Related Experiment Video
Updated: May 1, 2026
15:18
Near Infrared Optical Projection Tomography for Assessments of β-cell Mass Distribution in Diabetes Research
Published on: January 12, 2013
16.5K
Integrating non-linear radon transformation for diabetic retinopathy grading
Farida Mohsen1, Samir Belhaouari1, Zubair Shah2
1College of Science and Engineering, Hamad Bin Khalifa University, Doha, Qatar.
Scientific Reports
|August 21, 2025
Summary
This study introduces RadFuse, a new deep learning framework for detecting and grading diabetic retinopathy. RadFuse significantly improves accuracy by combining fundus images with novel sinogram representations, outperforming existing methods.
Area of Science:
- Ophthalmology
- Medical Imaging
- Artificial Intelligence
Background:
- Diabetic retinopathy is a leading cause of vision loss, necessitating early detection and grading.
- Current deep learning models struggle with the complex lesion patterns in retinal images.
- Accurate classification is crucial for timely intervention and preventing blindness.
Purpose of the Study:
- To introduce RadFuse, a multi-representation deep learning framework for enhanced diabetic retinopathy detection and grading.
- To improve the capture of subtle, complex retinal lesion features.
- To leverage both spatial and transformed domain information for better classification.
Main Methods:
- Developed RadEx transformation, a non-linear extension of the Radon transform, to create sinogram representations.
- Integrated sinogram images with traditional fundus images within a deep learning framework (RadFuse).
- Evaluated RadFuse using ResNeXt-50, MobileNetV2, and VGG19 on APTOS-2019 and DDR datasets.
Main Results:
- RadFuse significantly improved diabetic retinopathy detection and grading across all tested CNN architectures.
- Achieved 93.24% kappa for five-stage severity grading and 99.09% accuracy for binary classification.
- Outperformed state-of-the-art methods on benchmark datasets, demonstrating superior feature capture.
Conclusions:
- RadFuse effectively captures complex non-linear features, advancing diabetic retinopathy classification.
- The framework demonstrates the potential of integrating advanced mathematical transforms in medical image analysis.
- This approach offers a promising direction for improving automated screening and diagnosis of diabetic retinopathy.
