Multi-task Machine Learning Model Retraining via User Feedback
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models struggle with generalization to unseen data, lack of user validation for proper behavior, and difficulty in providing feedback to improve performance, particularly in predicting coronary artery disease severity and related tasks.
Innovation Solution
The system allows for online retraining of machine learning models using user feedback on secondary tasks that can be directly validated, improving the model's performance for both primary and secondary tasks by incorporating user corrections and confidence measures.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a machine learning model is trained for a specific task using labeled training images, then prediction accuracy for that task is improved, but generalization to unseen data deteriorates
Solution Approach 1:
The patent trains a single machine learning model to perform multiple related tasks simultaneously (primary task of predicting hemodynamic index and secondary tasks of predicting location of measurement points, common image points, stenosis locations, and vessel segmentation). This multi-task learning approach improves generalization while maintaining accuracy across all tasks by sharing learned features and representations.
2Measurement precision
If a machine learning model predicts hemodynamic index directly, then primary task performance is improved, but user validation capability deteriorates
Solution Approach 1:
The patent introduces secondary tasks as intermediary predictions that bridge the gap between the primary hemodynamic index prediction and user validation. These secondary tasks predict anatomical features (location of measurement points, common image points, stenosis locations, vessel segmentation) that users can directly verify on the images, providing a validation pathway while maintaining the primary prediction capability.
3Measurement precision
If input features are complex medical image data, then prediction capability is improved, but user feedback capability deteriorates
Solution Approach 1:
The patent segments the complex prediction task into a primary task (hemodynamic index) and multiple secondary tasks (anatomical feature predictions). The secondary tasks produce outputs that are segmented and localized on the medical images, making them visually interpretable and easier for users to understand and provide feedback on, while the primary task maintains high prediction capability using the full complexity of the input data.
Data Source
AI summary
Systems and methods for retraining a trained machine learning model are provided. One or more input medical images are received. Measures of interest for a primary task and a secondary task are predicted from the one or more input medical images using a trained machine learning model. The predicted measures of interest for the primary task and the secondary task are output. User feedback on the predicted measure of interest for the secondary task is received. The trained machine learning model is retrained for predicting the measures of interest for the primary task and the secondary task based on the user feedback on the output for the secondary task.