Paper: UDLGD: Unsupervised Deep Learning in Gradient Domain for Multi-contrast MRI Reconstruction

Authors: Tao Deng, Yu Guan, Shanshan Wang, Dong Liang, Qiegen liu

Date : 1/2021
Version : 1.0
The code and the algorithm are for non-comercial use only.
Copyright 2020, Department of Electronic Information Engineering, Nanchang University.

Multi-contrast magnetic resonance imaging (MRI) is a powerful imaging tool in clinical practice due to its ability to provide abundant contrast information. The recent end-to-end deep learning often required a large number of paired multi-contrast MR images, and unsupervised learning with flexible scheme is promising to alleviate the deficiency. Moreover, pursuing a universal learning that is quite urgent.This paper proposes an efficient unsupervised deep learning algorithm in gradient domain (UDLGD) for reconstructing multi-contrast images of the same anatomical cross section from partially sampled K-space data. The present UDLGD consists of two consequent stages. At the prior learning stage, score-based generative model is utilized to train gradient domain prior information from single-contrast image dataset. After the prior is learned, the data-consistency, gradient image and group sparsity are alternatively updated at the iterative reconstruction stage. Experimental results in synthetic and in-vivo MRI data demonstrated that the proposed reconstruction method can achieve lower reconstruction errors and better preserve image structures than competing methods.

% UDLGD 
python3.5 separate_SIAT.py --model ncsn --runner siat_multicontrast_compare_TSE_sag_random_R4 --config anneal_lr005_gradient4.yml --doc SIAT1_1dataaug4ch_lr005gradient4 --test --image_folder result_MultiContrast_Sag_4_random_R4
% UDLGD_GS
python3.5 separate_SIAT.py --model ncsn --runner siat_multicontrast_compare_TSE_sag_random_R4 --config anneal_lr005_gradient4.yml --doc SIAT1_1dataaug4ch_lr005gradient4 --test --image_folder result_MultiContrast_Sag_4_random_R4_GS --GS

Multi-contrast MR images and horizontal and vertical gradients structure information of the corresponding multi-contrast MR images. Here the value range of T1 and T2 images are different, while the corresponding gradient images are near to the same.

Unsupervised deep learning in gradient domain (UDLGD) for multi-contrast MRI reconstruction framework. UDLGD iterates between two alternative stages: Top: Training stage to learn the gradient distribution via denoising score matching. Bottom: Reconstruction to progressively remove aliasing and recover fine details via Langevin dynamics and group sparsity. and stand for extracting the real and imaging part, respectively. The gradient encoding means reconstructing the images from horizontal and vertical gradient estimates, and the gradient decoding means deriving the gradients from spatial images. It is worth noting that UDLGD is learned from single-modal dataset, while used for multi-contrast image reconstruction.

Reconstruction Results by Various Methods at 2D Random Undersampling, acceleration factor = 4.

Visual comparison of TSE_sag (256 × 256) reconstructions using same sampling schemes with acceleration . a: TSE scans at Nyquist rate sampling. b: Random sampling pattern. cd: BCS reconstruction and its absolute error. ef: GSMRI reconstruction and its absolute error. gh: FCSA-MT reconstruction and its absolute error. ij: UDLGD reconstruction and its absolute error. kl: UDLGD-GS reconstruction and its absolute error.

We provide pretrained checkpoints. You can download pretrained models from Baidu Drive. key number is "011k"

In file './Sag_Multicontrast', PD, T1 and T2-weighted sagittal brain datasets with size of 256x256 were acquired on a 3T scanner(SIEMENS MAGNETOM Trio), and the datasets were provided by Chinese Academy of Sciences.

• Multi-Contrast MR Reconstruction with Enhanced Denoising Autoencoder Prior Learning [Paper] [Code] [Slide]

• Multi-Channel and Multi-Model-Based Autoencoding Prior for Grayscale Image Restoration
  [Paper] [Code] [Slide]

• High-dimensional Embedding Network Derived Prior for Compressive Sensing MRI Reconstruction
  [Paper] [Code]

• Denoising Auto-encoding Priors in Undecimated Wavelet Domain for MR Image Reconstruction
  [Paper] [Code]

• REDAEP: Robust and Enhanced Denoising Autoencoding Prior for Sparse-View CT Reconstruction
  [Paper] [Code]

• Iterative Reconstruction for Low-Dose CT using Deep Gradient Priors of Generative Model
  [Paper] [Code]

• Progressive Colorization via Interative Generative Models
  [Paper] [Code]