Abnormality Detection in Histopathology via Density Estimation with Normalising Flows

Nick Pawlowski, Ben Glocker

Keywords: histopathology, out-of-distribution detection, normalising flows

TL;DR: We find lesions in histopathology images by only training on healthy regions.

Abstract: Diagnosis of cancer often relies on the time-consuming examination of histopathology slides by expert pathologists. Automation via supervised deep learning methods require large amounts of pixel-wise annotated data that is costly to acquire. Unsupervised density estimation methods that rely only on the availability of healthy examples could cut down the cost of annotation. We propose to use residual flows as density estimator and compare different tests for out-of-distribution (OOD) detection. Our results suggest that unsupervised OOD detection is a viable approach for detecting suspicious regions in histopathology slides.

Paper Type: both

Primary Subject Area: Application: Histopathology

Secondary Subject Area: Unsupervised Learning and Representation Learning

Paper Status: original work, not submitted yet

Source Code Url: We adapted https://github.com/rtqichen/residual-flows to add a data loader for PatchCamelyon.

Data Set Url: https://github.com/basveeling/pcam

Registration: I acknowledge that publication of this at MIDL and in the proceedings requires at least one of the authors to register and present the work during the conference.

Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.