從 2019 年底至今,新冠肺炎(Covid-19)造成全世界數以萬計的人染疫。能夠及早發現染疫的病患並且即時隔離及投入醫療資源照護,降低病患發生重症的機率,並且減少疾病持續傳遞的風險。目前最常用於檢測新冠肺炎的方式為,RT-PCR 試劑,但疫情爆發的地區可能發生試劑短缺的問題。所以我希望利用深度學習提出一電腦輔助系統,搭配由於新冠肺炎疫情進行大規模電腦斷層肺部掃描的影像,輔助包含新冠肺炎在內肺部疾病的診斷。
在研究中,使用專門應用於圖象分類的卷積神經網路 VGG16 以及VGG19 兩種模型,並且經過 ImageNet 預訓練。接上自己設計的分類器,對電腦斷層肺部影像進行分類,類別包含正常肺部影像、肺部腫瘤影像、一般肺炎影像、新冠肺炎影像共 4 種。最終得到最佳的結果,對一般正常肺部得到 96%的準確率,對於肺部腫瘤得到 89%準確率,對新冠肺炎得到 87%準確率,對一般肺炎得到 78%正確率。
關鍵字:新冠肺炎、卷積神經網路、電腦輔助診斷、遷移學習

Since the end of 2019, COVID-19 has infected tens of thousands of people around the world.It is important to detect infected patients as early as possible, isolate them immediately and devote medical resources to care, reduce the chance of severe illness, and reduce the risk of continuous transmission of the disease.
At present, the most common method used to detect COVID-19 is RT-
PCR(Real-time Polymerase Chain Reaction) reagents. But there may be
problems such as shortage of reagents in some place where COVID-19 outbreak severely.
Therefore, I hope to use deep learning to propose a computer-aided
system, using large number of computer tomography lung scans images ,helping the diagnosis of lung diseases including COVID-19.
n the research, two models of convolutional neural network VGG16
and VGG19, which are specially applied to image classification, are used,and they are pre-trained by ImageNet. Adding the classifier of my own design, the computerized tomography lung images are classified. The categories include normal lung images, lung tumor images, commom pneumonia images, and COVID-19 images.
In the end, the best results obtained, with 96% accuracy for normal
lungs, 89% accuracy for lung tumors, 87% accuracy for COVID-19 , and
78% accuracy for common pneumonia.
keywords:Covid-19,Convolutional neural networks, computer-aided
system,transfer learning

目錄
1. 前言 ...................................................................................... - 1 -
2. 類神經網路 ........................................................................... - 4 -
2.1
超參數 ......................................................................... - 5 -
2.1.1 學習率(Learning rate)........................................... - 6 -
2.1.2 批量大小(Batch size) ........................................... - 7 -
2.1.3 迭代(iteration)與時期(epochs) ............................. - 7 -
2.1.4 損失函數(Loss function) ...................................... - 7 -
2.1.5 優化器(Optimizer) ................................................ - 9 -
2.1.6 正則化(Regularization) ...................................... - 13 -
2.1.7 激活函數(Activation function) ........................... - 15 -
3.
2.2 倒傳遞類神經網路 ................................................... - 22 -
2.3 過度擬和 ................................................................... - 23 -
卷積神經網路 ..................................................................... - 25 -
3.1 卷積層(Convolution Layer)....................................... - 26 -
3.2 池化層(Pooling Layer) .............................................. - 28 -
i3.3
全連接層(Fully connected Layer) ............................. - 29 -
3.4 AlexNet ........................................................................ - 31 -
3.4.1 AlexNet 架構 ....................................................... - 32 -
3.5 VGG ............................................................................. - 33 -
3.5.1 VGG 架構與參數 ................................................ - 35 -
3.5.2 VGG 模型的訓練 ................................................ - 37 -
3.5.3 VGG 模型的測試 ................................................ - 37 -
4.
遷移學習 ............................................................................. - 40 -
4.1 Fine-tuning ................................................................... - 42 -
4.2 Multitask learning ......................................................... - 44 -
4.3 Domain-adversarial training .......................................... - 46 -
4.4 Zero-shot learning ......................................................... - 47 -
5.
