第一部分：高阶像差及光密度等角膜参数的圆锥角膜进展评估效力分析‌

目的‌：分析角膜高阶像差（HOA）、角膜光密度（CD）及角膜地形参数在圆锥角膜（KC）进展中的变化，并探讨它们在评估KC进展中的临床应用价值。

‌方法‌：本研究纳入2019年1月至2024年1月于北京协和医院眼科门诊就诊的282例（402眼）KC患者，通过Pentacam眼前节分析仪收集患者的角膜HOAs、CD及角膜地形参数，并验证相关参数在KC患者及飞秒激光小切口角膜基质透镜取出术（SMILE）前后的测量重复性。根据KC进展的定义，将患者分为进展组（96眼）和未进展组（306眼）。对两组随访前后数据以及两组间进行对比分析，通过单因素及多因素logistic回归分析确定与KC进展相关的参数，并评估其预测效能。

‌结果‌：KC进展组与未进展组在性别、年龄等基线特征上无显著差异，但进展组中央角膜厚度（CCT）更薄、角膜平均曲率（Km）更高。重复性检验显示所有参数在KC患者和SMILE手术前后的组内相关系数（ICC）均高于0.9，表明数据可靠性高。对比两组参数变化发现，未进展组中角膜垂直非对称性指数（IVA）增加、圆锥角膜指数（KI）下降，且角膜前表面HOAs、水平/垂直彗差（CA）及前层CD局部区域显著升高（P<0.05）。进展组则呈现更广泛且显著的角膜形态恶化（P<0.05）。多因素logistic回归模型筛选出Km、最薄角膜厚度（TCT）、平均角膜厚度进展指数（PPI Avg）、水平CA、垂直CA及前层2-6mm CD为KC进展的独立预测因子。其中，角膜水平CA和垂直CA的受试者工作特征曲线下面积（AUC）分别为0.832和0.820，显示出较高的检测效能。

‌结论‌：KC进展不仅伴随典型角膜曲率增加和厚度变薄，更与HOAs的快速增加及前基质层CD异常升高密切相关。角膜CA和CD参数的动态监测可为早期识别KC进展提供高精度指标，其诊断效能超越传统形态学参数。

第二部分：基于角膜图像的深度学习模型预测圆锥角膜进展的研究

目的：旨在开发一种基于深度学习（DL）的圆锥角膜（KC）进展预测模型，通过整合角膜图像数据，特别是角膜高阶像差（HOA）和角膜光密度（CD）图像，以更准确地量化评估KC进展风险。

方法：本研究纳入了来自多个数据集的508例（830眼）KC患者，收集其首次就诊时的Pentacam眼前节分析仪图像，包括角膜厚度图、角膜屈光力图、角膜前/后表面高度图、角膜HOA分布图及CD分布图。对图像进行标注和预处理后，构建基于NASNet架构的DL预测模型以及综合了6类图像的综合化DL预测模型。模型通过图像数据扩充、训练和验证，最终使用测试集评估其检测效能，并与人工检测进行对比。

结果：基于角膜后表面高度图的DL模型在基于角膜地形图像的DL模型中表现最优，准确度为0.804，特异度为94.9%，受试者工作特征曲线下面积（AUC）值为0.753。而基于角膜前表面HOA分布图的模型在基于HOAs及CD图像的DL模型对比中表现最佳，准确度达0.871，特异度为99.1%，AUC值为0.831。综合化DL模型通过整合多模态数据，进一步提升了预测能力，其中加载前表面HOA图像的综合DL模型准确度高达0.917，特异度为98.8%，AUC值为0.895。所有基于角膜HOAs及CD参数的DL模型的检测效能（除球差外）均超越人工检测。

结论：本研究成功构建了基于多模态角膜图像的DL预测模型，能够高精度地预测KC进展。角膜后表面高度图与前表面HOA分布图对KC进展具有重要鉴别价值，而综合化DL模型通过融合多维度特征，显著优于传统人工检测方法，为KC进展的无创精准预测提供了新策略。

第三部分：不同角膜交联术治疗进展期圆锥角膜的预后情况及深度学习预测模型的构建

目的‌：旨在对比传统（去上皮）角膜交联术（CXL）、跨上皮CXL及加速CXL在控制圆锥角膜（KC）进展方面的差异，并进一步通过综合各类角膜图像（尤其是高阶像差（HOA）及角膜光密度（CD）图像）构建CXL治疗效果的综合化深度学习（DL）预测模型，为未来CXL的治疗与否及方案选择提供参考。

