[Paper Review] Zero-Shot Self-Supervised Learning for MRI Reconstruction

Burhaneddin Yaman, Seyed Amir Hossein Hosseini+1 more|arXiv (Cornell University)|Feb 15, 2021

Medical Imaging Techniques and ApplicationsMedicine22 references32 citations

TL;DR

Proposes a zero-shot self-supervised learning (ZS-SSL) framework for subject-specific accelerated MRI reconstruction from a single undersampled scan, with an automated early-stopping validation and optional transfer learning. It avoids external training data and improves robustness to acquisition variations.

ABSTRACT

Deep learning (DL) has emerged as a powerful tool for accelerated MRI reconstruction, but often necessitates a database of fully-sampled measurements for training. Recent self-supervised and unsupervised learning approaches enable training without fully-sampled data. However, a database of undersampled measurements may not be available in many scenarios, especially for scans involving contrast or translational acquisitions in development. Moreover, recent studies show that database-trained models may not generalize well when the unseen measurements differ in terms of sampling pattern, acceleration rate, SNR, image contrast, and anatomy. Such challenges necessitate a new methodology to enable subject-specific DL MRI reconstruction without external training datasets, since it is clinically imperative to provide high-quality reconstructions that can be used to identify lesions/disease for \emph{every individual}. In this work, we propose a zero-shot self-supervised learning approach to perform subject-specific accelerated DL MRI reconstruction to tackle these issues. The proposed approach partitions the available measurements from a single scan into three disjoint sets. Two of these sets are used to enforce data consistency and define loss during training for self-supervision, while the last set serves to self-validate, establishing an early stopping criterion. In the presence of models pre-trained on a database with different image characteristics, we show that the proposed approach can be combined with transfer learning for faster convergence time and reduced computational complexity. The code is available at \url{https://github.com/byaman14/ZS-SSL}.

Motivation & Objective

• Motivate a subject-specific MRI reconstruction method that does not rely on external datasets.
• Develop a zero-shot self-supervised training framework that partitions k-space data for training, loss calculation, and validation.
• Provide an automated stopping criterion to prevent overfitting without external validation data.
• Demonstrate robustness and generalizability to varying sampling patterns, acceleration rates, SNR, contrast, and anatomy.
• Show how transfer learning can speed up convergence when a pretrained model is available.

Proposed method

• Partition the acquired k-space from a single scan into three disjoint sets: Θ for data-consistency updates, Λ for loss calculation, and Γ for self-validation.
• Train an unrolled physics-guided DL MRI reconstruction network using multiple (Θk, Λk) masks generated from Ω∖Γ (K partitions) with a data-consistency loss over Λk.
• Define an automated early-stopping criterion by monitoring a k-space validation loss on Γ while optimizing over weights θ.
• Optionally initialize from pretrained models via transfer learning (ZS-SSL-TL) to accelerate convergence and reduce computation.
• Use CG-SENSE/ResNet-based blocks in a 10-iteration unrolled architecture with complex-valued inputs (real+imag) for knee and brain MRI; end-to-end training with a normalized loss (ℓ1–ℓ2) and Adam optimizer.
• Demonstrate training without fully-sampled references and evaluate with PSNR and SSIM.

Experimental results

Research questions

• RQ1 Can zero-shot, subject-specific DL MRI reconstruction achieve competitive quality without external training data?
• RQ2 Does a three-set k-space partition enable effective self-supervision and automated stopping to prevent overfitting?
• RQ3 Can transfer learning further improve convergence and reduce computation when a pretrained model exists?
• RQ4 How does ZS-SSL-TL perform under domain shifts in sampling pattern, acceleration, contrast, SNR, and anatomy compared to traditional supervised methods?

Key findings

• ZS-SSL achieves artifact-free and high-quality reconstructions comparable to database-trained methods when training and testing data match.
• ZS-SSL-TL converges faster than ZS-SSL, reducing total training time.
• ZS-SSL-TL alleviates banding artifacts and residual aliasing more effectively than some supervised approaches under domain shifts.
• ZS-SSL and ZS-SSL-TL show superior robustness to changes in sampling pattern, acceleration, contrast, SNR, and anatomy compared to database-trained PG-DLR.
• Subject-specific ZS-SSL eliminates the need for large training datasets and is compatible with scan-specific transfer learning to balance quality and compute.