眼健康是国民健康的重要组成部分，随着人口老龄化和不良用眼习惯的增加，我国眼科疾病患者人数不断增长。超声成像是眼科临床中常用的诊断与筛查方法之一，尤其对于屈光介质混浊的眼科疾病至关重要。然而，超声成像高度依赖临床经验，而由于高水平医生数量有限，导致眼科超声影像面临巨大的识别压力。人工智能在眼科诊疗领域可以承担疾病诊断、病情预测以及手术规划和辅助等任务，推动眼科诊疗趋于规范化、个体化、精准化。因此，本文提出了一种基于深度学习的眼科超声影像智能辅助诊断方法，旨在提高判别准确性、减轻临床工作负担，并改善眼科诊疗现状。

本文首先构建疾病分类和病灶区域检测模型，实现基于眼科超声影像的疾病自动识别、静态图像病灶区域自动检测和动态视频影像病灶区域的实时跟踪；然后基于 B 型超声引导下的眼组织分段声速匹配技术，研究眼组织的自动定位方法，实现对角膜前沿、晶状体前后囊和黄斑中心凹的自动检测，并定量测量前房深度、晶状体厚度、玻璃体腔长度和眼轴长度；最后利用 PyQt 工具包开发眼科超声影像智能辅助诊断软件，实现对眼后节图像分类模型、病灶区域检测和眼组织自动测量模型的集成。

结果表明，用于疾病分类的 Inceptionv3-Xception 集成模型在测试集上的准确率、精确度、灵敏度和 F1 值分别为 96.73%、95.21%、95.28%和 95.23%，受试者工作特征曲线下面积达到了 0.9988。用于病灶区域检测的 MobileNetv2_Yolov5s 模型在测试集上的多类别平均精度、模型参数量和帧频分别为 99.39%、4.61×106 和41 张/秒。用于眼组织定位的Yolov5s_mobilenetv2_Ensembled 模型在测试集上检测角膜前沿、晶状体前后囊和视网膜黄斑中心凹的多类别平均精度、精确度、召回率、F1 值和帧频达到了92.69%、94.12%、92.92%、93.60%和 40 张/秒。使用自动测量算法计算前房深度、晶状体厚度、玻璃体腔长度和眼轴长度的结果与手动标注结果的误差平均值均小于 0.150mm，误差标准差均小于 0.283mm。辅助诊断软件可实现良好的人机交互功能，能够实现眼科超声影像分类、病灶区域检测和眼组织参数测量，且具有良好的可移植性。

基于深度学习的眼科超声影像智能辅助诊断方法能够快速准确地实现眼科超声影像的智能识别、病灶区域的实时跟踪检测以及眼组织参数自动测量，验证了深度学习识别眼后节图像的可行性，有助于眼后节疾病的早期筛选、诊断和管理。

Eye health is a crucial aspect of national well-being. The number of patients with ophthalmic diseases in China has been increasing due to the aging population and poor eye habits. Ultrasonography is a commonly used diagnostic and screening method in ophthalmology clinics, particularly for ophthalmic diseases with refractive media clouding. However, ultrasound imaging is reliant on clinical experience, and the limited number of highly qualified doctors creates significant pressure to accurately interpret ophthalmic ultrasound images. Artificial intelligence can assist in the diagnosis and treatment of diseases in the field of ophthalmology. It can perform tasks such as disease diagnosis, condition prediction, surgical planning and assistance. This promotesstandardization, individualization and precision in ophthalmology diagnosis and treatment. This paper proposes a deep learning-based method to assist in the diagnosis method of ophthalmic ultrasound images. It aims to improve discrimination accuracy, reduce clinical workload, and enhance the current state of ophthalmology diagnosis and treatment.

In this paper, we first constructed a disease classification and lesion region detection model. Automatic disease identification based on ophthalmic ultrasound images and the detection of lesion areas in static images and real-time tracking of lesion regions in dynamic video images were realized; Then, based on the segmented acoustic velocity measurement of ocular tissues under the guidance of B-mode ultrasound, we investigated an automatic localization and measurement of ocular tissues to realize the automatic detection of the corneal anterior edge, the anterior and posterior capsules of the lens and the central concavity of the macula. And quantitatively measure the anterior chamber depth, lens thickness, vitreous cavity length and ocular axis length; Finally, we developed ophthalmic ultrasound image intelligent auxiliary diagnosis software by using PyQt toolkit to realize the integration of the posterior segment image classification model, lesion area detection and ocular tissue automatic measurement model.

The results show that the Inceptionv3-Xception integrated model for disease classification on the test set achieved high accuracy, precision, sensitivity and F1 values of 96.73%, 95.21%, 95.28% and 95.23%, respectively, and the area under curve of the subjects reached 0.9988. The MobileNetv2_Yolov5s model for lesion area detection on the test set achieved the multi-category average precision, the number of model parameters and the frames per second of 99.39%, 4.61 × 106 and 41 frames/second, respectively. The Yolov5s_mobilenetv2_Ensembled model for ocular tissue localization reached 92.69%, 94.12%, 92.92%, 93.60%, and 40 frames/second for the multi-category average accuracy, precision, recall, F1 value and frame pre speed for detecting corneal frontal edge, anterior and posterior capsules of the lens, and the macular central pits of the retina on the test set. The mean error for caculating the anterior chamber depth, lens thickness, vitreous chamber length and ocular axis length using automatic measurement algorithm compared to manually annotated results were all less than 0.150 mm, with a standard deviation of less than 0.283 mm. The assisted diagnostic software enables effective human-computer interaction, ophthalmic ultrasound image classification, lesion area detection, and measurement of ocular tissue parameters, while also being highly portable.

The intelligent assisted diagnosis method of ophthalmic ultrasound images based on deep learning can quickly and accurately realize the intelligent recognition of ophthalmic ultrasound images, real-time tracking and detection of lesion areas, and automatic measurement of ocular tissue parameters, which verifies the feasibility of deep learning to recognize images of the posterior segment of the eye, and contributes to the early screening, diagnosis, and management of posterior segment diseases.