在血液透析療程中，動靜脈廔管扮演一個重要的角色，一個理想的動靜脈廔管其血流量為每分鐘500到1,500毫升之間，若在療程期間發生脫針，病患可能會有生命危險。本研究提出基於影像的方式去偵測紗布有無染血，偵測主要分為三個階段，依序為紗布偵測、紗布分割、分析出血嚴重性。首先使用YOLOv4從影像中偵測出紗布，然後使用Unet從影像中分割出紗布來做進一步分析，最後根據結果，分析出血嚴重性。然而對於深度學習來說，訓練一個好的影像偵測模型通常需要豐富的資料集，為了解決出血數據稀缺問題，本研究開發影像合成程式，產生多張出血紗布圖像，這大大地減少圖像收集階段所需之成本。為了證明基於影像合成之數據集可以應用於實境，自行模擬各種出血狀況，最後在此測試集得到精確率99.36%和召回率95.24%。希望藉此研究提高患者之安全性，同時減輕醫護人員的負擔。

Arteriovenous fistula play an important role in hemodialysis. The optimal blood flow of an arteriovenous fistula is between 500 and 1,500 mL/min. Once Venous needle dislodgement (VND) occurs during hemodialysis, the patient will be in danger. This study proposes an image-based approach to detect whether the gauze is stained with blood. The whole detection process is divided into three stages, which are gauze detection, gauze segmentation and analysis of bleeding severity. First, gauze detection from images using YOLOv4 deep learning network. Then gauze segmentation using Unet network for further analysis. Finally, analyzing the bleeding severity based on result. In general, when it comes to deep learning, the richer your dataset, the better your detection model performs. To solve data scarcity, this study develops an image synthesis program which generates lots of images of bloodstained gauze. It greater reduces the costs of data collection. To prove the model trained with artificial images can be applied to reality, this study simulates some blood leakage condition for testing. Finally, the system provided 99.36% precision and 95.24% recall. This study is expected to reduce the load on medical staff and improve patient safety.

第一章 緒論 1
1.1 研究背景與動機 1
1.2 研究目的 1
1.3 論文結構 2
第二章 相關文獻探討 3
2.1 血液透析出血偵測相關文獻 3
2.2 卷積神經網路 (CNN) 4
2.3 物件偵測模型YOLOV4 5
2.3.1 物件偵測網路 5
2.3.2 YOLOv4 6
2.3.3 有效率的訓練方法 13
2.4 影像分割模型UNET 14
2.5 失血量評估及相關文獻 15
2.6 模型評估指標 17
2.6.1 混淆矩陣 17
2.6.2 IOU 18
第三章 研究方法 21
3.1 資料集 21
3.2 紗布辨識 25
3.2.1 數據處理 25
3.2.2 訓練模型 30
3.3 紗布分割 32
3.3.1 數據處理 32
3.3.2 訓練模型 33
3.4 系統硬體及開發環境 34
第四章 實驗結果 35
4.1 紗布辨識結果 35
4.2 紗布分割結果 37
4.3 基於影像處理模擬出血可行性 40
4.4 結果討論 43
第五章 結論與未來展望 45
參考文獻 47

[1]許志成(民110年6月)。2020台灣腎病年報，40。取自https://lib.nhri.edu.tw/NewWeb/nhri/ebook/39000000465141/
[2]邱緯翔（2016）。利用陣列感測貼片實現穿戴式脫針漏血監測裝置。南臺科技大學生物醫學工程研究所碩士論文，台南市。 取自https://hdl.handle.net/11296/xrdn5x
[3]Nipro公司，https://pse.is/4dykf7
[4]Speranza-Reid, J., Brouwer-Maier, D., Cruz, C. M., & Inglese, M. (2021). Venous Needle Dislodgement and Access-Bloodline Separation. Nephrology Nursing Journal, 48(4), 347. https://doi.org/10.37526/1526-744x.2021.48.4.347
[5]Hu, H., Liu, C., Du, Y., Chen, K., Lin, H., & Lin, C. (2022). Real-time internet of medical things system for detecting blood leakage during hemodialysis using a novel multiple concentric ring sensor. Sensors, 22(5), 1988. doi:10.3390/s22051988
[6]Redmon, J., Divvala, S., Girshick, R., & Farhadi, A. (2016, May 09). You only look once: Unified, real-time object detection. Retrieved July 4, 2022, from https://arxiv.org/abs/1506.02640
[7]Bochkovskiy, A., Wang, C., & Liao, H. (2020, April 23). Yolov4: Optimal Speed and accuracy of object detection. Retrieved July 4, 2022, from https://arxiv.org/abs/2004.10934
[8]Wang, C., Liao, H., Yeh, I., Wu, Y., Chen, P., & Hsieh, J. (2019, November 27). CSPNet: A new backbone that can enhance learning capability of CNN. Retrieved July 4, 2022, from https://arxiv.org/abs/1911.11929
[9]Liu, S., Qi, L., Qin, H., Shi, J., & Jia, J. (2018). Path Aggregation Network for instance segmentation. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. doi:10.1109/cvpr.2018.00913
[10]Jeong, H., Park, K., & Ha, Y. (2018). Image preprocessing for efficient training of Yolo Deep Learning Networks. 2018 IEEE International Conference on Big Data and Smart Computing (BigComp). doi:10.1109/bigcomp.2018.00113
[11]Ronneberger, O., Fischer, P., & Brox, T. (2015). U-Net: Convolutional Networks for Biomedical Image Segmentation. Lecture Notes in Computer Science, 234-241. doi:10.1007/978-3-319-24574-4_28
[12]Editors, M. (2014, May 21). Medgadget. Retrieved July 5, 2022, from https://www.medgadget.com/2014/05/gauss-surgical-blood-loss-measurement-app-approved-by-fda-for-use-in-ors.html
[13]Li, Y. J., Zhang, L. G., Zhi, H. Y., Zhong, K. H., He, W. Q., Chen, Y., Yang, Z. Y., Chen, L., Bai, X. H., Qin, X. L., Li, D. F., Wang, D. D…(2020). A better method for the dynamic, precise estimating of blood/haemoglobin loss based on deep learning of artificial intelligence. Annals of translational medicine, 8(19), 1219. https://doi.org/10.21037/atm-20-1806
[14]Yolo_Label， https://github.com/developer0hye/Yolo_Label
[15]Darkent，https://github.com/pjreddie/darknet
[16]labelme，https://github.com/wkentaro/labelme