請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/70801
標題: | 使用3-D卷積類神經網路於乳房DCE-MRI遠端轉移預測 Distant Metastasis Prediction in Breast DCE-MRI using 3-D Convolutional Neural Networks |
作者: | Tzu-Chuan Lin 林子荃 |
指導教授: | 張瑞峰(Ruey-Feng Chang) |
關鍵字: | 遠端轉移,乳癌,核磁共振影像,電腦輔助診斷,深度學習,三維卷積神經網路,集成學習, Distant metastasis,Breast cancer,DCE-MRI,Computer-aided diagnosis,Deep learning,3-D convolutional neural network,Ensemble learning, |
出版年 : | 2018 |
學位: | 碩士 |
摘要: | 乳癌的遠端轉移是乳癌從原始腫瘤擴散到其他遠處器官的過程,是人類健康的重大威脅,同時也是乳癌致死的主要原因。透過準確預測來幫助患者能早期發現遠端轉移的情況,藉此能降低死亡率。有一些研究提取傳統特徵後,再透過統計分析方法來預測惡性腫瘤是否會產生遠端轉移,這些預測結果能夠幫助醫生優化乳癌的治療,醫生可以參考預測結果為患者制定適當的診療計畫。本研究提出了一個創新的預測系統來自動預測DCE-MRI上的惡性腫瘤在術後監測的期間是否可能發生遠端轉移,並與傳統方法比較。在傳統方法的實作中,提取傳統特徵之後,我們使用了兩種常見分析方法,邏輯回歸和支持向量機作為分類器來分析傳統手工特徵。而在本研究提出的預測系統中,所有研究資料會先進行影像前處理,以有效地進行腫瘤切割和獲得影像中感興趣的部分VOI (volume-of-interest),再來分別利用VOI、腫瘤遮罩影像以及腫瘤區域,來建構系統中的三個基底模型,分別是紋理卷積神經網路、形狀卷積神經網路和腫瘤卷積神經網路,最後,再利用集成學習來將三個基底模型合併成一個集成模型,來進行最後的預測。在最終結果上,本研究提出的預測系統可以達到準確率80.69%,靈敏性88.28%,特異性73.10%,ROC下的曲線面積為0.8517,總體而言,能比傳統方法得到更好的預測結果。 Distant metastasis is a significant threat to human health, and it is the principal cause of death from breast cancer. It is the process that the breast cancer has spread from the original tumor to distant organs. Accurate prediction and early detection of distant metastasis play an important role in decreasing the mortality rates of distant metastasis in breast cancer. There were several studies extracting conventional hand-crafted features in statistical analysis methods to predict distant metastasis of malignant tumors. This prediction can help to optimize the treatment, and physicians can also develop appropriate therapy plans for patients by referring to prediction results. In this study, an innovative prediction system was proposed to automatically predict the presence of distant metastasis on DCE-MRI, which was compared with conventional methods. In conventional methods, after extracting conventional features, two general analysis methods would be used, logistic regression and the support vector machine (SVM) as the classifiers to analyze conventional hand-crafted features. In the proposed prediction system, the input data was first performed image pre-processing to efficiently segment the tumor region and volume of interest (VOI). Subsequently, three convolutional neural network (CNN) models texture model trained with VOIs, shape model trained with masks, and the tumor model trained with tumor region images, were constructed to build the proposed system. Lastly, the ensemble model combined by the previous base models achieved 80.69% (234/290) in accuracy, 88.28% (128/145) in sensitivity, 73.10% (106/145) in specificity, 76.65% (128/167) in PPV, 86.18% (106/123) in NPV, and 0.8517 in the value of AUC. The results were superior to the conventional methods. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/70801 |
DOI: | 10.6342/NTU201802689 |
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-107-1.pdf 目前未授權公開取用 | 1.44 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。