Neuroimaging Dataset

Overview

The MRI component of BAARD leverages data from the OPTIMUM dataset and includes the following modalities:

• Structural MRI (T1-weighted)
  
  
• Resting-state functional MRI (rs-fMRI)
  
  
• Diffusion-weighted imaging (DWI)

Only baseline scans were used for training and modeling purposes. Features from these scans were extracted and aligned with clinical data using a shared record_id key. These features contribute to downstream modeling of treatment response and biotyping.

Processing Pipeline

Baseline OPTIMUM scans were processed using the HCP-ABCD pipeline. Freesurfer outputs from this were used to extract anatomical statistics. Functional scans were also processed via the ABCD pipeline, followed by denoising with XCP. The denoised outputs (TSVs) used Schaefer-156 parcellation and were processed into ROI-to-ROI and network-level connectivity matrices. Diffusion scans were preprocessed with QSIPREP, and tractography was performed using probabilistic methods.

All code used for the processing pipeline will be committed to the BAARD github repository. For more information or detailed breakdown of the processing pipeline please e-mail Hassan.Abdulrasul@camh.ca

Path to outputs

Each modality directory contains:

• A raw/ subfolder: cleaned and complete columns processed data.
  
  
• A processed/ subfolder: data aligned with clinical IDs (record_id), cleaned, and formatted for integration with the master sheet. Moreover, selected columns are used.

Processed Outputs

Structural MRI (sMRI) :

The sMRI dataset consists of FreeSurfer-derived metrics, quality control indicators, and corrected regional brain measurements for each participant at baseline. The processed data to use is OPT_baseline_selected_thickness.csv

Key Columns Include:

• record_id: Participant identifier
  
  
• mr_date: MRI acquisition date
  
  
• QC: Quality flag (1 = pass, 2 = noisy, 3 = fail)
  
  
• QC_SurfaceHoles: Mesh integrity indicator
  
  

Features Include:

• Global cortical thickness (e.g., lh_MeanThickness_thickness, total_meanthickness)
• Cortical thickness for Desikan-Killiany atlas regions
  Subcortical volumes normalized by eTIV and converted to liters
  (e.g., Left.Caudate_etiv = (Caudate / eTIV) * 1000)

Processed file location:
/external/rprshnas01/netdata_kcni/dflab/data/BAARD/mri/smri/processed

Functional MRI (fMRI) :

The fMRI dataset includes resting-state functional connectivity metrics derived from the OPTIMUM Neuro baseline data. Processing was conducted using the ABCD pipeline, followed by denoising with XCP, and connectivity metrics were extracted using the Schaefer 100-parcel atlas, along with subcortical and cerebellar ROIs.

 Output Files

Network-level connectivity: Average within- and between-network functional connectivity across 7 canonical brain networks (e.g., DMN, Control, Salience, Visual).File: OPT_baseline_connectivity_Network_Connectivity.csvNumber of features: 29 (includes record_id)

ROI-to-ROI matrix: Full pairwise connectivity matrix across Schaefer regions and additional subcortical/cerebellar ROIs. File: OPT_baseline_connectivity_Processed_Connectivity.csvNumber of features: 12091 (includes record_id)

Processed file location:
/external/rprshnas01/netdata_kcni/dflab/data/BAARD/mri/fmri/processed

MRI Data Dictionary

A full MRI data dictionary — including descriptions of parcellation variables, QC flags — is maintained in markdown and .txt format. These will be converted into human-readable HTML or CSV and integrated into Confluence once the formatting space is finalized.

Attached Here:

BAARD_structural_mri_dictionary.mdBAARD_connectivity_dictionary.md

MRI Quality Control (QC) Summary

QC was based on a standardized visual inspection protocol used across the Kimel and TIGR Lab. These criteria are consistent with established best practices for assessing ABCD pipeline outputs. Following this resting-state data were processed through the ABCD pipeline and denoised using XCP.

XCP QC Checks:

Participants with excessive head motion or corrupted outputs were automatically excluded during processing. If XCP failed to produce connectivity estimates, this indicates insufficient signal and/or excessive noise.

Correlation calculations.

Pairwise ROI-to-ROI correlation matrices and within-network connectivity scores were extracted per participant, this was performed using a python script (link to github here when ready).

A post-hoc filtering step was applied:
Participants with all-zero or near-zero correlation values were flagged and filtered out.These cases typically reflect unusable data due to low signal or failed denoising.