A Deep Cascade of Convolutional Neural Networks for Dynamic MR Image Reconstruction (1704.02422v2)

Abstract: Inspired by recent advances in deep learning, we propose a framework for reconstructing dynamic sequences of 2D cardiac magnetic resonance (MR) images from undersampled data using a deep cascade of convolutional neural networks (CNNs) to accelerate the data acquisition process. In particular, we address the case where data is acquired using aggressive Cartesian undersampling. Firstly, we show that when each 2D image frame is reconstructed independently, the proposed method outperforms state-of-the-art 2D compressed sensing approaches such as dictionary learning-based MR image reconstruction, in terms of reconstruction error and reconstruction speed. Secondly, when reconstructing the frames of the sequences jointly, we demonstrate that CNNs can learn spatio-temporal correlations efficiently by combining convolution and data sharing approaches. We show that the proposed method consistently outperforms state-of-the-art methods and is capable of preserving anatomical structure more faithfully up to 11-fold undersampling. Moreover, reconstruction is very fast: each complete dynamic sequence can be reconstructed in less than 10s and, for the 2D case, each image frame can be reconstructed in 23ms, enabling real-time applications.

A Deep Cascade of Convolutional Neural Networks for Dynamic MR Image Reconstruction

The paper "A Deep Cascade of Convolutional Neural Networks for Dynamic MR Image Reconstruction" proposes an innovative framework to tackle the challenges posed by reconstructing dynamic cardiac MR sequences from undersampled data. This framework leverages advances in deep learning and convolutional neural networks (CNNs) to accelerate the acquisition process significantly while maintaining high image quality.

Framework Overview

The central proposition of the paper is a deep cascade of CNNs applied to reconstruct dynamic sequences of 2D cardiac MR images. The objective is to outperform traditional compressed sensing (CS) techniques in terms of both reconstruction error and speed. The essential components of the framework include:

1. Independent Reconstruction using CNNs: Initially, each 2D image frame is reconstructed independently using CNNs. This method already demonstrates superior performance compared to state-of-the-art 2D compressed sensing methods, such as dictionary learning-based MR image reconstruction.
2. Joint Reconstruction using Cascaded CNNs: For enhanced performance, the authors propose a cascading network where CNNs iteratively refine the reconstruction by incorporating both spatial and temporal correlations. This strategy significantly boosts the fidelity of preserved anatomical structures even under aggressive 11-fold undersampling.
3. Data Consistency Layer: To integrate prior knowledge from the acquired $k$-space data effectively, a data consistency (DC) layer is introduced. This layer enforces that the reconstructed images remain faithful to the actual acquired measurements, modulated by noise levels. The DC layer ensures the network respects the original $k$-space data while refining the images iteratively.
4. Data Sharing Layer: For dynamic imaging scenarios, a data sharing (DS) layer is incorporated to exploit temporal redundancy. This layer creates aggregated images from neighboring frames to approximate missing $k$-space samples, thus aiding the CNNs in resolving aliasing artifacts more effectively.

Numerical Results

The numerical assessment demonstrates several strong points:

• Reconstruction Quality: The proposed method delivers superior reconstruction results. In particular, the framework consistently outperforms traditional methods like DLMRI, DLTG, kt-SLR, and L+S across various undersampling rates. For instance, in 9-fold acceleration factors, the CNN model adeptly preserves anatomical details, significantly reducing the mean squared error (MSE).
• Speed: The reconstruction times are drastically reduced. A 2D MR image can be processed in approximately 23 ms, enabling real-time applications. For dynamic sequences, the total reconstruction time is about 8.21 seconds per complete sequence, demonstrating substantial efficiency compared to existing methods like DLTG, which takes several hours per subject.

Implications and Future Developments

Practical Implications

The practical implications of this research are noteworthy:

• Clinical Workflow: The significantly reduced reconstruction times facilitate real-time applications, which can enhance the clinical workflow by allowing rapid assessment and diagnosis.
• Scalability: Given the capability of the model to generalize well across different undersampling patterns and noise levels, it can be readily adapted for various MR imaging protocols.
• Potential for Integration: The framework's design supports potential integration with existing MR systems, offering a clear pathway for implementation in clinical settings.

Theoretical Implications

From a theoretical perspective, the paper highlights several critical advancements:

• Learning Framework: The combination of DC and DS layers illustrates a powerful method for embedding domain knowledge into deep learning models, enhancing their robustness and accuracy.
• Implications for Deep Learning in Medical Imaging: The work contributes to a growing body of literature that demonstrates the effectiveness of deep learning frameworks in solving traditionally complex medical imaging problems.

Future Developments

Anticipating future developments:

• Multi-channel Data Integration: Extending the framework to handle multi-coil data could further improve reconstruction quality for parallel imaging scenarios.
• Optimized Sampling: Potential exists for combining the network with optimized sampling strategies such as radial or spiral trajectories, which could reduce aliasing artifacts further.
• Validation on Diverse Datasets: Expanding the validation on larger and more varied datasets would refine the model's generalization capabilities and confirm its robustness across different patient cohorts and MR machines.

In conclusion, the proposed deep cascade of CNNs presents a compelling approach to dynamic MR image reconstruction, offering improvements in both performance and speed. This framework exemplifies the transformative potential of deep learning in medical imaging, setting the stage for future advancements and integration in clinical practice.