Adaptive Knowledge Distillation for High-Quality Unsupervised MRI Reconstruction With Model-Driven Priors

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Deep Learning (DL) and Compressed Sensing (CS) have advanced Magnetic Resonance Imaging (MRI) reconstruction.
• Existing DL-based methods often require large, fully sampled datasets, which are not always available.
• Current unsupervised MRI reconstruction models face limitations in performance, speed, and distribution alignment.

Purpose of the Study:

• To develop an unsupervised method for training competitive MRI reconstruction models.
• To enable end-to-end generation of high-quality MRI samples without fully sampled data.
• To improve the efficiency and effectiveness of unsupervised MRI reconstruction.

Main Methods:

• Teacher models are trained via self-supervised learning on re-undersampled images.
• Knowledge distillation is employed to train a cascade model using undersampled k-space data.
• An adaptive distillation method re-weights samples based on teacher model variance for improved distillation quality.

Main Results:

• The proposed method significantly accelerates MRI reconstruction inference.
• Distilled models demonstrate preserved or improved performance compared to teacher models.
• Experiments show 5%-10% improvements in Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM).
• The distilled models achieve a 10x increase in speed over the teacher models.

Conclusions:

• The unsupervised method effectively trains high-performance MRI reconstruction models.
• The approach overcomes the need for fully sampled training data.
• This technique offers a faster and more efficient solution for MRI reconstruction.