Normalization and Clipping

[lcaronson]

Hello,

Just hoping for some insight as to why pixel values were clipped from -79 to 304 in the KiTS19 example. Does this relate to the window/level operation on images? I.e. if I am using images with a different contrast distribution, should I consider changing this range to be more appropriate?

Thanks a lot!

Luke

[muellerdo]

Heyho @lcaronson,

Just hoping for some insight as to why pixel values were clipped from -79 to 304 in the KiTS19 example.

sure. In the KiTS19 challenge, we were working on Computer Tomography (CT) scans.
CT scans are using a fixed standardized value intensity scale named Hounsfield scale.

Check out this Wikipedia entry for more information:
https://en.wikipedia.org/wiki/Hounsfield_scale

Summarized: These HU values in a CT scan provide exact and detailed information about the tissue.
Due to we want to find kidneys and tumor of kidneys, we can use this information and clip only to values near a HU range of this desired tissue type. E.g. According to this wikipedia entry, kidneys are between +20 to +45. However, tumor tissue can be a bit more dynamic and we also need a bit of "edge-range" at the top and bottom of our clipping to differentiate between desired tissue and "other" tissue. But overall, you can clip together various uninteresting tissues and drastically reduce complexity for your model by focusing just on the important tissue.

if I am using images with a different contrast distribution, should I consider changing this range to be more appropriate?

Clipping is only recommended if you have specific value ranges to exploit. If you are working also on CT data, than I would highly recommend utilizing clipping.
Still, the exact range of clipping is more a rule-of-thumb in my experience. I'm no radiologist which will probably have strong opinions on my last sentence 😄
Since we are working with normal CT normally and not with something fancy like contrast enhanced CTs (CECT), I would recommend to just have a little bit wider clipping ranges to allow some contrast bias. Then, it should be fine.

The best thing I can advise is to just apply some different clipping ranges on it and have a look on your images if you can still easily detect your ROIs or not. ;)

Hope that this answer makes the clipping a bit more clearer.

Cheers,
Dominik