Facing issues with saving nifti image with SaveImaged because of multi channel input

[lenlobo]

I am working with 3d segmentation using Monai and Nifti files as inputs and using a model which has 6 input channels, I get an error in the inference testing of the model.

class Net(pytorch_lightning.LightningModule):
    def __init__(self):
        super().__init__()
        self._model = UNet(
            dimensions=3,
            in_channels=6,
            out_channels=2,
            channels=(16, 32, 64, 128, 256),
            strides=(2, 2, 2, 2),
            num_res_units=2,
            norm=Norm.BATCH,
        )

So after using the tutorial code from "https://github.com/Project-MONAI/tutorials/blob/72728b678947bc21443d02d3a19d29f01c9eed01/3d_segmentation/spleen_segmentation_3d.ipynb", as a reference, the issue I end up is at post transforms to revert the pretransforms and save the output file.

As I am using 6 inputs, and end up with only one output, I get an error stating,

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
Input In [31], in <cell line: 1>()
      1 post_transforms = Compose([
----> 2     Invertd(
      3         keys="pred",
      4         transform=test_transforms,
      5         orig_keys=img_keys,
      6         meta_keys="pred_meta_dict",
      7         meta_key_postfix="meta_dict",
      8         nearest_interp=False,
      9         to_tensor=True,
     10     ),
     11     AsDiscreted(keys="pred", argmax=True, to_onehot=2),
     12     SaveImaged(keys="pred", meta_keys="pred_meta_dict", output_dir="./out", output_postfix="seg", resample=False),
     13 ])


File /opt/monai/monai/transforms/post/dictionary.py:592, in Invertd.__init__(self, keys, transform, orig_keys, meta_keys, orig_meta_keys, meta_key_postfix, nearest_interp, to_tensor, device, post_func, allow_missing_keys)
    590     raise ValueError("transform is not invertible, can't invert transform for the data.")
    591 self.transform = transform
--> 592 self.orig_keys = ensure_tuple_rep(orig_keys, len(self.keys)) if orig_keys is not None else self.keys
    593 self.meta_keys = ensure_tuple_rep(None, len(self.keys)) if meta_keys is None else ensure_tuple(meta_keys)
    594 if len(self.keys) != len(self.meta_keys):

File /opt/monai/monai/utils/misc.py:154, in ensure_tuple_rep(tup, dim)
    151 if len(tup) == dim:
    152     return tuple(tup)
--> 154 raise ValueError(f"Sequence must have length {dim}, got {len(tup)}.")

**ValueError: Sequence must have length 1, got 6.`**

Now, as I am passing img_keys which is a list of 6 keys, eg, input_1, input_2, etc, it requires one single key instead of 6, but in pre transform I have used 6 keys because of 6 input channels.
This is what my output dictionary keys look like after sliding_window_inference :
dict_keys(['input_1', 'input_2', 'input_3', 'input_4', 'input_5', 'input_6', 'input_1_meta_dict', 'input_2_meta_dict', 'input_3_meta_dict', 'input_4_meta_dict', 'input_5_meta_dict', 'input_6_meta_dict', 'pred'])
To summarize, I have 6 inputs, using all 6 inputs, I generate an output tensor, but when I try to save it using the post transform, I face issues with Value Error.
Any solution for this, multi input channel, single output channel?

[KumoLiu]

Hi @lenlobo, If you have images with 6 input channels, you can simply use one key to preprocess them like:

trans = Compose([
    LoadImaged(keys='image'),
    ......
])

Datalist like: [
{'image': [input_1, input_2, ....]
'label': label},
......

]
You can also refer to brats segmentation tutorial for reference which uses 4 modalities as input.
Hope it can help you, thanks!

[KumoLiu]

Hi @lenlobo, from the code you post, I noticed that you are using torch.utils.data.DataLoader. I would suggest using DataLoader in MONAI directly. If you are creating a DataLoader by yourself, you can also replace collate_fn and worker_init_fn with MONAI components.

MONAI/monai/data/dataloader.py

Lines 84 to 87 in 5b3b741

	if "collate_fn" not in kwargs:
	kwargs["collate_fn"] = list_data_collate
	if "worker_init_fn" not in kwargs:
	kwargs["worker_init_fn"] = worker_init_fn

And I also noticed that you add ToTensord in your transforms, you can just remove this line.
Hope it can help you, thanks!

[lenlobo]

Hi @KumoLiu, once I removed the ToTensord line, all my data did change from torch.Tensor to MetaTensor, now when I try to apply post transforms on my prediction by using test_data = [post_transforms(i) for i in decollate_batch(test_data)] I get these errors.

