(a) Upper triangle: plotting the overlap fraction for each pair of cells – the intersection over union of their bit masks, shows that the majority of overlaps occur for cells of different sizes. Almost all overlaps have the size of cell 2 greater than the size of cell 1 and lie off the diagonal. Lower triangle: With a single fuzzy size threshold, cells in the small category have sizes less than the upper threshold (; dashed line), and cells in the larger category have sizes greater than the lower threshold (; dotted line). We show the overlap fraction by the size of the dot. Within each category (green and blue dots), small overlap fractions dominate; between the two categories (red dots), large overlaps dominate. By converting the bit masks into two binary images, one for each size category, rather than a single binary image, we therefore eliminate most of the substantial overlaps. (b) The distribution of all mask areas in the same training data for comparison. We indicate the size thresholds as in a. (c) The distributions of overlap fractions for mask pairs grouped using the fuzzy size threshold of a. We omit pairs that do not overlap for clarity. (d) Applying morphological erosions of the cell masks reduces the number of overlapping cell pairs, but generates smaller masks. We judge masks with areas below 10 pixels squared to be too small. (e) The numbers of overlapping cell pairs remaining from the training, validation, and test sets either before, denoted None, or after splitting into size categories and applying an optimised number of erosions.