Retinal OCT Layer Segmentation With 2D-CRNN Inter-Slice Context
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing OCT retinal image segmentation methods struggle with anisotropic data and inter-B-scan gaps, leading to inaccurate and unreliable segmentation results, particularly in monitoring retinal thickness changes for conditions like multiple sclerosis.
Innovation Solution
An iterative registration and deep-learning-based segmentation method that utilizes 2D networks to extract intra-slice features and convolutional long-short-term memory networks for inter-slice context, incorporating longitudinal information by registering previous retinal layer segmentations as priors for current scans.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Volume of moving object
If interpolation between B-scans is used to fill gaps in OCT volumes, then 3D retinal images can be constructed, but segmentation accuracy deteriorates due to unreliable and inaccurate results caused by anisotropic data and inter-B-scan gaps
Solution Approach 1:
The patent segments the 3D OCT volume into individual 2D B-scans for processing, rather than attempting to interpolate across the entire volume. Each B-scan is processed independently with 2D segmentation networks, avoiding the propagation of interpolation errors through the 3D space while still constructing a complete 3D retinal structure from the segmented 2D slices.
Solution Approach 2:
The patent transitions from 3D volume processing to 2D B-scan processing by treating each B-scan as an independent segmentation problem. This dimensionality reduction eliminates the need for 3D interpolation while maintaining segmentation accuracy, as 2D networks can process each slice without being affected by the anisotropic gaps between B-scans.
2Reliability
If 3D convolution networks with isotropic kernels are applied to anisotropic OCT volumes, then inter-B-scan context can be utilized, but segmentation performance deteriorates due to inappropriate handling of inter-B-scan gaps
Solution Approach 1:
The patent applies different processing qualities to different spatial scales: 2D convolutional networks process local intra-slice features within each B-scan, while CRNN networks process inter-slice features across multiple B-scans. This local quality differentiation allows the system to handle anisotropic gaps appropriately by not applying isotropic 3D convolution kernels that would incorrectly interpolate across large inter-B-scan distances.
Solution Approach 2:
The patent uses a dynamic architecture where 2D encoders process individual B-scans and CRNN networks dynamically aggregate features across B-scans based on their sequential relationships. This dynamic feature aggregation allows the system to utilize inter-B-scan context without assuming uniform spacing, adapting to the anisotropic nature of OCT volumes by weighting and propagating features through the longitudinal dimension only when appropriate.
3Ease of manufacture
If 2D B-scan images are used for deep neural network segmentation, then processing simplicity is maintained, but 3D inter-B-scan context is lost leading to reduced segmentation accuracy
Solution Approach 1:
The patent merges 2D B-scan processing with 3D context utilization by combining 2D encoders that process individual B-scans with CRNN networks that aggregate features across B-scans. This merging allows the system to maintain the simplicity of 2D processing for each slice while incorporating 3D inter-B-scan context through the sequential feature aggregation in the CRNN, achieving both processing efficiency and improved accuracy.
Solution Approach 2:
The patent introduces CRNN networks as an intermediary between 2D B-scan processing and 3D volume segmentation. The CRNN acts as a mediator that takes 2D feature maps from individual B-scans, processes them in sequence to capture inter-B-scan relationships, and produces enhanced segmentation results that incorporate 3D context without requiring direct 3D convolution operations.
4Productivity
If standard OCT scanning is used, then data acquisition is efficient, but retinal thickness changes are concealed by noise and segmentation reliability deteriorates
Solution Approach 1:
The patent performs preliminary 2D segmentation on each B-scan before aggregating results across the volume. By segmenting individual B-scans first with 2D networks, the system establishes baseline segmentations that are then refined through CRNN-based inter-slice feature aggregation. This preliminary action allows noise reduction through temporal and spatial consistency checks across multiple B-scans while maintaining the efficiency of standard OCT acquisition.
Solution Approach 2:
The patent implements feedback mechanisms where segmentation results from one B-scan inform the processing of adjacent B-scans through the CRNN architecture. The CRNN uses sequential feature aggregation to propagate consistent segmentation boundaries across B-scans, providing feedback that reduces noise and improves reliability of retinal thickness measurements while maintaining efficient standard OCT scanning protocols.
Data Source
AI summary
Techniques for retinal layer segmentation are presented. The techniques include obtaining current optical coherence tomography (OCT) data for a retina; generating an estimated current en face image based on the current OCT data and an estimated current retinal layer segmentation; determining a registered previous retinal layer segmentation based on a previous retinal layer segmentation, a previous en face image, the estimated current en face image, and the estimated current retinal layer segmentation; updating the estimated current retinal layer segmentation using a deep neural network and based on the current OCT data and the registered previous retinal layer segmentation; repeating the generating, the determining, and the updating to obtain a current retinal layer segmentation as the estimated current retinal layer segmentation; and outputting a property of the retina determined at least in part from the current retinal layer segmentation.