Source-linked AI summary
Efficient and Accurate MRI Super-Resolution using a Generative Adversarial Network and 3D Multi-Level Densely Connected Network
Yuhua Chen, Feng Shi, Anthony G. Christodoulou, Zhengwei Zhou, Yibin Xie, Debiao Li
TL;DR
HR MRI offers detailed anatomy but entails longer scans, smaller coverage, and lower SNR, motivating SISR from a single LR input. The paper proposes a lightweight 3D mDCSRN with GAN-guided training; on 1,113 subjects, it recovers 4x downgraded images and runs 6x faster than other deep-learning methods.
Problem
3D MRI super-resolution must recover perceptually convincing HR detail while avoiding the memory and computational costs of 3D networks and the limitations of voxel-wise objectives.
Method
The paper combines a lightweight 3D multi-level densely connected super-resolution network with GAN-guided training using intensity and discriminator losses.
Results
6 times more quickly than current state-of-art deep learning approaches, mDCSRN-GAN recovered local textures and details more accurately in 4-fold resolution-reduced MRI.
Takeaways & Limitations
The technique would allow a 4-fold reduction in scan time while maintaining virtually identical image resolution and quality.
Abstract
from arXiv · showhide
High-resolution (HR) magnetic resonance images (MRI) provide detailed anatomical information important for clinical application and quantitative image analysis. However, HR MRI conventionally comes at the cost of longer scan time, smaller spatial coverage, and lower signal-to-noise ratio (SNR). Recent studies have shown that single image super-resolution (SISR), a technique to recover HR details from one single low-resolution (LR) input image, could provide high-quality image details with the help of advanced deep convolutional neural networks (CNN). However, deep neural networks consume memory heavily and run slowly, especially in 3D settings. In this paper, we propose a novel 3D neural network design, namely a multi-level densely connected super-resolution network (mDCSRN) with generative adversarial network (GAN)-guided training. The mDCSRN quickly trains and inferences and the GAN promotes realistic output hardly distinguishable from original HR images. Our results from experiments on a dataset with 1,113 subjects show that our new architecture beats other popular deep learning methods in recovering 4x resolution-downgraded im-ages and runs 6x faster.
1 Introduction
MRI super-resolution could recover detailed HR images from a single LR scan, avoiding the longer scan times and other costs of direct HR imaging. The problem remains underdetermined, while existing 3D and voxel-wise approaches face efficiency and perceptual-quality limitations.
- High-resolution MRI supports clinical diagnosis, decision making, and quantitative image analysis but requires longer scans, smaller coverage, and lower SNR.
- Single-image super-resolution aims to recover an HR image from one LR input without extra scan time, but infinitely many HR images can produce the same LR image.
- Previous medical-image SR methods often process 3D volumes slice by slice, while full 3D models increase memory use and computational expense.
- Using MSE alone optimizes voxel-wise error, although MSE and PSNR do not directly represent the visual quality of restored images.
- The proposed 3D mDCSRN is lightweight and fast, while GAN training further improves sharpness and realistic appearance.
2 Method
The method learns an inverse mapping from LR to HR MRI using a lightweight 3D multi-level densely connected network, optimized with intensity and GAN losses. Its design addresses 3D computational costs through dense connections, compression, and multi-level blocks, while evaluation uses simulated LR images derived from HR data.
- Background: SISR trains a model on HR references to map LR images to super-resolved outputs, then generates SR images from LR inputs during deployment.
- Background: The degradation model maps HR image X to LR image Y through f, while SISR learns an approximate inverse mapping g with reconstruction residual R.
- Background: CNN SISR jointly optimizes feature extraction, manifold learning, and image reconstruction, allowing reconstruction error to guide feature learning.
- Loss function: L1 or L2 voxel-wise losses can produce over-blurred outputs because they approximate averages of multiple possible HR images.
- Loss function: GAN training combines a generator that produces detailed images with a discriminator that distinguishes real from generated images.
- Efficient 3D network: 3D models learn volumetric MRI structure directly but require substantially more parameters and computation than 2D models.
- Efficient 3D network: mDCSRN adds an extra level of skip connections to dense connections, reducing parameter number and accelerating computation.
- Proposed network: The network splits one deep DenseBlock into several shallow blocks, with each block connected to preceding blocks and the reconstruction layer.
3 Experiments
Experiments used a 1,113-subject 3D brain MRI dataset and compared mDCSRN variants with interpolation and neural-network baselines. The largest mDCSRN variant achieved the best performance, while mDCSRN-GAN produced sharper, visually plausible reconstructions.
- 3.1 Dataset and Data Preparation: 1,113 subjects were divided into 780 training, 111 validation, 111 evaluation, and 111 test samples without cross-subset subject or patch duplication.The test set was reserved for final performance evaluation, while validation monitored training and evaluation supported hyper-parameter selection.
- 3.2 Experimental Setup: The experiments evaluated SSIM, PSNR, and NRMSE to measure reconstructed-image similarity.The study compared different DCSRN architectures and extended FSRCNN and SRResNet to 3D for baseline comparisons.
- 3.3 Results: The mDCSRN b2u4 variant slightly outperformed equally deep DCSRN b1u8 while reducing parameters and running time by more than 30%.Among the variants, the largest b4u4 network achieved the best performance without too much sacrifice in speed.
- 3.3 Results: The largest mDCSRN b4u4 network achieved the best variant performance with an average whole-subject 3D MRI runtime of around 20 seconds.This result is reported for the deepest network in the architecture comparison.
- 3.3 Results: mDCSRN was slightly better than SRResNet and ran more than 6x faster, while outperforming FSRCNN and bicubic interpolation in image quality.mDCSRN-GAN further sharpened reconstructions; vessels and gaps between vessels and gray matter were clearer, and its result was almost indistinguishable from ground truth.
4 Conclusion
The paper presents 3D mDCSRN-GAN for MRI single-image super-resolution. It reports more accurate texture and detail recovery at six times the speed of current state-of-the-art deep-learning approaches, with a potential four-fold scan-time reduction.
- 4 Conclusion: The proposed method is a 3D mDCSRN-GAN approach for MRI single-image super-resolution.The method is presented as a new technique for recovering MRI image details from lower-resolution inputs.
- 4 Conclusion: 6 times faster than current state-of-the-art deep-learning approaches, mDCSRN-GAN recovered local image textures and details more accurately.The conclusion presents speed and detail recovery together as the principal reported outcome.
- 4 Conclusion: A 4-fold scan-time reduction was reported while maintaining virtually identical image resolution and quality.This is the paper's stated practical implication of the proposed technique.