Content area
The rising prevalence of diabetes has made Diabetic Retinopathy (DR) a major cause of blindness, underscoring the necessity for a computer-aided diagnostic system that can support clinical diagnoses without requiring extensive human effort. Many researchers have turned to deep learning to create automated screening and diagnostic tools for DR. However, for such systems to be truly effective in clinical practice, they must provide highly accurate assessments and well-calibrated estimates of uncertainty. Unfortunately, deep neural networks often tend to be overconfident in their predictions and are not easily amenable to probabilistic approaches. In our study, we introduce a novel approach for evaluating diagnostic uncertainty in DR predictions by employing ensemble-based calibration techniques. What sets our approach apart from cutting-edge convolutional neural network models is our use of the EfficientNet architecture, which offers superior accuracy through transfer learning. We then apply a set of post-calibration techniques to transform the model's probabilistic output into a confidence level. To gauge the uncertainty of our forecasts, we compute the entropy of the calibrated confidence value. This approach greatly assists users in determining whether it is necessary to seek a second opinion. Our model achieves an impressive accuracy score of 96 %, and our ensemble technique exhibits a notable reduction in Expected Calibration Error (ECE), in addition to providing a reassuring uncertainty score.