Leveraging CNN and Transfer Learning With EfficientNet for Enhanced Optical Coherence Tomography (OCT)
TL;DR
Overall, these results suggest that compact EfficientNet feature extractors integrated with SVM offer accurate, stable, and computationally practical OCT‐driven retinal disease classification for clinical decision‐support settings.
Abstract
Optical coherence tomography (OCT) is an important imaging modality for detecting retinal disorders, including diabetic macular edema (DME), choroidal neovascularization (CNV), and drusen, as well as distinguishing healthy retinal patterns. Nevertheless, visual assessment of OCT scans is labor‐intensive, observer‐dependent, and dependent on expert clinical knowledge, encouraging reliable automated screening tools. This work introduces a hybrid Deep Transferred EfficientNet and Support Vector Machine (DTE‐SVM) approach for multiclass OCT image classification. In contrast to standard CNN models that commonly use a softmax layer for final prediction, the proposed framework employs transfer‐learned EfficientNet representations with an SVM decision model to strengthen interclass discrimination and robustness within the extracted feature space. Experiments used a working subset of 8767 OCT images selected from the OCT2017 dataset, representing nearly 10.4% of the complete 84,495‐image collection. This smaller yet representative subset was adopted to lower computational cost while maintaining class distribution among CNV, DME, drusen, and normal groups. Five EfficientNet architectures, from B0 through B4, were assessed using stratified 10‐fold cross‐validation. Among them, DTE‐SVM‐B0 produced the strongest overall results, obtaining an accuracy of 0.9076, precision of 0.9074, recall of 0.9076, and F1‐score of 0.9073. DTE‐SVM‐B2 showed the smallest fold‐level standard deviation, reflecting comparatively stable behavior. Comparative experiments indicated that DTE‐SVM‐B0 surpassed ResNet50, DenseNet‐121, and EfficientNetB0 softmax‐based classifiers, whereas ablation testing verified that substituting the softmax layer with SVM improved results for every EfficientNet backbone. Class‐level clinical evaluation also revealed strong sensitivity and specificity, with CNV identified most accurately and drusen remaining the hardest category. Overall, these results suggest that compact EfficientNet feature extractors integrated with SVM offer accurate, stable, and computationally practical OCT‐driven retinal disease classification for clinical decision‐support settings.