近年來，機器學習在各方面都有巨大的成功，像是物體識別、語意分析等，而在醫學影像方面也有突出的表現。周邊神經阻斷是一種局部麻醉技術，需要透過超音波掃描並找到影像上的神經位置，然而透過醫學專家人工圈選出在影像上的神經位置需要花費大量的時間。我們透過機器學習的方式訓練模型並能夠將即時切割出神經位置，有效減少大量人工診斷的時間。這項技術能夠應用在一種新的麻醉技術叫「超音波引導神經阻斷」，相較於全身麻醉，神經阻斷技術的副作用較少，我們與高雄榮民醫院(Kaohsiung Veterans General Hospital)與高雄醫學大學附設中和紀念醫院(Kaohsiung Medical University Chung-Ho Memorial Hospital)合作取得訓練資料。我們利用全卷積神經網絡並以U-net為架構的模型訓練資料;並且，我們將訓練好的模型進一步簡化它的參數量，透過簡化模型減少運算時間以及所需的記憶體，我們取得Dice coefficient 0.59的神經切割正確率、0.93的影像辨識準確率等結果。

In recent years, deep learning has achieved huge successes in many aspects such as object recognition and semantic analysis. Deep learning gets a great development in medical image as well. Peripheral Nerve Blocks is a type of regional anesthesia which need to find out the location of nerves and inject anesthetic nearby using ultrasound scanning. However, images recognition by medical experts is time-consuming. We trained a deep learning model which can perform segmentation of nerves in real time processing. It can be applied to an anesthesia technique called ‘Ultrasound-Guided Regional Anesthesia (UGRA). UGRA has less side-effect compare to general anesthesia. Our training dataset are acquired from two collaborative hospitals, Kaohsiung Veterans General Hospital (KVGH) and Kaohsiung Medical University Chung-Ho Memorial Hospital (KMUH). Our frameworks are based on the U-net model. Moreover, we prune our models using the Net-Trim algorithm, which is capable to reduce the parameters from a trained model. A simplified model consumes less prediction time and memory space. Our proposal achieves 0.59 Dice Coefficient and 0.93 accuracy for nerve segmentation.

中文摘要 I
Abstract II
誌謝 III
Contents 0
Chapter 1 Introduction 1
1.1 Motivation 1
1.2 Ultrasound Guided Regional Anesthesia 2
1.3 ANN and CNN 4
Chapter 2 Related Work 7
Chapter 3 Method 12
3.1 Pre-processing 13
3.2 Net-Trim algorithm 16
3.3 Proposed Network Architecture 17
3.4 Principal Component Analysis (PCA) 22
Chapter 4 Experiment Setup 24
4.1 Database 24
4.2 Setup 27
4.3 Evaluation 27
Chapter 5 Results and Discussion 29
5.1 Hyperparameters 29
5.2 Net-trim result 35
5.3 PCA removement result 38
5.4 Dataset 1 result and comparison to single U-net 41
5.5 Dataset 2 result and comparison to single U-net 47
Chapter 6 Conclusion and Future Work 53
Chapter 7 References 54

