Self-Supervised Model Pretraining for Lesion-Preserving Medical Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing self-supervised learning technologies struggle to effectively utilize unlabeled medical data for training models due to the loss of key lesion information in medical images during data augmentation, and the high cost and scarcity of labeled data hinder accurate model training.
Innovation Solution
A model generating device and method that combines self-supervised learning processes of reconstruction learning and contrastive learning, using different degrees of data augmentation to train a pre-trained model, allowing it to learn complete data information and retain lesion information, and requires minimal labeled data for fine-tuning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If strong data augmentation is used to expand training data diversity, then the quantity of training data is improved, but key lesion information is lost
Solution Approach 1:
The training process is segmented into two distinct phases: strong augmentation phase for learning complete data information, and weak augmentation phase for retaining lesion information. This segmentation allows each phase to optimize for its specific goal without compromising the other.
Solution Approach 2:
The data augmentation strength is dynamically adjusted based on the training phase. Strong augmentation (e.g., random masking, high-degree deformation) is applied in the first phase, then switched to weak augmentation (e.g., slight rotation, brightness adjustment) in the second phase, allowing the system to adapt to different learning objectives.
2Measurement precision
If supervised learning is used to train models with large amounts of labeled data, then model accuracy is improved, but training cost increases
Solution Approach 1:
The system uses self-supervised learning where the model learns from unlabeled data through self-supervision signals (reconstruction tasks and contrastive learning) without requiring expensive manual annotations. This eliminates the need for large amounts of labeled data while maintaining training effectiveness.
Solution Approach 2:
The model undergoes pre-training on unlabeled data using self-supervised learning before fine-tuning on small amounts of labeled data. This preliminary action allows the model to learn general features and patterns from abundant unlabeled data, reducing the need for extensive labeled data training later.
3Adaptability or versatility
If existing self-supervised learning is used to expand training data, then data diversity is improved, but lesion information is lost
Solution Approach 1:
The data augmentation strategy dynamically transitions from strong augmentation for diversity to weak augmentation for information retention. This dynamic adjustment ensures that lesion information is preserved when the model needs to learn from diverse data representations.
Solution Approach 2:
The augmentation parameters are changed between phases: strong augmentation with high transformation intensity is used initially, then parameters are adjusted to weak augmentation with low transformation intensity to preserve lesion characteristics while maintaining data diversity benefits.
Data Source
AI summary
A model generating device and method are provided. The device inputs strong data augmentation images corresponding to a plurality of sample images into an image restoration block in the self-supervised neural network to generate restoration inference vectors. The device generates a reconstructed image corresponding to each of the sample images based on the restoration inference vectors. The device calculates a reconstruction loss for each of the reconstructed images to train the image restoration block of the self-supervised neural network. The device inputs weak data augmentation images corresponding to the sample images into an image classification block in the self-supervised neural network to generate classification inference vectors. The device calculates a contrastive loss for the classification inference vectors based on clusters to train the image classification block of the self-supervised neural network.