How to load multiple nifti images per subject in ImageDataset ?

[mpsampat]

Hi All!
I have a basic question.
I have a classification task where I have 4 nifti files per subjects: these are FLAIR, T2w, T1w and T1w_CE.
I was able to build an ImageDataset with one one nifti file type as input:
train_ds = ImageDataset(image_files=images, labels=labels, transform=train_transforms)

where image files is
'/mnt/kg/rsna//train/00000/T1CE_to_SRI.nii.gz',
'/mnt/kg/rsna//train/00002/T1CE_to_SRI.nii.gz',
'/mnt/kg/rsna//train/00003/T1CE_to_SRI.nii.gz',
'/mnt//kg/rsna//train/00005/T1CE_to_SRI.nii.gz',
'/mnt/kg/rsna//train/00006/T1CE_to_SRI.nii.gz',

However I cannot figure out how to build an ImageDataset with 4 nifti files per subject,
Could you please help me with this issue ?
Thanks
Mehul

[Nic-Ma]

Hi @mpsampat ,

For this case, you can try to make every 4 files under 1 dir, use this dir as the image path for 1 data.
MONAI will detect it's a folder and stack the 4 images as 4 channels:
https://github.com/Project-MONAI/MONAI/blob/0.6.0/monai/data/image_reader.py#L383
I will try to enhance the doc-string to make it more clear.

Thanks.

[mpsampat]

thanks very much @Nic-Ma ! I had to make make some minor modifications but works well.
this is my same data-set with 1 or 2 nifti files.
#one nifti file per subject;
<class 'torch.Tensor'> torch.Size([2, 1, 240, 240, 155])
two nifti file per subject;
<class 'torch.Tensor'> torch.Size([2, 2, 240, 240, 155])

[mpsampat]

I made a custom Densenet model:
def get_new_net():
return DenseNet(
spatial_dims=3,
in_channels=1, # this is 1 for 1 nifti and 2 for 2 nifti and so on
out_channels=2,
init_features=16,
growth_rate=16,
block_config=(4, 4, 4, 4),
).to(device)

[mpsampat]

lastly I had to make changes to the transforms:

transforms for 1 nifti file

train_transforms = Compose([ScaleIntensity(),AddChannel(),EnsureType()])
val_transforms = Compose([ScaleIntensity(),AddChannel(),EnsureType()])

transforms for 2 or more nifti files

train_transforms = Compose([ScaleIntensity(),EnsureType()])
val_transforms = Compose([ScaleIntensity(),EnsureType()])

[Nic-Ma]

You can use EnsureChannelFirst() transform instead of AddChannel(), it can support all cases.

Thanks.

[Nic-Ma]

I submitted PR: #2876 to enhance the doc-string for multi-files reading.

Thanks.

[mpsampat]

@Nic-Ma can I read the 4 nifti files and save them as a 4-D nifti instead ?

[Nic-Ma]

Yes, the SaveImage transform can support it.

Thanks.