Liver Segmentation via Global-to-Local Shape Matching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current liver segmentation methods in CT images face challenges due to similar densities of liver tissue and adjacent organs, leading to weak edges and large inter-patient shape variability, which complicates accurate boundary extraction and requires time-consuming human interaction.
Innovation Solution
A two-phase segmentation method using a statistical shape model to estimate pose and global shape properties, followed by non-rigid template registration to recover local deformations, preserving liver topology even with large shape variations, and optimizing image terms for intensity likelihood within the region of interest and its background.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a statistical shape model is used to constrain segmentation, then robustness is improved, but accuracy deteriorates due to large inter-patient shape variability
Solution Approach 1:
The patent divides the segmentation process into two distinct phases: a first phase using a statistical shape model to obtain a preliminary segmentation with good robustness, and a second phase using non-rigid template registration to refine local shape properties and recover accurate liver boundaries. This segmentation of the segmentation process resolves the contradiction by applying different methods suited to different requirements at different stages.
Solution Approach 2:
The patent performs a preliminary segmentation using the statistical shape model before applying the more accurate but computationally intensive non-rigid template registration. This preliminary action provides a good initial estimate that guides the subsequent refinement process, ensuring both robustness and accuracy are achieved.
2Measurement precision
If non-rigid template registration is applied to recover local deformations, then accuracy is improved, but computational complexity increases
Solution Approach 1:
The statistical shape model performs a preliminary segmentation that provides a good initial estimate of liver pose and global shape. This preliminary action reduces the search space and initialization complexity for the subsequent non-rigid template registration, making the computationally intensive second phase more feasible.
Solution Approach 2:
The patent extracts and focuses computational effort on recovering only the local shape properties that were not captured by the global statistical shape model. By separating global and local shape recovery tasks, the method reduces unnecessary computational complexity while maintaining accuracy.
3Productivity
If automated segmentation is implemented, then productivity is improved, but reliability deteriorates due to challenges in detecting weak edges
Solution Approach 1:
The patent applies different quality standards and methods to different aspects of segmentation: using intensity-based methods for global pose estimation and shape-constrained methods for local boundary refinement. This local quality approach ensures reliable boundary detection even when edges are weak, while maintaining automation efficiency.
Solution Approach 2:
The patent combines multiple segmentation approaches (statistical shape modeling, intensity-based segmentation, and non-rigid template registration) into a composite method. Each component contributes its strengths, and their integration produces a reliable automated segmentation system that overcomes the limitations of individual methods.
Data Source
AI summary
A method for automatically segmenting a liver in digital medical images includes providing a 3-dimensional (3D) digital image I and a set of N training shapes {φi}i=1, . . . , N for a liver trained from a set of manually segmented images, selecting a seed point to initialize the segmentation, representing a level set function φα(θx+h) of a liver boundary Γ in the image asϕα(x)=ϕ0+∑i=1nαiVi(x),whereϕ0(x)=1N∑i=1Nϕi(x)is a mean shape, {Vi(x)}i=1, . . . , n are eigenmodes where n<N, αi are shape parameters, and h ε R3 and θε [0,2π]3 are translation and rotation parameters that align the training shapes, minimizing a first energy functional to determine the shape, translation, and rotation parameters to determine a shape template for the liver segmentation, defining a second energy functional of the shape template and a registration mapping weighted by image intensity histogram functions inside and outside the boundary, and minimizing the second energy functional to determine the registration mapping, where the registration mapping recovers local deformations of the liver.