實驗設計 ............................................................................. - 48 -
5.1 數據庫(Database) ...................................................... - 48 -
5.2 影像預處理(Image processing) ................................. - 49 -
5.2.1 影像大小調整(Resize) ....................................... - 49 -
5.2.2 正規化(Normalization) ....................................... - 50 -
ii5.2.3 影像資料增強(Data augumentation) .................. - 50 -
5.3
實驗架構 ................................................................... - 51 -
6. 評估指標 ............................................................................. - 53 -
7. 實驗結果與討論 ................................................................. - 57 -
7.1
一般肺炎、新冠肺炎、肺部腫瘤、肺部正常無異的分
- 58 -
類
7.2
肺部正常及肺部有異狀分類(包含肺部腫瘤、新冠肺炎
以及一般肺炎)............................................................................. - 62 -
7.3 肺部腫瘤及肺炎分類(包含一般肺炎及新冠肺炎) .. - 66 -
7.4 一般肺炎及新冠肺炎分類 ....................................... - 70 -
7.5 錯誤分類的影像 ....................................................... - 74 -
8. 結論 .................................................................................... - 76 -
9. 參考文獻 ............................................................................. - 80 -
圖目錄
圖 1、類神經網路架構 ........................................................ - 5 -
圖 2、學習率對損失值影響 ................................................. - 6 -
圖 3、局部極值 .................................................................... - 9 -
圖 4、使用 dropout 隨機抑制神經元 ............................... - 14 -
圖 5、Sigmoid 函數圖 ........................................................ - 17 -
圖 6、Tanh 函數圖 ............................................................. - 18 -
圖 7、Sigmoid 與 Tanh 函數比較 ...................................... - 18 -
圖 8、ReLu(線性修正單元)函數圖 ................................... - 19 -
圖 9、Softmax 函數轉換為機率示意圖 ............................. - 20 -
圖 10、Threshold 函數圖 .................................................. - 21 -
圖 11、倒傳遞類神經網路示意圖 ..................................... - 22 -
圖 12、訓練曲線擬和情形。(a)低度擬和(b)理想的擬和(c)過
度擬和 .......................................................................... - 23 -
圖 13、LeNet 架構示意圖 .................................................. - 25 -
圖 14、卷積層運算 ............................................................ - 27 -
圖 15、最大池化 ................................................................ - 28 -
圖 16、平均池化 ................................................................ - 29 -
圖 17、全連接層分類過程 ................................................. - 30 -
圖 18、AlexNet 架構示意圖 .............................................. - 31 -
圖 19、(a) VGG16 (b) VGG19 示意圖 ............................... - 34 -
圖 20、兩個 3×3 卷積核形成相當於 5×5 卷積核的感官視野 . -
36 -
圖 21、訓練階段與測試階段全連接層的差異(a)為訓練階段
(b)為測試階段 ............................................................ - 38 -
圖 22、固定所有卷積層參數 ............................................. - 42 -
圖 23、將預訓練模型的參數作為初始值衝新訓練 .......... - 43 -
圖 24、固定部分卷積層參數,另一部分重新訓練 .......... - 44 -
圖 25、Multitask learning 合併訓練示意圖 ....................... - 45 -
圖 26、實驗使用神經網路模型設計 ................................. - 52 -
圖 27、ROC 曲線 ............................................................... - 55 -
圖 28、閾值的調整影響曲線的座標 ................................. - 55 -
圖 29、(a)鑑別能力良好(b)鑑別能力普通(c)鑑別能力差 - 56 -
圖 30、第一階段的實驗分類 ............................................. - 57 -
圖 31、第二階段的實驗分類 ............................................. - 57 -
圖 32、VGG16 混淆矩陣 ................................................... - 60 -
圖 33、VGG19 混淆矩陣 ................................................... - 60 -
圖 34、VGG16 訓練過程 ................................................... - 60 -
圖 35、VGG19 訓練過程 ................................................... - 61 -
圖 36、未使用 Fine-tuning 方式 VGG16 的混淆矩陣 ...... - 61 -
圖 37、VGG16 混淆矩陣 ................................................... - 63 -
圖 38、VGG19 混淆矩陣 ................................................... - 63 -
圖 39、 VGG16 訓練過程 ................................................. - 63 -
圖 40、VGG19 訓練過程 ................................................... - 64 -
圖 41、未使用 Fine-tuning 方式 VGG16 的混淆矩陣 ...... - 64 -
圖 42、VGG16 混淆矩陣 ................................................... - 67 -
