August 7, we published a blog on software diagnosis in radiology and asked how fast it might come to pathology. Here. And we've already updated that once, August 14. Here.
Below, FDA approves de novo software that auto classifies MRIs as Parkinson disease vs several other disorders. Brave new world.
##
AI Corner
##
A few weeks ago, this blog reviewed DeepHealth SMART-B, an FDA-cleared breast-ultrasound system that detects and characterizes lesions and generates draft findings and impressions for radiologist review. That clearance prompted a broader discussion of the outlook for report generation in pathology—already visible in several U.S. research-use-only products and in some software cleared for clinical use in Europe.
Here is another timely example from a different corner of diagnostic medicine. FDA has granted De Novo classification to neuropacs, machine-learning software that analyzes diffusion MRI and produces a diagnostic classification report for Parkinson disease and two related parkinsonian disorders.
Neuropacs is not a general-purpose report-writing system, and it does not produce an open-ended neurological diagnosis from any patient with tremor. Its authorized use is considerably narrower. But that is precisely why it matters. It demonstrates that clinically consequential AI-generated reporting may arrive first as a tightly constrained, validated diagnostic work product—not as an eloquent narrative written by a large language model.
AI review follows.
Another FDA Signal: Neuropacs Turns Diffusion MRI Into a Parkinsonian-Syndrome Classification Report
The short version
On April 3, 2026, FDA granted De Novo classification to neuropacs, developed by Automated Imaging Diagnostics, a subsidiary of Neuropacs Corp. FDA created a new Class II device category called a “Parkinsonian syndrome diagnostic aid,” under product code SHO and 21 CFR 882.2000. The regulatory record is DEN240071.
The software receives diffusion-weighted MRI data from patients aged 40 or older and generates a classification report intended to help neurologists and neuroradiologists distinguish among:
Parkinson disease, or PD;
multiple system atrophy, parkinsonian variant, or MSAp; and
progressive supranuclear palsy, or PSP.
The underlying system uses an advanced diffusion-MRI technique known as free-water imaging, followed by a machine-learning classifier. The published model calculates imaging features across numerous brain regions and produces disease probabilities and a final classification.
This is a much more restricted claim than “AI diagnoses Parkinson’s.” FDA requires the software to be used as supplemental information alongside a conventional neurological assessment and other clinical testing. The labeling states that other causes of parkinsonism—including dementia with Lewy bodies, vascular parkinsonism, drug-induced parkinsonism, and corticobasal degeneration or syndrome—should first be excluded. Patient-management decisions must not be based solely on the software output. Those boundaries are explicit in the FDA classification order.
Still, the regulatory milestone is significant. The software analyzes a medical image and returns not merely a measurement or highlighted region, but a disease-oriented classification report. It therefore occupies some of the same conceptual territory as the emerging report-generation systems in radiology and pathology.
##
##
Why this differential diagnosis is difficult—and important
“Parkinsonism” is a clinical syndrome rather than a single disease. Slowness, rigidity, gait abnormalities, tremor, and postural instability can occur in Parkinson disease, but also in several less common neurodegenerative disorders.
MSA may initially resemble Parkinson disease but is often associated with prominent autonomic dysfunction and a different prognosis. PSP can also initially resemble PD but may later become recognizable through characteristic eye-movement abnormalities, early falls, axial rigidity, speech problems, and other features. Both disorders generally have different treatment expectations and usually respond less consistently to levodopa than typical PD.
The distinction matters for prognosis, counseling, treatment planning, selection for procedures such as deep-brain stimulation, and enrollment in disease-specific clinical trials. Yet even experienced movement-disorder specialists may need longitudinal follow-up before the diagnosis becomes clear.
Conventional MRI can identify certain characteristic findings in established MSA or PSP, but those signs are not uniformly present, especially early in disease. Dopamine-transporter SPECT imaging—commonly known by the DaTscan brand—can help distinguish neurodegenerative parkinsonism from conditions such as essential tremor. However, an abnormal dopamine-transporter scan does not reliably distinguish PD from MSA or PSP because all three can involve degeneration of the dopaminergic system.
Neuropacs is aimed at this second, more difficult question: once neurodegenerative parkinsonism is under consideration, does the pattern of brain microstructural injury look more like PD, MSAp, or PSP?
How the software works
The approach begins with a diffusion-weighted MRI scan acquired on a 3-Tesla scanner. According to the company, the analysis requires at least 30 diffusion-gradient directions and can work with qualifying scans from Siemens, GE HealthCare, and Philips systems. The published clinical protocol required less than ten minutes of additional MRI acquisition.
