Some beginner's question about cellmap segmentation challenge.

[ieellee]

Hi @rhoadesScholar , I write a script that directly download groundtruth and corresponding raw images, and checked the image and groundtruth, but have a few questions:

1. Each data chunk, such as jrc_mus-kidney-3 and jrc_mus-liver-zon-1, has a different number of categories in the groundtruth folder. The all folder represents the aggregation of all categories, but for jrc_mus-kidney-3/crop472, I found that only category indices [0, 37] are present, and for the chlor category, only index [0] is available. Is this normal?

2. For all downloaded all groundtruth files, does each unique index represent the same category?

3. Is the directly downloaded data type semantic segmentation? How are the labels for instance segmentation obtained?

4. I found some abnormal crop labels, such as jcr_hela-3/crop60, where the annotated groundtruth is all 0, and the original image appears to be random noise.

Sincerely looking forward your early reply!😄

[yuriyzubov]

Hi @FrankieMAN-YU!
Great questions, thank you!

1. Yes, it is expected. In this particular case chlor group class and its atomic classes (chlor_lum, chlor_mem, chlor_sg) were labeled during the annotation process as "not present". It is still useful information to know which organelle is certainly missing from the groundtruth array. So If you see an empty zarr array for a specific class, that means that this specific organelle class is definitely not in the groundtruth array (as in there are no positive labels). Additionally, for some of the crops, data is included for classes that will not be evaluated in the challenge - chlor is one of these classes.
2. Each index in all corresponds to a unique class, as indicated in the whole list that maps organelle class to its numeric equivalent in the cellmap-segmentation-challenge docs.
3. It is the mix of semantic and instance segmentations.
   The classes that have instance segmentations are: ['nuc', 'vim','ves','endo','lyso','ld','perox','mito' ,'np','mt','cell']. These instances were obtained by running skimage.measure.label, with connectivity=1, on the binary groundtruth masks.
   All atomic classes (e.g. mito_lum, mito_mem etc.) are semantic segmentations. We constrained the number of group classes with instance segmentations based on the morphology of the organelle and how countable it is. For example, we didn't provide instance segmentations for er due to the complexity of its shape and the fact that oftentimes the joining areas between two instances of er could be outside of groundtruth ROI, making instance labeling incorrect.
4. Yes, the crop60 is fully located in extracellular space and the corresponding class (ecs) is labeled as 'present', whereas all other classes are labeled as 'not present'.

[ieellee]

I think I'm starting to understand. Taking jrc_mus-heart-1 as an example, crop452 has a series of groundtruths, such as bm, chlor, chlor_lum, chlor_mem, etc. For instance, if the bm is an all-zero array, does that mean that bm exists in crop452 but hasn't been manually annotated?

@yuriyzubov looking forward your early reply!

[yuriyzubov]

For instance, if the bm is an all-zero array, does that mean that bm exists in crop452 but hasn't been manually annotated?

@FrankieMAN-YU all-zero array means that bm class was manually annotated, but no instances of this class were found in crop472 ROI. (i.e. "not present").