HT-Net: A Hybrid Transformer Network for Fundus Vessel Segmentation.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Accurate segmentation of fundus blood vessels is crucial for diagnosing eye diseases.
• Current deep learning methods struggle with low accuracy and capillary rupture in vessel segmentation.
• Improved segmentation can significantly aid in the early detection of various ocular conditions.

Purpose of the Study:

• To propose a novel hybrid Transformer network (HT-Net) for enhanced fundus imaging analysis.
• To improve the quality and accuracy of fundus vessel segmentation, particularly for microvessels.
• To address challenges like capillary rupture and varying vessel scales in segmentation tasks.

Main Methods:

• Developed HT-Net, a hybrid network combining Transformer and Convolutional Neural Network (CNN) components.
• Integrated Feature Fusion Blocks (FFB) to enrich feature spaces at shallow network levels.
• Incorporated Feature Refinement Blocks (FRB) to fuse multi-scale information and handle vessel scale variations.
• Utilized efficient self-attention mechanisms within the network architecture.

Main Results:

• HT-Net demonstrated superior performance on the DRIVE, CHASE_DB1, and STARE datasets.
• FFB and FRB effectively improved microvessel segmentation quality by extracting multi-scale information.
• The integration of self-attention mechanisms significantly boosted vessel segmentation accuracy.
• The proposed method outperformed most existing fundus vessel segmentation techniques.

Conclusions:

• HT-Net offers a competent and effective solution for fundus vessel segmentation.
• The hybrid approach successfully captures both local details and long-range dependencies in fundus images.
• The network's ability to refine features and handle scale variations leads to higher segmentation accuracy.
• This advancement holds promise for improving early diagnosis and management of eye diseases through better imaging analysis.