Overlapping nuclear boundaries

[rictoo]

Hi! Thanks for the amazing work. I believe there are overlapping nuclear boundaries in the CSV file, with overlapping regions seemingly assumed to represent a unique instance. Is this something you have noticed? Am I doing something wrong?

[cooperlab]

Hello @rictoo. We need an example if you want us to look into this. Is this a small overlap of several pixels or something larger?

[rictoo]

Thanks @cooperlab for the quick response.

Here's an example from TCGA-AC-A2BK-DX1_xmin15426_ymin8344_MPP-0.2500_xmin-2048_ymin-1024_xmax-3072_ymax-2048 subset to the top-left of the image (20% width and 20% height of whole mask/image).

Note the fragment to the right of the image (arrow for clarity).

This is the row in the csv/ directory corresponding to the fragment:
609,[NA],CancerEpithelium,polyline,166.0,172.0,104.0,109.0,"166,167,167,169,171,172,172,171,170","104,105,109,107,107,106,105,105,104"

This is the rgb/ image corresponding to that region:

When visualizing a mask chosen at random, I note quite a few of these "fragments" (this one's TCGA-A2-A0CM-DX1_xmin18562_ymin56852_MPP-0.2500_xmin-2048_ymin-2048_xmax-3072_ymax-3072 from the mask/ directory):

[cooperlab]

Thank you for the examples. I will look into this with @szolgyen

@szolgyen let's look at TCGA-A2-A0CM-DX1 and see if these fragments are present in the output of the segmentation model, or if they arise during post-processing.