深度卷積神經網路於牙科X光影像進行自動標記及齒顎矯正分析

Deep Convolution Neural Network for Automatic Cephalometric Landmark Detection and Orthodontics Analysis

楊必聖(Bi-Sheng Yang)

指導教授 ： 丁慧枝

共同指導教授 ： 許士彥(Shih-Yen Hsu)

義守大學/智慧科技學院/資訊工程學系/碩士(2022年)

摘要

臨床牙科側頭顱量測分析目前仍以人工方式標示參考點規劃齒顎矯正計畫，該方式常受主觀判斷與臨床經驗影響量測數據。本研究目的透過深度學習技術，進行頭顱X光標示點自動標記，提供齒顎矯正分析輔助資訊，改善人工作業耗時及主觀判斷之影響。
本研究透過深度神經網路進行前後位及側頭顱X光影像測量分析，提供自動標記及標記點位距離、角度量測與計算。在自動標記技術開發上，採用深度神經網路YOLO模型改善自動標記效率及誤差問題，同時經由參考點與點之間自動測量，自動計算臨床常用指標，最後透過多個統計指標評估模型之準確性與效能。
研究探討之15個標記點，經由深度神經網路訓練獲得自動標記模型，其平均精確度為90%、最小為74%、最大為97%；另外在自動量測分析上，分別計算五組臨床常用之基準線，結果顯示其中兩組自動量與手動量測值測誤差小於2 mm以內。
本研究提出之方法可提供臨床牙科側顱自動化分析，不僅可進行標示點定位，同時可進一步自動計算各個基準線進行數據分析，輔助矯正分析與改善人工作業之影響。

關鍵字

頭顱X光影像 ； 齒顎矯正分析

並列摘要

In clinical, the dental cephalometric measurement and analysis is still through manually marked the reference point in the orthodontic plan. This method which is often affected by subjective judgment and clinical experience. Hence, the purpose of this study is to use deep learning technology to automatically mark points on cephalometric. It can provide assistant information for orthodontic analysis, and improve the time-consuming and subjective judgment of manual operations.
In this study, cephalometric images were measured and analyzed through a deep neural network to provide automatic marking and measurement. Including the distance and angle of marked points. In the development of automatic labeling technology. The deep neural network YOLO model is used to improve the efficiency and error of automatic labeling. At the same time, through automatic measurement between reference points to points, it can automatically calculate the commonly indicators in clinical. Finally evaluate the accuracy and performance of the model through multiple statistical indicators.
For the 15 markers discussed in the study, an automatic label model was obtained through deep neural network training. The result of average precision was 90%. The minimum and maximum was 74%, and 97% with respectively. In addition, in the automatic measurement analysis of five groups of clinical. The results show that the measurement error between two sets of automatic and manual measurement values is less than 2 mm.
The method proposed in this study can provide automatic analysis of clinical dental cephalometric, which can not only locate marked points, but also further automatically calculate each baseline for data analysis. It can assist in correction analysis and improve the impact of manual operations.

並列關鍵字

Cephalometric ； Orthodontic Analysis ； YOLO

參考文獻

1. Tenti, F. Cephalometric analysis as a tool for treatment planning and evaluation. The European Journal of Orthodontics, 1981. 3(4): p. 241-245.

2. Downs, W.B. Analysis of the dentofacial profile. The Angle Orthodontist, 1956. 26(4): p. 191-212.

3. Wahl, N. Orthodontics in 3 millennia. Chapter 8: The cephalometer takes its place in the orthodontic armamentarium. American journal of orthodontics and dentofacial orthopedics, 2006. 129(4): p. 574-580.

4. Abdullah, R. et al., Steiner cephalometric analysis: predicted and actual treatment outcome compared. Orthodontics & craniofacial research, 2006. 9(2): p. 77-83.

5. 鄭信忠、裘薇藝, 基本齒顎矯正學. 中華民國(台灣): 合記圖書出版社.1980, p84-95.