Estimating Task-based Performance Bounds for Accelerated MRI Image Reconstruction Methods by Use of Learned-Ideal...

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence in Healthcare

Background:

• Objective image quality (IQ) measures are crucial for optimizing medical imaging systems.
• The ideal observer (IO) performance sets theoretical limits for image reconstruction, guiding system design.
• Deep learning (DL) methods for image reconstruction can obscure diagnostic information loss in under-sampled data.

Purpose of the Study:

• To explore the application of data space convolutional neural network (CNN) approximated IOs (CNN-IOs) for multi-coil magnetic resonance imaging (MRI).
• To establish task-based performance bounds for image reconstruction in MRI using CNN-IO analysis.
• To evaluate the impact of accelerated data acquisition on IO performance in MRI.

Main Methods:

• Utilized stylized multi-coil sensitivity encoding (SENSE) MRI systems.
• Employed deep-generated stochastic brain models for simulations.
• Applied signal-known-statistically and background-known-statistically (SKS/BKS) binary signal detection tasks.
• Investigated the performance of data space CNN-IOs under varying acceleration factors.

Main Results:

• Demonstrated the feasibility of using data space CNN-IO analysis for multi-coil MRI.
• Quantified the impact of acceleration factors on IO performance in MRI.
• Provided a method to identify MRI acquisition designs that preserve diagnostic information.

Conclusions:

• Data space CNN-IO analysis is a valuable tool for optimizing MRI systems and reconstruction methods.
• This approach can help ensure diagnostic image quality even with aggressive under-sampling and acceleration.
• The findings are critical for the development of advanced MRI techniques and medical device submissions.