HBCD MR Quality Control Procedures🔗

Raw MR Data QC🔗

Raw MR QC includes automated and manual checks to evaluate unprocessed MRI data. Raw data QC is performed to detect acquisition errors, image artifacts, or corrupted files early so that problematic scans are excluded from the released raw BIDS dataset and downstream processing.

Automated QC🔗

Automated QC is performed at the HBCD Data Coordinating Center (HDCC) immediately after data upload. Following protocol compliance and completeness checks, the following automated QC metrics available in the release are calculated:

Manual Review🔗

Automated metrics flag series for manual review using multivariate prediction and Bayesian classifiers. Trained technicians then score artifact severity on a 0–3 scale (0 = none, 1 = mild, 2 = moderate, 3 = severe). Series with severe artifacts (score = 3) are rejected (QC = 0) and excluded from processing. Data are selected from the remaining series based on manual ratings, notes, and automated scores.

BrainSwipes🔗

Manual visual inspection remains the gold standard for detecting artifacts in structural, functional (e.g., XCP-D), and diffusion (e.g., QSIPrep) derivatives. To support this, derivative visual reports are integrated into BrainSwipes, a gamified, crowdsourced QC platform built on the open-source Swipes For Science framework. BrainSwipes provides an intuitive interface for large-scale studies, guiding users through a short tutorial before they evaluate images and classify them as pass or fail.

Example quality assessment of surface delineation on BrainSwipes platform (displaying brain in axial plane at level of basal ganglia/putamen).

Surface Delineation:

For structural QA, swipers are presented with image slices in coronal, axial, and sagittal planes to assess the accuracy of T1w and T2w surface delineations in differentiating gray and white matter. Images are derived from XCP-D visual reports.

Atlas Registration:

In addition to surface delineation, structural QA also includes atlas registration quality, evaluated by overlaying delineations of the subject’s image onto the atlas, and vice versa. Swipes display nine T1w slices for visual inspection, with three slices per anatomical plane. Quality is assessed based on the alignment of the outer boundaries of the overlaid contours with those of the underlying image, ensuring minimal gaps or misalignments. Images are derived from XCP-D visual reports.

Functional Registration:

Functional registration is evaluated by overlaying outlines of functional images onto structural images and vice versa. Swipes display nine slices of the same functional image for visual inspection, with three slices per anatomical plane. Quality is assessed similarly to structural atlas registration, focusing on the alignment of the overlaid contours. Additional evaluation includes checking for artifacts such as signal dropout. Images are derived from XCP-D visual reports.

Diffusion Direction Encoding (to be included in future release):

Swipes display GIFs of full-resolution T2w images as a grayscale background, with the "Direction Encoded Color" (DEC) map overlaid. These GIFs sweep through a portion of the brain across the three anatomical planes. High-quality processed DWI images exhibit bands of color that closely follow the folds and contours of the grayscale background. These visuals are derived from the QSIPrep report.

Each visual report for a given modality is independently reviewed and rated as a pass or fail, which in the outputs are scored as values of 1 and 0 respectively. BrainSwipes generates a summary of these results that includes the average score as well as number of reviewers for each visual report of each modality.

BrainSwipes QC results are provided as tabulated data: img_brainswipes_xcpd-T2w and img_brainswipes_xcpd-bold. QC scores range from 0 (Fail) to 1 (Pass), averaged across reviewers. For example, a score of 0.6 indicates that 60% of reviewers rated the image as a pass. The data includes overall average QC scores and average number of reviewers for each session-level T2w and BOLD run (summarizing across all visual reports for a given run) as well as the average QC and number of reviewers for each visual report. A Python helper function is provided below to load a BrainSwipes TSV file into a Pandas DataFrame and filter runs with average QC scores above a user-specified threshold:

import pandas as pd

def read_and_filter_tsv(file_path, threshold):
    """
    Reads an HBCD BrainSwipes TSV file into a DataFrame and keeps subject rows where 
    the overall average QC value is greater than or equal to the specified threshold.

    Parameters:
    - file_path (str): Path to the .tsv file
    - threshold (float): Threshold value for filtering

    Returns:
    - pd.DataFrame: Filtered DataFrame
    """
    df = pd.read_csv(file_path, sep='\t')

    fourth_col = df.columns[3]
    return df[df[fourth_col] >= threshold]

# Example usage:
# Filter to keep only subjects with an average structural QC of at least 0.6 (60% pass rate)
filtered_df = read_and_filter_tsv("img_brainswipes_xcpd_T2w.tsv", 0.6)
print(filtered_df.head())

References🔗