Related Experiment Video
Updated: Nov 4, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.1K
Projection-Based cascaded U-Net model for MR image reconstruction.
Amir Aghabiglou1, Ender M Eksioglu2
1Graduate School of Science, Engineering and Technology, Istanbul Technical University, Istanbul, Turkey.
Computer Methods and Programs in Biomedicine
|May 30, 2021
Summary
This study introduces a novel cascaded U-Net framework for faster and more accurate Magnetic Resonance Imaging (MRI) reconstruction from undersampled k-space data, significantly improving image quality over existing methods.
Area of Science:
- Medical Imaging
- Deep Learning
- Image Reconstruction
Background:
- U-Net is an effective deep learning model for imaging inverse problems.
- U-Net was originally designed for biomedical image segmentation.
- MRI reconstruction benefits from deep networks for speed and artifact reduction.
Purpose of the Study:
- Develop a novel and efficient cascaded U-Net framework for MRI reconstruction.
- Improve reconstruction performance from undersampled k-space data.
- Compare the new framework against existing methodologies.
Main Methods:
- Proposed a novel cascaded framework using U-Net as a sub-block.
- Integrated U-Net cascade with a projection-based updated data consistency layer.
- Implemented the framework in PyTorch and trained/tested on the fastMRI dataset.
Main Results:
- The cascaded U-Net structure achieved 1.28 dB higher PSNR than the baseline U-Net.
- The new method showed a 3.32 dB improvement over standard CNNs.
- Quantitative and qualitative results were improved compared to conventional methods.
Conclusions:
- The proposed cascaded U-Net framework enhances MRI reconstruction performance.
- It outperforms CNN, cascaded CNN, and singular U-Net baseline methods.
- The projection-based data consistency layer further improves reconstruction quality.
