近年來隨著AI人工智慧蓬勃發展，導入深度學習架構用於輔助臨床診斷的工具也越來越受到青睞。在眼科方面，AI輔助青光眼的診斷上多以眼底影像作為對象，藉由深度學習從大量的眼底影像中找出有無罹患青光眼的特徵，作為模型判讀上的依據，而在僅透過單模態眼底影像來判定青光眼的情況上，AI診斷的準確度能達到90%以上的驚人效果。因此，本論文認為深度學習模型在眼底影像的判讀上有超越肉眼感官判斷極限的可能，而雖然很多相關研究都帶來很高的準確度，但在醫療診斷上缺乏可解釋性，也難以驗證臨床應用上的有效性。利用特徵可視化的方法來分析模型在判斷罹患青光眼與否所關注的特徵，並與臨床診斷上的領域知識相對應比較，提出可解釋AI，將能提高深度學習架構在臨床應用上的接受度與可靠性。
本論文主要透過深度學習模型對NTUH Dataset眼底影像資料集進行學習，藉由導入黃斑部神經節細胞複合體(GCC)厚薄度資訊、不同眼底鏡角度、資料集切分及影像裁切方式，來分析不同方式所提供給網路模型的特徵，對於判定效能上的影響外，更採用兩種不同面向的可視化方法進行分析，來評估與解釋網路模型所關注的區域及是否符合臨床診斷知識。
在實驗結果方面，導入GCC厚薄度資訊進一步對資料集進行切分及訓練，對於在青光眼判定的準確度有所提高。觀察模型可視化的結果，在罹患青光眼的案例上模型主要聚焦在視神經盤區域，這與眼科醫師在臨床上判讀青光眼眼底影像所關注的區域一致；而特別的是，在非青光眼的案例上，模型大多聚焦在黃斑部區域，進一步透過以視神經盤為中心和以黃斑部為中心兩種眼底影像配合不同影像裁切大小來訓練及分析，令人驚訝的是僅使用裁切黃斑部區域的影像，在判定青光眼時仍然可以實現很高的預測準確度，而從幾個模型關注區域對應GCC受損區域的案例中，結果可證明模型在僅黃斑部區域就能檢測到超過肉眼判斷極限且足夠代表青光眼的非強健性特徵。除此之外，導入GCC厚薄度資訊進一步對資料集進行切分及訓練，對於在青光眼判定的準確度也有所提高。

With the rapid development of deep learning in computer vision applications, AI-assisted diagnostic systems are increasingly favored to assist physicians with clinical diagnosis. Surprisingly, the AI-based systems for glaucoma inspection can achieve up to 90% in accuracy simply based on the single modality of color fundus phorographs. Thus, the deep features extracted by the deep learning models in glaucoma detection are visualized in this thesis, compared with the clinical knowledge to provide model interpretability. It will improve the acceptance and reliability of deep learning frameworks for clinical applications.
For the experiments, the deep learning models of ResNet50 were trained on the dataset of fundus images from National Taiwan University Hospital Hsin-Chu Branch, and the class-specific discriminative areas with various ganglion cell complex(GCC) thickness conditions, center focus areas, cropped patches from fundus, and dataset partitionss are discussed. In addition, two visualization methods were used to evaluate and explain the areas of interest of the network model and whether it conforms to the clinical diagnostic knowledge.
Experimental results showed that the accuracy of glaucoma determination was improved by incorporating GCC thickness information. Deep learning models mainly focus on the areas of the optic nerve head (ONH) for the diagnosis of glaucoma, which is accordant to the clinical rules in glaucoma assessment. Surprisingly, the deep learning models can still achieve high prediction accuracy in detecting glaucomatous cases with cropped images of macular areas only. Several cases showed the areas that the model focuses on the region with GCC impairment. The results implied that the deep learning models can detect the morphologically detailed alterations in fundus photographs, which may be beyond the visual diagnosis of experts.

[1] S. Kingman, "Glaucoma is second leading cause of blindness globally," Bulletin of the World Health Organization, vol. 82, no. 11, pp. 887-888, 2004-11 2004. [Online]. Available: https://apps.who.int/iris/handle/10665/269284.
[2] H. A. Quigley, "Glaucoma," The Lancet, vol. 377, no. 9774, pp. 1367-1377, 2011/04/16/ 2011, doi: https://doi.org/10.1016/S0140-6736(10)61423-7.
[3] Y. C. Tham, X. Li, T. Y. Wong, H. A. Quigley, T. Aung, and C. Y. Cheng, "Global prevalence of glaucoma and projections of glaucoma burden through 2040: a systematic review and meta-analysis," Ophthalmology, vol. 121, no. 11, pp. 2081-90, Nov 2014, doi: 10.1016/j.ophtha.2014.05.013.
[4] "Prevalence of Open-Angle Glaucoma Among Adults in the United States," Archives of Ophthalmology, vol. 122, no. 4, p. 532, 2004, doi: 10.1001/archopht.122.4.532.
