利用人工智慧模型分類暨分割超音波乳房影像腫瘤之研究

The Study of Classified and Segmented Tumors on Breast Sonography by Artificial Intelligence Models

朱姵蓉(Pei-Rong Zhu)

指導教授 ： 陳泰賓

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：臨床乳房超音波(Breast Ultrasound)影像品質受限於操作者經驗及技術熟練度而缺乏客觀性。然而，聲波脈衝回音 (Pulse-Echo)在不均勻人體組織環境中會產生深度衰減效應(Attenuation Effect)、散射效應(Scattering Effect)；因此超音波影像常具有斑點雜訊(Speckle Noise)及同組織影像強度不一致現象，造成影像診斷、分類(Classification)、標記(Labeling)或切割(Segmentation)具有一定影響及困難。因此透過人工智慧(Artificial Intelligence, AI)模型進行自動辨識乳房超音波腫瘤影像之研究。
材料與方法：本研究使用之乳房音波影像來自波蘭科學院(Polish Academy of Sciences)公開資料庫(Open Access Series of Breast Ultrasonic Data, OASBUD)。取得78位女性(52例惡性和48例良性)乳腺病變超音波射頻回音原始數據(Radio Frequency, RF)及腫瘤標記二元影像，共計200張RF及200張標記腫瘤二元影像。使用深度學習中的旋積類神經網路(Convolutional Neural Network, CNN)演算法建立影像分類與腫瘤切割之應用。採用隨機分組方法(Random Splitting Method)將樣本分成訓練與驗證組。影像分類模型效能評估採用驗證集之準確度、特異性、靈敏度及Kappa一致性統計量；影像腫瘤切割模型效能評估採用驗證集之整體準確度(Global Accuracy)、平均準確度(Mean Accuracy)、平均交併比Mean IoU (Intersection Over Union, Mean IoU)、加權交併(Weighted IoU)、平均BF值(Boundary F1 Score)。
結果：MobileNet-V2 結合Linear分類模型在測試集之分類效能在所有組合中具有最好的表現；其準確度、靈敏度、特異性、Kappa分別為0.950、0.947、0.952、0.900。而ResNet-18、ResNet-50、MobileNet-V2腫瘤分割平均邊界分數(Mean Boundary F-Score)分別為0.951、0.949以及0.926。
結論：ResNet-18、ResNet-50及MobileNet-V2不僅能做為影像特徵萃取結合機器學習建立分類模型；同時也能夠針對乳房超音波腫瘤影像進行分割。根據本研究結果顯示，這三個CNN模型適合乳房超音波腫瘤分類與分割之任務。

關鍵字

乳房超音波影像 ； 人工智慧 ； 深度學習 ； 旋積類神經網路 ； 全卷積類神經網路

並列摘要

Purpose: The image quality of clinical breast ultrasound is limited by the operator's experience and technical proficiency and lacks objectivity. Furthermore, the pulse-echo produces a depth attenuation effect and a scattering effect in the human tissue environment. Ultrasonic images often have speckle noise and a blurring effect due to imaging inhomogeneous tissues or organs. It has a certain influence and difficulty in imaging diagnosis, classification, labeling, or segmentation. Therefore, the artificial intelligent models were applied to identify breast tumor automatically.
Methods and Materials: The acoustic breast image used in this study comes from the Open Access Series of Breast Ultrasonic Data (OASBUD) of the Polish Academy of Sciences. We obtained 78 women (52 cases of malignant and 48 cases of benign) breast lesions with ultrasound radiofrequency echo raw data (Radio Frequency, RF) and tumor marker binary images, a total of 200 RF and 200 marker tumor binary images. The deep learning models with convolutional neural networks (CNN) and fully convolutional networks (FCN) were employed to establish image classification and segmentation. The random splitting method was used to divide the samples into training and validation groups. The evaluation of performance for the image classification model was the accuracy, specificity, sensitivity, PPV, NPV, and Kappa consistency statistics of the testing set. Meanwhile, the evaluation of performance for the tumor segmentation was the global accuracy, mean accuracy, mean IoU (Intersection Over Union, Mean IoU), Weighted IoU, and Boundary F1 Score.
Results: The MobileNet-V2 combined with the linear classifier provided the best performance among all possible models based on the test set. The accuracy, sensitivity, specificity, and Kappa were 0.950, 0.947, 0.952, and 0.900. The mean boundary F-scores of segmented tumors were 0.951, 0.949, and 0.926 with respective to ResNet-18, ResNet-50, and MobileNet-V2 in this study.
Conclusion: The ResNet-18, ResNet-50, and MobileNet-V2 were used to do image feature extraction and were applied to segment tumors in the ultrasound images. According to the performance of the testing set, these three CNN models are suitable for the classification and segmentation of breast ultrasound.

並列關鍵字

Breast Ultrasound ； Artificial Intelligence Model ； Deep Learning ； CNN ； FCN

參考文獻

[1] Siegel RL, Miller KD, Jemal A. Cancer statistics, 2019. CA Cancer J Clin. 2019;69(1):7-34.

[2] Løberg M, Lousdal ML, Bretthauer M, Kalager M. Benefits and harms of mammography screening. Breast Cancer Res. 2015;17(1):63.

[3] Cao Z, Duan L, Yang G, Yue T, Chen Q. An experimental study on breast lesion detection and classification from ultrasound images using deep learning architectures. BMC Med Imaging. 2019;19(1):51.

[4] Brem RF, Lenihan MJ, Lieberman J, Torrente J. Screening breast ultrasound: past, present, and future. AJR. 2015;204(2):234-40.

[5] Geisel J, Raghu M, Hooley R. The Role of Ultrasound in Breast Cancer Screening: The Case for and Against Ultrasound. Semin Ultrasound CT MR. 2018 ;39(1):25-34.