Parallel Fuzzy CNN for Retinal Image Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional retinal image processing methods are limited by manual design, leading to inaccuracies and subjectivity, and face challenges in obtaining sufficient and diverse retinal image samples, which affects the generalization of deep learning models.
Innovation Solution
A method involving chaotic supply-demand algorithm-based image enhancement, hybrid virtual retinal image generation, and a parallel multi-layer decomposed interval type-2 intuitionistic fuzzy convolutional neural network for enhanced feature extraction and classification, integrating outputs from multiple models for improved accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual design methods are used for retinal image processing, then the processing can be performed with simple systems, but the accuracy and objectivity of the processing results deteriorate
Solution Approach 1:
The patent replaces manual design methods with automated deep learning models. Specifically, it uses interval type-2 intuitionistic fuzzy convolutional neural networks to automatically extract features and classify retinal images, eliminating the need for manual feature engineering while significantly improving processing accuracy and objectivity.
Solution Approach 2:
The patent introduces interval type-2 intuitionistic fuzzy parameters to enhance the model's ability to handle uncertainty in retinal image data. By using fuzzy logic parameters and interval mathematics, the system can better represent and process the inherent ambiguities in medical images, leading to improved classification accuracy.
2Reliability
If deep learning models are trained with limited retinal image samples, then the model development can proceed with available data, but the generalization capability and model performance deteriorate
Solution Approach 1:
The patent applies data augmentation techniques as a preliminary action before model training. It uses geometric transformations (rotation, flipping, scaling) and intensity transformations to artificially expand the training dataset from limited real retinal images, ensuring the model can generalize better to unseen data.
Solution Approach 2:
The patent creates synthetic copies of real retinal images through various transformation operations. These copied and transformed images serve as additional training samples, effectively multiplying the available data without requiring additional physical samples, thus improving model generalization with limited original data.
3Adaptability or versatility
If virtual retinal images are generated to supplement training data, then the sample diversity increases, but the distribution difference between real and virtual images may worsen processing performance
Solution Approach 1:
The patent applies different processing strategies to different parts of the data pipeline. Real images are used for certain training phases while virtual images supplement in others, with each type optimized for its specific role. This local differentiation allows the system to leverage the strengths of both real and synthetic data while mitigating their respective weaknesses.
Solution Approach 2:
The patent uses domain adaptation techniques as an intermediary mechanism to bridge the distribution gap between real and virtual retinal images. The domain adaptation layer acts as a mediator that aligns the feature distributions of synthetic and real images, ensuring they can be trained together without causing performance degradation.
4Measurement precision
If a single convolutional neural network model is used for retinal image classification, then the model structure remains simple, but the classification accuracy and robustness deteriorate
Solution Approach 1:
The patent segments the classification task into multiple parallel neural network models, each specializing in different aspects of retinal image analysis. These parallel interval type-2 intuitionistic fuzzy CNNs process images independently and their outputs are aggregated, allowing the system to achieve higher accuracy through diverse specialized processors rather than a single general-purpose model.
Solution Approach 2:
The patent merges the outputs of multiple parallel convolutional neural networks through ensemble learning techniques. By combining the classification results from several specialized models, the system achieves superior accuracy and robustness compared to any individual model, while the fuzzy logic aggregation handles the complexity of integrating multiple outputs.
Data Source
AI summary
A method for parallel processing of retinal images includes: optimizing an objective function with a chaotic supply-demand algorithm to enhance a real retinal image; synthesizing a virtual retinal image by a hybrid image generation method; establishing a parallel multi-layer decomposed interval type-2 intuitionistic fuzzy convolutional neural network model based on the virtual retinal image and the enhanced real retinal image; and integrating outputs from a plurality of parallel multi-layer decomposed interval type-2 intuitionistic fuzzy convolutional neural network models as a final classification result.