An Improved Approach for Semantic Segmentation of Fundus Lesions using R2U-Net.

Area of Science:

• Medical Imaging
• Computer Vision
• Ophthalmology

Background:

• Diabetic Retinopathy (DR) affects 33% of diabetics and can cause irreversible vision loss if untreated.
• Early detection of DR relies on identifying fundus lesions like exudates, hemorrhages, and microaneurysms.
• Automated segmentation of these lesions is crucial for timely DR diagnosis.

Purpose of the Study:

• To develop and evaluate a computational method for segmenting diabetic retinopathy fundus lesions.
• To improve the accuracy and efficiency of lesion identification in retinal images.
• To address the scarcity of annotated datasets for diabetic retinopathy research.

Main Methods:

• Implementation of an R2U-Net architecture combined with data augmentation techniques.
• Training and validation of the model on the DDR dataset.
• Testing the model's performance on the IDRiD dataset.

Main Results:

• Achieved 99.87% accuracy and 59.69% mean Intersection over Union (mIoU) on the DDR dataset.
• Obtained an mIoU of 49.92% on the IDRiD dataset.
• Demonstrated the model's potential for lesion annotation in creating new DR datasets.

Conclusions:

• The proposed R2U-Net model shows significant potential for accurate diabetic retinopathy lesion segmentation.
• The model can assist in generating annotations for new datasets, crucial for advancing DR research.
• This approach contributes to the early diagnosis and management of diabetic retinopathy.