Semi-supervised Fundus Image Quality Assessment Using IR Tracking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current automated fundus image quality assessment tools face challenges in efficiently differentiating high-quality from low-quality images, particularly in retinal tracking applications, due to the need for vast labeled data and reliance on subjective or unreliable objective methods, which are costly and resource-intensive.
Innovation Solution
A deep learning-based approach that uses initial automated labeling and active learning to continuously update a machine learning model for fundus image quality assessment, leveraging similarity measures between images to select high-quality reference images and retrain the model in real-time, enabling accurate and efficient image quality evaluation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning methods are used for image quality assessment, then accuracy of quality assessment is improved, but vast amounts of labelled data are required which increases resource consumption
Solution Approach 1:
The patent applies preliminary action by using automated labeling methods (such as motion tracking algorithms or other assessment algorithms) to pre-label images before deep learning training. This preliminary labeling creates an initial training set that satisfies the performance requirements of the deep learning model, reducing the need for extensive manual labeling while still enabling accurate quality assessment.
2Measurement precision
If manual labeling is performed to train the learning model, then accuracy of quality assessment is improved, but the process becomes costly and resource-intensive
Solution Approach 1:
The patent implements self-service by enabling the system to automatically generate labeled training data through automated labeling algorithms. The system uses motion tracking or other assessment algorithms to self-label images, eliminating the need for expensive and time-consuming manual labeling while maintaining sufficient accuracy for deep learning training.
Solution Approach 2:
The patent uses an intermediary approach by introducing automated labeling algorithms as a mediator between raw images and the deep learning model. These algorithms generate intermediate labeled data that bridges the gap between unlabelled images and the labeled training data required by deep learning, reducing both cost and resource consumption.
3Reliability
If the learning model is retrained continuously with new data, then performance of quality assessment is improved, but computational resources and time are consumed
Solution Approach 1:
The patent applies continuity of useful action by implementing continuous retraining of the learning model using newly acquired labeled images from real-world operation. The system continuously incorporates new data into the training set, maintaining and improving model performance over time without interruption to the quality assessment functionality.
Solution Approach 2:
The patent implements periodic action by retraining the learning model at scheduled intervals or when specific conditions are met (such as accumulating a certain number of new labeled images). This periodic retraining approach balances the need for improved performance with the constraints of computational resources and time, preventing excessive retraining while maintaining model reliability.
Data Source
AI summary
System/Method/Device for labelling images in an automated manner to satisfy a performance of a different algorithm and then applying active learning to learn a deep learning model which would enable ‘real-time’ operation of quality assessment and with high accuracy.