請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/82572
標題: | 利用特徵融合與注意力機制3-D捲積神經網路偵測電腦斷層影像之肺部結節 One-Stage Pulmonary Nodule Detection using 3-D DCNN with Feature Fusion and Attention Mechanism in CT Image |
作者: | 周秉儒 Ping-Ru Chou |
指導教授: | 張瑞峰 Ruey-Feng Chang |
關鍵字: | 肺癌,卷積神經網路,電腦輔助偵測系統,感受野模組,單次聚合模組,特徵融合策略, Lung cancer,Convolution neural network,Computer-aided detection,Receptive field block,One-shot aggregation,Feature fusion scheme, |
出版年 : | 2021 |
學位: | 碩士 |
摘要: | 近年來,肺癌為最常見的癌症之一,其死亡率在男性中位居第一,而在女性中僅次於乳癌,若提早發現肺癌,將可大幅降低死亡率,這使得早期偵測肺癌日益重要。低劑量電腦斷層掃描(Low-dose Computed Tomography, LDCT)已被廣泛地使用作為肺癌篩檢的重要工具,透過LDCT影像找出在肺部內形成且日後有可能會演變成肺癌的異常組織-結節。然而,醫師從LDCT影像觀察可疑的結節是一項非常耗時的工作,故電腦輔助偵測系統(Computer-Aided Detection System, CADe)被提出用來當作第二閱片者來幫助醫師。近年來,許多基於卷積神經網路(Convolution Neural Network, CNN)實現的CADe系統被證實能輔助醫師進行偵測。因此,在這篇研究中,我們提出一個基於YOLOv3偵測方法的CADe系統-OSAF-YOLOv3輔助醫師進行LDCT影像的肺結節偵測。 本篇論文所提出的CADe系統包含了資料前處理、結節偵測、以及非極大值抑制(Non-maximum Suppression, NMS)。在資料前處理中,首先會將非肺部的區域去除,再將LDCT影像得空間資訊歸一化到相同地比例尺,最後切割LDCT影像成多個感興趣區域(Volume of Interest, VOI)。接著,將切割的VOI輸入到OSAF-YOLOv3模型中進行結節偵測,由於在YOLOv3中,因為下採樣而導致小結節的空間資訊遺失,故對小結節的敏感性較低,因此,在OSAF-YOLOv3中,使用單次聚合模組和感受野模組來增強空間資訊,且提出特徵融合策略將不同大小的特徵圖融合並增強空間資訊。最後,將模型偵測的結果進行NMS演算法以消除重疊的偵測結果。在我們的實驗中,OSAF-YOLOv3在平均偽陽性(False Positive per Scan)為8的前提下,敏感度為96.2%並且在競爭性能度量(Competitive Performance Metric, CPM)上達到0.905。此外,原生的YOLOv3以及利用DenseBlocks取代YOLOv3裡ResBlocks的DenseNet-YOLOv3皆被訓練用來驗證提出的模型效果。根據結果,YOLOv3和DenseNet-YOLOv3分別取得CPM分數0.850和0.865,證實我們所設計的系統能有效的提升偵測能力。 Lung cancer is the most common cause of cancer-related death in men and the second most in women after breast cancer. Early detection has become crucial to reduce the mortality rate. Low-dose computed tomography (LDCT) is a widely used image screening in lung cancer detection. The nodule is an abnormal tissue that forms in the lungs and may evolve into lung cancer. Hence, it is crucial to detect nodules in the early detection stage. However, reviewing the LDCT image by the radiologist to observe suspicious nodules is a time-consuming task. Therefore, the computer-aided detection (CADe) system is proposed as the second reviewer to assist radiologists. Recently, designing a CADe system with convolutional neural network (CNN) architecture has been proven that it is helpful for radiologists. Hence, in this study, a CADe system, OSAF-YOLOv3 improved from You Only Look Once version 3 (YOLOv3), is proposed for nodule detection in LDCT images. The proposed CADe system consists of data preprocessing, nodule detection, and the non-maximum suppression algorithm (NMS). In the data preprocessing, the background elimination removes the useless region (non-lung region), the spacing normalization, which normalizes the LDCT images into identity spacing, and the volume of interest (VOI) extraction, which divides LDCT into numerous VOIs, are conducted. Then, the extracted VOIs are fed into the OSAF-YOLOv3 model to detect the suspicious nodules. In YOLOv3, the spatial information of small nodules is missing due to downsampling; and as a result, the sensitivity rate on small nodules is low. Hence, in the OSAF-YOLOv3, the one-shot aggregation module version 2 (OSAv2) and receptive field block (RFB) are adopted to enhance spatial information. Furthermore, the feature fusion scheme (FFS) is proposed for combining the different resolution features to enrich the spatial information. Finally, the NMS algorithm is performed to eliminate the duplicated detection generated by the model. In our experiment result, the proposed system can achieve a sensitivities rate of 94.0% with the false positive rate of 8 and complete a competition performance metric (CPM) value of 0.877. Moreover, the primitive YOLOv3 and the DenseNet-YOLOv3, which replace the ResNet block of primitive YOLOv3 with dense connection block (DenseBlock), are tested to validate our model performance. The performance of the YOLOv3 and the DenseNet-YOLOv3 only achieves CPM scores of 0.828 and 0.843. The result indicates that the proposed system can significantly improve detection performance. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/82572 |
DOI: | 10.6342/NTU202101250 |
全文授權: | 同意授權(限校園內公開) |
電子全文公開日期: | 2023-07-29 |
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-109-2.pdf 授權僅限NTU校內IP使用(校園外請利用VPN校外連線服務) | 2.05 MB | Adobe PDF | 檢視/開啟 |
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。