近年來隨著深度學習結合電腦視覺技術在諸多應用上的成功，越來越多的研究與應用開始往醫療領域發展，專注於臨床上可使用的電腦輔助診斷系統開發。過去的研究顯示深度學習模型在眼底鏡影像分析中對青光眼的判讀上有著驚人的準確率，並有助於青光眼的早期篩檢。然而，訓練所得到的青光眼檢測模型對於不同來源的眼底鏡影像資料卻有泛化能力不佳的問題存在。也就是說，當一個檢測模型對不同資料集進行測試時，準確率往往有大幅度的下滑。為了讓青光眼檢測模型在面對不同資料集時都能維持高準確率與穩定度，提升青光眼檢測模型的泛化能力是最需要解決的問題。本論文透過探討穩健特徵與非穩健特徵對泛化能力的影響，進而解決深度學習模型泛化能力不佳的問題。實驗中先使用ResNet50模型架構對內湖三軍總醫院提供的TSGH Dataset進行模型訓練，並利用隨機森林模型取代ResNet50中的全連接層以拆解深度學習特徵與分類器。接著使用來自臺北榮民總醫院的TVGH Dataset中的驗證集對模型所萃取的所有特徵進行重要度排序，再籍著特徵篩選的方式找出泛化性較高的可遷移特徵，以降低對訓練集資料來源過擬合特徵的影響來達到提升模型的泛化能力的目的。實驗結果顯示，在特徵重要度篩選後，卷積網路模型對於跨院眼底鏡影像上的判讀準確率從 61.2% 提升至82.33%。此外，若再結合影像前處理的標準化方法以降低不同跨院影像間的差距，則模型的對於跨院資料集的預測準確率能提升至86.33%。藉由本論文所提出的特徵篩選以過濾過擬合特徵並結合影像前處理的標準化程序，將能有效改善深度學習模型在眼底鏡影像分析上其泛化能力不佳的問題。

With the advancement of deep learning in many computer vision applications, the development of computer-aided diagnosis systems has been focused on recently. The deep learning approaches for glaucoma assessment through fundus analysis have shown extremely high performance, especially for early detection, in many studies. However, the deep learning models also suffered from the problem of poor generalization for the cross-institutional datasets. That is, the model’s performance dropped dramatically on the input images from different institutions. So, it is the top issue to be tackled to improve the generalization ability of models. In this thesis, the influences of robust features and non-robust features on model generalization ability were mainly discussed. In the experiments, the ResNet50 models were trained on the fundus datasets of TSGH, and the fully connected layers in the ResNet50 were replaced by the classifier of Random Forest to perform feature extraction and classification separately. Then, the validation set from TVGH was used for the feature selection based on the extracted features’importance ranking. As a result, the robust features with better generalization were selected to reduce the influences from the non-robust features that overfit the training dataset from a specific source. With feature selection, the performance of the glaucoma assessment cross-institutional dataset can be improved from 61.2% to 82.33% in accuracy. In addition, the normalization process as image preprocessing was further applied to reduce the gaps among different institutional data, and the performance can be further boosted to 86.33%. In conclusion, the generalization ability of the deep learning model can be enhanced successfully with the proposed process of feature selection and image normalization.

