請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/83722
標題: | 電腦斷層肺血管攝影之肺栓塞偵測與分類:深度學習模型之建構 Detection and Classification of Pulmonary Embolism in CTPA Images Using Deep Learning Model |
作者: | 林彥廷 Yan-Ting Lin |
指導教授: | 陳中明 Chung-Ming Chen |
關鍵字: | 電腦斷層肺血管攝影,肺栓塞,深度學習,時間卷積網絡, computed tomography pulmonary angiography,pulmonary embolism,deep learning,temporal convolutional network, |
出版年 : | 2022 |
學位: | 碩士 |
摘要: | 肺栓塞為目前三種最常見的心血管疾病之一,其具有發病率高、臨床診斷率低與死亡率高的特性,隨著人口老化與生活水準上升等因素,肺栓塞的發病率逐年上升,且其發病的時間十分快速,需要緊急的治療,因此導致其未治療的30%高死亡率,若能及時進行診斷並加以治療,死亡率能降至2-10%,然而肺栓塞的診斷有其難度且耗時,電腦輔助診斷系統的引進可以幫助臨床醫師提供客觀的建議,以協助臨床醫師進行決策。 因此本研究開發一套基於深度學習的兩階段肺栓塞偵測分類演算法,針對肺栓塞的陽/陰性、急/慢性與出現的位置進行分類,使用RSNA STR Pulmonary Embolism Detection競賽中的CT Dataset作為樣本,演算法的第一階段為輸入CTPA二維影像以SE-ResNeXt-50進行特徵提取與初步肺栓塞陽陰性之分類,使用5-fold交叉驗證得到的平均AUC為0.962±0.003,而第二階段為輸入第一階段提取之特徵以本研究提出的PE-TCN模型針對整個案例的肺栓塞相關標籤進行分類,同時提升二維影像肺栓塞標籤分類的性能,另外本研究提出PE Weighting的方法加入到PE-TCN模型中,使用5-fold交叉驗證得到的Negative Exam for PE標籤平均AUC為0.924±0.007,Left-sided PE標籤平均AUC為0.908±0.013,Central PE標籤平均AUC為0.951±0.009,Right-sided PE標籤平均AUC為0.924±0.011,Chronic PE標籤平均AUC為0.654±0.025,Acute and Chronic PE標籤平均AUC為0.855±0.025,PE Present on Image標籤平均AUC為0.970±0.004,與常被用來處理序列問題的GRU模型進行比較,GRU模型使用5-fold交叉驗證得到的Negative Exam for PE標籤平均AUC為0.913±0.004,Left-sided PE標籤平均AUC為0.900±0.006,Central PE標籤平均AUC為0.938±0.009,Right-sided PE標籤平均AUC為0.917±0.013,Chronic PE標籤平均AUC為0.632±0.030,Acute and Chronic PE標籤平均AUC為0.843±0.028,PE Present on Image標籤平均AUC為0.925±0.039,PE-TCN模型的分類結果明顯優於GRU模型,因此,研究結果證明PE-TCN模型適用於肺栓塞相關的分類,本研究提出之基於深度學習的兩階段肺栓塞偵測分類演算法能有效協助醫師進行CTPA肺栓塞之診斷。 Pulmonary embolism is one of the three most common cardiovascular diseases. It has high morbidity, low clinical diagnosis rate and high mortality. With factors such as population aging and living standards, the incidence of pulmonary embolism is increasing year by year. Its time of onset is very fast and it requires urgent treatment. Therefore, it has high mortality rate of 30% without treatment. If timely diagnosis and treatment can be carried out, the mortality rate can be reduced to 2-10%. However, the diagnosis of pulmonary embolism is difficult and time-consuming. The use of computer-aided diagnosis systems can help physicians provide objective recommendations to assist physicians in decision-making. Therefore, this study will develop a deep learning based two-stage pulmonary embolism detection and classification algorithm to classify the positive/negative, acute/chronic, and location of pulmonary embolism. We use the CT Dataset by the RSNA STR Pulmonary Embolism Detection competition. The first stage of the algorithm inputs 2D CTPA images to extract features and preliminarily classify pulmonary embolism positive and negative with SE-ResNeXt-50 model. The average AUC using 5-fold cross-validation is 0.962±0.003. The second stage of the algorithm inputs the features extracted in the first stage to classify the study-level pulmonary embolism related labels and improves performance of 2D images pulmonary embolism classification with PE-TCN model proposed in this study. In addition, we propose the PE Weighting method adding to the PE-TCN model. The average AUC of Negative Exam for PE label is 0.924±0.007, the average AUC of Left-sided PE label is 0.908±0.013, the average AUC of Central PE label is 0.951±0.009, the average AUC of Right-sided PE label is 0.924±0.011, the average AUC of Chronic PE label is 0.654±0.025, the average AUC of Acute and Chronic PE label is 0.855±0.025, and the average AUC of PE Present on Image label is 0.970±0.004. We compare TCN-PE with GRU which is often used to deal with sequence problems, the average AUC of the Negative Exam for PE label obtained by GRU is 0.913±0.004, the average AUC of the Left-sided PE label is 0.900±0.006, the average AUC of the Central PE label is 0.938±0.009, the average AUC of the Right-sided PE label is 0.917±0.013, the average AUC of the Chronic PE label is 0.632±0.030, the average AUC of the Acute and Chronic PE label is 0.843±0.028, and the average AUC of the PE Present on Image label is 0.925±0.039. The classification results of the TCN-PE model are significantly better than the GRU model. Therefore, the research results prove that the PE-TCN model is suitable for the classification of pulmonary embolism. The deep learning based two-stage pulmonary embolism detection and classification algorithm proposed in this study can effectively assist physicians to diagnose CTPA pulmonary embolism. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/83722 |
DOI: | 10.6342/NTU202202129 |
全文授權: | 未授權 |
顯示於系所單位: | 醫學工程學研究所 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-110-2.pdf 目前未授權公開取用 | 3.98 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。