Retinal Layer Boundary Identification Using Guided Bidirectional Graph Search
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Wide-field optical coherence tomography (OCT) systems face challenges in accurately segmenting retinal layer boundaries due to decreased axial resolution, reduced back-scattered reflectance contrast, and increased retinal curvature, which complicates the identification of features for machine learning segmentation and leads to local errors in existing algorithms.
Innovation Solution
The implementation of a Guided Bidirectional Graph Search (GB-GS) method that uses guidance point arrays to guide the graph search in multiple directions, identifying retinal layer boundaries by generating candidate paths based on gradient maps and merging them to approximate the boundary positions, thereby improving accuracy and efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Area of stationary object
If wide-field OCT scanning is used to evaluate larger portions of the retina, then the field of view is improved, but axial resolution and back-scattered reflectance contrast decrease
Solution Approach 1:
The patent applies preliminary flattening of the retinal surface before segmentation to compensate for the decreased axial resolution in wide-field OCT. By pre-processing the image to flatten the curved retinal surface, the algorithm creates a standardized geometric foundation that improves subsequent boundary detection accuracy despite the inherent resolution limitations of wide-field scanning.
2Area of stationary object
If wide-field OCT scanning is used to evaluate larger portions of the retina, then the field of view is improved, but back-scattered reflectance contrast is reduced
Solution Approach 1:
The patent introduces gradient maps as an intermediary representation that enhances reflectance contrast. By computing gradients of the OCT signal, the algorithm transforms the low-contrast wide-field image into a gradient map where boundaries between retinal layers appear as prominent features, effectively mediating the contrast problem without requiring higher resolution imaging.
3Area of stationary object
If wide-field OCT scanning is used to evaluate larger portions of the retina, then the field of view is improved, but retinal curvature is increased
Solution Approach 1:
The patent performs preliminary flattening of the curved retinal surface to create a planar reference frame for segmentation. This pre-processing step transforms the naturally curved wide-field retinal image into a flattened coordinate system, eliminating curvature-induced segmentation errors while preserving the expanded field of view.
4Extent of automation
If traditional graph search algorithms are used for segmentation, then automated boundary identification is achieved, but local errors are propagated
Solution Approach 1:
The patent implements feedback mechanisms where the segmentation algorithm continuously refines boundary predictions by comparing gradient map features with anatomical constraints. The system uses feedback from multiple gradient directions and iterative optimization to correct local errors, preventing error propagation while maintaining automated operation.
5Extent of automation
If machine learning segmentation methods are used, then automated boundary identification is achieved, but large training datasets are required
Solution Approach 1:
The patent employs self-service segmentation where the algorithm extracts features directly from the OCT image gradients and anatomical relationships without requiring external training data. The method uses intrinsic image properties and gradient-based feature extraction to perform automated segmentation independently, eliminating the need for large training datasets while maintaining automation.
Data Source
AI summary
Methods for automatically identifying retinal boundaries from a reflectance image are disclosed. An example of the method includes identifying a reflectance image of the retina of a subject; generating a gradient map of the reflectance image, the gradient map representing dark-to-light or light-to-dark reflectance differentials between adjacent pixel pairs in the reflectance image; generating a guidance point array corresponding to a retinal layer boundary depicted in the reflectance image using the gradient map; generating multiple candidate paths estimating the retinal layer boundary in the reflectance image by performing a guided bidirectional graph search on the reflectance image using the guidance point array; and identifying the retinal layer boundary by merging two or more of the multiple candidate paths.