3-D匯總八度卷積神經網路使用於自動乳房超音波電腦輔助腫瘤診斷

Abstract

近年來乳癌已成為女性最常見的疾病之一。透過早期的偵測、診斷和治療可以顯著地改善死亡率。在乳癌的檢測中，全自動乳房超音波系統(Automated Breast Ultrasound System, ABUS)因其可以提供完整的乳房三維影像資訊，逐漸被採用於乳癌檢測。然而，即使是經驗豐富的醫生在檢查超音波圖像時也容易受到腫瘤形狀和內部紋理的影響導致初步的病灶誤判，為了幫助解決這個問題，基於卷積神經網路(Convolutional Neural Network, CNN)開發的電腦輔助診斷系統(Computer-Aided Diagnosis system)應運而生。本研究提出一個包含三維腫瘤切割網路和三維腫瘤分類網路的電腦輔助診斷系統，首先，透過從原始ABUS影像中提取含有紋理資訊的腫瘤區域，接著通過融合了殘差模塊(Residual Block)和巢狀U型網路(Nested U-Net)的腫瘤切割模型產生含有形狀資訊的腫瘤遮罩，最後將腫瘤遮罩和腫瘤區域同時放入由腫瘤分類網路取出形狀和紋理特徵以進行腫瘤良惡性的判斷。在腫瘤分類網路中使用了具有八度卷積以及擠壓和激發模組的匯總殘差網路瓶頸塊(Bottleneck Block from ResNeXt)作為基本結構來構建我們的模型，以獲得更準確的結果。本研究中總共使用了403顆腫瘤，其中包含了199顆良性腫瘤和204顆惡性腫瘤。實驗結果顯示，所提出的系統能達到準確率88.6%、靈敏性90.6%、特異性86.9%和ROC曲線下面積0.9333的成果，這樣的表現在臨床上與擁有三年ABUS經驗的醫生相同，顯示提出的系統有足夠的能力進行腫瘤良惡性的預測。

In recent years, breast cancer has become one of the most common diseases in women. Through early detection, diagnosis, and treatment, the mortality rate could be significantly improved. In breast cancer examination, the Automated Breast Ultrasound System (ABUS) was gradually adopted for breast cancer examination because it could provide complete information by recording the whole breast in the three-dimensional (3-D) image. However, even the experienced physician might be susceptible to the shape and internal texture of the tumor while reviewing ABUS images and misjudged the lesion. To solve this problem, the computer-aided diagnosis (CADx) system based on a convolutional neural network (CNN) was provided. In this study, a CADx system consisting of a 3-D tumor segmentation model and a 3-D tumor classification model was proposed for tumor diagnosis. First, the tumor region with texture information was extracted from the original ABUS image. Then, tumor masks containing shape information were generated by the 3-D tumor segmentation model, which fused the residual block and the U-net++. Finally, the tumor region and the corresponding tumor mask were both fed into our tumor classification model to extract the shape and texture feature maps for determining the tumor as benign or malignant. In our tumor classification model, the bottleneck block from ResNeXt with the octave convolution and the squeeze-and-excitation module was used as a basic structure to construct our model. In experiments, a total of 403 tumors, including 199 benign tumors and 204 malignant tumors, were used in this study to evaluate the proposed system performance. The experimental results showed that the proposed system could achieve 88.6% accuracy, 90.6% in sensitivity, 86.9% in specificity, and 0.9333 in the area under the ROC curve. Such performance was clinically the same as that of a doctor with three years of ABUS experience, showing that the proposed system had sufficient ability to predict tumors as benign or malignant.