CoVi-Net: A hybrid convolutional and vision transformer neural network for retinal vessel segmentation

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Accurate retinal vessel segmentation is vital for diagnosing ocular diseases.
• Existing methods struggle with integrating global and local features from fundus images, limiting diagnostic capabilities.

Purpose of the Study:

• To develop an advanced hybrid deep learning network, CoVi-Net, for improved retinal vessel segmentation.
• To address limitations in feature fusion and the simultaneous capture of multi-scale features in current segmentation techniques.

Main Methods:

• Introduced CoVi-Net, a hybrid network combining convolutional neural networks and vision transformers.
• Developed novel modules: local and global feature aggregation (LGFA), bidirectional weighted feature fusion (BWF), and adaptive lateral feature fusion (ALFF).
• Incorporated horizontal and vertical connections for enhanced multi-scale feature fusion.

Main Results:

• CoVi-Net achieved superior performance on DRIVE, CHASEDB1, and STARE datasets, outperforming state-of-the-art methods.
• Achieved high global accuracy (up to 0.9761) and area under the curve (up to 0.9915).
• Ablation studies confirmed the effectiveness of individual modules, and adaptability for lesion segmentation was demonstrated.

Conclusions:

• CoVi-Net offers a promising approach for accurate retinal vessel segmentation.
• The model's advanced feature fusion strategies enhance the diagnosis of retinal vascular disorders.
• CoVi-Net demonstrates potential for clinical application in ophthalmology.