AI powered Early Detection of diabetic retinopathy: A Deep Learning Approach for improved Clinical Decision-Making.

Abstract

Diabetic retinopathy (DR) is one of the primary reasons of vision loss conditions, and the prevention of irreparable loss is essential. The study involves deep learning models with applications to the detection of automatic DR of retinal fundus images, the purpose of which is to help specialists make a more accurate decision and enhance the accuracy of diagnoses. Four large-scale deep learning networks were tested by five severities of retina penalty namely CNN, ResNet50, DenseNet121 and InceptionV3 architecture on a sampled data set comprising 35,000 retina images. Preprocessing of data, additional data, and transfer learning were also implemented to improve model generalization. They were proven to be capable of 90.8% testing accuracy and 0.92 AUC-ROC with experimental results indicating CNN as a baseline. The RN50 saw a better increase in detection to 91.9% and 0.93 AUC-ROC, with DenseNet121 once again making a jump of 92.6% and 0.94 AUC-ROC for detection. The model that best did its work, InceptionV3 had the highest test accuracy, the model returned a cross of 93.2% and AUC-ROC was 0.95 that truly differentiated all levels of DR phases: Mild level, Severe level, and Proliferative. The article notes that many deep learning models, especially InceptionV3 and DenseNet121, are capable of autonomously diagnosing DR and reducing the workload of clinical staff, and they can enhance early treatment. These research results have proved the possibility of AI-based retomic ophthalmic systems being helpful tools in enhancing patient outcomes.