Figure 2.

The impact of algorithm parameters and cutoff thresholds on images. QuPath's ‘positive cell detection’ command is used to determine Ki67 labelling indices for the same field of view, acquired using two different scanners. This conventional image‐processing algorithm uses multiple adjustable parameters, although here only the thresholds for nucleus detection and DAB positivity are varied. Horizontally adjacent images are from the same scanner, while vertically adjacent images are generated using the same thresholds. Detected nuclei are shown as red or blue, depending upon whether they are classified as positive or negative, respectively. Changing either threshold or the scanner can substantially change the results, although typically in predictable ways (e.g. a high detection threshold leads to negative nuclei being missed, and the labelling index is inflated; a high DAB threshold leads to positive nuclei being misclassified as negative, and the labelling index is reduced). Combining this knowledge with a careful evaluation of the markup images, it is possible for a user to identify and address many errors by adjusting algorithm parameters accordingly.