曾明性Ming-Hseng Tseng洪嘉宏Chia-Hung Hung醫學資訊學系2026-10-132021-02-1720212021-01-27https://handle.ncl.edu.tw/11296/jaafvwhttps://scholars.csmu.edu.tw/handle/123456789/21201民國108年國人十大死因中，惡性腫瘤（癌症）位居第一，氣管、支氣管和肺癌為前十大惡性腫瘤排名之首。現今肺癌較新的篩檢方式，為透過低劑量電腦斷層（low dose spiral CT scan）產出的CT影像，再由醫療人員以人工的方式進行診斷。隨著時代的進步，影像辨識技術更趨成熟，其中深度學習在影像辨識的任務上有很大的貢獻。同時，在醫學上有一些影像辨識的輔助診斷工具開始發展。深度學習模型又被稱作「黑盒子」，在深度學習模型的分類任務上，往往不知道電腦是如何去做判斷，所以深度學習模型的可詮釋性顯得非常重要。本研究使用Lung Image Database Consortium（LIDC）的肺部CT影像資料集，訓練一個可詮釋性的3D卷積神經網絡。並以分散式平行化運算來加速深度學習模型的訓練，進行肺部結節良性、惡性之分類，並與其他學者提出的分類模型進行比較。Cancer took the top spot in Taiwan’s leading causes of death in 2019, and the three deadliest types of cancer in 2019 were lung, trachea and bronchus cancer. Currently, it is recommended to use Low-Dose spiral CT scan (also called LDCT) to produce CT image and manually diagnosed by medical personnel. With the ever-changing nature of image recognition technology, deep learning also has made a significant contribution to image recognition. Meanwhile, there are many kinds of image recognition auxiliary diagnostic tools starting to develop. Deep learning models also called “black box”. Since understanding how the final output was arrived at is exceedingly difficult, it makes the interpretability of deep learning models more important. This research uses Lung Image Database Consortium (LIDC) lung CT image data set to train an interpretable 3D convolutional neural network. Moreover, the study uses decentralized distributed parallel computing to accelerate deep learning models training and classify benign and malignant lung nodules.肺癌CT影像深度學習可詮釋性3D卷積神經網絡分散式平行化運算Lung cancerCT imagesdeep learninginterpretability3D convolutional neural networkdistributed parallel computingthesis