An end-to-end deep learning solution to perform motion correction (MC) and super-resolution (SR) concurrently in CMR SAX slices
Author: Zhennong Chen, PhD
This is the GitHub repo for the paper:
Motion Correction and Super-Resolution for Multi-slice Cardiac Magnetic Resonance Imaging via an End-to-End Deep Learning Approach.
Authors: Zhennong Chen, Hui Ren, Quanzheng Li, Xiang Li
Citation: TBD
Accurate reconstruction of a high-resolution 3D volume of the heart is critical for comprehensive cardiac assessments. However, cardiac magnetic resonance (CMR) data is usually acquired as a stack of 2D short-axis (SAX) slices, which suffers from the inter-slice misalignment due to cardiac motion and data sparsity from large gaps between SAX slices. Therefore, we aim to propose an end-to-end deep learning (DL) model to address these two challenges simultaneously, employing specific model components for each challenge.
Overall, the objective of this study is to reconstruct a high-resolution 3D volume of the heart from acquired CMR SAX slices.
The deep learning model can be found in end2end/model.py, which is the core innovation of our study.
To be more specific, this model generates three outpus: combined_results, final_LR_img, final_HR_img, which correspond to the slice-wise displacement vector, motion-corrected low-resolution volume and final high-resolution volume respectively. (refer to section 2.2 in the paper)
The users can easily create their own data generator as well as model training/testing scripts so we decide not to provide.
The script is based on tensorflow. You can build your own docker from the provided dockerfile docker/dockerfile.
We utilized a public dataset: Cardiac Super-resolution Label Maps. This dataset contains high-spatial resolution 3D balanced steady-state free precession cine CMR sequences. We simulated the inter-slice motion misalignment and z-axis downsampling and thus prepared paris of high-resolution and low-resolution data for supervised training. More details can be found in section 2.5.1 in the paper.
Please contact zchen36@mgh.harvard.edu or chenzhennong@gmail.com for any further questions.