Dataset for longitudinal imaging data

In the medical domain, patients often undergo repeated image screening. Yet, all datasets so far implemented in MONAI only consider cross-sectional data. I think a dataset that handles longitudinal imaging data would therefore be a great addition to the repository.

I have already built a longitudinal dataset with and without caching of non-random transforms in my own research project and I would be happy to share it with this great repository. Just like other MONAI datasets, the dataset expects a list of dictionaries with the expectation that the data is ordered longitudinally:

[{                                                                         {
'T1': ['t1image1_1.nii.gz', 't1image1_2.nii.gz', 't1image1_3.nii.gz'],     'T1': ['t1image2_1.nii.gz', 't1image2_2.nii.gz', 't1image2_3.nii.gz']
'T2': ['t2image1_1.nii.gz', 't2image1_2.nii.gz', 't2image1_3.nii.gz'],     'T2': ['t2image2_1.nii.gz', 't2image2_2.nii.gz', 't2image2_3.nii.gz']
'extra': [123, 456, 789]                                                   'extra': [123, 456, 789]
},                                                                         }]

Additionally, an argument specifying the image keys (e.g., T1 and T2 in the above example) needs to be added during the datasets initialisation:

dataset = SeqDataset(
   data=data,
   image_keys=['T1', 'T2'],
   transforms=transforms)

Note: If there is interest in the longitudinal dataset, I also implemented a custom collate function (i.e., SequenceBatchCollater) that automatically pads or truncates the sequence of images to the desired length and can be used with any MONAI dataloader:

dataloader = monai.data.DataLoader(
    dataset=dataset,
    ...,
    collate_fn=SequenceBatchCollater(keys=['image', 'label'], seq_length=8)

With the custom collate function, the dataloader returns a batch of shape [B x S, C, H, W, ...], where S is the sequence length.

Let me know what you think of the implementation. In the case of interest, I would be happy to make a contribution to the project.

Start by reviewing the proposed SeqDataset in data/datasets.py and SequenceBatchCollater in data/utils.py, then compare them with MONAI's existing dataset and DataLoader entry points. Check how image_keys, longitudinal ordering, caching, padding, and truncation would fit the repository's current APIs; done means an agreed design and integrated dataset behavior with suitable tests.