Segmentation in Diabetic Retinopathy using Deeply-Supervised Multiscalar Attention.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic Retinopathy (DR) is a leading cause of irreversible blindness.
• Early detection of DR requires identifying subtle pathological features.
• Current detection methods may lack the precision for minute pathologies.

Purpose of the Study:

• To introduce a novel Deeply-Supervised Multiscale Attention U-Net (Mult-Attn-U-Net) for DR pathology segmentation.
• To evaluate the model's performance on a public dataset.
• To compare the proposed model against state-of-the-art methods.

Main Methods:

• Development of a Deeply-Supervised Multiscale Attention U-Net (Mult-Attn-U-Net).
• Utilizing the publicly available Indian Diabetic Retinopathy Image Dataset (IDRiD).
• Segmentation of diabetic retinopathy pathologies: Microaneurysms (MA), Hemorrhages (HE), Soft Exudates (SE), and Hard Exudates (EX).

Main Results:

• The Mult-Attn-U-Net achieved superior performance compared to four other state-of-the-art models.
• Achieved segmentation accuracies of 0.65 for MA, 0.70 for HE, and 0.72 for SE.
• Demonstrated effective segmentation of key diabetic retinopathy pathologies.

Conclusions:

• The proposed Mult-Attn-U-Net is a highly effective tool for segmenting diabetic retinopathy pathologies.
• This model shows significant potential for improving early DR detection and preventing blindness.
• The study highlights the advancement of AI in medical image analysis for eye diseases.