嗜中性球細胞外陷阱自2004年首次發現，也稱為Neutrophil “NETosis”，是免疫系統對抗微生物的重要免疫反應。使用嗜中性球細胞外陷阱的DNA螢光值配合螢光影像人工計數，來找到不同的藥物介入的與發生嗜中性球細胞外陷阱的相關性，仍然是目前關於嗜中性球細胞外陷阱研究的主要方式，但耗費時間、需要實驗技術。探討嗜中性球細胞外陷阱的研究每年明顯增加，使用自動化且準確有效率的定量方法顯得非常重要。以機器學習應用於嗜中性球細胞外陷阱的研究越來越多，但需要相當大量的樣本數做條件訓練，且主要是探討細胞形態上是否發生嗜中性球細胞外陷阱進而定量，一但發生細胞交疊、時間過長、或是嗜中性球細胞外陷阱太強則不易判讀；本研究希望跳脫大量樣本、型態學、分割和分類的方式，以機器學習合併半監督學習來預測嗜中性球細胞外陷阱的螢光影像輸入（fluorescence image input）與DNA螢光值輸出(value output)之間的關係。
材料與方法
添加SYTOX Green dye，通過螢光顯微鏡觀察嗜中性球細胞外陷阱的形成，並通過分光光度法對其進行觀察。使用螢光讀取器對發生嗜中性球細胞外陷阱的DNA片段進行螢光值定量。本研究提出捲積神經網路的模型，並以tensorflow 2.0在Google Colab實作，由於低倍率下螢光值對視窗輕度偏移敏感性低的假設，經預處理後的影像作爲輸入，螢光值作為標記，來訓練模型；因為翻轉和旋轉後螢光值不變的認知，對採樣進行水平和垂直翻轉，1張增多為4張；輸入為336 x 336的256階灰階影像，經過重複4次的conv + BN + ReLU後，接著最後一次的conv + global average pooling + fully connected layer。由於標記資料稀缺，本研究嘗試加入同一實驗室未標記的146張熒光攝影影像用於半監督學習。本研究收集259張樣本，其中113張有標記，隨機分配85張進行全監督學習為模型A，另外28作為測試組。146張沒有標記數值加上模型A的資料進行半監督學習作為為模型B，再以測試組進行模型A和模型B的回歸分析。
結果
兩組模型訓練完成之後進行以測試組進行回歸分析，模型A 、B的R值分別為0.84和0.88；模型A、B的誤差率平均值（95%信賴區間）分別為4.0378% (2.6720~5.4036%)，和3.9923 % (2.6324~5.3522%)；將這兩個模型誤差值進行paired samples t-test，雖然數值上有差異，半監督學習組結果數值較佳且較為穩定，但是未達到統計上的差異（p= 0.8434）。
結論
使用卷積神經網路結合半監督學習來預測嗜中性球細胞外陷阱螢光影像和DNA螢光值的關係，是準確且有效率的，並且節省時間與樣本數；使用半監督學習習的優勢是使無目標標記、稀少且昂貴的樣本變得可用，且不會傷害模型，本研究顯示數值上性能有提升，雖然未達統計上的差異；未來仍需要更多不同的實驗室配置與樣本進行條件訓練來強化和應證本研究的模型。

Summary
Neutrophil extracellular traps (NETs) are crucial for the immune response to microorganisms. Studies have revealed a close correlation between NETs and autoimmune diseases, cancer, and aging. The use of fluorescence images and DNA fluorescence values to analyze drugs’ inhibition effects on NETs is time-consuming. We therefore seek an automated, accurate, and efficient trap quantification method to resolve this. Phorbol myristate acetate was added to neutrophils and further cultured to induce the formation of NETs. Dye was added to samples to observe the extracellular DNA of the neutrophils through spectrometry. DNA fragments were quantified using a fluorescence reader. A convolutional neural network model was formulated and trained using preprocessed data as input data, and fluorescence values were applied as labels. A combination of machine learning and semi-supervised learning were used to predict the NETs’ DNA fluorescence values and to identify the relationship between fluorescence image input and DNA fluorescence value output for the NETs. The method used in this study was confirmed to be feasible, with a deviation rate of 3.9923% and a 95% CI of 2.6324%–5.3522%. Even with insufficient data, semi-supervised learning did not negatively affect the models but, rather, improved numerical performance. Although no statistical significance was achieved, semi-supervised learning enabled effective identification of numerous unlabeled clinical data. Semi-supervised learning enables unlabeled, rare, or expensive samples to be effectively used with no negative effects on models. The proposed method can replace analysis methods that require numerous samples, morphology, segmentation, and classification.
