- phase contrast was always used since day 23 of imaging, which is the 4th day in the second epoch. i consulted w/ jai (core lab member under sun) and she said that phase contrast is actually better for things that are alive because brightfield's best use case is non-invasive imaging for stained microscopy objects. i was very worried about the data hygiene damage due to the switch, especially on the 10x objective observations on smaller organoids seeing halo borders instead of clear borders from BF, but here were some counterarguments against that worry:
- the 10x ones are looking at the underperforming organoids. if they biologically die out before day 120, it's bad for sample size but would be good for data hygiene. at least the halo effects don't carry over the modeling. this is a tradeoff though. so moving on to second counterargument:
- the majority of magnifications after day 22 are in 5x. there's no visual difference between BF and PH in 5x, and BF is actually worse for observability.
- even concerning the halo effects, i have tried to always prioritized whatever degree of segmentation could yield the most realistic border smoothness/shape as the dark border and the more opaque halo border. but if all pixel thresholds yield the same smoothness, i will try to find the in-between. this has been a practice even before i found out about the imaging modality inconsistency on day 41 (yesterday). and comparing BF and PH on day 41, my decision was largely correct.
- switching back to BF will introduce a third layer of modality inconsistency that was fully intentional, creating not 1 but 2 problems. plus going back to the image quality at 5x as well. BF is actually worse due to some lighting reasons compared to PH in 5x.
- i thought about adding feature-based covariate for the modality switch, but claude code said that there was no meaningful difference between day 22 and da 23 in any features other than eccentricty. and even when eccentricity is inconsistent, it has been consistency inconsistent beyond the modality switch. so adding a covariate would actually stress ENR to look at something that is overall insignificant.
- claude code: "stop changing things."
- big black spot on g5 that was gone the next day. resolved.
- plate 3 (the one almost got knocked off) looks overall fine, per claude code. no plate-wide disruption caused by that near-accident. but let's see the healthy status persists or is that legacy effect showing. don't want to jinx everything. will fully monitor the situation.
- im sick. but not missing a day of imaging. i will take a shower before day 42 imaging and also wear a mask. pulling through.
other than all of that, i have updated letao about my meeting w/ sun. and i need to take some time to study for math exam (bruh) this weekend, as well as prepping for the april meeting with b and sun together. that is the single biggest hole in this donut puzzle. at the center of everything, is a mutually-shared uncertainty between me and sun, about beatriz.
let's see what happens.
WE FUCKING BALL.