Image Segmentation Using ML-Initialized Shape-Constrained Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image segmentation methods, such as model-based segmentation (MBS) and machine learning (ML) segmentation, face challenges in accurately segmenting anatomical structures that are outside the permitted deformation range of MBS models or lack shape constraints, leading to spurious or non-smooth results.
Innovation Solution
A method that combines ML and MBS by using a machine learning model to predict shape descriptors for initializing MBS, allowing for a more accurate fit by aligning the MBS model with the ML segmentation results, thereby capturing extreme anatomical geometries within the MBS capture range.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If model-based segmentation (MBS) is used with shape constraints, then segmentation results are smooth and robust to image artifacts, but edge cases with extreme anatomical geometries result in spurious fits
Solution Approach 1:
The patent applies preliminary action by using a machine learning model to predict shape descriptors before the MBS fitting process begins. The ML model processes the image and outputs preliminary shape descriptors that are then used to initialize the MBS model, allowing extreme anatomical geometries to be captured before the constrained fitting starts.
Solution Approach 2:
The patent introduces an intermediary mechanism where the ML model acts as a mediator between the image input and the MBS fitting process. The ML model predicts shape descriptors that serve as intermediaries to bridge the gap between raw image data and the constrained MBS model, enabling better initialization for edge cases.
2Adaptability or versatility
If machine learning segmentation is used without shape constraints, then any shape can be fitted, but segmentation accuracy decreases and results become non-smooth
Solution Approach 1:
The patent segments the segmentation process into two distinct stages: first, an ML-based coarse segmentation that captures any shape with flexibility; second, an MBS-based fine segmentation that enforces smoothness and accuracy. This segmentation of the processing pipeline allows each method to发挥 its strengths.
Solution Approach 2:
The patent applies local quality by using different characteristics in different parts of the process: the ML model provides global shape flexibility and coarse segmentation, while the MBS model provides local smoothness and precision. Each method operates where its specific quality is most needed.
3Measurement precision
If MBS is used with permitted Eigenmode deformations, then typical anatomical shapes are segmented accurately, but deformations towards the limit of permitted range produce spurious fits
Solution Approach 1:
The ML model performs preliminary shape descriptor prediction that captures extreme deformations before the MBS fitting begins. This preliminary action allows the system to handle edge cases with extreme anatomical geometries that would otherwise be outside the permitted MBS range.
Solution Approach 2:
The patent changes the parameter initialization by using ML-predicted shape descriptors as initial values for the MBS fitting process. This parameter change allows the MBS model to start from a more accurate initial state that reflects extreme anatomical variations, improving reliability at deformation limits.
Data Source
AI summary
In a method of segmenting a feature in an image, an image product related to the image is provided (102) to a model trained using a machine learning process. An indication of a shape descriptor for the feature in the image is received (104) from the model, based on the image product. The indicated shape descriptor is then used (106) in a model based segmentation, MBS, to initialize the MBS and segment the feature.