Deep transfer learning and contour-based morphological analysis for detection of eye-hypertensive diseases from fundus images

Hypertensive eye disease produces subtle retinal changes that are often challenging for automated systems to recognize reliably. This paper proposes a hybrid diagnostic framework that combines contour-based morphological descriptors with deep transfer learning to classify hypertensive versus non-hypertensive fundus images from a binarized version of the ODIR dataset. Images were preprocessed using Gaussian filtering (σ = 1.0) and normalized prior to feature extraction. Various CNN architectures were evaluated, and all performance metrics were computed across five independent training seeds and reported as mean ± 95% confidence intervals. EfficientNetB4 achieved the strongest validation performance 96.61%, outperforming EfficientNetB2 and InceptionResNetV2, while maintaining the lowest loss and RMSE. Incorporating contour-derived geometric and intensity descriptors improved precision and F1-scores by approximately 1–2% across models, demonstrating that structural vessel cues provide complementary discriminative value beyond deep features alone. Calibration analysis further showed that EfficientNetB4 yielded the most reliable probability estimates, with the lowest ECE and Brier scores. External validation on unseen fundus images confirmed the model’s generalization potential. These results indicate that integrating morphological analysis with transfer learning enhances both performance and interpretability for hypertensive retinal change detection.