Related Experiment Video
Updated: Jul 1, 2025
04:48
Swin-PSAxialNet: An Efficient Multi-Organ Segmentation Technique
Published on: July 5, 2024
396
IWNeXt: an image-wavelet domain ConvNeXt-based network for self-supervised multi-contrast MRI reconstruction.
Yanghui Yan1, Tiejun Yang2,3,4, Chunxia Jiao1
1School of Information Science and Engineering, Henan University of Technology, Zhengzhou 450001, People's Republic of China.
Physics in Medicine and Biology
|March 13, 2024
Summary
This study introduces IWNeXt, a self-supervised deep learning method for faster multi-contrast MRI (MC MRI) reconstruction. It significantly improves image quality by leveraging cross-domain information, reducing the need for fully sampled data.
Area of Science:
- Medical Imaging
- Deep Learning
- Magnetic Resonance Imaging
Background:
- Multi-contrast MRI (MC MRI) offers comprehensive anatomical details but suffers from long acquisition times compared to single-contrast MRI.
- Current deep learning methods for MC MRI reconstruction often require fully sampled target contrast data and may overlook valuable prior information from reference contrasts in sparse domains.
- Conventional CNNs struggle with capturing non-local dependencies due to limited receptive fields, impacting reconstruction quality.
Purpose of the Study:
- To develop a self-supervised, cross-domain deep learning framework for accelerated MC MRI reconstruction.
- To enhance reconstruction accuracy and reduce the dependency on fully sampled target contrast images.
- To improve the capture of non-local information and high-frequency details in reconstructed MC MRI scans.
Main Methods:
- Proposed an Image-Wavelet domain ConvNeXt-based network (IWNeXt) for self-supervised MC MRI reconstruction.
- Employed separate ConvNeXt-based networks (INeXt and WNeXt) for image and wavelet domain reconstruction, respectively.
- Integrated reference contrast wavelet sub-bands and a novel attention ConvNeXt block for enhanced feature extraction and non-local information capture.
- Implemented a cross-domain consistency loss, including frequency, image, and wavelet domain losses, for self-supervised learning.
Main Results:
- Experimental validation on HCP and M4Raw datasets demonstrated superior performance.
- The proposed IWNeXt model achieved a 1.651 dB improvement in peak signal-to-noise ratio compared to DuDoRNet.
- The method effectively reconstructs undersampled MC MRI data with enhanced detail and accuracy.
Conclusions:
- IWNeXt presents a promising cross-domain approach for MC MRI reconstruction.
- The method enhances reconstruction accuracy and reduces the need for fully sampled target contrast data.
- This technique has the potential to significantly accelerate MC MRI acquisition while maintaining high image quality.
