使用三維卷積神經網路於全乳房自動超音波腫瘤偵測

Abstract

乳癌為女性最常見的癌症之一，根據過去的研究顯示，早期發現可以降低乳癌死亡率，超音波檢查為主要乳癌篩檢工具之一，自動化乳房超音波使用大型探測端掃描乳房三維影像，搭配電腦輔助診斷系統以協助醫生判讀超音波影像，進一步篩檢乳癌。而近年來卷積類神經網路已被廣泛應用在各個領域，有鑒於此，本研究提出一套應用卷積類神經網路的電腦輔助診斷系統，以突破過往使用手動提取特徵的限制。 本研究中首先使用滑動視窗的方式來提取三維超音波影像中感興趣的區塊，再者，將提取出的區塊作為三維整體學習卷積類神經網路的輸入，在類神經網路中學習影像特徵後計算判讀為腫瘤的機率為何，最後，加入影像後處理以避免腫瘤候選區塊的過多重疊。該研究中使用的資料集共包含了246份病例，其中分別切為訓練、驗證及測試資料集，並將測試資料集用於評估提出的電腦輔助診斷系統效能，測試資料集共包含81份病例、473通道影像，實驗結果發現在21.6/126.2、6.0/34.8和4.6/27.1三種不同偽陽性(一通道/一病例)的情況下，靈敏度分別達到了100% (81/81)、95.3% (77/81)和90.9% (74/81)。除此之外，系統的執行時間在每一通道影像下為28.3秒，完成一病例共需174.6秒。總結來說，應用三維卷積類神經網路於電腦輔助診斷系統可有效提升整體偵測表現並降低執行時間。

The breast cancer is one of the most common cancers in the female. According to the previous studies, the detection in the early stage can help reduce the breast cancer mortality. The automated breast ultrasound (ABUS) is useful for breast cancer examination that utilizes a larger transducer to record the whole breast images and provides the 3-D large volume of the breast. Besides, the computer-aided detection (CADe) system is introduced to help the radiologists while reviewing and diagnosing in ABUS images. With the success of convolutional neural network (CNN) in image tasks, a CADe system using 3-D CNN is proposed to overcome the limitations of conventional CADe system that requires hand-crafted features. In this study, a sliding window based volume of interest (VOI) extraction approach is adopted. Then, the extracted VOIs are considered as the input of 3-D ensemble CNN for estimating the tumor likelihood. Finally, the post-processing method is employed to address the overlapping tumor candidates’ issues. In this study, the dataset contains 246 cases and is divided into training, validation, and testing sets. The proposed CADe system is evaluated on the testing set that consists of 81 cases with 473 passes and 104 tumors. This study achieves the sensitivities of 100% (81/81), 95.3% (77/81), and 90.9% (74/81) with the false positives per pass/per case 21.6/126.2, 6.0/34.8, and 4.6/27.1 respectively. Besides, the execution time is 28.3 seconds per pass and 174.6 seconds per case. In conclusion, the proposed CADe system using 3-D CNN is much more time efficient and obtains better performance than previous works.