Structure-Guided Cross-Attention Network for Cross-Domain OCT Fluid Segmentation.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Accurate retinal fluid segmentation in Optical Coherence Tomography (OCT) is crucial for diagnosing eye diseases.
• Deep learning models excel with annotated data but struggle with domain shift from different OCT devices.
• This limitation hinders real-world application due to device variability across hospitals.

Purpose of the Study:

• To develop a model for effective cross-domain OCT fluid segmentation.
• To address the challenge of domain shift in OCT image analysis.
• To enable robust fluid segmentation regardless of the OCT device used.

Main Methods:

• Proposed a novel Structure-guided Cross-Attention Network (SCAN) for cross-domain OCT fluid segmentation.
• Leveraged robust retinal layer structure for domain alignment.
• Employed a multi-task approach, jointly learning structure prediction and fluid segmentation.
• Introduced a cross-attention module for feature correlation and an adaptation difficulty map for adversarial learning.

Main Results:

• Demonstrated the effectiveness of SCAN on the RETOUCH dataset across three domains.
• Achieved state-of-the-art performance in cross-domain OCT fluid segmentation.
• Showcased the robustness of the retinal layer structure for domain adaptation.

Conclusions:

• The proposed SCAN method effectively overcomes domain shift in OCT fluid segmentation.
• Jointly learning structure and segmentation with cross-attention enhances model performance.
• SCAN offers a promising solution for reliable OCT image analysis in diverse clinical settings.