Self-Adjusting Model Update System for Medical Diagnosis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning and deep learning models in medical diagnosis face challenges in adaptability across different hospitals due to varying judgment criteria, leading to low user acceptance and willingness to use the systems, as the training process is time-consuming and requires labeled data that may not align with individual hospital standards.
Innovation Solution
An electronic device with a processor and storage circuit that includes modules for prediction, intelligent labeling, automatic training, model evolution decision, and automatic process management and control, allowing for self-adjustment and training of models based on user feedback and performance metrics, enabling the system to adapt and improve prediction accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a highly parametric model is established with large number of parameters trained on labeled data, then the model can be used for predicting new data, but the training process is extremely time-consuming and requires manual labeling which reduces productivity
Solution Approach 1:
The system enables models to automatically generate training data and perform self-training without manual intervention. The model generates synthetic labeled data through its own predictions and automatically updates its parameters, eliminating the need for time-consuming manual labeling while maintaining prediction accuracy
Solution Approach 2:
The system pre-generates synthetic training data using the model's current predictions before actual training occurs. By preparing training data in advance through automated generation rather than waiting for manual labeling, the system accelerates the training process while ensuring data quality
2Measurement precision
If a model is trained on labeled data from one hospital, then the prediction results align with that hospital's judgment criteria, but the model cannot be effectively popularized to other hospitals with different judgment criteria, reducing adaptability
Solution Approach 1:
The system designs a universal model architecture that can adapt to multiple hospitals' different judgment criteria through configurable parameters and domain-specific training. The same base model can be customized for different hospitals by training on their specific data characteristics without requiring complete retraining, enabling both accuracy and adaptability
Solution Approach 2:
The system allows different parts of the model to have different specialization levels. The core model maintains general medical knowledge while allowing hospital-specific adaptations through localized training data and parameters. This enables the model to maintain consistent performance across hospitals while adapting to local judgment criteria
3Measurement precision
If manual data labeling is performed to train models, then the model learns labeling rules for accurate prediction, but the process requires significant time and human resources, increasing loss of time
Solution Approach 1:
The model automatically generates its own training data by producing predictions that are then used as synthetic labels. This self-generated labeled data eliminates the need for manual labeling while maintaining the quality and consistency of training signals, significantly reducing time loss
Solution Approach 2:
The system implements a feedback loop where the model's predictions are continuously evaluated and used to generate new training data. This automated feedback mechanism replaces manual labeling by using the model's own performance to create training examples, reducing time consumption while preserving labeling accuracy
Data Source
AI summary
An electronic device and a model updating method are provided. The method includes: inputting a plurality of files to a first model and outputting a predicted result of each of the plurality of files; receiving a corrected result for correcting the prediction result of at least one first file in the plurality of files, and generating a first label file corresponding to the at least one first file according to the corrected result and the first file; training a plurality of models according to the first label file to generate a plurality of a trained model; testing the plurality of trained models using at least one test set; and replacing the first model with a first trained model when the predicting accuracy of the first trained model of the plurality of trained models is higher than the predicting accuracy of the first model.