Why photon counts matter: Every pixel value in a fluorescence image is built from individual photons arriving at the detector. If you detect 100 photons at a pixel, the noise is √100 = 10, giving you 10% measurement uncertainty. Detect only 25 photons and your image has roughly 5 distinguishable intensity levels — barely enough for reliable thresholding. This is Poisson noise, and it sets the fundamental floor for detection accuracy.
Otsu's method — the intuition: Imagine trying every possible intensity threshold from 0 to 255. At each threshold, you split all pixels into two groups — "background" and "nuclei." For each split, you measure how uniform each group is internally (its variance). Otsu picks the threshold that makes the two resulting groups internally most uniform — minimizing within-class variance, which is mathematically equivalent to maximizing the separation between the group means.
The Rose criterion: A nucleus is reliably detectable only when its contrast above background exceeds 5× the local noise level (SNR > 5). Below this threshold, what looks like a nucleus might be nothing more than a noise fluctuation. This is why dim nuclei near the detection threshold are the most error-prone — they hover near the visibility limit where signal and noise become indistinguishable.
Watershed as landscape: The marker-controlled watershed treats the intensity image as a topographic surface — pixel intensity equals elevation. Each seed point (local maximum inside a nucleus) sits at the bottom of a "catchment basin" (the intensity is inverted for this purpose). Water floods upward from each basin simultaneously. Where water from adjacent basins meets, a dam is built. These dams become the segmentation boundaries. The key mathematical property: watershed lines always form closed contours, guaranteeing that every pair of touching nuclei gets a complete separation boundary.
Dobrucki's warning: "Inexperienced microscopists often take the grainy structure of an area in the image for a real variation of the fluorescence signal. However, such features of the image may merely be a consequence of a very low number of photons." When evaluating detection quality, remember that apparent intensity variation in dim regions may be pure Poisson noise, not biological signal.
Every fluorescence image is built from individual photons. When few photons are collected, the image becomes grainy — not because the tissue varies, but because of the statistical scatter in counting small numbers of photons. A nucleus needs to be at least 5 times brighter than the noise to be reliably detected.
Otsu's method finds the best threshold by trying every possible cutoff and choosing the one that makes "nuclei" and "background" the most internally consistent groups. The watershed algorithm then separates touching nuclei by treating intensity as a landscape and finding the ridges between peaks.
