Automated Detection and Quantification of Brain MRIs

Title XVIII of the Social Security Act, §1862 (a)(1)(A) allows coverage and payment for only those services that are considered to be reasonable and necessary for the diagnosis or treatment of illness or injury or to improve the functioning of a malformed body member.

Title XVIII of the Social Security Act, §1862 (a)(1)(D) Items and services related to research and experimentation.

Title XVIII of the Social Security Act, §1862 (a)(7) states Medicare will not cover any services or procedures associated with routine physical checkups.

Title XVIII of the Social Security Act, §1833 (e) prohibits Medicare payment for any claim which lacks the necessary information to process the claim.

42 CFR §410.32 indicates that diagnostic tests may only be ordered by the treating physician (or other treating practitioner acting within the scope of his or her license and Medicare requirements).

CMS Internet-Only Manual, Pub. 100-03, Medicare National Coverage Determinations (NCD) Manual, Chapter 1, Part 4, §220 Radiology.

The Protecting Access to Medicare Act (PAMA) of 2014, Section 218(b), established a new program to increase the rate of appropriate advanced diagnostic imaging services provided to Medicare beneficiaries.

42 CFR §414.92 codifies the Appropriate use Criteria Program policies. Data quality and AI model development with external testing.

CMS Internet-Only Manual, Pub. 100-03, Medicare National Coverage Determinations (NCD) Manual, Chapter 1, Part 4, §200.3 Monoclonal Antibodies Directed Against Amyloid for the Treatment of Alzheimer's Disease (AD)

This is a non-coverage policy for artificial intelligence (AI) assistive software tools for automated detection and quantification of the brain. 

Background

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.¹

Structural imaging has the potential to reduce variance and improve diagnostic and prognostic inferences from magnetic resonance imaging (MRI) scans. However, the tools must be trained and validated in a manner that provides generalizability to broader populations. Most available tools rely on multivariate machine learning approaches, offering promise for personalized medicine, but currently available studies are limited by small sample sizes, methodological limitations, and lack of longitudinal validation. Study authors emphasize that imaging outputs must be interpreted within the broader clinical context and that future progress will depend on addressing challenges such as biological heterogeneity, scanner variability, data integration, and robust validation.² Of eight tools identified, two were not approved for medical use and one had no associated references. Most tools were validated using a small number of cases and/or a single dataset. The tools were compared based on the number of validation methods conducted. None of the tools accounted for intrascanner variability resulting from differences in scanners, magnetic field strength, or acquisition parameters; therefore, the tools lack generalizability. The authors concluded that most available tools use multivariate machine learning methods and have the potential to open new possibilities in personalized medicine. However, they cautioned that results should be interpreted with vigilance because of study limitations, especially those related to small sample size and poor methodology. They also cautioned that results must be interpreted in light of the patient’s clinical history and symptomatology.

The American Society of Functional Neuroradiology (ASFNR) and the American Society of Neuroradiology (ASNR) acknowledge the challenges of AI in neurology and created an Artificial Intelligence Workshop Technology Workgroup.³ This group published a critical appraisal of AI-enabled imaging tools using the levels of evidence system demonstrated in the American Journal of Neuroradiology. They call for critical appraisal of AI-enabled imaging tools throughout the life cycle, from development to implementation, using systematic, standardized, and objective approaches that can verify both the technical and clinical efficacy of the tool. A challenge in developing AI models is accessing comprehensive and large datasets that can be used to train the technology. These datasets should represent the intended population and provide a diverse group from which the data may be extrapolated. This paper provides a resource for clinicians to aid in the critical assessment of AI technologies to ensure safe and effective implementation into healthcare practices. ³

Food and Drug Administration (FDA)-cleared devices as of the publication date of this LCD include:

• NeuroQuant^TM Medical Image Processing Software is registered as a Class II device under FDA 510(k) and intended for “automatic labeling, visualization and volumetric quantification of segmentable brain structures from a set of magnetic resonance images (MRI)”. NeuroQuant 4.0 uses AI modalities of machine learning and deep learning to aid in identifying complex patterns in imaging data.⁴
• Icobrain is registered as a Class II device under FDA 510(k) pathway.^7-10 It is intended for “automatic labeling, visualization and volumetric quantification of segmental brain structures from a set of MRI images”. The predicate device is NeuroQuant.
• Icobrain aria is registered as a Class II device under FDA 510(k) pathway and described as a software-only device for assisting radiologists with the detection and quantification of amyloid-related imaging abnormalities (ARIA) on brain MRI scans in patients undergoing an amyloid-beta directed antibody therapy. ⁵ Icobrain aria automatically processes inputs from brain MRI scans from two time points and calculates the ARIA-E (edema/sulcal effusion): the length of the longest axis computed from the segmented ARIA-E abnormalities, and the number of brain sites affected by ARIA-E; and ARIA-H (hemorrhage/superficial siderosis): which is the count of stable and new T2*-GRE hypo-intensities indicated as microhemorrhages or superficial siderosis.⁵ Using these measurements the ARIA radiographic severity is automatically derived based on deep learning technology and reported electronically. The intended use of the device is “as a computer-assisted detection and diagnosis software to be used as a concurrent reading aid to help trained radiologists in the detection, assessment, and characterization of ARIA. The software provides information about the presence, location, size, severity and changes of ARIA-E and ARIA-H. Patient management decisions should not be made solely based on analysis by Icobrain aria.” The device is not intended to replace radiologist review of images or clinical judgment and is not intended to be used to segment macro hemorrhages (diameter 10mm or more).⁶
• DeepBrain is registered as a Class II device under FDA 510(k) pathway.¹¹ The device is intended for “automatic labeling, quantification and visualization of segmental brain structures from MRI images.” It is intended to be used by trained health professionals.
• Siemens Morphometry Analysis is registered as a Class II device under FDA 510(k) pathway.¹² This product is a syngo-based post-acquisition imaging processing software for “viewing, manipulating, evaluating and analyzing MRI, MR PET, CT, PE, CT-PET and MR spectra using deep learning algorithms”.

Other FDA-cleared devices were not found at the time of this LCD literature search.

The literature search for evidence related to quantitative analysis of brain MRI was conducted using PubMed and Hayes with the following search terms: Alzheimer’s, ARIA, imaging, artificial intelligence or AI, automated or software or computation or deep learning, machine learning or artificial neural network, multiple sclerosis, MRI and AI, brain or neurology, AI or MRI. Searches for known commercially available products in the United States (U.S.) included NeuroQuant, NeuroGage, Icobrain, Icobrain aria, Jung Diagnostic, Quantib, Qure, and volBrain. No randomized controlled trials (RCTs) were identified. Unpublished reports, posters, abstracts, case reports, and small case series were omitted from the review unless no other evidence was available for consideration. Review papers were used for background and summary but were not considered for evidence review. There were no guidelines or recommendations on the use of AI assistive software tools for automated detection and quantification of the brain.

Alzheimer’s Disease

Amyloid beta (Aβ)-directed antibody therapies, such as aducanumab*, lecanemab, and donanemab, are approved in the U.S. for the treatment of patients with mild cognitive dementia due to Alzheimer’s disease. There is evidence of slowing disease progression and improvement in clinical outcomes for patients treated for mild cognitive impairment due to Alzheimer’s dementia (AD); however, it is known that Aβ-directed antibody therapies increase the risk of ARIA and ARIA-related complications.¹³ ARIA complications include autoimmune or inflammatory conditions, seizures, and other disorders associated with extensive white matter pathology. Symptoms and signs of ARIA can include new focal neurological signs, headache, confusion, altered mental status, dizziness, nausea, vomiting, fatigue, blurred vision, vision disturbances, gait disturbance, and/or seizure.

Radiologists have become familiar with the appearance of amyloid-related imaging abnormalities and have developed imaging protocols to classify abnormalities related to Aβ-directed antibody therapies. Clinicians must also determine the optimal pathways for the management of ARIA.¹⁴ This is an area of ongoing investigation, and several studies have contributed to the current knowledge of these challenges, but the literature remains sparse.^14-18

Multiple grading schemes to determine the severity of ARIA have been proposed.^15-18 ARIA radiographic severity is categorized as mild, moderate, or severe, and this classification system was used in the pivotal clinical trials for anti-amyloid immunotherapies and by the FDA for drug approval. Although it is becoming the accepted standard for classification, there has been little peer-reviewed literature as of the date of this LCD.^19,20,30,31 A comparison of the Barkhof Grand Total Scale (BGTS) and the 3- or 5-point Severity Scales of ARIA-E, SSAE-3 and SSAE-5 respectively, demonstrated a high degree of correlation between the scales. Despite the strong correlation between SSAE-3, SSAE-5, and the BGTS for assessing ARIA-E, these approaches are not yet ready for routine clinical adoption. While threshold alignment across scales suggests the potential to translate existing BGTS-based dosing suspension rules to SSAE-based criteria, further validation is required.¹⁷

To monitor for ARIA, the Appropriate Use Criteria for aducanumab* recommend MRI before the 5^th, 7^th, 9^th, and 12^th infusions to improve detection.²¹ The criteria recommend discontinuation of aducanumab* for any macrohemorrhage, more than one area of superficial siderosis, more than 10 microhemorrhages occurring since the initiation of treatment, more than two episodes of ARIA, severe symptoms of ARIA, or development of any medical condition requiring anticoagulation. The protocol allows continuation of aducanumab* for mild ARIA-E or ARIA-H with monthly MRI monitoring and discontinuation for worsening symptoms.

The Appropriate Use Criteria for lecanemab recommend obtaining MRI scans at baseline and before the 5^th, 7^th, 14^th, and 26^th infusions. The authors explain that 81% of ARIA-E cases occur early and resolve spontaneously within 4 months of radiographic detection.²² The criteria allow continuation of aducanumab* for mild ARIA-E or ARIA-H with monthly MRI monitoring and discontinuation for worsening symptoms. Once the ARIA resolves or stabilizes, monthly imaging can be discontinued. The criteria state that the imaging should be read by knowledgeable MRI readers proficient in the detection and interpretation of ARIA or clinicians skilled in the identification and interpretation of cerebrovascular lesions and ARIA.

The Appropriate Use Recommendations for donanemab recommend MRI before the 2^nd, 3^rd, 4^th, and 7^th infusions, and before the 12^th infusion in patients at high risk for ARIA. The protocol allows continuation of aducanumab* for mild ARIA-E or ARIA-H with monthly MRI monitoring and discontinuation for worsening symptoms.⁶

Since ARIA is a new entity, standard radiographic interpretation of ARIA is still under development. Those interpreting the imaging need education and training to ensure accuracy and consistency in reporting. The American College of Radiology, the Alzheimer’s Association, and the Radiological Society of North America are all offering training and continued medical education.¹⁹

Radiology

A retrospective report on ARIA reviewed imaging from 262 subjects in Phase 2 studies of patients with mild-to-moderate AD treated with bapineuzumab, a humanized monoclonal antibody against amyloid β. Two neuroradiologists independently reviewed 2,572 MRI scans from 262 participants. The readers were masked to the patients’ treatment arms. Patients were included in the risk analysis (n=210) if they did not have evidence of ARIA-E on pretreatment MRI, had received bapineuzumab, and had at least one MRI scan after treatment. Thirty-six patients (17%) developed ARIA-E during treatment with bapineuzumab; 28 patients (78%) were asymptomatic, while eight were symptomatic. Fifteen patients with ARIA-E detected on reread of MRI (42%) were not detected previously. All of these patients were asymptomatic and had fewer brain regions involved (mean 1.3, SD=0.5) than patients identified during the clinical studies (mean 2.6, SD=2.4; p=0.0193). Thirteen patients whose ARIA-E findings were not detected during the clinical trial continued bapineuzumab infusions for up to 2 years and remained asymptomatic. This study suggests that ARIA encompasses a spectrum of imaging findings with variable clinical relevance, including many subtle and asymptomatic cases that were not detected during clinical trials. The frequent co-occurrence of ARIA-E and ARIA-H supports a shared underlying pathophysiology, potentially related to transient vascular permeability and vascular amyloid. Associations with APOE ε4 status, treatment dose, and amyloid imaging further indicate a link between ARIA and amyloid burden and clearance, underscoring the need for continued investigation in ongoing amyloid-modifying therapy studies. ¹⁸

Using the same study population, investigators sought to describe the imaging characteristics of identified ARIA-E and ARIA-H.¹⁵ The rate of ARIA-H was reported as 12.4% (26/210). They also found that 49% of those with ARIA-E had associated ARIA-H. The authors concluded that this may suggest a common pathophysiologic mechanism. All scans were reviewed by local MR imaging readers and subsequently independently reviewed by the same two neuroradiologists as part of the study protocol. They reported an inter-reader kappa value of 0.76, indicating high inter-reader reliability, with 94% agreement between neuroradiologists regarding the presence or absence of ARIA-E.

Using this same population, a retrospective analysis of 242 patients found that ARIA-E was detected more frequently by trained neuroradiologists than by local site radiologists.²³ MRI was performed in patients with mild-to-moderate AD in a Phase III trial of bapineuzumab. Seventy-six cases of ARIA-E were not detected on the initial read and were reported on the final MRI review by the expert radiologist, including 51 cases not identified by central or local readers. These cases represented low radiologic severity. A final read analysis found that readers’ ability to detect ARIA-E improved as study duration increased, resulting in the majority of later ARIA-E occurrences being identified by the local site radiologist. This suggests that the ability to detect ARIA improved with increasing experience. It is unclear whether the outside readers were using the same imaging criteria, what duration and type of training they received, or whether this finding resulted in any treatment changes between groups. This supports the need for appropriate training for radiologists reading these studies and continued standardization of findings to ensure consistency between readers. The clinical significance of this finding has not been determined, as none of those with mild ARIA-E had symptoms despite continuation of therapy, and the sample size was too small to generalize to a larger population.

A volumetric analysis of structural MRI images of the brain was tested and reported to have a high correlation with independent computer-aided manual segmentation for detection of atrophy changes seen in mild AD.²⁴ Using the Open Access Series of Imaging Studies database, 40 subjects with mild probable AD were compared with healthy controls. These images were processed by the NeuroQuant software package. The investigators reported that volumetric results obtained by the software showed a high correlation and could benefit the evaluation of brain atrophy. The study is limited by its very small sample size, lack of human subjects and correlation with expert readers in the same population, and represents very low-quality evidence.

A study reviewing MRIs from 122 patients with dementia compared NeuroQuant readings with radiology readings in an effort to determine whether the automated software could identify whether dementia was AD or another type. The study involved patients with various cognitive conditions, including Dementia of the Alzheimer’s Type (DAT), other dementia types, mild cognitive impairment (MCI), and subjective cognitive impairment (SCI). NeuroQuant had moderate ability to differentiate DAT from non-dementia patients (SCI + MCI), with the hippocampus and amygdala showing moderate discriminatory power (AUC 0.80 and AUC 0.79, respectively). They concluded that the software could not be used alone because changes in brain segments were not specific for AD.²⁵

A prospective study evaluated brain MRIs from 40 patients across six scanners at five institutions using both NeuroQuant quantitative analysis and neuroradiologist readings. Image processing was conducted with FAST-DL (FAST = accelerated scan, DL = deep learning), a Digital Imaging and Communications in Medicine (DICOM)-based convolutional neural network-dependent deep learning AI enhancement software product called SubtleMR. Clinical classification performance was compared with standard-of-care scans, FAST-DL, and NeuroQuant. The authors reported that FAST-DL was statistically superior to standard-of-care in subjective image quality for perceived signal-to-noise ratio (SNR), sharpness, imaging artifacts, anatomic/lesion conspicuity, and image contrast (all P values < 0.008), despite a 60% reduction in sequence scan time. They concluded that deep learning can provide 60% faster image acquisition with statistically perceived image quality and accuracy comparable to standard-of-care scans.²⁶ This study is limited by small sample size, use of vendor-based software for comparison, risk of bias, and investigator conflicts of interest.

