A multi-modal multi-branch framework for retinal vessel segmentation using ultra-widefield fundus photographs.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Ultra-WideField (UWF) fundus images offer expanded views for disease analysis but present segmentation challenges due to high resolution and low contrast.
• Accurate retinal vessel segmentation is crucial for diagnosing various ocular diseases from fundus images.

Purpose of the Study:

• To develop a novel deep learning framework, M3B-Net, for improved retinal vessel segmentation in UWF fundus images.
• To leverage fundus fluorescence angiography (FFA) to enhance segmentation accuracy, particularly for fine vessels and low-contrast regions.

Main Methods:

• Introduced M3B-Net, a multi-modal, multi-branch framework integrating FFA images for UWF fundus image segmentation.
• Developed an enhanced UWF segmentation network incorporating a Selective Fusion Module (SFM) for improved feature extraction.
• Implemented a Local Perception Fusion Module (LPFM) to address context loss in high-resolution images and an Attention-Guided Upsampling Module (AUM) for enhanced performance.

Main Results:

• M3B-Net demonstrated significantly superior performance compared to existing state-of-the-art methods in UWF fundus image segmentation.
• The proposed modules effectively addressed challenges related to low contrast and high resolution in UWF images.

Conclusions:

• M3B-Net offers a robust solution for accurate retinal vessel segmentation in UWF fundus images.
• The framework's multi-modal approach and specialized modules pave the way for more precise disease analysis in ophthalmology.