在本研究中，我們探討了如何利用深度學習技術與眼底影像資料，來輔助診斷與分類視網膜與黃斑部退化疾病，包括錐-桿狀細胞失養症 (cone-rod dystrophy, CRD) 與老年性黃斑部病變 (age-related macular degeneration, AMD)。我們將眼底影像透過前處理技術進行轉換與標準化，以便神經網路模型更有效地捕捉並提取影像中的病灶特徵。為了提升醫療診斷的可信度，我們更進一步導入了可解釋性人工智慧（explainable AI, XAI）技術，以確保模型決策的合理性能夠被視覺化且嚴格驗證。面對醫療資料常見的資料不平衡挑戰，我們運用了訓練資料過採樣與資料擴增策略來平衡樣本分佈，從而提高模型的準確性。接著，我們應用了多種經典的卷積神經網路架構（包含 ResNet、Inception、DenseNet 及 EfficientNet）來建立影像辨識模型。為了進一步釐清「黑箱」決策，我們結合了 Grad-CAM++、Integrated Gradients 等多種現今主流解釋性分析方法，並透過模型參數隨機化測試（sanity check）驗證其可靠性。這些方法的結合旨在充分發揮深度學習特徵萃取的優勢，同時確保預測結果具備高度的準確性與臨床解釋力。實驗結果顯示，經過資料增強、模型訓練架構設計與優化以及解釋性驗證等一系列處理後，我們所建立的多種模型在預測與分類雙病種視網膜疾病的準確性上均取得了優異的綜合效能。研究結果不僅驗證了深度學習在易混淆病症的診斷上的潛力，模型解釋性分析亦獲得了醫師的評估認可，為未來眼科醫療輔助診斷系統的實際落地提供了新的參考與方法。

This study utilizes deep learning and fundus images to assist in the diagnosis and classification of retinal and macular degenerative diseases, specifically cone-rod dystrophy (CRD) and age-related macular degeneration (AMD). Fundus images are standardized via preprocessing to enhance lesion feature extraction by neural networks. To improve clinical credibility, explainable AI (XAI) techniques are introduced to visualize and validate model decisions. To address class imbalance common in medical datasets, training data oversampling and data augmentation strategies are applied to balance sample distribution and improve accuracy. Multiple convolutional neural network architectures, including ResNet, Inception, DenseNet, and EfficientNet, were implemented for image classification. To demystify ”black-box” decisions, mainstream explainability methods (Grad-CAM++ and Integrated gradients) are integrated and verified using sanity check. This framework leverages deep learning for accurate feature extraction while ensuring clinical interpretability. Experimental results demonstrate that the optimized models achieve excellent comprehensive performance in classifying both retinal diseases. These findings validate the potential of deep learning for dual-disease diagnosis. Furthermore, the explainability analysis was evaluated and endorsed by professional ophthalmologists, offering valuable references and methodologies for future ophthalmic computer-aided diagnosis systems.

[1] 陳達慶. “視網膜失養症之基因流行病學及分子病理學暨其臨床應用”. 博士論文. 國立臺灣大學, 2021, pp. 1–191. DOI: 10.6342/NTU202100781.
[2] 陳達慶. “遺傳性視網膜失養症”. 台大醫院基因分子診斷實驗室, 2021. URL: https://www.ntuh.gov.tw/gene-lab-mollab/Fpage.action?muid=4052&fid=3870.
[3] June-Goo Lee et al. “Deep learning in medical imaging: general overview”. In: Korean journal of radiology 18.4 (2017), pp. 570–584.
[4] Kaiming He et al. “Deep residual learning for image recognition”. In: Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778.
[5] Christian Szegedy et al. “Inception-v4, inception-resnet and the impact of residual connections on learning”. In: Proceedings of the AAAI conference on artificial intelligence, Vol. 31. 1, 2017.
[6] Gao Huang et al. “Densely connected convolutional networks”. In: Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 4700–4708.
[7] Haihan Zhang et al. “Quickly diagnosing Bietti crystalline dystrophy with deep learning”. In: Iscience 27.9 (2024).
[8] Maria Skublewska-Paszkowska et al. “Application of convolutional gated recurrent units u-net for distinguishing between retinitis pigmentosa and cone–rod dystrophy”. In: acta mechanica et automatica 18.3 (2024).
[9] Mateusz Buda, Atsuto Maki, and Maciej A Mazurowski. “A systematic study of the class imbalance problem in convolutional neural networks”. In: Neural networks 106 (2018), pp. 249–259.
[10] Robert Geirhos et al. “Shortcut learning in deep neural networks”. In: Nature Machine Intelligence 2.11 (2020), pp. 665–673.
[11] Deshant Singh and Anurag Sharma. “A Comparative Analysis of Oversampling and Undersampling Techniques for Image Data Classification Across Varying Imbalance Levels”. In: International Conference on Data Science and Applications, Springer, 2024, pp. 371–385.
[12] Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. “Deep inside convolutional networks: Visualising image classification models and saliency maps”. In: arXiv preprint arXiv:1312.6034 (2013).
[13] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. “Axiomatic attribution for deep networks”. In: International conference on machine learning, PMLR, 2017, pp. 3319–3328.
[14] Aditya Chattopadhay et al. “Grad-cam++: Generalized gradient-based visual explanations for deep convolutional networks”. In: 2018 IEEE winter conference on applications of computer vision (WACV), IEEE, 2018, pp. 839–847.
[15] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. “Why should I trust you? Explaining the predictions of any classifier”. In: Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, 2016, pp. 1135–1144.
[16] Julius Adebayo et al. “Sanity checks for saliency maps”. In: Advances in neural information processing systems 31 (2018).
[17] Ilham Maulana, Siti Ernawati, and Muhammad Indra. “IMPROVING IMAGE CLASSIFICATION ACCURACY WITH OVERSAMPLING AND DATA AUGMENTATION USING DEEP LEARNING: A CASE STUDY ON THE SIMPSONS CHARACTERS DATASET”. In: Jurnal Riset Informatika 6.4 (2024), pp. 201–210.
[18] Satyapriya Krishna et al. “The disagreement problem in explainable machine learning: A practitioner’s perspective”. In: arXiv preprint arXiv:2202.01602 (2022).