Morphological filtering: Remove Objects implements what Gonzalez & Woods call morphological filtering — using object properties to separate real structures from noise. This is conceptually related to morphological opening (erosion followed by dilation), which removes structures smaller than the structuring element. But Remove Objects is more flexible: it can filter on any measurable property, not just size.
The role of shape descriptors: Simple area thresholds can't distinguish a 200-pixel nucleus from a 200-pixel staining streak. Shape descriptors add discriminating power. Compactness (4π × Area / Perimeter²) equals 1 for a perfect circle and approaches 0 for extremely elongated shapes. Most nuclei have compactness between 0.6 and 0.95; most linear artifacts fall below 0.3. Adding a compactness filter catches artifacts that pass area and intensity filters.
Order of operations matters: Remove Objects should typically run before measurements and classification, not after. False detections that survive into the measurement stage introduce outliers that can skew population statistics (mean intensity, coefficient of variation) and confuse classification algorithms. In statistical terms, removing obvious non-cells before analysis is analogous to removing obvious outliers before computing summary statistics — it improves the accuracy of all downstream analysis.
The not-too-aggressive principle: Overzealous removal is as harmful as insufficient removal. If your area filter removes objects below 80 pixels, you might eliminate genuine small lymphocytes (which can be 50-70 pixels in some images). If your intensity filter removes objects below a threshold, you might eliminate real cells in dimly stained regions. The goal is to remove clearly artificial objects while preserving the full biological range of real cells.
Remove Objects is quality control — filtering detected objects by measurable properties to separate real cells from artifacts. Shape descriptors like compactness help distinguish round nuclei from elongated staining streaks. The key is finding the right balance: remove obvious false positives without accidentally eliminating real but unusual cells.
