Hyperreflective Foci Detection With OCT Segmentation and Size Thresholds
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing computational-based models struggle to accurately segment and quantify hyperreflective foci (HRF) in retinal OCT scans due to challenges such as inadequate training data, human error, and lack of a unifying standard, often misidentifying small hyperreflective entities like hard exudates and speckle noise.
Innovation Solution
A machine-learning-based HRF segmentation and classification pipeline using a U-Net architecture and image-processing algorithms to segment and classify hyperreflective entities, defining HRF as entities with a diameter of 50 μm or less, and utilizing the ETDRS grid for volumetric measurements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If computational-based models are used to segment HRF, then the ability to identify and study HRF is improved, but the models struggle to distinguish HRF from retinal blood vessels, hard exudates, and speckle noise
Solution Approach 1:
The patent divides the retinal image analysis into multiple segmentation stages: first segmenting hyperreflective material broadly, then further segmenting into specific categories (HRF, hard exudates, blood vessels) based on additional features like diametral size, location, and morphological characteristics. This multi-level segmentation approach resolves the contradiction by breaking down the complex classification task into manageable steps, improving both accuracy and reliability.
Solution Approach 2:
The patent applies different classification criteria and features to different regions and types of hyperreflective entities. Specifically, it uses diametral size thresholds (e.g., HRF typically <50μm), location relative to retinal layers, and local morphological features to differentiate HRF from other structures. This localized quality assessment improves distinction accuracy without requiring a single unreliable universal model.
2Measurement precision
If manual annotation of HRF is performed for training data, then training data quality is improved, but the process is time-consuming, costly, and susceptible to immense human error
Solution Approach 1:
The patent implements preliminary automated preprocessing steps that prepare images for annotation, including initial detection of hyperreflective regions and generation of candidate segments. This preliminary action reduces the burden on annotators, allowing them to focus on verifying and refining rather than creating annotations from scratch, thereby improving data quality while reducing time and error.
Solution Approach 2:
The patent introduces an intermediary automated detection system that generates initial annotations or candidate regions, which then serve as input for manual verification. This intermediary layer filters out obvious errors and highlights ambiguous cases, making the manual annotation process more efficient and accurate while reducing overall time investment.
3Quantity of substance
If existing computational models identify any material less than 100 microns as HRF, then the quantity of detected entities is improved, but the accuracy of HRF identification deteriorates due to misidentification of hard exudates and speckle noise
Solution Approach 1:
The patent changes multiple parameters simultaneously to improve identification accuracy: it refines the size threshold (e.g., HRF <50μm rather than <100μm), adds location parameters (depth within retinal layers), and incorporates morphological parameters (shape, texture, boundary characteristics). This multi-parameter approach maintains sensitivity to small entities while improving specificity to true HRF.
Solution Approach 2:
The patent employs dynamic classification thresholds and criteria rather than fixed rules. The classification of an entity as HRF depends on multiple dynamic factors including its size distribution pattern, location relative to retinal layers, and comparison with surrounding tissue characteristics. This dynamic approach allows accurate identification across varying conditions while filtering out false positives.
Data Source
AI summary
A method for identifying hyperreflective foci (HRF) in an eye of a patient includes accessing one or more optical coherence tomography (OCT) scans of a retina of the eye of the patient, and inputting the one or more OCT scans into one or more machine-learning models trained to segment the one or more OCT scans to identify a set of hyperreflective entities detectable from the one or more OCT scans. The method further includes determining, based on the segmented one or more OCT scans, one or more diametral measurements corresponding to each of the identified set of hyperreflective entities, and identifying hyperreflective foci (HRF) in the retina of the eye of the patient based on whether at least one of the one or more diametral measurements satisfy a diametral threshold.