LeNet-TSA: A Robust Framework for Automated Detection of Hard Exudates in Retinal Images

Diabetic retinopathy (DR) is a diabetes-related eye issue that occurs when the blood vessels in the retina are damaged. Bright lesions known as hard exudates (HEs) appear in the retinal fundus images of DR patients. Therefore, HE identification is a crucial task in preventing vision loss. However, HEs are difficult to detect because of their distinct appearance characteristics. It takes a lot of time and effort to manually detect HEs in order to diagnose many diabetic patients. An innovative framework, the LeNet-based tunicate swarm optimization algorithm (LeNet-TSA), is proposed in this research as an emerging mechanism for the robust detection of HEs in retinal images using an optic disk (OD) elimination operation. Various pre-processing strategies are applied to enhance image quality and mitigate distortions, along with data augmentation mechanisms to address overfitting difficulties. To find the best thresholds, Kapur’s thresholding approach is implemented for optimal OD segmentation, followed by an erosion operation that eliminates the OD in retinal images. Afterward, a modified convolutional neural network (CNN) strategy is employed for essential feature extraction, which reduces processing time and enhances detection accuracy. Additionally, the LeNet-5 approach identifies HEs in extracted retinal characteristics, and the TSA algorithm fine-tunes the parameters to enhance generalization, processing speed, and accuracy in HE diagnosis. The efficacy of the LeNet-TSA framework is evaluated on five retinal image datasets. The LeNet-TSA framework diagnoses HEs with simple calculations and a greater accuracy of 99.08% compared to other HE diagnosis strategies.