臺灣博碩士論文加值系統

糖尿病視網膜病變是糖尿病患者常見之健康問題，而適當的監控能有助於減緩此病情之傷害。目前臨床上主要之診斷方式為透過眼底影像讓醫生進行判別，進而進行病徵的診斷。然而此做法往往使得處於醫療資源不足的患者無法達到早期診斷，進而造成病情惡化之問題發生。有鑑於此，近期有許多學者成功將深度學習技術應用於此領域並證明其有良好的辨識效益，然而此架構仍有改善之空間。舉例來說，其架構於開發過程中需耗費龐大的臨床資源來取得及標記模型訓練所需之影像，而這也增加了許多臨床人員的工作負擔。因此，本研究採用近期熱門之生成對抗網路技術來試圖減少模型訓練時所需之影像數量，期望能減少系統開發之經費與臨床人員在收集影像上之時間。實驗結果顯示，本論文所發展之方法能有效將訓練資料擴增為兩倍，並證明其生成之影像能幫助以卷積神經網路為基礎之診斷系統提升其準確率。更具體來說，本研究提出之方法相較於未採用生成對抗網路之方法能提升7%的診斷準確率(靈敏度與特異度)，而這也表示其提出之方法將具有潛力能來減緩臨床人員收集資料所需之時間，並能提升自動化診斷之準確率。

Diabetic retinopathy (DR) is a common health problem for those who suffer from diabetes, but the damage caused can be alleviated with appropriate monitoring. Current clinical diagnosis method is mainly through fundus images detected by ophthalmologists so as to diagnose the disease. However, this approach usually causes the patients who live without sufficient medical resources unable to obtain early diagnosis, which results in deteriorating in the disease. As a result, recently many researchers had successfully applied the deep learning technology and had shown to have good effect on DR recognition tasks; however, there is still rooms for improvement in this framework. For example, during the process of developing the framework, huge amount of clinical resources is needed for acquiring and labeling the images required for training the model, which increase much more burden for clinicians. Therefore, in this study, we applied the Generative Adversarial Network (GAN) technique in an attempt to reduce the amount of images required for training the model to see if similar DR detection accuracy can be achieved. Experimental results showed the amount of training data can be effectively augmented by twice, and the synthesized images can help improving the accuracy of the diagnosis system based on Convolutional Neural Network (CNN). Specially, compared with the results without GAN, the method proposed in this study can increase the diagnosis accuracy (sensitivity and specificity) by 7%, which shows the proposed method have the potential to reduce the time needed for clinicians to collect and preprocess the data, and to maintain the accuracy for AI-assisted automatic diagnosis.

目 錄

誌 謝.........................................i
摘 要........................................ii
Abstract.....................................iii
目 錄........................................iv
圖 目 錄.....................................v
表 目 錄....................................vi
第一章 緒論....................................1
1.1 研究背景與動機..............................1
1.2 研究目的...................................5
第二章 文獻探討.................................6
2.1 糖尿病視網膜病變............................7
2.2 深度類神經網路..............................9
2.3 卷積神經網路...............................19
2.4 生成對抗網路...............................30
第三章 研究材料與方法...........................33
3.1 研究資料來源................................34
3.2 本研究提出之DR診斷架構.......................35
3.3 以CycleGAN技術為基礎之影像生成架構............41
第四章 研究結果與討論............................44
4.1 實驗一: DR診斷模型基準線建立..................44
4.2 實驗二: DR診斷架構效能測試....................47
4.3 實驗三: 提出DR診斷架構結合生成影像之效益........53
第五章 結論......................................62
參考文獻.........................................65


圖 目 錄

圖1: IDx-DR (AI診斷DR系統)。......................4
圖2: DR發病流程圖。...............................7
圖3: DR病患之眼底鏡影像示意圖。....................8
圖4: (a)DNN架構圖；(b)感知器示意圖。..............11
圖5: (a)範例1座標；(b)範例2座標；(c)範例2維度轉換後座標。 ................................................13
圖6: 使用SGD做為優化方式完全取決於當下的梯度值以及學習率，因此容易產生困於鞍點的問題。..........................18
圖7: 應用於影像辨識之CNN架構。....................20
圖8: (a)卷積運算；(b)完成卷積運算；(c)影像經由卷積產生特徵圖。.............................................22
圖9: (a) ReLU函數；(b)經由ReLU運算後特徵圖中的物體形狀更明顯。.............................................23
圖 10: (a)最大池化運算；(b)完成最大池化完運算。.....24
圖11: 經由兩次最大池化做次取樣後提升了神經元的視野大小。 .................................................25
圖12: (a) CNN與全連接層架構；(b) CNN與全局平均池化架構。 .................................................26
圖13: GAN運作流程圖。.............................30
圖14: 本論文提出之系統架構圖。.....................33
圖15: 本研究之ResNet-50架構圖以及殘差學習塊示意圖。 .................................................38
圖16: CAM利用判別為狗之各權重值回推出CNN對原始影像的關注圖。 .................................................40
圖17: 將原始眼底鏡影像針對黃斑部位置做切割。.........41
圖18: CycleGAN用於馬與斑馬影像生成示意圖。..........43
圖 19: 以數量佔比呈現之Caltech-101資料集辨識結果之混淆矩陣。 .................................................46
圖20: 以299*299解析度之眼底鏡影像訓練及測試DR診斷模型之測試結果。..............................................47
圖21: 以960*960解析度之眼底鏡影像訓練及測試DR診斷模型之測試結果。..............................................50
圖22: (a)原始「RDR」眼底鏡影像；(b)以960*960解析度影像訓練之模型測試此影像之熱度圖；(c)以299*299解析度影像訓練之模型測試此影像之熱度圖。.....................................52
圖23: 實驗組與對照組之10次DR診斷模型測試準確率折線圖。 ..................................................57
圖24: 實驗組與對照組之10次DR診斷模型測試準確率之平均值帶標準誤差線長條圖。.......................................57


