人工智慧應用於錐狀射束電腦斷層掃描進行牙科植體治療規劃之研究

Abstract

研究目標 本研究目標為建立一個透過錐狀射束電腦斷層掃描cone-beam computerized tomography (CBCT) 影像自動判讀植牙骨缺損型態分類的深度學習模型，以提供植牙治療計畫建議。

研究方法 使用177名患者（83名男性和 94名女性，年齡 56.49 ± 12.76 歲）的 630 個下顎CBCT影像切片，虛擬植入植體（直徑 4-5 毫米，長度 10 毫米），切片皆根據虛擬植體的長軸定向並透過人工標記。齒槽嵴缺損型態分成五類，採用ResNet-50卷積神經網路模型進行訓練。

研究結果 該模型在未經預處理影像切片的準確度為98.91 ± 1.45%，優於經預處理影像的準確度92.85 ± 2.60%。以各缺損型態分類的模型性能而言，未經預處理的影像切片之準確率和 F1 分數分別為 99.05%–99.84% 和 97.30%–99.66%，經預處理的切片之準確度和 F1 分數分別為 96.03%–98.89% 和 87.92%–96.71%。

結論 透過深度學習判讀CBCT影像，無需經過預處理即可高準確率識別出植牙骨缺損型態分類。

Objectives The aim of this study was to establish a deep learning model for the automatic identification of the ridge deficiency around dental implants based on an image slice from cone-beam computerized tomography (CBCT) in order to provide recommendations for dental implant treatment plan.

Materials and Methods A total of 630 CBCT image slices of the mandible from respective implants in 117 patients (83 males and 94 females, age 56.49 ± 12.76 years) were used as the datasets in this study. Virtual implant fixtures (4-5 mm diameter, 10mm length) were placed in the implant treatment planning software. In each implant, one slice crossing the central long-axis implant was chosen and the classifications of edentulous ridges were labeled. Alveolar ridge defect types were divided into five classifications. A convolutional neural network with ResNet-50 architecture was employed for deep learning.

Results The model achieved an accuracy of 98.91 ± 1.45% on the unpreprocessed image slices and was found to be superior to the accuracy of 92.85 ± 2.60% observed on the preprocessed slices. In terms of model performance for each defect type classification, the accuracy and F1 scores for unpreprocessed image slices were 99.05%–99.84% and 97.30%–99.66%, respectively, and 96.03%–98.89% and 87.92%–96.71%, respectively, for preprocessed slices.

Conclusion By identifying CBCT images through deep learning, the ridge deficiency around dental implants could be correctly classified with high accuracy without preprocessing.