Abstract
The diagnosis of retinal diseases traditionally requires significant clinical expertise, creating a bottleneck in healthcare delivery, particularly in underserved regions. This paper introduces a hybrid machine learning framework that integrates deep learning feature extraction with classical statistical classification. Utilizing a modified U-Net for vessel segmentation, Principal Component Analysis (PCA) for dimensionality reduction, and Support Vector Machines (SVM) for final classification, the proposed model achieves a peak accuracy of 89.73% across 32 disease categories. The research also introduces EyeNet, a comprehensive clinical dataset, and demonstrates the efficacy of hybrid voting strategies in improving diagnostic reliability.
Introduction
Retinal disease identification is a high-stakes visual task where accuracy is often tied to an ophthalmologist's years of experience. In rural or underserved areas, the scarcity of specialists leads to delayed diagnoses. While automated disease detection (ADD) systems exist, they frequently focus on single conditions or require massive labeled datasets. This work addresses these challenges by developing a robust system capable of identifying 32 distinct retinal diseases with high accuracy and reduced computational demands.
Methods
The core of the proposed system is a multi-stage hybrid pipeline that blends deep learning with classical statistical methods to maximize diagnostic performance.
- Feature Extraction (U-Net): A modified U-Net architecture is employed to segment retinal blood vessels. This process isolates "semantic features"—morphological structures that carry significant diagnostic weight—from noisy raw images.
- Dimension Reduction (PCA): High-dimensional features from the U-Net are compressed using PCA to prevent overfitting and reduce computational complexity. The authors identified the 62nd principal component as providing optimal performance.
- Classification (SVM): Instead of a standard neural network output layer, the compressed features are fed into an SVM to perform the final classification.
- Hybrid Voting: The model employs a "Hybrid-Ratio" approach, using weighted voting between two different SVMs: one trained on raw RGB images and one trained on the U-Net's vessel-segmented features.
Datasets
The study introduces EyeNet, a clinical collection derived from the Retina Image Bank (RIB). EyeNet encompasses 32 disease classes, including rare conditions such as Acute Macular Neuroretinopathy (AMN). Preprocessing involves generating a subset of images focused specifically on blood vessel morphology via the U-Net extractor to highlight critical disease markers.
Key Results
The hybrid model achieved a peak diagnostic accuracy of 89.73%, significantly outperforming pure-CNN baselines such as AlexNet (79.03%) and SqueezeNet (82.26%). The modified U-Net demonstrated exceptional robustness, achieving 95.27% accuracy in vessel segmentation, proving its utility as a feature extractor.
As shown in the performance table above, the optimal configuration utilized a 47%:53% hybrid ratio with an RBF kernel. Furthermore, the study confirmed the value of transfer learning; models initialized with ImageNet weights showed a 5% to 15% accuracy boost over random initialization, suggesting that low-level visual features are highly transferable to the medical domain.
Conclusion
By combining the vision capabilities of CNNs with the classification strengths of SVMs, this hybrid approach offers a scalable and efficient tool for clinical decision support. The reduction in computing complexity via PCA makes the model particularly suitable for point-of-care medical devices. Future work aims to integrate Visual Question Answering (VQA) to provide natural language explanations for diagnostic outputs, further bridging the gap between automated systems and clinical practice.
Summary updated