Multi-Stage Diabetic Retinopathy Detection With Comparative Study of CNN, ResNet50, and EfficientNetB2

Abstract

Diabetic retinopathy (DR) is a leading cause of preventable blindness among working-age adults with diabetes, yet systematic fundus screening remains inaccessible across much of the developing world due to an acute shortage of trained ophthalmologists. Early detection of the eye DR helps the doctors to decide the severity of the case. We evaluate automated five-class DR severity grading using three deep convolutional neural network architectures trained on the APTOS 2019 benchmark dataset supplemented with clinical fundus images from collaborating hospitals: a randomly initialized baseline CNN, ResNet50 with ImageNet pre-training, and EfficientNetB2 with ImageNet pre-training. All models are trained under identical conditions to isolate architectural effects. The CNN achieved 53% validation accuracy, near the 49.3% majority-class baseline. ResNet50 reached 96% and EfficientNetB2 reached 98%, with the lowest validation loss among the three, indicating superior calibration as well as accuracy. Macro-averaged F1 and AUC are reported throughout to account for a 9.3:1 class imbalance between Grade 0 and Grade 3 images. We additionally demonstrate that Gaussian blur must be applied before, not after, circular cropping to prevent a ring artifact at the retinal boundary that arises from the alternative preprocessing order.