基于OCT图像视网膜关键层多特征融合的玻璃膜疣轻量化识别方法

DRUSEN Recognition Based on Random Forest and Grayscale Features of OCT Image Layer Regions

摘要: 玻璃膜疣(DRUSEN)是早期年龄相关性黄斑变性(AMD)的典型病理特征，其精准识别对基层医疗AMD筛查至关重要。针对现有OCT图像DRUSEN识别方法技术复杂、部署成本高的问题，本研究提出多特征融合的轻量化集成学习识别方案。以Kermany公开OCT数据集的500例样本(DRUSEN与正常样本各250例)为对象，经U-net网络分割视网膜9层结构后，聚焦ONL、MEZ、RPE三大病理关键层，提取灰度统计、形态学、高阶纹理共63维特征构建特征集，以LightGBM、XGBoost为核心模型开展实验验证。结果表明，LightGBM模型性能最优，测试集准确率88.0%、AUC 0.925、DRUSEN召回率89.3%；形态学特征贡献度达42.3%为核心特征，且模型抗高斯噪声干扰能力良好。该方法轻量化、部署门槛低、可解释性强，非线性特征捕捉能力优于传统单一机器学习模型，为基层医疗DRUSEN快速筛查提供了实用技术支撑。

Abstract: Drusen is a typical pathological feature of early age-related macular degeneration (AMD), and its accurate identification is crucial for AMD screening in primary healthcare. Addressing the issues of complex technology and high deployment costs in existing OCT image drusen identification methods, this study proposes a lightweight ensemble learning scheme based on multi-feature fusion. Using 500 samples (250 drusen and 250 normal samples) from the publicly available OCT dataset by Kermany et al., the retinal structure was segmented into nine layers using a U-net network. Focusing on the three key pathological layers—ONL, MEZ, and RPE—63-dimensional features (including grayscale statistics, morphology, and higher-order texture) were extracted to construct a feature set. Experimental validation was conducted using LightGBM and XGBoost as the core models. Results show that the LightGBM model exhibits the best performance, with a test set accuracy of 88.0%, AUC of 0.925, and a drusen recall of 89.3%. Morphological features contributed 42.3% to the core features, and the model demonstrated good resistance to Gaussian noise. This method is lightweight, has low deployment threshold, and is highly interpretable. Its nonlinear feature capture capability is superior to traditional single machine learning models, providing practical technical support for rapid screening of DRUSEN in primary healthcare.