運用深度學習針對牙科環景全口X光影像進行品質管控之研究

臨床上常利用環景全口X光（panoramic x-ray, PANO）對病人的齒槽狀況進行評估。一張擺位正確的PANO影像，可提供醫師完善的判讀資訊，故在影像產生後傳送至醫師前，應進行品質控制（quality control, QC）以確保影像符合診斷標準。然而，過去對影像進行QC的方式，通常是由放射師進行目視檢查，此種仰賴雙眼的品管方法可能存在缺陷，如：1.目視檢查取決於放射師對影像判讀的感知與經驗，影響一致性。2.工作負擔增加造成疲勞感增加，易使判斷能力下降。3.大量影像進行品管較費時。本研究自院內醫學影像儲存通訊系統（Picture archiving and communication system, PACS）選取PANO影像553張，使用7種卷積神經網路（convolutional neural network, CNN）模型：alexnet、googlenet、mobilenet、resnet50、resnet101、vgg16及vgg19，並分別結合三種機器學習分類器：邏輯迴歸（logistic regression）、貝氏分類器（naïve Bayes classifier）、支持向量機（support vector machine, SVM），共21種組合分類模型，利用人工智慧（artificial intelligence, AI）進行自動辨識影像中常見的各種擺位錯誤，並比較各模型分類結果找出最佳分類模型。模型效能評比項目包含準確度（accuracy）及一致性係數（kappa）。實驗結果證明，resnet101結合logistic regression分類器模型之Kappa平均值為0.211，為最佳分類模型。

關鍵字

PANO ； CNN ； QC

並列摘要

In clinical practice, panoramic x-ray (PANO) is often used to evaluate the alveolar condition of patients. A correctly positioned PANO image can provide physicians with complete interpretation information. Therefore, quality control (QC) should be performed to ensure that the images meet the diagnostic criteria before being transmitted to physicians. However, in the past, the way of QC of images was usually visual inspection by radiologists. This quality control method relying on both eyes may have defects, such as 1. Visual inspection depends on the radiologist's perception and experience of image interpretation, Affect consistency. 2. The increased workload results in increased fatigue, which efficiently reduces the ability to judge. 3. Quality control of a large number of images is time-consuming. In this study, 553 PANO images were selected from the in-hospital medical image storage and communication system (PACS), and seven convolutional neural networks (CNN) models were used: alexnet, googlenet, mobilenet, resnet50, resnet101, vgg16 and vgg19, and combine three machine learning classifiers respectively: logistic regression (logistic regression), Bayes classifier (naïve Bayes classifier), support vector machine (support vector machine, SVM), a total of 21 combined classification models, using Artificial intelligence (AI) automatically identifies various common placement errors in images, and compares the classification results of each model to find the best classification model. Model performance evaluation items include accuracy (accuracy) and consistency coefficient (Kappa). The experimental results show that the average Kappa of resnet101 combined with the logistic regression classifier model is 0.211, which is the best classification model.

並列關鍵字

PANO ； CNN ； QC

主題瀏覽