基于深度学习的视网膜OCT图像的疾病自动分类

Automated disease classification in retinal OCT images using deep learning

摘要:

目的 实现视网膜疾病光学相干断层扫描图像（OCT）的自动化分类。方法 整合公开数据集和临 床医院收集的视网膜OCT图像，构建包含3类视网膜疾病的标准化数据集：正常（Normal）、息肉样脉络膜血管病变 （PCV）和糖尿病黄斑水肿（DME）。对图像进行系统性预处理，包括数据规范化、分类标签筛查和数据增强，随后系 统评估并比较了6种主流卷积神经网络（VGG、ResNet、DenseNet、MobileNet 、EfficientNet和Inception）的分类性能。 结果 VGG16模型实现了98.92%的准确率；ResNet50和ResNet50_V2均实现了99.79%的准确率，但ResNet50的精度 和召回率均稍微优于 ResNet50_V2；Densenet121 的准确率为 99.94%，精度为 99.90%，召回率为 99.95%，F1 分数为 0.999 2，均 高 于 其 他 经 典 卷 积 神 经 网 络 模 型 ；EfficientNetB0、EfficientNetV2S、MobileNetV1、MobileNetV2、 MobileNetV3 和 Inception_V3 的 准 确 率 分 别 为 99.75%、99.83%、99.72%、99.85%、99.43% 和 99.35%。 结 论 Densenet121模型在视网膜OCT图像自动化分类中性能最优异，实现了对视网膜OCT图像的高效准确诊断。

Abstract:

Objective To develop an automated classification system for retinal diseases using optical coherence tomography (OCT) images.Methods A standardized dataset was constructed from both public repositories and clinical hospitals, comprising OCT images categorized into three classes: Normal, Polypoidal Choroidal Vasculopathy (PCV), and Diabetic Macular Edema (DME). Systematic preprocessing steps including data normalization, label verification, and augmentation were applied. Six convolutional neural network (CNN) architectures (VGG, ResNet, DenseNet, MobileNet, EfficientNet, Inception) were rigorously evaluated for classification performance. Results The VGG16 model achieved an accuracy of 98.92%, while ResNet50 and ResNet50_V2 attained 99.79%. Notably, ResNet50 exhibited marginally superior precision and recall compared to ResNet50_V2. DenseNet121 outperformed all other models, achieving an accuracy of 99.94%, precision of 99.90%, recall of 99.95%, and an F1-score of 0.999 2. Among EfficientNet variants, EfficientNetB0 and EfficientNetV2S reached accuracies of 99.75% and 99.83%, respectively. MobileNetV2 achieved the highest accuracy (99.85%) within its variants, followed by MobileNetV1 (99.72%) and MobileNetV3 (99.43%). Inception_V3 yielded an accuracy of 99.35%.Conclusion DenseNet121 demonstrates optimal performance for automated retinal OCT image classification, offering high accuracy and diagnostic efficiency for clinical applications.