LoDoPaB-CT Grand Challenge Code
The task of this challenge is to reconstruct CT images of the human lung from (simulated) low photon count measurements. For evaluation, the PSNR and SSIM values are computed w.r.t. the images that were used as ground truth.
python3 -m venv lowdosect
source lowdosect/bin/activate
- Install pytorch first
pip install torch==1.6.0+cu101 torchvision==0.7.0+cu101 -f https://download.pytorch.org/whl/torch_stable.html
- Use requirements.txt for other requirements
pip install -r src/requirements.txt
- Install Astra toolbox (recommended for RayTransform)
If astra toolbox is not installed please change line 13 in RayTransform.py from:
impl = 'astra_cpu',
to:
impl = 'skimage',
Everything under the /src/libs folder is external libraries I wasn't able to install through pip. This includes LoDoPaB-CT challenge utilities (https://github.com/jleuschn/lodopab_challenge) and astra toolbox (https://github.com/astra-toolbox/astra-toolbox) I used the code from the master branch found on their respective github pages.
The astra toolbox was a little tricky to build. Using conda would have been better, but I was using venv. To install you will need to do the follow the guide found on their guide github page for building from source.
To run the training, inference and submission code you will need to have the data save in src/data/ folder. The folder should contain the unzipped data with the name ground_truth_{phase}, observation_{phase} and observation_challenge.
python src/train.py
usage: train.py [-h] [--batch-size BATCH_SIZE] [--epochs EPOCHS] [--lr LR] [--workers WORKERS] [--patience PATIENCE] [--path PATH] [--load-model]
Training Low Dose CT Scans
optional arguments:
-h, --help show this help message and exit
--batch-size BATCH_SIZE
input batch size for training (default: 4)
--epochs EPOCHS number of epochs to train (default: 8)
--lr LR learning rate (default: 1e-5)
--workers WORKERS number of workers for data loader (default: 4)
--patience PATIENCE random seed (default: 15)
--path PATH Path for loading Model
--load-model For loading the Model
Note: There is a patience parameter and a scheduler in the code but it is unused.
python src/inference.py
usage: inference.py [-h] [--test-batch-size B] [--workers WORKERS] [--path PATH]
Inference for Low Dose CT Scans
optional arguments:
-h, --help show this help message and exit
--test-batch-size B input batch size for test (default: 1)
--workers WORKERS number of workers for data loader (default: 12)
--path PATH Path for loading Model
python src/submission.py
The final architecture used for the challenge was a UNet++ implemented in PyTorch. The model can be found under src/model/Models.py under BiggerUnetPlusPlus. Paper can be found here: https://arxiv.org/pdf/1807.10165.pdf
Filtered Backprojection was used on the observed data and no other preprocessing techniques were used. Vertical flip, horizontal flip and random crop-resize was used during training.
The hyperparameters used are the following:
- batch size = 4
- epochs = 8
- learning rate = 1e-5
- num_workers = 4
- Image size = (362,362) (No changes to the dimensions)
The lose function was a mixed MSE and SSIM described by this (1-alpha)*ssim_loss + alpha*mse where alpha was 0.35.
The optimizer was RMSprop with weight_decay=1e-8 and momentum=0.9
Pytorch API was employed to implement the architecture. Training was done on Ubuntu with a GeForce GTX 2070Super GPU with 8 GB memory and 16 GB of RAM.
Deep Inversion Validation (dival) Library (https://github.com/jleuschn/dival) along with LoDoPaB-CT challenge utilities (https://github.com/jleuschn/lodopab_challenge) were used for submission. I would like to thank the following authors of the libraries for its use.
Daniel Otero Baguer otero@math.uni-bremen.de
Mateus Baltazar
David Erzmann erzmann@uni-bremen.de
Johannes Leuschner jleuschn@uni-bremen.de
Maximilian Schmidt maximilian.schmidt@uni-bremen.de