A Kernel PCA-Based Ensemble Deep Learning Approach for Foveal Avascular Zone Classification

Accurate classification of foveal avascular zone (FAZ) from optical coherence tomography angiography (OCTA) images is critical for the early detection and management of retinal pathologies such as diabetic retinopathy and myopia. In this study, we propose a Kernel Principal Component Analysis (KPCA) deep ensemble approach for classifying FAZ in OCTA images. Our framework first extracts the regions of interests (ROIs) and then generates deep features from multiple pre-trained models (ResNet50, VGG16, EfficientNetB0 and DenseNet201) to capture diverse, high-level representations of FAZ regions. Then, KPCA is used to fuse these features into a compact, non-linear representation. On top of this fused feature set, is build a deep classifier, a fully connected feedforward neural network (FFN), to differentiate among normal, diabetic and myopic conditions. Experimental results on the FAZID dataset demonstrate outstanding performance, with a training accuracy of 98.58% and a test accuracy of 98.36%. These results highlight the effectiveness of combining Kernel PCA with ensemble learning in capturing subtle yet clinically significant variations within the FAZ region. The proposed approach not only enhances classification accuracy and generalization but also supports more reliable and automated clinical decision support in retinal diagnostics.