Application of the Deep Reconstruction Framework with Self-Calibration Mechanisms (DEISM) in Accelerated Chemical Exchange Saturation Transfer Imaging

Academic Background

Chemical Exchange Saturation Transfer (CEST) imaging is a highly sensitive molecular magnetic resonance imaging technique capable of detecting biomolecules associated with various diseases, such as cancer, epilepsy, and stroke. However, a major drawback of CEST imaging is its prolonged scan time, which results from multiple data acquisitions over varying saturation frequency offsets. This long scanning duration limits the widespread clinical adoption of CEST imaging. To address this issue, researchers have been working on developing techniques to accelerate CEST imaging, primarily by exploiting redundancy in the obtained data to reconstruct images from undersampled k-space data.

Although existing parallel imaging and compressed sensing (CS) technologies have achieved some success in accelerating CEST imaging, these methods still have limitations. For example, the acceleration rates achievable with parallel imaging are constrained by noise amplification and inaccurate coil sensitivity estimation, while CS techniques are hindered by cumbersome parameter tuning and long reconstruction times that may be impractical for clinical settings. In recent years, deep learning algorithms have shown great potential in medical image reconstruction, offering new possibilities for faster CEST imaging by implicitly learning from mass historical data to leverage more adaptable priors.

Source of the Paper

This paper was co-authored by Jianping Xu, Tao Zu, Yi-Cheng Hsu, Xiaoli Wang, and Kannie W. Y. Chan, among others, with primary authors from several renowned research institutions. The paper has been published in the journal IEEE Transactions on Biomedical Engineering and was first presented as an abstract at the 2023 annual meeting of the International Society for Magnetic Resonance in Medicine (ISMRM).

Research Workflow and Details

Research Objectives

This study proposes a novel deep learning framework called DEISM (Deep Reconstruction Framework with Self-Calibration Mechanisms), aimed at accelerating CEST imaging through self-calibration mechanisms. The DEISM framework combines a model-based image reconstruction network with a data-driven artifact suppression network, leveraging spatial-frequential redundancy in the artifact field to significantly improve the quality of CEST image reconstruction.

Research Workflow

1. DEISM Framework Design

The DEISM framework consists of two main modules: - CEST-VN: A model-based image reconstruction network responsible for generating initial reconstructed images from undersampled multi-coil k-space data. - AS-Net: A data-driven artifact suppression network that estimates and corrects residual artifacts in the initial reconstructed images through a self-calibration mechanism.

2. Development of Artifact Suppression Network (AS-Net)

AS-Net adopts a novel multi-scale feature fusion convolutional neural network (Muff-CNN) for artifact estimation (AE) and artifact correction (AC). AS-Net extracts artifact information from fully-sampled calibration frames and uses this information to correct artifacts in undersampled frames.

3. Training and Optimization

The DEISM framework is trained end-to-end using simulated multi-coil CEST data for pre-training and fine-tuning. The training process is divided into four steps: - Step 1: Pre-train CEST-VN using a CEST-specific loss function for 40 epochs. - Step 2: Pre-train AS-Net using 3-fold undersampled image data. - Step 3: Fine-tune AS-Net with the weights of CEST-VN frozen, tailored to different acceleration factors. - Step 4: End-to-end training of the entire DEISM framework, optimized using a CEST-specific loss function.

4. Experiments and Evaluation

The study conducted retrospective and prospective experiments on data from five healthy volunteers and five brain tumor patients, evaluating the performance of the DEISM framework at various acceleration factors. Experimental results demonstrated that DEISM outperformed conventional parallel imaging and CS algorithms in reconstructing CEST source images, molecular maps, and CEST spectra.

Main Results

1. Artifact Suppression Effectiveness

On 3-fold accelerated CEST images, AS-Net effectively corrected artifacts, preserved more details, and exhibited lower errors and higher peak signal-to-noise ratio (PSNR) in quantitative evaluations. Compared to traditional linear artifact correction methods, AS-Net showed significant advantages in both image quality and artifact correction performance.

2. Performance of the DEISM Framework

On 8-fold accelerated CEST images, the source images and Amide Proton Transfer-weighted (APTw) maps reconstructed by the DEISM framework were highly consistent with fully-sampled reference images. The DEISM framework excelled in preserving structural details and correcting residual artifacts, especially under high acceleration factors, where its performance significantly surpassed other methods.

3. Prospective Experiments

In prospective experiments, the DEISM framework demonstrated high reconstruction quality on 4-fold and 7-fold accelerated CEST images, indicating its feasibility for clinical applications.

Conclusions and Implications

Conclusions

By combining model-based image reconstruction with data-driven artifact suppression, the DEISM framework significantly improved the quality of CEST image reconstruction, particularly under high acceleration factors. The framework leverages spatial-frequential redundancy in the artifact field to effectively correct residual artifacts through self-calibration mechanisms, providing a new solution for accelerating CEST imaging.

Scientific and Application Value

The introduction of the DEISM framework not only advances the development of CEST imaging technology scientifically but also provides clinicians with a more efficient imaging tool. By reducing scan time, DEISM is expected to promote the widespread application of CEST imaging in disease diagnosis and treatment monitoring.

Research Highlights

• Novel Self-Calibration Mechanism: The DEISM framework is the first to incorporate learned artifact priors into the CEST image reconstruction process, significantly improving image quality.
• Multi-Scale Feature Fusion Network: The multi-scale feature fusion mechanism adopted by AS-Net effectively enhances the accuracy of artifact estimation and correction.
• End-to-End Training: The DEISM framework achieves global optimization of image reconstruction and artifact suppression through end-to-end training, further enhancing reconstruction performance.

Other Valuable Information

This study also explored the impact of different calibration frame selections on artifact suppression effectiveness, providing important references for future optimization of CEST imaging techniques. Additionally, the open-source code for the DEISM framework has been made publicly available for further validation and improvement by other researchers.

---

Through this study, the DEISM framework offers an efficient and reliable method for accelerating CEST imaging, holding significant scientific and clinical value. In the future, this framework is expected to play an important role in more disease diagnoses and treatment monitoring.