Fig. 2.

Comparison of the NuCLS dataset with canonical ‘natural’ object detection datasets. Nucleus detection datasets typically contain objects that are much smaller and more densely packed than imaging datasets of natural or day-to-day scenery. NuCLS images are ∼380 × 380 pixel patches at 0.2 microns-per-pixel resolution, and contain on average 34 nuclei, each of which filling only ∼1% of the image area. These systemic differences motivate the adaptation of existing methods like Mask R-CNN to accommodate numerous small objects and to revisit some of the assumptions about object morphology that do not apply in the context of nucleus detection. Modified with permission from Lin et al. (2014)