Synthetic Data Generator for Neural Network Retraining

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional independent training of operational neural networks at different clinical sites results in inconsistent or inaccurate outputs due to limited diversity in training data sets, especially for rare medical conditions, and requires excessive resources for retraining, as they cannot share data or parameters across sites.

Innovation Solution

A neural network system that uses a synthetic data generator to create a synthetic retraining data set, which is used to retrain operational neural networks, ensuring consistent outputs without sharing actual data or parameters, by optimizing the loss function based on the difference in medical metric values from multiple networks.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If operational neural networks are trained independently at different clinical sites using locally acquired data sets, then data security and privacy constraints are satisfied, but the diversity of training data is limited resulting in inconsistent or inaccurate outputs for rare medical conditions

Engineering Contradiction:
Improveconsistency and accuracy of medical metric outputsVSAvoiddiversity of training data
Core Design Contradiction:
ReliabilityVSAdaptability or versatility

Solution Approach 1:

A central server acts as an intermediary that collects training data from multiple clinical sites, generates synthetic training data sets, and distributes them to operational neural networks at various sites. This mediator enables data diversity without requiring direct data sharing between sites, resolving the contradiction between data privacy and model consistency.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The system creates synthetic copies of training data through a central server that generates artificial training data sets based on data collected from multiple sites. These synthetic data copies provide diversity to local models without requiring access to actual patient data from other sites, maintaining privacy while improving consistency.

Inventive Principle:
Principle #26Copying

2Adaptability or versatility

If large amounts of training data are collected to central servers to improve model training, then the diversity of training data increases, but user data privacy and security risks increase

Engineering Contradiction:
Improvediversity of training dataVSAvoiddata privacy and security risks
Core Design Contradiction:
Adaptability or versatilityVSObject-affected harmful factors

Solution Approach 1:

The central server serves as a trusted intermediary that collects data in aggregated form, processes it into synthetic training data sets, and distributes the processed data back to sites. This intermediary architecture enables data diversity while minimizing privacy risks by never exposing individual patient data to external parties.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The system extracts only the necessary statistical patterns and features from collected data at the central server, then uses these extracted patterns to generate synthetic data sets. This extraction approach obtains the essential training value while removing identifiable patient information, reducing privacy risks.

Inventive Principle:
Principle #2Taking out (Extraction)

3Reliability

If parameters are exchanged between operational neural networks to improve robustness, then some consistency is achieved, but the problem of scarce training data with rare features remains and retraining resources increase exponentially with the number of networks

Engineering Contradiction:
Improverobustness of operational neural networkVSAvoidretraining resources and computational overhead
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The system merges the training data collection capability of multiple sites at a central server, consolidating diverse data sources into unified synthetic training data sets. This merging approach achieves the benefits of multiple networks' data diversity without requiring exponential parameter exchange or retraining resources.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The central server performs preliminary data processing and synthetic data generation before distribution to operational neural networks. This preliminary action prepares comprehensive training data in advance, eliminating the need for repeated parameter exchange and extensive retraining when networks need to be updated.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4068161A1Technique for retraining operational neural networks using synthetically generated retraining data
Publication Date: 2022.10.05 SIEMENS HEALTHINEERS AG
  • EP4068161A1 patent drawingFigure 1
  • EP4068161A1 patent drawingFigure 2
  • EP4068161A1 patent drawingFigure 3

AI summary

A neural network system (100) for retraining operational neural networks (130-1, 130-2) using a synthetic data set generated by a synthetic data generator neural network (110) is provided. The synthetic data generator network (110) comprises an input layer (112) for receiving (202) an input data set; an output layer (118) for outputting (208) the synthetic data set; and a loss function (124) for receiving (204) from each operational network (130-1, 130-2) a value of a medical metric. The operational networks (130-1, 130-2) each comprise an input layer for receiving (210) the synthetic data set; and an output layer for outputting (214) the value of the medical metric. The synthetic data generator network (110) is trained for generating the synthetic data set based on the loss function (124) comprising a difference of the values of the medical metric. Each operational network (130-1, 130-2) is retrained using the synthetic data set.