Medical Data Processing Apparatus for Model Adaptation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models in medical examinations face challenges in balancing performance consistency across examinations, as offline learning models are inflexible and online learning models risk being affected by previous examination data.
Innovation Solution
A medical data processing apparatus that generates a derived model from a trained model using first medical data, and applies second medical data with different acquisition parameters to the derived model to produce third medical data, thereby optimizing inference results while maintaining performance consistency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If online learning type is used to update parameters every time an examination is performed, then the capability of machine learning inference model in individual examinations is improved, but the performance may be affected by preceding examinations undesirably
Solution Approach 1:
The patent segments the learning process into two distinct phases: offline learning for general model training and online learning for individual examination adaptation. This segmentation allows the system to maintain a stable base model while enabling personalized adjustments, resolving the contradiction between adaptability and reliability by isolating the risks of online learning to specific examination contexts only.
Solution Approach 2:
The patent introduces an intermediary mechanism (the derived model generation process) that mediates between the trained model and the examination data. This intermediary layer processes and isolates the adaptation, allowing individual examination capabilities to be enhanced without direct contamination from preceding examinations, thus maintaining performance consistency while improving individual capability.
2Reliability
If offline learning type is used with parameters totally unupdated after training, then performance consistency is maintained, but the capability in individual examinations cannot be optimized
Solution Approach 1:
The patent implements a dynamic learning approach where the model transitions from static offline learning to dynamic online learning based on the examination context. The system dynamically generates derived models with updated parameters for each examination while maintaining the original trained model, enabling adaptability without sacrificing the reliability of the base model through controlled parameter updates.
Solution Approach 2:
The patent applies parameter changes selectively by generating derived models with updated parameters only when needed for individual examinations. This controlled parameter modification allows the system to optimize capability for specific examinations while maintaining the original parameter set for consistency, resolving the contradiction through conditional and localized parameter updates.
3Device complexity
If a single trained model is used for all examinations, then device complexity is reduced, but the inference accuracy for individual cases is limited
Solution Approach 1:
The patent uses copying by generating derived models as copies of the trained model with specific parameter adjustments for individual examinations. This approach maintains simplicity by basing all derived models on a single trained model structure, while improving accuracy through customized parameter sets copied and adapted for each specific examination case.
Solution Approach 2:
The patent applies preliminary action by pre-training a general model that serves as a foundation for all examinations. This preliminary trained model reduces complexity by eliminating the need to train from scratch for each examination, while enabling accurate individual case processing through subsequent derived model generation with examination-specific parameter adjustments.
Data Source
AI summary
According to one embodiment, a medical data processing apparatus includes a processing circuitry. The processing circuitry is configured to: obtain first medical data; generate, based on the first medical data, a derived model from a trained model; obtain second medical data for a same subject as a subject of the first medical data and with an acquisition parameter different from an acquisition parameter of the first medical data; and apply the second medical data to the derived model to generate third medical data.