BrainQCNet: A Deep Learning attention-based model for the automated detection of artifacts in brain structural MRI...

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Data Science

Background:

• Structural MRI (sMRI) analysis requires rigorous quality control (QC) for reproducible and generalizable research.
• Manual QC is time-consuming and resource-intensive, especially with growing Big Data in neuroimaging.
• Automated tools are needed to efficiently assess sMRI scan quality.

Purpose of the Study:

• To develop and validate an interpretable Deep Learning (DL) model for automated artifact detection and quality assessment of brain sMRI scans.
• To reduce the time and effort required for manual QC, thereby maximizing data retention and research efficiency.
• To provide a reliable and fast QC tool for the neuroimaging community.

Main Methods:

• Trained an interpretable DL model (ProtoPNet) on manually annotated 2D sMRI slices from the ABIDE 1 dataset (n=980).
• Evaluated the model on ABCD T1-weighted MRI scans (n=2141) with gold-standard manual QC annotations.
• Validated the model's performance on ABIDE 2 (n=799) and ADHD-200 (n=750) T1w MRI datasets.

Main Results:

• Achieved high accuracy in classifying scan quality: 82.4% for good quality (Pass) and 91.4% for medium to low quality (Fail).
• Demonstrated comparable or superior accuracy to existing Machine Learning (ML) models.
• Exhibited fast processing and prediction times (1 minute per scan on a GPU), outperforming traditional DL models in detecting poor quality scans.

Conclusions:

• The developed interpretable DL model provides fast, accurate, and reliable automated quality control for structural MRI scans.
• This tool enhances efficiency in neuroimaging research by enabling maximal data retention and reducing manual QC burden.
• The model is shared as a BIDS-app to support the neuroimaging community in faster and more accurate QC prediction.