Pre-screening of healthy retinal using geometrical features extracted from OCT retinal images

Optical Coherence Tomography (OCT) retinal imaging is essential in ophthalmology for the diagnosis of eye diseases. To date, identifying healthy OCT retinal images can be challenging due to the large volume of images but limited time for ophthalmologists. In this study, we developed a classification technique that differentiates between healthy and unhealthy OCT images of the retina. We applied the Standard Deviation Profiling (SDP) method that can accurately detect the inner limiting membrane (ILM) and retinal pigment epithelium (RPE), determining the region of interest for observing the structural information. First of all, the outer nuclear layer (ONL) was detected using k-means clustering segmentation, applied to the region between the ILM and RPE layers. Then, 24 geometrical features involving area-based, distance-based, intensity-based, and symmetry-based features, so-called “feature sets,” were extracted with the assistance of the three layers. A number of machine learning classifiers were built using such 24 geometrical features with a train:test data splitting of 80:20 among the dataset containing 80 OCT images. The results show that the extracted features quantitatively represented the structural conditions of the retina using geometrical relevance. We achieved 100% classification accuracy in various models, including fine tree, linear discriminant, Naïve Bayes networks, linear support vector machine, and ensemble subspace discriminant. Whilst models trained using each feature set on its own only yield above 80% accuracy using these five classifiers, with area-based and distance-based features presenting relatively higher accuracy. The results of this study emphasized the novel contribution of extracting the 24 geometrical features, which successfully trained machine learning classifiers to categorize healthy versus diseased retinal OCT images.

光学相干断层扫描（OCT）视网膜成像是眼科学领域用于眼部疾病诊断的核心技术。迄今为止，由于图像数据量庞大而眼科医生的接诊时间有限，甄别健康OCT视网膜图像仍颇具挑战。本研究开发了一种分类技术，用于区分健康与病变的视网膜OCT图像。我们采用了标准差剖面分析法（SDP），可精准检测内界膜（ILM）与视网膜色素上皮（RPE），以此确定用于观察视网膜结构信息的感兴趣区域。首先，针对内界膜与视网膜色素上皮之间的区域，通过k均值聚类分割检测外核层（ONL）。随后，依托这三层结构，提取了涵盖基于面积、基于距离、基于强度及基于对称性的共24项几何特征，即所谓的“特征集”。本数据集包含80张OCT图像，我们以80:20的比例划分训练集与测试集，基于上述24项几何特征构建了多款机器学习分类器。结果表明，所提取的特征可通过几何相关性定量表征视网膜的结构状态。我们在多种模型中均实现了100%的分类准确率，包括精细树分类器、线性判别分类器、朴素贝叶斯网络、线性支持向量机以及集成子空间判别分类器。而仅使用单一特征集训练的模型，在这五类分类器上的准确率仅能达到80%以上，其中基于面积和基于距离的特征展现出相对更高的准确率。本研究结果凸显了提取24项几何特征的创新性贡献，该方法可成功训练机器学习分类器以区分健康与病变的视网膜OCT图像。