Anatomy-Aware Deep Unrolling for Task-Oriented Acceleration of Multi-Contrast MRI.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Multi-contrast magnetic resonance imaging (MC-MRI) is vital in clinical practice but suffers from long scan times.
• Current accelerated MC-MRI methods often lack personalization, failing to prioritize specific clinical targets.
• Existing approaches focus on general image quality enhancement, neglecting specific pathologies or anatomical regions of interest.

Purpose of the Study:

• To develop a personalized and accelerated MC-MRI method tailored to individual clinical needs.
• To enhance the imaging quality of specific targets of interest (TOIs) in MC-MRI.
• To integrate deep learning with iterative reconstruction for efficient and task-oriented MC-MRI.

Main Methods:

• An anatomy-aware unrolling-based deep network (A²MC-MRI) was proposed, integrating a learnable group sparsity and an anatomy-aware denoising prior.
• A segmentation network within the denoising prior provides location information for TOI-enhanced denoising.
• The unrolled network was jointly learned with k-space sampling patterns for task-oriented reconstruction.

Main Results:

• A²MC-MRI demonstrated state-of-the-art performance in MC-MRI reconstruction under high acceleration rates.
• The method achieved notable enhancements in imaging quality for specific targets of interest (TOIs).
• Comprehensive evaluations on public and in-house datasets validated the network's effectiveness.

Conclusions:

• A²MC-MRI offers a promising solution for fast, personalized MC-MRI tailored to clinical requirements.
• The proposed network provides improved interpretability and learning capacity for accelerated MC-MRI.
• This approach significantly enhances TOI imaging quality, addressing limitations of current methods.