TiM-Net: Transformer in M-Net for Retinal Vessel Segmentation.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Retinal images offer insights into cardiovascular and cerebrovascular health.
• Accurate retinal vessel segmentation is vital for clinical diagnosis.
• Existing Convolutional Neural Network (CNN) models struggle with noise and underutilize long-range image features.

Purpose of the Study:

• To develop an advanced model for robust retinal vessel segmentation.
• To improve the utilization of features in fundus images.
• To enhance the clinical applicability of automated retinal image analysis.

Main Methods:

• Proposed TiM-Net, integrating M-Net with diverse attention mechanisms and weighted side outputs.
• Implemented a dual-attention mechanism (channel and spatial) to mitigate noise interference.
• Incorporated Transformer's self-attention into skip connections to model long-range relationships.
• Introduced weighted side output layers for superior feature utilization.

Main Results:

• TiM-Net demonstrated superior effectiveness and robustness in retinal vessel segmentation across three public datasets.
• Quantitative and qualitative analyses confirmed the model's clinical practicality.
• TiM-Net variants showed competitive performance, indicating strong scalability and generalization.

Conclusions:

• TiM-Net offers a significant advancement in retinal vessel segmentation, outperforming existing methods.
• The model's ability to handle noise and capture long-range dependencies makes it clinically valuable.
• The proposed architecture provides a scalable and generalizable solution for automated retinal image analysis.