PriorNet lesion segmentation in PET-CT including prior tumor appearance information Simone Bendazzoli12and Mehdi Astaraki13

2025-05-02 0 0 1.15MB 8 页 10玖币

侵权投诉

PriorNet: lesion segmentation in PET-CT

including prior tumor appearance information

Simone Bendazzoli12 and Mehdi Astaraki13

1Department of Biomedical Engineering and Health Systems , KTH Royal Institute

of Technology, Stockholm , Sweden

2Department of Clinical Science, Intervention and Technology, Karolinska Institutet,

Stockholm, Sweden

3Department of Oncology Pathology, Karolinska Institutet, Stockholm, Sweden

Abstract. Tumor segmentation in PET-CT images is challenging due

to the dual nature of the acquired information: low metabolic informa-

tion in CT and low spatial resolution in PET. U-Net architecture is the

most common and widely recognized approach when developing a fully

automatic image segmentation method in the medical ﬁeld. We proposed

a two-step approach, aiming to reﬁne and improve the segmentation per-

formances of tumoral lesions in PET-CT. The ﬁrst step generates a prior

tumor appearance map from the PET-CT volumes, regarded as prior

tumor information. The second step, consisting of a standard U-Net, re-

ceives the prior tumor appearance map and PET-CT images to generate

the lesion mask. We evaluated the method on the 1014 cases available

for the AutoPET 2022 challenge, and the results showed an average Dice

score of 0.701 on the positive cases.

Keywords: Tumor segmentation ·deep-learning ·PET-CT.

1 Introduction

Tumor lesion segmentation is one of the primary tasks performed on PET-CT

scans in oncological practice. The main aim is to identify and delineate the

tumor region, enabling a quantitative assessment, performing feature extraction

and planning the treatment strategy accordingly.

U-Net model [5] training is the most common supervised deep learning ap-

proach that yielded promosing results for diﬀerent medical image segmentation

tasks [2] [6]. We propose a 2-step approach, where the ﬁrst step aims at gener-

ating and providing prior tumor appearance information to the second step, a

conventional U-Net employed for the tumor segmentation.

2 Method description

The proposed method consists of two main modules. First, inspired by our recent

Normal Appearance Autoencoder (NAA) model [1], the appearance of healthy

arXiv:2210.02203v1 [eess.IV] 5 Oct 2022

2 S. Bendazzoli, M. Astaraki

anatomies from PET-CT images is learned by training an inpainting model.

In speciﬁc, a Partial Convolution Neural Network (PCNN) [4] was employed

to capture the distributions of healthy anatomies. Prior information regarding

the appearance of the tumors was, then, estimated by calculating the residuals

between the reconstructed fake healthy images from the tumoral ones. Second,

the prior information highlighting the presence of candidate tumoral regions was

added as an additional channel into a supervised segmentation network in order

to guide the attention of the model to the candidate regions.

2.1 Estimating the tumor appearance

Estimating the prior information regarding the appearance of tumors can be

achieved by, ﬁrst, modeling the healthy anatomies and then detecting the tumors

as anomalies. To model the distribution of complicated healthy anatomies from

whole-body PET-CT volumes, we employed a PCNN model as a robust inpaint-

ing network. This inpainting model can replace the pathological regions with the

characteristics of nearby healthy tissues and generate plausible pathology-free

images with realistic-looking and anatomically meaningful visual patterns. This

can be achieved by the following two steps: 1) forcing the model to learn the

appearance of healthy anatomies, and 2) guiding the model to inpaint only the

tumoral regions.

To learn the attributes of healthy anatomies, healthy image slices from the PET-

CT dataset were employed as the training set of the inpainting model. In speciﬁc,

more than 30000 healthy image slices were used for the training set, while the

pathological slices were employed for the testing set. Considering the large di-

versity in the shape, size, and location of the tumoral regions, random irregular

shapes were synthesized by combining regular geometrical shapes, including cir-

cles, squares, and ellipses, to corrupt the healthy images. The PCNN model is

trained until it ﬁlls the random holes and replaces the gaps with meaningful

anatomical and imagery patterns. The objective function of the PCNN model is

constructed based on several loss terms, including per-pixel loss, perceptual loss,

style loss, and total variation loss. This multi-objective optimization leverage

the quality of the inpainted images and reconstructs high-quality image while

preserving the anatomical details. The performance of the PCNN model was

evaluated by quantifying the following metrics: peak signal-to-noise ratio, mean

square error, and structural similarity index.

During the training step, the model learns to ﬁll the random holes with the

attributes of the nearby healthy tissues. This process enforces the model to

model the distribution of the healthy anatomies. Therefore, in the test phase,

the learned model can be used to replace the tumoral regions with visual char-

acteristics of healthy tissues. However, such a tumoral removal step essentially

requires tumoral masks. While in the NAA model, a second autoencoder model

was employed to remove the tumors automatically, in this study, we utilized

the learned PCNN model to directly inpaint the tumoral regions by gaining

from the hyperintensity patterns of PET images. In speciﬁc, tumoral regions

in PET volumes often appear with higher FDG uptake with respect to nearby

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：

PriorNet:lesionsegmentationinPET-CTincludingpriortumorappearanceinformationSimoneBendazzoli12andMehdiAstaraki131DepartmentofBiomedicalEngineeringandHealthSystems,KTHRoyalInstituteofTechnology,Stockholm,Sweden2DepartmentofClinicalScience,InterventionandTechnology,KarolinskaInstitutet,Stockholm,Sweden...

展开>> 收起<<

PriorNet lesion segmentation in PET-CT including prior tumor appearance information Simone Bendazzoli12and Mehdi Astaraki13.pdf

共8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！

PriorNet lesion segmentation in PET-CT including prior tumor appearance information Simone Bendazzoli12and Mehdi Astaraki13

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: