intensity normalisation transform about monai

[Nic-Ma]

Hi @atbenmurray ,

I committed a PR #25 for this task, could you please help review it?
Glad to see your feedback.
If I misunderstand anything, please feel free to contact me.
Thanks in advance.

[ericspod]

Currently we have the functionality to rescale arrays to [0,1] range, this suffices for the segmentation example so we feel it's beyond the MVP scope.

[Nic-Ma]

Hi @ericspod @atbenmurray @wyli @yanchengnv ,

Thanks for the discussion of the design principle on transforms.
I think we are on the same page to use basic PyTorch only. Then I have several questions about ShapeFormat, could you please help make them clear before we start to develop?

1. torchvision supposes the 2D input images are PIL image which is "channel last", and converts to torch.Tensor which is "channel first" in "toTensor()" transform. Can we also suppose all 3D input images are "channel last"?
2. If not all input images are "channel last", how does every transform know the input image is "channel last" or "channel first"? In previous design, we defined a MediaImage object to record ShapeFormat information with image data together, then we can support 12 kinds of ShapeFormats in transforms.
3. Should the transform support to process both numpy data and torch.Tensor data? If only support numpy data, I think we need to define an implicit transform "toTensor()" at the end of chain.

Welcome your discussion and please feel free to correct me if I misunderstand anything.
Thanks in advance.

[wyli]

@Nic-Ma as I understand it, for the current sprint we assume the input image is in NifTI format, which uses 'channel last'. Before feeding it to the network we add another 'ToTensor' transform to convert it into a 'channel first' torch tensor. Let's keep the implementation "minimal", for example, following torch vision's implementation.

[wyli]

#25 #28 jointly resolve this issue.