Fundus Image Classification Using Cascaded Feature Fusion

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing image processing methods for diabetic retinopathy screening, particularly those relying on manual feature extraction and deep learning, fail to fully utilize fundus photograph information, leading to limited accuracy and robustness in distinguishing diabetic retinopathy from other retinopathies.

Innovation Solution

An image processing method utilizing cascaded feature extractors and classifiers, including SE-ResNet, ResNext, and InceptionV3 networks, to extract and fuse multiple features from fundus photographs, followed by Xgboost classification, enhancing the screening performance for diabetic retinopathy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If manual feature extraction methods are used, then the system complexity is low, but the screening accuracy and robustness are limited

Engineering Contradiction:
Improvescreening accuracyVSAvoidsystem complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the feature extraction process into multiple specialized extractors (first feature extractor for diabetic retinopathy features, second feature extractor for other retinopathy features, third feature extractor for normal fundus features). Each extractor focuses on specific features, and their outputs are fused to achieve comprehensive and accurate classification, resolving the contradiction between accuracy and complexity.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent changes the parameter of feature extraction by transitioning from manual feature extraction to deep learning-based automatic feature extraction using cascaded neural networks. This parameter change enables the system to automatically learn hierarchical features from raw images, significantly improving screening accuracy while the modular architecture manages the increased complexity.

Inventive Principle:
Principle #35Parameter changes

2Reliability

If single-model classification is used, then the device complexity is low, but the ability to distinguish diabetic retinopathy from other retinopathies is limited

Engineering Contradiction:
Improverobustness in distinguishing diabetic retinopathyVSAvoidclassification system complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent divides the classification task into three separate classification models, each specialized for a specific category: diabetic retinopathy, other retinopathies, and normal fundus. This segmentation allows each classifier to focus on specific features and patterns, improving the reliability of distinction between different conditions while managing complexity through modular design.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces feature fusion as an intermediary step between feature extraction and classification. The feature fusion module combines features from multiple extractors before passing them to the classification models, enabling the system to leverage complementary information from different feature sources and improve robustness in distinguishing diabetic retinopathy from other conditions.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Loss of information

If deep learning methods are used, then the feature extraction capability is improved, but the ability to fully utilize fundus photograph information is still limited

Engineering Contradiction:
Improveutilization of fundus photograph informationVSAvoidfeature extraction complexity
Core Design Contradiction:
Loss of informationVSDevice complexity

Solution Approach 1:

The patent segments the deep learning feature extraction into multiple parallel extractors, each designed to capture different aspects of fundus photograph information. This segmentation allows the system to extract diverse high-level features simultaneously, reducing information loss while the modular structure manages the complexity of processing multiple feature streams.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP4109328B1Image processing method and apparatus, screening system, and computer-readable storage medium
Publication Date: 2025.11.05 BOE TECHNOLOGY GROUP CO LTD
  • EP4109328B1 patent drawingFigure 1
  • EP4109328B1 patent drawingFigure 2~4A
  • EP4109328B1 patent drawingFigure 4B~5A

AI summary

The present disclosure relates to an image processing method and apparatus, a screening system, and a computer-readable storage medium. An image processing method includes: acquiring an image to be processed; extracting a first feature of the image to be processed by using a first feature extractor; performing first classification on the image to be processed according to the first feature by using a first classifier; under the condition that a classification result is a first result, extracting a second feature of the image to be processed by using a second feature extractor; performing second classification on the image to be processed according to the second feature by using a second classifier; and outputting the classification result of the second classification.