System Overview.

Slides were scanned using Hamamatsu’s NanoZoomer microscope. During the first pass, ARLO extracted random image windows that human experts tagged as “pollen” or “not pollen”. Using this information, a segmentation or “Pollen Spotting” model was built to distinguish pollen pixels from the background of an image. During the second pass, ARLO applied this model to a new set of random windows. The pool of samples from which these windows were drawn varied in Experiments 1 and 2. Random windows with enough pollen content were then used to build a pollen species classification model, which was used to predict species ratios in the testing slides in Experiment 1 and to produce species classifications in Experiment 2. See text for details.