Learning Task-Specific Strategies for Accelerated MRI.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Compressed sensing magnetic resonance imaging (CS-MRI) traditionally separates subsampling, reconstruction, and prediction, leading to suboptimal performance.
• Existing CS-MRI methods struggle with end-to-end optimization for downstream diagnostic tasks.

Purpose of the Study:

• To introduce Tackle, a unified co-design framework for jointly optimizing CS-MRI subsampling, reconstruction, and prediction.
• To enhance the performance of CS-MRI on various downstream diagnostic tasks through integrated optimization.

Main Methods:

• Developed a unified framework (Tackle) for co-designing subsampling, reconstruction, and prediction strategies.
• Implemented a two-stage training procedure: generic pre-training (image reconstruction) followed by task-specific fine-tuning.
• Validated the framework on multiple public MRI datasets and a newly collected dataset with different acquisition setups.

Main Results:

• Tackle demonstrated improved performance across various diagnostic tasks compared to traditional CS-MRI methods.
• The framework exhibited robustness to distribution shifts, generalizing effectively to new datasets.
• A 4x accelerated MRI sequence was implemented, reducing scan time from 335 to 84 seconds with maintained high performance.

Conclusions:

• Tackle offers a superior, unified approach to CS-MRI, enhancing both diagnostic accuracy and efficiency.
• The co-design framework provides significant speed improvements and maintains high image quality, crucial for clinical applications.
• The study highlights the potential of integrated optimization for advancing medical imaging technologies.