參考文獻

[1] "Glaucoma " National Health Service. https://www.nhs.uk/conditions/glaucoma/#:~:text=Glaucoma%20is%20a%20common%20eye,not%20diagnosed%20and%20treated%20early. (accessed.
[2] "Types of Glaucoma." National Eye Institute. https://www.nei.nih.gov/learn-about-eye-health/eye-conditions-and-diseases/glaucoma/types-glaucoma (accessed.
[3] I. J. Goodfellow, J. Shlens, and C. Szegedy, "Explaining and harnessing adversarial examples," arXiv preprint arXiv:1412.6572, 2014.
[4] V. Nair and G. E. Hinton, "Rectified linear units improve restricted boltzmann machines," in Icml, 2010.
[5] 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.
[6] M. D. Zeiler, D. Krishnan, G. W. Taylor, and R. Fergus, "Deconvolutional networks," in 2010 IEEE Computer Society Conference on computer vision and pattern recognition, 2010: IEEE, pp. 2528-2535.
[7] K. He, G. Gkioxari, P. Dollár, and R. Girshick, "Mask r-cnn," in Proceedings of the IEEE international conference on computer vision, 2017, pp. 2961-2969.
[8] R. Girshick, "Fast r-cnn," in Proceedings of the IEEE international conference on computer vision, 2015, pp. 1440-1448.
[9] 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.
[10] V. S. Mary, E. B. Rajsingh, and G. R. Naik, "Retinal fundus image analysis for diagnosis of glaucoma: a comprehensive survey," IEEE Access, 2016.
[11] S. Akbar, S. A. Hassan, A. Shoukat, J. Alyami, and S. A. Bahaj, "Detection of microscopic glaucoma through fundus images using deep transfer learning approach," Microscopy Research and Technique, vol. 85, no. 6, pp. 2259-2276, 2022.
[12] M. Balasubramanian et al., "A framework for detecting glaucomatous progression in the optic nerve head of an eye using proper orthogonal decomposition," IEEE Transactions on Information Technology in biomedicine, vol. 13, no. 5, pp. 781-793, 2009.
[13] A. J. Patterson, D. F. Garway-Heath, N. G. Strouthidis, and D. P. Crabb, "A new statistical approach for quantifying change in series of retinal and optic nerve head topography images," Investigative ophthalmology & visual science, vol. 46, no. 5, pp. 1659-1667, 2005.
[14] D. C. DeBuc, "A review of algorithms for segmentation of retinal image data using optical coherence tomography," Image Segmentation, vol. 1, pp. 15-54, 2011.
[15] A. Herzog, K. L. Boyer, and C. Roberts, "Extracting the optic disk endpoints in optical coherence tomography data," in 2005 Seventh IEEE Workshops on Applications of Computer Vision (WACV/MOTION'05)-Volume 1, 2005, vol. 1: IEEE, pp. 263-268.
[16] M. S. Miri et al., "Multimodal segmentation of optic disc and cup from SD-OCT and color fundus photographs using a machine-learning graph-based approach," IEEE transactions on medical imaging, vol. 34, no. 9, pp. 1854-1866, 2015.
[17] P. M. Burlina, N. Joshi, M. Pekala, K. D. Pacheco, D. E. Freund, and N. M. Bressler, "Automated grading of age-related macular degeneration from color fundus images using deep convolutional neural networks," JAMA ophthalmology, vol. 135, no. 11, pp. 1170-1176, 2017.
[18] K. Shankar, Y. Zhang, Y. Liu, L. Wu, and C.-H. Chen, "Hyperparameter tuning deep learning for diabetic retinopathy fundus image classification," IEEE Access, vol. 8, pp. 118164-118173, 2020.
[19] Y. Peng et al., "DeepSeeNet: a deep learning model for automated classification of patient-based age-related macular degeneration severity from color fundus photographs," Ophthalmology, vol. 126, no. 4, pp. 565-575, 2019.
[20] Z. N. Riya and M. F. Hossain, "Active Contouring with Morphology Assisted Automated Glaucoma Detection," in 2021 International Conference on Computer, Communication, Chemical, Materials and Electronic Engineering (IC4ME2), 2021: IEEE, pp. 1-4.
[21] P. Elangovan and M. K. Nath, "En‐ConvNet: A novel approach for glaucoma detection from color fundus images using ensemble of deep convolutional neural networks," International Journal of Imaging Systems and Technology.
[22] J. M. Ahn, S. Kim, K.-S. Ahn, S.-H. Cho, K. B. Lee, and U. S. Kim, "A deep learning model for the detection of both advanced and early glaucoma using fundus photography," PloS one, vol. 13, no. 11, p. e0207982, 2018.
[23] L. Li, M. Xu, X. Wang, L. Jiang, and H. Liu, "Attention based glaucoma detection: a large-scale database and CNN model," in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 10571-10580.
[24] 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.
[25] Y.-C. Ko et al., "Deep learning assisted detection of glaucomatous optic neuropathy and potential designs for a generalizable model," PLoS One, vol. 15, no. 5, p. e0233079, 2020.
[26] A. Ilyas, S. Santurkar, D. Tsipras, L. Engstrom, B. Tran, and A. Madry, "Adversarial examples are not bugs, they are features," Advances in neural information processing systems, vol. 32, 2019.