請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/67039
標題: | 使用3-D注意力卷積神經網路診斷電腦斷層掃描影像上的肺結節 Diagnosis of Lung Nodules on Computed Tomography Images Using 3-D Attention Convolution Neural Network |
作者: | Sheng-Zhi Huang 黃聖智 |
指導教授: | 張瑞峰(Ruey-Feng Chang) |
關鍵字: | 肺結節,電腦斷層掃描,電腦輔助診斷,殘差網路,注意力機制,特徵網路金字塔, Lung nodules,computed tomography scan,computer-aided diagnosis,residual network,attention mechanism,feature pyramid network, |
出版年 : | 2020 |
學位: | 碩士 |
摘要: | 肺癌是全世界死亡率最高的癌症之一,而肺結節是早期肺癌的一個重要表現,及早發現與治療能有效降低肺癌致死率。低劑量電腦斷層掃描(Low-dose Computed Tomography, LDCT)是肺癌偵測與診斷的重要工具,不僅輻射劑量低,同時能a提供完整的3-D胸部影像與保有肺結節解析度,因此廣泛用於肺癌的檢測。過去幾年,卷積神經網路(Convolutional Neural Network, CNN)在醫學影像領域中蓬勃發展,基於卷積神經網路實現的電腦輔助診斷系統(Computer-Aided Diagnosis, CAD)已被證實能以機器學習方式擷取圖像特徵並幫助醫師進行初步診斷。因此,我們提出以ResNeXt作為模型骨幹的CAD提供醫生診斷方面的協助,為防止卷積神經網路生成過多冗餘的特徵,我們導入輕量化的注意力機制模組使網路能專注在重要的特徵,此外,為了提升較小的惡性肺結節判斷,結合金字塔網路以融合不同尺度的特徵資訊,最後使用通道優化的混合型損失函數作為系統訓練時的損失函數,使網路能學習更細微的辨識特徵用於區分極端病例。在此研究中,使用880筆來自美國國家肺篩檢試驗的低劑量電腦斷層掃描影像進行系統的訓練與測試,其中良性結節與惡性結節各為440筆影像。實驗結果顯示,提出的系統可以達到85.3%的準確率、86.8%的靈敏性、83.9%的專一性,且ROC曲線下面積可達0.9042,證實所設計的系統有不錯的診斷能力。 Lung cancer was one of the cancers with the highest mortality rate in the world, and the lung nodule was one of the crucial symptoms of early lung cancer. Early detection and treatment could effectively reduce lung cancer mortality. Low-dose Computed Tomography (LDCT) is an important tool for lung cancer detection and diagnosis. It not only has a low radiation dose but also provides a complete three-dimensional (3-D) chest image with a high resolution of lung nodules. In recent years, Convolutional Neural Network (CNN) had flourished in the field of medical images. It had been proven that a CNN-based computer-aided diagnosis system could extract the features through the machine learning method and help radiologists to make a preliminary diagnosis. Therefore, a Computer-Aided Diagnosis (CAD) system that used ResNeXt as the backbone was proposed to assist radiologists in diagnosis. Moreover, to prevent the backbone network from generating redundant features, a lightweight network module with an attention mechanism was embedded in our model to help focus on important features. Furthermore, a multi-level feature pyramid network that incorporates feature information of different scales was added to our model to enhance the prediction performance of small malignant nodules. Finally, a hybrid loss based on channel optimization was proposed to replace a binary cross-entropy (BCE) loss. It enabled the network to learn more detailed information to classify extreme cases. In this research, there were a total of 880 low-dose CT images from the American National Lung Screening Trial (NLST) for system training and testing, including 440 benign and 440 malignant lung nodules. The results of the experiments showed that our system could achieve an accuracy of 85.3%, the sensitivity of 86.8%, the specificity of 83.9%, and the area-under-curve (AUC) value was 0.9042. It was confirmed that the designed system had a good diagnostic ability. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/67039 |
DOI: | 10.6342/NTU202003614 |
全文授權: | 有償授權 |
顯示於系所單位: | 資訊網路與多媒體研究所 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
U0001-1608202020430800.pdf 目前未授權公開取用 | 1.73 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。