Comparative study of deep learning models for optical coherence tomography angiography.

Area of Science:

• Biomedical Imaging
• Deep Learning
• Microvasculature Analysis

Background:

• Optical coherence tomography angiography (OCTA) is vital for studying microvasculature.
• Deep learning excels at image-to-image translation, with prior work exploring its use in OCTA reconstruction.
• Existing research often focuses on limited deep learning architectures for OCTA.

Purpose of the Study:

• To conduct a comparative investigation of various deep learning models for OCTA reconstruction.
• To evaluate the performance of different network architectures in OCTA image processing.
• To identify optimal deep learning approaches for enhanced microvasculature imaging.

Main Methods:

• Four representative deep learning architectures were investigated: single-path, U-shaped, generative adversarial network (GAN)-based, and multi-path models.
• The models were trained and tested on a dataset of OCTA images from rat brains.
• Three potential solutions were explored to assess performance improvement feasibility.

Main Results:

• U-shaped and multi-path deep learning models demonstrated suitability for OCTA reconstruction.
• Comparative analysis revealed distinct performance characteristics among the investigated architectures.
• The study identified merging phase information as a promising avenue for future OCTA reconstruction enhancement.

Conclusions:

• U-shaped and multi-path architectures are effective for OCTA reconstruction.
• Deep learning offers a powerful toolkit for advancing OCTA-based microvasculature studies.
• Future research directions include integrating phase information to further improve OCTA image quality.