Endoscope Anomaly Detection via Autoencoder Restoration Error
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing endoscope systems rely on visual judgment for detecting lesion parts, which can lead to overlooked important inspections, and lack effective techniques for accurately identifying images indicating a target object not in a normal state.
Innovation Solution
An inspection device and method that acquires and analyzes images using an autoencoder learned from normal state images to detect anomalies, determining if a target image indicates an abnormal state by calculating restoration errors and comparing them to a threshold.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If visual judgment is used to detect lesion parts from continuous images, then the inspection process is simple and requires minimal computational resources, but important inspections may be overlooked and detection reliability is low
Solution Approach 1:
The patent replaces the mechanical/visual judgment process with an automated image processing system using deep learning neural networks. The detection unit automatically analyzes endoscope images to identify abnormal regions, substituting human visual judgment with computational algorithms that provide consistent and reliable detection without overlooking important inspections.
Solution Approach 2:
The system performs self-service by automatically detecting and marking abnormal regions in endoscope images without requiring continuous human intervention. The neural network model processes images independently, automatically identifying lesion parts and normal parts, thereby maintaining high detection reliability while reducing operational complexity.
2Measurement precision
If template matching or similar techniques are used to detect areas of interest, then the detection process is automated, but the system cannot accurately detect images indicating target object is not in normal state
Solution Approach 1:
The patent changes the detection parameters by using deep learning neural networks instead of traditional template matching. The system learns from training data containing both normal and abnormal images, enabling it to accurately detect and distinguish between normal and abnormal states with high precision and reliability.
Solution Approach 2:
The system creates a digital copy of the endoscope images and processes these copies through the neural network model. By analyzing replicated image data, the system can accurately identify abnormal patterns without affecting the original images, achieving both high measurement precision and reliability in abnormal state detection.
3Reliability
If all continuously obtained images are reviewed, then no inspection is overlooked, but the inspection time and computational resources increase significantly
Solution Approach 1:
The patent extracts and isolates only the abnormal regions from the continuous image sequence using the neural network detection unit. Instead of reviewing all images, the system extracts and focuses only on the abnormal parts that require attention, thereby maintaining inspection completeness while significantly reducing the time and computational resources needed.
Solution Approach 2:
The system segments the inspection process into two parts: automatic detection of abnormal regions using neural networks, and manual review only of the extracted abnormal parts. This segmentation allows the system to maintain high inspection completeness by ensuring all abnormalities are caught, while reducing overall inspection time by avoiding unnecessary review of normal images.
Data Source
AI summary
The acquisition unit 41B acquires a target image indicating a target object of inspection. The detection unit 42B detects, on the basis of a group of images each of which indicates the target object in a normal state, the target image that indicates the target object that is not in the normal state among the target images that the acquisition unit 41B acquires.