eduzhai > Applied Sciences > Engineering >

Switching Loss for Generalized Nucleus Detection in Histopathology

  • king
  • (0) Download
  • 20210506
  • Save

... pages left unread,continue reading

Document pages: 24 pages

Abstract: The accuracy of deep learning methods for two foundational tasks in medicalimage analysis -- detection and segmentation -- can suffer from classimbalance. We propose a `switching loss function that adaptively shifts theemphasis between foreground and background classes. While the existing lossfunctions to address this problem were motivated by the classification task,the switching loss is based on Dice loss, which is better suited forsegmentation and detection. Furthermore, to get the most out the trainingsamples, we adapt the loss with each mini-batch, unlike previous proposals thatadapt once for the entire training set. A nucleus detector trained using theproposed loss function on a source dataset outperformed those trained usingcross-entropy, Dice, or focal losses. Remarkably, without retraining on targetdatasets, our pre-trained nucleus detector also outperformed existing nucleusdetectors that were trained on at least some of the images from the targetdatasets. To establish a broad utility of the proposed loss, we also confirmedthat it led to more accurate ventricle segmentation in MRI as compared to theother loss functions. Our GPU-enabled pre-trained nucleus detection software isalso ready to process whole slide images right out-of-the-box and is usablyfast.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×