Source-linked AI summary
Unsupervised domain adaptation in brain lesion segmentation with adversarial networks
Konstantinos Kamnitsas, Christian Baumgartner, Christian Ledig, Virginia F. J. Newcombe, Joanna P. Simpson, Andrew D. Kane, David K. Menon, Aditya Nori, Antonio Criminisi, Daniel Rueckert, Ben Glocker
TL;DR
Biomedical segmentation systems can lose accuracy when deployed on data from different scanners or imaging protocols, while annotating every new target domain is impractical. This paper uses adversarial domain adaptation with a multi-connected discriminator to learn domain-invariant features without target labels, achieving performance close to supervised domain adaptation.
Problem
Segmentation performance can degrade across biomedical imaging domains, but unsupervised adaptation lacks target labels needed to learn the unknown target function.
Method
The paper trains a 3D segmentation CNN adversarially against a domain discriminator operating on activations at multiple layers to learn domain-invariant features.
Results
Without target labels, the method closes much of the performance gap with supervised target-label adaptation on traumatic brain injury segmentation.
Takeaways & Limitations
Multi-connected adversarial networks provide an unsupervised segmentation adaptation approach for data acquired with different scanners and MR protocols.
Takeaways & Limitations
The unsupervised setting assumes labeled source data and an unlabeled target database from a different but related domain.
Abstract
from arXiv · showhide
Significant advances have been made towards building accurate automatic segmentation systems for a variety of biomedical applications using machine learning. However, the performance of these systems often degrades when they are applied on new data that differ from the training data, for example, due to variations in imaging protocols. Manually annotating new data for each test domain is not a feasible solution. In this work we investigate unsupervised domain adaptation using adversarial neural networks to train a segmentation method which is more invariant to differences in the input data, and which does not require any annotations on the test domain. Specifically, we learn domain-invariant features by learning to counter an adversarial network, which attempts to classify the domain of the input data by observing the activations of the segmentation network. Furthermore, we propose a multi-connected domain discriminator for improved adversarial training. Our system is evaluated using two MR databases of subjects with traumatic brain injuries, acquired using different scanners and imaging protocols. Using our unsupervised approach, we obtain segmentation accuracies which are close to the upper bound of supervised domain adaptation.
1 Introduction
The paper addresses unsupervised domain adaptation for segmentation when biomedical training and test data differ, making source-only models unreliable and new target annotations impractical. It proposes adversarially learning domain-invariant representations without target labels.
- Motivation: Different scanners, protocols, contrasts, resolutions, noise levels, and sequence types can make source-trained models underperform on target biomedical data.The paper notes that this distribution shift affects even recent representation-learning models.
- Problem setting: Domain adaptation assumes shared label spaces while learning a function that performs across differing source and target domains.Unsupervised domain adaptation uses labeled source data and unlabeled target data.
- Problem setting: Because target labels are unavailable in unsupervised adaptation, the method learns representations invariant to domain differences using source labels and unlabeled target samples.The adapted representation and predictor are trained from {XS, YS, XT}.
- Contribution: The paper develops adversarial training between a segmentation network and domain classifier to remove domain-specific factors from the learned representation.This work applies domain-adversarial networks to biomedical image segmentation.
- Contribution: Multi-connected adversarial networks adapt layers at various depths and improve domain adaptation by supporting adversarial training throughout the segmentation network.The evaluation targets traumatic brain injury segmentation across databases acquired with different scanners and MR sequences.
2 Unsupervised domain adaptation with adversarial nets
The method combines a 3D CNN segmenter with a domain discriminator that processes intermediate activations and adversarially encourages domain-invariant features. Connecting the discriminator at multiple depths addresses segmentation’s need for both fine and high-level features.
- Segmentation system: The segmenter is a fully convolutional CNN that predicts labels for multiple voxels from an input image or sub-segment.Its parameters are learned by minimizing a segmentation loss with stochastic gradient descent.
- Domain discriminator: The domain discriminator classifies whether segmenter activations come from source or target distributions, using classification accuracy to indicate source-specific representation.It is a second 3D CNN operating on a selected hidden representation.
- Training considerations: Segmentation and discriminator batches require care because source data may be class-imbalance sampled using masks unavailable for unlabeled target data.The discriminator should avoid comparing differently sampled source and target examples.
- Adversarial training: The discriminator is trained to minimize domain-classification loss, while the segmenter minimizes segmentation loss and maximizes domain-classification loss.The adversarial weight α controls the relative importance of domain adaptation for the segmenter.
- Multi-connected adversarial networks: Multi-connected networks attach the discriminator at multiple segmenter layers to adapt early fine patterns while preventing source-specific patterns from being recovered at deeper layers.The design also gives the discriminator varied features, improving adversarial-gradient quality.
3 Experiments
Experiments evaluate transfer from a GE-equipped source database to an SWI-equipped target database, comparing supervised, non-adapted, and unsupervised adaptation settings. The proposed method improves target-domain lesion segmentation without target labels and analyzes layer depth and adaptation strength.
- 3.1 Material: The study uses two TBI databases acquired with different scanners and MR protocols, treating GE and SWI as interchangeable input channels.The source database contains GE, whereas the target database contains SWI; other sequences also differ between databases.
- 3.3 Evaluation: Training on source data and directly replacing GE with SWI at test time totally fails, while using only common sequences provides a practical lower bound.The common-sequence experiment excludes GE and SWI from both training and testing.
- 3.3 Evaluation: Using separate GE and SWI channels performs better than supervised training on T alone but transfers less knowledge than placing the sequences in the same channel.The separate-channel configuration uses six CNN channels and fills unavailable sequences with −4 after normalization.
- 3.3 Evaluation: The proposed unsupervised adaptation improves DSC by 3% over the non-adapted common-sequence baseline and covers 44% of the gap to the supervised upper bound.Target data are used for adaptation without manual labels, and GE and SWI share one input channel.
- 3.4 Analysis of system: Multiple discriminator connections improve adaptation, whereas shallow connections increase recall but significantly decrease precision.The multi-connected design closes to within approximately 1.5% DSC of the 66.5% practical UDA upper bound.
- 3.4 Analysis of system: Final target accuracy remains stable for αmax values from 0.05 to 1.0, while strongly countering the discriminator does not guarantee better performance.The schedule sets α to zero for the first 10 epochs, increases it linearly until epoch 35, and later refines the segmenter.
4 Conclusion
The paper presents unsupervised adversarial domain adaptation for biomedical image segmentation using two 3D neural networks. It proposes multi-connected adversarial networks, analyzes adaptation settings, and reports performance close to supervised baselines without target-domain labels.
- 4 Conclusion: The method adapts image-segmentation representations through adversarial training of two 3D neural networks.The segmenter and domain-classifier are trained adversarially to reduce domain-specific information in the learned representation.
- 4 Conclusion: Multi-connected adversarial networks improve adaptation by enabling higher-quality adversarial gradients throughout the adapted network.The paper also investigates adapted-layer depth and adaptation strength.
- 4 Conclusion: Without target-domain labels, the method performs close to supervised baselines.The conclusion frames the approach as relevant to multi-center studies with substantial domain differences.