Clinical Validity (or Technical Efficacy)

Sima et al.²⁷ conducted a diagnostic study to assess the clinical performance of an AI–based software tool for assisting radiological interpretation of brain MRI scans in patients monitored for ARIA. This study enrolled 16 U.S. Board of Radiology certified radiologists to perform radiologic image interpretation with the software, referred to as assisted reads, and without the software, referred to as unassisted reads. A total of 199 cases, where each case consisted of a pre-dosing baseline and a postdosing follow-up MRI of patients from aducanumab* clinical trials PRIME (NCT01677572), EMERGE (NCT02484547), and ENGAGE (NCT02477800) were retrospectively evaluated. End points were the difference in diagnostic accuracy between assisted and unassisted detection of ARIA-E and ARIA-H independently, assessed using the area under the receiver operating characteristic curve (AUC). Demographics included mean age was 70.4 (7.2) years; 105 (52.8%) were female; 23 (11.6%) were Asian, 1 (0.5%) was Black, 157 (78.9%) were White, and 18 (9.0%) were other or unreported race and ethnicity. Among the 16 radiological readers included, two were specialized neuroradiologists (12.5%), 11 were male individuals (68.8%), seven were individuals working in academic hospitals (43.8%), and they had a mean (SD) of 9.5 (5.1) years of experience. Radiologists assisted by the software were significantly superior in detecting ARIA than unassisted radiologists, with a mean assisted AUC of 0.87 (95% CI, 0.84-0.91) for ARIA-E detection (AUC improvement of 0.05 [95% CI, 0.02-0.08]; P = 0.001) and 0.83 (95% CI, 0.78-0.87) for ARIA-H detection (AUC improvement of 0.04 [95% CI, 0.02-0.07]; P = 0.001). Sensitivity was higher in assisted reading compared with unassisted reading (87% vs 71% for ARIA-E detection; 79% vs 69% for ARIA-H detection). Specificity remained above 80% for the detection of both ARIA types. The software had the greatest improvement in detection of mild cases (70% compared to 47%). The unassisted readers distinguished ARIA grades well but had higher inter-reader agreement with software assistance. Time for reading was similar in both groups. The authors concluded that radiological reading performance for ARIA detection and diagnosis was significantly better when using the AI-based assistive software. The study is limited by concerns for indirectness (small sample size), lack of generalizability (limited representation in the population), and risk of software errors due to computer learning from individual readers which may have been trained on discrepant data. The software was not trained to detect cerebral hemorrhages larger than 1 cm. The software assists but does not replace the need for qualified radiologists to read the studies. An additional validation study compared Icobrain to FreeSurfer with comparable results.²⁸ It demonstrated that Icobrain dm provided more accurate and reliable segmented brain volume measurements than FreeSurfer, particularly evident in hippocampal segmentation, which showed higher Dice similarity coefficients (0.86–0.88 vs 0.80–0.83). The study concluded that Icobrain dm could significantly improve clinical diagnostics for AD, with a recommendation for further research into its effectiveness at earlier stages of the disease and differentiation from other dementia types. However, the lead author is employed by the device maker.

Another study compared the MRIs from healthy controls (n=90) to those with subjective cognitive decline (n=930), mild cognitive impairment (n=357), and AD (n=820). Icobrain dm not only achieved a high area under the curve (AUC) of 0.914 for distinguishing healthy controls from AD dementia patients but also outperformed FreeSurfer in terms of diagnostic performance and processing time. Icobrain dm demonstrated an impressive specificity of 83.0% and sensitivity of 86.3%. Neuroquant showed effectiveness in distinguishing Alzheimer's type dementia from non-dementia but was less reliable in differentiating it from other dementia types.²⁹ FreeSurfer is not available clinically.

An empirical study in South Korea analyzed MRI for 98 patients with AD using VUNO Med-DeepBrain AD (DBAD) deep learning algorithm. They compared the results of the DBAD imaging reads to that of three panels of medical experts and reported comparable accuracy (87.1% for DBAD and 84.3% for ME), sensitivity (93.3% for DBAD and 80.0% for ME), and specificity (85.5% for DBAD and 85.5% for ME).³²

Clinical Utility (or Clinical Efficacy)

There are no published studies to date on clinical utility. The software was found to be particularly helpful in improving detection of mild cases of ARIA-E, however it is not determined if this improves patient outcomes.²⁷ The Appropriate Use Recommendations allows continuation of aducanumab* for mild ARIA of ARIA-E or ARIA-H with monthly MRI monitoring and discontinuation for worsening symptoms. There is a lack of evidence on if the continuation of medication in these cases correlates with improved outcomes or if this increases the risk of serious adverse events.

Multiple Sclerosis

Multiple sclerosis (MS) is an autoimmune demyelinating disease impacting the central nervous system and diagnosis is made by MRI findings, laboratory findings and clinical data. This has led to investigations into developing machine learning tools to aid in the diagnosis of MS.

Several studies examine the connection between brain atrophy and disability progression in MS. The studies call for standardized assessment techniques to effectively utilize brain volume measurements in managing MS.^33-38

Several review papers have identified multiple investigations for AI models and the potential this technology may bring but acknowledge these investigations did not yield a clinically usable model.^39-43

A 2018 systematic review identified English language articles published between 2000 and 2018, screening 614 records and analyzing 85 studies, with 30 meeting inclusion criteria related to system scope, reasoning methods, and evaluation. The review found that although all reasoning methods demonstrated high efficiency in MS diagnosis, performance varied by methodological characteristics, leading many studies to combine approaches to address limitations and improve diagnostic accuracy, highlighting the broader potential of computational methods in MS clinical practice.⁴⁴

A 2022 systematic review included 38 studies focusing on deep learning or AI to analyze any modalities with purpose of diagnosing MS. ³⁶ The majority of these studies utilized MRI (20 studies), with others employing OCT (6 studies), serum/CSF markers (6 studies), and motor function assessments (3 studies). AI, especially using MRI, shows promise in improving MS diagnosis, monitoring, and treatment methodologies. They conclude this is a growing field and can result in drastic improvements in the future with larger multicenter studies across diverse populations and MS subtypes needed to validate these findings before clinical adoption .³⁶ Another systematic review with meta-analysis from the same group in 2023 included 41 articles (n=5989) and reported a high precision in MS diagnosis for AI studies (95%CI: 88%, 97%) suggesting that AI can aid the clinician in accurate diagnosis of MS and has strong potential to improve MS diagnosis, particularly when based on MRI data.³⁷ The meta-analysis is limited by very high heterogeneity with overall I²=93% limiting the validity of these results. The authors conclude that further research is needed to develop robust and generalizable algorithms that adequately address preprocessing, data augmentation, feature selection, optimization, regularization, and model design and architecture.

