Oriented Active Shape Model for Medical Image Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing medical image segmentation methods, such as Active Shape Models (ASM), face challenges including inaccurate boundary matching, high requirements for landmark points and training samples, sensitivity to initialization, and reliance on statistical information, leading to inefficient and inaccurate segmentations.
Innovation Solution
The Oriented Active Shape Models (OASM) method combines an active shape model with a live wire cost function and dynamic programming to improve segmentation accuracy by automatically initializing landmarks and determining the smallest cost contour, reducing human interaction and improving boundary detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If Active Shape Model methods are used for segmentation, then the segmentation process can be automated, but the boundary detection accuracy deteriorates due to poor matching between parametric descriptions and perceptual boundaries
Solution Approach 1:
The patent combines Active Shape Model (ASM) with Live Wire methodology to create a hybrid approach. The ASM provides automated landmark initialization while Live Wire performs accurate boundary tracing between landmarks. This merging resolves the contradiction by retaining automation from ASM while achieving precision through Live Wire's pixel-level boundary detection that matches perceptual boundaries.
Solution Approach 2:
The patent introduces dynamic programming as an intermediary optimization technique that selects the minimum cost path between ASM-initialized landmarks. This intermediary process refines the boundary detection by choosing optimal pixel sequences that accurately follow perceptual boundaries, thereby improving measurement precision while maintaining the automated initialization benefit.
2Measurement precision
If many landmark points are used to model the shape, then the shape representation accuracy improves, but the training time and computational cost increase
Solution Approach 1:
The patent uses a limited number of strategically placed landmarks from ASM rather than densely sampling the entire boundary. This partial action approach places landmarks at key anatomical locations sufficient for shape characterization, avoiding the excessive computation that would result from using many more landmarks, thus reducing training time while maintaining adequate shape representation accuracy.
3Productivity
If a small search region is used around each landmark, then the computational efficiency improves, but the segmentation accuracy deteriorates as the ASM may miss the true boundary
Solution Approach 1:
The patent introduces Live Wire as an intermediary process that operates between landmark placement and final boundary determination. Live Wire performs an optimized search along the image gradient, effectively extending the search capability beyond the small ASM search region while maintaining computational efficiency through dynamic programming. This intermediary step ensures the true boundary is found without sacrificing productivity.
4Measurement precision
If manual landmark placement is used, then the initialization accuracy improves, but the human interaction requirement increases
Solution Approach 1:
The patent implements self-service initialization through Automatic Landmark Placement (ALP) that uses image intensity gradients and anatomical priors to automatically position landmarks without human intervention. This self-service mechanism achieves sufficient initialization accuracy to kickstart the automated segmentation process, eliminating the need for manual landmark placement while maintaining adequate starting precision for subsequent refinement.
Data Source
AI summary
An improved method of segmenting medical images includes aspects of live wire and active shape models to determine the most likely segmentation given a shape distribution that satisfies boundary location constrains on an item of interest. The method includes a supervised learning portion to train and learn new types of shape instances and a segmentation portion to use the learned model to segment new target images containing instances of the shape. The segmentation portion includes an automated search for an appropriate shape and deformation of the shape to establish a best oriented boundary for the object of interest on a medical image.