Hard Attention Net for Automatic Retinal Vessel Segmentation.

Area of Science:

• Ophthalmologic image analysis
• Medical image segmentation
• Deep learning applications

Background:

• Automated retinal vessel segmentation is crucial in ophthalmology.
• Existing deep learning models struggle with challenging vessels like thin or ill-defined ones.
• A need exists for advanced models that address these segmentation difficulties.

Purpose of the Study:

• To propose a novel end-to-end deep learning architecture, HAnet, for improved retinal vessel segmentation.
• To enhance the focus on difficult-to-segment regions using a hard attention mechanism.
• To evaluate HAnet's performance across diverse datasets and modalities.

Main Methods:

• Developed HAnet, an architecture with three decoders for segmenting 'hard' and 'easy' regions independently.
• Integrated attention mechanisms to specifically reinforce focus on challenging image features.
• Fused outputs from all decoders to generate the final vessel segmentation map.

Main Results:

• HAnet demonstrated superior or comparable performance in segmentation accuracy, AUC, and F1-score against state-of-the-art models.
• Evaluated on multiple public and self-collected datasets including fundus photography, SLO, and ICG angiography.
• Cross-dataset and cross-modality evaluations confirmed the model's generalization ability and extendibility.

Conclusions:

• HAnet effectively addresses the challenge of segmenting difficult retinal vessels.
• The proposed architecture offers improved accuracy and robustness in retinal vessel segmentation.
• HAnet shows significant potential for clinical applications in ophthalmologic image analysis.