Improve comparison plot and diagnostic plots for autoROI

[raacampbell]

We need to improve the comparison plot (evaluate.compareResults) and the diagnostic plots (evaluate.plotResults.m). In general, consider that the comparison function will need to be a sort of unit test to evaluate whether changes to the code have broken anything. It therefore needs to be very clear whether things have got worse. The plotResults needs to highlight very clearly any problems there might be.

Compare

• Large increases or decreases in coverage area
• Add option to exclude indices. 556aaea3e342f3bd80d7f72a184e8386d40a2672

Plot

• Make clear brains with missing sections; but also add a field to the pStack to indicate the minimum section we expect to pass. In some acquisitions the sample legitimately vanishes and these shouldn't trigger a warning. Better: calculate the number of missed sq mm differently. We do it section by section now and so missing sections don't appear. If we do the total imaged sq mm divided by all available tissue, then we won't ever run into problems. fdd6e4cac28a487486c3abbc54e1e46af3eb4067
• Highlight cases where the ROI expands beyond the FOV 32f7f055306f105f37c8cfc466b7af19b7edae9b
• Highlight cases where the ROI changes size suddenly
• Difference between the ground truth area and the area above threshold. This could bring out problems where lots of background is labelled as tissue.
• Consider SNR vs missing tissue plot.

Both

• Highlight problemCases so we know how they're doing. These are files that exhibit some awkward issue and are grouped in their own directory for this reason.

[raacampbell]

fdd6e4cac28a487486c3abbc54e1e46af3eb4067 makes changes to evaluateBoundingBoxes so that we count stats over all planes in the image stack. This means that cases where the ROI finding finish early do indeed end up generating lots of missed pixels. It also means there is no penalty for the few cases where the brain itself is just absent in the final few sections.

[raacampbell]

Removing small areas as above unfortunately messes up other recordings:

127/129. log_HMV_NN01_previewStack.mat -- 165 unprocessed sections. tThresh SD=2.72
129/129. log_sample_191120_164248_previewStack.mat -- 239 unprocessed sections. tThresh SD=10.34

BW=bwareaopen(BW, 20); or BW=bwareaopen(BW, 50); seem safe, but more is a bad idea.