PatchDataset/GridPatchDataset or SlidingWindowInferer for patch-based train with large volume

[Littlehhh]

Description

Hi
I found that there are several ways in MONAI to implement patch-based(grid patch) training for large medical volume.

1. PatchDataset/GridPatchDataset with SimpleInferer
2. Dataset with SlidingWindowInferer
3. Dataset with RandCropByPosNegLabel/(or extra transfrom)

So which way should be better? And these ways have different loss(single patch/full volume/several patch) backpropagations.
I am confused about this and I think maybe the first way is a more reasonable way for patch-based system.

In addition, from the perspective of framework design, which way is more appropriate, and what are the uses of other ways for patch-based train?

Thanks.

[Nic-Ma]

Hi @Littlehhh ,

Thanks for your interest here.

1. RandCropByPosNegLabel is a transform and usually used to generate random patches during training, the following transforms will apply to every patch separately.
2. SlidingWindowInfer is used to split big image into patches for the network forward(), so it doesn't contain random logic and the input data is already handled by transforms, usually used for validation.
3. PatchDataset can do global shuffling on the patches.

You can start from this example:
https://github.com/Project-MONAI/tutorials/blob/master/3d_segmentation/spleen_segmentation_3d.ipynb

Thanks.