In February, our lab at the North Bay Imaging Lab replaced its entire imaging pipeline — twelve workstations running five different acquisition packages, all feeding into a single analysis stack that nobody fully understood. The migration took six weeks and surfaced 1,400 hours of archival data we had never examined frame by frame.
The frames we throw away
Every fluorescence protocol includes a step called background subtraction. You image a region with no sample, record the dark current and read noise, then subtract that from every frame. Standard practice since 1998. But when we ran the subtraction pipeline on our migration test data and diffed the outputs against the raw frames, the subtracted pixels were not random.
The problem is not that our images are noisy. The problem is that we have trained ourselves to see noise as failure.
Mira Vale's 2006 Field Notes on Confocal Signal dedicates an entire chapter to detector noise sources — read noise, dark current, photon shot noise, quantization error — each treated as an engineering defect to be minimized. Two decades later, our data suggests that at least one of those defects carries reproducible spatial structure across sessions.