Anatomy of Domain Shift Impact on U-Net Layers in MRI Segmentation

Ivan Zakazov, Boris Shirokikh, Alexey Chernyavskiy, Mikhail Belyaev

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

1 Citation (Scopus)

Abstract

Domain Adaptation (DA) methods are widely used in medical image segmentation tasks to tackle the problem of differently distributed train (source) and test (target) data. We consider the supervised DA task with a limited number of annotated samples from the target domain. It corresponds to one of the most relevant clinical setups: building a sufficiently accurate model on the minimum possible amount of annotated data. Existing methods mostly fine-tune specific layers of the pretrained Convolutional Neural Network (CNN). However, there is no consensus on which layers are better to fine-tune, e.g. the first layers for images with low-level domain shift or the deeper layers for images with high-level domain shift. To this end, we propose SpotTUnet – a CNN architecture that automatically chooses the layers which should be optimally fine-tuned. More specifically, on the target domain, our method additionally learns the policy that indicates whether a specific layer should be fine-tuned or reused from the pretrained network. We show that our method performs at the same level as the best of the non-flexible fine-tuning methods even under the extreme scarcity of annotated data. Secondly, we show that SpotTUnet policy provides a layer-wise visualization of the domain shift impact on the network, which could be further used to develop robust domain generalization methods. In order to extensively evaluate SpotTUnet performance, we use a publicly available dataset of brain MR images (CC359), characterized by explicit domain shift. We release a reproducible experimental pipeline (https://github.com/neuro-ml/domain_shift_anatomy ).

Original languageEnglish
Title of host publicationMedical Image Computing and Computer Assisted Intervention – MICCAI 2021 - 24th International Conference, Proceedings
EditorsMarleen de Bruijne, Philippe C. Cattin, Stéphane Cotin, Nicolas Padoy, Stefanie Speidel, Yefeng Zheng, Caroline Essert
PublisherSpringer Science and Business Media Deutschland GmbH
Pages211-220
Number of pages10
ISBN (Print)9783030871987
DOIs
Publication statusPublished - 2021
Event24th International Conference on Medical Image Computing and Computer Assisted Intervention, MICCAI 2021 - Virtual, Online
Duration: 27 Sep 20211 Oct 2021

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume12903 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference24th International Conference on Medical Image Computing and Computer Assisted Intervention, MICCAI 2021
CityVirtual, Online
Period27/09/211/10/21

Keywords

  • Deep learning
  • Domain adaptation
  • MRI
  • Segmentation

Fingerprint

Dive into the research topics of 'Anatomy of Domain Shift Impact on U-Net Layers in MRI Segmentation'. Together they form a unique fingerprint.

Cite this