SPECHT: Self-tuning Plausibility Based Object Detection Enables
Quantification of Conflict in Heterogeneous Multi-scale Microscopy
Abstract
We introduce a novel method that is able to localize fluorescent
labelled objects in multi-scale 2D microscopy, and is robust to highly
variable imaging conditions. Localized objects are then classified in a
novel way using belief theory, requiring only the image level label.
Each object is assigned a ‘belief’ that describes how likely it is to
appear in an image with a given set of labels. We apply our method
successfully to identify amyloid-beta deposits, associated with
Alzheimer’s disease, and to discover caveolae and their modular
components in superresolution microscpy. We illustrate how our approach
allows the fusion or combination of models learned across markedly
different datasets. We show how we can compute the ‘conflict’, or
disagreement between the models, an insight that can allow the domain
expert to interpret the composite model.