Add SIIM ACR Pneumothorax dataset

constantinpape / torch-em

Deep-learning based semantic and instance segmentation for 3D Electron Microscopy and other bioimage analysis problems based on pytorch.

MIT License

72 stars 22 forks source link

Add SIIM ACR Pneumothorax dataset #256

Closed anwai98 closed 4 months ago

anwai98 commented 4 months ago

This is a dataset for segmenting pneumothorax in chest x-rays.

anwai98 commented 4 months ago

In addition, this PR also adds the feature to download files from kaggle using the Kaggle API.

anwai98 commented 4 months ago

Hi @constantinpape

This is a nice spot to review a few new features:

Downloads are now possible from Kaggle (some simple steps in the beginning to make the API work)
In ImageCollectionDataset, it allows now to accept patch_shape as None, and load the entire image as it is.
Added a new transformation function called ResizeInputs, which is the skimage.transform.resize operation to resize the inputs based on a desired patch shape.

Let me know if you spot something.

PS. The dataset is around 4GB, in case you want to test it out.

anwai98 commented 4 months ago

Thanks for the mentions!

Re: This logic should also be implemented for the SegmentationDataset. Otherwise there will be a disconnect in the functionality.

I was thinking of doing it in another relevant PR once this is merged, where segmentation dataset is used (ideally for a 3d case like ACDC). Would that work, or should I take care of this in this PR?

constantinpape commented 4 months ago

Would that work, or should I take care of this in this PR?

Yes, that works!