Source-linked AI summary
Brain MRI Super Resolution Using 3D Deep Densely Connected Neural Networks
Yuhua Chen, Yibin Xie, Zhengwei Zhou, Feng Shi, Anthony G. Christodoulou, Debiao Li
TL;DR
High-resolution MRI improves anatomical detail but entails longer scans, reduced coverage, and lower SNR, motivating reconstruction from a single LR scan. The paper introduces DCSRN, a 3D densely connected neural network for brain-MRI SISR. On 1,113 subjects, DCSRN produced better-quality super-resolved images more efficiently than interpolation and previous deep-learning methods.
Problem
High-resolution MRI provides important structural detail but typically requires longer scans, less spatial coverage, and lower SNR.
Method
The paper introduces DCSRN, a 3D densely connected neural network that reconstructs HR brain MRI from a single LR input.
Results
DCSRN produced significantly better-quality super-resolved images more efficiently than popular interpolation and previous deep-learning methods.
Takeaways & Limitations
DCSRN restored structural brain-MRI images more effectively than the compared interpolation and deep-learning approaches.
Abstract
from arXiv · showhide
Magnetic resonance image (MRI) in high spatial resolution provides detailed anatomical information and is often necessary for accurate quantitative analysis. However, high spatial resolution typically comes at the expense of longer scan time, less spatial coverage, and lower signal to noise ratio (SNR). Single Image Super-Resolution (SISR), a technique aimed to restore high-resolution (HR) details from one single low-resolution (LR) input image, has been improved dramatically by recent breakthroughs in deep learning. In this paper, we introduce a new neural network architecture, 3D Densely Connected Super-Resolution Networks (DCSRN) to restore HR features of structural brain MR images. Through experiments on a dataset with 1,113 subjects, we demonstrate that our network outperforms bicubic interpolation as well as other deep learning methods in restoring 4x resolution-reduced images.
1. INTRODUCTION
High-resolution MRI supports detailed analysis but requires longer scans, less coverage, and lower SNR. SISR offers a single-scan route to reconstruct HR images, while interpolation cannot recover lost high-frequency details and multi-scan methods are motion-sensitive.
- High-resolution MRI provides structural detail needed for accurate image analysis and quantitative measurement.
- Higher MRI spatial resolution typically increases scan time and reduces spatial coverage and signal-to-noise ratio.
- SISR reconstructs a high-resolution image from one low-resolution scan, making it an attractive alternative for MRI super-resolution.
- Interpolation methods fail to recover high-frequency information such as fine object edges.
- Multi-scan reconstruction is vulnerable to interscan motion and is neither time- nor cost-optimal in practice.
- Learning-based methods learn image priors directly from examples rather than imposing assumptions about the data distribution.
2. SUPER-RESOLUTION NEURAL NETWORKS 2.1. Background
The paper formulates super-resolution as learning an inverse mapping from LR to HR images, then proposes a densely connected 3D network that reuses features and exploits volumetric structure. Its evaluation uses MRI-like k-space degradation and compares multiple interpolation and neural-network baselines.
- 2.1. Background: SISR estimates HR image X from LR image Y by learning an approximate inverse of the degradation mapping.The learning pipeline extracts features, maps them into a feature space, and reconstructs X while minimizing training reconstruction error.
- 2.2. Proposed 3D Densely Connected Super-Resolution Networks (DCSRN): DCSRN uses densely connected layers, skip connections, weight sharing, and feature reuse to target faster training, a lightweight model, and less overfitting.
- 2.3. 3D model vs 2D model: A 3D model can learn volumetric structural context that slice-wise 2D processing may confuse with noise, such as a vessel edge extending into a neighboring slice.
- 3.1. Evaluation Setting: LR images were generated by transforming HR images to k-space, zeroing outer regions along two phase-encoding axes, and applying the inverse FFT.The resulting LR images retain the HR image size and avoid checkerboard artifacts.
- 3.2. Dataset: The dataset contained 1,113 brain MRI subjects, split into 780 training, 111 validation, 111 evaluation, and 111 test samples.The structural T1w images had 0.7 mm isotropic spatial resolution and a 320x320x256 matrix.
4. RESULTS
DCSRN achieved the strongest overall image-quality performance among the evaluated methods and was also the fastest in testing.
- DCSRN achieved the best performance across SSIM, PSNR, and NRMSE, outperforming interpolation and other deep learning methods (p<0.01).The comparison included nearest neighbor, bicubic, 2D-FSRCNN, and 3D-FSRCNN.
- DCSRN trained 4x faster than 3D FSRCNN.
- DCSRN processed a 64x64x64 patient image patch in 23.31s, compared with 35.52s for 2D FSRCNN and 63.95s for 3D FSRCNN.
5. CONCLUSION
The paper presents DCSRN for single-image super-resolution of 3D brain MRI and reports better-quality images with greater efficiency than the compared methods.
- DCSRN is a convolutional neural network for single-image super-resolution of 3D brain MRI.
- DCSRN produced significantly better-quality super-resolution images than popular interpolation and previous deep learning methods.
- DCSRN performed the reconstruction more efficiently than the compared methods.