author does substantial work, distinguishes applications, and reports favorable results—including improved ARIA detection. That willingness to present positive findings argues against a crude prosecution brief. The skepticism also has substance: software agreement, diagnostic accuracy, and patient benefit are different accomplishments.
However, the burden of uncertainty consistently falls on the technology. Positive studies establish promise; limitations establish reasons to withhold coverage. The author gives less attention to the competing uncertainty: whether withholding assistance leaves avoidable errors in ordinary practice. Expert interpretation functions as the practical baseline, although the cited evidence shows that unaided readers miss findings.
One revealing feature is the treatment of physician oversight. The author correctly notes that software cannot replace qualified radiologists, but its placement among limitations gives an intended feature of assistive technology a somewhat negative cast. Similarly, calling out manufacturer involvement is appropriate, but the decisive question should remain how sponsorship affects the credibility of particular results.
I would characterize the stance as methodologically serious, institutionally conservative, and more comfortable identifying insufficiency than defining sufficiency. It reads as though the governing question is, “Has the manufacturer earned coverage?” Less prominent is, “Under what circumstances could this improve the care Medicare already purchases?”
That is an interpretation of the document’s framing, not the author’s private motives. The strongest criticism is an asymmetric treatment of uncertainty—not dishonesty, laziness, or failure to review the literature.
###
Comparing February 2 026 CGS FInAL and October 2026 Palmetto Draft
Yes—and you wrote about it on January 22, 2026: “Two MACs Block Coverage of AI in Radiology.” Your post identified both CGS and NGS. discoveriesinhealthpolicy.com
Both earlier policies have the same title, “Automated Detection and Quantification of Brain MRIs.”
| MAC | LCD | Timing |
|---|---|---|
| CGS—Ohio and Kentucky | L40224 | Proposed September 25, 2025; final notice December 4; effective January 19, 2026. |
| NGS | DL40332 | Proposed January 22, 2026; comments closed March 8. |
| Noridian—another concurrent proposal | DL40454 | Proposed October 1, 2026, alongside Palmetto; comments close November 14. |
The dates are documented in CMS’s policy records. cms.gov
These are direct predecessors, not merely related policies: the earlier documents take the same broad noncoverage position for AI-assisted brain detection and quantification. cms.gov
One striking detail you flagged in January: CGS received no public comments. Its official response-to-comments article confirms that. cms.gov
####
Yes—there are substantial additions to the evidence review, but no meaningful change in the coverage decision. Palmetto’s October proposal is an expanded and partly revised version of the earlier CGS policy, rather than simply a contractor-name substitution.
| Feature | CGS final, effective January 2026 | Palmetto proposed, October 2026 |
|---|---|---|
| Numbered references | 54 | 75 |
| “Summary of Evidence” length | Approximately 4,400 words | Approximately 6,900 words—about 57% longer |
| Organization | Primarily disease-based, followed by comparisons between technologies | Retains that structure and adds a substantial “Technologies” section, with separate NeuroQuant, Icobrain, Icobrain Aria, and DeepBrain subsections |
| Coverage | Broad noncoverage | Same broad noncoverage |
| Central justification | Insufficient clinical utility/validity; standardization and generalizability concerns | Essentially unchanged |
The word counts cover the evidence section only, excluding the bibliography, administrative material, and concluding rationale. l40224 CgS MAC mRI AI Feb Proposed LCD - Automated Detect…
The meaningful changes are:
A larger bibliography, mostly adding previously available literature. The original 54 references are retained, although many are renumbered. The additional 21 entries represent 20 additional works and one duplicate: Klein’s ARIA severity-scale paper appears as both references 17 and 31. Many additions date from 2012–2024, so this is substantially a broader review of existing literature, rather than simply nine months of new publications.
More product-specific assessment. The new Technologies section discusses validation, diagnostic performance, software-version effects, acquisition parameters, and limitations of particular products. Some material repeats or expands studies already discussed elsewhere; the increased length does not represent an equivalent increase in independent evidence.
Additional recent evidence. Examples include two 2025 MS papers on atrophy, disability, and methodological consistency, and Bregman 2026 on AI-derived MRI predictors during lecanemab treatment. Palmetto acknowledges Bregman’s potential relevance to risk stratification but emphasizes that it did not assess effects on physician decisions or patient management.
Expanded ARIA severity-scale discussion. CGS said the severity classification lacked peer-reviewed publication beyond a poster. Palmetto adds the 2023 Bracoud validation paper, softens that assertion, and discusses correlations between scales and the need for further validation. Proposed LCD - Automated Detect…
The most consequential change in the concluding rationale concerns ARIA. CGS followed its description of the AI reader study with objections about non-U.S. readers, an unapproved drug, and a mild-to-moderate AD population—apparently mixing limitations of the older bapineuzumab studies into the discussion of the AI study. Palmetto removes those assertions and substitutes concerns about population diversity, image quality, acquisition protocols, and the software’s assistive role. That is a substantive revision to the argument. l40224 CgS MAC mRI AI Feb Proposed LCD - Automated Detect…
My assessment: expanded evidence review, some corrected or refined reasoning, unchanged policy outcome. The favorable ARIA sensitivity figures were already in CGS’s final LCD. Palmetto adds detail and literature but still supplies neither a covered exception nor a prospective qualification framework.
####
###
* The CGS draft and final LCDs were turned around nearly instantly in Dec/Jan/Feb 2025-2026.
