DR-TrustNet: Enhancing diabetic retinopathy detection using reliable efficient networks and uncertainty quantification

Abstract

Diabetic retinopathy (DR) is one of the main reasons people lose their vision, and catching it early is key to stopping permanent damage. Right now, doctors rely on manual screening, which takes a lot of time and it is not always consistent. The introduction of deep neural networks (DNNs) is a revolutionary step in analyzing high-precision DR detection, but there are concerns: these models can be over-confident in their prediction, leading to mistakes, especially in critical health care. Another problem is that the current method of deep learning does not respond well to uncertainties, which makes it difficult to trust them in the real medical environment. To address these challenges, we have developed a new system of three components. First, we improved the quality of retinal images using the Adaptive Fundus Enhancement Pipeline (AFEP). Then we will extract more useful features from the image using a modified version of EfficientNet-B0. Finally, we add steps to calibrate the model's prediction to ensure that its level of confidence is actually accurate. This step reduces the chances of incorrect diagnosis by utilizing a test time data augmentation and temperature scaling. The results of the IDRiD dataset test were promising. The model achieved 96% accuracy and showed a much better uncertainty calibration, with an expected calibration error of only 0.030. In other words, it is not only accurate, but also more reliable in the real world. Overall, our methodology can make AI-based DR screening more practical and reliable for both doctors and patients.