Rethinking the neighborhood information for deep learning-based optical coherence tomography angiography

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Optical coherence tomography angiography (OCTA) is crucial for noninvasive microvasculature studies.
• Deep learning shows promise for OCTA reconstruction but is limited by 2D architectures and local B-scan analysis.
• Current methods fail to fully leverage neighborhood information and temporal characteristics of OCTA data.

Purpose of the Study:

• To address limitations in current deep learning-based OCTA reconstruction.
• To introduce a novel network, the neighborhood information-fused Pseudo-3D U-Net (NI-P3D-U), for improved OCTA reconstruction.
• To enhance the utilization of both spatial neighborhood information and temporal dynamics in OCTA imaging.

Main Methods:

• Developed and implemented the neighborhood information-fused Pseudo-3D U-Net (NI-P3D-U) architecture.
• Evaluated the NI-P3D-U on an in vivo animal dataset using cross-validation.
• Compared the proposed method against state-of-the-art OCTA reconstruction techniques under fully and weakly supervised learning pipelines.

Main Results:

• The NI-P3D-U significantly outperformed existing deep learning OCTA algorithms in visual quality and quantitative metrics.
• Demonstrated effective generalization across different training strategies and imaging protocols.
• The neighborhood information fusion concept improved other network architectures, showing broad applicability.

Conclusions:

• The NI-P3D-U effectively reconstructs OCTA by integrating neighborhood information and temporal modeling.
• The proposed network shows potential for advanced OCTA reconstruction in clinical settings.
• Neighborhood information fusion is a valuable strategy for improving deep learning-based OCTA analysis.