Dual-Path CNN for Automatic EZ Loss Detection in SD-OCT Imaging
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for detecting ellipsoid zone loss in SD-OCT imaging for hydroxychloroquine retinal toxicity are subjective and prone to variability due to reliance on qualitative inspection and manual adjustments, leading to potential errors in diagnosing retinal toxicity.
Innovation Solution
A deep-learning based system using a dual-path convolutional neural network for scan-by-scan ellipsoid zone loss detection, combining horizontal and vertical projections to generate an enface EZ loss map, which automatically detects and quantifies EZ loss without relying on entire layer segmentation, enhancing objectivity and precision.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If qualitative inspection of individual OCT B-scans is used to identify EZ loss, then diagnostic capability is achieved, but subjectivity and variability are introduced
Solution Approach 1:
The patent replaces the manual mechanical inspection process with an automated deep learning system. The convolutional neural network automatically detects and segments EZ loss regions in OCT B-scans, eliminating human subjectivity and variability while maintaining high diagnostic accuracy through learned features from training data
Solution Approach 2:
The system enables self-service by allowing the algorithm to automatically perform EZ loss detection and quantification without requiring manual adjustments or expert intervention. The deep learning model independently processes OCT images, generates segmentation masks, and provides quantitative metrics, making the diagnostic process autonomous and consistent
2Measurement precision
If manual adjustments to algorithm-generated contours are performed, then segmentation accuracy is improved, but time consumption and operational complexity increase
Solution Approach 1:
The system achieves self-service by implementing an automated deep learning pipeline that performs EZ loss detection and segmentation without requiring manual contour adjustments. The convolutional neural network generates accurate segmentation masks directly from OCT B-scans, eliminating the time-consuming manual adjustment process while maintaining high segmentation accuracy through learned features
Solution Approach 2:
The patent introduces a deep learning model as an intermediary between the raw OCT images and the final segmentation results. This intermediary automatically learns the complex patterns of EZ loss and generates accurate contours without human intervention, bridging the gap between image input and diagnostic output efficiently
3Measurement precision
If retinal layer segmentation is performed to derive EZ loss metrics, then EZ loss quantification is achieved, but segmentation fails in the presence of disease requiring manual adjustments
Solution Approach 1:
The patent extracts EZ loss detection as a separate, independent task from the overall retinal layer segmentation process. Instead of relying on segmenting entire retinal layers and deriving EZ loss metrics, the system directly detects and segments EZ loss regions using a dedicated deep learning model, making the process more robust to disease-related segmentation failures
Solution Approach 2:
The deep learning model serves as an intermediary that directly processes OCT B-scans to identify EZ loss regions without requiring successful segmentation of intermediate retinal layers. This direct approach bypasses the vulnerability of traditional methods where layer segmentation failures propagate to EZ loss quantification errors
4Reliability
If entire retinal layer segmentation is performed to define ground truth, then comprehensive annotation is achieved, but time consumption and operational difficulty increase significantly
Solution Approach 1:
The patent extracts only the critical EZ loss regions for annotation rather than requiring segmentation of entire retinal layers. By focusing annotation efforts specifically on EZ loss boundaries in diseased regions, the system achieves comprehensive ground truth for the target pathology while dramatically reducing the time and complexity required compared to complete layer segmentation
Solution Approach 2:
The system applies local quality by concentrating annotation resources on the specific regions of interest (EZ loss areas) rather than uniformly annotating all retinal layers. This localized approach ensures high-quality ground truth where it matters most for diagnosis while reducing overall annotation burden through selective focus on pathological regions
Data Source
AI summary
Various embodiments for systems and methods for automatically detecting ellipsoid zone loss in SD-OCT imaging are disclosed herein.