About & method

NeuroVol takes a brain MRI you already have and shows you per-region volumes compared to a published reference for your age and sex. Here is how it works under the hood and where the limits are.

The pipeline

1. De-identification. Identifying DICOM header fields are removed before anything is persisted. We also regenerate study, series, and instance UIDs.
2. DICOM → NIfTI. dcm2niix converts the de-identified series to a single NIfTI volume.
3. Segmentation. We run SynthSeg on a serverless GPU. SynthSeg is a contrast- and resolution-agnostic deep segmentation model trained with extensive data augmentation, which is what lets it work on heterogeneous consumer-scanner MRIs.
4. Volumetry. We sum voxels per label and multiply by the voxel volume from the affine to get cubic millimetres per region.
5. Percentiles. For each region we look up the expected mean and SD for your age and sex in our reference table and compute a percentile under a normal model.
6. Plain-language report. Each region in the report has a curated anatomical description and a percentile sentence written by a human and stored in a versioned text file in this repository.

About the normative reference

The default normative model in NeuroVol is a parametric, anchor-based reference. For each region, we store two age anchors (30 years and 70 years) per sex with population-level mean values drawn from published volumetric studies (Coupé et al. 2017, Walhovd et al. 2011, Habes et al. 2016) and a conservative coefficient of variation. We interpolate linearly between the anchors and compute a percentile under a normal distribution.

This is a population-level approximation built for an educational MVP. It is not a cohort-trained model, it does not correct for intracranial volume or scanner differences, and it should not be read clinically. The interface is pluggable: replacing it with a cohort-trained model (for example, one fit to IXI or built on the Bethlehem et al. 2022 lifespan brain charts) is a matter of implementing the NormativeModel interface and selecting the new class — no callers change.

What NeuroVol does not do

• It does not diagnose disease.
• It does not make risk predictions.
• It does not interpret your scan visually or look at lesion load.
• It does not correct for differences between scanners or acquisition protocols.
• It does not compare your scan to a personalised baseline (the reference is a population mean, not your own historical mean).

Source & updates

NeuroVol is open source. The full pipeline, the curated text bank, the normative model, and the deployment configuration live in the GitHub repository.