Adaptive Neural Network Selection for Breast CAD
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing computer-assisted diagnosis (CAD) systems using uniform neural networks for image recognition may fail to achieve sufficient performance in recognition accuracy and processing speed due to varying usage conditions.
Innovation Solution
A diagnosis support program that utilizes a storage medium to store a program enabling multiple recognizers with different neural networks to be used for breast image recognition, with a selection mechanism choosing the appropriate recognizer based on examination information such as examination type, display performance, and interpreter proficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a uniform neural network is used for recognition processing, then the system structure is simple, but the recognition accuracy and processing speed are insufficient for varying usage conditions
Solution Approach 1:
The patent divides the recognition system into multiple independent neural networks (first neural network for category recognition, second neural network for characteristic recognition). Each network is specialized for a specific recognition task, allowing the system to select appropriate networks based on examination type. This segmentation enables adaptability to different usage conditions while maintaining manageable complexity through modular design.
Solution Approach 2:
The patent implements dynamic selection of neural networks based on examination information. The system can switch between different neural networks (first or second) depending on whether the examination is a screening examination or thorough examination. This dynamic adaptation allows the system to optimize recognition performance for each specific usage scenario rather than using a fixed uniform network.
2Measurement precision
If multiple specialized recognizers are used for different examination types, then the recognition accuracy is improved, but the system complexity increases
Solution Approach 1:
The patent applies local quality by creating neural networks with specialized functions tailored to specific recognition needs. The first neural network is optimized for category recognition in screening examinations, while the second neural network is optimized for characteristic recognition in thorough examinations. Each network has locally optimized parameters and structures suited to its specific purpose, improving recognition accuracy for each examination type.
Solution Approach 2:
The patent creates a universal recognition system that can handle multiple examination types through a selection mechanism. The system includes a plurality of neural networks that can be selectively activated based on examination information, making the overall system multi-functional. This universality allows a single system to serve both screening and thorough examination purposes with appropriate network selection.
3Reliability
If a neural network with high sensitivity is used, then the detection capability is improved, but the processing speed may be reduced
Solution Approach 1:
The patent applies partial action by implementing different levels of recognition based on examination type. For screening examinations, the system uses the first neural network for category recognition, which provides sufficient detection capability with reasonable processing speed. For thorough examinations, the system uses the second neural network for more detailed characteristic recognition. This partial application of recognition depth optimizes the balance between sensitivity and speed for each scenario.
Data Source
AI summary
A storage medium, a diagnosis support apparatus and a diagnosis support method that enable presenting a recognition result suitable to a status of use of a CAD (computer-assisted diagnosis/detection) function are provided. A diagnosis support apparatus performs recognition processing of a breast image showing a projection image or a section image of a breast of a subject, using one or more recognizers from among a plurality of recognizers each including a neural network, and selects a recognizer to be used for the recognition processing or a recognizer that is to output a recognition result of the recognition processing, from among the plurality of recognizers, according to examination information relating to an examination of the subject.