ImageDataset: accept list of lists of str as image_files

[grudloff]

I am fairly new to monai, so keep that in mind.

I am working with a classification task. The inputs are volumes (Height, Width, Depth), that are loaded by passing folder paths.

The classification are done over a patient, which can have multiple sessions (multiple volumes). Given the constraint of ImageDataset only accepting a single path for each label, the dataset obtained with the code bellow points to training the model on a per session manner.

# paths: List[str] of length N
# labels: nd.array of shape (N, C)
# N: Number of unique patient_id, session_id pairs
# C: Number of classes of label array (one-hot-encoded)

check_ds = ImageDataset(image_files=paths, labels=labels)

im, label = check_ds[0]
# im: metatensor of shape (1, H,W,D)
# label: ndarray of shape (1, C)

Is there an straightforward way in monai to construct a dataset that accepts as input a list of lists of strs, and outputs a list of tensors? Or should I just build my own logic with a custom dataset that internally uses LoadImage?

In summary, I need a CustomDataset like this:

# paths: List[List[str]] of length M
# labels: nd.array of shape (M, C)
# M: Number of unique patient_id
# C: Number of classes of label array (one-hot-encoded)

check_ds = CustomDataset(image_files=paths, labels=labels)

im, label = check_ds[0]
# im: List[metatensor]
# label: ndarray of shape (1, C)

[grudloff]

I am thinking about something like this:

import numpy as np
import monai
from monai.data import Dataset
from monai.transforms import LoadImage, Compose, Spacing, EnsureChannelFirst

class CustomDataset(Dataset):
    def __init__(self, image_files, labels, transform=None):
        self.image_files = image_files
        self.labels = labels
        self.transform = transform

    def __len__(self):
        return len(self.image_files)

    def __getitem__(self, index):
        img_paths = self.image_files[index]
        img_list = [self.transform(path) for path in img_paths]
        label = self.labels[index]
        return img_list, label


transform = Compose([LoadImage(reader="PydicomReader", image_only=True, dtype=np.float32),
                     EnsureChannelFirst(channel_dim='no_channel'),
                     Spacing(pixdim=[RESIZE_DIM]*3),
                    ])

check_ds = CustomDataset(image_files=paths, labels=labels, transform=transform)

im, label = check_ds[0]
print(type(im), [i.shape for i in im], label, label.shape)

Any advise or more elegant solution would be appreciated!

[KumoLiu]

Hi @grudloff, I think you can directly use Dataset which can accept a list of data paths or a list of dictionaries.
https://github.com/Project-MONAI/MONAI/blob/5fd23b8c6a5bcb378063b99523af89c81880060c/monai/data/dataset.py#L64

Hope it helps, thanks!

[KumoLiu]

Yes, could you share a case and then I can help take a deep look?

[grudloff]

Sure! Thanks a lot for the help!

Here is a kaggle notebook with the minimum required code.

[grudloff]

Hello @KumoLiu, I just added a new version to the notebook with a section at the end with an implementation that achieves what I wanted.

[grudloff]

I am still puzzled about pad_list_data_collate not working. In my transform, I normalize the spacing of each volume, shouldn't this make the affine transform the same? Sorry if I am missing something, as I said, I am new to monai and also to working with this kind of data.

[KumoLiu]

Hi @grudloff, I think the reason is that you have nested list in your paths.

And the spacing you have in the nested list are different so when they stack, there will be error throw out.
https://github.com/Project-MONAI/MONAI/blob/281cb0119c01eaa8e6c841880b91f92f45e8d7f7/monai/transforms/io/array.py#L247

Hope it helps, thanks!