Explainable deep learning-based meta-classifier approach for multi-label classification of retinal diseases

Early diagnosis of retinal diseases is important to prevent vision loss. This study introduces a novel multi-label classification system for detecting multiple retinal diseases using two publicly available datasets. The process begins with data collection and preprocessing, including image resizing and noise filtering to enable effective feature extraction. To develop and train the models, we apply a transfer learning approach to several state-of-the-art deep learning models, including MobileNetV2, InceptionV3, NASNetMobile, DenseNet169, EfficientNetB4, DenseNet121, ConvNeXt, and Xception. The two best-performing models were selected based on the validation results and were used as base models, which are subsequently combined using a meta-classifier. The experimental results demonstrate that the proposed model achieved an impressive performance, with 0.981 accuracy, 0.982 precision, 0.981 sensitivity, 0.981 F1 score and 0.994 specificity in the Eye Diseases Classification dataset and 0.977 accuracy, 0.978 precision, 0.977 sensitivity, 0.977 F1 score, and 0.978 specificity on the Retinal Fundus Images dataset. These results highlight the model’s high accuracy, reliability, and robustness, with statistically significant improvements validated by a paired t-test, outperforming state-of-the-art methods in retinal disease classification. Given the importance of model interpretability, especially in the healthcare field, this study utilizes Local Interpretable Model Agnostic Explanation to visually evaluate the model predictions using superpixels. This approach enhances transparency and trust in the model’s decision-making process. With excellent accuracy, statistical robustness, and interpretability, the proposed system assists medical practitioners in the early diagnosis of retinal diseases and contributes to improved patient care outcomes through the advancement of automated diagnostic systems in ophthalmology.