1 Dual-Stage Deeply Supervised Attention-based Convolutional Neural Networks for Mandibular

2025-04-28 0 0 9.52MB 7 页 10玖币

侵权投诉

Dual-Stage Deeply Supervised Attention-based

Convolutional Neural Networks for Mandibular

Canal Segmentation in CBCT Scans

Muhammad Usman1,2, Azka Rehman1, Amal Saleem1, Rabeea Jawaid3, Shi-Sub Byon1, Sung-Hyun Kim1,

Byoung Dai Lee3, Byung-il Lee1, and Yeong-Gil Shin2

1Center for Artiﬁcial Intelligence in Medicine and Imaging, HealthHub Co. Ltd., Seoul, 06524, South Korea

2Seoul National University, Seoul, Republic of Korea

3Division of AI and Computer Engineering, Kyonggi University, Suwon, Republic of South Korea

Abstract—Accurate segmentation of mandibular canals in

lower jaws is important in dental implantology. Medical experts

determine the implant position and dimensions manually from

3D CT images to avoid damaging the mandibular nerve inside

the canal. In this paper, we propose a novel dual-stage deep

learning-based scheme for the automatic segmentation of the

mandibular canal. Particularly, we ﬁrst enhance the CBCT scans

by employing the novel histogram-based dynamic windowing

scheme, which improves the visibility of mandibular canals. After

enhancement, we design 3D deeply supervised attention U-Net

architecture for localizing the volumes of interest (VOIs), which

contain the mandibular canals (i.e., left and right canals). Finally,

we employed the multi-scale input residual U-Net architecture

(MS-R-UNet) to segment the mandibular canals using VOIs

accurately. The proposed method has been rigorously evaluated

on 500 scans. The results demonstrate that our technique out-

performs the current state-of-the-art segmentation performance

and robustness methods.

Index Terms—Mandibular Canal, 3D Segmentation, Jaw Lo-

calization

I. INTRODUCTION

Inferior alveolar nerve (IAN), also known as mandibular

canal, is the most critical structures in the mandible region that

supplies sensation to the lower teeth. The sensation, provided

to lips and chin is via the mental nerve which passing through

the mental foramen [1]. One of a very critical steps in implant

placement, third molar extraction, and various other craniofa-

cial procedures including orthognathic surgery, is determining

the position of the mandibular canal. Patients may experience

aches and pain and temporary paralysis if the mandibular canal

get injured [2] [3] during any of these process. Localization

of the mandibular canal is important not only for diagnosis

of vascular and neurogenic diseases associated with the nerve,

but also for diagnosis of lesions near the mandibular canal,

and planning of oral and maxillofacial procedures. Therefore,

preoperative treatment planning and simulation are necessary

to avoid nerve injury. The identiﬁcation of exact location of

can assist in achieving the planning strategy required for the

task at hand [4].

One of the most frequently used three-dimensional (3D)

imaging modalities for preoperative treatment planning and

postoperative assessment in dentistry is Cone Beam Computed

Tomography which is also known as CBCT [5]. The CBCT

volume is reconstructed using projection images realized from

different angles with a cone-shaped beam and stored as a

sequence of axial images [6]. A clinical replacement is multi-

detector computed tomography (MDCT), but its application

is limited by high radiation dose and insufﬁcient spatial

resolution. In contrast, the CBCT allows more precise imaging

of hard tissues in the dentomaxillofacial area and its effective

radiation dosage is lower than that of the MDCT1. In addition,

CBCT is inexpensive and readily available. However, in prac-

tice, there are certain challenges associated with mandibular

canal segmentation from CBCT images, such as inaccurate

density and large amount of noise [7].

Surgical planning and pre-surgical examination are crucial

in dental clinics. One of the standard imaging tools used for

such assessments and planning is a panoramic radiograph,

constructed from a dental arch to provide all the relevant infor-

mation in a single view. These radiographs bear disadvantages

such as difﬁculties in determining the 3D rendering of an entire

canal and connected nerves [8]. One of the most common

approaches for preoperative assessment is to annotate the canal

in 3D images to produce the segmentation of the canal. This

kind of manual annotation is a very knowledge-intensive, time-

consuming, and tedious task. Thus, there is a need for a tool to

assist the radiologist and reduce the burden by using automatic

or semi-automatic segmentation of the canal.

Kwak et. al. [9] studied different models based on 2D and

3D techniques such as on 2D SegNet, and 2D and 3D U-Nets.

Their study also involved detailed pre and post-processing

steps including thresholding of teeth as well as bones. Jaskari

et. al. [10] presented an FCNN-based model to extract IAN.

Dhar et. al. [11] used a model based on 3D UNet to segment

the canal. They used pre-processing techniques to generate

the center lines of the mandibular canals and used them as

ground truths in the training process. Verhelst et. al. [12]

used a patch-based technique to localize the jaw and then

used the 3d UNet model to segment the canal in that ROI.

Lahoud et. al. [13] ﬁrst coarsely segmented the canal, and

performed ﬁne segmentation of the canal on patches that

arXiv:2210.03739v4 [eess.IV] 2 Nov 2022

Fig. 1: Proposed dual-stage scheme for mandibular canal segmentation, describing the models architectures utilized at each stage. Firstly, deeply supervised attention UNet model

is used for jaw localization which coarsely segments the canal. This coarsely segmented canal is further utilized to extract VOIs which is used to produce the ﬁne segmentation of

mandibular canals (i.e., left and right canals) by employing Residual UNet with multi-scale inputs.

are extracted based on the coarse segmentation. The network

utilized 3D UNet with skip connections. In all the above

studies, the models utilized were 3D UNet. Verhelst et al.

[12] and Lahoud et al. [13] utilized the localization technique

and then segmented the canal. However, they used patches for

ﬁne segmentation i.e. canal is divided into multiple patches

before segmentation. This reduces the visibility of the model

resulting in high chances of error. Other methodologies do not

take into account the varying size of the canal while training

the model. Previously work has been done in deep learning

In this study, we propose a cascade technique to segment the

mandibular canal in 3D CBCT scans. We ﬁrst localize the

jaw using a naively segmented canal in the form of volume

of interest (VOIs). After that, we divide the canal into left

and right parts and trained a multi-scale input Residual UNet

model for segmenting the canal. The purpose of using multi-

scale input patches is to take into account the varying size of

canals.

The rest of the paper is organized as follows; In Section

I, we present background and related work. In Section II, the

detail on each step of our proposed method as well as the

materials is described. In Section III, we present the obtained

results and comparison of our study. Finally, we analyze and

discuss our work in Section IV and conclude in Section V.

II. MATERIALS AND METHODS

A. Study Design

The objective of this study is to design a deep-learning

approach for automatic mandibular canal segmentation. The

study design consists of pre-processing, model training, and

post-processing each discussed in detail in the following

sections. The detailed network design is discussed in section

II-D Network Architecture. The network was validated on 500

scans.

B. Data Aquisition

For this study, 1010 dental CBCT scans were obtained from

the PACS of a dental hospital. The data was annotated in

two stages; annotation in the ﬁrst stage was carried out by

28 trained medical students and at the second stage 6 doctors

conducted the validation of annotated data. The CBCT scans

were in DICOM format with voxel spacing ranging from

0.3mm to 0.39mm. The annotated data was available as a

set of ﬂoating point polygon coordinates each for the left and

right canal separately, stored in JSON ﬁle format for each

patient. The spatial resolution of scans ranged from 512 ×

512 ×460 voxels to 670 ×670 ×640 voxels. The CBCT

scans are divided into three different types based on Hounsﬁeld

unit (HU) values ranges. The three different scans have ranges

from -1000 to +1000, -1000 to +2000, and 0 to 5000 HU .

However, not all dataset is utilized for testing and training.

Many experiments were conducted using 100, 200, 300, and

400 scans and tested on 500 samples.

C. Data Pre-processing

Although CT scans follow a worldwide standard for ranges

of HU values for different body parts like teeth, gums, bones,

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：

1Dual-StageDeeplySupervisedAttention-basedConvolutionalNeuralNetworksforMandibularCanalSegmentationinCBCTScansMuhammadUsman1,2,AzkaRehman1,AmalSaleem1,RabeeaJawaid3,Shi-SubByon1,Sung-HyunKim1,ByoungDaiLee3,Byung-ilLee1,andYeong-GilShin21CenterforArticialIntelligenceinMedicineandImaging,HealthHubCo....

展开>> 收起<<

1 Dual-Stage Deeply Supervised Attention-based Convolutional Neural Networks for Mandibular.pdf

共7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！

1 Dual-Stage Deeply Supervised Attention-based Convolutional Neural Networks for Mandibular

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: