CMU-NET A STRONG CONVMIXER-BASED MEDICAL ULTRASOUND IMAGE SEGMENTATION NETWORK Fenghe Tang1 Lingtao Wang1 Chunping Ning2 Min Xian3 and Jianrui Ding1

2025-04-29 0 0 3.07MB 5 页 10玖币

侵权投诉

CMU-NET: A STRONG CONVMIXER-BASED MEDICAL ULTRASOUND IMAGE

SEGMENTATION NETWORK

Fenghe Tang1, Lingtao Wang1, Chunping Ning2, Min Xian3, and Jianrui Ding1?

1School of Computer Science and Technology, Harbin Institute of Technology, Harbin, China.

2Ultrasound Department, The Afﬁliated Hospital of Qingdao University, Qingdao, China.

3Department of Computer Science, University of Idaho, Idaho Falls, ID 83401, USA

ABSTRACT

U-Net and its extensions have achieved great success in med-

ical image segmentation. However, due to the inherent local

characteristics of ordinary convolution operations, U-Net en-

coder cannot effectively extract global context information.

In addition, simple skip connections cannot capture salient

features. In this work, we propose a fully convolutional seg-

mentation network (CMU-Net) which incorporates hybrid

convolutions and multi-scale attention gate. The ConvMixer

module extracts global context information by mixing fea-

tures at distant spatial locations. Moreover, the multi-scale

attention gate emphasizes valuable features and achieves ef-

ﬁcient skip connections. We evaluate the proposed method

using both breast ultrasound datasets and a thyroid ultrasound

image dataset; and CMU-Net achieves average Intersection

over Union (IoU) values of 73.27% and 84.75%, and F1

scores of 84.16% and 91.71%. The code is available at

https://github.com/FengheTan9/CMU-Net.

Index Terms—Ultrasound image segmentation, U-Net,

ConvMixer, multi-scale attention

1. INTRODUCTION

Ultrasound imaging is non-invasive, non-radiative, cost effec-

tive and real-time for disease detection. It has been widely

used in the detection of breast tumor, thyroid nodule, fetal,

and gonadal tissue [1]. Conventional disease detection using

ultrasound images depended on manual labeling, which is la-

borious and time-consuming. The results were sensitive to

subjective factors such as radiologists’ experience and men-

tal state. With the emergence of deep learning approaches,

automatic medical image segmentation has been rapidly de-

veloped in the ﬁeld of image analysis, which can effectively

overcome the above limitations.

The segmentation of medical ultrasound images is chal-

lenging. As shown in Fig.1, most ultrasound images only

contain one lesion, and binary segmentation approaches could

be applied. But the sizes, shapes, and texture patterns of le-

sions from different cases vary greatly. In addition, ultrasound

images usually have low contrast, high speckle noises, and

(a) Breast ultrasound image

(b) Thyroid ultrasound image

Fig. 1. Examples of ultrasound image segmentation. The

pink contours denote lesion boundaries.

shadow artifacts, and conventional segmentation approaches

used to perform poorly.

U-Net [2] has an encoder-decoder based segmentation ar-

chitecture. It can effectively ﬁt scarce medical image data. In

recent years, many medical segmentation networks based on

U-Net have been proposed, such as U-Net++ [3], Attention

U-Net [4], Unet3+ [5], and UNeXt [6]. Due to the locality of

ordinary convolution operations in U-Net, a number of net-

works based on Transformer [7] have recently been applied to

medical image segmentation tasks [8, 9, 10] to extract global

information of images. TransUnet [8] employed Vit [11] to

obtain global context with CNN, but it required massive med-

ical images and computing overhead.

In order to solve the limitation of ordinary convolution lo-

cality, Trockman et al. proposed the ConvMixer [12] which

used large convolutional kernels to mix remote spatial loca-

tions to obtain global context information. Compared with

the Transformer, the ConvMixer is more efﬁciency and adapt

to computer vision tasks better, and its computational over-

head is less than that of the self-attention mechanism.

Inspired by the U-shape architectural design and Con-

vMixer, we propose an efﬁcient fully convolutional image

arXiv:2210.13012v4 [eess.IV] 10 Nov 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：

CMU-NET:ASTRONGCONVMIXER-BASEDMEDICALULTRASOUNDIMAGESEGMENTATIONNETWORKFengheTang1,LingtaoWang1,ChunpingNing2,MinXian3,andJianruiDing1?1SchoolofComputerScienceandTechnology,HarbinInstituteofTechnology,Harbin,China.2UltrasoundDepartment,TheAfliatedHospitalofQingdaoUniversity,Qingdao,China.3Departmen...

展开>> 收起<<

CMU-NET A STRONG CONVMIXER-BASED MEDICAL ULTRASOUND IMAGE SEGMENTATION NETWORK Fenghe Tang1 Lingtao Wang1 Chunping Ning2 Min Xian3 and Jianrui Ding1.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！

CMU-NET A STRONG CONVMIXER-BASED MEDICAL ULTRASOUND IMAGE SEGMENTATION NETWORK Fenghe Tang1 Lingtao Wang1 Chunping Ning2 Min Xian3 and Jianrui Ding1

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: