Data format when having thousands of targets per image

[FiReTiTi]

Hi,
I would like to use the Mask R-CNN version of MMdet, so I looked at how to transform my data to the coco format. It seems that the 'segmentation' contains the contour of an object.
As I work on biomedical images, I have thousands (between 5K and up to 40K) objects (from the same category/class) with non-smooth contours per training image. I'm not sure that this data format would be appropriate for my problem.
Is there a more suited format I should use?
Thanks in advance