Create “image” for ImageDataset

[relyativist]

I am trying to upload “image” with 8 channels for each subject (8,294,300,270) (1 is for T1w and 7 are for extracted Feature maps), to do this I just stack volumes to nn.Stack. I got following question:

• What should be the values of key “image”? Referencing doc, it should be paths to nifti images, but I got tensors (3,294,300,270). How it could be uploaded to ImageDataset?

[Nic-Ma]

Hi @relyativist ,

Thanks for your interest here.
I would suggest you to use the dict based transforms and dataset:
https://github.com/Project-MONAI/MONAI/blob/dev/monai/data/dataset.py#L49
So you can put any data in the dict with "image" key.

Thanks.

[Nic-Ma]

You can refer to this example: https://github.com/Project-MONAI/tutorials/blob/master/3d_segmentation/torch/unet_training_dict.py#L86.

Thanks.

[relyativist]

Hey @Nic-Ma, thanks for quick response. I am being unclear about details. What I was asking is how data with dict can be uploaded to Dataset?
I have input list of key, "image" and "seg" pair. With multiple "image" channels:

[{'image': {'t1': '8.nii.gz',
   'curv': 'curv.nii.gz',
   'sulc': 'sulc.nii.gz',
   'contrast1: 'contrast1.nii.gz',
   'contrast2': 'contrast2.nii.gz',
   'glcm': 'glcm.nii.gz',
   'contrast3': 'contrast3.nii.gz',
   'contrast4': 'contrast4.nii.gz'},
  'label': 'label_8.nii.gz'},
  ...
  {}]

I tried to follow https://github.com/Project-MONAI/tutorials/blob/master/3d_segmentation/torch/unet_training_dict.py#L86, but stuck on check_ds = monai.data.Dataset(data=train_files, transform=train_transforms), as i get multiple maps for 'image'.

[relyativist]

@Nic-Ma Thanks, That worked!

[relyativist]

@Nic-Ma Little detail I missed, if I need to make Transform on particular channel, e.g. 'curv.nii.gz', 'sulc.nii.gz',, for example:

im_trans = Compose(
    [
        ScaleIntensityd(keys=['curv.nii.gz', 'sulc.nii.gz'], minv=0.0, maxv=1.0, channel_wise=True),
        LoadImaged(keys=["image", "label"]),
        Resized(keys=['curv.nii.gz', 'sulc.nii.gz'], spatial_size=(241)),
    ]
)

It's seems is will throw an error, because there is no such keys in this dict, as they are values of key 'image'. Is it possible to apply transform on particular channels?

[Nic-Ma]

Hi @relyativist ,

If you resize images only on particular channels, you can't stack them together as they have different spatial shape anymore.
If you don't want to stack images, you can save every channel with a key in the dict directly, then apply resize transform on expected keys, like:

[
  {
   't1': '8.nii.gz',
   'curv': 'curv.nii.gz',
   'sulc': 'sulc.nii.gz',
   'contrast1: 'contrast1.nii.gz',
   'contrast2': 'contrast2.nii.gz',
   'glcm': 'glcm.nii.gz',
   'contrast3': 'contrast3.nii.gz',
   'contrast4': 'contrast4.nii.gz',
   'label': 'label_8.nii.gz'
  },
  ...
  {}
]

Thanks.

[relyativist]

@Nic-Ma Thanks

If you resize images only on particular channels, you can't stack them together as they have different spatial shape anymore. If you don't want to stack images, you can save every channel with a key in the dict directly, then apply resize transform on expected keys, like:

Is it possible to stack each key to create dataset, after transform separate keys?

[Nic-Ma]

Yes, if the transformed images have same spatial shape, you can concat them with ConcatItemsd transform:
https://github.com/Project-MONAI/MONAI/blob/dev/monai/transforms/utility/dictionary.py#L951

Thanks.