As advances in deep learning have quickly permeated into the field of ophthalmology, supervised deep learning approaches have achieved great performance in eye disease classification tasks such as diabetic retinopathy, age-related macular degeneration, and glaucoma. However, disease-specific classification models are limited in scope with regard to detecting other ocular diseases that are not included in the training set. One solution to this is to develop a multiclass model trained on a broad array of ocular diseases. However, this solution would require large data sets comprising comprehensive and balanced distributions of many ocular diseases, as well as necessity for labor- and cost-intensive sample annotations. A more scalable solution would be to use a semisupervised or unsupervised approach in order to train a general anomaly detection algorithm. In addition to reducing the cost of image annotation, AD models could also screen large populations ahead of a more comprehensive workup.

中文翻译：

---

眼科中基于机器学习的异常检测技术

随着深度学习的进步迅速渗透到眼科领域，有监督的深度学习方法在糖尿病视网膜病变、年龄相关性黄斑变性和青光眼等眼病分类任务中取得了很好的成绩。然而，疾病特异性分类模型在检测其他未包含在训练集中的眼部疾病方面的范围有限。对此的一种解决方案是开发一个针对各种眼部疾病进行训练的多类模型。然而，该解决方案需要包含许多眼部疾病的全面和平衡分布的大型数据集，以及对劳动和成本密集型样本注释的必要性。一个更具可扩展性的解决方案是使用半监督或无监督方法来训练通用异常检测算法。除了降低图像注释的成本外，AD 模型还可以在更全面的检查之前筛选大量人群。