Directional Gabor Filter for Retinal Image Registration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current algorithms for registering pairs or sequences of retinal images are computationally inefficient and unable to register images obtained from different imaging modes, such as reflectance and auto-fluorescence, due to insufficient feature detection and noise sensitivity in high-resolution retinal images with smooth vascular features.
Innovation Solution
A method involving directional filtering with a rotating Gabor kernel to enhance vascular features, followed by corner detection and cross-correlation for registering digital vascular images, allowing for accurate alignment across different imaging modes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If iterative searches and decision trees are used for vasculature tracking to achieve registration accuracy, then registration precision is improved, but computational efficiency deteriorates
Solution Approach 1:
The patent applies preliminary action by performing gradient calculation and thresholding before the main registration process. The method pre-processes images to create binary masks and identifies candidate feature points in advance, which significantly reduces the computational burden during iterative registration searches.
Solution Approach 2:
The patent extracts only the essential vascular features needed for registration by applying gradient-based thresholding to create binary masks. This extraction process isolates relevant vascular structures from the full image data, eliminating unnecessary computational processing of non-vascular regions during registration.
2Device complexity
If corner extraction is used for matching in high-resolution retinal images, then feature detection is simplified, but reliability deteriorates due to insufficient corner features and noise sensitivity
Solution Approach 1:
The patent changes the detection parameter from corner-based features to gradient-based features. By calculating intensity gradients and applying thresholding, the method transforms the feature detection approach to identify vascular structures based on intensity changes rather than geometric corners, which are more abundant and reliable in smooth retinal vasculature.
Solution Approach 2:
The patent creates binary mask copies of the original images that represent vascular structures. These binary masks serve as simplified representations that can be efficiently processed and compared for registration without requiring complex corner detection on the original high-resolution images.
3Adaptability or versatility
If known registration algorithms are used, then registration of common imaging modes is achieved, but adaptability deteriorates as they cannot register images from different imaging modes
Solution Approach 1:
The patent implements universality by creating a registration method that works across multiple imaging modes (reflectance, auto-fluorescence, and other retinal imaging modes). The gradient-based feature detection and binary mask approach is mode-agnostic, allowing the same algorithm to successfully register images from different imaging modalities by focusing on universal vascular structural features.
Data Source
Figure 1
Figure 2a~2b
Figure 3a~3b
AI summary
An image processing apparatus uses first and second digital vascular image data to register two images. The two images may be from different imaging modes. The first and second images are processed with a two- dimensional, directional filter (500) that has the effect of producing clusters of orthogonally adjacent image data points in which the magnitude of an intensity gradient between each orthogonally adjacent image data point is less than a predetermined value. Subsequently, common clusters are identified between the first and second image data using a corner detecting algorithm (600). The directional filter produces "stepping" features, where vascular features would otherwise appear with smooth edges. These numerous features are identified by the corner detecting algorithm and can be used (1000) for registering common clusters between the first and second image data. The filter may be a rotating Gabor filter matched to vascular features in the images.