Endoscope Processor Region Identification via Deep Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In endoscopic examinations, the insertion length of an endoscope does not correspond to the examination region due to variations in the stomach or large intestine, making it difficult to select an appropriate reference image for endoscopic systems that rely on insertion length for image recording.
Innovation Solution
A processor for an endoscope that includes an image acquisition unit, a region acquisition unit using a first learning model to identify the target region, and an image output unit that superimposes the endoscope image with an index indicating the target region, allowing for accurate selection and recording of reference images based on the examination region.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If insertion length is used to determine examination region, then reference image selection is simplified, but accuracy of region identification deteriorates due to anatomical variations
Solution Approach 1:
The patent replaces the mechanical measurement system (insertion length-based region determination) with an image processing system (deep learning model-based region identification). The processor uses convolutional neural networks to analyze endoscope images and automatically identify examination regions, eliminating the need for manual insertion length measurement and reference image selection. This substitution resolves the contradiction by providing both automated operation and high identification accuracy through AI-based image analysis.
2Measurement precision
If deep learning model is implemented for region identification, then examination region identification accuracy is improved, but device complexity increases
Solution Approach 1:
The patent implements a universal deep learning model that can identify multiple types of examination regions (stomach, large intestine, etc.) and perform various functions (region detection, boundary identification, reference image selection) through a single processor system. The model is trained on diverse datasets covering different anatomical structures, enabling it to adapt to various examination scenarios without requiring separate specialized systems. This multi-functionality reduces the overall complexity compared to having multiple dedicated systems for different examination types.
Solution Approach 2:
The deep learning model operates autonomously to identify examination regions and select reference images without requiring manual intervention. The processor automatically processes endoscope images, detects regions of interest, and determines appropriate reference images based on the identified regions. This self-service capability eliminates the need for complex manual操作流程 and reduces the operational complexity of the system while maintaining high identification accuracy.
3Device complexity
If manual reference image selection is performed, then system complexity is reduced, but productivity of examination process deteriorates due to time-consuming operations
Solution Approach 1:
The patent implements preliminary action by pre-training deep learning models on extensive datasets of endoscope images before actual examination use. The models are prepared in advance with learned features and patterns for accurate region identification. During the examination process, the pre-trained models immediately process images without requiring manual setup or configuration, enabling rapid automated reference image selection. This preliminary preparation resolves the contradiction by providing both system simplicity and high productivity through automated real-time processing.
Data Source
AI summary
A processor for an endoscope or the like that assists an endoscopic examination using an appropriate reference image is provided. A processor for an endoscope includes an image acquisition unit that acquires an endoscope image; a region acquisition unit that inputs the endoscope image acquired by the image acquisition unit to a first learning model that outputs a target region corresponding to a predetermined region to be photographed when the endoscope image is input to acquire the target region; and an image output unit that outputs the endoscope image acquired by the image acquisition unit and an index indicating the target region acquired by the region acquisition unit with the endoscope image and the index superimposed.