Diffusion MRI measures the movement of water through tissue. Conventional diffusion-tensor methods combine several contributors to that signal. Free-water imaging instead uses a two-compartment model to distinguish water moving relatively freely in the extracellular space from diffusion occurring within or around brain tissue.
That separation may reveal subtle microstructural changes before they become obvious as gross atrophy on an ordinary anatomical MRI.
In the pivotal published model, free-water and free-water-corrected fractional-anisotropy measurements were calculated across 132 brain regions and tracts. These included areas in the basal ganglia, thalamus, cortex, brainstem, cerebellum, corpus callosum, and sensorimotor pathways. Age and sex were also included in the principal feature set, although a sensitivity analysis found that removing them did not materially reduce performance.
A linear support-vector-machine model then performed a two-stage classification:
PD versus atypical parkinsonism, meaning MSAp or PSP; and
if atypical parkinsonism was predicted, MSAp versus PSP.
Examples in the published study display disease-probability estimates followed by a final classification. FDA’s order more generally describes the output as a “classification report.”
The company says that the software is delivered through a secure cloud workflow and can integrate with existing picture-archiving and communication systems. In practical terms, an eligible MRI study can be routed for remote analysis and the resulting report returned to the imaging or neurological workflow. Additional technical descriptions are available on the company’s AIDP product page.
The principal clinical study
The central evidence is a 2025 prospective, multicenter cohort study published in JAMA Neurology: Vaillancourt et al., “Automated Imaging Differentiation for Parkinsonism”.
The prospective portion was conducted from July 2021 through January 2024 at 21 Parkinson Study Group centers in the United States and Canada. Investigators screened 316 patients, of whom 249 met the study criteria:
99 with Parkinson disease;
53 with MSAp; and
97 with PSP.
For the prospective cohort, the reference diagnosis required unanimous agreement among three neurologists specializing in movement disorders. One was the examining site neurologist, while two off-site experts independently reviewed videotaped examinations, clinical scales, and conventional MRI information. The Neuropacs analysis was performed after the clinical assessment and was not used to establish the reference diagnosis.
The researchers also incorporated a retrospective training cohort of 396 patients: 211 with PD, 98 with MSA, and 87 with PSP.
Altogether, the primary model used 500 patients for training—104 prospective patients plus all 396 retrospective patients. An independent test set consisted of the remaining 145 prospective patients: 60 with PD, 27 with MSA, and 58 with PSP.
The independent test-set performance was strong:
| Diagnostic comparison | AUROC | Sensitivity | Specificity |
|---|---|---|---|
| PD vs. atypical parkinsonism | 0.961 | 87.1% | 88.3% |
| MSAp vs. PSP | 0.983 | 89.7% | 96.2% |
| PD vs. MSAp | 0.983 | 96.7% | 85.2% |
| PD vs. PSP | 0.984 | 98.3% | 91.4% |
The analysis was repeated with a second diffusion scan and produced similar results. The investigators also conducted 49 additional train-test splits and analyses that held out entire clinical sites. Performance declined modestly in some of the more demanding site-holdout analyses but remained generally strong, with AUROCs ranging from approximately 0.87 to 0.96.
A neuropathology analysis provided another encouraging result. Among 49 patients with postmortem diagnoses, the imaging classification agreed with pathology in 46, or 93.9%. The last clinical diagnosis agreed with pathology in 81.6%. However, the autopsy set was heavily weighted toward PSP: it contained 39 PSP brains but only five PD and five MSA brains.
An important caution about “96% accurate”
The company’s public materials sometimes describe the system as having “over 96% accuracy” or “up to 98% precision.” The underlying publication more carefully reports AUROC, sensitivity, specificity, positive predictive value, and negative predictive value.
These terms are not interchangeable.
An AUROC of 0.98 is an excellent discrimination result, but it does not mean that the device will provide the correct diagnosis in exactly 98% of ordinary clinical patients. Actual predictive performance depends on the classification threshold, the prevalence of each disease in the tested population, the patient-selection rules, and the clinical setting.
The principal study deliberately assembled substantial numbers of MSA and PSP cases. That was appropriate for developing and testing the classifier, but it does not reproduce the prevalence encountered in a general neurology practice, where PD is much more common. Positive and negative predictive values may therefore be different in routine use.
The most defensible summary is that the study produced AUROCs of approximately 0.96 to 0.98 in its independent prospective test set, with sensitivity and specificity varying by diagnostic comparison.
Strong evidence, but not yet every kind of evidence
The study is considerably stronger than the typical single-center retrospective AI paper. It was prospective, multicenter, adequately powered, tested on scanners from all three major manufacturers, included an independent test set, performed repeat-scan analyses, and included both site-holdout and neuropathology evaluations.
Several limitations nevertheless matter.
