An Explainable Ensemble Based Approach to Diabetic Retinopathy Grading

Diabetic retinopathy is a dangerous eye anomaly that may cause vision loss and blindness in people with diabetes (more than 500 million adults in 2021). This study presents a novel approach to diabetic retinopathy grading using explainable deep learning techniques to create a diagnosis aid tool that provides a fully detailed final report with alternative explanations for the result. This is clear novelty over existing solutions. The proposed approach is based on an ensemble of two very efficient deep learning networks: efficientNetV2 and ConvNeXt. This deep learning model is evaluated using a public available dataset and compared with the results obtained from previous works. Also, explainable deep learning techniques are applied to present the final report. The architecture is capable of achieving state-of-the-art performance for both the two-class problem and the five-class international clinical diabetes retinopathy (ICDR) classification problem. This work achieves a 96.7% accuracy and an AUC over 96% for all classes. The ensemble provides explainability records which are significantly improved when compared with those obtained from a single network. The ensemble architecture provides good quality explanations to the ophthalmologist with several configurable superimposed heat maps and two probabilityordered diagnostic suggestions, including a quality factor indicating the estimation of the probability ratio between the alternatives. This makes it a valuable diagnostic assistance tool for ophthalmologists and other healthcare professionals. This study introduces a novel approach to diabetic retinopathy (DR) grading, utilizing explainable deep learning techniquesto create a diagnostic aid tool. The system provides a detailed final report with alternative explanations, which is a clear advancement over existing solutions. Our method is based on an ensemble of two highly efficient deep learning networks: EfficientNetV2 and ConvNeXt. This ensemble approach achieves a 96.7% accuracy and an AUC over 96% for the ICDR classification problem, outperforming single-network models in both accuracy and explainability. The system offers ophthalmologists configurable heat maps and probability-ordered diagnostic suggestions, making it a valuable diagnostic tool.