From my understanding, in a tissue where nuclei are present and need to be detected, we need to predict bounding boxes (either rectangular/circular or in the shape of the nucleus, i.e. as in instance segmentation). However, a lot of research papers start with semantic segmentation. Again, what I understood is semantic segmentation won't give the location, bounding box or count of nuclei. It will just tell that some stuff is probably nuclei and rest is probably background.
So, what is the bridging that I am missing when trying to detect nuclei from semantic segmentation. I have personally done semantic segmentation but I can't seem to count/predict bounding boxes because I can't understand how to do that (for example if semantic segmentation gave a probable region for nuclei which is actually a mixture of 3 nuclei overlapping). Semantic segmentation (in the example) just stops right there.
- Thresholding algorithm like Watershed might not work in some cases as demonstrated in [Nuclei Detection][1] at 23:30 onwards.
- Edge detection between segmented nuclei and background would not separate overlapping nuclei.
- Finding local maxima and putting a dot there might give rise to false positives.
- Finding IoU but what if the output of segmentation is not a region of classification (1s and 0s) but a continuous probability map from values between 0 to 1.
- Isn't finding contours and getting bounding boxes from masks using opencv a parametric method? What I mean is, it being an image processing technique, there are chances it will work for some images and won't work for some.