Fundus Image Detection Using Multi-Model Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning technologies for fundus image detection are limited in accurately detecting a wide variety of features in fundus images due to their focus on single or few associated features, leading to inaccurate detection results and increased computational requirements.
Innovation Solution
A method using multiple classification models, where a first classification model detects high-saliency features and second classification models detect finer features in specific regions, with parallel processing and decision modules to improve accuracy and efficiency, and image quality analysis to filter out substandard images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a single deep learning model is trained to detect multiple fundus features, then the model complexity increases and computational load sharply increases, but the detection accuracy for each feature may be compromised
Solution Approach 1:
The patent divides the fundus image detection task into multiple independent detection models, each specialized for specific feature types (e.g., one model for hemorrhage detection, another for exudate detection). This segmentation allows each model to focus on particular features, maintaining high detection accuracy while avoiding the complexity of a single universal model that must handle all features.
Solution Approach 2:
The patent employs batch normalization layers and pooling layers that are universally applicable across different detection models. These components serve multiple functions: batch normalization standardizes input data across different models, and pooling layers extract common features that can be reused. This multi-functionality reduces overall system complexity while maintaining versatility across different feature detection tasks.
2Measurement precision
If multiple separate models are used to detect different fundus features, then detection accuracy for each feature improves, but the computational load and processing time increase
Solution Approach 1:
The patent applies batch normalization and pooling layers as preliminary processing steps before the main detection processes. Batch normalization pre-processes the input data by standardizing it, and pooling layers pre-extract common features that will be useful across multiple detection models. This preliminary action reduces the computational burden on subsequent detection models, improving overall efficiency while maintaining high detection accuracy.
Solution Approach 2:
The patent merges common processing components (batch normalization layers and pooling layers) that are shared across multiple detection models. Instead of having completely separate models for each feature, the system combines common feature extraction and normalization operations, reducing redundant computations and improving processing efficiency while maintaining specialized detection capabilities for each feature type.
3Use of energy by moving object
If traditional detection methods are used for fundus images, then the computational requirements are lower, but the detection accuracy for complex features decreases
Solution Approach 1:
The patent transforms the input fundus images through batch normalization, which changes the statistical parameters (mean and variance) of the pixel data. This parameter transformation standardizes the input distribution, enabling more efficient processing by subsequent detection models. The pooling layers also transform the data by aggregating information across spatial regions, reducing dimensionality while preserving important features. These parameter changes enable accurate detection of complex features with optimized computational requirements.
Data Source
Figure 1~3
Figure 4~5
Figure 6
AI summary
The present invention discloses a method, apparatus, and system for analysing a fundus image based on machine learning. The method comprises the following steps: acquiring a fundus image to be analysed; classifying the entire region of the fundus image by using a first classification model to determine whether the fundus image contains any first feature; classifying a specific region in the fundus image by using at least one-second classification model to determine whether the fundus image contains any second feature, wherein the saliency of the first features is greater than that of the second features, and determining a analysis result at least according to the classification results of the first classification model and the second classification model.