Performance of ResNet-18 and InceptionResNetV2 in Automated Detection of Diabetic Retinopathy

Abstract

Background

Diabetic retinopathy (DR) is a leading cause of blindness worldwide, particularly in diabetic individuals. Manual detection of DR by ophthalmologists is time-consuming and resource-intensive, making early automated detection essential for mitigating the risk of vision impairment. This study evaluates the effectiveness of two deep learning models, ResNet-18 and InceptionResNetV2, for detecting and classifying DR from retinal fundus images, with the aim of identifying the most suitable model for clinical application.

Methods

A dataset of 3662 retinal fundus images, divided into five DR severity classes, was used to train and test ResNet-18 and InceptionResNetV2. The key performance metrics used to assess classification across the DR stages included testing accuracy, precision, recall, specificity, F₁ score, and area under the curve (AUC).

Results

ResNet-18 achieved a testing accuracy of 83% and an AUC of 0.946, showing robust generalization across DR stages. InceptionResNetV2 achieved a testing accuracy of 70.4% and an AUC of 0.9305, with high precision in distinguishing “No DR” cases. However, it exhibited overfitting, particularly in “Mild” and “Proliferative DR” classifications, whereas ResNet-18 demonstrated a more stable performance across categories.

Conclusions

Our results suggest that ResNet-18 holds significant potential as an automated DR detection tool, providing reliable classification and superior generalization across DR stages. Integrating deep learning models such as ResNet-18 into clinical workflows may enhance early DR diagnosis and timely intervention, reducing the risk of vision impairment among patients with diabetes.