Modality Features: Structural MRI (sMRI)

Source Inputs

• FreeSurfer / CIVET outputs path.
• Atlas / parcellation used (e.g., Desikan, Schaefer-400 volumetric maps).
• QC reports (Euler number thresholds, manual edits).

Feature Extraction

• ROI metrics: cortical thickness, volume, surface area, subcortical volumes.
• Derived composites (asymmetry indices, global measures) if needed.
• File formats (.tsv, .csv, HDF5) and loader scripts.

Preprocessing & Harmonization

• Z-score within train folds; residualize covariates (age, sex, ICV, site/scanner, SES).
• Optionally apply ComBat or mixed-effects models; document rationale.
• Handle missing ROIs (impute, drop subject, or add mask).

Embedding / Projection

• Stack ROI vectors per subject; optionally reduce via PCA to 512-D (fit per train fold).
• Alternative: use pretrained encoders (Brain Harmony hub tokens) when available; document pooling strategy.
• Persist embeddings + scalers in artifacts/generated/embeddings/smri/.

Covariates

• Minimum set: age, sex, intracranial volume, site/scanner, scanner software version.
• Additional: handedness, SES, acquisition batch.

Integration Notes

• Use same subject IDs as genomics/fMRI intersections; log any exclusions.
• Provide YAML hooks for experiments: kb/datasets/<cohort>.yaml, configs/projectors/smri_pca512.yaml.
• Mention reliability metrics (test-retest ICC if available).

References

• Documentation for FreeSurfer pipeline version.
• Papers motivating ROI selection / harmonization steps.