ConnectedUNets Mass Segmentation from Whole Mammographic Images Prithul Sarker Sushmita Sarker George Bebis and Alireza Tavakkoli

2025-05-02 1 0 462.5KB 12 页 10玖币

侵权投诉

ConnectedUNets++: Mass Segmentation from

Whole Mammographic Images

Prithul Sarker∗, Sushmita Sarker∗, George Bebis, and Alireza Tavakkoli

Department of Computer Science and Engineering,

University of Nevada, Reno, United States

{prithulsarker, sushmita}@nevada.unr.edu

Abstract. Deep learning has made a breakthrough in medical image

segmentation in recent years due to its ability to extract high-level fea-

tures without the need for prior knowledge. In this context, U-Net is

one of the most advanced medical image segmentation models, with

promising results in mammography. Despite its excellent overall perfor-

mance in segmenting multimodal medical images, the traditional U-Net

structure appears to be inadequate in various ways. There are certain

U-Net design modiﬁcations, such as MultiResUNet, Connected-UNets

and AU-Net, that have improved overall performance in areas where the

conventional U-Net architecture appears to be deﬁcient. Following the

success of UNet and its variants, we have presented two enhanced ver-

sions of the Connected-UNets architecture: ConnectedUNets+and Con-

nectedUNets++. In ConnectedUNets+, we have replaced the simple skip

connections of Connected-UNets architecture with residual skip connec-

tions, while in ConnectedUNets++, we have modiﬁed the encoder decoder

structure along with employing residual skip connections. We have eval-

uated our proposed architectures on two publicly available datasets, the

Curated Breast Imaging Subset of Digital Database for Screening Mam-

mography (CBIS-DDSM) and INbreast.

Keywords: Convolutional Neural Network ·Mammogram ·Semantic

Segmentation ·U-Net ·ConnectedU-Nets ·MultiResUNet

1 Introduction

Breast cancer is the most frequent type of cancer that causes death in women,

with 44,130 instances reported in the United States in 2021 [1]. The need for

frequent mammography screening has been stressed in many studies in order

to reduce mortality rates by ﬁnding breast malignancies before they spread to

other normal tissues and healthy organs. A mammogram is an X-ray image of

the breast to record changes in the tissue. The disease is typically identiﬁed by

the presence of abnormal masses and microcalciﬁcations in mammograms [2], [3].

Radiologists examine a high number of mammograms on a daily basis looking for

abnormal lesions and assessing the location, shape, and type of any suspicious

*Prithul Sarker and Sushmita Sarker have equal contribution and are co-ﬁrst authors

arXiv:2210.13668v3 [eess.IV] 4 Nov 2022

2 Sarker et al.

area in the breast. This is an important procedure which requires high precision

and accuracy, however, it is still costly and prone to errors since detecting these

regions is challenging as their pixel intensities often coincide with normal tissue.

Deep learning advances [4], especially Convolutional Neural Networks (CNN)

[5], have shown a lot of promise in addressing these issues. Despite being a game-

changer in computer vision, CNN architectures have a key drawback: they require

an enormous amount of training data. In order to solve this problem, U-Net [6] is

introduced which is built on a simple encoder-decoder network with multiple sets

of CNN. Even with a limited quantity of labeled training data, U-Net has demon-

strated tremendous promise in segmenting breast masses, to the point where it

has become the de-facto standard in medical image segmentation [7]. In light of

the success of U-Net, various U-Net versions, such as Connected-UNets [8] and

AU-Net [9], have been proposed. These variations have demonstrated promising

results but appear to be ineﬃcient in terms of fully recovering the region of

interest in a given image.

In this work, we have proposed and experimented with two enhanced versions

of the Connected-UNets architecture. Although the proposed networks share an

architectural similarity, they are designed for diﬀerent use cases which are crucial

in real-world scenario. The proposed architectures take the entire mammogram

image as input and perform mass segmentation along with mass boundary ex-

traction. The main contributions of our work include:

1. We have proposed ConnectedUNets+and ConnectedUNets++, two novel and

improved versions of the Connected-UNets, by utilizing residual skip connec-

tions and enhanced encoder-decoder in order to achieve better convergence.

2. We have assessed the proposed architectures using full mammogram images

in contrast to the baseline model which operates on cropped images of cor-

rectly detected and classiﬁed masses by an object detection model.

3. We have experimented using all the images from two publicly available

datasets, the Curated Breast Imaging Subset of Digital Database for Screen-

ing Mammography (CBIS-DDSM) [10] and INbreast [11] for segmenting

the region of interest (ROI) of breast mass tumors.

