BakingTray autoThresh acquisitions yield threshold values totally different to those by autoROI

[raacampbell]

Image stacks processed by autoROI.runOnStackStruct produce tThresh values that are really different from those generated by doing a BT.bake in auto-ROI mode. Even more problematically, is that minimalBakeWithGUI is different from the bake. The thresholds from BakingTray are much higher and lead to loss of the tissue. I would expect some differences between BakingTray and autoROI, but what I see is huge and not acceptable:

Likely one cause is that "blank" pixels are making their way into autoROI when run from BakingTray. The easiest path is probably to get the minimal example code working correctly then to fix BT.bake itself.

[raacampbell]

The first issue I am seeing is that the bounding boxes are possibly too narrow in some cases. So it isn't just that the location is wrong. For example, I did minimalBake(1) with CA_514_A__514_C and I see a bounding box with a width of 435 pixels but the sample its bounding has imaged pixels with a width of 440 pixels.

I guess what is happening is that autoROI.boundingBoxToTiledBox is doing a slightly different thing to recipe.tilePattern. They therefore produce slightly different bounding box sizes.

[raacampbell]

dummyScanner was saving tiles that were too small. They were 44 pixels across not 48. It was saving the FOV minus the overlap. This is fixed now (f671881) and after doing this I get a ROI width of the 438 pixels for the stitched image. The preview is 439. Off by a pixel, could be a rounding error.

I also note that the bounding box size generated in autoROI is shorter by one overlap width. That explains the missing four pixels.

Correcting this doesn't get is exactly there. Our ROIs go to 440 pixels, so larger by 1 and the tThresh is still really high. Bake still stops early.

Maybe the easiest way to proceed right now it to figure out why in the first section it has such a high tThresh. Even after the preview scan it's high, even though no black pixels are being removed here. Let's work out this first.

Ah! The reason is that there are plenty of non-imaged pixels in what should be background. They aren't being removed and this is causing the problems.

[raacampbell]

Right... I think the way autoROI is choosing background pixels will have to change. The problem is that non-imaged pixels contaminate borders when median filtering happens. Then we use those borders to calculate the threshold. It is therefore really hard to get rid of them.

[raacampbell]

Substantial progress made on this: have to remove pixels labelled as background and don't interpolate when resampling. This clears up a lot of problems. A formal comparison between BakingTray and pure autoROI is not done yet.

[raacampbell]

Looked at two acquisitions and they are now pretty much the same between autoROI and BakingTray.