First, the principal reference standard remained an expert clinical diagnosis rather than neuropathology. Clinical diagnosis is unavoidable in most living-patient studies, but it is imperfect—especially early in disease.
Second, requiring unanimous agreement among three expert movement-disorder neurologists created a rigorous reference standard but also selected relatively classifiable cases. The patients in greatest need of a diagnostic aid may be precisely those for whom expert reviewers disagree. The authors appropriately identified evaluation in clinically ambiguous cases as a future research need.
Third, the study was conducted largely in specialist movement-disorder centers. Community neurology and general radiology practices may encounter different referral patterns, imaging quality, comorbidities, and diagnostic uncertainty.
Fourth, the study did not include all disorders that can cause parkinsonism. That limitation is reflected in FDA’s labeling, which requires clinicians to rule out several alternative conditions before using the Neuropacs classification.
Fifth, the autopsy findings are promising but still based on a small and unbalanced pathology cohort. The result is particularly preliminary for PD and MSA, with only five brains in each category.
Finally, the pivotal study established diagnostic discrimination. It did not establish that use of the software changes treatment, reduces other testing, shortens the time to a stable diagnosis, improves patient outcomes, or lowers total cost. Those are clinical-utility and health-economic questions that frequently become important after FDA authorization.
The study was supported by the National Institutes of Health. Relevant commercial relationships were disclosed: David Vaillancourt reported a licensed patent and support from Automated Imaging Diagnostics, and Angelos Barmpoutis reported being a co-founder and shareholder of Neuropacs. The disclosures do not negate the results, but they are part of a complete reading of the evidence.
De Novo classification: more than a one-product event
Neuropacs did not enter an existing FDA category through the usual 510(k) substantial-equivalence process. FDA instead granted a direct De Novo request and created a new generic device type: the Parkinsonian syndrome diagnostic aid.
FDA placed the device in Class II, concluding that its risks could be managed through general controls and new special controls. The identified risks are straightforward but consequential: false-positive or false-negative classifications, inappropriate treatment, delayed diagnosis, misinterpretation, and overreliance on the software output.
The special controls require:
clinical validation under anticipated conditions of use;
diagnostic-accuracy and reproducibility testing against a clinically relevant reference standard;
objective performance measures;
software verification, validation, and hazard analysis;
a technical description of model inputs and outputs;
disclosure of the population used for model development;
labeling of limitations and unvalidated subpopulations;
instructions for incorporating the output into the diagnostic workflow; and
a clear warning that the device is not a stand-alone diagnostic.
FDA also determined that future devices in this category will require premarket notification. In other words, the De Novo decision does not merely authorize Neuropacs. It creates a potential 510(k) pathway for later competitors that can demonstrate substantial equivalence while satisfying the special controls.
The FDA database indicates that no predetermined change-control plan was authorized with this De Novo. Material future algorithm changes therefore may require additional regulatory evaluation rather than being automatically covered by a preauthorized update plan.
How does Neuropacs relate to alpha-synuclein testing?
The timing is particularly interesting because laboratory testing for abnormal alpha-synuclein is moving toward clinical use at the same time.
Parkinson disease and MSA are synucleinopathies, while PSP is primarily a tauopathy. Alpha-synuclein seed-amplification assays can detect misfolded alpha-synuclein in cerebrospinal fluid and, through other methods, in tissue such as skin. These tests address a molecular question: is pathological alpha-synuclein present?
Neuropacs addresses a different question: what pattern of neurodegeneration is visible across the brain, and does that pattern more closely resemble PD, MSAp, or PSP?
Those are not redundant measurements.
A newly published Annals of Neurology study examined this relationship directly: Chiu et al., “Diffusion MRI and α-Synuclein Seed Amplification Status in Parkinson’s Disease”.
The investigators evaluated 462 participants with early clinically diagnosed PD from the Parkinson’s Progression Markers Initiative. Of these, 421 were alpha-synuclein SAA-positive and 41 were SAA-negative. Neuropacs’ underlying AIDP method classified 427 participants, or 92.4%, as PD and 35, or 7.6%, as having an atypical parkinsonian pattern.
SAA positivity was associated with worse olfaction and one focal free-water difference in the superior cerebellar peduncle, but it did not correspond to broad differences across the diffusion-MRI measurements. In other words, molecular evidence of alpha-synuclein aggregation did not simply map onto the global neurodegenerative patterns measured by the MRI method.
That finding supports a complementary-biomarker model:
SAA may provide evidence about pathological protein biology.
Diffusion MRI may provide evidence about the anatomical pattern and extent of neurodegeneration.
Clinical examination continues to define the patient’s syndrome and functional state.
