Comparative analysis of novel preprocessing techniques and deep learning based multi modal feature fusion for diabetic retinopathy grading

Abstract

Abstract Fundus images are crucial for the detection and monitoring of retinal diseases such as diabetic retinopathy (DR). However, issues such as uneven illumination, low contrast, and noise often degrade image quality, impacting the accuracy of automated grading systems. This study introduces three novel preprocessing techniques Adaptive Sigmoid Enhancement, LAB-ACE Image Enhancement, and Multi-channel Image Enhancement designed to address these challenges. Adaptive Sigmoid Enhancement adaptively adjusts local contrast to highlight subtle lesions, LAB-ACE operates in the LAB color space to selectively enhance the lightness channel while preserving color fidelity, and Multi-channel Image Enhancement applies targeted green-channel optimization combined with contrast stretching and channel recombination. These methods extend beyond conventional contrast enhancement and normalization by integrating multi-stage adaptive processing and color-channel-specific optimization to improve lesion visibility and vessel delineation while minimizing background noise. Following pre-processing, handcrafted features (LBP, GLCM) and deep features from a pre-trained ResNet-50 are fused in a multi-modal framework and evaluated using multiple classifiers, including SVM, KNN, Random Forest, and XGBoost. Results demonstrate that XGBoost with fused features and Adaptive Sigmoid Enhancement achieves the highest accuracy (96.39%), outperforming other combinations. The findings highlight the effectiveness of the proposed pre-processing strategies in enhancing DR grading performance, paving the way for improved computer-aided diagnosis systems.