Related Experiment Video
Updated: Jun 28, 2026
09:17
Using Retinal Imaging to Study Dementia
Published on: November 6, 2017
22.4K
BigEye: a clinically interpretable deep learning framework for diabetic retinopathy detection and stage prediction.
Hunter Mathias Gill1, Doaa Hassan Salem1,2, Okiemute Beatrice Omoru1
1Department of Biomedical Engineering & Informatics, Luddy School of Informatics, Computing and Engineering, Indiana University Indianapolis, 535 West Michigan Street, Indianapolis, IN, 46202, USA.
Scientific Reports
|March 9, 2026
Summary
BigEye, a novel framework, uses extracted retinal lesion features to predict Diabetic Retinopathy (DR) stages. This explainable AI approach aligns with clinical criteria, aiding in DR diagnosis.
Area of Science:
- Ophthalmology
- Medical Imaging
- Artificial Intelligence
Background:
- Diabetic Retinopathy (DR) is a leading cause of vision loss in diabetic patients.
- Conventional DR diagnosis relies on manual assessment of retinal lesions from fundus images using scales like the International Classification of Diabetic Retinopathy (ICDR).
- The increasing prevalence of DR necessitates advanced diagnostic tools, particularly those offering explainability.
Purpose of the Study:
- To develop and evaluate BigEye, a novel deep learning framework for predicting ICDR stages of Diabetic Retinopathy.
- To leverage extracted retinal lesion features for accurate and explainable DR staging.
- To demonstrate the clinical relevance of AI-driven DR diagnosis.
Main Methods:
- A DeepLabV3+ model was trained on a dataset of fundus images with annotated segmentation masks for six retinal lesions.
- Extracted features, including lesion quantities and pixel areas, were integrated into a classifier model.
- The framework's performance was evaluated using 10-fold nested cross-validation, with Shapely Additive Explanations (SHAP) used for interpretability.
Main Results:
- The BigEye framework achieved high performance metrics: 0.77 precision, 0.71 recall, 0.72 F1 score, 0.95 ROC-AUC, and 0.83 accuracy.
- SHAP value analysis confirmed that the discriminative lesions identified by the model closely correspond to established ICDR staging criteria.
- The results indicate strong alignment between the AI model's predictions and clinical knowledge.
Conclusions:
- BigEye provides an effective and explainable method for predicting Diabetic Retinopathy stages using retinal lesion features.
- The framework's ability to align with clinical criteria suggests its potential utility in augmenting DR diagnosis.
- This approach offers a promising direction for developing AI tools in medical imaging for ophthalmology.
