Computer Vision News - August 2020

The inputs to the model are digitized microscopic slides. For a whole-slide image, it answers a yes/no question: is there cancer present? To do this, it makes a segmentation mask and estimates uncertainty pixelwise, resulting in an uncertainty heatmap. The heatmap identifies regions in the whole-slide image where there is high uncertainty, and therefore, out-of- distribution data. This provides highly detailed pixel-level information to the clinician. There are two popular approaches to tackle the problem of uncertainty estimation: Monte Carlo dropout (MC-dropout) and deep ensembles . However, Jasper argues that a better approach is to use multi-head CNNs (M-heads). He explains: “M-heads are similar to the well- known deep ensembles algorithm, but more computationally efficient through extensive parameter sharing. The model shares its parameters in the first few layers and then the final layers are randomly initialized multiple times, similar to what you would do in a deep ensemble. For the multi-head approach, you train them all at the same time using a specialized objective function.” Jasper is excited by the novel aspects of this research. “People need to see this work!” he exclaims. “They need to see how effective uncertainty estimation techniques using a multi- head model are, and the value that this can bring to the field of digital pathology, especially in terms of out- of-distribution detection.” Jasper Linmans 35 Best of MIDL 2020