-Net: Retinal OCTA Image Stitching via Multi-Scale Representation Learning and Dynamic Location Guidance.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Optical coherence tomography angiography (OCTA) is vital for analyzing retinal vascular diseases.
• Limited field of view (FOV) in OCTA restricts analysis of large retinal areas.
• Existing stitching algorithms struggle with OCTA image noise, artifacts, and dense vasculature.

Purpose of the Study:

• To develop a novel deep learning network for accurate stitching of OCTA images.
• To overcome limitations of current stitching methods for enhanced retinal imaging.
• To enable high-resolution, wide-field OCTA analysis for better disease assessment.

Main Methods:

• Proposed a novel retinal OCTA image stitching network integrating multi-scale representation learning and dynamic location guidance.
• Implemented an image registration network with progressive multi-resolution feature fusion.
• Introduced a dynamic guidance strategy for foveal avascular zone (FAZ) localization and registration error constraint.
• Developed an image fusion network utilizing multiple mask constraints and adjacent image aggregation (AIA) for artifact elimination.

Main Results:

• The proposed network effectively addresses challenges like noise, artifacts, and dense vasculature in OCTA images.
• Achieved precise vessel alignment and artifact elimination in overlapping regions.
• Demonstrated superior performance compared to existing image stitching methods.
• Generated high-quality wide-field OCTA images with high structural similarity index (SSIM) scores (0.8264 and 0.8014).

Conclusions:

• The novel OCTA stitching network provides a promising solution for extending the field of view in OCTA imaging.
• Enables precise alignment and fusion of adjacent OCTA scans, leading to high-quality wide-field images.
• Facilitates more comprehensive analysis of retinal vascular diseases by allowing examination of larger retinal regions.