請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/67126
標題: | 使用3D卷積神經網路之乳房自動超音波電腦輔助腫瘤診斷 Using 3D Convolutional Neural Network on Automated Breast Ultrasound for Computer-Aided Tumor Diagnosis |
作者: | Hsin-Yi Peng 彭馨儀 |
指導教授: | 張瑞峰 |
關鍵字: | 乳癌,三維卷積神經網路,電腦輔助診斷, breast cancer,3-D convolutional neural network,computer-aided diagnosis system, |
出版年 : | 2017 |
學位: | 碩士 |
摘要: | 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,為了能提高判診斷的正確率,許多的研究會專注於腫瘤特徵的提取,例如:腫瘤紋理、腫瘤形狀等,好的特徵往往能大幅提升診斷系統的準確率,因此,此篇研究主要的目的是使用三維卷積神經網路自動提取腫瘤特徵,進而提升電腦輔助診斷的效能。此篇研究我們提出的方法是利用兩個不同架構的三維卷積神經網路分別做紋理特徵和形狀特徵,紋理卷積神經網路的輸入是從超音波影像取出的腫瘤VOI (volume-of-interest),形狀卷積神經網路的輸入是利用全卷積神經網路切割出的腫瘤遮罩影像,然後分別從這兩個訓練好的神經網路中取出特徵,最後將兩種類型的特徵結合輸入類神經網路裡做良惡性的分類。本篇研究使用了77筆腫瘤資料,其中35顆良性腫瘤,42顆惡性腫瘤,研究顯示,使用卷積神經網路取出的特徵分類結果確實比手動提取的特徵還好,可達到準確率85.7% (66/77),靈敏性 92.8% (39/42),特異性77.1% (27/35),ROC曲線面積0.8876,能比手動提取的特徵得到更準確的良惡性分類結果。 The breast cancer is a common cancer among female. In the past years, the mortality of the breast cancer has steadily declined because the earlier treatment can effectively increase the survival rate. In the clinical, breast ultrasound are often utilized to diagnosis tumor between benign or malignant. For increasing the diagnosis accuracy, there were many studies focusing on the tumor feature extraction. A proper feature can efficiently advance the performance of the diagnosis system. As a result, the main idea of this study was using 3-D convolutional neural network (CNN) for automatic feature extraction instead of the hand-crafted features to increase the performance of the system. The proposed method was utilized two 3-D CNNs with different architecture to obtain the texture and the morphology features respectively. The input of the texture CNN was the VOI (volume of interest) extracted from the original ABUS image, and the input of the shape CNN was the mask image of the VOI generated from the fully convolutional network (FCN). Then, the features extracted from the two CNNs were concatenated as the input of an artificial neural network (NN) for classification. There were total 77 tumors used in this study, including 35 benign tumors and 42 malignant tumors. According to the experimental results, the performance of the features learned from the CNNs were better than the hand-crafted features. The accuracy, sensitivity and the specificity of the proposed method were 85.7% (66/77), 92.8% (39/42) and 77.1% (27/35) respectively, and the area under the ROC curve was 0.887. In conclusion, the auto-learned features by the 3-D CNN performed better than the hand-crafted features. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/67126 |
DOI: | 10.6342/NTU201703030 |
全文授權: | 有償授權 |
顯示於系所單位: | 資訊工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-106-1.pdf 目前未授權公開取用 | 1.59 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。