Skip to content

PROJECT_OVERVIEW.md

Latest commit

 

History

History
71 lines (60 loc) · 3.35 KB

PROJECT_OVERVIEW.md

File metadata and controls

71 lines (60 loc) · 3.35 KB
Raw

Project Overview

What This Project Does

This repository implements an end-to-end, BIDS-native fMRI analysis platform for two study settings:

  • Resting-state schizophrenia vs healthy controls (SchizConnect-style cohorts)
  • Naturalistic movie fMRI ISC in healthy controls (Algonauts-style data)

The pipeline intentionally stops at group-level results. It does not perform symptom prediction, clinical outcome modeling, or any downstream machine-learning classifier beyond the requested inferential metrics.

Why It Is Structured This Way

The system follows a modular product-platform pattern:

  • Reliability layer: preprocessing + quality control so all downstream features consume standardized, motion-screened signals.
  • Feature modules: independent metric generators (ROI, ReHo, static FC, dFC, ICA, PCA, ISC).
  • Decision layer: group-level inference and correction that turns features into interpretable findings.

This design allows adding/removing metrics without rewriting ingestion or group statistics.

Milestones Implemented

  1. BIDS ingestion and harmonization
  2. Preprocessing and QC
  3. ROI time-series extraction (Schaefer-200)
  4. ReHo map computation
  5. Static FC (Pearson + Fisher-z)
  6. Dynamic FC (30 TR / 5 TR)
  7. Subject ICA + cross-subject matching (hierarchical clustering on spatial correlation)
  8. Subject PCA (top-5 explained variance)
  9. ISC leave-one-out + permutation null
  10. Group-level statistics + visualization/tables
  11. Reproducibility validation suite (Table 1: FC, ReHo, ICA, graph, dFC, network anchor)

Core Scientific Rules Enforced

  • Confounds: Friston-24 + WM/CSF (if present)
  • Scrubbing threshold: FD > 0.5 mm
  • Exclusion: >20% censored OR max translation >3 mm OR max rotation >3 deg
  • Temporal filtering: 0.01-0.10 Hz (TR-aware)
  • Spatial smoothing: 6 mm FWHM
  • Atlas: Schaefer-200
  • Edge-level multiple testing: FDR q<0.05
  • Voxelwise correction: FDR-based maps in current implementation

Inputs and Outputs

Inputs

  • BIDS roots (raw)
  • fMRIPrep derivatives (required at runtime)
  • Optional phenotypic TSV (diagnosis + covariates)

Outputs

Under derivatives/metrics:

  • manifests/: reusable run/subject manifests between milestones
  • qc/: run-level QC tables and plots
  • metric folders: roi_timeseries, reho, static_fc, dynamic_fc, ica, pca, isc
  • tables/: group-level inferential tables
  • figures/: group-level figures
  • group_stats/: voxelwise statistical maps

Under reports/reproducibility/:

  • scorecard.md / scorecard.csv: Table 1 pass/fail summary
  • fc_within_vs_between.csv, reho_summary.csv, ica_stability_seeds.csv
  • ica_stability_lorocv.csv, graph_metrics_bootstrap.csv
  • dfc_sensitivity.json, network_anchor_summary.csv
  • manifest.json: git SHA, package versions, run timing

Validated Run Context

Pipeline has been validated on OpenNeuro ds007318 (working-memory removal task, Northwest Normal University):

  • 3 subjects (sub-01, sub-02, sub-03); 5 runs total (sub-01: 2 sessions, sub-02: 1, sub-03: 2)
  • All participants are labeled as clinical patients; no healthy control arm
  • Group-level statistics disabled for this dataset (stats.diagnosis_column: __no_groups__)
  • ISC skipped (task paradigm, not naturalistic movie stimulus)
  • Full reproducibility suite completed; results reported in Table 1 of the manuscript

Mean QC: FD 0.10–0.15 mm; 0–1.2% volumes scrubbed; DVARS 3.68–4.41