DEO-Fusion: Differential evolution optimization for fusion of CNN models in eye disease detection

Abstract

Eye diseases pose a significant health concern globally, emphasizing the need for accurate and efficient diagnostic methods. The manual recognition of eye disorders is both time-consuming and challenging. Deep learning (DL) techniques have demonstrated their effectiveness in the analysis of medical images, underscoring their capability to improve the identification and categorization of eye-related conditions. This study introduces DEO-Fusion, a pioneering approach aimed at enhancing the accuracy of eye disease detection through a Weighted Averaging Ensemble (WEAE) technique. In contrast to previous research focusing on individual models, our work delves into the largely unexplored potential of ensemble learning. Initially, Transfer Learning (TL) is employed with four base models, bolstering their image representation capabilities via additional layers. The WEAE scheme combines their outputs, and novel weight allocation is achieved through an Evolutionary Algorithm-based Differential Evolution Optimization (DEO) approach. In contrast to the commonly employed experimental weight assignments in the literature, DEO optimally allocates weights to each model, leading to a substantial improvement in performance. The comparison with other optimization algorithms was also conducted to evaluate the performance and effectiveness of the DEO algorithm in weight optimization for ensemble model, providing a comprehensive assessment of its capabilities in the context of eye disease detection. The proposed approach underwent evaluation using two publicly available datasets—one comprising digital camera images with cataract and normal classes, and the other containing fundus images with four classes (cataract, glaucoma, diabetic retinopathy, and normal). The method attained impressive accuracy rates of 98.34 % and 94.92 % on the digital camera images dataset and retinal fundus images datasets, respectively. These results underscore the superior performance of DEO-Fusion compared to existing methods and widely employed ensemble techniques. Grad-CAM analyses were conducted to elucidate infected areas in the eye, providing clinicians with valuable insights for prompt and accurate diagnoses of eye diseases.