眼科超声生物显微镜用来成像角膜、前房和晶状体等眼前节结构，为青光眼的早期诊断提供重要参考。我国青光眼基层诊疗压力大，建设人工智能与分级诊疗相结合的医疗体系能够推动医疗资源的共享。本文基于深度学习方法定量测量UBM图像中的各眼前节生理参数，期望为屈光手术、白内障超声乳化手术和青光眼等眼科疾病的诊治提供参考信息。

本文制作了全景UBM图像眼前节参数测量数据集。足够数量且质量符合要求的临床医学图像能够保证模型在临床应用场景中的泛化性能，使用临床采集的全景UBM图像作为数据来源，并让高年资眼科专家在保持数据脱敏的情况下进行数据清洗，使用labelme标注软件进行标注，并在标注之后由另一位眼科专家检查标注的合理性，确保了研究数据的规范性和有效性，最终共有717幅临床全景UBM图像及其金标准被纳入研究。

由于医学数据本身的稀缺性，为防止由于数据不足导致的模型泛化能力不强，本文使用StyleGAN3生成虚拟UBM图像，扩充数据集，提高了算法的泛化性能，缓解模型过拟合。使用临床UBM图像训练可将随机噪声向量生成为虚拟UBM图像的生成对抗网络StyleGAN3，同时使用临床UBM图像训练一个可为UBM图像生成十个眼前节关键点坐标的目标检测网络。使用StyleGAN3生成2210幅虚拟UBM图像，使用目标检测网络推理虚拟UBM图像的伪标签，经眼科专家筛除其中图像或伪标签不合格的数据，最终共有717幅虚拟UBM图像及其伪标签被纳入研究。虚拟UBM图像数据集能够避免医学深度学习中伦理隐私、法律法规方面的风险，降低医学数据获取的门槛，让更多对医学深度学习感兴趣的研究者参与其中。

本文使用YOLOv5、YOLOv8、DETR等目标检测算法作为眼前节关键点预测算法，结合生成对抗网络、迁移学习、数据增强、遗传算法等技术，通过对照实验筛选出最优的算法组合，实现了实时精确的眼前节关键点定位。由于直接使用目标检测算法进行多分类会引入分类误差，且实际分类效果并不能满足需求，于是采用单分类算法。由于目标检测算法输出眼前节关键点坐标时，各个点之间的排序是无序的，因此根据眼前节的生理结构设计眼前节关键点排序算法，以得到各个图像中的预测点的医学名称。在测量出眼前节生理参数之后，使用各参数的临床参考值范围作为判定依据，以筛除超声扫查切面不正的UBM图像或病理性结构变化的UBM图像。

结果表明，以深度学习算法为核心的眼前节参数测量系统对眼前房关键点的定位误差为66.27±66.25 μm，对中央角膜厚度的相对误差为9.61%，其它眼前节参数的预测相对误差均在3%以下,在配备Ryzen 5 4600U处理器的笔记本电脑上预测一张UBM图片仅需要140 ms，精度和速度都符合应用场景的要求。以DETR为核心算法的眼前节参数测量系统对眼前房关键点的准确度则更高，为61.84±37.38 μm，但对算力要求较高。本文结合目标检测算法和生成对抗网络制作了虚拟数据集，使用虚拟数据集训练出的模型在真实数据上的预测精度为62.83±38.21 μm，与真实数据集训练的模型精度相比无统计学上的显著差异。

       基于以上实验，本文实现了高精度与实时的眼前节参数测量算法，并提出了一种虚拟数据集的制作方法，具有较好的临床应用场景。

Ultrasound biometrics in ophthalmology are used to image the anterior segment structures such as cornea, anterior chamber and lens, providing important reference for the early diagnosis of glaucoma. The primary diagnosis and treatment of glaucoma in China is under great pressure, and the construction of a medical system combining artificial intelligence and hierarchical diagnosis and treatment can promote the sharing of medical resources. In this paper, the physiological parameters of each anterior segment in UBM images were quantitatively measured based on deep learning method, hoping to provide reference information for the diagnosis and treatment of eye diseases such as refractive surgery, cataract phacoemulsification surgery and glaucoma.

In this paper, a data set for measurement of anterior segment parameters in panoramic UBM images is developed. Sufficient quantity and quality of clinical medical images can ensure the generalization performance of the model in clinical application scenarios. Panoramic UBM images collected in clinical practice are used as data sources, and senior ophthalmologists are allowed to carry out data cleaning while maintaining data desensitization. labelme labeling software is used for labeling. After the annotation, another ophthalmologist checked the rationality of the annotation to ensure the standardization and validity of the study data, and a total of 717 clinical panoramic UBM images and its gold standard were finally included in the study.

Due to the scarcity of medical data itself, in order to prevent the poor generalization ability of the model caused by insufficient data, StyleGAN3 was used in this paper to generate virtual UBM images, expand the data set, improve the generalization performance of the algorithm, and alleviate the overfitting of the model. Using clinical UBM image training, a random noise vector can be generated into a generative adjunct network StyleGAN3 for virtual UBM images, while using clinical UBM images to train an object detection network that can generate ten focal point coordinates for UBM images. StyleGAN3 was used to generate 2210 virtual UBM images, the object detection network was used to deduce false labels of virtual UBM images, and after ophthalmology experts screened out unqualified data of images or false labels, a total of 717 virtual UBM images and false labels were finally included in the study. Virtual UBM image dataset can help reduce the cost of obtaining medical data, avoid the risks of ethical privacy and laws and regulations in medical deep learning, lower the threshold of obtaining medical data, and allow more researchers interested in medical deep learning to participate in it.

YOLOv5, YOLOv8, DETR and other target detection algorithms are used as the key point prediction algorithm for the anterior segment. Combined with generative adversarial network, transfer learning, data enhancement, genetic algorithm and other technologies, the optimal combination of algorithms is selected through controlled experiments to achieve real-time and accurate key point location for the anterior segment. Because the direct use of object detection algorithm for multiple classification will introduce classification errors, and the actual classification effect can not meet the needs, so the single classification algorithm is used. Since the sequence of the key points in the anterior segment is disordered when the object detection algorithm outputs the key points in the anterior segment, the key points sorting algorithm is designed according to the physiological structure of the anterior segment to obtain the medical name of the predicted points in each image. After the physiological parameters of the anterior segment were measured, the clinical reference value range of each parameter was used as the basis to screen out the UBM images with abnormal section or the UBM images with pathological structural changes.

The results show that the positioning error of the anterior node parameter measurement system with deep learning algorithm as the core is 66.27± 66.25μm, and the prediction time is less than 150ms on the portable laptop computer without independent graphics card, both accuracy and speed meet the requirements of the application scenario. The anterior segment parameter measurement system with DETR as the core algorithm has a higher accuracy of 61.84± 37.38μm for the key point of the anterior chamber, but it has a higher requirement for computing power. In this paper, a virtual data set is created by combining the target detection algorithm and the generative adversarial network. The prediction accuracy of the model trained by the virtual data set is 62.83± 38.21μm on the real data, which has no statistically significant difference compared with the model trained by the real data set.

Based on the above experiments, this paper realizes the high-precision and real-time anterior segment parameter measurement algorithm, and proposes a virtual data set production method, which has a good clinical application scenario.