青光眼 (Glaucoma) 被公認為是最棘手的眼科疾病，原因是 (1) 目前醫學上僅能阻止或延緩青光眼的惡化，卻無法根治青光眼;
(2) 早期的青光眼難以被偵測到。
根據台灣高雄長庚醫院的資料顯示，青光眼早期偵測之準確度皆在92% 以下。
基於青光眼早期偵測之困難，方有與高雄長庚醫院產學合作之契機。
高雄長庚醫院提供資料有 163人, 共1020 張眼底圖片 (其中青光眼有115人, 共899張青光眼眼底圖片; 正常眼有48人, 共121張青光眼眼底圖片)。

本研究整合大數據分析方法, 包括
(a) 顏色增強算法 (所謂的 Retinex), 將照相機照到的眼底圖片去除光線影響, 還原成眼底圖片之原色，(b) 萃取重要區域 (包含視神經杯, 視神經盤, 黃斑部, 與視神經纖維層之最小區域), 與(c) 卷積類神經網絡 (Convolutional Neural Network, CNN)。
本研究再利用實驗設計得到最佳的 CNN 的參數(如: 輸入CNN 之眼底相片尺寸, 卷積層數, 與隱藏層數等)。
本研究建議一種有效可以輔助醫生預測青光眼的模型, 其敏感度 (Sensitivity) 為97%、特異度 (Specificity)為96% 與準確度 (Accuracy)為97%。
本研究提出的策略優於現有文獻的研究結果。

Glaucoma is an ophthalmic disease that is notoriously known to be incurable,
with medical procedures that can only delay the aggravation of glaucoma but not
restore ophthalmic health.
According to data from Kaohsiung Chang Gung
Hospital in Taiwan, the accuracy of early detection of glaucoma is below 92%.
Motivated by the problem of low accuracy of early Glaucoma detection, we
cooperated with Kaohsiung Chang Gung Hospital which provided a total of 1020 fundus images (163 patients), 899 glaucoma fundus images and 121 healthy images (115 glaucoma patients and 48 Healthy patients).

This study integrates big data analysis methods, including (a) a color enhancement algorithm (so-called Retinex)
which removes the effects of fundus photography flash and restores the original colors of the fundus image,
(b) extraction of important areas (including optic nerve cup, optic disc, macula, and minimal area of optic nerve fiber layer),
and (c) Convolutional Neural Networks (CNN).
This study uses experimental design to obtain the optimal parameters combinations for CNN
(such as input image size, convolution layer, and hidden layer).
We propose an effective classification model based on the full fundus image with sensitivity 97%, specificity 96%, and accuracy 97%. The proposed strategy outperforms existing methodologies.

目錄
中文摘要 1
Abstract 2
目錄 4
表目錄 6
圖目錄 6
1 緒論 10
1.1 研究背景 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.2 研究動機 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.3 青光眼的檢測儀器 ( 方法 ) . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.4 重要區域 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.5 績效指標 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2 文獻回顧 14
2.1 青光眼預測方法比較 . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3 研究方法與步驟 16
3.1 資料前處理 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.1.1 圖片清理 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.1.2 Retinex 演算法 . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.1.3 重要區域萃取 . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.1.4 圖片擴增 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.1.5 尺寸縮減 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.2 卷積類神經網路 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.2.1 卷積 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.2.2 整流 (ReLU) . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.2.3 池化 (Pooling) . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.2.4 陣列化資料 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.2.5 神經元 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4
3.2.6 輸入層 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.2.7 隱藏層 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.2.8 輸出層 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.2.9 初始參數 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.2.10 正向傳播 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.2.11 損失函數 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.2.12 反向傳播 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.2.13 Adam 演算法 . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.2.14 Keras 邏輯 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.2.15 資料集介紹 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.3 類神經網路之參數 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.3.1 參數選擇 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.3.2 實驗設計 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
4 結果與結論 36
4.1 資料分群 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.2 策略類型 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.3 結果比較 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4.4 結論 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4.5 未來研究 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
附錄 40
參考文獻 41

參考文獻
[1] Rüdiger Bock et al. “Classifying glaucoma with image-based features from
fundus photographs”. In: Joint Pattern Recognition Symposium. Springer.
2007, pp. 355–364.
[2] Xiangyu Chen et al. “Glaucoma detection based on deep convolutional neural
network”. In: Engineering in Medicine and Biology Society (EMBC), 2015
37th Annual International Conference of the IEEE. IEEE. 2015, pp. 715–
718.
[3] Google. The code of Keras. 2019. url: https://reurl.cc/Wm3dk.
[4] Michael J Greaney et al. “Comparison of optic nerve imaging methods to
distinguish normal eyes from those with glaucoma”. In: Investigative oph-
thalmology & visual science 43.1 (2002), pp. 140–145.
[5] Sek Tien Hoh et al. “Optical coherence tomography and scanning laser polarimetry
in normal, ocular hypertensive, and glaucomatous eyes”. In: Amer-
ican Journal of Ophthalmology 129.2 (2000), pp. 129–135.
[6] Mei-Ling Huang, Hsin-Yi Chen, and Jian-Jun Huang. “Glaucoma detection
using adaptive neuro-fuzzy inference system”. In: Expert Systems with Ap-
plications 32.2 (2007), pp. 458–468.
[7] Fauzia Khan et al. “Detection of glaucoma using retinal fundus images”. In:
Biomedical Engineering International Conference (BMEiCON), 2013 6th.
IEEE. 2013, pp. 1–5.
[8] Ana Maria Mendonca and Aurelio Campilho. “Segmentation of retinal blood
vessels by combining the detection of centerlines and morphological reconstruction”.
In: IEEE transactions on medical imaging 25.9 (2006), pp. 1200–
1213.
9] This paper. The optimal model. 2019. url: https://reurl.cc/VbypZ.
[10] S. Parthasarathy and P. Sankaran. “An automated multi Scale Retinex with
Color Restoration for image enhancement”. In: 2012 National Conference
on Communications (NCC). Feb. 2012, pp. 1–5. doi: 10.1109/NCC.2012.
6176791.
[11] Jyotika Pruthi and Saurabh Mukherjee. “Computer based early diagnosis of
glaucoma in biomedical data using image processing and automated early
nerve fiber layer defects detection using feature extraction in retinal colored
stereo fundus images”. In: International Journal of Scientific & Engineering
Research 4.4 (2013), pp. 1822–28.
[12] R Geetha Ramani, S Sugirtharani, and B Lakshmi. “Automatic Detection
of Glaucoma in Retinal Fundus Images through Image Processing and Data
Mining Techniques”. In: International Journal of Computer Applications
166.8 (2017).
[13] PF Sharp et al. “Laser imaging of the retina”. In: British journal of ophthal-
mology 83.11 (1999), pp. 1241–1245.
[14] Wheyming T Song and Bruce W Schmeiser. “Displaying statistical point
estimates using the leading-digit rule”. In: IIE Transactions 43.12 (2011),
pp. 851–862.
[15] Yih-Chung Tham et al. “Global prevalence of glaucoma and projections of
glaucoma burden through 2040: a systematic review and meta-analysis”. In:
Ophthalmology 121.11 (2014), pp. 2081–2090.
[16] Linda M Zangwill et al. “Discriminating between normal and glaucomatous
eyes using the Heidelberg retina tomograph, GDx nerve fiber analyzer, and
optical coherence tomograph”. In: Archives of ophthalmology 119.7 (2001),
pp. 985–993.