Motion Artifact Detection for T1-Weighted Brain MR Images Using Convolutional Neural Networks

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Quality assessment (QA) of magnetic resonance imaging (MRI) is crucial but often lacks standardization, relying on manual inspection.
• Large datasets exacerbate challenges in manual MRI quality evaluation due to time and expertise requirements.
• Automated methods, particularly machine learning with convolutional neural networks (CNNs), offer a promising solution for consistent MRI QA.

Purpose of the Study:

• To develop and evaluate a CNN for automated detection of random head motion artifacts (RHM) in T1-weighted MRI.
• To assess the performance of the CNN in identifying images with pronounced motion artifacts.
• To explore the feasibility of a multi-class classification for nuanced artifact evaluation.

Main Methods:

• A CNN model was trained to detect RHM in 420 T1-weighted whole-brain MRI volumes.
• A two-step approach was employed: first, identifying images with significant artifacts, then evaluating a three-class classification.
• Human experts manually classified artifact prominence to create ground truth labels.

Main Results:

• The CNN achieved 95% accuracy in identifying MRI volumes with pronounced random head motion artifacts.
• A subsequent three-class classification, including an intermediate artifact level, maintained 76% accuracy.
• The model demonstrated high efficacy in flagging images requiring closer inspection.

Conclusions:

• CNN-based automated QA shows significant potential for enhancing efficiency in post-hoc analysis of large MRI datasets.
• Automated detection of motion artifacts can streamline the identification of lower-quality scans.
• This approach can improve the reliability and scalability of MRI quality control.