青光眼是一类以进行性视神经损害为特征的致盲性眼科疾病，患者在初期并没有明显的视力障碍症状，直至出现可察觉的视功能缺损时才会寻求医疗服务，所以实现青光眼的早期筛查对于保护患者的视力至关重要。基于深度学习的青光眼诊断虽然已取得巨大进展，但因其预测手段无法提供诊断依据和病因推理，不能被医生和患者完全信任和接受。因此，面向青光眼诊断的深度学习可解释性研究能够为医学知识和疾病辅助诊断的深度融合提供有效且可交互的途径，有力推动医疗的智能化。

（1）医学研究表明青光眼的结构性损伤和功能性损伤之间存在空间对应关系，所以在眼底图像中定位和量化区域变得尤为重要，本文第一个工作提出了一种基于卷积神经网络的青光眼诊断可视化方法。首先，在多尺度特征金字塔结构的基础上引入动态感受野模块，通过内嵌坐标注意力机制解决多尺度下特征信息无法被注意力准确聚焦的问题；其次，使用融合后的梯度加权类激活热力图来提供较详细的注意力图，解除了可视化卷积层特征图方法中存在的高层连接性约束。实验结果表明，本文提出的方法不仅具有更高的分类准确率，还提供了病灶区域可视化证据，更容易被医生和患者认可。

（2）仅基于眼底图像评估青光眼的信息较为有限，本文的第二个工作借鉴临床诊断经验，提出了一种基于知识图谱的青光眼病因推理方法。该方法分为问题关系分析模块、病因推理模块、答案预测模块。为了从知识图谱中抽取出相关信息，问题关系分析模块利用图神经网络和条件随机场等方法从问题中提取待求关系及目标实体；抽取出信息之后，病因推理模块估计出每条信息的条件期望和倾向分数；答案预测模块则利用每条信息的倾向分数和条件期望计算问题中待求关系的因果效应并完成预测分析；最后，本文借助Neo4j工具实现了数据可视化，并搭建了基于知识图谱的青光眼推理问答系统。实验结果表明，本文提出的方法为节点中不存在答案的问题提供了一条新的解决途径，设计的因果推理外部知识载体在公开数据集上的答案准确性有着一定的先进性。

Glaucoma is a kind of blinding eye disease characterized by progressive optic nerve damage. Patients do not have obvious visual impairment symptoms in the early stage, and will not seek medical services until they have noticeable visual impairment. Although glaucoma diagnosis based on deep learning has made great progress, it cannot be fully trusted and accepted by doctors and patients because its prediction methods cannot provide diagnostic basis and etiological reasoning. Therefore, the interpretability research of deep learning for glaucoma diagnosis can provide an effective and interactive way for the deep integration of medical knowledge and disease auxiliary diagnosis, and effectively promote the intelligence of medical treatment.

(1) Medical studies have shown that there is a spatial correspondence between structural damage and functional damage in glaucoma, so it is particularly important to locate and quantify the regions in fundus images. In the first work of this thesis, a visualization method for glaucoma diagnosis based on convolutional neural network is proposed. Firstly, a dynamic receptive field module is introduced on the basis of the multi-scale feature pyramid structure to solve the problem that feature information cannot be accurately focused by attention through the embedded coordinate attention mechanism. Secondly, the fused gradient-weighted class activation heat map is used to provide more detailed attention maps, which removes the high-level connectivity constraints in the visualization convolutional layer feature map method. Experimental results show that the method proposed in this thesis not only has higher classification accuracy, but also provides visual evidence of focal areas, which is easier to be recognized by doctors and patients.

(2) The information of evaluating glaucoma based only on fundus images is relatively limited. The second work of this thesis draws on the experience of clinical diagnosis and proposes a knowledge map-based etiological inference method for glaucoma. The method consists of problem relation analysis module, etiology reasoning module and answer prediction module. In order to extract relevant information from the knowledge graph, the problem relation analysis module uses graph neural network and conditional random field to extract the desired relation and target entity from the problem. After extracting the information, the etiological reasoning module estimated the conditional expectation and propensity score of each information. The answer prediction module uses the propensity score and conditional expectation of each piece of information to calculate the causal effect of the relationship to be solved in the problem and completes the prediction analysis. Finally, the Neo4j tool was used to realize data visualization and build a glaucoma reasoning question and answer system based on knowledge map. The experimental results show that the method proposed in this thesis provides a new way to solve the problems that do not have answers in the nodes, and the designed external knowledge carrier of inference has a certain advancement in the accuracy of the answers on the open data set.

