Segmented Retinal Fundus Image Classification Approach with Multiple Instance Learning

Diabetic Retinopathy (DR) is a leading cause of vision loss, and its early detection is essential for preventing unchanging damage. This study proposes an efficient classification framework using segmented retinal fundus images from the MESSIDOR dataset, focusing on vessel-level segmentation to enhance disease-specific feature extraction and minimize extraneous background information. The pipeline includes preprocessing steps such as image resizing, green channel extraction, binary masking, and Contrast Limited Adaptive Histogram Equalization (CLAHE), followed by segmentation of both thick and thin vessels using traditional image processing techniques. Segmented features are classified within a Multiple Instance Learning (MIL) framework, evaluating MIL with Lagrangian Relaxation (MIL-RL), Semiproximal SVM (mi-SPSVM), Kernel-based Semiproximal SVM (mi-KSPSVM), and conventional SVM (linear and RBF kernels). Performance is evaluated using Testing Correctness (TC), Sensitivity, Specificity, F1-score, and computational time under 5-fold, 10-fold, and Leave-One-Out (LOO) cross validation (CV). Results show that preprocessing significantly improves accuracy across all models, with mi-SPSVM under LOO-CV achieving the best at 68.00% in TC and a 68.32% in F1-score. At the same time, mi-KSPSVM achieved the highest sensitivity of 84.00%. The findings reveal that a segmentation-first approach in the classification task can enhance both diagnostic accuracy and computational efficiency, making it suitable for real-time and resource-limited environments. This approach offers a worthwhile, scalable solution for automated DR screening, with future work aimed at integrating deep learning- based segmentation to further improve precision.