Neural Network Model Training via Server-Client Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Medical imaging systems require significant computing resources for training neural network models, which can occupy large amounts of CPU capacity on client terminals, making it desirable to distribute the training process to reduce resource usage.
Innovation Solution
A system that utilizes a server device with stronger computing capacity to train a neural network model using training samples received from client terminals, transmitting the model back for testing and updating based on evaluation scores and quality parameters, allowing for efficient image processing tasks like denoising, enhancement, and artifact removal without overburdening the client terminals.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If neural network model training is performed on the client terminal, then the model can be generated locally, but it occupies a large number of computing resources (CPUs) on the client terminal
Solution Approach 1:
The system divides the model training task into two parts: data preprocessing is performed on the client terminal (lightweight operation), while the actual neural network training is segmented and executed on the server (resource-intensive operation). This segmentation allows the client to participate in the process without consuming excessive local computing resources.
Solution Approach 2:
The server acts as an intermediary between the client terminal and the model generation process. The client sends training data to the server, the server performs the computationally intensive training, and then returns the trained model to the client. This intermediary approach transfers the resource burden from the client to the server.
2Adaptability or versatility
If neural network model training is performed on the client terminal, then the model can be customized locally, but the training time is extended due to limited computing capacity
Solution Approach 1:
The training process is segmented into data preparation (client-side) and model training (server-side). This allows the time-consuming training operation to be performed on the server with its superior computing capacity, significantly reducing the effective training time experienced by the client while preserving local customization capabilities.
Solution Approach 2:
The server provides self-service model training capabilities that clients can access remotely. Instead of requiring clients to perform resource-intensive training themselves, the server automatically handles the training process using its own computing resources, making the service available to multiple clients simultaneously.
3Manufacturing precision
If computing resources are allocated to training neural network models, then model accuracy can be improved, but other image processing tasks may experience resource shortages
Solution Approach 1:
The system segments computing tasks by function: resource-intensive model training is performed on the server, while lighter image processing inference tasks are executed on the client terminal. This segmentation allows high-accuracy model generation without compromising the client's ability to perform real-time image processing operations.
Solution Approach 2:
The server acts as an intermediary that handles all resource-intensive training operations, freeing up client terminal resources for image processing tasks. The client only needs to send data and receive results, maintaining high productivity for image processing while the server handles the accuracy-critical training operations.
Data Source
AI summary
The disclosure relates to a system and a method for generating a neural network model for image processing by interacting with at least one client terminal. The method may include receiving via a network, a plurality of first training samples from the at least one client terminal. The method may also include training a first neural network model based on the plurality of first training samples to generate a second neural network model. The method may further include transmitting, via the network, the second neural network model to the at least one client terminal.