eduzhai > Applied Sciences > Engineering >

Orientation-Disentangled Unsupervised Representation Learning for Computational Pathology

  • king
  • (0) Download
  • 20210507
  • Save

... pages left unread,continue reading

Document pages: 20 pages

Abstract: Unsupervised learning enables modeling complex images without the need forannotations. The representation learned by such models can facilitate anysubsequent analysis of large image datasets.However, some generative factors that cause irrelevant variations in imagescan potentially get entangled in such a learned representation causing the riskof negatively affecting any subsequent use. The orientation of imaged objects,for instance, is often arbitrary irrelevant, thus it can be desired to learn arepresentation in which the orientation information is disentangled from allother factors.Here, we propose to extend the Variational Auto-Encoder framework byleveraging the group structure of rotation-equivariant convolutional networksto learn orientation-wise disentangled generative factors of histopathologyimages. This way, we enforce a novel partitioning of the latent space, suchthat oriented and isotropic components get separated.We evaluated this structured representation on a dataset that consists oftissue regions for which nuclear pleomorphism and mitotic activity was assessedby expert pathologists. We show that the trained models efficiently disentanglethe inherent orientation information of single-cell images. In comparison toclassical approaches, the resulting aggregated representation ofsub-populations of cells produces higher performances in subsequent tasks.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×