Improved Deep Learning Framework with hybrid feature selection model for Exact Identification and Categorization of Multiple Eye Diseases using Retinal Images

Abstract

Nowadays, different models are developed to detect and classify eye disease using Deep Learning (DL) but still issues of misclassification, complexity, and error. Also, multiple eye disease detection is challenging because of complex data. So, in this paper, design a comprehensive DL framework to detect and classify multiple eye diseases using retinal images. Initially, multiple eye disease images are collected and trained in the system. Image augmentation is performed using geometric transformation to enhance the dataset. The preprocessing technique used the Isolation Forest Algorithm (IFA) for outlier detection, BM3D for image denoising, CLAHE for contrast enhancement and Macenko's method for color normalization. To enhance the performance of ROI identification, propose a Loss and Attention-Augmented Multi-Scale U-Net++ (LAAMS-UNet++) architecture. Then extract retinal vessel features, optic disc features, texture features, and deep features in the feature engineering phase. Moreover, a Hybrid Seahorse and Beluga Whale (HSHBW) optimization is employed for the feature selection phase to enhance the performance of classification results. Finally, classification is performed using a Dense-CSPDarkNet53 + LinkNet-34 model with an EfficientNetB7 encoder. The developed model accurately detects and classifies multiple eye diseases with high accuracy. The attained results are authorized with existing techniques regarding precision, accuracy, specificity, sensitivity, etc. The model performs high accuracy of 98.521%, precision of 98.812%, sensitivity (97.102%), and specificity (97.553%) demonstrating its reliable performance over a variety of datasets.