Pseudo-Label Filtering for Medical Image Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for generating training data for deep neural networks in data-scarce fields, such as the medical field, suffer from noise in pseudo-labels due to unfiltered model outputs, leading to potential deterioration in pseudo-label quality and reduced learning model performance.
Innovation Solution
A training data generation method that involves acquiring unlabeled images and texts, using learning models to generate pseudo-labels, and filtering these pseudo-labels based on information from both image and text analysis to improve accuracy, specifically for medical images by refining pseudo-labels for lesion regions, organ names, and disease types.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If pseudo-labels are generated without filtering, then the quantity of training data is increased, but the quality of pseudo-labels deteriorates due to noise
Solution Approach 1:
The patent applies preliminary action by performing filtering processing on pseudo-labels before they are used for training. The system generates pseudo-labels from unlabeled data using a trained model, then filters these pseudo-labels to remove noise and low-quality entries before incorporating them into the training dataset. This preliminary filtering ensures that only high-quality pseudo-labels are used, maintaining data quality while still expanding the training dataset size.
2Measurement precision
If filtering processing is applied to pseudo-labels, then the quality of training data is improved, but the device complexity increases
Solution Approach 1:
The patent implements feedback by using the trained model to evaluate and filter its own generated pseudo-labels. The system creates a feedback loop where the model's output is assessed against established criteria, and only pseudo-labels that meet quality thresholds are retained for training. This self-evaluating feedback mechanism improves data quality without requiring complex external filtering systems.
3Measurement precision
If more filtering criteria are applied, then the accuracy of pseudo-labels is improved, but the productivity of training data generation is reduced
Solution Approach 1:
The patent applies parameter changes by adjusting the stringency of filtering criteria based on the specific application domain and data characteristics. The system can modify filtering thresholds and criteria to balance accuracy and productivity requirements. For example, in critical applications like medical imaging, more stringent filtering is applied, while in less critical applications, more lenient filtering maintains higher productivity.
Data Source
AI summary
There are provided a training data generation device, a training data generation method, and a learning device. First images having no label and first texts attached to the first images having no label are acquired, first information on at least one of a region, a type, or a state of the first image having no label is acquired from the first image having no label, second information on at least one of a region, a type, or a state of the first image having no label is acquired from the first text, an output of a first learning model to which the first images having no label are input is acquired as first pseudo-labels of the first images having no label, and the first pseudo-labels are filtered on the basis of the first information and the second information.