Image thresholds bias significance of tissue biomarker results.

<p><b>a</b>. Thresholds for CD99 and Ki67 were split into centiles of the log₂ (nuclear/cytoplasmic ratio) distribution. The heat map shows the Cox regression <i>p</i> value (log₁₀) for each pair of centiles with respect to survival outcome of the whole imaged cohort. The square marker shows the optimal threshold with respect to the <i>p</i> value (CD99; −0.18 [62%tile], Ki67; 0.31 [22%tile]. <b>b</b>. Using this pair of thresholds, the Ki67 index and the log-rank test is calculated for all possible dichotomised splits of patient groups. <b>c</b>. Optimal threshold values defined in <b>a</b>. and <b>b</b>. lead to the Kaplan-Meier plot of overall survival outcome for good (low Ki67 index, blue) and poor prognosis (high Ki67 index, red) cases (dotted lines; 95% confidence intervals). Cox regression (b = 1.6, z = 4.8, p = 1.6×10⁻⁶, log-rank p = 2.6×10⁻⁷). <b>d</b>. A single observer (CB) scored the same images for Ki67 labelling by eye based on a binary low (blue) or high (red) score, resulting in the Kaplan Meier plot (Cox regression = 1.3, z = 3.7, p = 2.5×10⁻⁴, log-rank p = 1×10⁻⁴).