- 學位論文
應用三維注意力拆分機制和三維空間分組增強模塊搜尋神經網路架構於肺部電腦斷層掃描結節診斷
Diagnosis of Lung Nodules on Computed Tomography Images Using Neural Architecture Search with 3-D Split-attention Mechanism and 3-D Spatial Group-wise Enhance Block
摘要
近年來，在全球因癌症而死亡的人中，肺癌已佔據極大比例，透過早期肺結節診斷，可以有效降低肺癌的死亡率。電腦斷層掃描(computed tomography, CT)可以呈現整個胸腔的完整三維視圖且具有高解析度等優點，被廣泛運用在肺結節診斷中。但醫師透過電腦斷層掃描影像診斷肺結節時，依然會因為不同肺結節的形狀、紋理等特徵而影響判斷，甚至出現診斷錯誤的可能性。而近年來卷積神經網路(convolutional neural network, CNN)在醫學影像領域蓬勃發展，其自動提取影像特徵的特性可以在不同任務上達到良好表現，因此我們提出一個以卷積神經網路為基礎的電腦輔助診斷系統(computer-aided diagnosis system)來協助醫生對肺結節進行診斷。我們的系統由影像前處理和肺結節分類兩個部分組成，前處理的部分我們將感興趣的區域從原始電腦斷層掃描影像中擷取出來，包含肺結節及其周圍組織，再把它丟到我們提出的模型中。我們提出的3-D NAS-SGE-SANet是利用三維注意力拆分機制和三維空間分組增強模塊(spatial group-wise enhance and split-attention block, SGE-SA)和梯度提升機(gradient boosting machine, GBM)，搭配神經網路架構搜索(neural architecture search, NAS)尋找出來的架構。本研究總共使用了716張電腦斷層掃描影像，其中包含302個良性結節和414個惡性結節。實驗結果顯示，所提出的系統能達到準確率93.44%、靈敏性92.51%、特異性94.70%和ROC曲線下面積0.9663的成果，顯示出我們提出的系統有非常好的診斷能力。
並列摘要
Recently, lung cancer has been pointed out as a leading cause of cancer death worldwide. Through early lung nodules diagnosis, the mortality rate of lung cancer could be effectively reduced. Computed Tomography (CT) was an essential tool for lung nodules diagnosis since it provided a complete three-dimensional (3-D) chest image with high resolution and detailed information. However, even the experienced physician might be susceptible to the nodule's shape, texture, and other hand-craft features while reviewing the CT images and made the wrong judgment. In a recent, the convolutional neural network (CNN) has flourished in medical images. It could automatically extract the features from the input images and achieve high performance on the task. Therefore, the computer-aided diagnosis (CADx) system based on CNN architecture was proposed to assist the physicians in lung nodule diagnosis. In this study, the proposed CADx system was composed of image preprocessing and nodule classification. In the image preprocessing, the volume of interest containing the nodule and surrounding tissue was extracted from the CT images. After that, in the nodule classification, a novel 3-D NAS-SGE-SANet designed by neural architecture search (NAS) with 3-D spatial group-wise enhance and split-attention block (SGE-SA), and the gradient boosting machine (GBM) was present to determine the nodule as benign or malignant. In experiments, a total of 716 CT images, including 302 benign and 414 malignant lung nodules, were used in this study to evaluate the performance of the proposed system. The results showed that our system could achieve an accuracy of 93.44%, sensitivity of 92.51%, specificity of 94.70%, and the area under the ROC curve (AUC) of 0.9663. It was confirmed that the proposed system had an excellent diagnostic capability.