‌方法‌：本研究选取了2023年7月至2024年7月在北京协和医院门诊就诊并被诊断为KC的患者，同时从多个数据集收集了相关数据。共纳入342例（582眼）接受了不同CXL手术的KC患者。收集患者的Pentacam眼前节分析图像及相关角膜参数，并进行不同CXL手术之间及CXL术后KC进展与否之间对比。采用基于6类角膜图像的综合化DL模型，构建并验证不同CXL手术方式的术后KC进展预测模型。

‌结果‌：传统CXL、跨上皮CXL和加速CXL组的KC进展率分别为11.85%、16.39%和11.70%。跨上皮组的进展率显著高于其他两组（P<0.05）。角膜参数分析显示，传统和加速CXL在减少角膜曲率和HOAs方面优于跨上皮CXL。CXL术后KC进展组在绝大多数角膜地形参数、总HOA和前后表面HOA参数中均较术后未进展组出现恶化（P<0.05）。而角膜前层CD在CXL术后进展组和未进展组间未发现明显差异。DL模型预测效能评估表明，基于跨上皮CXL的DL模型的预测准确度最高（0.867），其次为传统CXL（0.853）和加速CXL（0.842），均显著优于人工检测（0.713）。

‌结论‌：传统与加速CXL在控制KC进展及改善角膜形态方面效果相近，而跨上皮CXL疗效相对较弱。DL模型在预测CXL术后KC进展方面表现出高精度，尤其是基于跨上皮CXL的DL模型。这为个体化术后评估提供了有力工具，有助于优化CXL手术方案的选择，提高治疗效果。

Part I: Analysis of the efficacy of corneal parameters such as high-order aberration and densitometry in the assessment of keratoconusl progression

Purpose: To analyse the changes of corneal higher-order aberration (HOA), corneal densitometry (CD) and corneal topographic parameters in the progression of keratoconus (KC) and to explore the value of their clinical application in assessing KC progression.

Methods: This study included 282 cases (402 eyes) of KC patients who attended the outpatient ophthalmology clinic of Peking Union Medical College Hospital from January 2019 to January 2024, and the corneal HOAs, CDs, and corneal topographic parameters of the patients were collected by the Pentacam anterior segment analyser, and verify the measurement repeatability of relevant parameters from KC patients and before and after small incision lenticule extraction (SMILE). Based on the definition of KC progression, the patients were divided into a progressive group (96 eyes) and a non-progressive group (306 eyes). Comparative analyses of the data before and after follow-up and between the two groups were performed, and parameters related to KC progression were identified by univariate and multivariate logistic regression analyses, and the predictive efficacy was also assessed. 

Results: There were no significant differences between the KC progression and non-progression groups in baseline characteristics such as gender and age, but the progression group had thinner central corneal thickness (CCT) and higher mean keratometry (Km). Repeatability tests showed that the intragroup correlation coefficients (ICC) for all parameters were higher than 0.9 from KC and before and after SMILE, indicating high reliability of the data. Comparison of the parameter changes between the two groups revealed an increase in the corneal index of vertical asymmetry (IVA), a decrease in the keratoconus index (KI), and a significant increase in HOAs, horizontal/vertical coma aberrations (CA), and localised areas of anterior CD in the nonprogressed group (P<0.05). The progressive group showed a more extensive and significant corneal morphological deterioration (P<0.05). A multifactorial logistic regression model screened Km, thinnest corneal thickness (TCT), average pachymetry progression index (PPI Avg), horizontal CA, vertical CA, and anterior layer 2-6 mm CD as independent predictors of KC progression. Among them, the area under the curve of the receiver operating characteristic (AUC) with horizontal and vertical corneal CA was 0.832 and 0.820, respectively, which showed a high detection efficacy. 

Conclusions: KC progression is not only accompanied by the typical keratometry increase and thickness thinning, but also closely associated with the rapid increase of HOAs and abnormal elevation of CD in the anterior stromal layer. Dynamic monitoring of corneal CA and CD parameters can provide a highly accurate indicator for early identification of KC progression, and its diagnostic efficacy exceeds that of traditional morphological parameters.

