OCT Retinal Layer Differentiation with Anatomical Hierarchy Constraints
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for differentiating retinal layers in OCT images using deep neural networks fail to account for the anatomical hierarchical relationship between layers, leading to inaccurate predictions and errors in layer identification, particularly affecting diagnoses of retinal diseases like diabetic retinopathy and glaucoma.
Innovation Solution
A method involving a deep neural network that calculates probabilities and brightness changes to determine boundary lines between retinal layers by considering the anatomical hierarchy, using trained data sets and a graph theory-based optimization algorithm to enhance accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If deep neural network is used for image segmentation to predict retinal layer regions, then the automation and speed of retinal layer differentiation is improved, but the accuracy of layer identification deteriorates due to ignoring anatomical hierarchical relationships
Solution Approach 1:
The patent introduces an anatomical hierarchy constraint module as an intermediary between the deep neural network prediction and final layer identification. This module acts as a mediator that enforces anatomical rules (e.g., NFL must be above GCL, RPE must be below ONL) on the probability maps generated by the network, correcting predictions that violate hierarchical relationships while preserving the automation benefits of AI-based segmentation.
Solution Approach 2:
The system implements feedback by using the predicted probability maps to generate initial layer predictions, then checking these predictions against anatomical hierarchy constraints. When violations are detected, the system provides feedback to adjust the predictions, creating an iterative refinement process that improves accuracy while maintaining automated operation.
2Speed
If conventional image segmentation is used without considering anatomical hierarchy, then the processing speed is improved, but the reliability of diagnosis deteriorates due to erroneous layer predictions
Solution Approach 1:
The patent applies preliminary action by pre-establishing anatomical hierarchy rules and constraints before performing the segmentation process. These anatomical relationships (such as the fixed sequential order of retinal layers from NFL to choroid) are encoded as prior knowledge that guides and constrains the deep neural network predictions, ensuring that speed gains do not compromise diagnostic reliability.
Solution Approach 2:
The system changes parameters by adjusting the probability thresholds and confidence levels for layer identification based on anatomical constraints. When the deep neural network produces predictions that violate anatomical hierarchy, the system modifies the decision parameters to favor predictions that conform to known anatomical relationships, thereby maintaining reliability while preserving fast automated processing.
3Device complexity
If deep neural network predicts probability for each pixel independently, then the computational complexity is reduced, but the manufacturing precision of boundary lines deteriorates due to local errors
Solution Approach 1:
The patent merges independent pixel-level predictions by combining them with anatomical hierarchy constraints and spatial context information. Instead of treating each pixel prediction in isolation, the system integrates multiple sources of information including neighboring pixel probabilities, anatomical rules, and spatial relationships to produce more accurate boundary line detections while maintaining computational efficiency through the use of efficient constraint propagation algorithms.
Data Source
AI summary
A method for differentiating retinal layers includes obtaining an optical coherence tomography retinal layer image of an eye to be examined and inputting it into a deep neural network, calculating probability P that each pixel is included in each layer from the inputted layer image, calculating probability P1(x,y) that a pixel is in or above (k−1)-th layer and probability P2(x,y) that the pixel is in or below k-th layer from probabilities P(x,y) that the pixel is included in each layer, calculating probability Fk−1,k(x,y) that the pixel becomes a boundary line between the (k−1)-th layer and the k-th layer from P1(x,y) and P2(x,y), obtaining a node cost Ck−1,k(x,y) of the pixel by multiplying Fk−1,k(x,y) with an image brightness change ΔI at the pixel, and determining the boundary line between the (k−1)-th and k-th layers by connecting coordinates (x,y) of the pixels where Ck−1,k(x,y) calculated for each position x is the maximum.