表 目 錄

表1: Gulshan等人針對AI應用於DR診斷之研究成果。........2
表2: Messidor-2之DR分期標準。......................34
表3: 模型基準線建立之比較參考對象研究結果準確率(%)。 ..................................................45
表4: 另一項以複製Google Research的DR診斷研究為目的之測試結果。..............................................49
表5: 利用CAM對DR診斷模型進行ROI可視化。.............51
表6: 利用CycleGAN將「no_RDR」轉為「RDR」影像........54
表7: 利用CycleGAN將「RDR」轉為「no_RDR」影像。......55
表8: 實驗組與對照組進行10次DR診斷模型訓練及測試之整體準確率結果。..............................................56
表9: 模型判斷為「no_RDR」之黃斑部影像及其熱度圖。....60
表10: 模型判斷為「RDR」之黃斑部影像及其熱度圖。......61

[1] Vision impairment and blindness, World Health Organization. (April 25, 2019). Retrieved from https://www.who.int/news-room/fact-sheets/detail/blindness-and-visual-impairment.
[2] Resnikoff S., et al. (2004) Global data on visual impairment in the year 2002. Bulletin of the world health organization, 82: 844-851.
[3] Yau J.W.Y., et al. (2012) Global prevalence and major risk factors of diabetic retinopathy. Diabetes care, 35.3: 556-564.
[4] Diabetes, World Health Organization. (April 26, 2019). Retrieved from https://www.who.int/news-room/fact-sheets/detail/diabetes.
[5] Facts about Diabetic Eye Disease, National Eye Institute. (April 27, 2019). Retrieved from https://nei.nih.gov/health/diabetic/retinopathy.
[6] Gulshan V., et al. (2016) Development and Validation of a Deep Learning Algorithm for Detection of Diabetic Retinopathy in Retinal Fundus Photographs. Jama, 316.22: 2402-2410.
[7] Diabetic Retinopathy Detection with Data from EyePACS, Kaggle. (April 29, 2019). Retrieved from https://kaggle.com/c/diabetic-retinopathy-detection.
[8] Messidor, ADCIS. (April 29, 2019). Retrieved from http://www.adcis.net/en/third-party/messidor/.
[9] Abràmoff M.D., et al. (2016) Improved Automated Detection of Diabetic Retinopathy on a Publicly Available Dataset Through Integration of Deep Learning. Investigative ophthalmology & visual science, 57.13: 5200-5206.
[10] IDx-DR EU. (April 28, 2019). Retrieved from https://www.eyediagnosis.co/idx-dr-eu-1.
[11] FDA permits marketing of artificial intelligence-based device to detect certain diabetes-related eye problems, U.S. FOOD & DRUG Administration. (Jun 2, 2019). Retrieved from http://www.fda.gov/news-events/press-announcements/fda-permits-marketing-artificial-intelligence-based-device-detect-certain-diabetes-related-eye.
[12] LeCun Y., Bengio Y., and Hinton G.E. (2015) Deep learning. Nature, 521.7553: 436.
[13] Esteva A., et al. (2017) Dermatologist-level classification of skin cancer with deep neural networks. Nature, 542.7639: 115.
[14] Bejnordi B.E., et al. (2017) Diagnostic Assessment of Deep Learning Algorithms for Detection of Lymph Node Metastases in Women With Breast Cancer. JAMA, 318.22: 2199-2210.
[15] Sirinukunwattana K., et al. (2016) Locality Sensitive Deep Learning for Detection and Classification of Nuclei in Routine Colon Cancer Histology Images. IEEE Trans. Med. Imaging, 35.5: 1196-1206.
[16] LeCun Y., and Yoshua B. (1995) Convolutional networks for images, speech, and time series. The handbook of brain theory and neural networks, 3361.10: 1995.
[17] Goodfellow I., et al. (2014) Generative adversarial nets. Advances in neural information processing systems, pp 2672-2680.
[18] Diabetic Retinopathy, Diabetes.co.uk the global diabetes community. (Jun 3, 2019). Retrieved from https://www.diabetes.co.uk/diabetes-complications/diabetic-retinopathy.html.
[19] Tarr J.M., Kaul K., Chopra M., Kohner E.M., and Chibber R. (2013) Pathophysiology of Diabetic Retinopathy. ISRN ophthalmology, pp 1-13.
[20] Diabetic Retinopathy Treatment, Eagle Eye Centre. (April 27, 2019). Retrieved from http://www.eagleeyecentre.com.sg/service/diabetic-retinopathy/.
[21] Ivakhnenko A.G. (1973) Cybernetic Predicting Devices. CCM Information Corporation.
