Endoscopic Lesion Bounding Box Adjustment for Accurate AI Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing medical image processing systems face challenges in accurately detecting lesions using bounding boxes due to the irregular shapes of lesions, particularly in tubular organs like the esophagus, leading to incorrect learning data and false detections.
Innovation Solution
A medical image processing device adjusts bounding boxes based on brightness information and pixel analysis to fit the shape of the lesion, reducing the area and aligning the centroid with the lesion centroid, and excludes low-brightness regions to improve accuracy for machine learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a bounding box is used to represent a lesion with irregular shape (e.g., circumferential lesion in esophagus), then the lesion can be detected and highlighted, but most of the region inside the bounding box becomes unrelated to the lesion, reducing machine learning accuracy
Solution Approach 1:
The patent segments the bounding box region into multiple sub-regions based on brightness information and pixel characteristics. By dividing the bounding box into multiple smaller regions and analyzing their brightness values, the system can identify which sub-regions actually contain lesion information and which are empty or contain irrelevant tissue, thereby reducing information loss while maintaining detection accuracy
Solution Approach 2:
The patent applies local quality analysis by examining the brightness distribution and pixel characteristics at different locations within the bounding box. Regions with specific brightness patterns (indicating lesion presence) are identified and separated from regions with different brightness patterns (normal tissue or shadow areas), allowing the system to focus machine learning on relevant local areas only
2Ease of operation
If a bounding box is used for circumferential lesions, then the lesion position can be indicated, but normal mucous membrane portions and shadow boundaries are incorrectly learned as lesion features, causing false detections
Solution Approach 1:
The patent utilizes brightness changes (analogous to color changes) as a key feature to distinguish lesion regions from normal tissue and shadow areas. By analyzing brightness value distributions and identifying regions with specific brightness characteristics, the system can reliably differentiate between actual lesion features and artifacts, significantly reducing false detections while maintaining ease of operation
Solution Approach 2:
The patent introduces brightness information analysis as an intermediary step between the bounding box generation and machine learning processes. This intermediary analysis filters out false features by examining brightness patterns, acting as a mediator that prevents normal mucous membrane portions and shadow boundaries from being incorrectly learned as lesion features
3Loss of information
If the bounding box area is reduced to focus on the lesion, then machine learning data quality improves, but the bounding box may no longer encompass the entire lesion, reducing detection coverage
Solution Approach 1:
The patent applies partial action by focusing the bounding box adjustment on the most relevant regions identified through brightness analysis, rather than uniformly reducing the entire bounding box area. By selectively adjusting the bounding box to encompass only the regions with relevant lesion information (identified through brightness patterns), the system maintains detection coverage while improving data quality for machine learning
Data Source
AI summary
A medical image processing device includes an image acquisition unit that acquires an endoscopic image, an bounding box acquisition unit that acquires a bounding box which corresponds to an object to be detected shown in the endoscopic image and in which at least a part of the object to be detected is included, an adjustment unit that changes the position of the bounding box and reduces the area of the bounding box on the basis of the endoscopic image, and a storage controller that associates a new adjusted bounding box with the endoscopic image and stores the new adjusted bounding box and the endoscopic image.