基於深度注意力卷積神經網路之乳房自動超音波腫瘤診斷

Tumor Diagnosis on Automated Breast Ultrasound Using Attention Inception Neural Network

張鈞(Jun Zhang)

指導教授 ： 張瑞峰

國立臺灣大學/電機資訊學院/生醫電子與資訊學研究所/碩士(2020年)

https://doi.org/10.6342/NTU202003377

摘要

自動乳房超音波(Automated Breast Ultrasound, ABUS)是一種廣泛應用於乳癌檢測與診斷的超音波造影技術，可以一次掃描整個乳房，提供完整的三維乳房影像資訊。然而大量的影像需要放射科醫生花費較多的時間來診斷，也存在著初步誤判的風險。為了減少誤判的情況，電腦輔助診斷系統(Computer-Aided Diagnosis, CADx)應運而生，近年來卷積神經網路(Convolutional Neural Network, CNN)能夠自動提取特徵，在醫學影像上發展迅速，基於卷積神經網路的電腦輔助診斷系統在診斷上也有出色的表現。因此，本研究提出了一個基於三維卷積神經網路的電腦輔助診斷系統，系統內包含一個增強型三維U形網路與一個三維注意力卷積神經網路，其中注意力卷積神經網路由Inception網路、ResNeXt網路及SE (Squeeze-and-Excitation)網路組成。首先，從ABUS影像中擷取腫瘤區域並縮放至固定大小，同時進行影像增強的前處理；接著，利用增強型三維U形網路對腫瘤區域進行切割，取得腫瘤遮罩；最後，將腫瘤區域原始影像、腫瘤區域增強影像及腫瘤遮罩匯入三維注意力卷積神經網路進行腫瘤的分類進行初步診斷腫瘤的良惡性。本論文所提出的系統可達到89.2%的準確率、90.3%的靈敏性、88.1%的特異性以及0.9255的曲線下面積的結果，實驗結果顯示提出的系統利用ABUS影像能超越3年經驗醫師的判斷結果。

關鍵字

乳癌 ； 自動乳房超音波 ； 電腦輔助診斷系統 ； 卷積神經網路 ； 分組卷積 ； 注意力機制

並列摘要

The automated breast ultrasound (ABUS) has been widely used in the detection and diagnosis of breast cancer since it could scan the whole breast and provide the complete three-dimensional (3-D) volume of the breast. However, it was a time-consuming task for a radiologist to diagnosis by reviewing an ABUS image, and there was a risk of misdiagnosis. To eliminate the risk, the computer-aided diagnosis (CADx) systems were proposed to assist the physicians. In recent years, the convolutional neural networks (CNN), which could extract features automatically, has developed rapidly in the field of medical images, and the CNN-based CADx could have outstanding performance. Hence, in our study, a CADx system based on 3-D CNN was proposed for ABUS tumor classification. Our CADx system was composed of the tumor volume of interest (VOI) extraction, the tumor segmentation, and the tumor classification. First, the tumor VOI extracted from the ABUS image was resized to fixed size. Then, the VOI was enhanced by histogram equalization. Second, in the tumor segmentation, the 3-D U-Net++ was applied to the resized VOI for generating the tumor mask. Finally, for the tumor classification, the VOI, the enhanced VOI, and the corresponding tumor mask were fed into the 3-D SE-Inception-ResNeXt network, which was composed of the Inception model, the ResNeXt model and the Squeeze-and-Excitation (SE) module, to classify the tumor as benign or malignant. In our experiment result, the accuracy, sensitivity, and specificity could reach 89.2%, 90.3% and 88.1% respectively, and the area under the receiver operating characteristic curve (AUC) was 0.9255, which demonstrated that our CADx system for ABUS image was comparable to a 3-year-experience reader in tumor diagnosis tasks.

並列關鍵字

breast cancer ； ABUS ； CADx ； CNN ； group convolution ； attention mechanism

參考文獻

[1] R. L. Siegel, K. D. Miller, and A. J. C. a. c. j. f. c. Jemal, "Cancer statistics, 2019," vol. 69, no. 1, pp. 7-34, 2019.

[2] N. Antropova, B. Q. Huynh, and M. L. J. M. p. Giger, "A deep feature fusion methodology for breast cancer diagnosis demonstrated on three imaging modality datasets," vol. 44, no. 10, pp. 5162-5171, 2017.

[3] A. Vourtsis and A. J. E. r. Kachulis, "The performance of 3D ABUS versus HHUS in the visualisation and BI-RADS characterisation of breast lesions in a large cohort of 1,886 women," vol. 28, no. 2, pp. 592-601, 2018.

[4] A. Jalalian, S. B. Mashohor, H. R. Mahmud, M. I. B. Saripan, A. R. B. Ramli, and B. J. C. i. Karasfi, "Computer-aided detection/diagnosis of breast cancer in mammography and ultrasound: a review," vol. 37, no. 3, pp. 420-426, 2013.

[5] M. Elter and A. J. M. p. Horsch, "CADx of mammographic masses and clustered microcalcifications: a review," vol. 36, no. 6Part1, pp. 2052-2068, 2009.