Generative Model Training for Medical Images via Automated Labeling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for training generative models for medical images require expertise for selecting body parts, classifying training data, and generating models for each body part, which is time-consuming and costly.
Innovation Solution
A method and apparatus for training a generative model using 2D CT slice real images and label data, where a machine learning Body Part Regression (BPR) model estimates BPR scores, allowing for automated labeling and sampling of data to train a single model capable of generating medical images for multiple body parts based on user input, such as BPR scores or text indicating anatomical locations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a single generative model is trained to handle multiple body parts, then device complexity is reduced, but measurement precision for specific anatomical locations may be compromised
Solution Approach 1:
The training data is segmented by anatomical location using BPR scores, and the model learns to generate images for specific body parts by processing queries with location-specific parameters. This allows a single model to handle multiple body parts while maintaining precision through parameterized segmentation.
Solution Approach 2:
The model uses BPR score parameters to control which anatomical locations are generated. By changing the BPR score parameter in the input query, the model can selectively generate images for different body parts, maintaining precision through parameter-driven control rather than requiring separate models.
2Loss of time
If automated labeling with machine learning BPR model is used, then loss of time is reduced, but measurement precision of anatomical location may be affected
Solution Approach 1:
The system uses an automated machine learning BPR model to generate labels for training data without requiring manual annotation by experts. This self-service approach dramatically reduces labeling time while the model is trained to maintain precision through iterative optimization on available data.
3Productivity
If sampling is performed based on specific BPR score ranges, then productivity is improved, but loss of information about other anatomical locations occurs
Solution Approach 1:
The training process uses sampling with specific BPR score ranges to focus on particular anatomical locations during different training phases. This partial action approach improves training efficiency for specific body parts while the overall training regimen ensures comprehensive coverage across all anatomical locations.
Data Source
AI summary
Provided is a method for training a generative model for medical images associated with a plurality of body parts, which is performed by one or more processors and includes receiving training medical image data, acquiring label data associated with the training medical image data, and training a generative model for medical images based on the training medical image data and the acquired label data.