Part II: A deep learning model based corneal images for predicting progression of keratoconus

Purpose: The aim was to develop a deep learning (DL)-based prediction model for keratoconus (KC) progression by integrating corneal images, particularly corneal higher-order aberration (HOA) and corneal densitometry (CD) images, in order to more accurately and quantitatively assess the risk of KC progression.

Methods: This study included 508 patients (830 eyes) with KC from multiple datasets, whose Pentacam anterior segment analyser images were collected at the time of their first visit, including corneal thickness maps, corneal refractive maps, corneal anterior/posterior surface height maps, corneal HOA distribution maps and CD distribution maps. After annotation and preprocessing of the images, a DL prediction model based on the NASNet architecture and an integrated DL prediction model that combined the six types of images were constructed. The models were expanded, trained and validated with images, and finally their detection efficacy was evaluated using a test set and compared with manual detection.

Results: The DL model based on the corneal posterior surface height map performed best among the DL models based on corneal topographic images, with an accuracy of 0.804, a specificity of 94.9%, and an area under the curve of the receiver operating characteristic (AUC) value of 0.753. Whereas, the model based on the HOA distribution map of the corneal anterior surface performed best in the comparison of the DL models based on the HOAs and the CD images, with an accuracy of 0.871, a specificity of 99.1% with an AUC value of 0.831. The integrated DL model further improved its predictive ability by integrating multimodal data, with the integrated DL model loaded with anterior surface HOA images having a high accuracy of 0.917, a specificity of 98.8%, and an AUC value of 0.895. The detection efficacy of all the DL models based on corneal HOAs and CD parameters (except for spherical aberration) exceeded manual detection.

Conclusions: In this study, we successfully constructed a DL prediction model based on multimodal corneal images, which can predict KC progression with high accuracy. The posterior surface height map and the anterior surface HOA distribution map are important for identifying KC progression, and the integrated DL model significantly outperforms the traditional manual detection method by integrating multidimensional features, which provides a new strategy for non-invasive and accurate prediction of KC progression.

Part III: Prognosis of different corneal cross-linking procedures for the treatment of progressive keratoconus and construction of a deep learning prediction model

Purpose: The aim is to compare the differences between conventional (epi-off) corneal cross-linking (CXL), epi-on CXL and accelerated CXL in controlling the progression of keratoconus (KC), and to further construct an integrated deep learning (DL) of the treatment effect of CXL by combining various types of corneal images, especially higher order aberration (HOA) and corneal densitometry (CD) images. prediction model for future CXL treatment or not and protocol selection.

Methods: Patients who attended outpatient clinics at Peking Union Medical College Hospital from July 2023 to July 2024 and were diagnosed with KC were selected for this study, and relevant data were also collected from other datasets. A total of 342 patients (582 eyes) with KC who underwent different CXL procedures were included. Pentacam anterior segment analysis images and related corneal parameters of patients were collected and compared between different CXL surgeries and between KC progression or not after CXL. A synthesised DL model based on 6 types of corneal images was used to construct and validate a predictive model for postoperative KC progression between different CXL surgical procedures.

Results: The KC progression rates in the conventional CXL, epi-on CXL, and accelerated CXL groups were 11.85%, 16.39%, and 11.70%, respectively. The progression rate in the epi-on CXL group was significantly higher than the other two groups (P<0.05). Analysis of corneal parameters showed that conventional and accelerated CXL were superior to epi-on CXL in reducing keratometry and HOAs. The post-CXL KC progression group showed deterioration in the vast majority of corneal topographic parameters, total HOA, and anterior/posterior surface HOA parameters compared with the postoperative non-progressed group (P<0.05). In contrast, no significant difference in anterior corneal CD was found between the post-CXL progressed and non-progressed groups. Evaluation of the predictive efficacy of the DL model showed that the DL model based on epi-on CXL had the highest predictive accuracy (0.867), followed by conventional CXL (0.853) and accelerated CXL (0.842), which were significantly better than the manual test (0.713).

Conclusions: Conventional and accelerated CXL were similarly effective in controlling KC progression and improving corneal morphology, whereas the efficacy of epi-on CXL was relatively weaker. DL models demonstrated high accuracy in predicting KC progression after CXL, especially the DL model based on epi-on CXL. This provides a powerful tool for individualised postoperative assessment, which can help optimise the choice of CXL surgical plan and improve treatment outcomes.