臺灣博碩士論文加值系統

探討醫學臨床研究問題時，常需要量測有興趣的特徵，並進行統計來驗證假說。然而每一種特徵的量測需要花很多人力及時間，加上由於需要人工篩選，只有少量的資料能進行統計，並且找尋的特徵有時會受限於人的想像力，無法更廣泛的探索。因此若能從大量資料表現結果，縮小特徵找尋的範圍，並提供可能的線索特徵，便能加快研究進展。

卷積神經網路能從大量的資料中，主動學習重要特徵，並且透過可視化工具來協助找尋此特徵的位置。因此本研究採用深度學習演算法，利用ResNet、VGG及Inception等卷積神經網路，透過3134位正常人眼睛黃斑部之光學同調斷層掃描影像，學習分辨年齡及性別。經過訓練及驗證後，模型能在測試資料集中達高於75%以上的性別預測準確率以及低於6年的年齡預測誤差。從可視化工具Grad-CAM、Guided Grad-CAM及影像裁切實驗，深度探討後發現(1)男女的黃斑部中央凹有差異，(2)差異的特徵應為中央凹的色素上皮層及內界膜輪廓，(3)黃斑部各層組織與脈絡膜應會隨著年齡而有變化。

過去在眼科學界的認知中，黃斑部的結構並沒有具臨床意義的性別差異。本研究透過深度學習發現黃斑部的中央凹輪廓具有性別差異，而由於根據研究顯示，在OCT上呈現寬中心凹的眼睛，雖然視力完全正常，但其對側眼卻有極高的比例產生特發性黃斑前膜(51.9 %)及黃斑部裂孔(9.6 %)等黃斑部疾病，暗示中央凹輪廓可能是具有臨床意義之性別差異。期望本研究的結果能提供眼科醫師對於黃斑部新的研究方向，未來在流行病學、致病機轉、或是疾病的預防、治療及預後等各方面有所幫助。

When solving clinical problems, features are selected and measured to support a hypothesis. However, it takes a lot of manpower and time to manually measure these features, therefore the amount of measured data is relatively small. Moreover, the characteristics to be searched for are sometimes limited by human imagination. Therefore, convolutional neural networks are trained using deep learning methods, which can automatically detect important features that are difficult or impossible to be recognized by humans, and the amount of analyzed data is much higher.

In this thesis, ResNet, VGG, and Inception are used to distinguish the gender and age of 3134 healthy Taiwanese based on images of the macular region by means of optical coherence tomography. The gender-related and age-related features in the images are detected and showed using grad-CAM, a model visualization tool.

After training and verification, the model can achieve a gender prediction accuracy of more than 75% and an age error of less than 6 years in the test data set. A thorough analysis with Grad-CAM, Guided Grad-CAM, and image cropping experiments, shows that (1) the fovea in the macula has differences between men and women, (2) the differences should be characterized by the retina pigment epithelium and the inner limiting membrane contours of the fovea, and (3) the tissues and choroids of the macula should change with age.

In the past, there was no clinically meaningful gender difference in the structure of the macula. This study found that the contours of the fovea are different for men and women, and according to the research, the eyes with wide fovea pit on the OCT, although the vision is completely normal, the fellow eyes have a very high proportion of macular diseases such as epiretinal membrane (51.9%) and macular hole (9.6%), implicating that the foveal contours may be a clinically meaningful gender difference. It is hoped that the results of this study can provide ophthalmologists with a new research direction for the macula, and will help in the future in epidemiology, pathogenic mechanism, or disease prevention, treatment, and prognosis.

誌謝 I
摘要 II
Abstract III
目錄 V
圖目錄 VIII
表目錄 XII
第一章 緒論 1
1.1背景及研究動機 1
1.2本文內容概述 2
第二章 頻域式光學同調斷層掃描及視網膜結構介紹 4
2.1頻域式光學同調斷層掃描 4
2.1.1 譜域式光學同調斷層掃描儀之原理 4
2.1.2 海德堡光學同調斷層掃描顯微鏡 6
2.1.3 掃頻式光學同調斷層掃描儀 6
2.2眼睛視網膜介紹 7
2.2.1 視網膜組織結構 7
2.2.2 視網膜的生理功能 9
2.2.3 黃斑區與黃斑中心凹 11
2.3黃斑部與年齡性別關聯 11
第三章 卷積神經網路在OCT影像之應用 14
3.1卷積神經網路用於影像判讀 14
3.1.1 卷積神經網路的基本架構及工作原理 15
3.1.2 反向傳播用於梯度計算及權重計算 17
3.1.3 Dropout、正則化及數據增強 20
3.1.4 遷移學習 24
3.2卷積神經網路之解釋工具-類激活圖 29
3.3視網膜OCT影像資訊及預處理方法 33
3.3.1 影像資訊-數量、標記、分類 34
3.3.2 影像數值預處理 37
3.3.3 自動偵測並裁剪影像 40
第四章 影像預測結果分析 41
4.1性別預測結果及分析 42
4.1.1 性別差異之特徵位置 44
4.1.2 中央凹特徵-色素上皮層及內界膜輪廓 46
4.1.3 色素上皮層及內界膜輪廓與厚度之關聯 53
4.2年齡預測結果及分析 63
4.3年齡性別同時預測結果及分析 65
4.4減緩過擬合之效果 67
第五章 結論與未來展望 69
5.1結論 69
5.2未來展望 70
參考文獻 72
附錄1本論文使用之模型架構與訓練細節 77
VGG11 78
VGG19 79
ResNet18 81
ResNet34 82
Inception V3 84
附錄2 海德堡資料之Python讀取程式說明 85
附錄3 訓練結果補充 87