---------------------------------------------------------------------------
IndexError                                Traceback (most recent call last)
File /opt/monai/monai/transforms/transform.py:90, in apply_transform(transform, data, map_items, unpack_items, log_stats)
     89         return [_apply_transform(transform, item, unpack_items) for item in data]
---> 90     return _apply_transform(transform, data, unpack_items)
     91 except Exception as e:

File /opt/monai/monai/transforms/transform.py:54, in _apply_transform(transform, parameters, unpack_parameters)
     52     return transform(*parameters)
---> 54 return transform(parameters)

File /opt/monai/monai/transforms/utility/dictionary.py:523, in ToTensord.inverse(self, data)
    522     # Remove the applied transform
--> 523     self.pop_transform(d, key)
    524 return d

File /opt/monai/monai/transforms/inverse.py:206, in TraceableTransform.pop_transform(self, data, key, check)
    192 """
    193 Return and pop the most recent transform.
    194 
   (...)
    204     - RuntimeError: data is neither `MetaTensor` nor dictionary
    205 """
--> 206 return self.get_most_recent_transform(data, key, check, pop=True)

File /opt/monai/monai/transforms/inverse.py:188, in TraceableTransform.get_most_recent_transform(self, data, key, check, pop)
    187 if check:
--> 188     self.check_transforms_match(all_transforms[-1])
    189 return all_transforms.pop() if pop else all_transforms[-1]

IndexError: list index out of range

The above exception was the direct cause of the following exception:

RuntimeError                              Traceback (most recent call last)
File /opt/monai/monai/transforms/transform.py:90, in apply_transform(transform, data, map_items, unpack_items, log_stats)
     89         return [_apply_transform(transform, item, unpack_items) for item in data]
---> 90     return _apply_transform(transform, data, unpack_items)
     91 except Exception as e:

File /opt/monai/monai/transforms/transform.py:54, in _apply_transform(transform, parameters, unpack_parameters)
     52     return transform(*parameters)
---> 54 return transform(parameters)

File /opt/monai/monai/transforms/post/dictionary.py:664, in Invertd.__call__(self, data)
    663 with allow_missing_keys_mode(self.transform):  # type: ignore
--> 664     inverted = self.transform.inverse(input_dict)
    666 # save the inverted data

File /opt/monai/monai/transforms/compose.py:183, in Compose.inverse(self, data)
    182 for t in reversed(invertible_transforms):
--> 183     data = apply_transform(t.inverse, data, self.map_items, self.unpack_items, self.log_stats)
    184 return data

File /opt/monai/monai/transforms/transform.py:114, in apply_transform(transform, data, map_items, unpack_items, log_stats)
    113         _log_stats(data=data)
--> 114 raise RuntimeError(f"applying transform {transform}") from e

RuntimeError: applying transform <bound method ToTensord.inverse of <monai.transforms.utility.dictionary.ToTensord object at 0x7fada5dffd90>>

The above exception was the direct cause of the following exception:

RuntimeError                              Traceback (most recent call last)
Input In [75], in <cell line: 1>()
----> 1 test_data = [post_transforms(i) for i in decollate_batch(test_data)]

Input In [75], in <listcomp>(.0)
----> 1 test_data = [post_transforms(i) for i in decollate_batch(test_data)]

File /opt/monai/monai/transforms/compose.py:173, in Compose.__call__(self, input_)
    171 def __call__(self, input_):
    172     for _transform in self.transforms:
--> 173         input_ = apply_transform(_transform, input_, self.map_items, self.unpack_items, self.log_stats)
    174     return input_

File /opt/monai/monai/transforms/transform.py:114, in apply_transform(transform, data, map_items, unpack_items, log_stats)
    112     else:
    113         _log_stats(data=data)
--> 114 raise RuntimeError(f"applying transform {transform}") from e

RuntimeError: applying transform <monai.transforms.post.dictionary.Invertd object at 0x7fab21bc78b0>

Pretty close to fixing the issue here but need a bit more assistance! :) Thanks!

[KumoLiu]

Hi @lenlobo, if you remove ToTensord in test post transforms, you should also remove it in train post transforms. :) Thanks!

[lenlobo]

Hi @KumoLiu, you mean the transforms when I trained my model? Because I do not really have that code, I was using a model which is trained by someone else at my organization and just testing inputs by creating an inference script, or, are you talking about to_tensor=True in the post transforms mentioned by me in the above reply?

[lenlobo]

@KumoLiu Hi, this has kind off solved my issues, thank you so much! this helped a lot!