How can I adjust hovernet to 0.5

[OmarAshkar]

Hello,

I am trying to use hovernet implementation in MONAI. I successfuly trained a model but it seems that some objects are being removed cause of 0.5 mpp resolution I have.

the original mpp of hovernet is 0.25, then there is hovernetplus with mpp 0.5. I believe it was a single line of code change:
https://github.com/TissueImageAnalytics/tiatoolbox/blob/aab2df9c8169ae87c433eedf42d9232f526c5339/tiatoolbox/models/architecture/hovernetplus.py#L293

I cannot find any equivalent for this postprocess in MONAI implementation of hovernet.

Any help is appreciated.

Thanks!

[KumoLiu]

Hi @OmarAshkar, from the code you shared, he proposed a way to calculate the relationship between mpp and the kernel size of sobel, the size of small object, which is currently not added in MONAI implementation.
https://github.com/TissueImageAnalytics/tiatoolbox/blob/aab2df9c8169ae87c433eedf42d9232f526c5339/tiatoolbox/models/architecture/hovernet.py#L535-L536

But you could achieve the same effect by specifying min_object_size and sobel_kernel_size these two parameters in HoVerNetInstanceMapPostProcessing directly.

MONAI/monai/apps/pathology/transforms/post/array.py

Line 652 in 1a018a7

class HoVerNetInstanceMapPostProcessing(Transform):

Hope it can help you, thanks!

[OmarAshkar]

@KumoLiu Super helpful! to reiterate, there is no post processing step in MONAI that would remove the small objects anyway, so I can use any mpp, right?

[KumoLiu]

Hi @OmarAshkar, any mpp is OK, in fact. I'm not sure what you mean about removing small objects. There are two places using remove small objects, IMO it doesn't affect what mpp you use.

MONAI/monai/apps/pathology/transforms/post/array.py

Lines 693 to 695 in 1a018a7

	self.generate_watershed_mask = GenerateWatershedMask(
	activation=activation, threshold=mask_threshold, min_object_size=min_object_size
	)

MONAI/monai/apps/pathology/transforms/post/array.py

Lines 698 to 703 in 1a018a7

	self.generate_watershed_markers = GenerateWatershedMarkers(
	threshold=marker_threshold,
	radius=marker_radius,
	postprocess_fn=marker_postprocess_fn,
	min_object_size=min_object_size,
	)

Hope it can help you, thanks!

[OmarAshkar]

@KumoLiu I followed your tutorial and my segmentation results does not include small cells. My resolution is 20X. So, I will play around the min_object_size and see if it enhances the result.

Many thanks!

[KumoLiu]

Hi @OmarAshkar, that make sense, maybe you could set min_object_size=-1, then it will not remove small objects.

MONAI/monai/apps/pathology/transforms/post/array.py

Line 319 in 1a018a7

self.remove_small_objects = RemoveSmallObjects(min_size=min_object_size) if min_object_size > 0 else None

Thanks!