MON-855 - Evaluating the Clinical Value of Resnet152 for Detection and Differentiation of Diabetic Retinopathy from Common Ocular Conditions using Fundus Imaging

Introduction
Diabetic retinopathy is a major microvascular complication of metabolic disease and a leading cause of preventable vision loss worldwide. Artificial intelligence–based retinal screening has shown promise in improving access to early detection; however, the clinical value of progressively deeper neural architectures remains uncertain. While residual learning enables stable optimization at extreme depth, increasing architectural complexity may yield diminishing diagnostic returns relative to computational cost. Evaluating ultra-deep networks as benchmarks can clarify whether additional depth translates into clinically meaningful benefit for endocrine screening workflows.

Aim
To evaluate ResNet152 for multiclass diabetic retinopathy screening and to assess the diagnostic gains associated with extreme network depth.

Methods
We trained a residual convolutional neural network based on ResNet152 on a pooled dataset of 50,000 color fundus images reviewed by expert ophthalmologists. Images were labeled across seven categories: diabetic retinopathy, age-related macular degeneration, cataract, glaucoma, hypertensive retinopathy, pathological myopia, and normal fundus. Diabetic retinopathy severity was graded according to the International Clinical Diabetic Retinopathy classification. 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 evaluated using accuracy, precision, recall, F1 score, and class-wise area under the receiver operating characteristic curve.

Results
ResNet152 demonstrated strong and reproducible diagnostic performance across heterogeneous datasets. On validation, overall accuracy exceeded 98 percent with balanced precision, recall, and F1 scores approaching 0.98. High discriminatory capacity was observed across all ocular categories, including diabetic retinopathy and other vision-threatening conditions. On independent testing, accuracy remained high at approximately 97.6 percent with an F1 score of 0.98, confirming robust generalization. Despite excellent performance, incremental gains relative to shallower residual architectures were modest and accompanied by substantially increased computational complexity.

Conclusion
ResNet152 establishes a robust upper-bound benchmark for multiclass diabetic retinopathy screening, demonstrating that extreme architectural depth yields only incremental diagnostic improvement beyond lighter residual networks. These findings highlight diminishing returns with increasing depth and support for prioritizing efficient architectures for scalable, real-world deployment in endocrinology-led retinal screening programs, where performance must be balanced against computational cost and accessibility.

*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.*