Scan-Specific MRI Reconstruction using Zero-Shot Physics-Guided Deep Learning

Burhaneddin Yaman, Seyed Amir Hossein Hosseini, Mehmet Akçakaya

Abstract

Deep learning (DL) has emerged as a powerful tool for accelerated MRI reconstruction, but these methods often necessitate 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 recently developed translational acquisitions. Moreover, 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 that can enable scan-specific DL MRI reconstruction without any external training datasets. In this work, we propose a zero-shot self-supervised learning approach to perform scan-specific accelerated MRI reconstruction to tackle these issues. The proposed approach splits available measurements for each scan into three disjoint sets. Two of these sets are used to enforce data consistency and define loss during training, while the last set is used to establish 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 to further improve reconstruction quality.

Introduction

Magnetic Resonance Imaging (MRI) is a vital diagnostic tool, but its utility is often limited by long acquisition times. To accelerate the process, clinicians "undersample" k-space data, which typically results in aliasing artifacts. While Deep Learning (DL) has revolutionized the reconstruction of these undersampled images, standard supervised models suffer from two major bottlenecks: a dependency on massive databases of "fully sampled" reference scans and a lack of generalization when faced with new scanners, different anatomies, or varying noise levels.

Methodology: Zero-Shot Self-Supervised Learning (ZS-SSL)

The proposed ZS-SSL framework eliminates the need for external training data by treating each individual scan as its own mini-training set. The core of the approach involves partitioning the acquired undersampled k-space data ($\Omega$) into three disjoint sets:

  1. Training Set ($\Theta$): Used within the network's Data Consistency (DC) units to ensure the reconstruction adheres to the physical measurements.
  2. Loss Mask ($\Lambda$): Used to calculate the self-supervised training loss by comparing the network output to the actual acquired data at these locations.
  3. Validation Mask ($\Gamma$): An independent set used to monitor for overfitting and trigger an automated early stopping criterion.
The ZS-SSL framework partitions acquired k-space data into three sets for training, loss calculation, and validation, feeding into an unrolled network architecture.

This framework utilizes Physics-Guided Deep Learning Reconstruction (PG-DLR), where a traditional optimization algorithm is "unrolled" into a series of network layers. This ensures that the reconstruction process remains grounded in the actual physics of MRI while benefiting from the representational power of neural networks.

Key Insights: Early Stopping and Transfer Learning

A critical challenge in scan-specific training is avoiding the "memorization" of noise. The authors discovered that while training loss consistently decreases, the validation loss on the $\Gamma$ set eventually begins to rise. This behavior provides a mathematically grounded "early stopping" point, preventing the model from overfitting to the specific noise patterns of the scan.

Training and validation loss curves for ZS-SSL and ZS-SSL-TL. Red arrows indicate the early stopping points determined by the validation loss.

To address the computational time required for per-scan training, the authors introduced ZS-SSL-TL (Transfer Learning). By using weights from a pre-trained model (even one from a different anatomy) as a starting point, the network converges significantly faster, reducing reconstruction time from roughly 640 seconds to 85 seconds per slice.

Results and Performance

Evaluated on the fastMRI database (Knee and Brain scans), ZS-SSL achieved reconstruction quality nearly identical to supervised models trained on hundreds of full datasets. For instance, on Coronal PD knee scans, ZS-SSL reached a PSNR of 39.55, compared to 39.97 for a fully supervised model.

More importantly, ZS-SSL demonstrated superior robustness to "domain shift." In tests where a brain-trained model was used to reconstruct knee images, standard supervised models failed and produced heavy artifacts. In contrast, ZS-SSL-TL successfully adapted to the new data, producing clean, high-fidelity images without the "banding" artifacts common in traditional DL methods.

Conclusion

ZS-SSL represents a significant step toward making deep learning MRI reconstruction clinically practical. By enabling high-quality, scan-specific reconstruction without external databases, it overcomes the hurdles of data scarcity and model brittleness. While future work is needed to reduce computational overhead and extend the method to 3D and dynamic imaging, ZS-SSL provides a robust path forward for personalized, physics-guided medical imaging.

Summary updated

Abstract

Deep learning (DL) has emerged as a powerful tool for accelerated MRI reconstruction, but these methods often necessitate 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 recently developed translational acquisitions. Moreover, 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 that can enable scan-specific DL MRI reconstruction without any external training datasets. In this work, we propose a zero-shot self-supervised learning approach to perform scan-specific accelerated MRI reconstruction to tackle these issues. The proposed approach splits available measurements for each scan into three disjoint sets. Two of these sets are used to enforce data consistency and define loss during training, while the last set is used to establish 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 to further improve reconstruction quality.
BibTeX
@inproceedings{yaman2022scan,
  author = {Burhaneddin Yaman and Seyed Amir Hossein Hosseini and Mehmet Akçakaya},
  title = {{Scan-Specific MRI Reconstruction using Zero-Shot Physics-Guided Deep Learning}},
  booktitle = {International Conference on Learning Representations},
  year = {2022},
  url = {https://openreview.net/forum?id=085y6YPaYjP},
}
Figure