Retrieving filename and coordinates from WSIReader

[codybum]

When using WSIReader is there a simple way to retrieve the image coordinates and source filename from the dataset? I am trying to assign a confidence score to a tile and it is not obvious how to retrieve the metadata required to do so.

Thanks

[Nic-Ma]

Hi @drbeh ,

Could you please help share some best practices about this question?

Thanks in advance.

[codybum]

Thanks! I am specifically interested in extracting patch locations from the following tutorial: https://github.com/Project-MONAI/tutorials/blob/bf06be2d2ca7843ff1d8cc01f5b56aa8ac831a93/pathology/multiple_instance_learning/panda_mil_train_evaluate_pytorch_gpu.py#L86

[codybum]

This looks like a repeat of this: #4090

[codybum]

The dataloader batch_data instance seems to represent data for a single slide. Batch_data (from loader) contains dict_keys(['image', 'label', 'image_meta_dict', 'original_spatial_shape', patch_location, patch_size, num_patches, 'offset', 'label_transforms']), and batch_data[image_meta_dict] contains dict_keys(['backend', 'original_channel_dim', 'spatial_shape', 'num_patches', 'path', 'patch_location', 'patch_size', 'patch_level', 'filename_or_obj', affine, space]. This provides all the metadata I need on the slide-level, but I can't find anything on the patch level. I don't see any metadata associated with the tiles themselves.

print(batch_data[image_meta_dict]) {'backend': ['cucim'], 'original_channel_dim': tensor([0]), 'spatial_shape': tensor([[19291, 20916]]), 'num_patches': tensor([1]), 'path': ['/slides/1.svs'], 'patch_location': tensor([[0, 0]]), 'patch_size': tensor([[19291, 20916]]), 'patch_level': tensor([1]), 'filename_or_obj': ['/slides/1.svs'], affine: tensor([[[1., 0., 0., 0.],
[0., 1., 0., 0.],
[0., 0., 1., 0.],
[0., 0., 0., 1.]]], dtype=torch.float64), space: [RAS]}

[bhashemian]

Hi @codybum,

patch_location and patch_size are both patch-related metadata and you should be able to use them directly. Also path metadata shows the filename of the corresponding slide. However, from you output exapmle, I can see that you only have one patch (the same size as the whole slide) as 'num_patches': tensor([1]). If you want to extract a patch from WSIReader, you should provide both location and size. Otherwise, it loads the entire slide.

On the other hand, you may load the entire slide but use GridPatch or GridSplit to split the slide into patches, which seems not to be the case in your example.

Anyways, if you share the code that you are using, I might be help more.

[codybum]

Hi @drbeh,

I am trying to extract coordinates from the pathology MIL tutorial (https://github.com/Project-MONAI/tutorials/blob/main/pathology/multiple_instance_learning/panda_mil_train_evaluate_pytorch_gpu.py) during val_epoch, which uses GridPatchd (https://github.com/Project-MONAI/tutorials/blob/283cfb1858c6426e29442d94d23b8fcafdb471c2/pathology/multiple_instance_learning/panda_mil_train_evaluate_pytorch_gpu.py#L303). I had considered using WSIReader directly, but GridPatchd at does eliminate some unneeded tiles through its thresholding, at least if I understand it correctly.

You are correct in that the first num_patches = 1, but the second shows the correct patch size (256) and num_patches = 6075.

Here is a print of batch_data, notice the last line of output (https://github.com/Project-MONAI/tutorials/blob/283cfb1858c6426e29442d94d23b8fcafdb471c2/pathology/multiple_instance_learning/panda_mil_train_evaluate_pytorch_gpu.py#L108)

{'image': tensor([[[[[0.0431, 0.0431, 0.0431, ..., 0.0431, 0.0431, 0.0431],
[IMAGE DATA REMOVED]
[0.0471, 0.0471, 0.0471, ..., 0.0471, 0.0471, 0.0471]]]]]), 'image_meta_dict': {'backend': ['cucim'], 'original_channel_dim': tensor([0]), 'spatial_shape': tensor([[19291, 20916]]), 'num_patches': tensor([1]), 'path': ['/slides/443043.svs'], 'patch_location': tensor([[0, 0]]), 'patch_size': tensor([[19291, 20916]]), 'patch_level': tensor([1]), 'filename_or_obj': ['/slides/443043.svs'], affine: tensor([[[1., 0., 0., 0.],
[0., 1., 0., 0.],
[0., 0., 1., 0.],
[0., 0., 0., 1.]]], dtype=torch.float64), space: [RAS]}, 'label': tensor([[0., 0., 0., 0., 0.]]), 'original_spatial_shape': [tensor([19291]), tensor([20916])], patch_location: tensor([[0, 0]])
Metadata
backend: cucim
original_channel_dim: 0
spatial_shape: [19291 20916]
num_patches: 1
path: /slides/443043.svs
patch_location: [0 0]
patch_size: [19291 20916]
patch_level: 1
filename_or_obj: /slides/443043.svs
affine: tensor([[1., 0., 0., 0.],
[0., 1., 0., 0.],
[0., 0., 1., 0.],
[0., 0., 0., 1.]], dtype=torch.float64)
space: RAS

Applied operations
[]
Is batch?: False, patch_size: [tensor([256]), tensor([256])], num_patches: tensor([6075]), 'offset': [tensor([0]), tensor([0])], 'label_transforms': [{class: ['ToTensord'], id: tensor([139883705466832]), orig_size: []}]}

Thanks for the help!

[bhashemian]

@codybum, thanks for clarifying on the issue. I know what is happening. There are two patch_location here:

1. the location of patch extracted by WSIReader, which is [0, 0] here because we are loading the entire whole slide image.
2. the location of patch generated by GridPatch.

The patch_location that you are seeing in the metadata is first one (produced by WSIReader). The one that you need is the second one and this is being overwritten by the custom collate function in the pipeline. I'll make a quick fix for you and will let you know.

[bhashemian]

I made the changes here: Project-MONAI/tutorials#841
Please let me know if your problem is solved. Thanks

[codybum]

Fantastic, I will give it a try.

[codybum]

That did the trick! Thanks for your efforts.

[bhashemian]

Glad that it worked! The Project-MONAI/tutorials#841 is going to be merged soon, so it will be available on main branch too.