Improved Retinopathy Detection Using ResNet-50 Optimized AI Deep Learning Models

In this study, we improve a pretrained Resnet, a-50 network to provide a better model for identifying diabetic retinal detachment. The model leverages the strength of Resnet-50’s feature extraction capabilities and the potency of transfer learning to enhance performance by modifying the final layers using attention approaches. The model may focus on the most relevant areas of the retina by employing these attention techniques, improving the accuracy and efficacy of its diagnosis. To optimise model performance, several critical factors will be taken into account. High-resolution images are used to first capture the incredibly small features of retinal scanning, which improves the model’s ability to detect subtle signs of retinopathy. In order to minimise biases and prevent overfitting, careful data collection techniques are also employed. Extra effort is made to ensure that there is no leakage throughout training and validation sets because sample(replace $=$ True) has the ability to significantly bias model evaluation. Age normalisation is also included in the model pipeline because it has a significant influence on the progression of diabetic retinopathy and must be carefully controlled to ensure accurate forecasts. Combining prepared models, high-resolution data, suitable sampling strategies, and attention processes results in a more reliable and efficient diagnostic tool. By focussing on the most relevant retinal regions and minimising potential errors in the data pipeline, the proposed approach shows promise in enhancing early detection of diabetic retinopathy and, eventually, improving patient care in hospitals.