This is my implementation of the AUTOMAP algorithm described in the following paper: B. Zhu, J. Z. Liu, B. R. Rosen, and M. S. Rosen. Image reconstruction by domain transform manifold learning. arXiv preprint arXiv:1704.08841, 2017. https://arxiv.org/abs/1704.08841
NB1: I run the code at AWS cluster, using the following AMI: Deep Learning AMI (Ubuntu), and the following instance: p3.2xlarge. In addition, I use CPU’s memory to initialize the second fully-connected layer for 128x128 images (otherwise, there is memory error) NB2: I use the following Python package to download images from ImageNet: imagenetscraper 1.0 (https://goo.gl/QK6f8p)
I encourage you to contact me if you have any questions, comments, or suggestions: email@example.com.
The code uses data in image space and corresponding frequency space to teach a CNN model to do a reconstruction of an MRI image. The architecture consists of fully-connected (FC) and convolutional (Conv) layers and is the following: FC1 -> tahn activation -> FC2 -> tanh activation -> Conv1 -> ReLU activation -> Conv2 -> ReLU activation -> de-Conv
This includes function load_images_from_folder, which creates training set for a model. It loads images into array Y and performs a Fourier transform and saves both real and imaginary parts of it into array X. Optional normalizing of data and rotation of input images are available.
This includes function load_images_from_folder, which creates training set for a model. It loads images into array Y and "moves" it by 8 pixels the performs a Fourier transform and combined the frequency space of both Y images (before and after it was moved) - as if the patient moved by 8 pixels in one direction in the middle of the acquisition. Then the function saves both real and imaginary parts of motion-corrupted frequency space into array X. Optional normalizing of data and rotation of input images are available.
This includes the CNN model using TensorFlow.
Uses forward propagation to reconstruct image from frequency space using the trained model, which was saved in myAutomap.py
(Very) preliminary results:
Y_dev - original images; X_iFFT - images reconstructed from frequency space corrupted by motion - ghosting artifacts are clearly seen; Y_recon - images reconstructed using trained model - ghosting is gone!, however, images look very blurry - the cost was still quite high - needs optimizing.
Hyperparameters: learning rate - 0.00002, 7500 images (30000 after augmentation), 80x80 resolution, 200 epochs.