圖 43、VGG19 混淆矩陣 ................................................... - 67 -
圖 44、VGG16 訓練過程 ................................................... - 67 -
圖 45、VGG19 訓練過程 ................................................... - 68 -
圖 46、未使用 Fine-tuning 方式 VGG16 的混淆矩陣 ...... - 68 -
圖 47、VGG16 混淆矩陣 ................................................... - 71 -
圖 48、VGG19 混淆矩陣 ................................................... - 71 -
圖 49、VGG16 訓練過程 ................................................... - 71 -
圖 50、VGG19 訓練過程 ................................................... - 72 -
圖 51、未使用 Fine-tuning 方式 VGG16 的混淆矩陣 ...... - 72 -
圖 52、肺部腫瘤分類正確影像及錯誤影像 ..................... - 74 -
圖 53、新冠肺炎分類正確影像及錯誤影像 ..................... - 74 -
圖 54、一般肺炎分類正確影像及錯誤影像 ..................... - 75 -
圖 55、正常無異狀分類正確影像及錯誤影像.................. - 75 -
圖 56、最佳結果使用的架構 ............................................. - 78 -
表目錄
表 1、遷移學習類別 .......................................................... - 41 -
表 2、混淆矩陣 .................................................................. - 53 -
表 3、兩個模型對四種類別的精確度、召回值、F1 值及正確
率 .................................................................................. - 59 -
表 4、肺部有異狀及肺部無異狀兩分類模型準確度、敏感度
及特異性 ...................................................................... - 65 -
表 5、肺部有異狀及肺部無異狀影像兩分類精確度、召回率
及 F1 值 ........................................................................ - 65 -
表 6、肺部腫瘤及肺炎影像兩分類模型準確度、敏感度及特
異性 .............................................................................. - 69 -
表 7、肺部腫瘤及肺炎影像兩分類精確度、召回率及 F1 值 . -
69 -
表 8、一般肺炎及新冠肺炎影像兩分類模型準確度、敏感度
及特異性 ...................................................................... - 73 -
表 9、一般肺炎及新冠肺炎影像兩分類模型精確度、召回率
及 F1 值 ........................................................................ - 73 -

[1] Zhu, N., Zhang, D., Wang, W., Li, X., Yang, B., Song, J., ... & Tan, W.(2020). A novel coronavirus from patients with pneumonia in China,2019. New England journal of medicine.
[2] Guan, W. J., Ni, Z. Y., Hu, Y., Liang, W. H., Ou, C. Q., He, J. X., ... & Zhong, N. S. (2020). Clinical characteristics of coronavirus disease 2019 in China. New England journal of medicine, 382(18), 1708-1720.
[3] World Health Organization. Coronavirus disease (COVID-19) pandemic. Geneva:World Health Organization;2000
https://www.who.int/emergencies/diseases/novel-coronavirus-2019
(accessed 25 May 2020) .
[4] Arevalo-Rodriguez, I., Buitrago-Garcia, D., Simancas-Racines, D.,
Zambrano-Achig, P., Del Campo, R., Ciapponi, A., ... & Zamora, J.
(2020). False-negative results of initial RT-PCR assays for COVID-19:
a systematic review. PloS one, 15(12), e0242958.
[5] Purohit, K., Kesarwani, A., Kisku, D. R., & Dalui, M. (2020). Covid-19 detection on chest x-ray and ct scan images using multi-imageaugmented deep learning model. BioRxiv.
[6] Hasan, A. M., Al-Jawad, M. M., Jalab, H. A., Shaiba, H., Ibrahim, R. W.,& AL-Shamasneh, A. A. R. (2020). Classification of covid-19
coronavirus, pneumonia and healthy lungs in ct scans using q-deformed
entropy and deep learning features. Entropy, 22(5), 517.
[7] Fang, Y., Zhang, H., Xie, J., Lin, M., Ying, L., Pang, P., & Ji, W. (2020).Sensitivity of chest CT for COVID-19: comparison to RT-
PCR. Radiology, 296(2), E115-E117.
[8] Li, X., Fang, X., Bian, Y., & Lu, J. (2020). Comparison of chest CT findings between COVID-19 pneumonia and other types of viral
pneumonia: a two-center retrospective study. European radiology, 30,
5470-5478.
[9] Deng, L., & Yu, D. (2014). Deep learning: methods and
applications. Foundations and trends in signal processing, 7(3–4), 197-387.
[10] Yang, X., He, X., Zhao, J., Zhang, Y., Zhang, S., & Xie, P. (2020).
COVID-CT-Dataset: a CT image dataset about COVID-19. arXiv
preprint arxiv:2003.13865, 3.
[11] Peng, Y., Tang, Y. X., Lee, S., Zhu, Y., Summers, R. M., & Lu, Z. (2020).
COVID-19-CT-CXR: a freely accessible and weakly labeled chest X-
ray and CT image collection on COVID-19 from biomedical
literature. arXiv preprint arXiv:2006.06177.
[12] Jin, C., Chen, W., Cao, Y., Xu, Z., Zhang, X., Deng, L., ... & Feng, J. Development and evaluation of an AI system for COVID-19 diagnosis.medRxiv2020. preprint[https://doi.org/10.1101/2020.03.
