Related Experiment Video
Updated: Dec 22, 2025
14:08
Automated Midline Shift and Intracranial Pressure Estimation based on Brain CT Images
Published on: April 13, 2013
43.3K
Deep neural network inspired by iterative shrinkage-thresholding algorithm with data consistency (NISTAD) for fast
Wenyuan Qiu1, Dongxiao Li2, Xinyu Jin1
1College of Information Science and Electronic Engineering, Zhejiang University, Hangzhou, China.
Magnetic Resonance Imaging
|May 1, 2020
Summary
A new deep neural network, NISTAD, accelerates MRI reconstruction from undersampled k-space data. This method achieves high-quality images comparable to state-of-the-art techniques with a simpler architecture.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Signal Processing
Background:
- Undersampled k-space data in Magnetic Resonance Imaging (MRI) necessitates advanced reconstruction algorithms.
- Conventional Compressed Sensing MRI (CS-MRI) methods are often time-consuming and require extensive hyper-parameter tuning.
- Existing deep learning methods for MRI reconstruction can be complex with numerous parameters.
Purpose of the Study:
- To develop a fast and high-quality MRI reconstruction algorithm from undersampled k-space data.
- To propose a novel deep neural network, NISTAD, inspired by the Iterative Shrinkage Thresholding Algorithm with Data consistency (ISTA-DC).
- To offer a simpler yet effective alternative to existing CS-MRI and deep learning reconstruction methods.
Main Methods:
- NISTAD employs a three-block architecture: encoding (modeling ISTA flow graph), decoding (sparse representation recovery), and data consistency (adaptive enforcement based on learned noise levels).
- The Iterative Shrinkage Thresholding Algorithm with Data consistency (ISTA-DC) is mapped into an end-to-end deep neural network.
- The method was validated on retrospectively undersampled diencephalon challenge data using various acceleration factors.
Main Results:
- NISTAD demonstrated significantly reduced reconstruction time compared to conventional model-based CS-MRI methods.
- The proposed network architecture is simpler and has fewer parameters than general deep learning-based reconstruction methods.
- NISTAD achieved comparable image quality (PSNR, SSIM, subjective assessment) to state-of-the-art methods like DAGAN and Cascade CNN.
Conclusions:
- NISTAD offers an efficient and effective deep learning approach for high-quality MRI reconstruction from undersampled data.
- The method balances reconstruction speed, image quality, and network simplicity.
- NISTAD presents a promising alternative for accelerating MRI acquisition and improving diagnostic capabilities.