[1] 王宁利, 辛晨, 张敬学,等. 中国青光眼防治工作展望[J]. 眼科学报, 2021, 36(6): 5.

[2] 钱朝旭, 钟华. 人工智能在青光眼领域的研究进展[J]. 国际眼科杂志, 2021, 21(12): 5.

[3] 张巧丽, 赵地, 迟学斌. 基于深度学习的医学影像诊断综述[J]. 计算机科学, 2017, 44(B11): 7.

[4] Ting D S W, Pasquale L R, Peng L, et al. Artificial intelligence and deep learning in ophthalmology[J]. British Journal of Ophthalmology, 2019, 103(2): 167-175.

[5] Agarwal A, Gulia S, Chaudhary S, et al. A novel approach to detect glaucoma in retinal fundus images using cup-disk and rim-disk ratio[C]//2015 4th international work conference on bioinspired intelligence(IWOBI). IEEE, 2015: 139-144.

[6] Lee M H, Kim H K, Kim S S. Risk Factors Associated with a Large Vertical Cup-to-Disc Ratio: Korean National Health and Nutritional Examination Survey[J]. Journal of Glaucoma, 2023, 32(3): 221-226.

[7] Issac A, Parthasarthi M, Dutta M K. An adaptive threshold based algorithm for optic disc and cup segmentation in fundus images[C]//2015 2nd international conference on signal processing and integrated networks(SPIN). IEEE, 2015: 143-147.

[8] Osareh A, Mirmehdi M, Thomas B, et al. Comparison of Colour Spaces for Optic Disc Localisation in Retinal Images[C]//2002 16th International Conference on Pattern Recognition. IEEE Computer Society, 2002.

[9] Blanco M, Penedo M G, Barreira N, et al. Localization and extraction of the optic disc using the fuzzy circular hough transform[C]//International conference on artificial intelligence and soft computing. Springer, Berlin, Heidelberg, 2006: 712-721.

[10] Mahapatra D, Buhmann J M. A field of experts model for optic cup and disc segmentation from retinal fundus images[C]//2015 12th international Symposium on Biomedical Imaging(ISBI). IEEE, 2015: 218-221.

[11] Abramoff M D, Alward W L M, Greenlee E C, et al. Automated segmentation of the optic disc from stereo color photographs using physiologically plausible features[J]. Investigative ophthalmology & visual science, 2007, 48(4): 1665-1673.

[12] Noronha K P, Acharya U R, Nayak K P, et al. Automated classification of glaucoma stages using higher order cumulant features[J]. Biomedical Signal Processing and Control, 2014, 10: 174-183.

[13] Sevastopolsky A. Optic disc and cup segmentation methods for glaucoma detection with modification of U-Net convolutional neural network[J]. Pattern Recognition and Image Analysis, 2017, 27: 618-624.

[14] Ahn J M, Kim S, Ahn K S, et al. Correction: A deep learning model for the detection of both advanced and early glaucoma using fundus photography[J]. PlOS ONE, 2019, 14(1): e0211579.

[15] Bander B, Nuaimy W, Taee M A, et al. Automated glaucoma diagnosis using deep learning approach[C]//2017 14th international Multi-Conference on Systems, Signals & Devices(SSD). IEEE, 2017: 207-210.

[16] Paschali M, Naeem M F, Simson W, et al. Deep learning under the microscope: improving the interpretability of medical imaging neural networks[J]. arXiv preprint arXiv: 1904.03127, 2019.

[17] Zhang Z, Chen P, Sapkota M, et al. TandemNet: Distilling Knowledge from Medical Images Using Diagnostic Reports as Optional Semantic References[J]. Springer, Cham, 2017.

[18] Niu Y, Gu L, Lu F, et al. Pathological evidence exploration in deep retinal image diagnosis[C]//Proceedings of the AAAI conference on artificial intelligence, 2019, 33(01): 1093-1101.

