Endoscopic Image Visualization Using Multi-Model Activation Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Endoscopic image analysis systems face challenges in accuracy due to false positives and inconsistencies in detection caused by endoscope movement and varying processing speeds, and existing methods for reducing false detections are limited in indicating detailed target areas.
Innovation Solution
A system and method that utilizes multiple AI models to determine analysis targets through an agree-all or voting method for probability values, generates common activation maps (CAMs) and adjacent common activation maps (ACAMs), applies weights to these maps, and sets detection confirmation thresholds to reduce false detections, and indicates targets in detail using closed curves.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a single AI model is used to analyze endoscopic images, then the analysis process is simple and fast, but detection accuracy decreases due to false positives and inconsistencies caused by endoscope movement
Solution Approach 1:
The patent combines multiple AI models (first AI model and second AI model) to analyze the same endoscopic image. The system integrates predictions from both models and uses a voting mechanism to determine the final detection result, thereby improving detection accuracy while maintaining processing speed through parallel execution of multiple models.
2Reliability
If a fixed number of frames N is used to confirm detection, then false detections are reduced, but processing speed varies and detection accuracy is affected by examiner movement speed
Solution Approach 1:
The patent replaces the fixed frame count approach with a dynamic voting mechanism that evaluates predictions from multiple AI models simultaneously. The system uses probability values from multiple models and applies a voting threshold to determine detection confirmation, allowing flexible adaptation to different examination speeds and conditions while maintaining consistent detection accuracy.
3Ease of operation
If square coordinates are used to display detection results, then the output is simple and easy to process, but detailed target area indication is insufficient
Solution Approach 1:
The patent enhances the simple square coordinate output by adding activation map visualization. The system generates activation maps that provide spatial information about the detected target areas, adding a visual dimension to the coordinate data. This allows the system to maintain simple coordinate output while simultaneously providing detailed target area indication through the activation map overlay.
Data Source
AI summary
A method of visualizing an analysis result of an endoscopic image according to a first embodiment includes loading, by a model loading/condition setting unit, a plurality of analysis models and setting analysis conditions, analyzing, by an image analyzer, the image using the plurality of analysis models when the situation requires image analysis, extracting, by an AM extractor/generator, AMs of the plurality of analysis models and generating a CAM when an image analysis result satisfies a set analysis condition, generating an ACAM by applying weights to CAMs extracted in previous and current image frames, extracting a peak activation value from the ACAM by an activation value extractor, and extracting, by a closed curve extractor, a closed curve of an AM and drawing the closed curve on a corresponding analysis image when the peak activation value is greater than or equal to a threshold.