Watershed

The Watershed engine implements the flooding simulation algorithm on a grayscale image interpreted as a topographic surface:

1. Initialization — Find all local minima in the image. Each minimum becomes the starting point of a catchment basin.
2. Flooding — Imagine water rising uniformly from below. At each water level, examine newly flooded pixels:
   • If a new pixel connects to exactly one existing basin, it joins that basin
   • If it connects to two or more basins, a dam (boundary) is built to prevent the basins from merging
   • If it doesn't connect to any existing basin, it starts a new one
3. Completion — When water reaches the maximum intensity, all pixels have been assigned to a basin or a boundary. The boundaries form the segmentation result.

In marker-controlled mode, internal markers (seed points) replace the natural local minima. The image is modified so that only marker locations are minima — eliminating the hundreds of spurious minima that cause over-segmentation. In StrataQuest's Nuclei Detection, local intensity maxima serve as markers, and the algorithm operates on the inverted image so that bright nuclei become basins.

The over-segmentation problem: A raw gradient image of tissue might have hundreds of local minima — every small dip in intensity creates a catchment basin. Solomon & Breckon describe this as "noise and small-scale structures result in many local, small catchment basins (broadly analogous to puddles on the landscape)." These "puddles" produce hundreds of tiny regions where you expected only a few objects.

The marker-controlled solution: Instead of using every natural minimum, you pre-define seed points — one per expected object. The algorithm modifies the image so that only these seeds are minima, then floods as before. This is a five-step recipe: (1) compute gradient magnitude, (2) find foreground markers via morphological operations, (3) find background markers, (4) modify the gradient to have minima only at markers, (5) compute watershed on the modified gradient.

Mathematical guarantee: Watershed boundaries are always closed contours. This is a topological necessity — if a boundary were open, water from adjacent basins could flow through the gap and merge, contradicting the algorithm's construction. This guarantee is valuable because edge detection methods (Sobel, Canny) often produce broken contours that require additional linking steps.

The distance transform connection: When the input is a distance transform rather than an intensity image, watershed performs shape-based segmentation. Each object's skeleton becomes a ridge in the distance transform, and the watershed boundaries follow these ridges. This is how StrataQuest separates touching objects that have similar intensities — by segmenting based on shape rather than brightness.

Computational note: The flooding algorithm processes each pixel exactly once (each pixel is examined when the water level reaches its intensity). For an image with L intensity levels and N pixels, the time complexity is O(N log L) — efficient enough for large tissue images.

Separating densely packed tumor cells in a Ki-67-stained section where nuclei frequently touch:

1. Nuclei Detection creates initial seeds at local intensity maxima within each nucleus
2. The DAPI intensity image is inverted (bright nuclei become valleys)
3. Marker-controlled Watershed floods from each seed, building boundaries where adjacent nuclear regions meet
4. Each resulting basin becomes a separate nucleus in the coded image

Result: Even in regions where 10+ nuclei form a connected cluster, each individual nucleus receives its own boundary and label. Without watershed, the entire cluster would be detected as a single massive object, yielding one measurement instead of ten.