Bhatt Vaidehi D, Shah Nikhil, Bhatt Deepak C, Dabir Supriya, Sheth Jay, Berendschot Tos T J M, Erckens Roel J, Webers Carroll A B
UBM Institute, Mumbai, India.
Pursuing Masters in Computer Science at Stevens Institute of Technology, Jersey City, NJ, USA.
Clin Ophthalmol. 2025 Mar 18;19:939-947. doi: 10.2147/OPTH.S501316. eCollection 2025.
The purpose of this study was to create and test a deep learning algorithm that could identify and distinguish between membranes attached to optic disc [OD; retinal detachment (RD)/posterior vitreous detachment (PVD)] based on ocular ultrasonography (USG).
We obtained a database of B-scan ultrasonography from a high-volume imaging center. A transformer-based Vision Transformer (ViT) model was employed, pre-trained on ImageNet21K, to classify ultrasound B-scan images into healthy, RD, and PVD. Images were pre-processed using Hugging Face's AutoImage Processor for standardization. Labels were mapped to numerical values, and the dataset was split into training and validation (505 samples), and testing (212 samples) subsets to evaluate model performance. Alternate methods, such as ensemble strategies and object detection pipelines, were explored but showed limited improvement in classification accuracy.
The AI model demonstrated high classification performance, achieving an accuracy of 98.21% for PVD, 97.22% for RD, and 95.83% for normal cases. Sensitivity was 98.21% for PVD, 96.55% for RD, and 92.86% for normal cases, while specificity reached 95.16%, 100%, and 95.42%, respectively. Despite the overall strong performance, some misclassification occurred, with seven instances of RD being incorrectly labeled as PVD.
We developed a transformer-based deep learning algorithm for ocular ultrasonography that accurately identifies membranes attached to the optic disc, distinguishing between RD (97.22% accuracy) and PVD (98.21% accuracy). Despite seven misclassifications, our model demonstrates robust performance and enhances diagnostic efficiency in high-volume imaging settings, thereby facilitating timely referrals and ultimately improving patient outcomes in urgent care scenarios. Overall, this promising innovation shows potential for clinical adoption.
本研究的目的是创建并测试一种深度学习算法，该算法能够基于眼部超声检查（USG）识别并区分附着于视盘[OD；视网膜脱离（RD）/玻璃体后脱离（PVD）]的膜。
我们从一个大容量成像中心获取了B超超声检查数据库。采用基于变压器的视觉变换器（ViT）模型，该模型在ImageNet21K上进行了预训练，用于将超声B超图像分类为健康、RD和PVD。使用Hugging Face的自动图像处理器对图像进行预处理以实现标准化。将标签映射为数值，并将数据集分为训练集和验证集（505个样本）以及测试集（212个样本）子集，以评估模型性能。还探索了诸如集成策略和目标检测管道等替代方法，但在分类准确性方面显示出有限的提升。
人工智能模型表现出较高的分类性能，PVD的准确率为98.21%，RD的准确率为97.22%，正常病例的准确率为95.83%。PVD的灵敏度为98.21%，RD的灵敏度为96.55%，正常病例的灵敏度为92.86%，而特异性分别达到95.16%、100%和95.42%。尽管总体性能强劲，但仍发生了一些误分类，有7例RD被错误标记为PVD。
我们开发了一种基于变压器的眼部超声深度学习算法，该算法能够准确识别附着于视盘的膜，区分RD（准确率97.22%）和PVD（准确率98.21%）。尽管有7例误分类，但我们的模型在大容量成像环境中表现出强大的性能并提高了诊断效率，从而有助于及时转诊并最终改善紧急护理场景中的患者预后。总体而言，这项有前景的创新显示出临床应用的潜力。