[1]Y. Hsieh, Ma IH., "Wide Foveal Pit - Predilection for Maculopathy," presented at the APVRS Congress 2019, Shanghai, China, 2019.

[2]M. Wagner-Schuman et al., "Race-and sex-related differences in retinal thickness and foveal pit morphology," Investigative Ophthalmology & Visual Science, vol. 52, no. 1, pp. 625─634, 2011.

[3]A. Wong, C. Chan, and S. Hui, "Relationship of gender, body mass index, and axial length with central retinal thickness using optical coherence tomography," Eye, vol. 19, no. 3, pp. 292─297, 2005.

[4]C. A. McCannel, J. L. Ensminger, N. N. Diehl, and D. N. Hodge, "Population-based incidence of macular holes," Ophthalmology, vol. 116, no. 7, pp. 1366─1369, 2009.

[5]W. Xiao, X. Chen, W. Yan, Z. Zhu, and M. He, "Prevalence and risk factors of epiretinal membranes: a systematic review and meta-analysis of population-based studies," BMJ open, vol. 7, no. 9, p. e014644, 2017.

[6]S. J. Park, N.-K. Choi, K. H. Park, and S. J. Woo, "Nationwide incidence of clinically diagnosed retinal vein occlusion in Korea, 2008 through 2011: preponderance of women and the impact of aging," Ophthalmology, vol. 121, no. 6, pp. 1274─1280, 2014.

[7]R. Poplin et al., "Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning," Nature Biomedical Engineering, vol. 2, no. 3, p. 158, 2018.

[8]R. J. Sidd, S. L. Fine, S. L. Owens, and A. Patz, "Idiopathic preretinal gliosis," American Journal of Ophthalmology, vol. 94, no. 1, pp. 44─48, 1982.

[9]D. Pascolini et al., "2002 global update of available data on visual impairment: a compilation of population-based prevalence studies," Ophthalmic Epidemiology, vol. 11, no. 2, pp. 67─115, 2004.

[10]N. Congdon et al., "Causes and prevalence of visual impairment among adults in the United States," Archives of Ophthalmology (Chicago, Ill.: 1960), vol. 122, no. 4, pp. 477─485, 2004.

[11]W. H. Organization, "Prevention of blindness and visual impairment," Priority Eye Diseases, 2017.

[12]李新華, 驗光學──眼鏡驗光與加工職業技能基礎教程. 新文京, 2013.

[13]"視網膜介紹." [Online]. Available:
https://smallcollation.blogspot.com/2013/06/retina.html?m=1&fbclid=IwAR2ArLWq4gHAytpJuQTvWgNf2FhZatmAvAogww4khKfPmZaC6c4cHGeafj0#gsc.tab=0

[14]"視網膜." [Online]. Available:
https://zh.wikipedia.org/wiki/%E8%A7%86%E7%BD%91%E8%86%9C

[15]A. Koh, "Optical Coherence Tomography of the Outer Retinal Layers," in Optical Coherence Tomography, vol. 4: Karger Publishers, 2014, pp. 26─33.

[16]B. K. B. Stanton, Berne & Levy Physiology, Updated Edition 6th Edition. Mosby, 2010.

[17]L. W. Tao, Z. Wu, R. H. Guymer, and C. D. Luu, "Ellipsoid zone on optical coherence tomography: a review," Clinical & Experimental Ophthalmology, vol. 44, no. 5, pp. 422─430, 2016.

[18]"視網膜病變." [Online]. Available: http://www.eyedoc.com.tw/retina_06.html

[19]"脈絡膜." [Online]. Available:
https://zh.wikipedia.org/wiki/%E8%84%89%E7%BB%9C%E8%86%9C

[20]眼科学 惠延年主编, 第5版 ed. (新世纪课程教材). 北京: 人民卫生出版社, 2001.

[21]A. Pokharel, G. S. Shrestha, and J. B. Shrestha, "Macular thickness and macular volume measurements using spectral domain optical coherence tomography in normal Nepalese eyes," Clinical Ophthalmology (Auckland, NZ), vol. 10, p. 511, 2016.

