Anomaly Detection using Generative Models and Sum-Product Networks in Mammography Scans Marc Dietrichstein1 David Major1 Martin Trapp2 Maria Wimmer1

2025-04-27 0 0 533.9KB 10 页 10玖币

侵权投诉

Anomaly Detection using Generative Models and

Sum-Product Networks in Mammography Scans

Marc Dietrichstein1?, David Major1?, Martin Trapp2, Maria Wimmer1,

Dimitrios Lenis1, Philip Winter1, Astrid Berg1, Theresa Neubauer1, and

Katja Bühler1

1VRVis Zentrum für Virtual Reality und Visualisierung Forschungs-GmbH,

Vienna, Austria

david.major@vrvis.at

2Department of Computer Science, Aalto University, Espoo, Finland ??

Abstract. Unsupervised anomaly detection models which are trained

solely by healthy data, have gained importance in the recent years, as

the annotation of medical data is a tedious task. Autoencoders and gen-

erative adversarial networks are the standard anomaly detection methods

that are utilized to learn the data distribution. However, they fall short

when it comes to inference and evaluation of the likelihood of test sam-

ples. We propose a novel combination of generative models and a proba-

bilistic graphical model. After encoding image samples by autoencoders,

the distribution of data is modeled by Random and Tensorized Sum-

Product Networks ensuring exact and eﬃcient inference at test time. We

evaluate diﬀerent autoencoder architectures in combination with Ran-

dom and Tensorized Sum-Product Networks on mammography images

using patch-wise processing and observe superior performance over uti-

lizing the models standalone and state-of-the-art in anomaly detection

for medical data.

Keywords: Anomaly Detection ·Generative Models ·Sum-Product

Networks ·Mammography.

1 Introduction

Acceleration of the detection and segmentation of anomalous tissue by auto-

mated computer aided approaches is a key for enhancing cancer screening pro-

grams. It is especially important for mammography screening, as breast cancer

is the most common cancer type and the leading cause of death in women world-

wide [20]. Training an artiﬁcial neural network in a supervised way needs a high

?Equal contribution

?? This preprint has not undergone peer review (when applicable) or any post-

submission improvements or corrections. The Version of Record of this contribution

is published in LNCS 13609, and is available online at https://doi.org/10.1007/978-

3-031-18576-2 8.

arXiv:2210.06188v1 [cs.CV] 12 Oct 2022

2 M. Dietrichstein and D. Major et al.

amount of pixel-wise annotated data. As data annotation is very costly, meth-

ods which involve as less annotation as possible, are of high demand. Anomaly

detection approaches are good representatives of this type, as they only uti-

lize healthy cases for learning, and anomalous spots are detected as a deviation

from the learned data distribution. The deviation is measured either by straight-

forward metrics such as reconstruction error of input and output samples or by

more sophisticated constructs such as log-likelihood in probabilistic models.

Unsupervised anomaly detection methods have been evaluated on a plethora

of diﬀerent pathologies and medical imaging modalities. A state-of-the-art method

in this area is f-AnoGAN [14] which leverages Generative Adversarial Networks

(GANs) to model an implicit distribution of healthy images and detect outliers

via a custom anomaly score based on reconstruction performance. f-AnoGAN

has been utilized to detect anomalies in OCT scans [14], Chest X-rays [1] and

3D Brain scans [16]. However, it requires the training of a separate encoder

module to obtain latent codes of images, which are used by the generator for

reconstruction. The autoencoder architecture (AE) on the other hand jointly

trains an encoder and decoder and is thus able to directly map an input to

its corresponding latent representation. AE variants have been applied to le-

sion detection in mammography images [19] and brain scans [8,21] as well as

head [13] and abdomen [8] CT scans. However, the practical applicability of all

those models is limited by the fact that the respective anomaly scores are not

easily interpretable by a human decision maker. Here, to remedy the situation,

it would be desirable for the model to provide some degree of certainty for its

decision. To this end, density estimation models can be employed. Such models

learn an explicit probability density function from the training data and assume

that anomalous samples are located within low-density regions. Examples are

the application of Gaussian Mixture Models (GMMs) [2] for brain lesion detec-

tion as well as Bayesian U-Nets for OCT anomaly detection [15]. Although these

approaches are similar to ours, they are tailored to speciﬁc image modalities and

can thus not be directly and easily applied to our domain.

In this work, we introduce a novel and general method for anomaly detec-

tion which combines AE architectures with probabilistic graphical models called

Sum-Product Networks (SPNs). A recent powerful SPN-type called Random and

Tensorized SPN (RAT-SPN) [11] was chosen, as they are easy to integrate into

deep learning frameworks and are trained by GPU-based optimization. More

than that, while standard and variatonal AEs only provide either simple image

reconstruction based or the evidence lower bound based inference metrics for

anomaly detection, SPNs, on the other hand, allow exact and eﬃcient likeli-

hood inference by imposing special structural constraints on modeling the data

distribution. Therefore, we compare diﬀerent AE architectures combined with

RAT-SPNs on the unsupervised lesion and calciﬁcation detection use-case in

public mammography scans and demonstrate improved performance.

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：

AnomalyDetectionusingGenerativeModelsandSum-ProductNetworksinMammographyScansMarcDietrichstein1?,DavidMajor1?,MartinTrapp2,MariaWimmer1,DimitriosLenis1,PhilipWinter1,AstridBerg1,TheresaNeubauer1,andKatjaBühler11VRVisZentrumfürVirtualRealityundVisualisierungForschungs-GmbH,Vienna,Austriadavid.major@v...

展开>> 收起<<

Anomaly Detection using Generative Models and Sum-Product Networks in Mammography Scans Marc Dietrichstein1 David Major1 Martin Trapp2 Maria Wimmer1.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！

Anomaly Detection using Generative Models and Sum-Product Networks in Mammography Scans Marc Dietrichstein1 David Major1 Martin Trapp2 Maria Wimmer1

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: