基於深度學習之乳房超音波單階段腫瘤偵測

Abstract

全乳房自動超音波(Automated Breast Ultrasound, ABUS)已被廣泛應用在乳房篩檢上，然而，其產生的數百張超音波影像反而造成醫師需耗費較多時間閱片。近來，基於卷積類神經網路(Convolutional Neural Network, CNN)開發的電腦輔助系統(Computer-aided System)已證實在醫學影像上能有效協助醫師進行腫瘤偵測與診斷。因此本研究提出了一個單階段的三維卷積神經網路電腦輔助偵測系統，先行找到影像中可能的病灶位置，協助醫師檢閱影像。本系統特點是一次性地(One Take)掃描整個ABUS影像進行腫瘤偵測，因此偵測速度極快，平均每個全乳房超音波掃描影像在0.8秒內可以完成掃描，此外，我們更針對小腫瘤問題設計系統架構，減少遺漏的情況。 在系統設計上，我們除了修改單階段的YOLOv3 (You Only Look Once Version 3) 偵測架構外，更應用焦點損失(Focal Loss)概念修改損失函數與提出循環訓練方法(Cycle Learning)改善資料不平衡問題(Data Imbalance)。在偵測時，先調整全乳房超音波影像大小，接著系統掃描影像並產生具有腫瘤的邊界框。最後，為了避免重複偵測的問題，我們使用非最大值抑制(Non-maximal Suppression, NMS)來移除重疊的邊界框得到最後預測結果。系統在偵測率98%、95%、90%的情況下，每個全乳房自動超音波掃描影像分別平均產生3.8、2.0、1.0個偽陽性(False Positive, FP)腫瘤。相較過去的方法，我們提出的系統效果更好且執行速度更快。

The automated breast ultrasound (ABUS) had been widely used in breast examination. However, it is time-consuming for the physician to review hundreds of slices produced by ABUS. In recent, the computer-aided system based on the convolutional neural network (CNN) had been proven that it can assist effectively the physician for tumor detection and diagnosis on the medical image. Therefore, a computer-aided detection (CADe) system based on the one-stage 3-D convolutional neural network (CNN), 3-D You Only Look Once (3-D YOLO), is proposed in this study to locate suspicious lesions for the physician reviewing image. The particular characteristic of our system is that the whole ABUS is detected quickly in one take and the average detecting time of per ABUS image is 0.8 second. Moreover, the proposed system is also designed for smaller tumors to reduce the misdetection rate. In the proposed system, for achieving better performance, not only the one-stage YOLOv3 (You Only Look Once Version 3) object architecture is used and redesigned for smaller tumors, but also the essence of focal loss is applied to our loss function, and the scheme of cycle learning is designed for solving data imbalance problem. Before the tumor detection, the ABUS images are resized firstly to match the model shape. Then, the bounding boxes of tumor candidates are generated by the detection system. After that, the non-maximal suppression (NMS) is performed to eliminate the overlapping for determining the final tumor bounding box. Finally, our method achieves sensitivities of 98%, 95%, 90% with 3.8, 2.0, 1.0 false positives (FP) per pass, respectively. Compared to the previous works, the proposed CADe system is much better and faster.