Information-Theoretic Optimization for Task-Adapted Compressed Sensing Magnetic Resonance Imaging.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Information Theory

Background:

• Compressed sensing MRI (CS-MRI) reduces scan times but faces challenges in uncertainty prediction for clinical tasks.
• Existing methods lack adaptive sampling and end-to-end optimization for versatile clinical applications.

Purpose of the Study:

• To develop the first task-adapted CS-MRI framework using information theory for probabilistic inference and adaptive sampling.
• To address uncertainty in medical diagnosis and adapt to various sampling ratios and clinical tasks.

Main Methods:

• Formalized task-adapted CS-MRI by maximizing mutual information between undersampled k-space data and clinical tasks.
• Utilized amortized optimization and variational bounds for joint optimization of sampling, reconstruction, and task-inference models.
• Enabled flexible sampling ratio control and addressed joint task/reconstruction and privacy-preserving task implementation.

Main Results:

• Achieved competitive performance on standard metrics (e.g., Dice) compared to deterministic methods.
• Demonstrated superior distribution matching to ground-truth posterior distributions using generalized energy distance (GED).
• Showcased adaptability to arbitrary sampling ratios and diverse clinical scenarios within a unified framework.

Conclusions:

• The proposed information-theoretic framework offers a robust solution for task-adapted CS-MRI, improving uncertainty prediction and diagnostic accuracy.
• The method provides flexibility in sampling strategies and supports both enhanced reconstruction and privacy-preserving task implementation.
• This approach advances CS-MRI by integrating probabilistic inference and adaptive sampling for broader clinical utility.