Fundus Image Classification via Regional Cues and Interpretation Text
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current convolutional neural networks (CNNs) for classifying fundus images are limited in accurately localizing lesions and distinguishing individual findings, leading to suboptimal diagnostic performance and labor-intensive data collection due to biased feature extractors and passive annotations.
Innovation Solution
A method that uses regional annotations to train a CNN architecture, enabling better localization and classification of fundus image findings by generating interpretation text and providing it to external entities, while improving classification performance through guidance from regional cues.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If hand-crafted feature extractors are used for segmentation, then the detection process is simpler, but the detection accuracy is limited due to human designer bias
Solution Approach 1:
The patent replaces hand-crafted mechanical feature extraction with a neural network-based automated feature learning system. The CNN automatically learns optimal features from data, substituting the manual mechanical process with an intelligent system that adapts to complex patterns without human bias.
Solution Approach 2:
The patent transforms the approach by changing from fixed hand-crafted parameters to dynamic parameters learned through training. The neural network adjusts its internal parameters (weights and biases) based on training data, enabling adaptive feature extraction that improves detection accuracy while maintaining implementation feasibility.
2Measurement precision
If CNN for segmentation or detection is used, then the detection accuracy is improved, but the data collection procedure becomes very expensive due to labor-intensive passive annotations
Solution Approach 1:
The patent applies preliminary action by pre-processing fundus images to automatically generate candidate lesion regions before the main detection task. This preliminary segmentation reduces the annotation burden by providing pre-identified regions that require less manual verification, thereby improving data collection efficiency while maintaining detection accuracy.
Solution Approach 2:
The patent introduces an intermediary automated segmentation module that bridges the gap between raw images and final detection. This intermediary process generates preliminary annotations that guide the main detection algorithm, reducing the need for extensive manual annotations while preserving high detection accuracy.
3Productivity
If CNN is trained to directly derive diagnosis content, then the diagnostic speed is improved, but the ability to distinguish individual findings and localize lesions is insufficient
Solution Approach 1:
The patent segments the diagnostic task into multiple specialized components: one network for classification (diagnosis) and another for segmentation (lesion localization). This segmentation allows each network to specialize in its specific function, improving both diagnostic speed and lesion localization accuracy simultaneously.
Solution Approach 2:
The patent adds another dimension to the diagnostic process by introducing spatial localization as a separate output dimension alongside classification. Instead of a single classification output, the system produces both diagnostic labels and spatial masks, enabling simultaneous improvement in speed and localization precision through multi-dimensional processing.
Data Source
AI summary
The present invention relates to a method for classifying a fundus image and a device using same. Specifically, according to the method of the present invention, a computing device acquires a fundus image of a subject, generates classification information of the fundus image, generates an interpretation text on the basis of the classification information, and provides the interpretation text to an external entity.