Eye diseases are a major global cause of blindness and vision impairment, highlighting the vital need for accurate and early diagnosis to prevent further deterioration. Despite advancements in medical imaging, the retinal diseases classification using fundus images remains challenging due to the complexity and subtlety of visual features. This study aims to develop and validate EDDNet30, a novel 30-layer deep learning model, to improve the classification of eye diseases from fundus images. The model incorporates advanced architectural features, such as spatial attention and multi-scale fusion modules, to enhance diagnostic accuracy and robustness. To validate the proposed model, a diverse dataset of 5531 images comprising 9 different disease categories is collected from various sources has been utilized. To enhance image quality, several pre-processing approaches were applied, including histogram equalization, color space conversion and contrast adjustment, ensuring high-resolution final images. Furthermore, image data augmentation was employed to increase the dataset’s count, enhancing the model’s ability to generalize during training. The spatial attention module helps the model to concentrate on the most relevant regions of the images, while multi-scale fusion modules capture and integrate details at different scales, significantly improving classification performance. To evaluate its effectiveness, the model was compared to various transfer learning models. The results show that EDDNet30 consistently outperforms transfer-learning models with 95.29% accuracy tested on 10% split test data, demonstrating superior accuracy and robustness in eye disease classification. Furthermore, model interpretability was enhanced by employing various explainable AI methods, such as Grad-CAM, Grad-CAM++ and LIME to identify and emphasize key features influencing the decision-making process. EDDNet30 represents a promising advancement in automated ocular disease detection, offering improved diagnostic reliability for clinical use.
Nirob et al. (Thu,) studied this question.