Another 2022 systematic review included 66 papers which addressed developing classifiers for MS identification or measuring its progression. They also acknowledge the potential benefits of this approach if applied appropriately and provide guidance for further research.¹

A retrospective study compared the MRIs for patients with MS analyzed with Icobrain software platform for 6826 MRIs with 1207 MRI pairs meeting inclusion criteria. The investigators reported that Icobrain could be utilized for percentage brain volume change (PBVC) based on strict selection criteria. ⁴⁷ Another explored the potential role of Icobrain and use of an MS app MS to inform treatment changes in a small population. Given the heterogeneity of MS and the need for more personalized, data driven, and standardized care pathways, the findings demonstrate the potential and feasibility of linking multiple digital and AI tools into an overarching MS care pathway, while also highlighting limitations related to real world data reliance, patient engagement, and clinical implementation.⁴⁸

Another 2021 study introduced a digital care management platform for MS, comprising the Icompanion app for patients and Icobrain ms for MRI analysis. The study highlighted the platform's capacity to enhance disease activity detection through patient-reported outcomes and improved MRI analytics, with patient interest in such digital tools peaking at 95.6% for MS apps and 98.2% for tracking MRI changes. However, the study relied on patient-reported outcomes, and the effectiveness of the platforms were dependent on the patients’ digital literacy. The study was observational limiting generalizability.⁴⁸

Volumetric data for patients with MS were analyzed on the same and different-scanner MRI pairs. Of 6826 MRIs, 85% had appropriate volumetric sequences and 4446 serial MRI pairs were analyzed and 3335 (75%) met inclusion criteria. The PBVC of the included MRI pairs showed variance of 0.78 % for same-scanner pairs and 0.80 % for different-scanner pairs, but further selection of included MRI pairs with the best variance resulted in 1885 (42%) MRI pairs with PBVC variance of 0.34%. The authors acknowledge the challenges and limitations for brain volumetry measurements and need for standardization to perform adequately. The authors conclude that Icobrain should be utilized for PBVC determination only on selected MRIs with best alignment similarity and with strict selection criteria for the included MRI pairs to reduce PBVC variability.⁴⁷

The integration of AI into the MS clinical setting has strong potential to transform diagnosis and prognosis, improving patient outcomes through enhanced decision support, efficiency, and care effectiveness. Continued research, data sharing, advances in computational technologies, and progress in interpretability, validation, regulatory, and ethical frameworks will be critical to enabling trustworthy and effective clinical adoption.⁴¹

Brain Tumors

Two poster abstracts and several review papers were identified.^49,50,51

This retrospective study developed and validated a deep learning model using multimodal MRI to segment enhancing and nonenhancing cellular tumor in patients with glioblastoma and to predict overall and progression free survival. While the model showed strong performance across internal and external cohorts, its reliance on retrospective data, advanced imaging sequences, and expert annotated reference standards highlights limitations related to generalizability, data availability, and real-world clinical implementation.⁵⁰

Another abstract outlines the evolution of diffusion weighted imaging from its early success in stroke diagnosis to its expanded role in oncologic imaging, particularly in neuro oncology, while addressing common misconceptions about the biophysical basis of diffusion contrast. ⁵¹

A novel AI-driven application to aid in brain tumor detection from MRI images reports on the development of EfficientNetB2. This report focuses on the proposed technology and evaluation of performance in non-clinical setting.⁵² EfficentNetB2 is not FDA-cleared as a medical device.

Epilepsy

Comparison of MRI images read by neuroradiologists and analyzed with NeuroQuant software in 144 patients with temporal lobe epilepsy (TLE) was performed. The investigators found similar specificity to neuroradiologist visual MR imaging analysis (90.4% versus 91.6%; P = 0.99) but a lower sensitivity (69.0% versus 93.0%, P < 0.001). Visual MR imaging analysis by a neuroradiologist with expertise in epilepsy had a higher sensitivity than NeuroQuant analysis, likely due to the inability of NeuroQuant to evaluate changes in hippocampal T2 signal or architecture. The positive predictive value of NeuroQuant analysis was comparable with visual MR imaging analysis (84.0% vs 89.1%), whereas the negative predictive value was not comparable (79.8% vs 95.0%). While NeuroQuant showed comparable specificity and may be useful when results are positive, negative findings require further expert evaluation to avoid false negatives, limiting its use as a standalone diagnostic tool.⁵³ They conclude the technology may aid in evaluations when a neuroradiologist is not available, however product information states this is intended as an adjuvant, not replacement for the radiologist.

A prospective study measured volumetric MRI imaging data for 34 patients with TLE and compared to 116 control subjects.⁵⁴ Structural volumes were calculated using automated quantitative MRI imaging analysis software (NeuroQuant). Results of the quantitative MRI imaging were compared with visual detection of atrophy and histological specimens if available. Quantitative MRI imaging results compared to visual inspection of the volumetric MRI imaging studies by two experienced neuroradiologists had a concordance between hippocampal asymmetry (91-97%). They reported the software discriminated patients with TLE from control subjects with high sensitivity (86.7%–89.5%) and specificity (92.2%–94.1%). The authors conclude that the software can provide “an expert eye” in centers that lack expertise, however the FDA indications for this software indicates it is intended to be an adjunct to the radiologist reader, not a substitute. Limitations of the study include lack of generalizability to non-expert readers, small sample size, lack of confirmative histological confirmation available for 12 patients (35%).

A retrospective report included 36 patients with mesial temporal sclerosis (MTS) which is important to detect for TLE as it often guides surgical intervention. One of the features of MTS is hippocampal volume loss. Using electronic medical records, researchers scanned patients with proven MTS and analyzed the imaging with volumetric assessment software (NeuroQuant). They reported an estimated accuracy of the neuroradiologist as 72.6% with Kappa statistic of 0.512 (95% CI, 0.388–0.787). They conclude that the NeuroQuant software compared favorably with trained neuroradiologists in predicting MTS.⁵⁵

Other literature identified was limited to review papers, case reports and series and not included in this assessment.

Traumatic Brain Injury

Twenty MRI images from patients with mild-to-moderate traumatic brain injury (TBI) were analyzed with NeuroQuant automated software and compared to attending radiologist interpretation. The investigators reported radiologist’s traditional approach found at least one sign of atrophy in 10.0% of patients compared to NeuroQuant finding this in 50.0% of patients concluding higher sensitivity of Neuroquant.⁵⁶ A subsequent expanded study with 24 subjects found similar results.⁵⁷ These studies are limited by very small sample size, and lack of knowledge if the atrophy was caused by TBI or other conditions that can cause similar findings. The authors state “we have never seen an MRI report on a patient that used a qualitative rating scale to assess level of atrophy or ventricular enlargement. With the rapid advances in computer-based technology, instead of focusing on understanding and developing the approach based on qualitative ratings, it may be more advantageous to focus on the computer-automated approaches.” In the absence of qualitative ratings, the comparison of the gold standard visual approach and qualitative approach is not established, and to the claim that the software was 50% more accurate is not validated.

