ORF33-04 - Efficient and Scalable Diabetic Retinopathy Screening using EfficientNetB0: Multiclass Detection, Severity Staging, and Differential Diagnosis from Fundus Imaging

Introduction

Diabetic retinopathy is a leading cause of preventable blindness worldwide, and timely retinal screening is central to effective diabetes care. Although artificial intelligence has improved automated detection, many models remain challenging to deploy in resource-limited settings due to computational demands. Architectures that achieve high diagnostic accuracy with minimal computational burden are essential to enable scalable and equitable screening across diverse healthcare environments.

Aim

To develop and validate a computationally efficient EfficientNetB0-based deep learning model for diabetic retinopathy detection, severity staging, and differential diagnosis from fundus imaging using globally curated datasets.

Methods

We trained a compound-scaled convolutional neural network based on EfficientNetB0 on 50,000 color fundus images pooled from nine international datasets spanning six continents (EyePACS, APTOS 2019, DDR, IDRiD, FGADR, MESSIDOR-1/2, ODIR-5K, DeepDRiD, and OIA-DDR). Images were independently reviewed by board-certified ophthalmologists, with diabetic retinopathy graded according to the International Clinical Diabetic Retinopathy severity scale. Data were partitioned into training, validation, and independent test cohorts. Transfer learning with ImageNet initialization and end-to-end fine-tuning were applied. Model performance was assessed using accuracy, precision, recall, F1-score, and class-wise area under the receiver operating characteristic curve.

Results

EfficientNetB0 demonstrated stable learning and strong generalizability across heterogeneous datasets. On validation, accuracy exceeded 98 percent with balanced precision, recall, and F1-scores approaching 0.99. High discriminative performance was maintained across diabetic retinopathy and other common ocular conditions, including normal fundus, cataract, age-related macular degeneration, glaucoma, hypertensive retinopathy, and pathological myopia. On independent testing, overall accuracy approached 98 percent with an F1-score near 0.98, confirming reproducible performance outside the training environment. Inference was computationally lightweight, enabling rapid analysis suitable for deployment on widely available clinical hardware.

Conclusion

EfficientNetB0 demonstrates that architectural efficiency and high diagnostic performance are complementary rather than competing objectives. By delivering reliable multiclass diabetic retinopathy detection and severity staging with minimal computational requirements, this model supports scalable deployment in endocrinology-led screening programs, teleophthalmology platforms, and resource-constrained clinical settings, advancing equitable access to vision-preserving care.