Counting Statistics, Error Propagation & ROC

Quick Answer

Radioactive decay is a Poisson process, so the standard deviation of a count N is √N and the fractional (percent) error is 1/√N. Achieving 1% precision requires 10,000 counts; 5% requires 400. Errors propagate — variances add for sums/differences, fractional variances add for products/quotients — which governs net-count (signal-minus-background) uncertainty. The chi-square test checks system reproducibility, and ROC/AUC analysis compares diagnostic performance independent of any single threshold.

Poisson Statistics — the Foundation

For a Poisson-distributed count N:

• Standard deviation: σ = √N
• Fractional (relative) error: σ/N = 1/√N; percent error = 100/√N
• 95% confidence ≈ N ± 1.96√N

The practical corollary: precision improves only as the square root of counts — a 10× improvement in precision needs 100× the counts (or time). This is why low-count studies (delayed images, small ROIs, short frames) are inherently noisier.

Error Propagation

• Sum or difference (A ± B): add variances → σ = √(σ²_A + σ²_B) = √(A + B) for counts.
• Product or quotient (A×B or A/B): add fractional variances (relative errors in quadrature) → (σ/R)² = (σ_A/A)² + (σ_B/B)².
• Multiply by a constant k: σ scales by k; fractional error unchanged.

Worked: Net Counts (Signal − Background)

Net = Gross − Background. Because both are counts, the variance of the net = Gross + Background (not Gross − Background):

• Gross = 1,000; Background = 200 → Net = 800
• σ_net = √(1000 + 200) = √1200 ≈ 34.6
• Fractional error = 34.6/800 = 4.3%

Note the net count (800) has more relative uncertainty than the gross count alone — background subtraction always increases relative error, the reason high background (or low target-to-background) degrades quantitation.

The Chi-Square Test (Reproducibility)

Repeated measurements of a constant source should vary only by Poisson statistics. χ² = Σ(Nᵢ − N̄)² / N̄ over n measurements; the value is compared to tabulated ranges for (n−1) degrees of freedom. An out-of-range χ² indicates the system is varying more (instability) or less (something suppressing normal variation) than Poisson statistics allow — a QC red flag for camera/counter reproducibility.

Dead Time & Count-Rate Losses

At high count rates, detectors miss events during the processing ("dead") time:

• Nonparalyzable model: measured rate saturates toward a maximum.
• Paralyzable model: measured rate rises, peaks, then falls at very high rates (each event extends the dead period) — dangerous because a very high true rate can read as a low one.
• Relevant to first-pass, Rb-82 cardiac, and high-activity therapy imaging; corrected by system dead-time models.

Minimum Detectable Activity

The smallest signal distinguishable from background depends on background counts and counting time (e.g. Currie's L_D). Practically: lower background and longer counting lower the detection limit — the basis for wipe-test and contamination-survey sensitivity.

Diagnostic Performance — Beyond a Single Threshold

PPV/NPV depend on prevalence (Bayes): the same test performs differently in a screening vs a high-prevalence population — a normal MPI's high NPV (the "warranty period") reflects low pretest risk as much as test performance.

ROC / AUC — Threshold-Independent Comparison

A receiver-operating-characteristic (ROC) curve plots sensitivity (TPF) vs 1 − specificity (FPF) across all thresholds. The area under the curve (AUC) summarizes discrimination:

• AUC 1.0 = perfect; 0.5 = chance (diagonal).
• ROC compares observers, systems, or tracers independent of any single cutoff — the standard method for imaging-performance studies (and for justifying a chosen operating threshold).

Board-Level Synthesis

Decay is Poisson: σ = √N, percent error = 1/√N — so 10,000 counts = 1%, 400 = 5%, and precision improves only as the square root of counts (10× precision needs 100× counts). Net-count variance = Gross + Background, so background subtraction always raises relative error.

Error propagation: add variances for sums/differences, add fractional variances for products/quotients. The chi-square test flags a system varying more (or less) than Poisson statistics allow — a reproducibility QC check.

PPV/NPV are prevalence-dependent (Bayes) — the same sensitivity/specificity gives different predictive values across populations. ROC/AUC compares diagnostic performance independent of threshold (AUC 1.0 perfect, 0.5 chance), the correct way to compare tracers/observers/systems. At high count rates, paralyzable dead time can make a very high true rate read falsely low.

Related Pages

• Physics: Physics of nuclear medicine, radiation detection & instrumentation (dead time, QC).
• Reference: Quality control procedures (chi-square/constancy in practice).

Figure / Diagram Suggestions

• A √N precision curve (counts vs % error) with the 400/10,000 landmarks.
• A net-count error worked schematic (gross, background, propagated σ).
• An ROC curve with AUC and the sensitivity/specificity trade-off.

Self-Check