Multi-Scale Deformable Image Registration via Prediction-Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current deformable image registration methods, such as VoxelMorph and TransMorph, face challenges with large deformations due to high degrees of freedom in transformation parameters, leading to performance degradation, especially in medical image registration for radiotherapy applications where fast and accurate registration is crucial.
Innovation Solution
A multi-scale approach is introduced, where features are extracted at multiple scales, and a prediction-correction framework is used to progressively predict and correct mappings between images, leveraging a pyramidal feature extraction and diffeomorphic registration to achieve improved accuracy and speed for both small and large deformations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If deep neural networks (e.g., VoxelMorph, TransMorph) are used for deformable image registration, then registration speed is improved, but performance degrades for large deformations due to high degrees of freedom in transformation parameters
Solution Approach 1:
The patent segments the transformation parameter space into discrete bins along the deformation trajectory. Instead of optimizing continuous transformation parameters with high degrees of freedom, the method discretizes the deformation path into multiple segments or bins, constraining the search space at each level. This segmentation reduces the complexity of the optimization problem while maintaining the ability to model large deformations through multiple discrete steps.
Solution Approach 2:
The patent introduces a new dimension to the transformation parameter space by adding the trajectory or deformation path as an additional parameter. Rather than directly optimizing the final transformation, the method optimizes over the path taken to achieve it, effectively moving from optimizing end-state parameters to optimizing a sequence of intermediate states. This dimensional extension allows the model to capture large deformations more effectively.
2Measurement precision
If classical optimization methods are used for deformable image registration, then registration accuracy is improved, but processing time becomes excessively long
Solution Approach 1:
The patent performs preliminary discretization of the deformation trajectory before optimization. By pre-defining the bin structure and deformation path segments, the method prepares the parameter space in advance, reducing the computational burden during the actual optimization process. This preliminary structuring allows for faster convergence while maintaining accuracy.
Solution Approach 2:
The patent employs a dynamic optimization approach where the transformation parameters are optimized in a sequential manner through multiple bins or stages. Rather than optimizing all parameters simultaneously in a static framework, the method dynamically progresses through deformation stages, allowing for efficient exploration of the parameter space and faster convergence to accurate solutions.
3Adaptability or versatility
If transformation parameters with high degrees of freedom are used, then flexibility in modeling deformations is improved, but performance degrades for large deformations
Solution Approach 1:
The patent segments the continuous transformation parameter space into discrete bins, organizing the high-dimensional parameter space into manageable segments. This segmentation maintains flexibility by allowing independent optimization within each bin while reducing the overall complexity. The segmented structure enables the model to handle large deformations by composing multiple smaller, more manageable transformation steps.
Solution Approach 2:
The patent adds the deformation trajectory as an additional dimension to the parameter space, transforming the problem from optimizing a single transformation state to optimizing a sequence of states. This dimensional change provides a structured way to model large deformations while controlling the degrees of freedom at each step, balancing flexibility with performance.
Data Source
AI summary
A method for performing deformable image registration on first and second volumetric medical images comprises (i) for each image, extracting features at each of a plurality of different scales, (ii) initiating a mapping between the first and second volumetric medical images at the lowest of the plurality of different scales, and (iii) sequentially, for each scale of the plurality of different scales that is above the lowest scale, predicting a mapping between the first and second volumetric medical images at a given scale, and correcting the predicted mapping between the first and second volumetric medical images at the given scale. The method further comprises (iv) predicting the deformation field between the first and second volumetric medical images at full resolution using the corrected prediction of the mapping at the highest of the plurality of different scales.