Optimal MRI undersampling patterns for ultimate benefit of medical vision tasks

Area of Science:

• Medical Imaging
• Computer Vision
• Magnetic Resonance Imaging

Background:

• Compressed sensing accelerates MRI by undersampling k-space, typically prioritizing reconstructed image quality.
• Current methods focus on enhancing overall image fidelity post-reconstruction.

Purpose of the Study:

• To reframe MRI acceleration by optimizing k-space undersampling patterns for downstream image analysis outcomes.
• To improve the detection and localization of pathologies in accelerated MRI scans.

Main Methods:

• Developed a novel approach to optimize undersampling patterns based on performance in medical vision tasks (reconstruction, segmentation, classification).
• Introduced an iterative gradient sampling routine for universal applicability across these tasks.
• Validated the paradigm on three medical datasets.

Main Results:

• Demonstrated significant improvements in target metrics at high acceleration factors.
• Achieved up to a 12% improvement in Dice score for segmentation at 16x acceleration compared to other patterns.
• Showcased the effectiveness of the proposed MRI acceleration strategy.

Conclusions:

• Optimizing k-space undersampling for specific downstream analysis tasks enhances diagnostic utility.
• The proposed method offers a superior MRI acceleration paradigm, particularly for high acceleration factors.
• This approach advances medical image analysis by prioritizing clinically relevant outcomes.