Related Experiment Video
Updated: Jan 4, 2026
13:19
Deep Neural Networks for Image-Based Dietary Assessment
Published on: March 13, 2021
9.9K
MR image reconstruction using deep learning: evaluation of network structure and loss functions.
Vahid Ghodrati1,2, Jiaxin Shao1, Mark Bydder1
1Department of Radiological Sciences, David Geffen School of Medicine, University of California, Los Angeles, CA, USA.
Quantitative Imaging in Medicine and Surgery
|November 1, 2019
Summary
Residual network (Resnet) and Unet convolutional neural networks (CNNs) offer similar cardiac MRI image quality. Resnet requires 10x fewer parameters, enabling efficient training for accelerated imaging.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Cardiovascular MRI
Background:
- Accelerated cardiac MRI is crucial for clinical practice.
- Convolutional Neural Networks (CNNs) show promise for image reconstruction.
Purpose of the Study:
- Evaluate CNN reconstruction approaches for accelerated cardiac MRI.
- Compare Unet and Resnet architectures and various loss functions.
Main Methods:
- Assessed Unet and Resnet CNNs using quantitative and qualitative radiologist evaluations.
- Compared L1, L2, Dssim, and perceptual loss functions.
- Utilized retrospectively and prospectively undersampled cardiac MRI data.
Main Results:
- Resnet and Unet achieved comparable image quality.
- Resnet used 100,000 parameters versus 1.3 million for Unet.
- Perceptual loss significantly outperformed L1, L2, and Dssim based on radiologist scores.
Conclusions:
- Resnet CNNs provide comparable image quality to Unet for cardiac MRI reconstruction with substantially fewer parameters.
- This efficiency has implications for reduced data requirements in network training.
- Perceptual loss function aligns better with radiologist assessments than other tested loss functions.
