Endoscopic Image Processing With Multi-Granularity Lesion Masks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing endoscopic image processing systems struggle to accurately detect areas of interest, such as lesions, particularly flat lesions, due to difficulties in identifying these areas correctly.
Innovation Solution
An image processing device that generates multiple mask images with varying levels of granularity to identify attention parts, like lesions, using a mask image generation model based on Feature Pyramid Networks, and adjusts granularity according to user input for precise detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a single mask image with fixed granularity is used to detect attention parts, then the detection process is simple and fast, but the detection accuracy deteriorates for flat lesions and other challenging attention parts
Solution Approach 1:
The patent divides the detection task into multiple granularity levels by generating multiple mask images with different granularities (coarse to fine). Each mask image targets different scales of attention parts, allowing the system to detect both large and small lesions effectively. This segmentation of the detection space resolves the contradiction by maintaining simplicity at each level while achieving high overall accuracy.
Solution Approach 2:
The patent introduces a new dimension of granularity level beyond the traditional single-scale detection. By creating mask images across multiple granularity dimensions, the system can capture attention parts at various scales simultaneously, improving detection accuracy for flat lesions without significantly increasing processing complexity.
2Measurement precision
If pixel-level identification is used for attention parts, then the identification precision is high, but the detection reliability deteriorates for flat lesions that cannot be correctly detected
Solution Approach 1:
The patent segments the identification process into multiple granularity levels rather than relying on a single pixel-level identification. By generating mask images at different granularities and integrating their results, the system maintains high identification precision while improving reliability for flat lesions that are difficult to detect at any single scale.
Solution Approach 2:
The patent merges the detection results from multiple mask images with different granularities to produce a final identification. This combination approach allows the system to leverage the strengths of each granularity level, maintaining high precision while improving reliability by cross-validating detections across multiple scales.
3Measurement precision
If multiple mask images with different granularities are generated, then the detection accuracy for flat lesions is improved, but the processing time and computational load increase
Solution Approach 1:
The patent segments the processing into hierarchical granularity levels, where coarse-granularity mask images are generated first to quickly identify potential attention areas, followed by finer-granularity processing only in those regions. This segmented approach improves detection accuracy for flat lesions while minimizing overall processing time by avoiding full-resolution processing of the entire image.
Solution Approach 2:
The patent performs preliminary detection using coarse-granularity mask images before proceeding to fine-granularity processing. This preliminary action identifies regions of interest that require detailed analysis, allowing the system to achieve high detection accuracy while reducing total processing time by focusing computational resources only where needed.
Data Source
AI summary
The image processing device 1X includes an acquisition means 31X, a mask image generation means 32X, and an identification means 33X. The acquisition means 31X acquires a captured image acquired by photographing an inspection target by a photographing unit provided in an endoscope. The mask image generation means 32X generates a plurality of mask images which indicate candidate areas of an attention part in the captured image by different levels of granularity. The identification means 33X identifies the attention part based on the plurality of the mask images. The image processing device is able to support decision-making based on identification results of the attention part in endoscopic examination.