Key words: neutrophil extracellular traps, NETs, machine learning, convolutional neural network, semi-supervised learning

1 Boeltz, S. et al. To NET or not to NET:current opinions and state of the science regarding the formation of neutrophil extracellular traps. Cell Death Differ 26, 395-408, doi:10.1038/s41418-018-0261-x (2019).
2 Elsherif, L. et al. Machine Learning to Quantitate Neutrophil NETosis. Sci Rep 9, 16891, doi:10.1038/s41598-019-53202-5 (2019).
3 Cools-Lartigue, J., Spicer, J., Najmeh, S. & Ferri, L. Neutrophil extracellular traps in cancer progression. Cell Mol Life Sci 71, 4179-4194, doi:10.1007/s00018-014-1683-3 (2014).
4 Ortmann, W. & Kolaczkowska, E. Age is the work of art? Impact of neutrophil and organism age on neutrophil extracellular trap formation. Cell Tissue Res 371, 473-488, doi:10.1007/s00441-017-2751-4 (2018).
5 Sur Chowdhury, C. et al. Enhanced neutrophil extracellular trap generation in rheumatoid arthritis: analysis of underlying signal transduction pathways and potential diagnostic utility. Arthritis Res Ther 16, R122, doi:10.1186/ar4579 (2014).
6 Chen, M. S., Lin, W. C., Yeh, H. T., Hu, C. L. & Sheu, S. M. Propofol specifically reduces PMA-induced neutrophil extracellular trap formation through inhibition of p-ERK and HOCl. Life Sci 221, 178-186, doi:10.1016/j.lfs.2019.02.030 (2019).
7 Gonzalez, A. S., Bardoel, B. W., Harbort, C. J. & Zychlinsky, A. Induction and quantification of neutrophil extracellular traps. Methods Mol Biol 1124, 307-318, doi:10.1007/978-1-62703-845-4_20 (2014).
8 van Breda, S. V. et al. Computational Methodologies for the in vitro and in situ Quantification of Neutrophil Extracellular Traps. Front Immunol 10, 1562, doi:10.3389/fimmu.2019.01562 (2019).
9 Dong Wang, Y. Z., Kexin Zhang, Liwei Wang. FocalMix Semi-Supervised Learning for 3D Medical Image Detection. Computer Vision and Pattern Recognition (2020).
10 van der Linden, M., Westerlaken, G. H. A., van der Vlist, M., van Montfrans, J. & Meyaard, L. Differential Signalling and Kinetics of Neutrophil Extracellular Trap Release Revealed by Quantitative Live Imaging. Sci Rep 7, 6529, doi:10.1038/s41598-017-06901-w (2017).
11 Coelho, L. P. et al. Automatic determination of NET (neutrophil extracellular traps) coverage in fluorescent microscopy images. Bioinformatics 31, 2364-2370, doi:10.1093/bioinformatics/btv156 (2015).
12 Manda-Handzlik, A., Fiok, K., Cieloch, A., Heropolitanska-Pliszka, E. & Demkow, U. Convolutional Neural Networks-Based Image Analysis for the Detection and Quantification of Neutrophil Extracellular Traps. Cells 9, doi:10.3390/cells9020508 (2020).
13 Mohanty, T., Sorensen, O. E. & Nordenfelt, P. NETQUANT: Automated Quantification of Neutrophil Extracellular Traps. Front Immunol 8, 1999, doi:10.3389/fimmu.2017.01999 (2017).
14 Szegedy, S. I. C. Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. Cornell University (2015).
15 Agarap, A. F. Deep Learning using Rectified Linear Units (ReLU). Cornell University (2019).
16 Avital Oliver, A. O., Colin Raffel, Ekin D. Cubuk & Ian J. Goodfellow. Realistic Evaluation of Deep Semi-Supervised Learning Algorithms. Neural Information Processing Systems (2018).
17 Cheplygina, V., de Bruijne, M. & Pluim, J. P. W. Not-so-supervised: A survey of semi-supervised, multi-instance, and transfer learning in medical image analysis. Med Image Anal 54, 280-296, doi:10.1016/j.media.2019.03.009 (2019).