Anatomically Constrained Image Registration via Pre-computed Transformation Fields
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current deformable image registration methods require explicit landmark setting, which is tedious and application-specific, or rely on intensity invariance assumptions that are often violated, limiting their effectiveness in registering atlas images with clinical images.
Innovation Solution
A system that generates a candidate transformation field based on displacements of landmarks in training images relative to an atlas image, using principal component analysis to compute elementary transformation fields, which are then combined linearly to register the atlas image with an objective image without requiring explicit landmark setting in the objective image.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual landmark setting is used for registration, then registration accuracy is improved, but operation time and complexity increase significantly
Solution Approach 1:
The patent pre-computes a transformation field from training images that encodes anatomical correspondence relationships before actual registration is needed. This preliminary computation stores the deformation patterns that will later enable rapid automatic landmark-free registration, trading initial computation time for operational efficiency.
Solution Approach 2:
The patent uses intensity information from the objective image to sample and copy transformation values from the pre-computed transformation field, creating a registration transformation without requiring explicit landmark correspondence. This copying approach bypasses manual landmark setting while maintaining accuracy.
2Loss of time
If automated landmark setting methods are used, then operation time is reduced, but application-specific design effort and system complexity increase
Solution Approach 1:
The patent creates a universal transformation field from training images that can be applied to multiple different objective images and anatomical structures without requiring application-specific customization. The pre-computed field serves multiple registration tasks, eliminating the need for separate automated landmarking algorithms for each application.
Solution Approach 2:
The system uses the intensity information inherently present in the objective image to automatically determine where to sample from the transformation field, making the registration process self-sufficient without requiring application-specific landmarking algorithms or manual intervention.
3Extent of automation
If intensity-based deformable registration is used, then automation is achieved, but reliability decreases due to violated intensity invariance assumptions
Solution Approach 1:
The patent pre-computes the transformation field using training images with known anatomical correspondences, establishing a reliable deformation model before registration. This preliminary step captures accurate anatomical relationships that are then applied to the objective image, avoiding direct reliance on intensity invariance assumptions during the actual registration process.
Solution Approach 2:
The patent introduces a pre-computed transformation field as an intermediary that mediates between the atlas and objective images. This transformation field, derived from training data, serves as a reliable bridge that transfers anatomical correspondence information without requiring direct intensity-based matching between atlas and objective images.
4Measurement precision
If standard landmark-based registration is used, then registration accuracy is maintained, but ease of operation deteriorates due to tedious manual process
Solution Approach 1:
The patent copies transformation values from the pre-computed transformation field to the objective image based on intensity information, automatically establishing landmark correspondences without manual intervention. This copying mechanism preserves registration accuracy while eliminating the tedious manual landmark setting process.
Solution Approach 2:
The system automatically performs registration by using the objective image's own intensity information to guide sampling from the transformation field, making the process self-serving and eliminating the need for operator intervention in landmark placement while maintaining accuracy.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The invention relates to a system (100) for registering an atlas image from an atlas of multidimensional images with an objective image, the system comprising a generation unit (105) for generating a candidate transformation for transforming a first region of the atlas image, a transformation unit (110) for transforming the first region of the atlas image using the candidate transformation, a similarity unit (120) for computing a measure of similarity of the transformed first region of the atlas image and a corresponding first region of the objective image, an evaluation unit (130) for evaluating the candidate transformation using a criterion based on the computed measure of similarity and determining an optimal transformation based on the evaluation of the candidate transformation, an extension unit (140) for extending the optimal transformation of the first region of the atlas image to a second region of the atlas image, wherein the second region comprises the first region, thereby creating a registration transformation and a registration unit (150) for transforming the second region using the registration transformation, thereby registering the atlas image with the objective image. Advantageously, the system (100) does not require setting landmark positions in the objective image. A further advantage of the system (100) is that since the similarity measure is computed locally, i.e., based on the first region of the atlas image, the registration is fast and thus attractive for clinical use.