卷積神經網絡的集成學習之乳房超音波電腦輔助診斷

Abstract

乳癌是女性癌症中很常見的一種，而近年來，乳癌的死亡率已大大下降，因為及早發現及早治療能有效的提高存活率。而在臨床上，乳房超音波影像常用來判斷腫瘤的良惡性，並且搭配電腦輔助診斷系統以協助醫生偵測及診斷，也能夠降低不同醫生對於相同腫瘤診斷的變異性。本研究主要的目的是使用卷積神經網路自動提取腫瘤特徵，並且搭配集成學習結合多個網路以增進效能。此篇研究我們提出的方法是利用四種不同架構的卷積神經網路分別學習不同腫瘤特徵，並且我們還提出利用全卷積神經網路自動切割出腫瘤遮罩影像的方法提供腫瘤形狀的特徵達成更精確的診斷。本篇研究使用了1687筆腫瘤資料，其中有953顆良性腫瘤，有734顆惡性腫瘤，研究顯示，使用集成學習結合學習到不同特徵的卷積神經網路的分類結果確實比單一卷積神經網路的分類結果還要好，可達到準確率91.10%，靈敏性 85.14%，特異性 95.77%，ROC曲線面積0.9697，比所有單一網路的分類結果還要高，因此，使用集合方法可以減少分類的偏差，並且使用代表性特徵可以提高診斷效果。

Breast cancer is the most common malignancy of the total cancer cases in United States females. However, early diagnosis leads to early treatment and reduces mortality rates. In the clinical usage, breast ultrasound and computer aided diagnosis (CAD) is usually used to diagnosis tumors into benignancy or malignancy. In addition, CAD has been used to decrease the diagnosis variation of different physicians and assist to classify or detect the tumors. In our study, we use the convolutional neural network (CNN) for automatic feature extraction and the ensemble method to combine multi CNN models for better diagnostic performance. The CNN-based method proposed in this study includes VGG-Like, VGG-16, ResNet, and DenseNet. Also, we proposed a fully convolutional network (FCN) to employ tumor segmentation and extract tumor shape features automatically. There were total 1687 tumors used in this study, including 953 benign tumors and 734 malignant tumors. The accuracy, sensitivity, and the specificity of the proposed method were 91.10%, 85.14%, and 95.77%, respectively, and the area under the ROC curve was 0.9697. In conclusion, the ensemble method can improve the performance by using multiple CNN methods and the tumor shape feature can improve the diagnostic effect.