Technologies

NeuroQuant

Choi et al. (2024), in a retrospective study, compared NeuroQuant with FreeSurfer and Heuron AD in 78 patients with normal cognition, mild cognitive impairment, and Alzheimer disease. Inter-software agreement was substantial to almost perfect for most brain regions, but significant systematic differences in measured brain volumes were seen for some regions. The authors concluded that volumes produced by different automated platforms should not be used interchangeably in clinical practice.⁵⁸

Ochs et al. (2015) analyzed 60 MRI scans from normal controls, patients with AD, and patients with mild TBI using automated brain volumetric comparisons generated by NeuroQuant and FreeSurfer. They concluded that there was good-to-excellent agreement between the two software programs for most brain regions, although NeuroQuant tended to overestimate volumes compared with FreeSurfer.⁵⁹

Wang et al. (2016), in a validation study, evaluated NeuroQuant and MSmetrix against the reference standard Structural Image Evaluation using Normalization of Atrophy – Cross-sectional (SIENAX) for whole-brain volumetry in 63 patients withMS. NeuroQuant showed excellent precision and good agreement versus SIENAX, although it systematically overestimated brain volume by approximately 5.5% compared with the reference method.⁶⁰

Yim et al. (2021), in a retrospective study, compared NeuroQuant and FreeSurfer volumetric measurements in 80 patients with mild cognitive impairment. They demonstrated good-to-excellent agreement between the two software programs for most brain regions using both slice thicknesses, with slightly better reliability when using 1-mm slices. Differences in slice thickness affected volumetric measurements and agreement between software programs, suggesting that MRI acquisition parameters must be considered when viewing volumetric results.⁶¹

Kim et al. (2021) performed a prospective multicenter study involving 113 participants with healthy cognition, mild cognitive impairment, or Alzheimer's disease to evaluate whether combining NeuroQuant MRI volumetry with tau PET improved diagnostic performance. NeuroQuant-derived regional brain volumes demonstrated good discrimination of Alzheimer's disease, and combining MRI volumetry with tau PET improved performance compared with PET alone, although the combined model did not significantly outperform MRI volumetry alone.⁶²

Stelmokas et al. (2017) evaluated NeuroQuant version 1 and version 2 in 132 healthy older adults and patients with mild cognitive impairment. Although both versions demonstrated clinically meaningful associations between medial temporal lobe volumes and memory performance, volumetric differences were observed between software versions. Age and sex-adjusted z-scores from NeuroQuant v2 provided stronger prediction of memory impairment than intracranial volume-corrected measurements. They note that software version differences may affect longitudinal comparisons.⁶³

Persson et al. (2018) compared automated hippocampal volumetry performed with NeuroQuant with the Scheltens visual medial temporal atrophy (MTA) scale in a cohort of 56 patients with Alzheimer's disease or mild cognitive impairment. There was a strong correlation between NeuroQuant volumetry and visual MTA score. Both tests were similarly effective at distinguishing Alzheimer's disease from non-dementia controls, although NeuroQuant had a slightly higher area under the curve. In terms of overall diagnostic accuracy, the visual rating scale (MTA) was favored.⁶⁴

Ross et al. (2012) in a preliminary study demonstrated that NeuroQuant showed high test–re-test reliability and concluded that NeuroQuant is a reliable and valid method for assessing the anatomic effects of TBI.⁶⁵

Ross et al. (2013), in another study, compared NeuroQuant volumetric analysis against routine radiologist visual interpretation of imaging studies from 20 patients with mild/moderate TBI. NeuroQuant detected brain atrophy in 50% of patients, while radiologist reports documented brain atrophy in only 10% of patients. They concluded that automated volumetric analysis might identify structural changes not visually apparent.⁵⁶

Icobrain

Beadnall et al. (2019), in an observational study, showed excellent agreement between Icobrain long and Structural Image Evaluation using Normalization of Atrophy (SIENA) for longitudinal whole-brain atrophy tracking in MS; however, this comparison reflects analytical agreement and not accuracy, as SIENA was used as the reference standard. They also demonstrated that Icobrain-derived brain atrophy metrics tracked brain volume loss and disease progression longitudinally in MS but only modestly correlated with disability. Evidence supporting clinical utility is limited. The authors concluded that automated measurements of whole-brain atrophy have promise for incorporation into routine MS care but require further optimization before use at the individual patient level.⁶⁶

Simarro et al. (2024) in a retrospective validation study demonstrated that Icobrain-dl attained segmentation accuracy on par with age-specific models and significantly greater reproducibility than comparable software in pediatric and adult populations.⁶⁷

Struyfs et al. (2020) showed superior segmentation accuracy, lower volumetric error, and greater test-retest precision compared with FreeSurfer. They also found excellent discrimination between Alzheimer's disease and cognitively healthy controls (AUC 0.99) using Icobrain-derived volumetric measurements.²⁸

Nguyen et al. (2024) showed that reliable measurement of longitudinal brain volume changes, including with Icobrain, required strict image quality control and scanner alignment, highlighting the need for imaging criteria to be met.⁴⁷

Wittens et al. (2021) published results from the REMEMBER study and demonstrated that Icobrain outperforms FreeSurfer and that combined volumetric measures can be used to achieve high diagnostic accuracy for healthy control vs Alzheimer's disease classification (AUC 0.914; sensitivity 86.3%; specificity 83.0%).²⁹

Van Hecke et al. (2021) demonstrated Icobrain assisted MRI interpretation decreased reader variability and increased sensitivity for detection of MS disease activity and progression from 24% to 76% compared with standard radiology interpretation.⁴⁸

Bregman et al. (2026), in a retrospective study, found that Icobrain-quantified gray matter volume predicted cognition, while microhemorrhage burden predicted ARIA during lecanemab treatment. Icobrain-determined microhemorrhage counts showed excellent agreement with expert readers (ICC 0.89). The authors concluded that Icobrain-derived MRI biomarkers may enhance ARIA risk stratification and identification of patients most likely to benefit from lecanemab treatment; however, they did not assess its impact on physician decision-making or patient management.⁶⁸

Icobrain Aria

Sima et al. (2024) conducted a diagnostic study assessing the clinical performance of Icobrain aria for assisting radiological interpretation of brain MRI scans in patients monitored for ARIA. The study enrolled 16 U.S. Board of Radiology–certified radiologists to perform radiological reading with the software (assisted) and without the software (unassisted). The study included 199 retrospective cases, where each case consisted of a predosing baseline and a postdosing follow-up MRI from patients in aducanumab clinical trials. Radiologists who used the software showed significantly improved detection of ARIA compared with radiologists without AI support. Sensitivity also improved for both ARIA-E and ARIA-H. The authors concluded that radiological reading performance for ARIA detection and diagnosis was significantly better when using the AI-based assistive software (Icobrain aria) and that it could help improve safety monitoring and management of patients with AD treated with amyloid-β–directed monoclonal antibody therapies.²⁷ The evidence is limited by its study design, relatively small sample sizes for some outcomes, lack of assessment for macrohemorrhages, and lack of prospective evidence that AI-assisted interpretation improves management or clinical outcomes for patients.

DeepBrain

Kim et al. (2022) performed a prospective diagnostic accuracy study of VUNO Med-DeepBrain AD (DBAD). Participants were 98 adults aged ≥60 years enrolled from two Korean hospitals who completed a standardized clinical evaluation, neuropsychological assessment, amyloid PET scan, and MRI. DBAD was compared with a consensus diagnosis from two neuroradiologists and one neurologist. DBAD had similar diagnostic performance compared with the expert panel for AD versus cognitively normal controls. Diagnostic accuracy was lower when preclinical AD and MCI were included. The authors concluded that DBAD may support physicians who are not AD specialists but did not evaluate whether AI-assisted interpretation improves clinician diagnostic performance, influences treatment decisions, changes patient management, or improves patient outcomes.³²

Yang et al. (2023), in a retrospective study, compared brain volume and normative percentiles generated by NeuroQuant (NQ) and DeepBrain (DB) in 163 Korean participants without gross structural brain abnormalities and evaluated whether differences in normative percentiles were related to cranial shape. They concluded that the mean normative percentiles obtained from NQ and DB differed significantly for many brain regions and that these differences may be related to cranial shape, suggesting the need to correct normative percentiles using transverse (TS), anterorposterior (AP), and craniocaudal (CC) diameters on brain MRI to reflect individual cranial shape.⁶⁹

