Medical Image Learning with Pseudo Lesion Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Obtaining a large amount of medical image data with correct answers is challenging, especially for original images, which are often deleted, making it difficult to secure robustness in lesion detection using machine learning.
Innovation Solution
A medical image learning method that combines pre-task learning through self-supervised learning on both original and processed images, followed by target-task learning using augmented data with pseudo lesions, to enhance lesion detection accuracy with limited correct-answer data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If original medical images are used for machine learning, then robustness in lesion detection is improved, but availability of data is worsened because original images are often deleted after processing
Solution Approach 1:
The system performs preliminary action by creating and storing pseudo-ground truth data for original images before they are deleted. Virtual lesions are generated on original images using transformation models, and this pseudo-ground truth is saved for later use in machine learning training, ensuring data availability even after original images are removed.
Solution Approach 2:
The system creates copies of original images by generating pseudo-ground truth data that replicates the essential lesion information. These pseudo-labels serve as substitutes for the original images, allowing the model to train on original image characteristics without requiring the original images to be retained.
2Quantity of substance
If processed medical images are used for machine learning, then availability of data is improved, but robustness in lesion detection is worsened due to vendor-specific image processing differences
Solution Approach 1:
The system introduces an intermediary approach by using processed images as the base but applying style transfer or normalization techniques to remove vendor-specific characteristics. This intermediary processing creates a standardized representation that maintains data availability while reducing the harmful effects of vendor-specific variations on model robustness.
3Quantity of substance
If a large amount of data with correct answers is obtained for machine learning, then training quality is improved, but cost and time for data preparation are worsened due to manual annotation requirements
Solution Approach 1:
The system implements self-service by automatically generating pseudo-ground truth data through virtual lesion generation and transformation models. The machine learning model creates its own training data by synthesizing lesion information on original images, eliminating the need for manual annotation by radiologists while still providing sufficient training data.
Solution Approach 2:
The system performs preliminary action by pre-generating pseudo-ground truth data and storing it for future use. Virtual lesions are created and saved as training data before actual model training begins, allowing the system to have ready-to-use annotated data without requiring time-consuming manual annotation during the training phase.
Data Source
AI summary
A medical image learning method includes: pre-task learning in which a model performs self-supervised learning, based on first medical image data; and target-task learning in which the model that has learned in the pre-task learning learns to detect a lesion, based on second medical image data that has a correct answer. The first medical image data includes original image data on which predetermined image processing is not performed and/or processed image data on which the predetermined image processing has been performed. The second medical image data includes the original image data.