深度學習於超音波影像辨識之應用：以膽囊為例

Research on the Application of Deep Learning in Automatic Identification from Dynamic Ultrasound Imaging: Using Gallbladder as a Target

蘇心悅(Hsin-Yueh Su) ； 王友立(Yu-Li Wang) ； 黃詠暉(Yung-Hui Huang) ； 陳泰賓(Tai-Been Chen)

《臺灣醫事放射期刊》 10卷1期 (2022/07) Pp. 1-7

https://doi.org/10.6717/JTMRT.202207_10(1).0001

摘要

超音波是目前應用於腹部急症常見的檢查工具之一，在臨床實務操作上具有非侵入性、無輻射線、低成本、操作靈活且能提供即時影像的優點。但由於不同操作者之間對影像品質的要求、造影參數設定、操作經驗及受檢者的配合程度皆不盡相同，造成超音波影像之對比度與明暗度不一致，使診斷鑑別困難。因此，運用合適的輔助工具以快速取得標的正確的超音波影像，可提升臨床檢查之效率並降低操作門檻。本研究利用錄製30位受檢者之動態超音波影像（MPEG-4格式）轉換成單張靜態影像（JPG格式），挑選出共12,768張腹部超音波靜態影像，使用兩種卷積神經網路（convolutional neural network, CNN）模型：GoogLeNet及ResNet50，並分別結合三種機器學習分類器：線性回歸（linear regression）、貝氏分類器（Naïve Bayes）、支持向量機（support vector machine, SVM），共六種組合分類模型，進行自動辨識腹部超音波靜態影像視野內之膽囊影像，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，ResNet50結合SVM分類器之模型最佳，其準確度及kappa值皆大於90%。本研究利用錄製人體腹部超音波之動態影像，經由人工智慧（artificial intelligence, AI）建構影像自動辨識及分類模型，可輔助經驗不足的操作者能在較短時間內獲得最佳影像，並滿足無法配合閉氣者之造影需求。

關鍵字

超音波 ； GoogLeNet ； ResNet50

並列摘要

Ultrasound is one of the common inspection tools commonly used in abdominal emergencies. It has the advantages of being non-invasive, no radiation, low cost, flexible operation, and can provide real-time images in clinical practice. However, the quality of ultrasound images is often hampered by the complexity of parameter setting, inexperienced operators, and uncooperative patients resulting in difficulties in diagnosis. In this study, 30 subjects recorded the dynamic ultrasound images (MPEG-4 format) and converted them into a single static image (JPG format) to obtain 12,768 ultrasound images from the abdomen. Two convolutional neural networks (CNN) models, including GoogLeNet and ResNet50, were combined with three machine learning classifiers: linear regression, Naïve Bayes, and support vector machine (SVM). Six combined classification models were applied to automatically obtain the ultrasound static images of the gallbladder in the field of view. The accuracy and kappa value among the six models were compared, respectively. The results showed that ResNet50 combined with the SVM classifier was the best. The accuracy and kappa value were both more significant than 90%. Furthermore, the disadvantages of ultrasound can be overcome by using dynamic recording images of the abdomen through artificial intelligence (AI) to construct automatic image recognition and classification models. These models can be used to assist inexperienced operators and to obtain high-quality images in a shorter time.

並列關鍵字

Ultrasound ； GoogLeNet ； ResNet50