[22] Robert H.N. (1992) Theory of the backpropagation neural network. Neural networks for perception, pp 65-93. Academic Press.
[23] Hochreiter S. (1998) The vanishing gradient problem during learning recurrent neural nets and problem solutions. International Journal of Uncertainty. Fuzziness and Knowledge-Based Systems, 6.02: 107-116.
[24] Verma J.P. (2014) Statistics for Exercise Science and Health with Microsoft Office Excel. John Wiley & Sons.
[25] Rubinstein R.Y. and Kroese D.P. (2013) The Cross-Entropy Method: A Unified Approach to Combinatorial Optimization, Monte-Carlo Simulation and Machine Learning. Springer Science & Business Media.
[26] Ruder S. (2016) An overview of gradient descent optimization algorithms. arXiv preprint arXiv: 1609.04747.
[27] Li M., et al. (2014) Efficient mini-batch training for stochastic optimization. Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining, pp.661-670.
[28] Hubel D. H., and Wiesel T. N. (1962) Receptive fields, binocular interaction and functional architecture in the cat's visual cortex. The Journal of physiology, 160.1: 106-154.
[29] Krizhevsky A., Sutskever I., and Hinton G.E. (2012) Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, pp 1097-1105.
[30] Lin M., Chen Q., and Yan S. (2013) Network in network. arXiv preprint arXiv:1312.4400.
[31] Glorot X., Bordes A., and Bengio Y. (2011) Deep sparse rectifier neural networks. Proceedings of the fourteenth international conference on artificial intelligence and statistics, pp 315-323.
[32] Le H., and Borji A. (2017) What are the receptive, effective receptive, and projective fields of neurons in convolutional neural networks? arXiv preprint arXiv: 1705.07049.
[33] Hawkins D.M. (2004) The problem of overfitting. Journal of chemical information and computer sciences, 44.1: 1-12.
[34] Zhou B., et al. (2016) Learning deep features for discriminative localization. Proceedings of the IEEE conference on computer vision and pattern recognition, pp 2921-2929.
[35] Bishop C.M. (2006) Pattern Recognition and Machine Learning. Springer.
[36] Harris D. and Harris S. (2010) Digital Design and Computer Architecture. Morgan Kaufmann.
[37] Erhan D., et al. (2010). Why does unsupervised pre-training help deep learning? Journal of Machine Learning Research, 11.Feb: 625-660.
[38] He K., Zhang X., Ren S., and Sun J. (2016) Deep residual learning for image recognition. Proceedings of the IEEE conference on computer vision and pattern recognition, pp 770-778.
[39] Zhu J.Y., Park T., Isola P., and Efros A.A. (2017) Unpaired image-to-image translation using cycle-consistent adversarial networks. Proceedings of the IEEE international conference on computer vision, pp 2223-2232.
[40] Voets M., Møllersen K., and Bongo L.A. (2018) Replication study: Development and validation of deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs. arXiv preprint arXiv: 1803.04337.
[41] Szegedy C., et al. (2016) Rethinking the inception architecture for computer vision. Proceedings of the IEEE conference on computer vision and pattern recognition, pp 2818-2826.
[42] Hinton G.E., Simon O., and The Y.W. (2006) A fast learning algorithm for deep belief nets. Neural computation, 18.7: 1527-1554.
[43] Oquab M., et al. (2014) Learning and transferring mid-level image representations using convolutional neural networks. Proceedings of the IEEE conference on computer vision and pattern recognition, pp 1717-1724.
[44] Caltech101. (April 30, 2019). Retrieved from http://www.vision.caltech.edu/Image_Datasets/Caltech101/.
[45] Li J., et al. (2017) Visual representation and classification by learning group sparse deep stacking network. IEEE Transactions on Image Processing, 27.1: 464-476.
[46] Fay M.P. and Proschan M.A. (2010) Wilcoxon-Mann-Whitney or t-test? On assumptions for hypothesis tests and multiple interpretations of decision rules. Statistics surveys, 4: 1.