Biomarkers and Tools

1. A biomarker is defined by its context of use

A number is not "validated" in the abstract. It can be validated for one or more specific jobs:

• susceptibility/risk: identify people more likely to develop disease or an outcome;
• diagnostic: identify current disease or subtype;
• prognostic: predict course among people with disease;
• monitoring: track change over time;
• predictive: identify differential response or harm from an intervention;
• pharmacodynamic/response: show that a biological process changed after treatment;
• trial enrichment/stratification: select or balance participants.

Technical repeatability, association with cognition, pathology prediction, and treatment responsiveness are different validation layers.

2. ARTS: the arteriolosclerosis biomarker

What it is

ARTS is a fully automated, containerized MRI-based classifier designed to output a score related to the likelihood of brain arteriolosclerosis. It combines:

• 3D T1-weighted MRI;
• T2-weighted FLAIR for WMH-related features;
• diffusion tensor imaging, including regional fractional-anisotropy information;
• age at MRI and sex.

The public package accepts DICOM (preferred by its documentation) or NIfTI and was distributed for Singularity-based execution on Linux/macOS and high-performance computing. Check the current package requirements before using Apptainer as a drop-in replacement.

Development logic

ARTS was trained using ex vivo MRI and neuropathology, then translated to in vivo imaging. Its central scientific value is the attempt to learn a pathology-linked arteriolosclerosis signature rather than treating total WMH as if it were specific.

Evidence to date

• The initial publication reported prediction of arteriolosclerosis, scan-rescan reproducibility, and association with two-year cognitive decline in nondemented older adults.
• MarkVCID instrumental validation reported extremely high inter-rater reliability, high test-retest repeatability, and strong inter-scanner reproducibility in the validation design.
• Clinical/biological validation linked higher baseline ARTS scores with worse baseline performance and faster decline in processing-speed/executive measures in independent cohorts.
• A later diverse cohort of 1,226 older adults linked higher ARTS with incident MCI, dementia, stroke, and several vascular risk factors, with heterogeneity across racial/ethnic groups.

What ARTS does not yet justify

• It should not be called a definitive clinical diagnosis of arteriolosclerosis.
• Its output is not a direct image of the vessel wall.
• Validation in very old cohorts does not automatically generalize to midlife, rare genetic cSVD, unusual scanner/protocol distributions, or different disease spectra.
• Age and sex are inputs; fairness and transportability must be evaluated, not assumed.
• Association with outcomes does not prove that changing the ARTS score changes outcomes.
• The composite model can capture correlates of arteriolosclerosis, comorbid pathology, aging, or tissue consequences unless competing pathologies are explicitly tested.

Access and reproducibility record

• Tool page: NITRC ARTS biomarker
• MarkVCID protocol/resources: MarkVCID2 protocols
• Initial paper: ARTS: a novel in-vivo classifier
• Validation report: ARTS instrumental and clinical validation

For any ARTS analysis, preserve software version, container hash, input series identifiers, acquisition parameters, DICOM-to-NIfTI conversion provenance, all quality-control failures, demographic inputs, exclusions, and raw output before transformation.

3. MarkVCID biomarker kits

MarkVCID is an NIH-funded multisite program intended to move promising VCID biomarkers toward trial readiness using common protocols and staged instrumental, biological, and clinical validation.

Current major MRI/fluid kits include:

Resources: MarkVCID consortium, MarkVCID1 results, and MarkVCID2 protocols.

4. CAA biomarkers

Hemorrhagic MRI markers

Strictly lobar CMBs, cSS/cSAH, and lobar ICH are the most established clinical imaging signals. cSS extent is also prognostic for recurrent lobar ICH in CAA cohorts. Sequence sensitivity is a major confounder: SWI detects more lesions than many GRE protocols, and field strength/echo time affect counts.

Non-hemorrhagic MRI markers

• severe centrum semiovale PVS;
• posterior or multispot WMH patterns;
• cortical microinfarcts;
• diffusion damage and PSMD;
• reduced structural-network efficiency;
• altered cerebrovascular reactivity;
• cortical thinning and atrophy;
• BBB permeability and vascular-function measures.

These markers may occur earlier or correlate more closely with cognition than microbleed count. Most lack the etiologic specificity to stand alone.

Amyloid PET

Amyloid PET can be sensitive to cerebral amyloid but cannot readily separate vascular from parenchymal amyloid. A negative scan may be useful in selected diagnostic contexts; a positive scan is less specific because Alzheimer-type plaques are common. Tracer, reference region, partial-volume correction, disease stage, and coexisting AD must be modeled.

CSF

Across studies, CAA is often associated with lower CSF Aβ40 and Aβ42, while tau patterns help distinguish or identify coexisting AD. Meta-analyses and newer cohorts support signal, but assay platforms, diagnostic reference standards, and overlap limit stand-alone use. CSF biomarkers are not part of Boston v2.0.

