MON-583 - Compound Scaling Meets Clinical Precision: Efficientnetb3-Based AI Application for Accurate and Scalable Diabetic Retinopathy Detection

Introduction

Diabetic retinopathy is a common microvascular complication of diabetes mellitus and a major cause of vision loss worldwide. Early detection is essential for effective diabetes management, yet conventional screening remains resource intensive and inconsistently accessible. Artificial intelligence offers scalable screening solutions, provided models preserve high diagnostic accuracy with computational efficiency suitable for real-world clinical use.

Aim

To evaluate the performance, generalizability, and efficiency of the EfficientNetB3 deep learning architecture for diabetic retinopathy detection, severity grading, and differentiation from other common ocular diseases using globally diverse fundus image datasets.

Methods

We developed a compound-scaled convolutional neural network based on EfficientNetB3, which jointly adjusts network depth, width, and input resolution to optimize accuracy per computational unit. The model was trained on a pooled dataset of 50,000 color fundus images derived from multinational sources spanning six continents, including EyePACS, APTOS 2019, DDR, IDRiD, FGADR, MESSIDOR 1 and 2, ODIR-5K, DeepDRiD, and OIA-DDR. Each image was independently reviewed by at least two expert ophthalmologists, with diabetic retinopathy graded according to the International Clinical Diabetic Retinopathy classification. Data were apportioned into training, validation, and independent test cohorts with no dataset overlap. Transfer learning with ImageNet initialization and end-to-end fine-tuning were applied. Performance was assessed using accuracy, precision, recall, F1, F2 scores, and area under the receiver operating characteristic curve (AUROC).

Results

EfficientNetB3 demonstrated stable convergence and robust generalization across heterogeneous datasets. On validation, the model achieved >98 percent accuracy with balanced precision, recall, and F1 scores of approximately 0.98. Class-wise discrimination remained consistently high across diabetic retinopathy and other ocular conditions, including normal fundus, cataract, age-related macular degeneration, glaucoma, hypertensive retinopathy, and pathological myopia. On independent testing, overall accuracy was 97.1 percent with an F1 score of 0.97, confirming reproducible performance across geographically diverse data sources.

Conclusion

EfficientNetB3 demonstrates that compound-scaled deep learning can achieve high diagnostic accuracy for diabetic retinopathy without excessive computational demand. Its strong generalizability across multinational datasets supports its potential as a scalable screening tool to aid early detection of diabetic microvascular complications and enhance diabetes care within endocrine practice.

*Unless otherwise noted, all abstracts presented at ENDO must not be released to the press or the public until the date and time of presentation. For oral presentations, the abstracts are embargoed until the session begins. The Endocrine Society reserves the right to lift the embargo on specific abstracts that are selected for promotion prior to or during ENDO.*