Test-time unsupervised domain adaptation

TY - CHAP

T1 - Test-time unsupervised domain adaptation

AU - Varsavsky, Thomas

AU - Orbes-Arteaga, Mauricio

AU - Sudre, Carole H.

AU - Graham, Mark S.

AU - Nachev, Parashkev

AU - Cardoso, M. Jorge

PY - 2020

Y1 - 2020

N2 - Convolutional neural networks trained on publicly available medical imaging datasets (source domain) rarely generalise to different scanners or acquisition protocols (target domain). This motivates the active field of domain adaptation. While some approaches to the problem require labelled data from the target domain, others adopt an unsupervised approach to domain adaptation (UDA). Evaluating UDA methods consists of measuring the model’s ability to generalise to unseen data in the target domain. In this work, we argue that this is not as useful as adapting to the test set directly. We therefore propose an evaluation framework where we perform test-time UDA on each subject separately. We show that models adapted to a specific target subject from the target domain outperform a domain adaptation method which has seen more data of the target domain but not this specific target subject. This result supports the thesis that unsupervised domain adaptation should be used at test-time, even if only using a single target-domain subject.

AB - Convolutional neural networks trained on publicly available medical imaging datasets (source domain) rarely generalise to different scanners or acquisition protocols (target domain). This motivates the active field of domain adaptation. While some approaches to the problem require labelled data from the target domain, others adopt an unsupervised approach to domain adaptation (UDA). Evaluating UDA methods consists of measuring the model’s ability to generalise to unseen data in the target domain. In this work, we argue that this is not as useful as adapting to the test set directly. We therefore propose an evaluation framework where we perform test-time UDA on each subject separately. We show that models adapted to a specific target subject from the target domain outperform a domain adaptation method which has seen more data of the target domain but not this specific target subject. This result supports the thesis that unsupervised domain adaptation should be used at test-time, even if only using a single target-domain subject.

KW - Brain MRI

KW - Domain adaptation

KW - One-shot

UR - http://www.scopus.com/inward/record.url?scp=85093091748&partnerID=8YFLogxK

U2 - 10.1007/978-3-030-59710-8_42

DO - 10.1007/978-3-030-59710-8_42

M3 - Conference paper

AN - SCOPUS:85093091748

SN - 9783030597092

T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

SP - 428

EP - 436

BT - Medical Image Computing and Computer Assisted Intervention – MICCAI 2020 - 23rd International Conference, Proceedings

A2 - Martel, Anne L.

A2 - Abolmaesumi, Purang

A2 - Stoyanov, Danail

A2 - Mateus, Diana

A2 - Zuluaga, Maria A.

A2 - Zhou, S. Kevin

A2 - Racoceanu, Daniel

A2 - Joskowicz, Leo

PB - Springer Science and Business Media Deutschland GmbH

T2 - 23rd International Conference on Medical Image Computing and Computer-Assisted Intervention, MICCAI 2020

Y2 - 4 October 2020 through 8 October 2020

ER -