Super-Resolution Reconstruction of OCTA Via Multi-Field-of-View Representation Learning.

Area of Science:

• Ophthalmology
• Biomedical Imaging
• Medical Image Analysis

Background:

• High-resolution Optical Coherence Tomography Angiography (OCTA) is crucial for analyzing retinal vasculature and diagnosing eye diseases.
• A persistent challenge in OCTA instrumentation is the trade-off between high resolution (HR) and a large scanning field-of-view (FOV).
• Large FOV OCTA images offer more retinal data but typically suffer from low resolution (LR), noise, and poor contrast.

Purpose of the Study:

• To develop a novel method for generating HR OCTA images with a larger FOV.
• To enable LR OCTA images to learn HR feature representations for improved retinal analysis.
• To enhance the accuracy of retinal structure segmentation and eye disease classification using OCTA data.

Main Methods:

• A self-similar dynamic domain adaptation network utilizing cross-field-of-view representation learning was proposed.
• A multiple random degradation model was used to generate synthetic LR images from HR OCTA images.
• A dynamic domain adaptation framework and a self-similar supervision loss were employed for LR to HR reconstruction.

Main Results:

• The proposed method successfully generated HR OCTA images from LR inputs.
• Experimental results demonstrated superior performance compared to existing state-of-the-art methods on three OCTA datasets.
• Significant enhancements in retinal structure segmentation and disease classification were observed.

Conclusions:

• The novel network effectively addresses the HR/large FOV trade-off in OCTA imaging.
• The method shows significant potential for improving the diagnosis and monitoring of eye-related diseases.
• The study introduces the first OCTA dataset with paired $3\times 3$ and $6\times \text{6}~\text{mm}^{2}$ images, along with publicly available code.