Gradient Vector Intersection for Flat Polyp Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional medical image processing techniques, such as the divergent gradient field response (DGFR) candidate generator, struggle to accurately and efficiently detect asymmetric, flat, and difficult-to-detect anatomical structures like polyps in digital medical images due to their 2D limitations and high sensitivity to false positives, leading to missed candidates and computational inefficiencies.
Innovation Solution
A method that calculates an intersection score for gradient vectors at points on a polypoid structure, identifying candidates by determining if the gradient vectors of neighboring points intersect or nearly intersect, and merging candidate points to improve detection accuracy and reduce false positives, using a program storage device to execute these steps in a computer-aided detection system.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If conventional DGFR candidate generator is used, then detection speed is improved, but detection accuracy for flat and asymmetric structures deteriorates
Solution Approach 1:
The patent segments the detection process into multiple stages: initial candidate generation using DGFR, followed by refined candidate selection using gradient vector intersection analysis. This segmentation allows the system to maintain fast initial screening while applying more accurate but computationally intensive methods only to relevant regions, thus resolving the contradiction between speed and accuracy.
Solution Approach 2:
The patent transitions from 2D gradient analysis to 3D gradient vector field analysis by examining gradient vectors in three-dimensional space and calculating their intersection points. This dimensional enhancement enables accurate detection of flat and asymmetric structures that cannot be properly detected using conventional 2D methods, while maintaining computational efficiency through selective application to candidate regions.
2Productivity
If conventional 2D gradient analysis is used, then computational efficiency is improved, but detection capability for three-dimensional structures deteriorates
Solution Approach 1:
The patent extends gradient analysis from two-dimensional image space to three-dimensional gradient vector space by computing gradient vectors with both magnitude and direction components in 3D. The intersection of these 3D gradient vectors provides robust detection of volumetric structures like polyps, maintaining computational efficiency through hierarchical processing that applies 3D analysis only where needed.
Solution Approach 2:
The patent introduces gradient vector intersection points as intermediary markers that bridge 2D image data and 3D structural interpretation. These intersection points serve as computational mediators that encode three-dimensional structural information without requiring full 3D reconstruction, thus maintaining computational efficiency while enhancing detection capability.
3Reliability
If high sensitivity detection is applied, then detection coverage is improved, but false positive rate increases
Solution Approach 1:
The patent implements a feedback mechanism where gradient vector intersection analysis provides secondary validation of DGFR-generated candidates. Candidates that satisfy both DGFR criteria and gradient intersection criteria are confirmed as true positives, while those failing the intersection test are rejected as false positives. This feedback loop maintains high detection coverage while systematically reducing false positives.
Solution Approach 2:
The patent performs preliminary gradient vector intersection analysis on DGFR candidate regions before final candidate confirmation. This preliminary action filters out false positives early in the process, ensuring that only candidates with strong gradient convergence evidence proceed to final detection, thus maintaining high sensitivity while reducing false positive rates.
Data Source
AI summary
A method for generating candidates from a digital image includes considering at least one point x that may lie on a polypoid structure, determining whether the point x satisfies a first predetermined set of conditions, for each point x that satisfies the predetermined set of conditions, identifying each neighbor point y within a predetermined distance of point x that satisfies a second predetermined set of conditions, determining a gradient vector v1 for point x and identifying a first half-line to which the gradient vector v1 belongs, determining a gradient vector v2 for point y and identifying a second half-line to which the gradient vector v2 belongs, calculating an intersection score that represents how close the first and second half-lines come to intersecting, and identifying point x as a candidate when a candidate score is greater than a predetermined value, wherein the candidate score is the sum of intersection scores for all neighbor points y.