Joo et al. (2022) conducted a retrospective study to investigate the diagnostic performance of VUNO Med-DeepBrain white matter hyperintensity (WMH) volume segmentation for classifying Fazekas scale grades and differentiating subcortical vascular dementia. Sensitivity, specificity, and accuracy for detecting moderate/severe WMH were 96.4%, 89.9%, and 91.7%, respectively, when using WMH volume. Similarly, WMH volume showed strong discrimination for subcortical ischemic vascular dementia (SVaD). The authors concluded that deep learning–based automatic WMH segmentation (DeepBrain) may be promising for classifying Fazekas scale grades and differentiating subcortical vascular dementia. The study was retrospective and lacked external validation.⁷⁰

Comparison Between Technologies

A review paper focused on AI for brain neuroanatomical segmentation in MRI imaging concludes high accuracy and fast performance overall. The application shows strong potential to enhance clinical diagnostics and surgical planning by improving the efficiency and accuracy of brain image analysis. Despite advances in automated segmentation, key challenges remain, including robustness to anatomical variability and maintaining high accuracy across diverse patient populations and imaging conditions.⁷¹ One of the challenges in developing automated volumetry software is lack of a gold standard for similar brain measurements to establish if the software correlates with reality. In current software product measurements of the brain segments are made with different methods and tools and therefore lack standardized measures for comparison. Efforts to understand the performance of different software modalities are undergoing as consistency between programs is an important component to create reliable standards which can be applied to clinical practice.

Multiple investigations compared the inter-method reliability between NeuroQuant and FreeSurfer computer-automated programs for measuring MRI brain volume. These demonstrate high inter-method reliability between the modalities yet often differ substantially in absolute volume estimates, making untransformed measures unsuitable for combined statistical analyses because method differences can masquerade as group effects. Linear transformations can reduce these discrepancies for some regions, but their usefulness is study specific, requires independent validation, and is inadequate when moderate or large effect size differences remain. While Bayesian regression methods appear to resolve most method related differences, they are not yet widely available or routinely used, and the relative accuracy of NQ versus FS across brain regions remains unclear.^72,73 Using 56 MRIs in patients with AD or MS, investigators compared results between NeuroQuant and volBrain automated brain analysis software and found high reliability except in the thalamus and amygdala where reliability has been proven to be poor. Using 115 MRIs with clinically isolated syndrome both measured with NeuroQuant and FMRIB's Integrated Registration Segmentation Tool (FIRST) found some variability between the modalities with larger volumes achieving better agreement. The presence of MS lesions affected volume estimates and must be accounted for to obtain reliable results. Differences between groups varied by software, with FIRST identifying more significant regional differences, underscoring the need for more accurate clinically robust MRI volumetric tools before routine clinical use.⁷⁴ Another investigation compared NeuroQuant to DeepBrain and found significant differences in many brain regions.⁶⁹ It found that NQ reported significantly higher normative percentiles for the thalamus, putamen, and parietal lobe compared to DB, while DB reported higher values for the occipital lobe. The study highlighted a significant correlation between these differences and cranial shape, indicating that skull morphology should be considered in brain volumetry assessments, especially across different ethnicities. A retrospective report compared 87 patients’ images with memory impairment with FreeSurfer, NeuroQuant, and Heuron AD and found significant differences between the programs. Heuron AD indication is for PET amyloid scans.⁷⁵

A study compared SIENA to Icobrain longitudinal pipeline (Icobrain long) for assessing longitudinal whole brain atrophy (WBA) in MS patients. They found a strong correlation between SIENA and Icobrain long (r = 0.805, p < 0.001) with Icobrain long consistently quantifying longitudinal WBA in MS patients. The study highlights that WBA can serve as a potential biomarker for neurodegeneration in MS, aiding individual patient management and treatment monitoring.⁶⁶

Another study compared brain volumetrics in MS using SIENAX to NeuroQuant and MSmetrix.⁶⁰ SIENAX is widely used in cross-sectional MS studies, but clinical applications limited. The authors compare the performance of NeuroQuant and MSmetrix to SIENAX and concluded comparable results.

FreeSurfer, volBrain and FIRST are not FDA-cleared and used for research purposes.

*Aducanumab has been discontinued as a treatment for Alzheimer’s.

Monoclonal antibody therapies directed against amyloid are a novel treatment option for Alzheimer’s disease. As of the publication date of this LCD, this drug is currently covered under National Coverage Determination (NCD) 200.3 as part of the coverage with evidence development (CED) pathway. To receive coverage patients must meet the criteria within the NCD for eligibility and the study must be CMS-approved and registered on the ClinicalTrails.gov website. These investigations will aid in the understanding of complications as a result of these treatments and optimal management. It is expected those conducting the studies have the expertise required for the study, including appropriately trained radiologists and clinicians to diagnose and manage ARIA and other complications.

Software has been developed to support automated detection and quantification of ARIA. Investigators have conducted a clinical validity study demonstrating that the software can detect ARIA and may offer improved diagnostic performance compared with community-based readers, particularly in identifying mild ARIA findings. However, the authors noted several important limitations, including the need for validation in broader and more diverse clinical populations, reliance on image quality and standardized acquisition protocols, and the intended role of the tool as decision support rather than a replacement for expert clinical interpretation.²⁷ There is no evidence of clinical utility demonstrating that automated detection and quantification of ARIA improves patient outcomes or informs patient management decisions. The use of AI augmented technology does not eliminate the need for appropriate reader training, and the results are not intended to be used independently of physician interpretation. Noridian does not consider automated detection and quantification of brain MRI imaging to be reasonable and necessary.

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. Noridian will continue to monitor the evolving research related to these devices.

