3D Deep Learning for Retinal Layer Boundary Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current OCT systems face challenges in quantitative analysis of retinal layers due to the lack of well-defined boundaries between layers like the choroid and sclera, and the transparency of these layers, leading to inefficiencies in segmentation techniques.
Innovation Solution
A machine learning system is trained using 2D en face images generated from 3D OCT volume data to identify texture differences between retinal layers, allowing for the generation of probability maps and accurate boundary identification between layers like the choroid and sclera, using techniques such as shortest path search and thresholding.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional boundary detection methods (graph-search theory) are used to identify retinal layer boundaries, then the method can handle well-defined boundaries, but it fails when boundaries are poorly defined or layers are transparent (e.g., choroid-sclera interface, GCL-IPL interface)
Solution Approach 1:
The patent transforms the boundary detection problem from relying on intensity gradients to relying on textural parameters. By training a convolutional neural network to recognize textural patterns in OCT images, the system can identify boundaries in transparent layers where traditional intensity-based methods fail. The network learns textural features that characterize different retinal layers, enabling accurate boundary detection even when optical boundaries are poorly defined.
2Stability of the object's composition
If complex preprocessing steps (3D model building, thin plate spline) are used to construct the choroid-sclera interface, then the interface can be made continuous and smooth, but the assumption of continuity contradicts the biological reality of no clear physical boundary
Solution Approach 1:
The patent extracts and removes the problematic thin plate spline smoothing step from the traditional pipeline. Instead of forcing continuity through mathematical smoothing, the system directly segments the choroid and sclera based on learned textural characteristics. This extraction eliminates the contradiction between mathematical continuity and biological reality, allowing the interface to be represented as a probabilistic transition zone rather than a forced continuous curve.
3Device complexity
If 2D cross-sectional images (B-scans) are used for layer segmentation, then the process is computationally simpler, but the lack of contextual information reduces segmentation accuracy
Solution Approach 1:
The patent transitions from analyzing 2D B-scan cross-sections to processing 3D volumetric OCT data. By incorporating the third dimension (depth information across multiple B-scans), the convolutional neural network gains access to contextual information from surrounding tissue volumes. This dimensional enhancement allows the network to distinguish retinal layers more accurately by considering spatial relationships and textural patterns in three dimensions, resolving the trade-off between computational simplicity and segmentation accuracy.
Data Source
AI summary
A machine learning model is trained to identify the texture difference between the different layers of a multilayer object. By training with data in full 3D space, the resulting model is capable of predicting the probability that each pixel in a 3D image belongs to a certain layer. With the resulting probability map, comparing probabilities allows one to determine boundaries between layers, and/or other properties and useful information such as volume data.