eduzhai > Applied Sciences > Engineering >

First U-Net Layers Contain More Domain Specific Information Than The Last Ones

  • king
  • (0) Download
  • 20210506
  • Save

... pages left unread,continue reading

Document pages: 14 pages

Abstract: MRI scans appearance significantly depends on scanning protocols and,consequently, the data-collection institution. These variations betweenclinical sites result in dramatic drops of CNN segmentation quality on unseendomains. Many of the recently proposed MRI domain adaptation methods operatewith the last CNN layers to suppress domain shift. At the same time, the coremanifestation of MRI variability is a considerable diversity of imageintensities. We hypothesize that these differences can be eliminated bymodifying the first layers rather than the last ones. To validate this simpleidea, we conducted a set of experiments with brain MRI scans from six domains.Our results demonstrate that 1) domain-shift may deteriorate the quality evenfor a simple brain extraction segmentation task (surface Dice Score drops from0.85-0.89 even to 0.09); 2) fine-tuning of the first layers significantlyoutperforms fine-tuning of the last layers in almost all supervised domainadaptation setups. Moreover, fine-tuning of the first layers is a betterstrategy than fine-tuning of the whole network, if the amount of annotated datafrom the new domain is strictly limited.

Please select stars to rate!


0 comments Sign in to leave a comment.

    Data loading, please wait...