Keywords: Digital pathology, representation learning, classification, similar image search
TL;DR: Creating small but meaningful representations of whole slide digital pathology images by aggregating tiles in a simple but efficient way.
Abstract: Representation learning is a popular application of deep learning where an object (e.g.,
an image) is converted into a lower-dimensional representation that still encodes relevant
features of the original object. In digital pathology, however, this is more difficult because
whole slide images (WSIs) are tiled before processing because they are too large to process
at once. As a result, one WSI can be represented by thousands of representations - one for
each tile. Common strategies to aggregate the “tile-level representations” to a “slide-level
representation” rely on pooling operators or even attention networks, which all find some
weighted average of the tile-level representations.
In this work, we propose a novel approach to aggregate tile-level representations into
a single slide-level representation. Our method is based on clustering representations from
individual tiles that originate from a large pool of WSIs. Each cluster can be seen as encoding
a specific feature that might occur in a tile. Then, the final slide-level representation is a
function of the proportional cluster membership of all tiles from one WSI. We demonstrate
that we can represent WSIs in parsimonious representations and that these aggregated
slide-level representations allow for both WSI classification and, reversely, similar image
search.
3 Replies
Loading