The Annals analysis should not be treated as a second pivotal validation study. All participants entered as clinical PD cases, and the 35 atypical imaging classifications were not shown to be confirmed reclassifications by pathology. It is better viewed as evidence that molecular and imaging biomarkers capture partly different dimensions of disease.
This distinction may become increasingly important as neurodegenerative medicine moves from traditional syndrome labels toward biological classification and staging systems.
What Neuropacs has to do with AI-generated reports
DeepHealth SMART-B and Neuropacs are not identical regulatory precedents.
SMART-B performs lesion detection and characterization and generates draft radiology findings and impressions. Neuropacs performs quantitative image processing and generates a constrained disease-classification report. It does not appear to be an unconstrained narrative-report generator, and its public evidence does not suggest that it composes a complete neurological consultation note.
Nevertheless, both products illustrate the same larger transition.
Earlier medical-imaging AI generally returned a heat map, contour, score, measurement, or alert. A human specialist then had to translate that output into the professional diagnostic work product.
These newer systems move further down the chain. They assemble analytical outputs into something closer to the assessment that enters the clinical record:
SMART-B: image findings and a draft impression;
Neuropacs: disease probabilities and a final diagnostic classification;
emerging pathology systems: cancer detection, grading, tumor measurements, biomarker quantification, and structured report fields.
The key step is not necessarily elegant prose. A report can be clinically transformative even if it consists of validated structured fields, probabilities, and a templated conclusion. Indeed, that constrained architecture may be easier to validate and regulate than free-form language generation.
Neuropacs reinforces a central observation from this blog’s updated review of AI-generated pathology reports: the first widely useful AI reports may come from a modular system in which validated image analysis establishes the facts and a controlled reporting layer packages them for specialist review.
Lessons for pathology
Several features of the Neuropacs authorization may foreshadow the FDA pathway for more ambitious pathology-reporting systems.
First, narrow claims may win before general claims. Neuropacs does not diagnose every cause of parkinsonism. A pathology system may similarly begin with a tightly bounded specimen type and diagnostic question—such as prostate core biopsies—rather than “autonomous surgical pathology.”
Second, structured reports may precede generative prose. A system that reliably outputs tumor presence, Gleason patterns, Grade Group, tumor length, percentage involvement, perineural invasion status, and uncertainty flags may be closer to authorization than one that writes unrestricted paragraphs.
Third, explicit workflow position matters. FDA specifies that Neuropacs is adjunctive, must be interpreted with other clinical evidence, and should not independently determine management. Similar labeling is likely for early pathology-report generation: the pathologist remains the reviewer and signer.
Fourth, the reference standard becomes a central problem. Neuropacs used three-expert consensus and a smaller pathology-confirmed subset. Pathology algorithms may have the apparent advantage of expert slide review, but difficult cases, interobserver variability, ancillary studies, sampling limitations, and diagnostic evolution still complicate “ground truth.”
Fifth, report design is itself a safety control. Probabilities, uncertainty, excluded diagnoses, warnings, and links back to the supporting image regions can reduce misinterpretation. A polished paragraph without provenance may be less safe than a structured report that visibly exposes its evidentiary basis.
Sixth, regulatory clearance is not reimbursement. FDA’s order establishes safety and effectiveness for the authorized use; it does not create Medicare coverage or payment. The public materials reviewed here do not identify a dedicated national coverage policy or product-specific reimbursement pathway for Neuropacs. Adoption will depend on whether providers can incorporate the software cost into existing imaging economics, obtain separate payment, or demonstrate enough clinical and operational value to justify institutional purchase.
The larger significance
Neuropacs is not autonomous neurology. It does not evaluate the complete patient, establish every differential diagnosis, decide treatment, or replace a movement-disorder specialist. The FDA labeling goes out of its way to prevent that interpretation.
But it is also more than another AI heat map.
The software takes a routinely recognizable form of clinical data, performs a technically sophisticated analysis invisible to ordinary visual inspection, and returns a disease-level classification report for physician use. FDA has created a dedicated regulatory category around that function.
That makes Neuropacs another marker of where diagnostic AI is heading. The decisive transition may not be from human-written reports to machine-written prose. It may be from algorithms that provide isolated measurements to systems that assemble a bounded, reviewable diagnostic conclusion.
Radiology is already crossing that boundary. Neuropacs shows it occurring in neurodegenerative diagnosis. Pathology is likely to follow through the same route: validated image features, disciplined diagnostic logic, structured report generation, traceable evidence, and specialist sign-off.
The resulting first-generation reports may look less like ChatGPT and more like an exceptionally sophisticated synoptic report. They may also arrive sooner—and carry more clinical consequence—than the phrase “AI report generation” initially suggests.
