Near-0 real signal gets convolved with near-0 background because of the background suppression --> uint8 conversion. We can tell true positives often times due to the non-random conformation of structures in the image upon inspection.
We have a problem of some channels with very low background , and this foreground segmentation approach to suppress low values has not been working well. What happens is a bad background estimate leads to very low values coming "up" and getting set at 1 in the uint8 image. These then contribute to the cell averages that are based on these uint8 images. One scenario where this comes up is when a slide is almost all-zero for a particular channel... or in the case of very rare cells that are not captured in background estimation. This leads to an actual reversing of the stain representation in near-all-0 slides to be near-all-1 instead. We could detect this post-hoc maybe, but added complexity ought to be avoided.
The desired behavior is a perfect segmentation of background signal. We think that the uint16 has enough range to correctly represent real signal albeit low valued, versus near-0 level background. So take average values directly from the uint16 image.
nathanin
commented
3 years ago
Near-0 real signal gets convolved with near-0 background because of the background suppression -->
uint8conversion. We can tell true positives often times due to the non-random conformation of structures in the image upon inspection.We have a problem of some channels with very low background , and this foreground segmentation approach to suppress low values has not been working well. What happens is a bad background estimate leads to very low values coming "up" and getting set at
1in theuint8image. These then contribute to the cell averages that are based on theseuint8images. One scenario where this comes up is when a slide is almost all-zero for a particular channel... or in the case of very rare cells that are not captured in background estimation. This leads to an actual reversing of the stain representation in near-all-0 slides to be near-all-1 instead. We could detect this post-hoc maybe, but added complexity ought to be avoided.The desired behavior is a perfect segmentation of background signal. We think that the uint16 has enough range to correctly represent real signal albeit low valued, versus near-0 level background. So take average values directly from the
uint16image.