IDEAL IMPROVED DENSE LOCAL CONTRASTIVE LEARNING FOR SEMI-SUPERVISED MEDICAL IMAGE SEGMENTATION Hritam Basak1 Soumitri Chattopadhyay2 Rohit Kundu2 Sayan Nag3 Rammohan Mallipeddi4

2025-05-08 0 0 667.39KB 5 页 10玖币

侵权投诉

IDEAL: IMPROVED DENSE LOCAL CONTRASTIVE LEARNING FOR SEMI-SUPERVISED

MEDICAL IMAGE SEGMENTATION

Hritam Basak1, Soumitri Chattopadhyay∗2, Rohit Kundu∗2, Sayan Nag∗3, Rammohan Mallipeddi 4

1Stony Brook University, 2Jadavpur University, 3University of Toronto, 4Kyungpook National University

ABSTRACT

Due to the scarcity of labeled data, Contrastive Self-Supervised

Learning (SSL) frameworks have lately shown great potential

in several medical image analysis tasks. However, the existing

contrastive mechanisms are sub-optimal for dense pixel-level

segmentation tasks due to their inability to mine local fea-

tures. To this end, we extend the concept of metric learning to

the segmentation task, using a dense (dis)similarity learning

for pre-training a deep encoder network, and employing a

semi-supervised paradigm to ﬁne-tune for the downstream

task. Speciﬁcally, we propose a simple convolutional pro-

jection head for obtaining dense pixel-level features, and a

new contrastive loss to utilize these dense projections thereby

improving the local representations. A bidirectional consis-

tency regularization mechanism involving two-stream model

training is devised for the downstream task. Upon compar-

ison, our IDEAL method outperforms the SoTA methods

by fair margins on cardiac MRI segmentation. Our source

codes are publicly accessible at: https://github.com/Rohit-

Kundu/IDEAL-ICASSP23.

Index Terms—Semi-supervised learning, Segmentation,

MRI, Contrastive learning

1. INTRODUCTION

The success of supervised deep learning approaches can be

attributed to the availability of large quantities of labeled data

that is essential for network training [1]. However, in the

biomedical domain [2,3], it is difﬁcult to acquire such large

quantities of annotated data as the annotations are to be done

by trained medical professionals. Although supervised learn-

ing has been used extensively in the past decade in biomedi-

cal imaging [1], Self-supervised Learning (SSL) [4,5,6,7,8]

provides more traction for sustaining deep learning methods

in the medical vision domain [9,10,11]. SSL-based pre-

training alleviates the data annotation problem by utilizing

only unstructured data to learn distinctive information which

can further be utilized in downstream applications, typically

in a semi-supervised fashion [12,13,9]. Typically, SSL have

shown to be exceedingly promising in domains with enor-

mous amounts of data such as natural images [6,7] with a

*Equal contribution.

focus on both contrastive [14,7,6,15] and non-contrastive

variants [16,17]. Recently, SSL has begun to be employed

for medical imaging as well, evident from the survey [2].

One of the most popular ways of employing SSL is con-

trastive learning [14,7,6], which owing to its proliﬁc per-

formance across various vision tasks has almost become a

de facto standard for self-supervision. The intuition of con-

trastive learning is to pull embeddings of semantically similar

objects closer and simultaneously push away the representa-

tions of dissimilar objects. In medical image segmentation,

contrastive learning has been leveraged in a few prior works

[13,18]. However, naive contrastive learning frameworks

such as SimCLR [6] and MoCo [7] learn a global represen-

tation and thus, are not suitable to be applied directly to seg-

mentation, a pixel-level task. The seminal work by Chaitanya

et. al. [13] utilized a contrastive pre-training strategy to learn

useful representations for a segmentation task in a low-data

regime. Despite having success, their method has a few limi-

tations, which we attempt to address and improve upon in this

paper. In particular, we take a cue from the aforementioned

work [13] to propose a novel contrastive learning-based med-

ical image segmentation framework.

Although our work is built upon [13], there are several

salient points of difference between the two models that ren-

der our contributions non-trivial and unique. First of all, our

proposed method preserves spatial information and constructs

a dense output projection that preserves locality. This is in

contrast to the projection used in [13] where a global pooling

is applied to the backbone encoder, thereby obtaining a sin-

gle global feature representation vector for every input image.

In other words, [13] employed global contrastive learning for

encoder pre-training and the local contrastive learning is only

restricted to the decoder ﬁne-tuning. In contrast, our method

involves pre-training of the encoder on dense local features.

The intuitive difference between our proposed framework and

that of [13] has been depicted in Fig. 1. We argue that this

will beneﬁt pixel-level informed downstream tasks such as

segmentation (supported by our ﬁndings in Section 3).

Secondly, our deﬁnition of ﬁnding positive and negative

pairs for contrastive learning is different from that of tradi-

tional contrastive [6,7] methods. This is because, in our

case, we ﬁnd dense correspondences across views and deﬁne

positives and negatives accordingly, i.e., we perform feature-

arXiv:2210.15075v2 [cs.CV] 2 Mar 2023

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：

IDEAL:IMPROVEDDENSELOCALCONTRASTIVELEARNINGFORSEMI-SUPERVISEDMEDICALIMAGESEGMENTATIONHritamBasak1,SoumitriChattopadhyay2,RohitKundu2,SayanNag3,RammohanMallipeddi41StonyBrookUniversity,2JadavpurUniversity,3UniversityofToronto,4KyungpookNationalUniversityABSTRACTDuetothescarcityoflabeleddata,Contra...

展开>> 收起<<

IDEAL IMPROVED DENSE LOCAL CONTRASTIVE LEARNING FOR SEMI-SUPERVISED MEDICAL IMAGE SEGMENTATION Hritam Basak1 Soumitri Chattopadhyay2 Rohit Kundu2 Sayan Nag3 Rammohan Mallipeddi4.pdf

共5页,预览1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！

IDEAL IMPROVED DENSE LOCAL CONTRASTIVE LEARNING FOR SEMI-SUPERVISED MEDICAL IMAGE SEGMENTATION Hritam Basak1 Soumitri Chattopadhyay2 Rohit Kundu2 Sayan Nag3 Rammohan Mallipeddi4

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: