Unsupervised Cross-Modality Domain Adaptation for Vestibular Schwannoma Segmentation and Koos Grade Prediction based

2025-05-06 0 0 2.01MB 10 页 10玖币

侵权投诉

Unsupervised Cross-Modality Domain

Adaptation for Vestibular Schwannoma

Segmentation and Koos Grade Prediction based

on Semi-Supervised Contrastive Learning

Luyi Han1,2, Yunzhi Huang3?, Tao Tan2( ), and Ritse Mann1,2

1Department of Radiology and Nuclear Medicine, Radboud University Medical

Center, Geert Grooteplein 10, 6525 GA, Nijmegen, The Netherlands.

2Department of Radiology, The Netherlands Cancer Institute, Plesmanlaan 121,

1066 CX, Amsterdam, The Netherlands.

3School of Automation, Nanjing University of Information Science and Technology,

Nanjing 210044, China.

{taotanjs}@gmail.com

Abstract. Domain adaptation has been widely adopted to transfer styles

across multi-vendors and multi-centers, as well as to complement the

missing modalities. In this challenge, we proposed an unsupervised do-

main adaptation framework for cross-modality vestibular schwannoma

(VS) and cochlea segmentation and Koos grade prediction. We learn the

shared representation from both ceT1 and hrT2 images and recover an-

other modality from the latent representation, and we also utilize proxy

tasks of VS segmentation and brain parcellation to restrict the consis-

tency of image structures in domain adaptation. After generating missing

modalities, the nnU-Net model is utilized for VS and cochlea segmenta-

tion, while a semi-supervised contrastive learning pre-train approach is

employed to improve the model performance for Koos grade prediction.

On CrossMoDA validation phase Leaderboard, our method received rank

4 in task1 with a mean Dice score of 0.8394 and rank 2 in task2 with

Macro-Average Mean Square Error of 0.3941. Our code is available at

https://github.com/fiy2W/cmda2022.superpolymerization.

Keywords: Domain Adaptation ·Semi-supervised Contrastive Learn-

ing ·Segmentation ·Vestibular Schwnannoma.

1 Introduction

Domain adaptation has recently been employed in various clinical settings to im-

prove the applicability of deep learning approaches. The goal of Cross-Modality

Domain Adaptation (CrossMoDA) challenge 4is to segment two key brain struc-

tures, namely vestibular schwannoma (VS) and cochlea. It also requires predict-

?Luyi Han and Yunzhi Huang contributed equally to this work.

4https://crossmoda2022.grand-challenge.org/

arXiv:2210.04255v1 [cs.CV] 9 Oct 2022

2 L. Han et al.

ing the Koos grading scale for VS. The two tasks are required for the measure-

ment of VS growth and evaluation of the treatment plan (surveillance, radio-

surgery, open surgery). Although contrast-enhanced T1 (ceT1) MR imaging is

commonly used for patients with VS in diagnosis and surveillance, the research

on non-contrast imaging, such as high-resolution T2 (hrT2), is growing due to

lower risk and more eﬃcient cost. Therefore, CrossMoDA aims to transfer the

model learned from annotated ceT1 images to unpaired and unlabeled hrT2

images based on domain adaptation.

2 Related Work

Unsupervised domain adaptation for VS and cochlea segmentation has been

extensively validated in previous research [6]. Most of them employ an image-

to-image translation method, e.g. CycleGAN [16], to generate pseudo-target do-

main images from source domain images. And then generated images and the

corresponding manual annotations are used to train the segmentation models.

Dong et al. [5] utilize NiceGAN [2], which is trained by reusing discriminators

for encoding, to improve the performance of domain adaptation and further

segmentation. Choi [3] proposes a data augmentation method by halving the in-

tensity in the tumor area for generated hrT2. Shin et al. [13] employ an iterable

self-training strategy in their method: (1) train the student model with anno-

tated generated hrT2 and pseudo-labeled real hrT2; (2) make the student a new

teacher and update the pseudo label for real htT2. Following these works, our

proposed method focuses more on extracting joint representations from multi-

modality MRIs, which can reduce the distance between diﬀerent modalities in

the latent space.

Classiﬁcation task for medical images is always more diﬃcult than segmen-

tation due to fewer annotations. In recent years, contrastive learning has led to

state-of-the-art performance in self-supervised representation learning [14,1,7,10].

The key idea is to reduce the distance between an anchor and a “positive” sample

in latent space, and distinguish the anchor from other “negative” samples. Based

on this, contrastive learning can be applied to multi-modality pretraining. In or-

der to improve the sensor setup ﬂexibility of the robot, Meyer et al. [11] propose

a multimodal approach based on contrastive learning to learn from RGB-Depth

images. Yuan et al. [15] develop an approach for joint visual-textural pretrain-

ing that focuses on both intra-modality and inter-modality learning. For medical

image analysis, Huang et al. [8] develop an attentional contrastive learning frame-

work for global and local representation learning between images and radiology

reports. Inspired by these works, contrastive learning is employed at the pre-

training phase to mine multi-modality representation strategically for diﬀerent

types of samples.

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

UnsupervisedCross-ModalityDomainAdaptationforVestibularSchwannomaSegmentationandKoosGradePredictionbasedonSemi-SupervisedContrastiveLearningLuyiHan1;2,YunzhiHuang3?,TaoTan2(),andRitseMann1;21DepartmentofRadiologyandNuclearMedicine,RadboudUniversityMedicalCenter,GeertGrooteplein10,6525GA,Nijmegen,The...

展开>> 收起<<

Unsupervised Cross-Modality Domain Adaptation for Vestibular Schwannoma Segmentation and Koos Grade Prediction based.pdf

共10页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Unsupervised Cross-Modality Domain Adaptation for Vestibular Schwannoma Segmentation and Koos Grade Prediction based

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: