ELEN E6876 Research Project - Columbia University Spring 2022
- Alex Thornton ([email protected])
The purpose of this research is to create a model which can efficiently store slices of MRI scans. Efficiency is ex- plored across two optimization domains; minimization of genera- tional loss through recursive compression, and compressed space minimization. Some techniques and frameworks explored include k-wavelet approximation, and various auto-encoder frameworks for neural networks. Training and test data was provided by FastMRI, a joint venture between NYU and Facebook (Meta).
Index Terms—Wavelet, Medical Imaging, Compression, Sparse Auto Encoder, Convolutional Neural Network
- JPEG compression is optimized for general images
- But what about images that share common features?
- Various deep learning and signal processing techniques are explored for new MRI compression schemes
- Approved use of FastMRI dataset from Facebook (Meta) and NYU
- Only using subset of the multicoil_test dataset ~34.2GB zipped
- 1 brain for test (16 slices), 26 brains for evaluation (14 - 16 slices each)
- Open source software suite for slicing and manipulating data
- Recursive JPEG compressions quickly erodes quality (photocopier effect)
- After only a few iterations, JPEG compression succumbs to noise corruption
- Wavelet basis is very resilient to recursive operations
- Subject k-wavelet coefficients to float32 quantization
- Results in quality loss in only first compression!
- Compression size of 14.4 KB
- Can compress further with sparse auto-encoder via convolutional neural network framework
- Captures image features well, but very lossy
- Read the full paper
- Check out the presentation
- Clone this repository
git clone www.github.com/athornton1618/MRI_Compression.git
cd MRI_Compression
pip install -r requirements.txt
- Obtain permission for FastMRI dataset
- Download multicoil_test dataset with following command:
curl -C - "https://fastmri-dataset.s3.amazonaws.com/brain_multicoil_test.tar.gz?AWSAccessKeyId=YOUR TOKEN" --output brain_multicoil_test.tar.gz
mkdir MRI_Compression/data
- Unzip multicoil_test
- Move files to MRI_Compression/data/multicoil_test
- You can run any of the scripts locally on your machine from this point, except sparse_autoencoder_cnn_colab.ipynb
- Run this python command to build the dataset for Colab:
cd utils
python make_dataset.py
- Open sparse_autoencoder_cnn_colab.ipynb in Google Colab
- Under the 'Files' tab in Colab, make directory /data, and upload the following from MRI_Compression/data:
- train_set_image.json
- train_set_wavelet.json
- test_set_image.json
- test_set_wavelet.json
- Run sparse_autoencoder_cnn_colab.ipynb
This project marks the completion of my M.Sc. at Columbia University.
I'd like to thank my Mom, Dad, sister Carolyn, brother Matt, and Murphy 🐕!
Roar Lion Roar 👑!
Copyright (c) 2022 Alex Thornton
Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.