利用FCN模型架構對動態B-Mode總頸動脈超音波影像進行分割之研究

A Fully Convolutional Network Method Applied to Segment Dynamic B-Mode Common Carotid Artery

林志憲(Chih-Hsien Lin)

指導教授 ： 黃詠暉

共同指導教授 ： 陳泰賓(Tai-Been Chen)

義守大學/智慧科技學院/資訊工程學系/碩士(2021年)

摘要

動機與目的：總頸動脈(Common Carotid Artery, CCA)在醫學診斷上扮演了重要的角色，而臨床B-Mode(灰階亮度模式)超音波影像則可以將總頸動脈血管內資訊即時的呈現出來。然而，B-Mode超音波影像上的散射以及斑點雜訊總是會干擾血管壁偵測的準確性。因此本研究採用FCN模型架構對於動態B-Mode總頸動脈超音波影像進行分割。
材料與方法：本研究為回顧性並包含了應用試驗性的實驗設計，材料分別由來自4位受試者之左、右總頸動脈共8組B-Mode動態超音波影像的10秒錄影檔，並將動態超音波影像錄影檔轉換為單張的BMP影像檔，取240張左側總頸動脈單張影像檔做為訓練集用以建立模型架構，同時以240張右側總頸動脈單張影像檔做為測試集來驗證模型。採用Fully Convolutional Networks (FCN)深度學習中的MobileNet-V2、ResNet-18以及ResNet-50演算法進行影像特徵萃取，再由Support Vector Machine 分類器建立總頸動脈血管分割模型；同時探討三種Batch Size (10、15、20)、三種Epoch Size(10、15、20)以及二種影像大小(224×224、256×256)；共計54種模型參數組合。最後，透過 Global Accuracy、Mean Accuracy、Mean IoU (Intersection Over Union)、Weighted IoU以及Mean BF Score (Boundary F1 Score)指標評估分割模型效能。
結果：ResNet-18分割模型在測試集之分類效能在所有組合中具有最好的表現，其中Global Accuracy、Mean Accuracy、Mean IoU、Weighted IoU以及Mean BF Score這5項指標的值分別為 0.969、0.941、0.894、0.942以及0.696。
結論：ResNet-18、ResNet-50與MobileNet-V2不僅能做為影像特徵萃取結合機器學習建立分割模型；同時也能夠對動態B-Mode總頸動脈超音波影像進行影像分割，並根據本研究結果顯示，這三個FCN模型皆具有很高的可行性及準確性。

關鍵字

總頸動脈 ； 動態B-Mode超音波影像 ； 深度學習

並列摘要

Common Carotid Artery (CCA) plays an essential role in medical diagnosis, and clinical B-Mode (gray-scale brightness mode) ultrasound imaging can present the intravascular information of the CCA in real-time. However, the scattering and speckle noise on the B-Mode ultrasound image will reduce segmented accuracy for the dynamic B-Mode ultrasound of CCA. Therefore, the fully convolutional network (FCN) is applied to segment the dynamic B-Mode ultrasound of CCA.
The retrospective study is designed in this application. Four healthy adults are involved in this study. The left and right CCA of the four healthy adults are imaging by ultrasound scanner with B-Mode dynamic ultrasound for 10 seconds and saving as MPEG format. The MPEG files are converted to series of single BMP images (i.e., 30 BMP images). A total of 480 BMP images of the left CCA are used as training sets, and another 480 BMP images of the right CCA are applied to validate the FCN models. The MobileNet-V2, ResNet-18, and ResNet-50 algorithms of FCN deep learning methods are adopted to extract image features. The Support Vector Machine (SVM) is utilized to build a classifier for segmented CCA. Meanwhile, three batch sizes (10, 15, 20), three Epoch Sizes (10, 15, 20), and two kinds of image sizes (224×224, 256×256) are investigated for segmented models. A total of 54 parametric combinations are inspected in this study. Finally, the Global Accuracy, Mean Accuracy, Mean IoU (Intersection Over Union), Weighted IoU, and Mean BF Score (Boundary F1 Score) are employed to evaluate the performance of the segmented models.
In the results, the ResNet-18 is the higher values of indexes among 54 parametric combinations. The Global Accuracy, Mean Accuracy, Mean IoU, Weighted IoU, and Mean BF Score as 0.969, 0.941, 0.894, 0.942, and 0.696 generated by ResNet-18.
ResNet-18, ResNet-50, and MobileNet-V2 could be used to extract image features with machine learning algorithms to segment the dynamic B-mode CCA ultrasound image. According to the results of this research, these three FCN models have high feasibility and accuracy.

並列關鍵字

FCN ； CCA ； Dynamic B-Mode Ultrasound

參考文獻

[1] Xiao Hu, Yufeng Zhang, Guanghui Cai, Kexin Zhang, Li Deng, Lian Gao, Suya Han, Jianhua Chen. A dynamic ultrasound simulation of a pulsating three-layered CCA for validation of two-dimensional wall motion and blood velocity estimation algorithms. Med Phys. 2018 Jan；45(1): 131-143.

[2] Jimmy Tat, Jason S Au, Peter J Keir, Maureen J MacDonald. Reduced common carotid artery longitudinal wall motion and intramural shear strain in individuals with elevated cardiovascular disease risk using speckle tracking. Clin Physiol Funct Imaging. 2017 Mar;37(2): 106-116.

[3] Sudha S, Jayanthi K B, Rajasekaran C, Nirmala Madian, Sunder T. Convolutional Neural Network for Segmentation and Measurement of Intima Media Thickness. J Med Syst. 2018 Jul 9;42(8): 154.

[4] Mainak Biswas, Venkatanareshbabu Kuppili, Luca Saba, Damodar Reddy Edla, Harman S Suri, Aditya Sharma, Elisa Cuadrado-Godia, John R Laird, Andrew Nicolaides, Jasjit S Suri. Deep learning fully convolution network for lumen characterization in diabetic patients using carotid ultrasound: a tool for stroke risk. Med Biol Eng Comput. 2019 Feb;57(2): 543-564.

[5] Meiyan Xie, Yunzhu Li, Yunzhe Xue, Randy Shafritz, Saum A. Rahimi, Justin W. Ady, Usman W. Roshan. Vessel lumen segmentation in internal carotid artery ultrasounds with deep convolutional neural networks. IEEE International Conference on Bioinformatics and Biomedicine (BIBM). 2019: 2393-2398.