Multi-View Polyp Matching via Local Appearance Features
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current systems for detecting polyps in medical images, particularly in Computed Tomography Colonography (CTC), face challenges in achieving high sensitivity while minimizing false positives, especially when dealing with collapsed or deformed colon segments, and require manual registration which is time-consuming and inaccurate.
Innovation Solution
A processor-based system that performs multi-view matching of regions of interest using local appearance features, learns a distance metric through feature selection and metric boosting, allowing for efficient matching of polyps across prone and supine views without the need for global geometric information or surface registration.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual registration of polyp findings or colon segments is performed, then accuracy can be assessed, but it is time-consuming and difficult
Solution Approach 1:
The patent replaces manual mechanical registration processes with an automated computational system that uses feature extraction and matching algorithms. The system automatically registers colon segments across multiple views by extracting local features (shape, texture, intensity) and matching them through computational methods, eliminating the need for manual radiologist registration while achieving comparable or superior accuracy.
Solution Approach 2:
The system performs self-registration by automatically aligning colon segments across different views without requiring external manual intervention. The automated algorithm independently extracts features, computes transformations, and registers segments, allowing the system to serve itself rather than requiring continuous manual guidance for registration tasks.
2Reliability
If current CAD systems detect polyps with high sensitivity, then detection rate improves, but false positives increase
Solution Approach 1:
The patent applies local quality by analyzing local appearance features (shape, texture, intensity characteristics) of specific colon segments rather than relying solely on global detection algorithms. This localized analysis allows the system to distinguish true polyps from false positives by examining the unique local characteristics of each segment, improving both sensitivity and specificity simultaneously.
Solution Approach 2:
The system segments the colon into multiple segments across different views (prone, supine, lateral) and analyzes each segment independently. By dividing the complex detection task into smaller segment-level analyses, the system can apply specialized feature extraction and matching to each segment, reducing false positives while maintaining high sensitivity through comprehensive coverage.
3Measurement precision
If multi-view matching is performed to assess polyp mobility, then detection accuracy improves, but system complexity increases
Solution Approach 1:
The patent segments the multi-view matching task into independent module: feature extraction module, feature matching module, and registration module. Each module processes specific aspects of the data independently, making the overall complex system manageable through modular architecture. The system processes prone, supine, and lateral views as separate but coordinated components, reducing computational complexity while maintaining comprehensive analysis.
Solution Approach 2:
The system manages complexity by dynamically adjusting parameter thresholds and feature weights based on the specific viewing conditions and colon segment characteristics. Rather than using fixed complex parameters, the system adapts feature extraction and matching parameters to the local context, simplifying the overall system while improving detection accuracy across different views.
Data Source
AI summary
Described herein is a framework for multi-view matching of regions of interest in images. According to one aspect, a processor receives first and second digitized images, as well as at least one CAD finding corresponding to a detected region of interest in the first image. The processor determines at least one candidate location in the second image that matches the CAD finding in the first image. The matching is performed based on local appearance features extracted for the CAD finding and the candidate location. In accordance with another aspect, the processor receives digitized training images representative of at least first and second views of one or more regions of interest. Feature selection is performed based on the training images to select a subset of relevant local appearance features to represent instances in the first and second views. A distance metric is then learned based on the subset of local appearance features. The distance metric may be used to perform matching of the regions of interest.