Deep Learning Semantic Segmentation of Gonioscopic Images

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current methods for analyzing RGB digital gonioscopic images of the irido-corneal interface layers are limited by their subjective nature, requiring substantial expertise and being inefficient for follow-up image analysis, which hampers accurate assessment of glaucoma risk and angle closure.

Innovation Solution

A computer-implemented method using a deep learning neural network for automatic semantic segmentation of RGB digital gonioscopic images, involving a training step with pre-processed images and corresponding ground truth data maps, and a per-pixel classification step to accurately distinguish anatomical layers.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If manual gonioscopy examination is used, then clinical expertise and subjective assessment are obtained, but time consumption and requirement for substantial expertise increase

Engineering Contradiction:
Improveassessment accuracyVSAvoidexamination time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent replaces the manual mechanical gonioscopy examination system with an automated deep learning-based image analysis system. The neural network automatically segments and classifies anatomical layers in gonioscopic images, substituting the clinician's manual inspection and subjective assessment with an objective computational system that provides precise measurements without requiring substantial clinical expertise for interpretation.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

2Extent of automation

If deep learning neural network is used for semantic segmentation, then automation and objectivity are improved, but algorithm complexity and training requirements increase

Engineering Contradiction:
Improveanalysis automationVSAvoidalgorithm complexity
Core Design Contradiction:
Extent of automationVSDevice complexity

Solution Approach 1:

The patent applies preliminary action by implementing an extensive training phase before the actual analysis. During this training step, the deep learning neural network is trained on a large dataset of annotated gonioscopic images with ground truth labels for each anatomical layer. This preliminary training enables the network to learn complex patterns and relationships, allowing it to perform accurate automated segmentation and classification during actual clinical use without requiring complex real-time adjustments.

Inventive Principle:
Principle #10Preliminary action

3Productivity

If automated image analysis is implemented, then productivity and follow-up efficiency are improved, but measurement precision and anatomical distinction accuracy may deteriorate

Engineering Contradiction:
Improveanalysis throughputVSAvoidlayer distinction accuracy
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The patent replaces subjective human assessment with an objective deep learning-based measurement system. The neural network performs pixel-wise classification to distinguish anatomical layers with quantitative precision, providing consistent and reproducible measurements that are not subject to human variability. The system outputs precise segmentation masks and classification probabilities for each anatomical structure, enabling accurate automated assessment of angle closure and trabecular meshwork status.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Data Source

PatentEP4179498B1A computer-implemented method for executing semantic segmentation of RGB digital images
Publication Date: 2025.03.05 NIDEK CO LTD
  • EP4179498B1 patent drawingFigure 1~2
  • EP4179498B1 patent drawingFigure 3~4
  • EP4179498B1 patent drawingFigure 5~7(c)

AI summary

A Computer-implemented method for executing a deep learning semantic segmentation of RGB digital gonioscopic images representing the irido-corneal interface layers. The method provides a training step and a pixel-wise classification step, both steps providing to enter said RGB digital gonioscopic images as input of a deep neural network (DNN) comprising an encoder and two decoders, comprising a plurality of levels.