Resources
The research behind NeuroMetric.
How the model works, what the dashboard includes, and where the data comes from. A technical reference for the precomputed research demo.
Model & pipeline
Research-only estimation of chronological age from T1 MRI using a pretrained, fine-tuned SFCN. Brain-age gap is predicted minus chronological age; it is not ground-truth biological age.
The local pipeline uses 3D T1-weighted NIfTI (.nii / .nii.gz); the OASIS preparation tool also reads Analyze .img/.hdr pairs. SynthStrip brain extraction, FSL FAST bias correction and 12-DOF FLIRT alignment to 1 mm MNI precede a 160 × 192 × 160 SFCN input. No browser upload or DICOM service is implemented.
Whole-brain predicted age and raw age gap. Selected examples include real signed Grad-CAM and MRI-derived FastSurfer regional volumes. Volumes are not regional brain ages, and attention maps are not disease-risk maps. Healthy-control percentiles and calibrated individual confidence intervals are not available.
Not a diagnostic device. Alzheimer-cohort evaluation, clinical validation, individual uncertainty, regional validity and longitudinal reliability remain unconfirmed. Repeated validation tuning can inflate estimates. Pretraining overlap with OASIS is unaudited; exploratory results cannot be called independent clinical validation.
Pretrained, fine-tuned 3D SFCN · contrast-cont4 · best checkpoint at epoch 22. PyTorch inference uses the saved checkpoint; the website displays precomputed results and does not run the model in the browser.
Implemented features
- PredictionimplementedPretrained SFCN fine-tuning, saved model checkpoints and held-out chronological-age evaluation.
- AttentionimplementedPatient-specific signed Grad-CAM with axial, coronal and sagittal view switching. Attribution is explanatory, not diagnostic.
- RegionsmeasuredSelected IXI and DLBS examples have bilateral hippocampal and cortical gray-matter volumes, with their own segmentation images. Technical review is distinct from clinical validation.
- Follow-upexploratoryOAS2_0007 displays its own recorded visits, 518 days apart. Registration review is pending; a validated aging rate is not claimed.
Data policy
Public-access research datasets have their own licenses and access conditions. Public availability does not mean unrestricted use.
Training/validation source and historical held-out test split. Selected dashboard cases: IXI170, IXI361 and IXI566. CC BY-SA 3.0 attribution applies.
External evaluation dataset ds004856. Selected dashboard example: sub-3898. Dataset license recorded in the project catalog: CC0 1.0.
OASIS-1 supports exploratory dementia-cohort work; OASIS-2 provides longitudinal research examples. Follow the applicable access agreement and citation requirements. A dementia label is not necessarily biomarker-confirmed Alzheimer disease.
This static site accepts no MRI uploads and performs no inference. Local research tools should receive only appropriately de-identified data; age is required to compute an age gap.
Training and evaluation rely on public-access/de-identified research data under their own terms. MRI conversion is not a certified anonymization or defacing service; do not assume processing removes every identifying feature.
Downloaded scans, prepared arrays and QC artifacts are retained locally for this research workflow. There is no implemented automatic deletion policy or hosted patient-data service.
IXI, SALD and NIMH supply training/validation data. IXI test and DLBS are held out from tuning; OASIS-1 is an exploratory dementia-cohort evaluation source, not model training data. OASIS use requires its data-use agreement, no reidentification and required acknowledgements.
Research demonstration only. Not a diagnostic device or a substitute for clinical assessment. No visitor MRI uploads, live inference or hosted patient database are implemented.