20.20039834].
[13] Zhang, Z., & Sabuncu, M. R. (2018). Generalized cross entropy loss for training deep neural networks with noisy labels. arXiv preprint arXiv:1805.07836.
[14] Kingma, D. P., & Ba, J. (2014). Adam: A method for stochastic
optimization. arXiv preprint arXiv:1412.6980.
[15] Glorot, X., Bordes, A., & Bengio, Y. (2011, June). Deep sparse rectifier neural networks. In Proceedings of the fourteenth internationalconference on artificial intelligence and statistics (pp. 315-323). JMLR Workshop and Conference Proceedings.
[16] Ramachandran, P., Zoph, B., & Le, Q. V. (2017). Swish: a self-gated activation function. arXiv preprint arXiv:1710.05941, 7, 1.
[17] Sermanet, P., Eigen, D., Zhang, X., Mathieu, M., Fergus, R., & LeCun,Y. (2013). Overfeat: Integrated recognition, localization and detection using convolutional networks. arXiv preprint arXiv:1312.6229.
[18] LeCun, Y. (2015). LeNet-5, convolutional neural networks. URL:
http://yann. lecun. com/exdb/lenet, 20(5), 14.
[19] Albawi, S., Mohammed, T. A., & Al-Zawi, S. (2017, August).
Understanding of a convolutional neural network. In 2017 International Conference on Engineering and Technology (ICET) (pp. 1-6). Ieee.
[20] Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). Imagenet
classification with deep convolutional neural networks. Advances in
neural information processing systems, 25, 1097-1105.
[21] Simonyan, K., & Zisserman, A. (2014). Very deep convolutional
networks for large-scale image recognition. arXiv preprint
arXiv:1409.1556.
[22] Pan, S. J., & Yang, Q. (2009). A survey on transfer learning. IEEE Transactions on knowledge and data engineering, 22(10), 1345-1359.
[23] Torrey, L., & Shavlik, J. (2010). Transfer learning. In Handbook of research on machine learning applications and trends: algorithms,
methods, and techniques (pp. 242-264). IGI global.
[24] Yosinski, J., Clune, J., Bengio, Y., & Lipson, H. (2014). How
transferable are features in deep neural networks?. arXiv preprint
arXiv:1411.1792.
[25] Ganin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H.,Laviolette, F., ... & Lempitsky, V. (2016). Domain-adversarial training of neural networks. The journal of machine learning research, 17(1),2096-2030.
[26] Johnson, M., Schuster, M., Le, Q. V., Krikun, M., Wu, Y., Chen, Z., ... & Dean, J. (2017). Google’s multilingual neural machine translation system: Enabling zero-shot translation. Transactions of the Association for Computational Linguistics, 5, 339-351.
[27] Ning, W., Lei, S., Yang, J., Cao, Y., Jiang, P., Yang, Q., ... & Wang, Z.(2020). Open resource of clinical data from patients with pneumonia for the prediction of COVID-19 outcomes via deep learning. Nature biomedical engineering, 4(12), 1197-1207.
[28] Yan, T., Wong, P. K., Ren, H., Wang, H., Wang, J., & Li, Y. (2020).Automatic distinction between covid-19 and common pneumonia using multi-scale convolutional neural network on chest ct scans. Chaos,Solitons & Fractals, 140, 110153.
[29] Al-Yasriy, H. F., AL-Husieny, M. S., Mohsen, F. Y., Khalil, E. A., & Hassan, Z. S. (2020, November). Diagnosis of Lung Cancer Based on CT Scans Using CNN. In IOP Conference Series: Materials Science and Engineering (Vol. 928, No. 2, p. 022035). IOP Publishing.
[30] Shorten, C., & Khoshgoftaar, T. M. (2019). A survey on image data augmentation for deep learning. Journal of Big Data, 6(1), 1-48.
[31] Ibrahim, D. M., Elshennawy, N. M., & Sarhan, A. M. (2021). Deep-chest:Multi-classification deep learning model for diagnosing COVID-19,pneumonia, and lung cancer chest diseases. Computers in biology andmedicine, 132, 104348.
[32] Hanley, J. A., & McNeil, B. J. (1982). The meaning and use of the area under a receiver operating characteristic(ROC)curve. Radiology, 143(1), 29-36.
[33] Huang, J., & Ling, C. X. (2005). Using AUC and accuracy in evaluating learning algorithms. IEEE Transactions on knowledge and Data Engineering, 17(3), 299-310.