BrainQCNet

BrainQCNet trains an interpretable Deep Learning model to classify minimally preprocessed 2D slices of brain structural MRI scans for automated artifact detection and quality control.

Key Features:

• Attention-based architecture: Uses ProtoPNet, a prototype-based neural network that provides interpretable predictions through learned attention mechanisms.
• Binary quality classification: Categorizes scans as Pass (good quality) or Fail (medium to low quality) based on artifact presence.
• Fast processing speed: Achieves prediction in approximately 1 minute per scan on GPU hardware with CUDA compatibility.
• BIDS-app distribution: Shared as a BIDS-compliant application for integration into neuroimaging workflows.

Scientific Applications:

• Large-scale neuroimaging quality control: Enables rapid, automated QC evaluation of big datasets such as ABIDE and ABCD to maximize data retention and reproducibility.
• Multi-cohort validation: Validated on T1-weighted MRI scans from ABIDE 1 (n=980), ABIDE 2 (n=799), ADHD-200 (n=750), and ABCD (n=2141).
• Artifact detection in structural MRI: Identifies poor quality scans with accuracy comparable to or exceeding existing machine learning models.

Methodology:

The model was trained on manually annotated, minimally preprocessed 2D slices from ABIDE 1. The ProtoPNet architecture was applied to classify slices, with evaluation against gold-standard manual QC annotations from ABCD and additional validation on ABIDE 2 and ADHD-200 cohorts.