Automatic CNN-based detection of cardiac MR motion artefacts using k-space data augmentation and curriculum learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• High-quality medical images are crucial for accurate analysis, especially in large population studies like the UK Biobank.
• Manual identification of motion artifacts in cardiac MRI (CMR) is time-consuming and impractical for large datasets.
• Automated techniques are needed for efficient and reliable medical image quality assessment.

Purpose of the Study:

• To develop and evaluate an automated method for detecting motion-related artifacts in CMR cine images.
• To compare the performance of 3D Convolutional Neural Networks (3D-CNN) and Long-term Recurrent Convolutional Networks (LRCN) for this task.
• To address the challenge of imbalanced data in population studies by proposing novel data augmentation and curriculum learning strategies.

Main Methods:

• Two deep learning architectures, 3D-CNN and LRCN, were investigated for classifying poor-quality CMR images.
• A novel data augmentation scheme involving synthetic k-space artifact creation was implemented.
• A curriculum learning approach using synthetic artifact severity was explored to handle data imbalance.

Main Results:

• The LRCN architecture demonstrated superior performance over the 3D-CNN.
• The LRCN model achieved rapid detection of motion artifacts in 2D+time short axis images (<1ms) with high recall.
• The combination of data augmentation and curriculum learning improved classification performance, yielding an Area Under the ROC Curve of 0.89.

Conclusions:

• Automated detection of motion artifacts in CMR images is feasible and highly effective using deep learning.
• The proposed LRCN-based method, enhanced with novel data augmentation and curriculum learning, significantly improves the efficiency and accuracy of medical image quality assessment.
• This approach offers a scalable solution for quality control in large-scale cardiac imaging studies like the UK Biobank.