Thanks for the implementation of new feature request.
I am wondering if it's possible to support the split patch functionality in a transform, which handles the output of PatchWSIDataset with other transforms?
@wyli @ericspod What do you think?

Thanks.

I lose the overall picture of the WSI-related reader/dataset modules, would be great to have some tutorials in https://github.com/Project-MONAI/tutorials/tree/master/modules to demonstrate the use cases, otherwise it's not easy to help review these PRs.

Thanks for the implementation of new feature request. I am wondering if it's possible to support the split patch functionality in a transform, which handles the output of PatchWSIDataset with other transforms? @wyli @ericspod What do you think?

Thanks.

Typically we would want to decouple behaviours so that we have the option to choose what to use, and reuse functionality elsewhere. Using GridSplit here as a hard-coded transform is something I would guess is done as an efficiency necessity which we had discussed in previous PRs, or just such common practice that it makes sense to have a built-in option. @drbeh correct me if I'm wrong!

(For an example of the tutorials: this PR proposes code changes #4236, and the corresponding usage is shown in a PR to the tutorials Project-MONAI/tutorials#690. These greatly help the reviewing/discussion process...)

@Nic-Ma @ericspod, this split does not have any randomness (like what we have discussed before) and I had in mind to use the transform only but the reason to have the split here is to managing the following transforms. What I mean is that the split transform generate several patches (from few patches to few hundreds or thousands of patches), and we want these patches treated independently like images in a batch. It was not clear to me how its output will be handle by its following transforms. Assume the following:

• input patch: (3, 16, 16)
• (3, 20, 24) -> GridSplit (2x2) -> (4, 3, 5, 6) [mind the additional dimension]
• (4, 3, 5, 6) -> OTHER TRANSFORMS -> (4, 3, 5, 6) [does all other transform acknowledge the batch dimension?]
• (4, 3, 5, 6) -> BATCHING (10) -> (40, 3, 5, 6) [not (10, 4, 3, 5, 6), can we achieve this without a new collate function?]

I would really like to make this as a separate transform so I'd appreciate it if you have any solution that can create the batches as we like (without implementing a new collate function).

Hi @drbeh ,

I think your problem is similar to our RandCropByPosNegLabeld transform which generates a list of patches:
https://github.com/Project-MONAI/MONAI/blob/dev/monai/transforms/croppad/dictionary.py#L1042
The following transforms will apply for each item of the patch list:
https://github.com/Project-MONAI/MONAI/blob/dev/monai/transforms/transform.py#L88

Thanks.

https://github.com/Project-MONAI/MONAI/blob/dev/monai/transforms/transform.py#L88

@Nic-Ma, this is what the current implementation is already leveraging and that's why I put GridSplit in the dataset so the transform chain can be applied on each of them.

However, this is not the case if a transform at the middle of chain creates a list of images, right?

I think your problem is similar to our RandCropByPosNegLabeld transform which generates a list of patches:
https://github.com/Project-MONAI/MONAI/blob/dev/monai/transforms/croppad/dictionary.py#L1042

I think the problem is similar, so how this is being batched? Is there any issue for the other transforms if it is a list?

RandCropByPosNegLabeld can be in the middle of a transform chain, and we also have several other transforms generate a list of samples.
monai.data.DataLoader can batch for this case with the built-in collate function.

Thanks.

@Nic-Ma @ericspod, I have updated the GridSplit and now the transform and dataset are completely separated. Thanks for your inputs.

@wyli, I have updated the tumor detection pipeline that uses these transforms and dataset: Project-MONAI/tutorials#697

@wyli, I have updated the tumor detection pipeline that uses these transforms and dataset: Project-MONAI/tutorials#697

thanks, @Nic-Ma could you please help run and verify the tutorial (perhaps with a subset of CAMELYON16 for a quick setup)?

@wyli, I have updated the tumor detection pipeline that uses these transforms and dataset: Project-MONAI/tutorials#697

thanks, @Nic-Ma could you please help run and verify the tutorial (perhaps with a subset of CAMELYON16 for a quick setup)?

That's a good idea. I'll add a subset dataset to the PR of the tutorial to facilitate the testing but we can remove it before merge.

@Nic-Ma @wyli, I have added a subset dataset ('training.txt') into this tutorial PR: Project-MONAI/tutorials#697

• You can use this dataset for both training and validation with a flag of --train-file ../training.txt --valid-file ../training.txt since we don't want to check any performance here.

• You also need to download one image tumor_091.tif and store it in IMAGE_ROOOT/training/images/, where IMAGE_ROOT is arbitrary and can be set using --root IMAGE_ROOT: https://drive.google.com/file/d/0BzsdkU4jWx9BVWx1eERwYnlVWVE/view?usp=sharing&resourcekey=0-Zd3y-ucNOjPq3nqsx-P3sQ

Please let me know if you have any question.

Hi @drbeh @wyli ,

I tried to run the tutorial, seems it can run normally:

Thanks.

/build