Stroke infarct growth prediction (3D, PyTorch 0.3)
Learning to Predict Stroke Infarcted Tissue Outcome based on Multivariate CT Images
The source code is working from within the IMI network at University of Luebeck, as the closed dataset of 29 subjects is only accessable if you are member of the bvstaff group. The filenames have been renamed and cases are represented as a subfolder. CTP modalities CBV and TTD are used as input, corresponding manual segmentations for core and penumbra, as well as follow-up lesion segmentation (FUCTMap). The directory contains more files since the work for the Master's thesis of Linda Aulmann.
The dataset specified in data.py is inherited from torch.utils.data.Dataset, thus can be exchanged with other datasets and loaders (At the moment there are two datasets with different transformations for training and validation). The existing Learners expect 3D pytorch tensors of shape BxCxDxHxW, but implementing an own Learner will enable the use of 2D data as well.
Set up a Python 3.5 environment including the packages of requirements.txt file.
The repository consists of the following subfolders:
- common: contains commonly used files such as DTOs, models to be learned, helper files or Inference classes
- learner: contains the Learner for running the different trainings
- tester: contains the Tester for running the different tests (i.e. inference-only)
Further, there are other important files:
- data.py: defines the dataset as mentioned under section Data and contains required transformations
- util.py: contains helper functions
Activate the above environment under section Setup.
For learning the shape space on the manual segmentations run the following command:
train_shape_reconstruction.py ~/tmp/shape_f3.model --lrsteps 200 250 --epochs 300 --outbasepath ~/tmp/shape --channelscae 1 16 24 32 100 200 1 --validsetsize 0.3 --fold 17 6 2 26 11 4 1 21 16 27 24 18 9 22 12 0 3 8 23 25 7 10 19
The --fold is an arbitrary but fixed list of indices between 0 and 28 to specify a fold out of the 29 dataset subjects, from which a fraction specified by --validsetsize will be used as validation data (e.g. for 0.275 and the above fold it means that 17 training and 6 validation cases are used by the Learner).
Always specify a --outbasepath to where files are being saved. This includes the *.model file once a new validation minimum has been reached, and *.png files that plot the losses, metrics and visualize some samples during the training run:
Train a Unet with the same fold as specified before, to use the Unet segmentation for further training of an adapted encoder to predict on segmentations of unseen CTP modalities:
train_unet_segmentation.py ~/tmp/unet_f3.model --epochs 200 --outbasepath ~/tmp/unet --channels 2 16 32 64 32 16 32 2 --validsetsize 0.275 --fold 17 6 2 26 11 4 1 21 16 27 24 18 9 22 12 0 3 8 23 25 7 10 19
The --channels arguments specifies the channels used for each layer, incl. input and output. E.g., for the above command the three scales Unet will process the input with 16, downsample and process with 32, and, again, downsample and process with 64 neurons before it is upsampled again with the reverse order of channel numbers per scale. To test the trained Unet on some cases, run:
test_unet_segmentation.py ~/tmp/unet_f3.model --outbasepath ~/tmp/tmp --channels 2 16 32 64 32 16 32 2 --fold 5 13 14 15 20 28
For comparison pruposes, you can run a shape interpolation via signed distance maps:
sdm_resampling.py /share/data_zoe1/lucas/Linda_Segmentations/tmp/tmp_unet_f3.model --fold 22 --downsample 0 --groundtruth 1
The experiments in the article "Learning to predict ischemic stroke growth on acute CT perfusion data by interpolating low-dimensional shape representations" have been conducted with the following parameters (command for fold 5):
python train_shape_reconstruction.py --channelscae 1 16 24 32 100 800 1 --outbasepath /tmp/shape_f5 --validsetsize 0.275 --epochs 200 --fold 17 6 2 26 11 4 1 21 16 27 24 18 15 20 28 14 5 13 9 22 12 0 3 8
Listed under: https://www.researchgate.net/project/Learning-to-predict-stroke-outcome-on-multivariate-CT-data