Grad Mix for nuclei segmentation and classification in imbalanced pathology image datasets Tan Nhu Nhat Doan1 Kyungeun Kim2 Boram Song2 and Jin Tae Kwak3

2025-05-06 2 0 647.29KB 10 页 10玖币

侵权投诉

GradMix for nuclei segmentation and classification in

imbalanced pathology image datasets

Tan Nhu Nhat Doan1, Kyungeun Kim2, Boram Song2, and Jin Tae Kwak3

1Department of Computer Science and Engineering, Sejong University, Seoul 05006, Korea

2Department of Pathology, Kangbuk Samsung Hospital, Sungkyunkwan University School of

Medicine, Seoul 05505, Korea

3School of Electrical Engineering, Korea University, Seoul 02841, Korea

jkwak@korea.ac.kr

Abstract. An automated segmentation and classification of nuclei is an essential

task in digital pathology. The current deep learning-based approaches require a

vast amount of annotated datasets by pathologists. However, the existing datasets

are imbalanced among different types of nuclei in general, leading to a substantial

performance degradation. In this paper, we propose a simple but effective data

augmentation technique, termed GradMix, that is specifically designed for nuclei

segmentation and classification. GradMix takes a pair of a major-class nucleus

and a rare-class nucleus, creates a customized mixing mask, and combines them

using the mask to generate a new rare-class nucleus. As it combines two nuclei,

GradMix considers both nuclei and the neighboring environment by using the

customized mixing mask. This allows us to generate realistic rare-class nuclei

with varying environments. We employed two datasets to evaluate the effective-

ness of GradMix. The experimental results suggest that GradMix is able to im-

prove the performance of nuclei segmentation and classification in imbalanced

pathology image datasets.

Keywords: nuclei segmentation and classification, data augmentation, data im-

balance.

1 Introduction

The assessment of nuclei is one of the primary tasks in digital pathology since nuclear

features, including shape, size, and density, have known to be related to disease diag-

nosis and prognosis [1]. In order to analyze nuclear features in pathology images, an

accurate and reliable segmentation and classification of nuclei is a pre-requisite. How-

ever, nuclei segmentation and classification remain a challenging task since there exists

an enormous number of nuclei in a relatively small pathology image and there is a sub-

stantial intra- and inter-variability in the morphology, texture, and intensity among nu-

clei of differing cell types as well as within the same cell type. Hence, a robust nuclei

segmentation and classification method is needed to expedite digital pathology analysis

and to improve diagnostic decisions on pathology images.

Recently, several research efforts have made to develop deep learning-based nuclei

segmentation and classification methods. Most of them focused on nuclei segmentation

where one of the most challenging tasks is to separate touching or overlapping nuclei

[2]. Some designed multi-resolution convolutional neural networks (CNNs) to preserve

contextual information at multiple resolutions [3] [4]. Some others proposed to exploit

morphology of nuclei. For example, [5] [4] utilized nuclear boundaries in identifying

individual nuclei. [6] formulated nuclei segmentation as a regression of the distance

map of nuclei. [7] exploited both the nuclear distance map and nuclear boundary for

nuclei segmentation. Moreover, [8] utilized dense steerable filters to achieve rotation-

symmetry within the network. Nuclei classification has been mainly studied as a down-

stream analysis of nuclei segmentation. For instance, [9] detected nuclear centroids and

classified nuclei using CNN. [10] proposed RCCNet that classifies nuclei image

patches into pertinent classes. [2] developed HoVer-Net that simultaneously performs

nuclei segmentation and classification by utilizing horizontal and vertical distance maps

of nuclei. Despite such recent advances, nuclei segmentation and classification still

need to be improved. There exists a high variability in both segmentation and classifi-

cation performance among different types of nuclei [2] [11]. This may be ascribable to

imbalance in the datasets among different nuclei types. The lesser the annotated nuclei

are available, the poorer the performance is in general. For nuclei segmentation, [12,

13] proposed to use a generative adversarial network (GAN) to synthesize pathology

images with known nuclei annotations; however, GAN is not only computationally ex-

pensive but also requires a sufficient number of supervised datasets. Mixup [14], Cut-

out[15], and CutMix [16] are regularization techniques to generate new images from

the existing images. These are computationally efficient and have been successfully

applied to image classification and object detection. But, no prior work exists for image

segmentation in digital pathology.

Herein, we propose a gradation mixup (GradMix) data augmentation technique for

an improved nuclei segmentation and classification in imbalanced pathology image da-

tasets. In the imbalanced datasets, there exist one or more major-classes of nuclei that

are prevalent in the datasets and one or more rare-classes of nuclei that are under-rep-

resented in the datasets. GradMix is a data augmentation technique that is tailored to

nuclei segmentation and classification tasks. The technique aims at increasing the num-

ber of rare-class nuclei by generating realistic nuclei under various environments. Grad-

Mix generates a new rare-class nucleus by combining a major-class nucleus with a rare-

class nucleus via a customized mixing mask ℳ. The rare-class nucleus is utilized as it

is and placed at the center of the major-class nucleus. Then, the neighboring pixels of

the rare-class nucleus and the corresponding pixels that are either major-class nucleus

or its neighbors are combined with the corresponding weights in ℳ. The weights for

the neighboring pixels of the rare-class nucleus are inversely related to the distance to

the boundary of the rare-class nucleus. In this manner, we are able to generate a set of

new, realistic rare-class nuclei with varying environments. This, in turn, aids in improv-

ing the performance of nuclei segmentation and classification.

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：

GradMixfornucleisegmentationandclassificationinimbalancedpathologyimagedatasetsTanNhuNhatDoan1,KyungeunKim2,BoramSong2,andJinTaeKwak31DepartmentofComputerScienceandEngineering,SejongUniversity,Seoul05006,Korea2DepartmentofPathology,KangbukSamsungHospital,SungkyunkwanUniversitySchoolofMedicine,Seoul0...

展开>> 收起<<

Grad Mix for nuclei segmentation and classification in imbalanced pathology image datasets Tan Nhu Nhat Doan1 Kyungeun Kim2 Boram Song2 and Jin Tae Kwak3.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！

Grad Mix for nuclei segmentation and classification in imbalanced pathology image datasets Tan Nhu Nhat Doan1 Kyungeun Kim2 Boram Song2 and Jin Tae Kwak3

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: