Dual-Classifier Image Processing for Medical Region Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing techniques based on machine learning struggle to accurately extract target regions, particularly in medical images where adjacent objects have similar density values, leading to misclassification or incomplete extraction of regions like the liver and heart in CT scans.
Innovation Solution
An image processing apparatus that employs a dual-classifier system, using a 2D-U-Net for initial classification and a 3D-U-Net for specific local regions, integrates results to enhance the accuracy of target region extraction by distinguishing between the liver and heart in three-dimensional medical images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a single classifier trained based on machine learning is used to extract target regions, then the extraction process is simple and fast, but the accuracy of extracting target regions deteriorates in regions where objects captured together with target objects vary
Solution Approach 1:
The patent divides the extraction process into multiple stages using different classifiers. A first classifier performs initial extraction, then a second classifier processes specific regions identified as problematic by the first classification result. This segmentation allows each classifier to specialize in different aspects, improving overall accuracy while maintaining efficiency through selective application of the more complex second classifier.
Solution Approach 2:
The patent applies different classification approaches to different regions of the image. The first classifier handles general regions, while the second classifier is specifically applied to regions where extraction is likely to fail. This local differentiation optimizes accuracy for problematic regions without unnecessarily processing entire images with the more complex second classifier, balancing speed and accuracy.
2Measurement precision
If multiple classifiers are used to improve extraction accuracy, then the accuracy of extracting target regions improves, but the device complexity increases
Solution Approach 1:
The patent segments the classification task into two parts: a first classifier for general extraction and a second classifier for refining specific regions. This segmentation reduces complexity compared to using a single complex classifier for the entire image, as the second classifier only processes identified problematic regions rather than the whole image.
Solution Approach 2:
The patent applies the second classifier only partially - specifically to regions where the first classifier is likely to fail, as determined by the determination unit. This partial application reduces computational complexity and system resource requirements compared to applying multiple classifiers to the entire image, while still achieving improved accuracy where needed.
3Loss of time
If a single classifier is used for classification, then the processing time is short, but the extraction accuracy deteriorates in regions with varying captured objects
Solution Approach 1:
The patent applies the more time-consuming second classifier only to specific regions identified as problematic, rather than processing the entire image with both classifiers. This partial application significantly reduces total processing time while still improving accuracy in the critical regions where single-classifier extraction is likely to fail.
Solution Approach 2:
The patent implements local quality enhancement by applying different processing intensities to different regions. High-quality processing with the second classifier is applied only where needed (regions with varying captured objects), while other regions use the faster first classifier, optimizing the balance between processing time and accuracy.
Data Source
AI summary
An image processing apparatus includes a first classification unit configured to classify each of a plurality of pixels included in a three-dimensional medical image using a first classifier for classifying each pixel into a plurality of classes including a class representing a first target region, a determination unit configured to determine an image region including the first target region and a second target region from the three-dimensional medical image based on a first classification result, a second classification unit configured to classify each of a plurality of pixels included in the determined image region using a second classifier for classifying each pixel into a plurality of classes including at least either one of a class representing the first target region and a class representing the second target region, and an integration unit configured to integrate the first and the second classification results to acquire a third classification result.