Endoscopic Lesion Classification With Separate Evidence Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing medical image processing systems struggle with inappropriate classification results due to images unsuitable for automatic analysis, such as hidden or blurred lesions, and observers find it difficult to determine the appropriateness of classification outcomes.
Innovation Solution
A medical image processing apparatus that classifies images into multiple classes, generates a region image highlighting areas contributing to classification, and displays this region image separately from the original image, allowing observers to assess classification appropriateness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If automatic classification is performed on all captured images, then classification productivity is improved, but classification reliability deteriorates due to inappropriate images (hidden or blurred lesions)
Solution Approach 1:
The system performs preliminary quality assessment of captured images before automatic classification by AI. Images are evaluated for suitability (clear depiction of lesions) before being submitted to the classification model, preventing inappropriate images from compromising classification reliability while maintaining high productivity for suitable images
2Ease of operation
If only classification results are displayed, then ease of operation is improved, but information completeness deteriorates as observers cannot assess classification appropriateness
Solution Approach 1:
The system highlights specific regions within images that contributed to the classification decision using visual indicators such as heat maps or bounding boxes. This allows observers to see exactly which areas of the image influenced the AI's classification, providing local detail about the classification basis without overwhelming the entire interface
Solution Approach 2:
The system introduces an intermediate visual representation (region highlighting overlay) between the original image and the classification result. This intermediary layer translates the AI's internal decision-making process into observable visual cues that help observers assess whether the classification is appropriate based on the highlighted regions
Data Source
AI summary
An apparatus, a system, a method, and a program for medical image processing are provided. The apparatus includes one or more processors configured to acquire an endoscopic image generated through imaging of a living body, perform classification of lesion regions contained in the endoscopic image into two or more classes, identify a classification contributing region in the endoscopic image, the classification contributing region contributing, with a degree of contribution, to the classification of one of the lesion regions contained in the endoscopic image, generate a region image displaying a region in the endoscopic image, the region image displaying the classification contributing region with a density or a heat map according to the degree of contribution, and display the region image along with the endoscopic image on a monitor, where the region image is displayed at a position different from a position at which the endoscopic image is displayed.