Fundus Oculi Image Analysis Model for Quality Assessment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for fundus oculi image analysis, particularly for diagnosing DR and macula lutea lesions, face challenges in accurately determining image quality due to reliance on overall edge histograms or vessel segmentation, which are insufficient for capturing all quality factors, leading to potential misdiagnoses and low reliability in automated systems.
Innovation Solution
A method utilizing a fundus oculi image analysis model that includes an image overall grade prediction sub-model and an image quality factor sub-model, employing multi-task learning with convolutional neural networks to analyze image features in both rectangular and polar coordinate systems, thereby predicting image gradability and extracting quality factors such as artifact, clarity, and position information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If manual feature extraction methods (overall edge histogram or vessel segmentation) are used to determine image quality, then the method is simple to implement, but the quality features of the overall image are not extracted sufficiently and multiple quality factors cannot be taken into account
Solution Approach 1:
The patent replaces manual feature extraction methods with a deep learning-based convolutional neural network system. The CNN automatically learns and extracts quality features from fundus images, substituting the mechanical/manual process of hand-crafted feature extraction with an intelligent automated system that can capture comprehensive quality factors including artifacts, clarity, and positional information simultaneously.
Solution Approach 2:
The patent employs multi-task learning where a single CNN model performs multiple quality assessment functions simultaneously - evaluating artifacts, clarity, and positional information in one unified framework. This multi-functional approach allows comprehensive quality factor extraction without requiring separate specialized methods for each quality dimension.
2Device complexity
If a single two-classification task is used to determine image quality, then the model training is simple, but the reliability of manual labeling is not high enough and affects training and network performance
Solution Approach 1:
The patent segments the image quality assessment task into multiple independent sub-tasks (artifact detection, clarity assessment, positional information evaluation) that are processed separately by different branches of the CNN. This segmentation allows each sub-task to be optimized independently while maintaining overall system reliability, avoiding the limitations of a single binary classification approach.
Solution Approach 2:
The patent transitions from a single-dimension binary classification (gradable/not gradable) to a multi-dimensional quality assessment framework that evaluates multiple quality factors simultaneously. This dimensional expansion enriches the training signal and improves prediction reliability by capturing the nuanced relationships between different quality attributes.
Data Source
AI summary
A method of fundus oculi image analysis includes acquiring a target fundus oculi image; analyzing the target fundus oculi image by a fundus oculi image analysis model determined by training to acquire an image analysis result of the target fundus oculi image; and the fundus oculi image analysis model includes at least one of an image overall grade prediction sub-model and an image quality factor sub-model. The method performs quality analysis on the target fundus oculi image by the fundus oculi image analysis model, and when the model includes the overall grade prediction sub-model, a prediction result of whether the target fundus oculi image as a whole is gradable can be acquired; when the model includes the image quality factor sub-model, the analysis result of the fundus oculi image quality factor can be acquired and the image analysis model is determined by extensive image training, and the reliability of the result of whether the image is gradable determined based on the above model is high.