Prepare image_files for mulitple channels and seg_files for ImageDataset

[relyativist]

Hi, I am trying to set up monai.data.ImageDataset, but can't understand how to create image_files and seg_files dicts for input. By now I have created dict train_files for 10 subjects:

[{'image': ['/input/sub-8/T1.nii.gz',
   '/input/sub-8/feauture_map0.nii.gz',
   '/input/sub-8/feauture_map1_mni.nii',
   '/input/sub-8/feauture_map2.nii.gz',
   '/input/sub-8/feauture_map3.nii.gz'],
  'mask': '/input/sub-8/sub-8T1_brain-final_mask.nii.gz',
  'seg': '/input//label/8.nii.gz'},
 {'image': ['/input/sub-27/T1.nii.gz',
   '/input/sub-27/feauture_map0.nii.gz',
   '/input/sub-27/feauture_map1_mni.nii',
   '/input/sub-27/feauture_map2.nii.gz',
   '/input/sub-27/feauture_map3.nii.gz'],
  'mask': '/input/sub-27/sub-27T1_brain-final_mask.nii.gz',
  'seg': '/input//label/27.nii.gz'},
 {'image': ['/input/sub-53/T1.nii.gz',
   '/input/sub-53/feauture_map0.nii.gz',
   '/input/sub-53/feauture_map1_mni.nii',
   '/input/sub-53/feauture_map2.nii.gz',
   '/input/sub-53/feauture_map3.nii.gz'],
  'mask': '/input/sub-53/sub-53T1_brain-final_mask.nii.gz',
  'seg': ‘/input/label/53.nii.gz'},
...

Then I am create dataset and dataloader with:

train_ds = monai.data.ImageDataset(train_files, seg_files=train_files['seg'], transform=train_transf)`  with no error.
check_loader = DataLoader(train_ds, batch_size=1, num_workers=0)

and have an error:
RuntimeError: LoadImage cannot find a suitable reader for file: ('image', 'mask', 'seg').

I assume that dict for seg_files should be for each subject, But how to pass multiple channels for image?
Thank you

[Nic-Ma]

Hi @relyativist ,

Thanks for your interest here.
The ImageDataset supposes to work with array data instead of dict data.
For your case, I would suggest to use Dataset and dict based transforms:
https://github.com/Project-MONAI/MONAI/blob/dev/monai/data/dataset.py#L57
You can refer to the Brats tutorial for more details:
https://github.com/Project-MONAI/tutorials/blob/main/3d_segmentation/brats_segmentation_3d.ipynb

Thanks.

[relyativist]

@Nic-Ma I assume that I should stick with 'ImageDataset' because I want to create PatchDataset from images. Following
https://docs.monai.io/en/stable/data.html#patchdataset, to create Patch Dataset, I have to assign data as ImageDataset.
Or there are options to Patch Dict dataset, like monai.data.Dataset?

[KumoLiu]

Hi @relyativist, first you may need to take the feature map out of the image and turn it into:

data = [
{'image': '/input/sub-8/T1.nii.gz',
  'map0': '/input/sub-8/feauture_map0.nii.gz',
  'map1': '/input/sub-8/feauture_map1_mni.nii',
  'map2':'/input/sub-8/feauture_map2.nii.gz',
  'map3':'/input/sub-8/feauture_map3.nii.gz',
  'mask': '/input/sub-8/sub-8T1_brain-final_mask.nii.gz',
  'seg': '/input//label/8.nii.gz'},
.....
]

And then you may try PatchIterd and GridPatchDataset like this to create Patch Dataset:

dataset = Dataset(data=data, transform=pre_transforms)
patch_iter = PatchIterd(keys=["image", "seg"], patch_size=(64, 64), start_pos=(0, 0))
ds = GridPatchDataset(data=dataset, patch_iter=patch_iter, transform=transforms, with_coordinates=True)

Hope it can help you! Thanks!

[KumoLiu]

Hi @relyativist, if you want to stack all the files in the image key, you can just keep your original train_files, ignore what I said about taking the feature map out of image, thanks!

[relyativist]

@KumoLiu Could you please clarify differences between transforms in GridPatchDataset and Dataset
Thanks

[KumoLiu]

Hi @relyativist, transforms in GridPatchDataset is a callable data transform that operates on the patches, and transforms in Dataset operates on the images if you need some preprocess on the images before you create patches. Thanks!