Background and theory

This page explains the clinical motivation, imaging measurements, and classification logic behind AutoHS. It is written for clinicians, researchers, and engineers who land on the documentation site without prior context.

Clinical context

Hippocampal sclerosis (HS) is a common pathological substrate in temporal lobe epilepsy (TLE). In surgical workup, MRI is used to detect asymmetric hippocampal atrophy on the side of seizure onset. Quantitative hippocampal volumetry and asymmetry metrics can support screening when visual read is equivocal.

AutoHS automates a reproducible workflow:

  1. Segment the T1-weighted scan (FreeSurfer or FastSurfer)

  2. Read left and right hippocampal volumes from segmentation statistics

  3. Compute the asymmetry index (AI)

  4. Apply published thresholds for volume laterality and HS screening

AutoHS is a decision-support tool. It does not replace clinical judgment, histopathology, EEG, or multidisciplinary epilepsy conference review.

What AutoHS measures

From each participant’s T1w scan, AutoHS extracts:

  • Left hippocampus volume (mm³) — Left-Hippocampus in aseg.stats or aseg+DKT.stats

  • Right hippocampus volume (mm³) — Right-Hippocampus

Volumes come from whole-brain subcortical segmentation (FreeSurfer recon-all volume-only stages, or FastSurfer --seg_only). AutoHS does not use FLAIR, diffusion, or hippocampal subfield segmentation in the default pipeline.

The asymmetry index (AI)

The asymmetry index quantifies relative left–right hippocampal volume difference:

\[\mathrm{AI} = \frac{V_{\mathrm{left}} - V_{\mathrm{right}}}{V_{\mathrm{left}} + V_{\mathrm{right}}}\]

Properties:

  • AI = 0 — equal volumes (perfect symmetry in this metric)

  • AI > 0 — left hippocampus larger than right

  • AI < 0 — right hippocampus larger than left

  • Range is approximately (−1, +1) when both volumes are positive

AutoHS reports AI rounded to four decimal places in derivative JSON and PDF reports.

Two layers of interpretation

AutoHS applies two related but distinct classification layers.

Volume laterality (descriptive)

Uses a symmetric band around zero with threshold ±0.05:

Condition

Label

AI > 0.05

Left > Right

AI < −0.05

Right > Left

−0.05 ≤ AI ≤ 0.05

Symmetric

This layer describes direction of volume imbalance without invoking HS pathology.

HS screening (pathology-oriented)

Uses asymmetric thresholds derived from the AutoHS validation study (Brain Communications, in press). These thresholds are stricter on the negative side because hippocampal atrophy in HS typically makes the affected side smaller:

Condition

HS screening label

AI > 0.046915816971433

Left-dominant (Right HS suspected)

AI < −0.070839747728063

Right-dominant (Left HS suspected)

otherwise

Balanced (No HS)

Naming convention: “Left HS suspected” means pathology on the left hippocampus is suspected, which manifests as right-dominant volumes (smaller left hippocampus, more negative AI). Similarly, “Right HS suspected” corresponds to left-dominant volumes.

These thresholds are fixed constants in ai_compute/asymmetry.py and match the values written to *_desc-autohs_metrics.json.

End-to-end workflow (theory view)

T1w MRI (BIDS)
     │
     ▼
┌─────────────────────────────────────┐
│ Step 1: Whole-brain segmentation    │
│  FreeSurfer vol-only OR FastSurfer    │
│  → subcortical volume labels          │
└─────────────────────────────────────┘
     │
     ▼
┌─────────────────────────────────────┐
│ Step 2: AI-compute                  │
│  Parse Left/Right Hippocampus mm³   │
│  AI = (L − R) / (L + R)             │
│  Laterality + HS screening labels   │
│  JSON / PDF / summary reports       │
└─────────────────────────────────────┘
     │
     ▼
BIDS derivatives under output/autohs/

Step 1 establishes anatomical correspondence and voxel-wise labels; Step 2 is purely tabular arithmetic and rule-based classification on hippocampal volumes. No additional machine-learning model is trained at runtime in Step 2.

Segmentation backends

FreeSurfer (default)

Uses recon-all -autorecon1 -autorecon2-volonly plus mri_segstats on the automated segmentation. This is the conventional neuroimaging approach for subcortical volumes and has extensive literature support.

FastSurfer (--fastsurfer)

Uses a deep-learning segmentation pipeline in --seg_only mode, reading aseg+DKT.stats. It is typically much faster on CPU and does not require a FreeSurfer license, but volumes may differ slightly from FreeSurfer. For research comparability, pick one backend per study and report it in methods (see Methods boilerplate).

Expected inputs and limitations

Inputs

  • One T1w NIfTI per subject/session (BIDS *_T1w.nii.gz)

  • FreeSurfer license when using the default backend

Limitations

  • Single time-point T1w only; no longitudinal atrophy rates

  • No FLAIR hyperintensity or hippocampal T2/T1W signal analysis

  • Thresholds were validated in the AutoHS publication cohort; external validation on your scanner population is recommended before clinical deployment

  • Group-level statistics are not computed (participant-level BIDS App only)

Scientific reference

The asymmetry index formulation and HS screening thresholds implemented in AutoHS are described in:

Ndagijimana P, Brennan D, Shinohara R, Gugger J. MRI derived hippocampal asymmetry identifies hippocampal sclerosis in epilepsy surgical specimens. Brain Communications. Accepted (in press).

See Citation for BibTeX and third-party tool references (FreeSurfer, FastSurfer, BIDS).

Further reading in this documentation