Reliability index of Morphological Chan-Vese

Hello, I am using Morphological Chan-Vese in order to describe an area. However, I need to find a way to tell if this has been done correctly.
For example: a more compete, well enclosed and dense output should be “good” (left image ) and a more dispersed one should be considered as bad output(right image).

This would be used as a reliability index for further decision making, if it should be trusted or not.
I am not uploading images from the source file that is being processed because they are surgical images.

Also, if you have any suggestion on finding a reliability measure for active contours please share!
Thanks in advance

Hi @Martin_Vasilkovski!

I’d suggest looking at skimage.measure.regionprops to see whether any metric in there can reliably predict quality in your examples. Things like circularity (perimeter over area), solidity (area over convex hull area), and euler number (number of holes in the region) might be good proxies for the fuzziness you are looking for. You could also come up with your own, such as binary-closing area over area.

Hope this helps!

3 Likes

To my knowledge there is no named metric specifically for this, but we can borrow the Jaccard metric and binary morphology to make one which should be relatively robust and usable for you:

Consider conducting a binary morphologal closing (dilation followed by erosion) with a moderately sized structuring element on your raw output from Chan-Vese. Then compute the Dice coefficient or, preferably, the Jaccard metric comparing the raw Chan-Vese result to the result after the closing operation.

Bad results as you see on the right will score poorly on Dice/Jaccard. Good results as on the left will score well. You most likely can set a threshold for this.

Results which score well should definitely be good. Results which score poorly MAY be bad, or may have closely spaced objects. This approach will only work robustly for relatively isolated objects as you see here. If you have multiple closely spaced objects (separation less than half the diameter of your circular structuring element), the above approach will connect them in the dilation step and the score will be incorrectly low.

1 Like

Thank you so much for this, it might be what I was looking for!