Diabetic retinopathy detection using a well-calibrated uncertainty aware convolutional neural network

The early detection of diabetic retinopathy is crucial in preventing irreversible vision loss, making it a critical concern in healthcare. While deep learning models have shown advancements in categorization tasks, they can sometimes exhibit overconfidence in incorrect outcomes, posing significant risks in the medical field. To tackle this challenge, a novel model is proposed, which leverages a CNN (Convolutional neural network) architecture to predict softmax probabilities. The model incorporates train time augmentation to enhance performance and subsequently calibrates the projected probabilities using multiple calibration techniques, including Platt scaling, temperature scaling, isotonic regression, histogram binning, and Platt histogram. Furthermore, Test Time Data Augmentation (TTAUG) is employed to improve prediction accuracy by presenting the same data in multiple variations. The proposed technique is rigorously evaluated on benchmark datasets, namely Kaggle and Messidor. Notably, TTAUG yields remarkable results, achieving 98% accuracy on the Kaggle dataset, a significant improvement compared to the 93% achieved without augmentation. The calibration of probabilities further enhances the performance, as indicated by the reduction of Expected Calibration Error (ECE) from 0.52 to an impressive 0.019. The results underscore the significance of quantifying uncertainty in predictions, which enhances the interpretability of the proposed method as a diagnostic aid tool. By effectively addressing overconfidence and providing calibrated probabilities, the model proves its potential in improving both accuracy and reliability in diabetic retinopathy detection.