[19] Tham Y C, Li X, Wong T Y, et al. Global prevalence of glaucoma and projections of gla-ucoma burden through 2040: a systematic review and meta-analysis[J]. Ophthalmology, 2014, 121(11): 2081-2090.

[20] 张芳, 赵东旭, 肖志涛,等. 眼底图像质量分类综述[J]. 计算机辅助设计与图形学学报, 2020, 32(3): 12.

[21] Brigatti L, Hoffman D, Caprioli J. Neural networks to identify glaucoma with structuraland functional measurements[J]. American journal of ophthalmology, 1996, 12(15): 511-521.

[22] Huang D, Swanson E A, Lin C P, et al. Optical coherence tomography[J]. science, 1991, 254(5035): 1178-1181.

[23] 熊荔. 基于彩色眼底图的图像分析与疾病自动诊断算法研究[D]. 北京: 北京理工大学, 2017.

[24] Krizhevsky A, Sutskever I, Hinton G E. Imagenet classification with deep convolutional neural networks[J]. Communications of the ACM, 2017, 60(6): 84-90.

[25] He K, Zhang X, Ren S, et al. Deep residual learning for image recognition[C]// Proceedings of the IEEE conference on computer vision and pattern recognition, 2016: 770-778.

[26] Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition[J]. arXiv preprint arXiv: 1409.1556, 2014.

[27] 梁蒙蒙, 周涛, 张飞飞,等. 卷积神经网络及其在医学图像分析中的应用研究[J]. 生物医学工程学杂志, 2018, 35(6): 9.

[28] David M S, Renjith S. Comparison of word embeddings in text classification based on RNN and CNN[J]. IOP Conference Series: Materials Science and Engineering, 2021, 1187(1): 56-66.

[29] Mohd S E, Ibrahim H, Lai C Q, et al. A Long Short-Term Memory Network Using Resting-State Electroence phalogram to Predict Outcomes Following Moderate Traumatic Brain Injury[J]. Computers, 2023, 12(2): 45.

[30] Wen C, Chen T, Jia X, et al. Medical Named Entity Recognition from Un-labelled Medical Records based on Pre-trained Language Models and Domain Dictionary[J]. Data Intelligence, 2021, 3(3): 402-417.

[31] 李小亚. 医学超声领域知识图谱的设计与实现[D]. 北京: 北京邮电大学, 2021.

[32] Yang X. A Historical Review of Collaborative Learning and Cooperative Learning[J]. TechTrends, 2023: 1-11.

[33] 刘峥峥, 蒋凡, 杨俊. 模型转换规则自动生成研究[J]. 计算机工程与应用, 2010, 46(8): 6.

[34] Humphreys K, Gaizauskas R, Azzam S, et al. University of Sheffield: Description of the LaSIE-II system as used for MUC-7[C]//Conference on Message Understanding. Association for Computational Linguistics, 1995.

[35] Liu Y, Culpepper S A, Chen Y. Identifiability of Hidden Markov Models for Learning Trajectories in Cognitive Diagnosis[J]. Psychometrika, 2023: 1-26.

[36] 王彦林, 金汉均, 梅洪洋. 二维最大熵模型在图像分类算法中的应用研究[J]. 华中师范大学学报: 自然科学版, 2015, 49(4): 4.

[37] Lv C, Pan D, Li Y, et al. A novel Chinese entity relationship extraction method based on the bidirectional maximum entropy Markov model[J]. Complexity, 2021, 2021: 1-8..

[38] 潘华山, 严馨, 余正涛,等. 基于支持向量机的越语新闻文本分类方法[J]. 山西大学学报: 自然科学版, 2013, 36(4): 5.

[39] Feng C L, Liu C, Jiang D X. Unsupervised anomaly detection using graph neural networks integrated with physical-statistical feature fusion and local-global learning[J]. Renewable Energy, 2023, 206.

[40] 李魁. 结合卷积神经网络与条件随机场的高光谱半监督分类[D]. 北京: 中国地质大学, 2021.

[41] 罗平. 基于深度学习的中文实体识别及关系抽取研究[D]. 兰州: 兰州交通大学, 2022.

