A supervised learning approach for diffusion MRI quality control with minimal training data.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Diffusion MRI quality control (QC) is challenging due to numerous artifacts and large data volumes.
• Manual data inspection, the current gold standard, is time-consuming and subjective.
• Supervised learning, particularly convolutional neural networks, shows promise but requires extensive manual labeling.

Purpose of the Study:

• To reduce the need for manual data labeling in diffusion MRI QC.
• To develop an AI-based QC method that utilizes simulated data for training.
• To assess the performance of this method in detecting movement artifacts.

Main Methods:

• Training a convolutional neural network classifier on simulated diffusion MRI data.
• Employing a small set of manually labeled real data for final model calibration.
• Comparing the performance against a classifier trained solely on manually labeled real data.

Main Results:

• Demonstrated that training on simulated data significantly reduces the need for manual labeling.
• Showcased the potential of the proposed method for detecting severe movement artifacts.
• Provided a performance comparison with traditional supervised learning approaches.

Conclusions:

• AI-based QC for diffusion MRI can be effectively trained using simulated data, minimizing manual annotation efforts.
• This approach offers a more efficient and potentially more objective alternative to manual inspection.
• Further development could enhance the reliability and applicability of automated QC in diffusion MRI.