MTPA_Unet: Multi-Scale Transformer-Position Attention Retinal Vessel Segmentation Network Joint Transformer and CNN.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Accurate retinal vessel segmentation is crucial for diagnosing and managing diseases like diabetic retinopathy and hypertension.
• Traditional methods struggle with complex vessel structures and limited receptive fields.
• Deep learning, particularly convolutional neural networks (CNNs) and Transformers, offers potential for improved segmentation.

Purpose of the Study:

• To develop a novel deep learning model, MTPA-Unet, for enhanced retinal vessel segmentation.
• To leverage the strengths of both CNNs (local features) and Transformers (long-range dependencies).
• To improve the accuracy and detail of blood vessel segmentation, especially for capillaries.

Main Methods:

• A hybrid network, MTPA-Unet, combining CNNs and Transformers was designed.
• Multi-resolution image input was utilized for comprehensive feature extraction.
• A novel Transformer Path Aggregation (TPA) module was introduced to capture long-distance dependencies and focus on vessel pixel location.
• The model was evaluated on the DRIVE, CHASE DB1, and STARE retinal image datasets.

Main Results:

• MTPA-Unet achieved high accuracy (0.9718-0.9773), sensitivity (0.8410-0.8938), and Dice coefficients (0.8318-0.8557) across all datasets.
• The TPA module effectively captured long-range dependencies and aided in capillary segmentation.
• The proposed method outperformed existing retinal image segmentation techniques on public datasets.

Conclusions:

• MTPA-Unet demonstrates superior performance in retinal vessel segmentation compared to current methods.
• The hybrid CNN-Transformer architecture with the TPA module is effective for detailed vessel analysis.
• This approach holds significant promise for improving early disease detection and treatment planning in ophthalmology.