Multi-task magnetic resonance imaging reconstruction using meta-learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Reconstructing Magnetic Resonance Imaging (MRI) data with single-task deep learning is challenging due to poor generalizability across different imaging sequences and contrasts.
• Dissimilarity among datasets with varying contrasts leads to suboptimal learning performance in conventional deep learning models.

Purpose of the Study:

• To propose a meta-learning approach for efficient feature learning from multiple MRI datasets.
• To enable simultaneous reconstruction of MRI images acquired with different sequences and contrasts using multi-task learning.

Main Methods:

• Developed a proximal gradient descent-inspired optimization method for learning image features in both image and k-space domains.
• Incorporated meta-learning ('learning-to-learn') to enhance the learning of shared features across multiple image contrasts.
• Implemented a multi-task learning framework to simultaneously reconstruct images from diverse MRI datasets.

Main Results:

• The proposed multi-task meta-learning approach significantly outperforms state-of-the-art single-task learning methods, especially at high acceleration rates.
• Achieved accurate and detailed reconstructions with the lowest pixel-wise errors and enhanced qualitative performance across all tested acceleration rates.
• Successfully reconstructed highly-undersampled k-space data from multiple MRI datasets simultaneously.

Conclusions:

• The multi-task meta-learning framework offers superior performance and generalizability for MRI reconstruction compared to single-task methods.
• This approach effectively addresses the challenges of reconstructing MRI data with diverse contrasts and high undersampling rates.
• The developed method demonstrates a significant advancement in automated MRI reconstruction.