To the best of our knowledge, our paper is the ﬁrst to address the shortcom-

ings of other papers’ methodologies and to conduct an unbiased comparison. We

applied the same loss function, optimizer, and image size to all architectures to

maintain objectivity. Additionally, to ensure a fair and accurate comparison, we

used full mammograms as input for all the models and adopted a comparable

preprocessing approach.

2 Related Works

U-Net [6], a deep learning network having an encoder-decoder architecture, is

among the most prominent deep neural networks commonly employed in med-

ical image segmentation. The network has a symmetric architecture, with an

encoder which extracts spatial information from the image and a decoder which

ConnectedUNets++: Mass Segmentation from Whole Mammographic Images 3

constructs the segmentation map from the encoded data. The encoder and de-

coder are linked by a series of skip connections which are the most innovative

component of the U-Net architecture since they enable the network to recover

spatial data that has been lost due to pooling procedures. Abdelhaﬁz et al. [12]

used a vanilla U-Net model to segment mass lesions in whole mammograms.

To segment suspicious regions in mammograms, Ravitha Rajalakshmi et al. [13]

presented a deeply supervised U-Net model (DS U-Net) combined with a dense

Conditional Random Field (CRF). Li et al. [14] proposed a Conditional Resid-

ual U-Net, named CRUNet, to improve the performance of the basic U-Net for

breast mass segmentation.

Though U-Net is among the most popular and successful deep learning mod-

els for biomedical image segmentation, several improvements are still possible.

Speciﬁcally, the concatenation of encoder and decoder features reveals a signif-

icant semantic gap despite the preservation of dispersed spatial features, which

is a shortcoming of the simple skip connections. To deal with this issue, Ibte-

haz et al. [15] proposed the MultiResUNet architecture by incorporating some

convolutional layers along with shortcut connections in U-Net. Instead of sim-

ply concatenating the feature maps from the encoder stage to the decoder stage,

they ﬁrst pass them through a chain of convolutional layers and then concatenate

them with the decoder features, which makes learning substantially easier. This

idea is inspired from the image-to-image conversion using convolutional neural

networks [16], where pooling layers are not favorable for the loss of informa-

tion. MultiResUNet has shown excellent results on diﬀerent biomedical images,

however, the authors did not experiment with mammograms.

Based on the U-Net architecture, Baccouche et al. [8] proposed an improved

architecture that connects two simple U-Nets, called Connected-UNets. In addi-

tion to the original idea of the U-Net architecture, which includes skip connec-

tions between the encoder and decoder networks, it cascades a second U-Net and

adds skip connections between the decoder of the ﬁrst U-Net and the encoder

of the second U-Net. The key idea was to recovering ﬁne-grained characteristics

lost in U-Net’s encoding process. However, the authors ﬁrst used YOLO [17] to

detect the location of masses in mammograms, and then applied their method

to segment only correctly localized masses. Such an approach is not optimum

in practical settings where it is desirable to simultaneously localize and segment

masses in whole mammograms rather than processing cropped mammograms.

Several modiﬁcations of the U-Net architecture have also been proposed by

incorporating an attention mechanism, which has shown to be extremely eﬀec-

tive in medical image segmentation. Oktay et al. [18] proposed a new attention

U-Net by adding an attention gate into the conventional U-Net. This enhanced

the accuracy of the predictions. However, they didn’t evaluate their model for

breast mass segmentation. Similarly, Li et al. [19] built an attention dense U-

Net for breast mass segmentation, which was compared to U-Net [6], Attention

U-Net [18], and DenseNet [20]. In another study by Sun et al. [9], an attention-

guided dense upsampling network, called AUNet, was built for breast mass seg-

mentation in full mammograms. The major drawback of the papers mentioned

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：

ConnectedUNets++:MassSegmentationfromWholeMammographicImagesPrithulSarker,SushmitaSarker,GeorgeBebis,andAlirezaTavakkoliDepartmentofComputerScienceandEngineering,UniversityofNevada,Reno,UnitedStatesfprithulsarker,sushmitag@nevada.unr.eduAbstract.Deeplearninghasmadeabreakthroughinmedicalimagesegmen...

展开>> 收起<<

ConnectedUNets Mass Segmentation from Whole Mammographic Images Prithul Sarker Sushmita Sarker George Bebis and Alireza Tavakkoli.pdf

共12页,预览3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！

ConnectedUNets Mass Segmentation from Whole Mammographic Images Prithul Sarker Sushmita Sarker George Bebis and Alireza Tavakkoli

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: