Assessment of alveolar bone quality in cone-beam computed tomographic images for dental implant placement using deep learning techniques

การประเมินคุณภาพกระดูกเบ้าฟันในภาพรังสีส่วนตัดอาศัยคอมพิวเตอร์ชนิดโคนบีมเพื่อการฝังรากเทียมโดยใช้การเรียนรู้เชิงลึก

Name: Thatphong Pornvoranant

LCSH: Teeth -- Roots

LCSH: Endodontics

LCSH: Teeth -- Radiography, Medical

LCSH: Teeth -- bones

Abstract: The characteristics of alveolar bone play a crucial role in treatment planning for dental implants. The Lekholm and Zarb Classification (1985) categorizes the cross-sectional morphology of alveolar bone into four types based on cortical bone thickness and trabecular bone density as observed in cone-beam computed tomography (CBCT) images. However, this classification lacks precise numerical criteria for type differentiation. This research aimed to evaluate various deep learning (DL) models for classifying alveolar bone characteristics from CBCT images and to compare the performance of the best-performing DL model with postgraduate students and experienced dental specialists. The researchers prepared 1,100 CBCT cross-sectional images, which were categorized into four types based on the Lekholm and Zarb Classification by two experienced dental radiologists. The dataset was divided into 1,000 training images and 100 testing images. Five DL models—AlexNet, GoogLeNet, ResNet-50, DenseNet-201, and Inception-ResNet-v2—were trained over 50 epochs using learning rates (LR) of 0.001 and 0.0001. The results showed that Inception-ResNet-v2 (LR=0.001) achieved the highest accuracy (86%), followed by GoogLeNet (LR=0.0001, 82%) and DenseNet-201 (LR=0.0001, 81%). The Receiver Operating Characteristic (ROC) curve analysis revealed that GoogLeNet, ResNet-50, and Inception-ResNet-v2 had area-under-the-curve (AUC) values exceeding 0.9, while all models achieved AUC values above 0.8. Thus, Inception-ResNet-v2 (LR=0.001) demonstrated the best performance in classifying alveolar bone characteristics from CBCT images. The test data were also evaluated by 23 postgraduate students and two experienced implantologists. The results showed that Inception-ResNet-v2 outperformed the first implantologist (75% accuracy), postgraduate students (68% accuracy), and the second implantologist (64% accuracy). These findings suggest that DL models, particularly Inception-ResNet-v2, exhibit superior performance in classifying alveolar bone characteristics compared to postgraduate students and experienced dental specialists.

Chiang Mai University. Library

Address: CHIANG MAI

Email: cmulibref@cmu.ac.th

Name: Pisaisit Chaijareenont

Role: Advisor

Name: Pimduen Rungsiyakull

Role: Advisor

Name: Wannakamon Panyarak

Role: Advisor

Name: Kittichai Wantanajittikul

Role: Advisor

Created: 2025

Modified: 2025-06-23

Issued: 2025-06-23

วิทยานิพนธ์/Thesis

application/pdf

eng

DegreeName: Master of Science

Level: Master's Degree

Descipline: Dentistry

Grantor: Chiang Mai University

©copyrights Chiang Mai University

ลำดับที่.	ชื่อแฟ้มข้อมูล	ขนาดแฟ้มข้อมูล	จำนวนเข้าถึง	วัน-เวลาเข้าถึงล่าสุด
1	650931044.pdf	2.01 MB

ใช้เวลา

0.043347 วินาที