Source-linked AI summary
A Transfer-Learning Approach for Accelerated MRI using Deep Neural Networks
Salman Ul Hassan Dar, Muzaffer Özbey, Ahmet Burak Çatlı, Tolga Çukur
TL;DR
Large domain-specific MRI datasets for training reconstruction networks are scarce. This study uses transfer learning with fine-tuning on tens of target-domain MR images, achieving nearly identical performance to direct training on thousands of images.
Problem
Large datasets acquired under a common MRI protocol are rare, and cross-contrast network generalization remains uncertain.
Method
Networks trained on large natural-image or brain-MRI datasets were fine-tuned end-to-end using only tens of target-domain brain MR images.
Results
Fine-tuned domain-transferred networks achieved nearly identical performance to networks trained directly in the testing domain using thousands of images, outperforming compressed-sensing methods.
Takeaways & Limitations
Transfer learning may support accelerated MRI reconstruction when large domain-specific imaging datasets are unavailable.
Takeaways & Limitations
The compressed-sensing comparator uses computationally complex nonlinear optimization that can make reconstruction times prohibitive as data size increases.
Abstract
from arXiv · showhide
Purpose: Neural networks have received recent interest for reconstruction of undersampled MR acquisitions. Ideally network performance should be optimized by drawing the training and testing data from the same domain. In practice, however, large datasets comprising hundreds of subjects scanned under a common protocol are rare. The goal of this study is to introduce a transfer-learning approach to address the problem of data scarcity in training deep networks for accelerated MRI. Methods: Neural networks were trained on thousands of samples from public datasets of either natural images or brain MR images. The networks were then fine-tuned using only few tens of brain MR images in a distinct testing domain. Domain-transferred networks were compared to networks trained directly in the testing domain. Network performance was evaluated for varying acceleration factors (2-10), number of training samples (0.5-4k) and number of fine-tuning samples (0-100). Results: The proposed approach achieves successful domain transfer between MR images acquired with different contrasts (T1- and T2-weighted images), and between natural and MR images (ImageNet and T1- or T2-weighted images). Networks obtained via transfer-learning using only tens of images in the testing domain achieve nearly identical performance to networks trained directly in the testing domain using thousands of images. Conclusion: The proposed approach might facilitate the use of neural networks for MRI reconstruction without the need for collection of extensive imaging datasets.
Abstract
The study proposes transfer learning to address scarce training data for accelerated MRI reconstruction. Networks pretrained on large natural-image or brain-MRI datasets and fine-tuned with few target-domain images approached the performance of networks trained directly on thousands of target-domain images.
- Motivation and approach: Transfer learning addresses accelerated-MRI data scarcity by pretraining networks on large source datasets and fine-tuning them with limited images from the target domain.The study targets the extensive training data typically required for neural-network MRI reconstruction when large common-protocol datasets are unavailable.
- Evaluation: Evaluations covered acceleration factors R=2-10 and domain transfer between T1- and T2-weighted brain images, and between ImageNet and T1- or T2-weighted images.Domain-transferred networks were compared with networks trained directly in the testing domain and with conventional reconstruction methods.
- T2-weighted reconstruction: After fine-tuning, average T2-reconstruction differences between ImageNet- and T2-trained networks decreased from (1.23dB, 3.40%) to (0.19dB, 0.40%).Differences between T1-trained and T2-trained networks decreased from (1.14dB, 2.80%) to (0.14dB, - 0.20%).
- T2-weighted reconstruction: Across R, domain-transferred networks outperformed CS by 5.21dB PSNR and 12.5% SSIM on T2-weighted reconstructions.Following fine-tuning, domain-transferred networks also produced visually similar reconstructions to networks trained directly on T2-weighted images after as few as 20 images.
ImageNet-trained T1-trained SPIRiT … Supplementary Materials
The study evaluates transfer learning for accelerated MRI reconstruction using ImageNet-, T1-, and T2-trained networks, including fine-tuning across contrasts and from natural images to MRI. Across representative reconstructions and PSNR analyses, domain-transferred networks approach directly trained networks while fine-tuning requirements vary with acceleration and training conditions.
- ImageNet-trained T1-trained SPIRiT: The proposed architecture sequentially trains calibration-consistency, convolutional, and data-consistency blocks for undersampled multi-coil reconstruction.Subnetworks are trained on synthetic multi-coil natural images from ImageNet before transfer to MRI.
- ImageNet-trained T1-trained SPIRiT: At R=4, representative T1-weighted reconstructions compare zero-filled Fourier, ImageNet-trained, T2-trained, and T1-trained networks before and after fine-tuning.The comparisons include reconstructed images, error maps, and a fully sampled reference.
- ImageNet-trained T1-trained SPIRiT: With 2000 training images and 20 contrast-matched fine-tuning images, domain-transferred networks maintain nearly identical performance to networks trained directly in the testing domain.The comparison includes ImageNet-trained, T1-trained, and T2-trained networks for T1-weighted reconstruction at R=4.
- PSNR SSIM PSNR SSIM PSNR SSIM: Average T1-weighted PSNR is evaluated across acceleration factors R=2, R=4, R=6, R=8, and R=10 while varying fine-tuning samples and comparing ImageNet-, T1-, and T2-trained networks.ImageNet-trained networks use 500, 1000, 2000, or 4000 training images; the T1-trained network uses 4k training images and 100 fine-tuning images.
- PSNR SSIM PSNR SSIM PSNR SSIM: ImageNet-trained networks converge when the percentage change in PSNR from increasing Ntune falls below 0.05% of the T1-trained network’s average PSNR.Convergence is measured for ImageNet networks trained on 500, 1000, 2000, and 4000 images.
- PSNR SSIM PSNR SSIM PSNR SSIM: For multi-coil T1-weighted acquisitions, PSNR is compared across R=2, R=4, R=6, R=8, and R=10 for a T1-trained network and a 2000-image ImageNet-trained network.The T1-trained network is trained and fine-tuned on 360 images, while the ImageNet-trained network is evaluated as fine-tuning samples vary.
- PSNR SSIM PSNR SSIM PSNR SSIM: At higher acceleration factors, ImageNet-trained networks require more fine-tuning samples for PSNR convergence.Convergence is defined using the 0.05% PSNR-change criterion relative to the T1-trained network.
- Supplementary Materials: At R=10, an ImageNet-trained network fine-tuned on 20 T1-weighted images maintains similar performance to a T1-trained network and is compared with SPIRiT.The representative multi-coil reconstruction also includes zero-filled Fourier reconstruction, error maps, and a fully sampled reference.