使用三維卷積神經網路與優先候選聚合於全乳房自動超音波腫瘤偵測

Abstract

全乳房自動超音波已被廣泛應用在乳房攝影，用以篩檢乳房病變。然而，檢視數百張的全乳房自動超音波影像十分耗時。因此，本研究提出一套基於三維卷積神經網路 (convolutional neural networks, CNN) 與優先候選聚合的電腦輔助偵測系統，輔助加速檢閱。首先，本系統採用有效的三維滑動視窗法掃描整個影像以提取三維的感興趣區域 (volumes of interest, VOI)。接著，三維 CNN 會估測每個 VOI 是腫瘤的機率，並且機率大於某閾值的 VOI 會留下成為候選腫瘤。因為候選腫瘤彼此可能重疊，本系統採用一套新穎的方法以聚合重疊的腫瘤。聚合時，腫瘤機率高者會賦予較高優先權，以避免過度聚合之問題。最後，上述步驟會針對不同目標腫瘤大小，重複執行數次。本研究有效利用目標腫瘤與 VOI 之間的相對大小關係，以優化各個步驟之效能。在偵測率 95% (184/194)、90% (175/194)、85% (165/194)、80% (155/194) 的情況下，本系統在每個全乳房自動超音波掃描平均分別產生 2.9、2.0、1.3 以及 0.8 個誤判區域。總結之，相較過去的方法，本研究設計的系統較為通用，執行速度較快，偵測效果也較佳。

Automated whole breast ultrasound (ABUS) has been widely used as a screening modality for examination of breast abnormalities. Reviewing hundreds of slices produced by ABUS, however, is time-consuming. Therefore, in this study, a fast and effective computer-aided detection (CADe) system based on 3-D convolutional neural networks (CNN) and prioritized candidate aggregation is proposed to accelerate this reviewing. Firstly, an efficient sliding window method is used to extract volumes of interest (VOIs). Then, each VOI is estimated the tumor likelihood with a 3-D CNN, and VOIs with higher estimated likelihood are selected as tumor candidates. Since the candidates may overlap each other, a novel scheme is designed to aggregate the overlapped candidates. During the aggregation, candidates are prioritized based on estimated tumor likelihood to alleviate over-aggregation issue. The relationship between the sizes of VOI and target tumor is optimally exploited to effectively perform each stage of our detection algorithm. On evaluation with a testing set of 192 tumors, our method achieved sensitivities of 95% (184/194), 90% (175/194), 85% (165/194), and 80% (155/194) with 2.9, 2.0, 1.3, and 0.8 FPs per pass, respectively. In summary, our method is more general and much faster than preliminary works, and demonstrates promising results.