The repository contains a source code for Resolution-Enhancement-Generative-Adversarial-Inline-Network (REGAIN). REGAIN is a generative adversarial neural network for REGAINing image sharpness and Spatial resolution. The trained network generates a resolution-enhanced image in the Phase-encoding (ky) direction in MRI.
k-space filtering should be turned off before using REGAIN.
For training and testing: torch, numpy
For GUI-based testing: torch, numpy, scipy, PyQt5, pydicom, PIL, qimage2ndarray, opencv-python
The pretrained model can be found at Dataverse of Harvard Cardiac MR center (https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/4PQS82). Please refer file name "Pretrained_Model_REGAIN_epoch500.pth" (Link : https://dataverse.harvard.edu/api/access/datafile/6922233).
The testing result from 181 patients can be found at Dataverse of Harvard Cardiac MR center (https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/4PQS82). The results present reconstructed image data 1) Breath-hold GRAPPA-accelerated ECG-gated segmented Cine 2) Free-breathing CS-accelerated Real-time Cine
-Config.py: The configuratioFile Descriptionsn file for defining a cuda machine, dataroot folder, and hyperparameters during training
-Dataset.py: Specefies the training dataset loading and data augumentation
-Models.py: Implementation of REGAIN
-Train.py: The numpy magnitude images are loaded and the training is performed. (This is what you run for training)
-Test.py: The numpy magnitude images are loaded and the testing is performed. (This is what you run for testing)
-GUI-Testing.py: The DICOM files are loaded and visualize the result of REGAIN (This is what you run for off-line Reconstruction)