Detecting and classifying cancers from image data using optimal transportation

Abstract

We describe a new approach to digital pathology that relies on measuring the optimal transportation (Kantorovich-Wasserstein) metric between pairs of nuclei obtained from histopathology images. We compare the approach to the standard feature space approach and show that our method performs at least as well if not better in automatically detecting and classifying different cancers of the liver and thyroid. 100% classification accuracy is obtained in 15 human test cases. In addition, we describe methods for using the geometric space framework to visualize and understand the differences in the data distribution that allow one to classify the data with high accuracy. © 2010 Springer-Verlag.