1. Hossain MZ, Daskalaki E, Brüstle A, Desborough J, Lueck CJ, Suominen H. The role of machine learning in developing non-magnetic resonance imaging based biomarkers for multiple sclerosis: A systematic review. BMC Med Inform Decis Mak. 2022;22(1):242.
2. Scarpazza C, Ha M, Baecker L, et al. Translating research findings into clinical practice: A systematic and critical review of neuroimaging-based clinical tools for brain disorders. Translational Psychiatry. 2020;10(1):107.
3. Pham N, Hill V, Rauschecker A, et al. Critical appraisal of artificial intelligence–enabled imaging tools using the levels of evidence system. American Journal of Neuroradiology. 2023;44(5):E21-E28.
4. FDA. K170981. 510(k) premarket notification. https://www.accessdata.fda.gov/cdrh_docs/pdf17/K170981.pdf . Published 9/7/2017. Accessed 8/20/26.
5. FDA. K240712. 510(k) premarket notification. https://www.accessdata.fda.gov/cdrh_docs/pdf24/K240712.pdf. Published 11/7/24. Accessed 8/20/26.
6. Rabinovici GD, Selkoe DJ, Schindler SE, et al. Donanemab: Appropriate use recommendations. J Prev Alzheimers Dis. 2025;12(5):100150.
7. FDA. K161148. 510(k) premarket notification. https://www.accessdata.fda.gov/cdrh_docs/pdf16/K161148.pdf. Published 8/9/2016. Accessed 8/20/26.
8. FDA. K181939. 510(k) premarket notification. https://www.accessdata.fda.gov/cdrh_docs/pdf18/K181939.pdf. Published 11/6/2018. Accessed 8/20/26.
9. FDA. K192130. 510(k) premarket notification. https://www.accessdata.fda.gov/cdrh_docs/pdf19/K192130.pdf. Published 12/13/2019.. Accessed 8/26/26.
10. FDA. K180326. 510(k) premarket notification. https://www.accessdata.fda.gov/cdrh_docs/pdf18/K180326.pdf. Published 3/8/2018. Accessed 8/26/26.
11. FDA. K231398. 510(k) premarket notification. https://www.accessdata.fda.gov/cdrh_docs/pdf23/K231398.pdf. Published 10/4/2023. Accessed 8/26/26.
12. FDA. K182904. 510(k) premarket notification. https://www.accessdata.fda.gov/cdrh_docs/pdf18/K182904.pdf. Published 7/5/2019. Accessed 8/26/26.
13. Wu W, Ji Y, Wang Z, et al. The FDA-approved anti-amyloid-β monoclonal antibodies for the treatment of Alzheimer's disease: A systematic review and meta-analysis of randomized controlled trials. European Journal of Medical Research. 2023;28(1):544.
14. Cogswell PM, Barakos J, Barkhof F, et al. Amyloid-related imaging abnormalities with emerging Alzheimer disease therapeutics: Detection and reporting recommendations for clinical practice. Am J Neuroradiol. 2022;43(9):E19-E35.
15. Barakos J, Sperling R, Salloway S, et al. MR imaging features of amyloid-related imaging abnormalities. American Journal of Neuroradiology. 2013;34(10):1958-1965.
16. Barkhof F, Daams M, Scheltens P, et al. An MRI rating scale for amyloid-related imaging abnormalities with edema or effusion. Am J Neuroradiol. 2013;34(8):1550-1555.
17. Klein G, Scelsi MA, Barakos J, et al. Comparing ARIA-E severity scales and effects of treatment management thresholds. Alzheimers Dement. 2022;14(1):e12376.
18. Sperling R, Salloway S, Brooks DJ, et al. Amyloid-related imaging abnormalities in patients with Alzheimer's disease treated with bapineuzumab: A retrospective analysis. The Lancet Neurology. 2012;11(3):241-249.
19. Cogswell PM, Andrews TJ, Barakos JA, et al. Alzheimer disease anti-amyloid immunotherapies: Imaging recommendations and practice considerations for monitoring of amyloid-related imaging abnormalities. American Journal of Neuroradiology. 2025;46(1):24-32.
20. Bracoud L, Fiebach JB, Purcell DD, et al. [P1–047]: Validation of a simple severity scale for assessing ARIA-E. Alzheimer's & Dementia. 2017;13(7S_Part_5):P253-P254.
21. Cummings J, Rabinovici GD, Atri A, et al. Aducanumab: Appropriate sse recommendations update. The Journal of Prevention of Alzheimer's Disease. 2022;9(2):221-230.
22. Cummings J, Apostolova L, Rabinovici GD, et al. Lecanemab: Appropriate use recommendations. The Journal of Prevention of Alzheimer's Disease. 2023;10(3):362-377.
23. Ketter N, Brashear HR, Bogert J, et al. Central review of amyloid-related imaging abnormalities in two phase III clinical trials of bapineuzumab in mild-to-moderate Alzheimer’s disease patients. Journal of Alzheimer’s Disease. 2017;57(2):557-573.
24. Brewer JB, Magda S, Airriess C, Smith ME. Fully-automated quantification of regional brain volumes for improved detection of focal atrophy in Alzheimer disease. American Journal of Neuroradiology. 2009;30(3):578-580.
25. Engedal K, Brækhus A, Andreassen OA, Nakstad PH. Diagnosis of dementia—automatic quantification of brain structures. Tidsskr Nor Laegeforen. 2012;132(15):1747-1751.
26. Bash S, Wang L, Airriess C, et al. Deep learning enables 60% accelerated volumetric brain MRI while preserving quantitative performance: A prospective, multicenter, multireader trial. Am J Neuroradiol. 2021;42(12):2130-29137.
27. Sima DM, Vân Phan T, Van Eyndhoven S, et al. Artificial intelligence assistive software tool for automated detection and quantification of amyloid-related imaging abnormalities. JAMA Network Open. 2024;7(2):e2355800.
28. Struyfs H, Sima DM, Wittens M, et al. Automated MRI volumetry as a diagnostic tool for Alzheimer's disease: Validation of icobrain dm. Neuroimage Clin. 2020;26:102243.
29. Wittens MMJ, Sima DM, Houbrechts R, et al. Diagnostic performance of automated MRI volumetry by icobrain dm for Alzheimer's disease in a clinical setting: A REMEMBER study. J Alzheimers Dis. 2021;83(2):623-639.
30. Bracoud L, Klein G, Lyons M, et al. Validation of 3- and 5-point severity scales to assess ARIA-E. Alzheimer’s Dement. 2023;15(4):e12503. doi:10.1002/dad2.12503
31. Klein G, Scelsi MA, Barakos J, et al. Comparing ARIA-E severity scales and effects of treatment management thresholds. Alzheimer’s Dement. 2022;14(1):e12376. doi:10.1002/dad2.12376
32. Kim JS, Han JW, Bae JB, et al. Deep learning-based diagnosis of Alzheimer's disease using brain magnetic resonance images: An empirical study. Sci Rep. 2022;12(1):18007. doi:10.1038/s41598-022-22917-3
33. Nguyen AL, Horakova D, Havrdova EH, et al. Relationship between brain atrophy and disability in a multi-site multiple sclerosis registry. BMJ Neurology Open. 2025;7:e001126. doi:10.1136/bmjno-2025-001126
34. Temmerman J, Vanremoortel A, Bjerke M,et al Real-world brain volumetry in multiple sclerosis: Importance of methodological consistency and clinical relevance of gray matter atrophy. Front Neurol. 2025;16:1637835. doi:10.3389/fneur.2025.1637835
35. Rocca MA, Battaglini M, Benedict RHB,et al. Brain MRI atrophy quantification in MS: From methods to clinical application. Neurology. 2017;88:403–413.
36. Akaishi T, Takahashi T, Fujihara K et al. Number of MRI T1-hypointensity corrected by T2/FLAIR lesion volume indicates clinical severity in patients with multiple sclerosis. PLoS ONE. 2020;15(4):e0231225. doi: 10.1371/journal.pone.0231225
37. D'hooghe MB, Gielen J, Van Remoortel A, et al. Single MRI-based volumetric assessment in clinical practice is associated. J Magn Reson Imaging. 2019;49(5):1312-1321.
38. Fragoso YD, Wille PR, Abreu M, et al.Correlation of clinical findings and brain volume data in multiple sclerosis. Journal of Clinical Neuroscience. 2017;44:155–157. doi: 10.1016/j.jocn.2017.06.006
39. Seccia R, Romano S, Salvetti M, Crisanti A, Palagi L, Grassi F. Machine learning use for prognostic purposes in multiple sclerosis. Life. 2021;11(2):122.
40. Shoeibi A, Khodatars M, Jafari M, et al. Applications of deep learning techniques for automated multiple sclerosis detection using magnetic resonance imaging: A review. Computers in Biology and Medicine. 2021;136:104697.
