Optic Nerve Head Segmentation Using Stereo Disparity Features
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for diagnosing glaucoma, such as Heidelberg Retinal Tomography and Optical Coherence Tomography, lack automation in determining optic nerve cupping from stereo photographs, necessitating the development of automated planimetry systems for objective and quantitative assessment.
Innovation Solution
A method that classifies optic nerve cup, rim, and blood vessel pixels in retinal images using a trained classifier, incorporating non-stereo and stereo disparity features, such as Gaussian derivatives and Gabor wavelets, to determine the cup-to-disc ratio.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If automated planimetry is implemented, then productivity and objectivity are improved, but device complexity increases
Solution Approach 1:
The classification process is segmented into multiple stages: initial pixel classification using a trained classifier, followed by refinement through morphological operations and boundary detection. This multi-stage segmentation approach breaks down the complex automated planimetry task into manageable components, improving overall system productivity while managing complexity through modular design.
Solution Approach 2:
A trained classifier is pre-trained on a database of stereo retinal images with known optic nerve head boundaries before deployment. This preliminary action allows the system to learn optimal classification parameters in advance, enabling rapid and accurate automated planimetry without requiring complex real-time decision-making, thus improving productivity while keeping the operational system relatively simple.
2Measurement precision
If multiple feature types are used for classification, then measurement precision is improved, but device complexity increases
Solution Approach 1:
The system incorporates both two-dimensional image intensity features and three-dimensional depth information from stereo disparity maps. By adding this extra dimensional information (depth), the classification precision is significantly improved as the system can distinguish optic nerve head boundaries more accurately, while the complexity is managed by integrating these features into a unified classification framework.
Solution Approach 2:
The feature set combines multiple types of image characteristics (intensity, texture, color, and depth) into a composite feature vector for each pixel. This composite approach allows the classifier to leverage complementary information from different feature types, improving measurement precision while the feature fusion process manages the overall complexity through systematic integration.
3Measurement precision
If stereo disparity features are incorporated, then classification accuracy is improved, but processing time increases
Solution Approach 1:
Stereo disparity maps and depth information are pre-computed from the stereo retinal image pair before the classification stage. This preliminary processing allows the main classification algorithm to directly utilize ready-to-use depth features without performing computationally intensive stereo matching during the time-critical classification phase, thus improving boundary detection accuracy while minimizing additional processing time.
Solution Approach 2:
The system computes stereo disparity features for all pixels in the image, even though only pixels near the optic nerve head boundary require high-precision depth information. This partial excess action ensures that sufficient depth information is available throughout the image for accurate boundary detection, while the classifier can efficiently focus computational resources on the critical boundary regions during the classification phase.
Data Source
AI summary
A method of classifying an optic nerve cup and rim of an eye from a retinal image comprising receiving a retinal image, determining a feature vector for a candidate pixel in the retinal image, and classifying the candidate pixel as a cup pixel or a rim pixel based on the feature vector using a trained classifier. The retinal image can be a stereo pair, the retinal image can be color or monochrome. The method disclosed can further comprise identifying an optic nerve, identifying an optic nerve cup and optic nerve rim, and determining a cup-to-disc ratio.