OOOE Only-One-Object-Exists Assumption to Find Very Small Objects in Chest Radiographs Gunhee Nam Taesoo Kim Sanghyup Lee and Thijs Kooi

2025-04-24 0 0 2.47MB 10 页 10玖币

侵权投诉

OOOE: Only-One-Object-Exists Assumption to

Find Very Small Objects in Chest Radiographs

Gunhee Nam, Taesoo Kim, Sanghyup Lee, and Thijs Kooi

Lunit Inc.

{ghnam, taesoo.kim, eesanghyup, tkooi}@lunit.io

Abstract. The accurate localization of inserted medical tubes and parts

of human anatomy is a common problem when analyzing chest radio-

graphs and something deep neural networks could potentially automate.

However, many foreign objects like tubes and various anatomical struc-

tures are small in comparison to the entire chest X-ray, which leads

to severely unbalanced data and makes training deep neural networks

diﬃcult. In this paper, we present a simple yet eﬀective ‘Only-One-

Object-Exists’ (OOOE) assumption to improve the deep network’s abil-

ity to localize small landmarks in chest radiographs. The OOOE enables

us to recast the localization problem as a classiﬁcation problem and

we can replace commonly used continuous regression techniques with a

multi-class discrete objective. We validate our approach using a large

scale proprietary dataset of over 100K radiographs as well as publicly

available RANZCR-CLiP Kaggle Challenge dataset and show that our

method consistently outperforms commonly used regression-based detec-

tion models as well as commonly used pixel-wise classiﬁcation methods.

Additionally, we ﬁnd that the method using the OOOE assumption gen-

eralizes to multiple detection problems in chest X-rays and the resulting

model shows state-of-the-art performance on detecting various tube tips

inserted to the patient as well as patient anatomy.

Keywords: Point detection ·Localization ·Object segmentation.

1 Introduction

A common and eﬀective application of deep neural networks in the domain of

automated Chest X-ray (CXR) analysis is the localization of foreign objects and

human anatomy [32]. For example, the ability to segment and locate foreign ob-

jects, such as catheters, tubes, and lines has tremendous potential to optimize

clinical workﬂow and ultimately improve patient care [6,7,8]. The innovations

in object detection and segmentation methods for natural images [5,21,25,26]

have sparked progress in detecting foreign objects and anatomy in CXR images

[4,18,28]. However, despite unique challenges associated with ﬁnding objects in

CXR images, many of the methods designed for equivalent tasks in natural im-

ages are applied to CXR images without signiﬁcant architectural modiﬁcations.

Compared to most objects in natural images, foreign objects and human

anatomy viewed in CXR images are much smaller in scale. Training deep neural

arXiv:2210.06806v1 [cs.CV] 13 Oct 2022

2 Nam et al.

network to detect small scale objects is challenging, because the number of back-

ground pixels far outweighs the foreground pixel count [3,22,29]. Frid-Adar et

al. [4] proposed to generate training data by synthesizing images with augmented

endotracheal tubes (ETT). Their method addresses data imbalance and improves

performance of an image level classiﬁcation, but does not provide a solution for

the small object detection problem. Kara et al.[15] proposed a regression based

cascade method to localize the tip of ETT and the carina. In comparison, we

provide a classiﬁcation based solution to the detection problem which is often

reported to outperform regression based methods for various detection tasks in

natural images [12,19,20,24,27].

In this work, we present a solution to the problem of detecting small foreign

objects or anatomical structures in chest radiographs. We introduce the ‘Only-

One-Object-Exists’ (OOOE) assumption, a simple yet eﬀective assumption, that

limits the number of observable instances of a particular object we want to detect

to one per image and reduces the detection problem to a point localization

problem. Using these assumptions, the localization problem can be cast as a

classiﬁcation problem that can be solved with a spatial-softmax operation.

We validate our approach for (1) detecting ETT tip and (2) detecting the

carina, on the publicly available RANZCR-CliP Kaggle Challenge dataset. Ad-

ditionally, we also provide results on a large scale proprietary dataset of over

100K chest X-ray images. Our method inspired by the OOOE assumption out-

performs two commonly used baselines: (1) a simple segmentation model [23,26]

and (2) a regression based detection approach [15]. We additionally demonstrate

that our approach leads to a model that generalizes better across datasets and

makes better use of global context information.

2 Methods

We address the problem of detecting small objects in an image, using the as-

sumption that they occur once and only once. We also observe that small objects,

such as the tip of a tube or a certain landmark of an anatomy, can essentially

be represented as a single point in an image.

Our solution to the point detection problem consists of two parts: a feature

extractor Fand a detection head g, which will be described in detail in the

following sections.

2.1 Feature Extractor

A feature extractor is a function that satisﬁes the following:

X=F(I),(1)

where I∈RH×W×Cis an input image with spatial dimensions H, W and with

Cchannels. The feature extractor Fis a transformation such that the output

feature X∈Rh×w×cis a tensor with spatial dimensions h, w such that h <

H, w < W and with cchannels. In this work, we implement Fwith a widely

used convolutional neural network with residual connections (ResNet34) [9].

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：

OOOE:Only-One-Object-ExistsAssumptiontoFindVerySmallObjectsinChestRadiographsGunheeNam,TaesooKim,SanghyupLee,andThijsKooiLunitInc.fghnam,taesoo.kim,eesanghyup,tkooig@lunit.ioAbstract.Theaccuratelocalizationofinsertedmedicaltubesandpartsofhumananatomyisacommonproblemwhenanalyzingchestradio-graphsands...

展开>> 收起<<

OOOE Only-One-Object-Exists Assumption to Find Very Small Objects in Chest Radiographs Gunhee Nam Taesoo Kim Sanghyup Lee and Thijs Kooi.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！

OOOE Only-One-Object-Exists Assumption to Find Very Small Objects in Chest Radiographs Gunhee Nam Taesoo Kim Sanghyup Lee and Thijs Kooi

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: