AI-Driven Diabetic Retinopathy Diagnosis Enhancement through Image Processing and Salp Swarm Algorithm-Optimized Ensemble Network

Saif Ur Rehman Khan, Muhammad Nabeel Asim, Sebastian J. Vollmer, Andreas DengelarXiv 2025 · DOI · arXiv

Introduction

Diabetic Retinopathy (DR) is a leading cause of vision impairment and blindness globally. Early detection is vital for effective treatment, yet the current gold standard—manual evaluation of retinal fundus images by ophthalmologists—is labor-intensive, costly, and subject to human error. While automated deep learning systems have emerged as a solution, they often struggle with the inherent complexities of retinal images, such as low contrast, noise, and the subtle nature of early-stage pathological features. Furthermore, standard ensemble methods typically rely on fixed weighting, which may not optimally leverage the strengths of individual neural architectures.

Methodology

The proposed framework addresses these challenges through a four-phase pipeline designed to enhance feature extraction and classification accuracy.

Image Pre-processing and Fusion

To improve the quality of input data, the researchers implemented a multi-stage preprocessing strategy. Contrast Limited Adaptive Histogram Equalization (CLAHE) was applied to enhance local contrast, while Gamma correction normalized image brightness. These two versions of the image were then integrated using Discrete Wavelet Transform (DWT) fusion. This process creates a "richer" composite image that preserves multi-resolution details from both enhancement techniques, providing a more informative input for the neural networks.

Overview of the image fusion framework using CLAHE, Gamma correction, and DWT.

Model Architecture and Improved Residual Blocks

The ensemble utilizes three diverse backbone architectures: DenseNet169, MobileNetV1, and Xception. To improve gradient flow and prevent the "dead neuron" problem associated with standard ReLU activations, the authors integrated an improved residual block into these backbones. By replacing ReLU with LeakyReLU, the network maintains a small, non-zero gradient for inactive units, which is critical for learning the intricate patterns necessary to distinguish between subtle DR stages.

Comparison between the original residual block and the improved block using LeakyReLU.

Salp Swarm Optimization (SSA)

A key innovation of this work is the use of the Salp Swarm Algorithm (SSA) to optimize the ensemble. SSA is a bio-inspired metaheuristic that mimics the foraging behavior of salp chains. In this framework, SSA is employed to dynamically search for the optimal weights assigned to each of the three backbone models. This ensures that the final prediction is weighted toward the model most capable of identifying specific features, rather than relying on a simple average.

Experimental Results

The framework was primarily evaluated on the Kaggle APTOS 2019 dataset (5,590 images) and further validated on the DDR dataset (13,673 images).

The SSA-optimized ensemble achieved a state-of-the-art accuracy of 89.07% on the APTOS 2019 dataset, significantly outperforming individual models like DenseNet169 (86.06%) and Xception (85.24%). On the larger DDR dataset, the model demonstrated strong generalizability with 84.35% accuracy. Statistical validation via McNemar’s test confirmed that these improvements were statistically significant (p < 0.05). Furthermore, the model achieved an Area Under the Curve (AUC) of 0.9673, reflecting high sensitivity and specificity across all five stages of DR.

Conclusion and Future Work

This research demonstrates that combining advanced image fusion with metaheuristic-optimized ensembles can significantly enhance the diagnostic accuracy of Diabetic Retinopathy. While the current model is computationally intensive, future efforts will focus on model pruning to facilitate deployment in real-time mobile applications for low-resource clinical settings. Additionally, the team plans to explore unsupervised learning techniques to further refine feature extraction capabilities.

Summary updated

BibTeX
@misc{khan2025ai,
  author = {Saif Ur Rehman Khan and Muhammad Nabeel Asim and Sebastian J. Vollmer and Andreas Dengel},
  title = {{AI-Driven Diabetic Retinopathy Diagnosis Enhancement through Image Processing and Salp Swarm Algorithm-Optimized Ensemble Network}},
  year = {2025},
  eprint = {2503.14209},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2503.14209},
}
Figure