[22]P. J. Kelty, J. F. Payne, R. H. Trivedi, J. Kelty, E. M. Bowie, and B. M. Burger, "Macular thickness assessment in healthy eyes based on ethnicity using Stratus OCT optical coherence tomography," Investigative Ophthalmology & Visual Science, vol. 49, no. 6, pp. 2668─2672, 2008.

[23]A. H. Kashani et al., "Retinal thickness analysis by race, gender, and age using Stratus OCT," American Journal of Ophthalmology, vol. 149, no. 3, pp. 496-502. e1, 2010.

[24]W. K. Song, S. C. Lee, E. S. Lee, C. Y. Kim, and S. S. Kim, "Macular thickness variations with sex, age, and axial length in healthy subjects: a spectral domain–optical coherence tomography study," Investigative Ophthalmology & Visual Science, vol. 51, no. 8, pp. 3913─3918, 2010.

[25]S. Grover, R. K. Murthy, V. S. Brar, and K. V. Chalam, "Normative data for macular thickness by high-definition spectral-domain optical coherence tomography (spectralis)," American Journal of Ophthalmology, vol. 148, no. 2, pp. 266─271, 2009.

[26]A. C. Sull et al., "Comparison of spectral/Fourier domain optical coherence tomography instruments for assessment of normal macular thickness," Retina (Philadelphia, Pa.), vol. 30, no. 2, p. 235, 2010.

[27]M. Çubuk, B. Kasım, Y. Koçluk, and E. A. Sukgen, "Effects of age and gender on macular thickness in healthy subjects using spectral optical coherence tomography/scanning laser ophthalmoscopy," International Ophthalmology, vol. 38, no. 1, pp. 127─131, 2018.

[28]M. Adhi, S. Aziz, K. Muhammad, and M. I. Adhi, "Macular thickness by age and gender in healthy eyes using spectral domain optical coherence tomography," PLoS One, vol. 7, no. 5, 2012.

[29]R. S. Ramrattan, T. L. van der Schaft, C. M. Mooy, W. De Bruijn, P. Mulder, and P. De Jong, "Morphometric analysis of Bruch's membrane, the choriocapillaris, and the choroid in aging," Investigative Ophthalmology & Visual Science, vol. 35, no. 6, pp. 2857─2864, 1994.

[30]"從卷積層、激活層、池化層到全連接層深度解析卷積神經網絡的原理." [Online]. Available: https://kknews.cc/zh-tw/tech/vvx2qeq.html

[31]"機器/深度學習: 基礎介紹-損失函數." [Online]. Available: https://reurl.cc/Mv5pKk

[32]周飞燕, 金林鹏, and 董军, "卷积神经网络研究综述," 计算机学报, vol. 40, no. 6, pp. 1229─1251, 2017.

[33]李宏毅. "ML Lecture 7: Backpropagation." [Online]. Available:
https://www.youtube.com/watch?v=ibJpTrp5mcE&list=PLJV_el3uVTsPy9oCRY30oBPNLCo89yu49&index=12

[34]"The Stanford CS class: Convolutional Neural Networks for Visual Recognition." [Online]. Available: https://cs231n.github.io/neural-networks-3/

[35]G. E. Hinton, N. Srivastava, A. Krizhevsky, I. Sutskever, and R. R. Salakhutdinov, "Improving neural networks by preventing co-adaptation of feature detectors," arXiv preprint arXiv:1207.0580, 2012.

[36]"理解dropout." [Online]. Available: https://blog.csdn.net/stdcoutzyx/article/details/49022443

[37]"深度學習中Dropout原理解析." [Online]. Available: https://www.itread01.com/content/1547209261.html.

[38]"深度學習（二）." [Online]. Available: https://blog.csdn.net/mdzzzzzz/article/details/78028123.

[39]K. Simonyan and A. Zisserman, "Very deep convolutional networks for large-scale image recognition," arXiv preprint arXiv:1409.1556, 2014.

[40]"Brilliant Code.net." [Online]. Available: https://www.brilliantcode.net/1646/convolutional-neural-networks-3-calculate-number-of-parameters/

[41]K. He, X. Zhang, S. Ren, and J. Sun, "Deep residual learning for image recognition," in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016, pp. 770─778.

[42]C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna, "Rethinking the inception architecture for computer vision," in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016, pp. 2818─2826.

[43]Y. Ding et al., "A deep learning model to predict a diagnosis of Alzheimer disease by using 18F-FDG PET of the brain," Radiology, vol. 290, no. 2, pp. 456─464, 2019.

[44]R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra, "Grad-cam: Visual explanations from deep networks via gradient-based localization," in Proceedings of the IEEE International Conference on Computer Vision, 2017, pp. 618─626.

[45]B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Torralba, "Learning deep features for discriminative localization," in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016, pp. 2921─2929.