Purpose

Age-related macular degeneration (AMD) is a common threat to vision. While classification of disease stages is critical to understanding disease risk and progression, several systems based on color fundus photographs are known. Most of these require in-depth and time-consuming analysis of fundus images. Herein, we present an automated computer-based classification algorithm.

Design

Algorithm development for AMD classification based on a large collection of color fundus images. Validation is performed on a cross-sectional, population-based study.

Participants

We included 120 656 manually graded color fundus images from 3654 Age-Related Eye Disease Study (AREDS) participants. AREDS participants were >55 years of age, and non-AMD sight-threatening diseases were excluded at recruitment. In addition, performance of our algorithm was evaluated in 5555 fundus images from the population-based Kooperative Gesundheitsforschung in der Region Augsburg (KORA; Cooperative Health Research in the Region of Augsburg) study.

Methods

We defined 13 classes (9 AREDS steps, 3 late AMD stages, and 1 for ungradable images) and trained several convolution deep learning architectures. An ensemble of network architectures improved prediction accuracy. An independent dataset was used to evaluate the performance of our algorithm in a population-based study.

Main Outcome Measures

κ Statistics and accuracy to evaluate the concordance between predicted and expert human grader classification.

Results

A network ensemble of 6 different neural net architectures predicted the 13 classes in the AREDS test set with a quadratic weighted κ of 92% (95% confidence interval, 89%–92%) and an overall accuracy of 63.3%. In the independent KORA dataset, images wrongly classified as AMD were mainly the result of a macular reflex observed in young individuals. By restricting the KORA analysis to individuals >55 years of age and prior exclusion of other retinopathies, the weighted and unweighted κ increased to 50% and 63%, respectively. Importantly, the algorithm detected 84.2% of all fundus images with definite signs of early or late AMD. Overall, 94.3% of healthy fundus images were classified correctly.

Conclusions

Our deep learning algoritm revealed a weighted κ outperforming human graders in the AREDS study and is suitable to classify AMD fundus images in other datasets using individuals >55 years of age.

中文翻译：

---

一种深度学习算法，用于预测与年龄有关的黄斑变性从彩色眼底摄影中得出的与年龄有关的眼疾研究严重程度量表

目的

年龄相关性黄斑变性（AMD）是视力的普遍威胁。虽然疾病阶段的分类对于了解疾病的风险和进展至关重要，但已知一些基于彩色眼底照片的系统。其中大多数都需要对眼底图像进行深入且费时的分析。在这里，我们提出了一种基于计算机的自动分类算法。

设计

基于大量彩色眼底图像的AMD分类算法开发。验证是在基于人群的横断面研究中进行的。

参加者

我们纳入了3654个与年龄有关的眼病研究（AREDS）参与者的120 656个手动分级的彩色眼底图像。AREDS参与者的年龄大于55岁，并且在招募时排除了非AMD威胁视力的疾病。此外，我们的算法的性能在来自奥格斯堡地区基于人口的Kooperative Gesundheitsforschung（KORA；奥格斯堡地区合作卫生研究）研究的5555个眼底图像中进行了评估。

方法

我们定义了13个类（9个AREDS步骤，3个AMD后期阶段以及1个用于不可分级图像的阶段），并训练了几种卷积深度学习架构。一组网络体系结构提高了预测准确性。在基于人口的研究中，使用了独立的数据集来评估我们算法的性能。

主要观察指标

κ统计和准确性，以评估预测的和专业的人类评分者分类之间的一致性。

结果

由6种不同的神经网络体系结构组成的网络集合预测AREDS测试集中的13个类别，其二次加权κ为92％（95％置信区间，89％–92％），总准确度为63.3％。在独立的KORA数据集中，被错误分类为AMD的图像主要是在年轻人中观察到的黄斑反射的结果。通过将KORA分析限制在55岁以上的个体以及先前排除其他视网膜病变的个体，加权和未加权κ分别增加到50％和63％。重要的是，该算法可检测到所有眼底图像中84.2％的早期或晚期AMD迹象明确。总体而言，正确分类了94.3％的健康眼底图像。

结论

我们的深度学习算法在AREDS研究中揭示了加权κ优于人类评分者，适用于使用超过55岁的个体对其他数据集中的AMD眼底图像进行分类。