Multi scale multi attention network for blood vessel segmentation in fundus images

Area of Science:

• Medical Imaging
• Ophthalmology
• Computer Vision
• Artificial Intelligence

Background:

• Accurate segmentation of retinal vasculature is essential for diagnosing vision-threatening conditions.
• Challenges include limited context, variable vessel thickness, complex structures, and lesion interference.
• Existing methods struggle with these inherent difficulties in retinal image analysis.

Purpose of the Study:

• To develop an advanced deep learning model for robust retinal blood vessel segmentation.
• To overcome limitations of traditional segmentation approaches in complex retinal images.
• To improve the accuracy and reliability of automated retinal vasculature analysis.

Main Methods:

• Introduced the MSMA Net model, featuring novel Multi-Scale Squeeze and Excitation (MSSE) blocks and Bottleneck residual (B-Res) paths with Spatial Attention Blocks (SAB).
• Replaced conventional convolution blocks and skip connections with these advanced architectural components.
• Validated the model on diverse public datasets: DRIVE, STARE, CHASE_DB1, HRF, and DR HAGIS.

Main Results:

• MSMA Net demonstrated superior performance compared to existing segmentation techniques across all tested datasets.
• Achieved higher accuracy, sensitivity, Dice score, and Area Under the Curve (AUC) for vessel segmentation.
• Effectively handled variations in vessel thickness, complex morphologies, and the presence of lesions.

Conclusions:

• The proposed MSMA Net model offers a significant advancement in retinal blood vessel segmentation.
• Its novel architecture effectively addresses key challenges, leading to improved diagnostic potential.
• This approach holds promise for enhanced early detection and management of ocular diseases.