Adaptive Multi-Scale Contrast Enhancement and Morphological Flow Integration for Diabetic Retinopathy Detection Using ELM-Based Classifier

Abstract

Introduction: Diabetic retinopathy affects 100 million individuals worldwide and represents a leading preventable cause of vision loss. Automated screening systems demonstrate suboptimal performance due to heterogeneous imaging conditions and insufficient preprocessing strategies. This study aimed to develop an integrated artificial intelligence pipeline that combines adaptive preprocessing, morphological feature extraction, and optimized classification methods for robust diabetic retinopathy severity assessment. Methods: The research employed the preprocessed "Diabetic Retinopathy Arranged" dataset from Kaggle platform containing 34,860 fundus images across five severity grades. Three methodological innovations were implemented: Adaptive Multi-Scale Contrast Limited Adaptive Histogram Equalization (AMS-CLAHE) for content-aware preprocessing, Morphological Transition Flow architecture for structural change modeling, and Bayesian optimization for Extreme Learning Machine variants. Comprehensive ablation studies evaluated preprocessing configurations, architectural components, and classification strategies through systematic parameter optimization. Results: The study proposes an AMS-CLAHE framework with dynamic threshold calibration and entropy-based attention mechanisms for content-aware preprocessing, achieving F1-score of 0.908 and AUC-ROC of 0.986 with processing efficiency below 250ms per image. The All-ELM configuration demonstrated superior performance (F1=0.970, accuracy=0.970) compared to conventional architectures. LAB color space processing outperformed RGB representation. Bayesian-optimized Random Forest delivered optimal classification results (F1=0.997, MCC=0.996) across all severity grades. Conclusions: The integrated pipeline demonstrated that systematic preprocessing optimization enables competitive diagnostic accuracy while maintaining computational efficiency. This approach facilitates scalable diabetic retinopathy screening implementation in diverse clinical environments where expert assessment remains limited.