EOCSA Predicting Prognosis of Epithelial Ovarian Cancer with Whole Slide Histopathological Images_2

2025-04-29 1 0 2.06MB 32 页 10玖币

侵权投诉

EOCSA: Predicting Prognosis of Epithelial Ovarian

Cancer with Whole Slide Histopathological Images

Tianling Liua, Ran Sua,∗, Changming Sunc, Xiuting Lid, Leyi Weib,∗

aSchool of Computer Software, College of Intelligence and Computing, Tianjin University,

China.

bSchool of Software, Shandong University, China.

cCSIRO Data61, Epping, NSW 1710, Australia.

dSingapore Bio-imaging Sciences Consortium (SBIC), Agency for Science, Technology and

Research (A*STAR), Singapore.

Abstract

Ovarian cancer is one of the most serious cancers that threaten women around

the world. Epithelial ovarian cancer (EOC), as the most commonly seen subtype

of ovarian cancer, has rather high mortality rate and poor prognosis among var-

ious gynecological cancers. Survival analysis outcome is able to provide treat-

ment advices to doctors. In recent years, with the development of medical

imaging technology, survival prediction approaches based on pathological im-

ages have been proposed. In this study, we designed a deep framework named

EOCSA which analyzes the prognosis of EOC patients based on pathological

whole slide images (WSIs). Speciﬁcally, we ﬁrst randomly extracted patches

from WSIs and grouped them into multiple clusters. Next, we developed a

survival prediction model, named DeepConvAttentionSurv (DCAS), which was

able to extract patch-level features, removed less discriminative clusters and

predicted the EOC survival precisely. Particularly, channel attention, spatial

attention, and neuron attention mechanisms were used to improve the perfor-

mance of feature extraction. Then patient-level features were generated from

our weight calculation method and the survival time was ﬁnally estimated using

∗Corresponding author

Email addresses: dling@tju.edu.cn (Tianling Liu), ran.su@tju.edu.cn (Ran Su),

changming.sun@csiro.au (Changming Sun), li_xiuting@sbic.a-star.edu.sg (Xiuting Li),

weileyi@sdu.edu.cn (Leyi Wei)

Preprint submitted to Journal of L

X Templates October 12, 2022

arXiv:2210.05258v1 [eess.IV] 11 Oct 2022

LASSO-Cox model. The proposed EOCSA is eﬃcient and eﬀective in predicting

prognosis of EOC and the DCAS ensures more informative and discriminative

features can be extracted. As far as we know, our work is the ﬁrst to analyze

the survival of EOC based on WSIs and deep neural network technologies. The

experimental results demonstrate that our proposed framework has achieved

state-of-the-art performance of 0.980 C-index. The implementation of the ap-

proach can be found at https://github.com/RanSuLab/EOCprognosis.

Keywords: Epithelial ovarian cancer, WSI, EOCSA, DCAS, Prognosis

prediction, Deep learning

1. Introduction

Ovarian cancer is a disease that seriously damages the well-being of women in

the world. It ranks the seventh among the most common cancers in women and

the eighth among all female deadly cancers (Lheureux et al., 2019). Epithelial

ovarian cancer (EOC), as the most commonly seen subtype of ovarian cancer,

accounts for more than 90% among all ovarian cancer patients (Reid et al.,

2017; Torre et al., 2018). Although the 5-year relative survival rate at the early

stage of EOC is as high as 93% (Torre et al., 2018), the symptoms at the early

stage of EOC are not obvious, which make EOC diﬃcult to be detected at the

early stage (Lheureux et al., 2019). In contrast, more than 75% of EOCs are

detected at an advanced stage and the 5-year relative survival rate of EOC at an

advanced stage is only about 29% (Lheureux et al., 2019; van Baal et al., 2018).

Survival analysis is a branch of statistics which concentrates on predicting the

time duration from the beginning of the follow-up study to the interest events

occur, for example, disease recurrence or death (Ohno-Machado, 2001). An

important application of survival analysis in the medical ﬁeld is to observe the

prognosis of diseases (Huang et al., 2018; Kather et al., 2019). Survival analysis

can make a rough prediction of patients’ survival status over time, and can also

roughly assess the development trend of the disease condition. There are many

factors in EOC patients’ data, such as age, gender, pathogenesis and tumor

stage (Holschneider & Berek, 2000; Tingulstad et al., 2003). Survival analysis

can determine which factor or set of factors has a greater impact on patient’s

survival, so that doctors can evaluate the eﬀectiveness of diﬀerent treatment

plans to tailor for each patient. Doctors can also change the treatment plans

or drugs according to the results of survival analysis during the treatment of

patients to ensure that patients have a high survival rate. With the development

of medical imaging technology, various types of cancer images are available such

as MRIs, CT images and pathological images. These high-quality images contain

rich information related to the characteristics of cancers, thus a large number of

image-based survival analysis methods have been proposed (Wang et al., 2014;

Yu et al., 2016; Lu et al., 2019). Pathological images, especially the whole slide

images (WSIs) have been adopted to perform survival analysis and have shown

impressive performance (Wang et al., 2014; Yu et al., 2016; Zhu et al., 2016a).

Most of these methods ﬁrstly extracted hand-crafted features and then used

machine learning algorithms to conduct survival prediction (Wang et al., 2014;

Yu et al., 2016). However, the hand-crafted features extracted from the WSIs

requires predeﬁnition of features, thus the performance largely depends on prior

knowledge which may bring bias into the results.

