Every year, diabetic retinopathy (DR) threatens the vision of millions, but the screening process just can't keep up. The current system moves slowly—specialists are overworked, results change from one doctor to the next, and too many patients learn they have DR only after their sight is already at risk. We wanted a fix. So, our team created an automated deep learning platform—a hybrid that stacks a Convolutional Neural Network (CNN) with Long Short-Term Memory (LSTM) layers. This model doesn't just detect DR; it also grades its severity from plain retinal fundus photos. We didn't stop after building one model. We tried three hybrid approaches—Custom CNN+LSTM, MobileNetV2+LSTM, and InceptionResNetV2+LSTM—and compared them to seven standard CNN-only baselines. To make sure even the subtle signs stand out, we used CLAHE (Contrast Limited Adaptive Histogram Equalization) on every image. Medical datasets are imbalanced by nature, so we rebalanced things through loss weighting, giving serious DR cases the extra attention they deserve. And for transparency, we turned to Grad-CAM, producing heatmaps so doctors can see exactly what our AI focused on. When it came down to results, the InceptionResNetV2+LSTM beat the rest: 91.4% accuracy, a Quadratic Cohen's Kappa of 0.89, and a Macro F1-Score of 0.86 for multi-class DR grading. More than just numbers—two ophthalmologists validated our Grad-CAM maps and agreed with the AI's focus 91% of the time. To make everything practical, we built a Streamlit web app layered with secure roles, live predictions, explainability, and instant PDF reports. This project pushes DR screening closer to where it needs to be. With smarter AI, clear explainability, and a clinic-ready platform, screening can be faster, fairer, and more dependable—catching DR cases that used to slip by, and backing up doctors with real confidence.
Patil et al. (Thu,) studied this question.