Semi-Supervised Medical Image Segmentation via Pseudo-Label Refinement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current medical image segmentation methods, particularly those using supervised deep learning and semi-supervised deep learning, face challenges with high labeling costs, low versatility, and accuracy issues when the amount of labeled data is small, leading to suboptimal performance in medical image processing.
Innovation Solution
A medical image processing method that involves training a deep neural network using labeled data, performing weak and strong data augmentations on unlabeled data to generate pseudo-labels, and updating the network based on both labeled and unlabeled data, enhancing the accuracy of segmentation results through optimized pseudo-label determination and network optimization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If semi-supervised deep learning with pseudo-labels is used to reduce labeling costs, then the amount of labeled data required is reduced, but the accuracy of prediction results deteriorates due to low quality pseudo-labels
Solution Approach 1:
The patent segments the image processing task into multiple stages: first performing processing on a first image to generate initial results, then using those results to guide processing of a second image. This staged approach allows the system to leverage limited labeled data effectively while maintaining accuracy through progressive refinement rather than attempting to process all images simultaneously with insufficient labeled examples.
Solution Approach 2:
The patent performs preliminary processing on a first image before using those results to process a second image. The first image processing establishes a baseline and generates reference information that improves subsequent processing accuracy. This preliminary action allows the system to accumulate quality information from limited labeled data before applying it to new images.
2Reliability
If consistency regularization is applied to improve prediction stability, then prediction consistency is improved, but device complexity increases due to multiple neural network structures and pretext tasks
Solution Approach 1:
The patent extracts and utilizes only the essential information from the first image processing results to guide second image processing. Rather than implementing full consistency regularization with multiple networks and pretext tasks, the system extracts key diagnostic information and uses that to improve predictions, achieving reliability without the complexity of comprehensive consistency frameworks.
3Quantity of substance
If data augmentation is applied to unlabeled data to improve training accuracy, then the amount of effective training data is increased, but the complexity of the training process increases
Solution Approach 1:
The patent applies data augmentation selectively and partially - transforming only the first image and using those transformed versions to guide processing of the second image. This partial application of augmentation increases effective training data without requiring comprehensive augmentation of all datasets, thereby reducing overall training complexity while still benefiting from enhanced data diversity.
Data Source
AI summary
A medical image processing method according to an embodiment of the present disclosure includes: training a deep neural network by using labeled image data; obtaining a first augmented image by carrying out a weak data augmentation on unlabeled image data; performing a predicting process on the first augmented image by using the deep neural network and determining whether each of the pixels in the first augmented image is able to serve as a pseudo-label on the basis of prediction information of the pixel; obtaining a second augmented image by carrying out a strong data augmentation on the first augmented image; training the deep neural network by using the second augmented image and the pseudo-labels; and updating the deep neural network on the basis of training results of the labeled image data and the unlabeled image data and processing a medical image by using the updated deep neural network.