Deep Learning Imaging System Gradient Update Aggregation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning algorithms for imaging systems require access to raw imaging data for training, which poses challenges in preserving patient privacy and leveraging improvements across different imaging systems.
Innovation Solution
A method where imaging systems perform scans, train deep neural networks, and transmit updates to a central server for training a global deep learning model, allowing for aggregation and averaging of updates to deploy an improved neural network across systems without sharing raw patient data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If raw imaging data is transmitted to train central deep learning models, then model training quality is improved, but patient privacy is compromised
Solution Approach 1:
The patent extracts only the essential training information (gradient updates) from the raw imaging data, transmitting only these extracted updates to the central server while leaving the sensitive raw data local to each imaging system. This resolves the contradiction by separating the useful training signal from the privacy-sensitive data.
Solution Approach 2:
The patent introduces local deep learning models as intermediaries that process raw imaging data locally and convert it into gradient updates. These intermediaries prevent direct transmission of raw data while still enabling central model training, thus mediating between privacy protection and model quality improvement.
2Productivity
If raw imaging data is shared across imaging systems, then global model improvement is accelerated, but data security is compromised
Solution Approach 1:
The patent creates copies of training information in the form of gradient updates that can be freely shared across imaging systems. These copies contain the essential learning signals without replicating the sensitive raw data, enabling rapid global model improvement while maintaining data security.
Solution Approach 2:
The patent transforms raw imaging data into a different parameter representation (gradient updates) that preserves training utility while removing sensitive information. This parameter transformation enables secure sharing across systems without compromising data security.
3Object-affected harmful factors
If deep learning models are trained locally without sharing data, then patient privacy is protected, but model generalization capability deteriorates
Solution Approach 1:
The patent merges the training efforts of multiple imaging systems by collecting gradient updates from various local models and aggregating them at a central server. This combination enables the global model to learn from diverse data sources while each local system maintains privacy protection.
Solution Approach 2:
The patent establishes a feedback loop where local models train on private data, send gradient updates to the central server, receive updated global models, and apply them locally. This feedback mechanism enables continuous improvement of generalization capability while maintaining privacy protection throughout the cycle.
Data Source
AI summary
Methods and systems are provided for generating deep learning training data with an imaging system. In one embodiment, a method for an imaging system comprises performing a scan of a subject to acquire imaging data, training a deep neural network on the imaging data to obtain updates to the deep neural network, and transmitting the updates to a server for training a central deep neural network. In this way, imaging data may be leveraged for training and developing global deep learning models without transmitting the imaging data itself, thereby preserving patient privacy.