Implementing Datastore to access a database without file URIs

[bhoeckendorf]

I'm attempting to add rudimentary support for the Flywheel database to a MONAILabel app by implementing monailabel.interfaces.datastore.Datastore, and I'm confused how Datastore.get_image and Datastore.get_image_uri are meant to be used.
In this case, the image files are located in the Flywheel database and I don't always have a HTTP URI, but I can read them using the Flywheel SDK. So now I seem to have the following situation:

class FlywheelDatastore(monailabel.interfaces.datastore.Datastore):

    # ...

    def get_image(self, image_id: str, params=None) -> bytes:
	"""
	Returns bytes from data file, using Flywheel SDK.
	Works when called explicitly, but is never called by MONAILabel.
	"""
	# ...
	    

    def get_image_uri(self, image_id: str) -> str:
	"""
	N/A, no valid URI, we must use the Flywheel SDK.
	MONAILabel calls this function, and then rightfully complains
	that the file doesn't exist.
	"""
	# ...

It isn't completely clear to me why we need both functions to begin with. Datastore should either provide a URI for us to read, or read the file for us but then we don't need its URI.

What am I missing and how do I proceed?

[SachidanandAlle]

it is normally local file/access path.. that can represent location of your image..

get_image_uri is used when you run inference or training.. this is the time it likes to check local existence of your image data so that it can run operations on that data..

i guess you have data in remote database.. and when someone calls get_image_uri you would like to download it from remote and cache it.. something similar to xnat/dsa datastore..

https://github.com/Project-MONAI/MONAILabel/blob/main/monailabel/datastore/xnat.py#L92

[SachidanandAlle]

If you want don't want to support cache.. then there is a workaround.. you need to use your own LoadImage transform that can fetch the image directly from your remote database.. but ideally this user-experience is slow.. specially when you run some interactions on the same image again and again..

so simplest way is to download and cache the image.. when someone calls get_image_uri method..

[SachidanandAlle]

to answer your first question.. who uses get_image.. this is when you want to upload the image via server to the client..

example.. you are using 3d slicer.. and wants to load an image (via next sample).. and the image is with monailabel server (via some datastore).. 3d slicer will call this method/api to download the image and open it it's viewer.. but again.. since this is 2 hop operation.. for now.. we just fallback on cached file.. so that it's useful in either case..

[bhoeckendorf]

Got it, thanks for your explanation. I did notice a lot of temporary file I/O, and this explains the design thoughts behind it. Are there any plans to improve upon this I/O pattern? In particular, I’d argue that all of this should be part of the datastore implementation. Or perhaps there should be a LocalCachingDatastore that wraps datastore implementations written without any caching concerns. The data store then just returns (and accepts) data dicts after LoadImage, or equivalent. In my opinion, this would really simplify a lot of MONAILabel’s internal code. No more I/O or file formats anywhere but in the datastore. It would also help handle use cases where reading the entire image isn’t feasible and we instead want to sample image patches. I’d be happy to open a dedicated issue for this conversation, or perhaps even a feature/improvement request. Let me know what you think.