41. Amin M, Martínez-Heras E, Ontaneda D, Prados Carrasco F. Artificial intelligence and multiple sclerosis. Curr Neurol Neurosci Rep. 2024;24(8):233-243.
42. Aslam N, Khan IU, Bashamakh A, et al. Multiple sclerosis diagnosis using machine learning and deep learning: Challenges and opportunities. Sensors. 2022;22(20):7856.
43. Sima DM, Esposito G, Van Hecke W, Ribbens A, Nagels G, Smeets D.Health economic impact of software-assisted brain MRI on therapeutic decision-making and outcomes of relapsing-remitting multiple sclerosis patients—A microsimulation study. Brain Sci. 2021:11(12):1570. doi: 10.3390/brainsci11121570
44. Arani LA, Hosseini A, Asadi F, Masoud SA, Nazemi E. Intelligent computer systems for multiple sclerosis diagnosis: A systematic review of reasoning techniques and methods. Acta Informatica Medica. 2018;26(4):258-264.
45. Nabizadeh F, Masrouri S, Ramezannezhad E, et al. Artificial intelligence in the diagnosis of multiple sclerosis: A systematic review. Mult Scler Relat Disord. 2022;59:103673.
46. Nabizadeh F, Ramezannezhad E, Kargar A, Sharafi AM, Ghaderi A. Diagnostic performance of artificial intelligence in multiple sclerosis: A systematic review and meta-analysis. Neurol Sci. 2023;44(2):499-517.
47. Nguyen AL, Sormani MP, Horakova D, et al. Utility of icobrain for brain volumetry in multiple sclerosis clinical practice. Mult Scler Relat Disord. 2024;92:106148.
48. Van Hecke W, Costers L, Descamps A, et al. A novel digital care management platform to monitor clinical and subclinical disease activity in multiple sclerosis. Brain Sci. 2021;11:1171.
49. White N, Baig S, Vidic I, et al. Segmentation of pre and post-treatment glioma tissue types including resection cavities. Paper presented at: Neuro-Oncology. 2022.
50. Gagnon L, Gupta D, Mastorakos G, et al. NIMG-30. Deep learning segmentation of infiltrative and enhancing cellular tumor in post-treatment glioblastoma patients with multi-shell diffusion MRI. Neuro-Oncology. 2023;25(Supplement_5):v191-v192.
51. White NS, McDonald C, Farid N, et al. Diffusion-weighted imaging in cancer: Physical foundations and applications of restriction spectrum imaging. Cancer Research. 2014;74(17):4638-4652.
52. Zubair Rahman AM, Gupta M, Aarathi S, et al. Advanced AI-driven approach for enhanced brain tumor detection from MRI images utilizing EfficientNetB2 with equalization and homomorphic filtering. BMC Med Inform Decis Mak. 2024;24(1):113.
53. Louis S, Morita-Sherman M, Jones S, et al. Hippocampal sclerosis detection with NeuroQuant compared with neuroradiologists. Am J Neuroradiol. 2020;41(4):591-597.
54. Farid N, Girard HM, Kemmotsu N, et al. Temporal lobe epilepsy: Quantitative MR volumetry in detection of hippocampal atrophy. Radiology. 2012;264(2):542-550.
55. Azab M, Carone M, Ying SH, Yousem DM. Mesial temporal sclerosis: Accuracy of NeuroQuant versus neuroradiologist. American Journal of Neuroradiology. 2015;36(8):1400-1406.
56. Ross DE, Ochs AL, Seabaugh JM, Shrader CR. Man versus machine: Comparison of radiologists' interpretations and NeuroQuant® volumetric analyses of brain MRIs in patients with traumatic brain injury. J Neuropsychiatry Clin Neurosci. 2013;25(1):32-39.
57. Ross DE, Ochs AL, DeSmit ME, Seabaugh JM, Havranek MD. Man versus machine part 2: Comparison of radiologists' interpretations and NeuroQuant measures of brain asymmetry and progressive atrophy in patients with traumatic brain injury. J Neuropsychiatry Clin Neurosci. 2015;27(2):147-152.
58. Choi KH, Heo YJ, Baek HJ, Kim JH, Jang JY. Comparison of inter-method agreement and reliability for automatic brain volumetry using three different clinically available software packages. Medicina. 2024;60(5):727. doi:10.3390/medicina60050727
59. Ochs AL, Ross DE, Zannoni MD, Abildskov TJ, Bigler ED. . Comparison of automated brain volume measures obtained with NeuroQuant and FreeSurfer. J Neuroimaging. 2015;25(5):721-727. doi:10.1111/jon.12229
60. Wang C, Beadnall HN, Hatton SN, et al. Automated brain volumetrics in multiple sclerosis: A step closer to clinical application. J Neurol Neurosurg Psychiatry. 2016;87(7):754-757. doi:10.1136/jnnp-2015-312304
61. Yim Y, Lee JY, Oh SW, et al. Comparison of automated brain volume measures by NeuroQuant vs. Freesurfer in patients with mild cognitive impairment: Effect of slice thickness. Yonsei Med J. 2021;62(3):255-261. doi:10.3349/ymj.2021.62.3.255
62. Kim M, Kim SJ, Park JE, et al. Combination of automated brain volumetry on MRI and quantitative tau deposition on THK-5351 PET to support diagnosis of Alzheimer's disease. Sci Rep. 2021;11(1):10343. doi:10.1038/s41598-021-89797-x
63. Stelmokas J, Yassay L, Giordani B, et al. Translational MRI volumetry with NeuroQuant: Effects of version and normative data on relationships with memory performance in healthy older adults and patients with mild cognitive impairment. J Alzheimers Dis. 2017;60(4):1499-1510. doi:10.3233/JAD-170306
64. Persson K, Barca ML, Cavallin L, et al. Comparison of automated volumetry of the hippocampus using NeuroQuant® and visual assessment of the medial temporal lobe in Alzheimer's disease. Acta Radiol. 2018;59(8):997-1001. doi:10.1177/0284185117743778
65. Ross DE, Ochs AL, Seabaugh JM, et al. Progressive brain atrophy in patients with chronic neuropsychiatric symptoms after mild traumatic brain injury: A preliminary study. Brain Inj. 2012;26(12):1500-1509. doi:10.3109/02699052.2012.694570
66. Beadnall HN, Wang C, Van Hecke W, Ribbens A, Billiet T, Barnett MH. Comparing longitudinal brain atrophy measurement techniques in a real-world multiple sclerosis clinical practice cohort: Towards clinical integration? Ther Adv Neurol Disord. 2019;12:1-20. doi:10.1177/1756286418823462
67. Simarro J, Meyer MI, Van Eyndhoven S, et al. A deep learning model for brain segmentation across pediatric and adult populations. Sci Rep. 2024;14(1):11735. doi:10.1038/s41598-024-61798-6
68. Bregman N, de Barros NP, Nathan T, et al. Lecanemab in practice: AI-derived MRI predictors of benefit and Amyloid Related Imaging Abnormalities (ARIA). Alzheimers Res Ther. 2026.doi:10.1186/s13195-026-02127-z
69. Yang MH, Kim EH, Choi ES, Ko H. Comparison of normative percentiles of brain volume obtained from NeuroQuant® vs. DeepBrain® in the Korean population: Correlation with cranial shape. J Korean Soc Radiol. 2023;84(5):1080-1090. doi:10.3348/jksr.2023.0006
70. Joo L, Shim WH, Suh CH, et al. Diagnostic performance of deep learning-based automatic white matter hyperintensity segmentation for classification of the Fazekas scale and differentiation of subcortical vascular dementia. PLoS One. 2022;17(9):e0274562. doi:10.1371/journal.pone.0274562
71. Andrews M, Di Ieva A. Artificial intelligence for brain neuroanatomical segmentation in magnetic resonance imaging: A literature review. J Clin Neurosci. 2025;134:111073.
72. Fujimori J, Fujihara K, Wattjes M, Nakashima I. Patterns of cortical grey matter thickness reduction in multiple sclerosis. Brain and Behavior. 2021;11:e02050. doi:10.1002/brb3.2050
73. Ross DE, Ochs AL, Tate DF, et al. High correlations between MRI brain volume measurements based on NeuroQuant® and FreeSurfer. Psychiatry Res Neuroimaging. 2018;278:69-76.
74. Pareto D, Sastre-Garriga J, Alberich M, et al. Brain regional volume estimations with NeuroQuant and FIRST: A study in patients with a clinically isolated syndrome. Neuroradiology. 2019;61(6):667-674.
75. FDA. K231642. 510(k) premarket notification. https://www.accessdata.fda.gov/cdrh_docs/pdf23/K231642.pdf Published 10/13/2023. Accessed 8/27/26.

• Creation of Uniform LCDs With Other MAC Jurisdiction