Related Experiment Video
Updated: Aug 5, 2025
05:32
Retrospective Cardiac Gating with A Prototype Small-Animal X-ray Computed Tomograph
Published on: February 21, 2025
353
A Deep Learning Framework for Cardiac MR Under-Sampled Image Reconstruction with a Hybrid Spatial and k-Space Loss
Walid Al-Haidri1, Igor Matveev1, Mugahed A Al-Antari2
1School of Physics and Engineering, ITMO University, Saint Petersburg 191002, Russia.
Diagnostics (Basel, Switzerland)
|March 29, 2023
Summary
This study introduces a deep learning framework for faster Magnetic Resonance Imaging (MRI) reconstruction, significantly improving image quality and reducing scan times. The Conditional Generative Adversarial Network (CGAN) model offers superior performance compared to traditional methods.
Area of Science:
- Medical Imaging
- Artificial Intelligence in Healthcare
- Deep Learning for Image Reconstruction
Background:
- Modern Magnetic Resonance Imaging (MRI) systems face challenges with long scanning times, leading to patient discomfort and motion artifacts.
- Accelerated or parallel MRI techniques aim to reduce scan duration, minimize patient stress, and lower medical costs.
- Accurate image reconstruction from under-sampled or aliased data is crucial for efficient MRI.
Purpose of the Study:
- To propose a novel deep learning framework for enhanced MR image reconstruction from under-sampled data.
- To improve the accuracy and quality of reconstructed MR images, reducing scan times and artifacts.
- To evaluate the proposed framework against existing reconstruction techniques.
Main Methods:
- A deep learning reconstruction framework based on Conditional Generative Adversarial Networks (CGANs) with an encoder-decoder U-Net generator.
- Implementation of a hybrid spatial and k-space loss function to optimize reconstruction in both domains.
- Comparative analysis against individual U-Net, CGAN models, L1-norm, and the traditional SENSE technique using SSIM and PSNR metrics on the OCMR dataset.
Main Results:
- The proposed CGAN framework achieved superior reconstruction quality, outperforming SENSE by 6.84% (PSNR with U-Net) and 9.57% (PSNR with CGAN).
- The hybrid loss function improved reconstruction performance by 6.84% (U-Net) and 9.57% (CGAN) compared to the simple L1-norm.
- Structural Similarity (SSIM) results were comparable to SENSE, indicating robust performance across different metrics.
Conclusions:
- The proposed deep learning framework, particularly using CGAN, provides superior MR image reconstruction performance compared to U-Net and SENSE.
- The hybrid loss function enhances reconstruction accuracy by considering both spatial and frequency domains.
- This framework shows significant potential for practical cardiac MR imaging, offering improved image quality and efficiency.
