Related Experiment Video
Updated: Jul 9, 2026
08:54
A Workflow to Quantitatively Determine Age-Related Macular Degeneration Lesion-Specific Variations in Fundus Autofluorescence
Published on: May 26, 2023
A Lesion-Aware Detection Framework for Red Lesions in Retinal Fundus Images
Pritam Mandal1, Swagata Kundu2, Aatreya Sengupta3
1Department of Electrical Engineering, National Institute of Technology, Durgapur, Durgapur, 713209, India. pritamkgei@gmail.com.
Journal of Imaging Informatics in Medicine
|July 7, 2026
Summary
This study introduces a new AI method to detect red lesions, early signs of diabetic retinopathy (DR). The refined feature attention module significantly improves the detection of these critical indicators, aiding in vision loss prevention.
Area of Science:
- Ophthalmology
- Medical Imaging
- Artificial Intelligence
Background:
- Diabetic retinopathy (DR) is a leading cause of vision loss.
- Early detection of red lesions (microaneurysms and hemorrhages) is crucial for managing DR.
- Current detection methods may struggle with diverse lesion morphologies and dataset variations.
Purpose of the Study:
- To develop an advanced deep learning model for accurate detection of red lesions in retinal images.
- To enhance the feature representation and detection capabilities for early DR signs.
- To improve the robustness and generalization of DR detection models across different datasets.
Main Methods:
- Implementation of a refined feature attention (RFA) module integrated with a feature pyramid network (FPN).
- Dataset-specific adaptive anchor ratios derived via quantile-based multi-stage clustering.
- A novel lesion-aware sampling strategy based on morphological characteristics for balanced training.
Main Results:
- Achieved FROC scores of 0.4650 (Messidor-A), 0.4429 (SaNMoD), and 0.549 (E-Ophtha MA).
- Demonstrated high precision (73.6%), recall (69.0%), and F1-score (71.2%) on the Messidor-A dataset.
- The combined RFA module, adaptive anchoring, and lesion-aware sampling significantly improved red lesion detection.
Conclusions:
- The proposed RFA module enhances feature refinement for improved lesion detection.
- Adaptive anchoring and lesion-aware sampling strategies optimize model performance for diverse retinal datasets.
- The developed network shows strong potential for clinical application in early diabetic retinopathy screening.
