GANCS

Compressed Sensing MRI based on Generative Adversarial Network

Introduction

First Authors:

Morteza Mardani, Enhao Gong

Arxiv Paper

Deep Generative Adversarial Networks for Compressed Sensing Automates MRI Arxiv Paper: https://arxiv.org/abs/1706.00051

References:

The basic code base is derived from super resolution github repo. https://github.com/david-gpu/srez

https://github.com/david-gpu/srez https://arxiv.org/abs/1609.04802 https://github.com/carpedm20/DCGAN-tensorflow http://wiseodd.github.io/techblog/2017/03/02/least-squares-gan/ http://wiseodd.github.io/techblog/2017/02/04/wasserstein-gan/

Method

Deep Generative Adversarial Networks for CS MRI

Magnetic resonance imaging (MRI) suffers from aliasing artifacts when it is highly undersampled for fast imaging. Conventional CS MRI reconstruction uses regularized iterative reconstruction based on pre-defined sparsity transform, which usually include time-consuming iterative optimization and may result in undesired artifacts such as oversmoothing. Here we propose a novel CS framework that permeates benefits from deep learning and generative adversarial networks (GAN) to modeling a manifold of MR images from historical patients. Extensive evaluations on a large MRI datasets of pediatric pateints show it results in superior perforamnce, retrieves image with improved quality and finer details relative to conventional CS and pixel-wise deep learning schemes.

Undersampling

1D undersampling

1D undersampling is generated using the variable density distribution. R_factor defines the desired reduction factor, which controls how many samples to randomly pick. R_alpha defines the decay of VD distribution with formula p=x^alpha. R_seed defines the random seed used, for negative values there is no fixed undersampling pattern.

1D/2D undersampling

sampling_pattern can be a path to .mat file for specific 1D/2D undersampling mask

Generator Model

Several models are explored including ResNet-ish models from super-resolution paper and encoder-decoder models.

Descriminator Model

Currently we are using 4*(Conv-BN-RELU-POOL)+2*(CONV-BN-RELU)+CONV+MEAN+softmax (for logloss)

descriminator related loss

We have been exploring different loss functions for GAN, including:

• log-loss
• LS loss (better than log-loss, use as default, easy to tune and optimize)
• Cycle-GAN/WGAN loss (todo)

Loss formulation

Loss is a mixed combination with: 1) Data consistency loss, 2) pixel-wise MSE/L1/L2 loss and 3) LS-GAN loss

FLAGS.gene_log_factor = 0 # log loss vs least-square loss

FLAGS.gene_dc_factor = 0.9 # data-consistency (kspace) loss vs generator loss

FLAGS.gene_mse_factor = 0.001 # simple MSE loss originally for forward-passing model vs GAN loss GAN loss = generator loss + discriminator loss

gene_fool_loss = FLAGS.gene_log_factor * gene_log_loss + (1-FLAGS.gene_log_factor) * gene_LS_loss

gene_non_mse_loss = FLAGS.gene_dc_factor * gene_dc_loss + (1-FLAGS.gene_dc_factor) * gene_fool_loss

gene_loss = FLAGS.gene_mse_factor * gene_mse_loss + (1- FLAGS.gene_mse_factor) * gene_non_mse_loss

Dataset

Dataset is parsed to pngs saved in specific folders

• Phantom dataset
• Knee dataset
• DCE dataset

Results

Multiple results are exported while traning

• loss changes
• test results (zero-fill VS Recon VS Ref) saved to png for each epoch
• some of the train results are exported to png, WARNING, there was a memory bug before when we try to export all train results
• layers (some layers are skipped) of generator and detectors are exported into json after each epoch.

Code structures

Files

• srez_main.py handles commendline arguments
• srez_input.py loads from input files, conducting undersampling or loading/extracting the undersampled patterns,
• srez_model.py defines Generator and Discriminator models
• srez_train.py trains model in batch and export checkpoints and results for progress summaries

Usage example:

Training example

(currently working on t2) python srez_main.py --dataset_input /home/enhaog/GANCS/srez/dataset_MRI/phantom --batch_size 8 --run train --summary_period 123 --sample_size 256 --train_time 10 --train_dir train_save_all --R_factor 4 --R_alpha 3 --R_seed 0

(currently working on t2 for DCE) python srez_main.py --run train --dataset_input /home/enhaog/GANCS/srez/dataset_MRI/abdominal_DCE --sample_size 200 --sample_size_y 100 --sampling_pattern /home/enhaog/GANCS/srez/dataset_MRI/sampling_pattern_DCE/mask_2dvardesnity_radiaview_4fold.mat --batch_size 4 --summary_period 125 --sample_test 32 --sample_train 10000 --train_time 200 --train_dir train_DCE_test