PatchInferer and Monai server

[mdrio]

Hi, I am trying to do some inference with a TorchScript model on whole slide images using a MONAILabelApp and DSA. I am able to use the model using the PatchInferer and the WSISlidingWindowSplitter. How can it be integrated in a MONAILabelApp? I checked the sample-apps/pathology, but apparentely only SimpleInferer is used.
Thanks

[MathijsdeBoer]

Perhaps you could make use of the monai.inferers.SlidingWindowInferer? I'm not experienced with whole slide images, so I'm not sure whether that'll work for you.

[MathijsdeBoer]

Rereading I realized that you already know how to run a prediction. I think you can just change the InferTask class in ".../lib/infers/" to use the PatchInferer instead. For example, apps/radiology has a few examples that define the inferer() method as follows:

    def inferer(self, data=None) -> Inferer:
        return SlidingWindowInferer(
            roi_size=self.roi_size,
            sw_batch_size=2,
            overlap=0.5,
            padding_mode="replicate",
            mode="gaussian",
        )

[mdrio]

Ok, thanks. Actually I am using PatchInferer and SlidingWindowSplitter. Anyway I got it working, but now I have the following problem: since the model is patch based , the output is rescaled. So if a 2560x2560 region is processed with 256x256 patches, the output is 10x10. I correctly get the annotation (segmentation) in DSA format using FindContoursd transform (present in MonaiLabel app code), but rescaled. Is there any utility I can use for getting the result with the right scale? O do I have to do it manually (perhaps with Shapely)?

[MathijsdeBoer]

Perhaps monailabel.transform.post.Restored can help you out? The radiology apps seem to use it to return an output to match the input(s).