Edge CAD Device Selective Learning Data Transmission
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current medical image processing systems require excessive processing capacity and time for retraining or additional training of computer-aided diagnosis (CAD) models, and large volumes of communication data are needed to transmit learning data from edge devices to external systems, leading to inefficient data processing and communication.
Innovation Solution
A medical image processing apparatus that performs additional training for CAD devices using input medical images, evaluates the effectiveness of the training, and selectively communicates learning difference information to reduce unnecessary data transmission, thereby reducing communication volume and processing load.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If learning data is collected from multiple edge devices and transmitted to external systems for batch retraining, then the CAD model can be improved through additional training, but the communication volume and processing time become excessively large
Solution Approach 1:
The patent performs additional training at edge devices before transmission, preparing learning data in advance with preprocessed features and gradients. This preliminary action reduces the computational burden on external systems and accelerates the overall retraining process by avoiding redundant computations during centralized training.
Solution Approach 2:
The patent extracts only the essential learning differences (parameter updates, feature gradients) from the additional training results and transmits only these extracted elements to external systems. This selective extraction significantly reduces communication volume while preserving the critical information needed for model improvement.
2Reliability
If learning data is collected from multiple edge devices and transmitted to external systems for batch retraining, then the CAD model can be improved through additional training, but the communication volume becomes excessively large
Solution Approach 1:
The patent extracts only the essential learning differences (parameter updates, feature gradients) from the additional training results and transmits only these extracted elements to external systems. This selective extraction significantly reduces communication volume while preserving the critical information needed for model improvement.
Solution Approach 2:
The patent transforms the learning data from raw image formats into compressed parameter representations (gradient vectors, parameter deltas) that convey the same training information in a much more compact form, thereby reducing communication requirements.
3Reliability
If additional training is performed using all collected learning data, then the CAD model performance can be improved, but the processing load becomes excessively high
Solution Approach 1:
The patent segments the training process into two parts: additional training performed locally at edge devices using local data, and centralized retraining performed at external systems using only the extracted learning differences. This segmentation distributes the computational load and avoids the need for any single system to handle the entire training workload.
Solution Approach 2:
The patent extracts only the essential learning differences (parameter updates, feature gradients) from the additional training results and transmits only these extracted elements to external systems. This selective extraction significantly reduces communication volume while preserving the critical information needed for model improvement.
Data Source
AI summary
There are provided a medical image processing apparatus, a medical image processing method, a machine learning system, and a program that can reduce the volume of communication and can reduce the processing load of retraining or additional training that is performed by a machine learning apparatus. A medical image processing apparatus (13) includes: a trainer (26) that performs additional training for a first computer-aided diagnosis device on the basis of an input medical image; an evaluation unit that compares a second computer-aided diagnosis device obtained by the additional training with the first computer-aided diagnosis device and evaluates whether learning difference information about the additional training contributes to improvement of performance of the first computer-aided diagnosis device; a communication determination unit that determines, on the basis of a result of evaluation, whether the learning difference information is to be communicated; and a communication unit (34) that outputs the learning difference information in accordance with a result of determination by the communication determination unit.