Learning Model Training via Mask State Change for Medical Imaging
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing methods for training segmentation models require a large amount of correct answer data, which is time-consuming to create, and often rely on simulation images that can introduce domain gaps, particularly in applications like medical imaging where accurate anatomical structure identification is crucial.
Innovation Solution
A method to efficiently generate learning data by creating pairs of masks representing normal and abnormal regions through state changes in a normal object's mask, allowing for the training of models that can differentiate between these regions, thereby reducing the need for extensive correct answer data and minimizing domain gaps.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a large amount of correct answer data is collected for training a segmentation model, then the model training accuracy is improved, but the data creation time is excessively long
Solution Approach 1:
The patent applies preliminary action by pre-defining normal regions in masks before training. The system creates masks where normal regions are predetermined, then generates abnormal region masks by modifying these pre-established normal masks. This approach eliminates the need to manually create extensive annotated training data, as the training data is synthesized from the pre-defined normal masks through automated modification processes.
Solution Approach 2:
The patent uses copying by generating abnormal region masks based on copies of normal region masks. Instead of creating entirely new annotated data for each training sample, the system copies existing normal masks and modifies them to create synthetic training pairs, dramatically reducing the time required to generate large volumes of training data while maintaining model training accuracy.
2Ease of manufacture
If simulation images are used instead of real masks for training, then the data creation cost is reduced, but a domain gap is introduced affecting model performance
Solution Approach 1:
The patent introduces masks as an intermediary between real images and simulation data. Instead of directly using simulation images that create domain gaps, the system uses masks (particularly normal region masks) as an intermediate representation. Training data is synthesized by combining these mask intermediaries with image data, allowing the model to learn from structured regional information without the domain gap issues of direct simulation image usage.
3Measurement precision
If extensive correct answer data is manually created, then the segmentation model accuracy is improved, but the manufacturing complexity of training data increases
Solution Approach 1:
The patent applies self-service by enabling the system to automatically generate its own training data from normal region masks. Instead of requiring manual annotation of each training sample, the system uses automated processes to create abnormal region masks by modifying normal masks, generating synthetic training pairs without human intervention. This self-service approach maintains high training accuracy while eliminating the complex manual data manufacturing process.
Data Source
AI summary
An object of the present invention is to provide a manufacturing method of a learning model, a learning model, an estimation method, an image processing system, and a program, which are capable of efficiently generating a large amount of learning data and of performing training on a learning model to which the efficiently generated large amount of learning data is applied.A manufacturing method of a learning model includes acquiring a region of a normal object included in a processing target image as a first mask, generating a second mask by changing a state of the first mask, and performing training to estimate a difference between the second mask and the first mask as a region-of-interest or performing training to estimate the first mask from the second mask, using the first mask and the second mask as learning data.