Cutoffs

Cutoffs apply configurable thresholds to measurement values:

1. Select measurement — Choose any column in the measurement table: raw intensity, derived measurement, ratio, density value, shape descriptor.
2. Define thresholds — Set one or more cutoff values that divide the measurement range into categories.
3. Classify — Each cell (or region) is assigned to the category corresponding to its measurement value.
4. Output — A new classification column with the category labels (low/medium/high, 0/1+/2+/3+, responder/non-responder, etc.).

Thresholding theory (Gonzalez & Woods): All thresholding — whether Otsu's automatic method on pixel intensities or manual cutoffs on derived measurements — faces the same fundamental challenge: finding the decision boundary that minimizes classification error. The optimal boundary depends on the distributions of the two (or more) classes and their relative prevalences. Gonzalez & Woods show that when class distributions are Gaussian, the optimal threshold has a closed-form solution that depends on means, variances, and prior probabilities.

Cutoff optimization (Dilbilir): When the "correct" cutoff is unknown, it can be optimized from outcome data. For each candidate cutoff, compute the sensitivity (true positive rate) and specificity (true negative rate) for predicting an outcome (treatment response, survival). The optimal cutoff maximizes some criterion — typically the Youden index (sensitivity + specificity − 1), which balances both types of error. ROC curve analysis visualizes this tradeoff: the cutoff corresponding to the ROC point closest to (0,1) provides the best overall discrimination.

The Rose criterion as a biological cutoff: Pawley's Rose criterion (SNR > 5 for reliable detection) is itself a cutoff — a threshold on signal-to-noise ratio below which detection becomes unreliable. In measurement terms, a biomarker intensity that barely exceeds the background noise (SNR ≈ 2-3) produces unreliable gate results regardless of where the threshold is set. The Rose criterion provides a principled lower bound for meaningful gating: don't gate a marker whose positive population has SNR < 5 relative to the negative population.

Implementing PD-L1 Combined Positive Score (CPS) for gastric cancer:

1. Gate PD-L1 on all cells → PD-L1+ cells (tumor + immune)
2. Count: PD-L1+ cells (all types) and total tumor cells in the analysis region
3. CPS = (PD-L1+ cells / total tumor cells) × 100
4. Cutoff: CPS ≥ 1 (eligible for pembrolizumab), CPS ≥ 10 (higher response rate)

The cutoffs of 1 and 10 come from clinical trial data — they represent the thresholds where treatment benefit was demonstrated. Implementing them precisely in the analysis pipeline ensures the digital pathology result matches the clinical guideline.