DARCNN: Domain Adaptive Region-based Convolutional Neural Network forUnsupervised Instance Segmentation in Biomedical Images

Code

• github.com/joyhsu0504/darcnn
  Officialpytorch★ 4

Abstract

In the biomedical domain, there is an abundance ofdense, complex data where objects of interest may be chal-lenging to detect or constrained by limits of human knowl-edge. Labelled domain specific datasets for supervisedtasks are often expensive to obtain, and furthermore dis-covery of novel distinct objects may be desirable for un-biased scientific discovery. Therefore, we propose leverag-ing the wealth of annotations in benchmark computer visiondatasets to conduct unsupervised instance segmentation fordiverse biomedical datasets. The key obstacle is thus over-coming the large domain shift from common to biomedicalimages. We propose a Domain Adaptive Region-based Con-volutional Neural Network (DARCNN), that adapts knowl-edge of object definition from COCO, a large labelled visiondataset, to multiple biomedical datasets. We introduce a do-main separation module, a self-supervised representationconsistency loss, and an augmented pseudo-labelling stagewithin DARCNN to effectively perform domain adaptationacross such large domain shifts. We showcase DARCNN’sperformance for unsupervised instance segmentation on nu-merous biomedical datasets.

Tasks

Reproductions