Blood

Plasma proteomic, vascular-injury, inflammatory, Aβ, and APOE-related signatures are active research areas. Recent exploratory studies report candidate protein panels associated with lobar CMBs or neuropathologic CAA. These require independent, prospective, spectrum-appropriate replication and comparison against AD, arteriolosclerosis, renal function, age, and systemic vascular disease.

Genetic information

• APOE ε4 is associated with CAA presence/amyloid deposition and capillary CAA.
• APOE ε2 is associated with severe vasculopathic/hemorrhagic manifestations in several cohorts.
• APP mutations cause hereditary CAA syndromes such as Dutch-type CAA; other genes/proteins cause rarer amyloid angiopathies.

Genetic association is not deterministic diagnosis in sporadic disease. Ancestry, allele frequency, selection, and treatment context matter.

5. Visual rating tools

Rating forms and atlases should be versioned. Training examples should include mimics and uncertainty, not only textbook-positive cases.

6. Open and common computational tools

WMH segmentation

• LST-LGA/LPA: SPM-based lesion growth/prediction algorithms; widely used, originally developed in multiple-sclerosis contexts and often adapted to aging WMH.
• BIANCA: FSL supervised classifier for white-matter hyperintensities; requires representative training labels and careful thresholding.
• SAMSEG: sequence-adaptive Bayesian segmentation in FreeSurfer; supports lesion segmentation in multimodal data.
• WMH Segmentation Challenge containers: standardized benchmark methods; benchmark rank does not guarantee transportability to a new cohort.
• nnU-Net/custom U-Net pipelines: powerful when trained and externally validated; susceptible to scanner and population domain shift.

For all automated WMH tools, perform blinded manual QC, record failure categories, assess small-lesion false positives, and compare performance across scanner, site, race/ethnicity, age, lesion burden, and co-pathology.

Diffusion and microstructure

• PSMD toolbox: miac-research/psmd, using FSL/TBSS-based skeletonization and histogram analysis.
• FSL diffusion tools: TOPUP, EDDY, DTIFIT, and TBSS; acquisition-specific QC is essential.
• Free-water models: implementation and fitting choices vary; keep code/version and parameter maps.

Microbleed detection

Automated CMB detectors are improving but face severe class imbalance and mimics from vessels, calcification, and susceptibility artifact. Evaluate lesion-level sensitivity and false positives per scan, not only voxel Dice or patient-level AUC. External validation on different SWI/GRE protocols is essential.

Structural/network analysis

• FreeSurfer: cortical thickness, volume, and surface reconstruction.
• FSL/ANTs: registration, segmentation, and image processing.
• MRtrix3: tractography and structural connectomics.
• QSIPrep/fMRIPrep-like reproducible pipelines: useful for standardized preprocessing, but cSVD lesions can break assumptions and require lesion-aware QC.

Digital pathology

The Perosa/Scherlek workflow trained separate deep-learning models for parenchymal Aβ plaques, vascular Aβ, iron/calcium, astrocytes, microglia, and fibrin extravasation. The key methodological lesson is to retain marker-specific outputs and expert QC rather than collapsing all staining into one pathology score.

Useful platforms include QuPath and slide-specific deep-learning frameworks. Cross-laboratory stain normalization, scanner differences, annotation policy, tissue folds, vessel-size definitions, and spatial registration to MRI are first-order problems.

7. Literature and evidence-archive tools

Discovery tools such as Connected Papers, ResearchRabbit, and Litmaps are useful for finding neighborhoods of papers but should not define completeness. Search strategy and eligibility rules should be documented independently.

8. Minimum tool-record schema

Every tool record should contain:

• name, version, owner, and stable URL;
• license and redistribution terms;
• intended context of use;
• required inputs and accepted formats;
• preprocessing and quality-control requirements;
• output definition and scale direction;
• training/development population;
• technical, biological, clinical, and external validation evidence;
• known failure modes and excluded populations;
• compute/runtime/container requirements;
• last verified date and maintainer status;
• primary citation and independent evaluations.

Use templates/TOOL_RECORD_TEMPLATE.md to add new entries.

9. A practical validation ladder

1. Analytical validity: does the measurement accurately quantify its intended signal on controlled data?
2. Repeatability: does it agree across raters and repeat scans?
3. Reproducibility: does it agree across scanners, sites, and implementations?
4. Biological validity: does it associate with tissue or a credible mechanistic reference?
5. Clinical validity: does it identify or predict a meaningful clinical state/outcome?
6. Incremental validity: does it add information beyond age, risk factors, and standard imaging?
7. Transportability and fairness: does it work in new sites, protocols, ancestries, ages, and comorbidity patterns?
8. Responsiveness: does it change over time or after an intervention in an interpretable way?
9. Clinical utility: does using it improve decisions or outcomes?

Most cSVD biomarkers have not completed the full ladder. That is not failure; it is the research map.