[42] Miao Y L, Cheng W F, Ji Y C, et al. Aspect-based sentiment analysis in Chinese based on mobile reviews for BiLSTM-CRF[J]. Journal of Intelligent & Fuzzy Systems, 2021, 40(5): 8697-8707.

[43] 邢毅雪, 朱永华, 高海燕,等. 基于注意力机制的远程监督实体关系抽取[J]. 上海大学学报:自然科学版, 2021, 27(5): 10.

[44] 董淼, 苏中琪, 周晓北,等. 利用Text-CNN改进PubMedBERT在化学诱导性疾病实体关系分类效果的尝试[J]. 现代图书情报技术, 2021(011): 005.

[45] 金留可. 基于递归神经网络的生物医学命名实体识别[D]. 大连: 大连理工大学, 2016.

[46] 胡怡然, 夏芳. 基于自注意力机制与BiLSTM的短文本匹配模型[J]. 武汉科技大学学报, 2023, 46(1): 6.

[47] Feng C L, Liu C, Jiang D X. Unsupervised anomaly detection using graph neural networks integrated with physical-statistical feature fusion and local-global learning[J]. Renewable Energy, 2023, 206.

[48] 杨霖. 半结构化数据蕴涵规则提取方法的研究[D]. 辽宁: 辽宁工业大学, 2018.

[49] 张吉祥, 张祥森, 武长旭,等. 知识图谱构建技术综述[J]. 计算机工程, 2022, 48(3): 15.

[50] 张婷婷, 马明栋, 王得玉. OCR文字识别技术的研究[J]. 计算机技术与发展, 2020, 30(04): 85-88.

[51] 蒋婷婷. 面向知识图谱的实体对齐方法研究[D]. 合肥: 合肥工业大学, 2022.

[52] Saad M, Zhang Y, Tian J, et al. A graph database for life cycle inventory using Neo4j[J]. Journal of Cleaner Production, 2023: 136344.

[53] Biran O, Cotton C. Explanation and justification in machine learning: A survey[C]// Proceedings of IJCAI-17 Workshop on Explainable AI (XAI). Melbourn: IJCAI, 2017, 8: 1-5.

[54] Yan J, Zhang B, Zhou M, et al. Multi-Branch-CNN: Classification of ion channel interacting peptides using multi-branch convolutional neural network[J]. Computers in Biology and Medicine, 2022, 147: 105717.

[55] 颜玉松, 尹芳洁, 王彩玲. 融合Xception特征提取和坐标注意力机制的血细胞分割[J]. 计算机系统应用, 2023, 32(1): 6.

[56] Salehi A, Balasubramanian M. DDCNet: Deep dilated convolutional neural network for dense prediction[J]. Neurocomputing, 2023, 523: 116-129.

[57] Selvaraju R R, Cogswell M, Das A, et al. Grad-cam: Visual explanations from deep networks via gradient-based localization[C]//Proceedings of the IEEE international conference on computer vision, 2017: 618-626.

[58] Acharya U R, Ng E Y K, Eugene L W J, et al. Decision support system for the glaucoma using Gabor transformation[J]. Biomedical Signal Processing and Control, 2015, 15: 18-26.

[59] Dua S, Acharya U R, Chowriappa P, et al. Wavelet-based energy features for glaucomatous image classification[J]. IEEE transactions on information technology in biomedicine, 2011, 16(1): 80-87.

[60] Bock R, Meier J, Nyúl L G, et al. Glaucoma risk index: automated glaucoma detection from color fundus images[J]. Medical image analysis, 2010, 14(3): 471-481.

[61] Cheng J, Liu J, Xu Y, et al. Superpixel classification based optic disc and optic cup segmentation for glaucoma screening[J]. IEEE transactions on medical imaging, 2013, 32(6): 1019-1032.

[62] 李飞. 基于知识图谱的问答系统研究与实现[D]. 南京: 南京邮电大学, 2022.

[63] Rytgaard H C W, Ekstrøm C T, Kessing L V, et al. Ranking of average treatment effects with generalized random forests for time‐to‐event outcomes[J]. Statistics in Medicine, 2023.

[64] 马智勤. 基于分布式 ElasticSearch 相似内容比对算法研究与应用[D]. 成都: 四川师范大学, 2021