Using nifty file as input for my MAP application

[ArpanGyawali]

I have created MAP for taking dicom data as input and get the segmentation for skull stripping task following the spleen seg inference application.
We have more data in nifty format than in dicom so I wanted to create a additional application or combine with the previous that takes nifty as well.
I tried using Nii_file_loader_operator but it returns numpy array(np.ndarray) and when i tried to give this output to skull_strip_operator(similar to spleen_seg_operator), as it takes monai.core.Image as input. So this raises an error

So i edited the Nii_file_loader_operator by just changing the type explicitly by
op_output.set(Image(image_np))

But this returns Image without any metadata headet so i got the error like
KeyError: 'row_pixel_spacing' thrown by LoadImaged pretransform.

What should i do so that my LoadImaged pretransform takes nifty image from my disk or numpy image returned by nifty_loader from InMemory.

[MMelQin]

@ArpanGyawali A Numpy array as the result of an image loader is never enough in the case of inference. The loader need to provide both the Numpy array as well as the metadata. Further, to fit what's expected by the built-in inference operators, an App SDK Image object is required, which really is just numpy array with a list of expected metadata.

A more efficient implementation is actually do what's done in MONAI core, i.e. passing the file path to the LoadImaged transform instead of using the custom in_mem image reader. The one thing is that this requires extending and creating a custom inference operator.

[ArpanGyawali]

thank you @MMelQin
But I havent used built in inference operator.
For creating the skull strip application I have used the following operator (version: sdk 0.5.1)

Dicom Loader
Dicom Series Selector
Dicom Series to Volume
Custom SkullStrip operator (similar to spleen seg operator)
Dicom Writer
and used the model that i got after training custom data in monai label.

So according to you, I can create a custom inference operator where the pretransform LoadImaged can take the nifty file path directly instead of using any other nifty loader operator?
How can I do this? Is there any references?
Any help would be deeply appreciated. Thank you