HPMI: A retinal fundus image dataset for identification of high and pathological myopia based on deep learning

Myopia is one of the leading causes of visual impairment worldwide and can progress to high or pathological myopia (HM or PM) if proper measures were not taken. Accurate identification of HM and PH plays an important role for their intervention and treatment, which can be implemented leveraging deep learning technology on sufficient annotation image data. However, few efforts have been made to construct public accessible annotation data for this task. In this paper, we constructed a retinal fundus image dataset (called HPMI) to identify the HM and PM. This dataset consists of 4011 fundus images with their corresponding annotations for HM and PM, which were confirmed by multiple ophthalmic examinations (e.g., visual acuity and axial length). To the best of our knowledge, this is the largest fundus image dataset for the classification of HM and PM. Based on the dataset, we further validated the classification potential of three representative deep learning networks (i.e., ResNet50, DenseNet121, and InceptionV3) and analyzed the consistency between prediction results and annotations

近视是全球视力损害的主要病因之一，若未采取恰当干预措施，可进展为高度近视（high myopia, HM）或病理性近视（pathological myopia, PM）。精准识别HM与PM对其干预与治疗具有重要价值，该任务可依托充足标注图像数据结合深度学习技术实现。然而，当前针对该任务构建可公开获取的标注数据集的相关研究较为匮乏。本文构建了一款用于识别HM与PM的视网膜眼底图像数据集（命名为HPMI）。该数据集包含4011张眼底图像及对应的HM与PM标注信息，所有标注均经多项眼科检查（如视力检测、眼轴长度测量）确认。据我们所知，这是目前用于HM与PM分类的规模最大的眼底图像数据集。基于该数据集，本文进一步验证了ResNet50、DenseNet121与InceptionV3三类典型深度学习网络的分类潜力，并分析了模型预测结果与人工标注之间的一致性。

搜集汇总

背景与挑战

背景概述

HPMI是一个包含4011张标注视网膜眼底图像的数据集，专门用于高度和病理性近视的深度学习识别研究，是目前该领域最大的公开数据集。数据集已用于验证ResNet50等三种深度学习网络的分类性能，标注结果经过多种眼科检查确认。