eduzhai > Applied Sciences > Engineering >

HATNet An End-to-End Holistic Attention Network for Diagnosis of Breast Biopsy Images

  • king
  • (0) Download
  • 20210505
  • Save

... pages left unread,continue reading

Document pages: 10 pages

Abstract: Training end-to-end networks for classifying gigapixel size histopathologicalimages is computationally intractable. Most approaches are patch-based andfirst learn local representations (patch-wise) before combining these localrepresentations to produce image-level decisions. However, dividing largetissue structures into patches limits the context available to these networks,which may reduce their ability to learn representations from clinicallyrelevant structures. In this paper, we introduce a novel attention-basednetwork, the Holistic ATtention Network (HATNet) to classify breast biopsyimages. We streamline the histopathological image classification pipeline andshow how to learn representations from gigapixel size images end-to-end. HATNetextends the bag-of-words approach and uses self-attention to encode globalinformation, allowing it to learn representations from clinically relevanttissue structures without any explicit supervision. It outperforms the previousbest network Y-Net, which uses supervision in the form of tissue-levelsegmentation masks, by 8 . Importantly, our analysis reveals that HATNet learnsrepresentations from clinically relevant structures, and it matches theclassification accuracy of human pathologists for this challenging test set.Our source code is available at url{this https URL}

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×