Effect of training data set size, tiling method and prediction threshold on segmentation performance for Mask R-CNN models with prediction threshold 0.95, average over six different virtual slides.

Blue arrows indicate shift from fixed-stride to object-based tile positioning, green arrows indicate shift to object-based positioning + object integrity constraint. FS = fixed-stride, OBP = object-based positioning, OBP+OIC = object-based positioning + object integrity constraint; Pred. threshold = prediction threshold. Please note the different axis scalings, i.e., that substantial improvements in precision with the object-based tiling schemes are usually associated with relatively much smaller decreases in recall. (TIFF)