Robust Physics-Based Deep MRI Reconstruction via Diffusion Purification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Science

Background:

• Deep learning (DL) methods significantly improve magnetic resonance imaging (MRI) reconstruction.
• Existing DL models are vulnerable to noise, variations in acquisition parameters, and distribution shifts from unseen data.
• Robustness remains a critical challenge for reliable DL-based MRI reconstruction.

Purpose of the Study:

• To develop a robustification strategy for DL-based MRI reconstruction using diffusion models (DMs).
• To enhance the resilience of DL MRI reconstruction against worst-case perturbations and distribution shifts.
• To introduce a novel method, RODIO (robust DL-based MRI with diffusion purification).

Main Methods:

• Leveraging pretrained diffusion models (DMs) as purifiers for MRI data.
• Implementing a robustification strategy that requires efficient fine-tuning on purified examples.
• Comparing the proposed method against standalone diffusion-based reconstructors and adversarial training (AT) and randomized smoothing (RS).

Main Results:

• The proposed RODIO method significantly improves the robustness of DL-based MRI reconstruction.
• RODIO outperforms existing leading robustification techniques, including AT and RS.
• Demonstrated adaptability across multiple DL reconstruction models and compatibility with accelerated diffusion samplers.
• Showcased robustness to unseen lesions and effectiveness in unsupervised generative reconstruction.

Conclusions:

• Diffusion model purification offers an effective and efficient approach to robustify DL-based MRI reconstruction.
• RODIO provides a promising solution to enhance the reliability and generalizability of DL MRI methods.
• The strategy successfully addresses vulnerabilities without the complexity of minimax optimization problems.