HeartSpot Privatized and Explainable Data Compression for Cardiomegaly Detection Elvin Johnson Shreshta Mohan Alex Gaudioyz Asim Smailagic Christos Faloutsos Aur elio Campilhoyz

2025-05-06 0 0 9.28MB 5 页 10玖币

侵权投诉

HeartSpot: Privatized and Explainable Data

Compression for Cardiomegaly Detection

Elvin Johnson∗, Shreshta Mohan∗, Alex Gaudio∗†‡ , Asim Smailagic∗, Christos Faloutsos∗, Aur´

elio Campilho†‡

∗Carnegie Mellon University, Pittsburgh, PA, USA

†Faculty of Engineering of the University of Porto, Porto, Portugal

‡INESC TEC, Porto, Portugal

Abstract—Advances in data-driven deep learning for chest

X-ray image analysis underscore the need for explainability,

privacy, large datasets and signiﬁcant computational resources.

We frame privacy and explainability as a lossy single-image

compression problem to reduce both computational and data

requirements without training. For Cardiomegaly detection in

chest X-ray images, we propose HeartSpot and four spatial bias

priors. HeartSpot priors deﬁne how to sample pixels based on

domain knowledge from medical literature and from machines.

HeartSpot privatizes chest X-ray images by discarding up to 97%

of pixels, such as those that reveal the shape of the thoracic cage,

bones, small lesions and other sensitive features. HeartSpot priors

are ante-hoc explainable and give a human-interpretable image

of the preserved spatial features that clearly outlines the heart.

HeartSpot offers strong compression, with up to 32xfewer pixels

and 11xsmaller ﬁlesize. Cardiomegaly detectors using HeartSpot

are up to 9xfaster to train or at least as accurate (up to +.01

AUC ROC) when compared to a baseline DenseNet121. HeartSpot

is post-hoc explainable by re-using existing attribution methods

without requiring access to the original non-privatized image.

In summary, HeartSpot improves speed and accuracy, reduces

image size, improves privacy and ensures explainability.

Index Terms—Explainability, Privacy, Compression, Domain

Knowledge, Chest X-ray, Medical Image Analysis, Deep Learning

I. INTRODUCTION

Cardiomegaly is a medical condition describing an enlarged

heart. It can indicate underlying or life-threatening heart prob-

lems. Medical domain knowledge of Cardiomegaly describes

its detection from chest X-ray images as a relation between the

volume of the heart to the volume of the thoracic cage. The

cardio-thoracic ratio in a two-dimensional image, for instance,

compares the transverse diameter of the heart in pixels to that

of the thoracic cage. A variation compares the area of the heart

to the lungs [1]. Our approach, based on sampling of lines of

pixels, encompasses the main idea of both techniques.

Existing literature on Cardiomegaly detection from chest X-

ray considers the problem via classiﬁcation or segmentation

[2]. For classiﬁcation on the CheXpert [3] dataset of chest

X-ray images, several standard deep networks were evaluated

The TAMI project is funded by the European Regional Development

Fund (ERDF) through the Programa Operacional Regional do Norte (NORTE

2020) and by National Funds through the Portuguese Foundation for Science

and Technology (FCT). I.P. within the scope of the CMU Portugal Pro-

gram LA/P/0063/2020 and NORTE-01-0247-FEDER-045905. Alex Gaudio

is funded by the FCT, SFRH/BD/143114/2018.

(a) Good Compression (b) Fast and Accurate

Fig. 1: HeartSpot compresses with no learning (1a) to give a

fast and accurate (1b) model. HeartSpot image compression

is ante-hoc explainable and privatized (1c) with useful post-

hoc saliency map explanations obtained without access to the

original chest X-ray image (1d).

to offer an expected benchmark [4]. The work of [5] trains a

U-Net to segment the heart and thorax, and then computes

the thoracic ratio from image segmentation. Segmentation

improves performance over classiﬁcation, but requires pixel-

wise labels. The existing literature does not consider privacy

or compression. Our approach improves both efﬁciency and

privacy through single-image compression and is compatible

with any of these classiﬁcation or segmentation approaches.

As in Figure 1, the main contributions of HeartSpot are:

•Compression of single images with no learning, giving

up to to 11xsmaller ﬁlesize and 32xfewer pixels.

•Speed and Accuracy Gains up to +0.01 area under ROC

curve or 9xfaster training throughput.

•Privatized and Ante-hoc explainable compression via a

spatial bias prior. Up to 97% of pixels are removed while

preserved pixels clearly visualize the heart.

•Post-Hoc Attribution Explanations without requiring

the original image and improved via a quantile ﬁlter.

arXiv:2210.02241v1 [eess.IV] 5 Oct 2022

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：

HeartSpot:PrivatizedandExplainableDataCompressionforCardiomegalyDetectionElvinJohnson,ShreshtaMohan,AlexGaudioyz,AsimSmailagic,ChristosFaloutsos,Aur´elioCampilhoyzCarnegieMellonUniversity,Pittsburgh,PA,USAyFacultyofEngineeringoftheUniversityofPorto,Porto,PortugalzINESCTEC,Porto,PortugalAbstrac...

展开>> 收起<<

HeartSpot Privatized and Explainable Data Compression for Cardiomegaly Detection Elvin Johnson Shreshta Mohan Alex Gaudioyz Asim Smailagic Christos Faloutsos Aur elio Campilhoyz.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！

HeartSpot Privatized and Explainable Data Compression for Cardiomegaly Detection Elvin Johnson Shreshta Mohan Alex Gaudioyz Asim Smailagic Christos Faloutsos Aur elio Campilhoyz

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: