MFI-Net: A multi-resolution fusion input network for retinal vessel segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Deep Learning

Background:

• Retinal vessel segmentation is crucial for diagnosing eye diseases.
• Current deep learning methods struggle with shallow feature extraction, leading to blurred boundaries and inaccurate capillary segmentation.
• Information fusion in existing models can isolate shallow layer features, introducing noise.

Purpose of the Study:

• To propose MFI-Net (Multi-resolution Fusion Input Network) to address limitations in current retinal vessel segmentation methods.
• To improve the accuracy and robustness of retinal vessel segmentation.
• To enhance the diagnosis of retinal diseases through improved segmentation.

Main Methods:

• Developed MFI-Net, incorporating a multi-resolution input module for comprehensive feature extraction (local and global).
• Implemented an information aggregation method for improved feature fusion between encoder and decoder layers.
• Validated MFI-Net on DRIVE, CHASE_DB1, and STARE datasets.

Main Results:

• MFI-Net demonstrated superior performance compared to U-Net and R2U-Net across multiple metrics.
• Achieved higher F1 scores than U-Net by 2.42%, 2.46%, and 1.61% on the datasets.
• Outperformed R2U-Net with F1 score improvements of 1.47%, 2.22%, and 0.08%.

Conclusions:

• MFI-Net effectively alleviates issues of incomplete shallow feature extraction and noisy segmentation.
• The proposed model shows high stability and generalization ability for retinal vessel segmentation.
• MFI-Net offers a robust and accurate solution for clinical applications in ophthalmology.