Endoscope Image Classification for Sensor-Specific Model Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing signal analysis systems struggle to generate high-quality annotations for machine learning models when dealing with data from sensors of varying characteristics and unknown conditions, making it difficult to handle diverse sensor types and conditions effectively.
Innovation Solution
An information processing device and method that classify image groups from different endoscopes to determine whether they were obtained at specific times, using previously annotated data to create inference models for image feature determination, thereby handling sensors of differing characteristics and unknown conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If deep learning is performed with data from sensors of different characteristics without considering sensor type differences, then the system can handle diverse sensor types, but the quality of annotations and inference model generation deteriorates
Solution Approach 1:
The patent segments the training data by sensor type and creates separate inference models for each sensor type. This allows the system to maintain versatility across different sensor types while ensuring high annotation quality for each specific type through dedicated processing and annotation workflows.
Solution Approach 2:
The patent applies local quality by tailoring the annotation process and inference model generation to the specific characteristics of each sensor type. Each sensor type receives customized processing that matches its unique properties, thereby maintaining high quality annotations while preserving the ability to handle diverse sensors.
2Productivity
If classification is continuously performed using specialist instructions, then effective use of data is achieved, but the system cannot handle sensors with unknown characteristics
Solution Approach 1:
The patent performs preliminary classification of image groups into known and unknown sensor types before the main processing. This preliminary action enables the system to handle unknown sensor characteristics by routing them to appropriate processing streams, while maintaining continuous effective data processing through specialist instructions for known sensor types.
Solution Approach 2:
The patent introduces an intermediary classification mechanism that acts as a mediator between raw sensor data and the main processing pipeline. This intermediary structure allows the system to accommodate sensors with unknown characteristics by first identifying their type and then routing them through appropriate processing, thus maintaining both productivity and adaptability.
3Quantity of substance
If all acquired images are used for training without time-based classification, then the training data volume is maximized, but the efficiency of model generation and annotation quality deteriorate
Solution Approach 1:
The patent segments the training data by time periods and sensor types, creating organized groups that can be efficiently processed. This segmentation allows the system to maintain adequate training data volume while improving model generation efficiency through structured processing of classified image groups.
Solution Approach 2:
The patent performs preliminary classification of images into time-based groups and sensor type categories before training. This preliminary action organizes the training data in a way that maintains sufficient data volume while significantly improving the efficiency of subsequent model generation processes.
Data Source
AI summary
An information processing device of the present invention is capable of collaboration with a learning device to determine whether an image group that has been obtained in time series by the first endoscope is an image group obtained at a first time or at a second time, and to create a first inference model for image feature determination of images for the first endoscope by performing learning with results of having performed annotation on the image group that was obtained at the second time as training data, the information processing device comprising at least one or a plurality of classifying processors that classify image groups constituting training data candidates, within an image group from the first endoscope that has been newly acquired, or an image group from a second endoscope, using the image group that has been obtained at the first time, when the first inference model was created.