piechaczekmyller / brats

BRATS

brats

BRATS

brats's Issues

Add inference script

Choose a system for experiments repeatability and reproducibility

Adjust the metrics and losses to fit multiclass output

Implement 3D Unet

Implement working 3D Unet architecture. One of my concerns is that volumes might have different number of slices. The network should deal with that.

Branch names convention

I suggest new branches would follow convention issue#X, otherwise in a moment we would have large amount of branches with similar names

What are your comments?

Implement/customize DICE and boundary loss to match 3D ground truth

Because of the fact that we are now working with 3D data, we need to check whether DICE and boundary losses could and should be calculated differently for volumes.

Add standardization with filtering background

Fix NaN issue while training

Find how to deal with uneven number of slices

As the image downscales x2 with each step along each dimension, it is impossible for it to deal with images that are uneven in any dimension. In decathlon dataset (and brats as well), the number of slices is 155.

How to deal with this problem? Add empty slice/copy 1 slice?
Please, comment @jnalepa

Describe data in wiki

Data, modalities order, statistics for each modality etc

Add dataset that would merge 2 NiftiDatasets for training

Try to run it on NVIDIA machine

Describe model in wiki

Prepare dataloader

Check if new Pillow works on ubuntu

Add trainers

Prepare training script

Implement preprocessing script

Write preprocessing code to structure, normalize and generate ready-to-feed data. One should keep in mind that it will be later needed to preprocessed Brats 2020(?) dataset. Also, final format of the data should be picked.

Find/implement Hausdorff distance

Find whether obtained data is preprocessed in any way

Check boundary loss

Add visualization tools

Add NLL loss

Implemented loss which uses DICE and Cross Entropy loss with conjunction.

Check whether torch.contiguous() does anything in our case

Run tumor detection training

Change training script so it uses ComposeLoss and standardization with filtering

Create transformations needed for training