1.Ilfeld, B.M., T.E. Morey, and F.K. Enneking, Continuous Infraclavicular Brachial Plexus Block for Postoperative Pain Control at HomeA Randomized, Double-blinded, Placebo-controlled Study. Anesthesiology: The Journal of the American Society of Anesthesiologists, 2002. 96(6): p. 1297-1304.
2.Long, J., E. Shelhamer, and T. Darrell. Fully convolutional networks for semantic segmentation. in Proceedings of the IEEE conference on computer vision and pattern recognition. 2015.
3.Marhofer, P. and V.W. Chan, Ultrasound-guided regional anesthesia: current concepts and future trends. Anesth Analg, 2007. 104(5): p. 1265-9, tables of contents.
4.Walker, F.O., et al., Ultrasound of nerve and muscle. Clin Neurophysiol, 2004. 115(3): p. 495-507.
5.Gai, S., et al., Speckle noise reduction in medical ultrasound image using monogenic wavelet and Laplace mixture distribution. Digital Signal Processing, 2018. 72: p. 192-207.
6.Smistad, E. and F. Lindseth, Real-Time Automatic Artery Segmentation, Reconstruction and Registration for Ultrasound-Guided Regional Anaesthesia of the Femoral Nerve. IEEE Trans Med Imaging, 2016. 35(3): p. 752-61.
7.Sarle, W.S., Neural networks and statistical models. 1994.
8.LeCun, Y., Y. Bengio, and G. Hinton, Deep learning. nature, 2015. 521(7553): p. 436-444.
9.Krizhevsky, A., I. Sutskever, and G.E. Hinton. Imagenet classification with deep convolutional neural networks. in Advances in neural information processing systems. 2012.
10.Niu, X.-X. and C.Y. Suen, A novel hybrid CNN–SVM classifier for recognizing handwritten digits. Pattern Recognition, 2012. 45(4): p. 1318-1325.
11.Noble, A. and D. Boukerroui, Ultrasound image segmentation: a survey. IEEE Transactions on medical imaging, 2006. 25(8): p. 987-1010.
12.Ronneberger, O., P. Fischer, and T. Brox. U-net: Convolutional networks for biomedical image segmentation. in International Conference on Medical image computing and computer-assisted intervention. 2015. Springer.
13.Zhao, H. and N. Sun, Improved U-Net Model for Nerve Segmentation, in Image and Graphics. 2017. p. 496-504.
14.Ultrasound Nerve Segmentation. 2016; Available from: https://www.kaggle.com/c/ultrasound-nerve-segmentation/data.
15.Cahall, D.E., et al., Inception modules enhance brain tumor segmentation. Frontiers in computational neuroscience, 2019. 13: p. 44.
16.Ioffe, S. and C. Szegedy, Batch normalization: Accelerating deep network training by reducing internal covariate shift. arXiv preprint arXiv:1502.03167, 2015.
17.Baby, M. and A. Jereesh. Automatic nerve segmentation of ultrasound images. in 2017 International conference of Electronics, Communication and Aerospace Technology (ICECA). 2017. IEEE.
18.Lawrence, S., et al., Face recognition: A convolutional neural-network approach. IEEE transactions on neural networks, 1997. 8(1): p. 98-113.
19.Zhang, Q., et al. Image segmentation with pyramid dilated convolution based on ResNet and U-Net. in International Conference on Neural Information Processing. 2017. Springer.
20.Szegedy, C., et al. Inception-v4, inception-resnet and the impact of residual connections on learning. in Thirty-first AAAI conference on artificial intelligence. 2017.
21.Srivastava, N., et al., Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research, 2014. 15(1): p. 1929-1958.
22.Aghasi, A., et al. Net-trim: Convex pruning of deep neural networks with performance guarantee. in Advances in Neural Information Processing Systems. 2017.
23.Wold, S., K. Esbensen, and P. Geladi, Principal component analysis. Chemometrics and intelligent laboratory systems, 1987. 2(1-3): p. 37-52.
24.Abdi, H. and L.J. Williams, Principal component analysis. Wiley Interdisciplinary Reviews: Computational Statistics, 2010. 2(4): p. 433-459.
25.Manikanta Reddy, D. and H.C. Karnick, On segmentation of Nerve Structures in Ultrasound Images. 2016.
26.Juang, P.-A. and M.-N. Wu. Ultrasound speckle image process using wiener pseudo-inverse filtering. in IECON 2007-33rd Annual Conference of the IEEE Industrial Electronics Society. 2007. IEEE.
27.Hiremath, P., P.T. Akkasaligar, and S. Badiger, Speckle noise reduction in medical ultrasound images, in Advancements and breakthroughs in ultrasound imaging. 2013, Intechopen.
28.Perez, L. and J. Wang, The effectiveness of data augmentation in image classification using deep learning. arXiv preprint arXiv:1712.04621, 2017.
29.Agarap, A.F., Deep learning using rectified linear units (relu). arXiv preprint arXiv:1803.08375, 2018.
30.Aghasi, A., A. Abdi, and J. Romberg, Fast convex pruning of deep neural networks. arXiv preprint arXiv:1806.06457, 2018.
31.Li, Y., et al., U-net Ensemble Model for Segmentation inHistopathology Images. 2019.
32.Milletari, F., N. Navab, and S.-A. Ahmadi, V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation, in 2016 Fourth International Conference on 3D Vision (3DV). 2016. p. 565-571.