[5] D. A. Lee and E. J. Higginbotham, "Glaucoma and its treatment: A review," American Journal of Health-System Pharmacy, vol. 62, no. 7, pp. 691-699, 2005, doi: 10.1093/ajhp/62.7.691.
[6] Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner, "Gradient-based learning applied to document recognition," Proceedings of the IEEE, vol. 86, no. 11, pp. 2278-2324, 1998, doi: 10.1109/5.726791.
[7] M.-C. Popescu, V. E. Balas, L. Perescu-Popescu, and N. Mastorakis, "Multilayer perceptron and neural networks," WSEAS Trans. Cir. and Sys., vol. 8, no. 7, pp. 579–588, 2009.
[8] J. Deng, W. Dong, R. Socher, L. J. Li, L. Kai, and F.-F. Li, "ImageNet: A large-scale hierarchical image database," in 2009 IEEE Conference on Computer Vision and Pattern Recognition, 20-25 June 2009 2009, pp. 248-255, doi: 10.1109/CVPR.2009.5206848.
[9] 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.
[10] W. Samek, T. Wiegand, and K.-R. Müller, "Explainable artificial intelligence: Understanding, visualizing and interpreting deep learning models," arXiv preprint arXiv:1708.08296, 2017.
[11] C. Molnar, Interpretable machine learning. Lulu. com, 2020.
[12] M. T. Ribeiro, S. Singh, and C. Guestrin, ""Why Should I Trust You?": Explaining the Predictions of Any Classifier," presented at the Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, San Francisco, California, USA, 2016. [Online]. Available: https://doi.org/10.1145/2939672.2939778.
[13] B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Torralba, "Learning Deep Features for Discriminative Localization," in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 27-30 June 2016 2016, pp. 2921-2929, doi: 10.1109/CVPR.2016.319.
[14] M. Lin, Q. Chen, and S. Yan, "Network in network," arXiv preprint arXiv:1312.4400, 2013.
[15] C. Szegedy et al., "Going deeper with convolutions," in 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 7-12 June 2015 2015, pp. 1-9, doi: 10.1109/CVPR.2015.7298594.
[16] A. Diaz-Pinto, S. Morales, V. Naranjo, T. Köhler, J. M. Mossi, and A. Navea, "CNNs for automatic glaucoma assessment using fundus images: an extensive validation," BioMedical Engineering OnLine, vol. 18, no. 1, p. 29, 2019/03/20 2019, doi: 10.1186/s12938-019-0649-y.
[17] K. Simonyan and A. Zisserman, "Very deep convolutional networks for large-scale image recognition," arXiv preprint arXiv:1409.1556, 2014.
[18] F. Chollet, "Xception: Deep Learning with Depthwise Separable Convolutions," in 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 21-26 July 2017 2017, pp. 1800-1807, doi: 10.1109/CVPR.2017.195.
[19] S. Phene et al., "Deep Learning and Glaucoma Specialists: The Relative Importance of Optic Disc Features to Predict Glaucoma Referral in Fundus Photographs," Ophthalmology, vol. 126, no. 12, pp. 1627-1639, 2019/12/01/ 2019, doi: https://doi.org/10.1016/j.ophtha.2019.07.024.
[20] M. A. Zapata et al., "Artificial intelligence to identify retinal fundus images, quality validation, laterality evaluation, macular degeneration, and suspected glaucoma," Clinical Ophthalmology (Auckland, NZ), vol. 14, p. 419, 2020.
[21] S. Phan, S. i. Satoh, Y. Yoda, K. Kashiwagi, and T. Oshika, "Evaluation of deep convolutional neural networks for glaucoma detection," Japanese journal of ophthalmology, vol. 63, no. 3, pp. 276-283, 2019.
[22] G. Huang, Z. Liu, L. V. D. Maaten, and K. Q. Weinberger, "Densely Connected Convolutional Networks," in 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 21-26 July 2017 2017, pp. 2261-2269, doi: 10.1109/CVPR.2017.243.
[23] H. Liu et al., "Development and validation of a deep learning system to detect glaucomatous optic neuropathy using fundus photographs," JAMA ophthalmology, vol. 137, no. 12, pp. 1353-1360, 2019.
[24] R. Hemelings et al., "Accurate prediction of glaucoma from colour fundus images with a convolutional neural network that relies on active and transfer learning," Acta ophthalmologica, vol. 98, no. 1, pp. e94-e100, 2020.
[25] F. Li et al., "Deep learning-based automated detection of glaucomatous optic neuropathy on color fundus photographs," Graefe's Archive for Clinical and Experimental Ophthalmology, vol. 258, no. 4, pp. 851-867, 2020/04/01 2020, doi: 10.1007/s00417-020-04609-8.
[26] R. Hemelings, B. Elen, J. Barbosa-Breda, M. B. Blaschko, P. De Boever, and I. Stalmans, "Deep learning on fundus images detects glaucoma beyond the optic disc," Scientific Reports, vol. 11, no. 1, p. 20313, 2021/10/13 2021, doi: 10.1038/s41598-021-99605-1.