The development of deep learning brings about the automated acquisition of

features and has increasingly demonstrated the ability to extract high-dimensional

features (Jin et al., 2019; Su et al., 2019). In recent years, there are many sur-

vival analysis studies based on deep learning. Zhu et al. proposed the Deep-

ConvSurv model which applied deep convolutional neural network to image-

based survival analysis tasks (Zhu et al., 2016a). For further solving WSI-based

survival analysis tasks, Zhu et al. proposed the WSISA framework that esti-

mated the survival using discriminative patches extracted from WSIs (Zhu et al.,

2017). Mobadersany et al. integrated WSIs and genetic data to make cancer

outcomes prediction using convolutional neural network (CNN) (Mobadersany

et al., 2018). Using WSIs and deep learning to analyze EOC survival still re-

quires exploration.

In this study, inspired by the WSISA, we proposed a EOCSA framework

for the analysis of EOC survival based on WSI data. Firstly, we randomly

picked a certain number of patches from each WSI and grouped them into

diﬀerent clusters. Then we trained a deep survival prediction model named

DeepConvAttentionSurv (DCAS) for each cluster. Next, we selected the clus-

ters which contained patches with discriminative information and extracted the

patch-level features based on the trained DCAS models. Finally, we generated

weighted patient-level features based on patch-level features to make ﬁnal sur-

vival analysis. Compared to the WSISA, the EOCSA is able to extract more

eﬀective features and make more accurate prediction. We have three contri-

butions in this study. Firstly, as far as we know, we are the ﬁrst to propose

the use of deep learning to process WSIs for survival analysis of EOC; Sec-

ondly, we developed an eﬃcient deep survival analysis model named DCAS. In

order to extract more eﬀective features, we added spatial, channel and neuron

attention modules into the architecture; Thirdly, we proposed a new weight cal-

culation method to obtain more discriminative patient-level features compared

with the existing methods. The last two contributions are algorithmic inno-

vations. And the ﬁrst contribution is application innovation. We apply two

algorithmic innovation to the survival analysis of EOC. Experiments demon-

strate the eﬃciency and eﬀectiveness of the proposed EOCSA in predicting

EOC patients’ survival. The implementation of the approach can be found at

https://github.com/RanSuLab/EOCprognosis.

2. Background

Survival analysis is frequently used to observe the prognosis of medical dis-

eases. The Kaplan-Meier survival estimate, a non-parameter method, predicts

survival rate by observed survival time (Dabrowska, 1988). It is the easiest way

to calculate survival rate over time (Goel et al., 2010). Mevlut et al. com-

bined Kaplan-Meier analysis with decision trees to predict survival of breast

cancer (Ture et al., 2009). But the drawback of Kaplan-Meier survival estimate

for survival analysis is that it is univariable analysis which can only perform

……………………

…… …… ……

......

Patch-level

Features

Weighted

Patient-level

Features

Results

Weight

Calculation

LASSO-Cox

Data Pre-processing Feature Extraction Final Survival Analysis

Randomly Patches Extraction

Cluster by K-means

DCAS Model Training

......

Figure 1: The overview of our EOCSA framework. There are three main steps in EOCSA: 1.

Data pre-processing, model training; 2. Feature extraction; 3. Weighted patient-level feature

generation and survival prediction.

single factor analysis and it ignores the eﬀects of other factors (Jager et al.,

2008). The interaction between multiple factors has a great impact on survival

analysis. Cox proportional hazards (CPH) model has the ability to perform

multivariate survival analysis (Cox, 1972) and is widely used in survival analy-

sis research. Guo et al. studied the link between genetic variation and survival

of breast cancer using the CPH model (Guo et al., 2015). With the advancing

of high-dimensional molecular data or clinical data, the survival analysis model

based on the CPH model has been gradually improved. Gui et al. proposed the

LARS-Cox model which classiﬁed high-risk and low-risk patient groups using

gene expression data relevant to the survival phenotypes (Gui & Li, 2005). Eric

et al. proposed a framework to predict cancer subtypes using the combination of

gene expression data and clinical data, and then proceeded with survival analy-

sis of cancer (Bair & Tibshirani, 2004). Park et al. proposed an L1-regularized

CPH model which was applicable to survival analysis of data in which the num-

ber of variables was much larger than that of the samples (Park & Hastie, 2007).

Tibshirani and Robert proposed a variable selection and compression method

used for the CPH model (Tibshirani, 1997). For high dimensional micro-array

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：

EOCSA:PredictingPrognosisofEpithelialOvarianCancerwithWholeSlideHistopathologicalImagesTianlingLiua,RanSua,,ChangmingSunc,XiutingLid,LeyiWeib,aSchoolofComputerSoftware,CollegeofIntelligenceandComputing,TianjinUniversity,China.bSchoolofSoftware,ShandongUniversity,China.cCSIROData61,Epping,NSW1710,A...

收起<<

EOCSA Predicting Prognosis of Epithelial Ovarian Cancer with Whole Slide Histopathological Images_2.pdf

共32页,预览5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！

EOCSA Predicting Prognosis of Epithelial Ovarian Cancer with Whole Slide Histopathological Images_2

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: