Deformable Medical Image Registration Using Time-Varying NFODE
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deformable image registration methods using convolutional neural networks (CNNs) or transformer networks are sensitive to data distribution changes and struggle to accurately model geometric transformations between medical images from different modalities, requiring computationally inefficient test-time optimization.
Innovation Solution
A method utilizing a Neural Field Ordinary Differential Equation (NFODE) parameterized by an Implicit Neural Representation (INR) to estimate a time-varying velocity field and integrate it to generate a deformation field, enabling efficient and accurate registration of volumetric medical images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If test-time optimization is used to address data distribution changes and improve registration accuracy, then registration accuracy is improved, but computation efficiency deteriorates due to requiring finetuning or optimizing a large network at test time
Solution Approach 1:
The model performs preliminary actions during training by learning robust features and transformation patterns from diverse datasets. The neural network is pre-trained to handle various imaging modalities and deformation types, so that at test time, the model can directly apply learned knowledge without requiring computationally intensive finetuning or optimization, thus resolving the contradiction between accuracy and computation efficiency
2Measurement precision
If SIREN network with large capacity is used to model complex deformation fields, then registration accuracy is improved, but device complexity increases
Solution Approach 1:
The patent changes the parameterization approach by using neural ordinary differential equations (Neural ODEs) with time-varying parameters instead of fixed-capacity SIREN networks. The Neural ODE framework models deformation fields through continuous time evolution, allowing the system to adapt its computational complexity dynamically based on the specific registration task requirements, thereby achieving high accuracy without consistently requiring large model capacity
Solution Approach 2:
The invention introduces dynamics into the model by using time-varying parameters in the Neural ODE formulation. The deformation field is modeled as a dynamic process evolving over time, allowing the network to adaptively adjust its computational resources during the integration process. This dynamic approach enables the system to achieve complex deformation modeling only when necessary, reducing overall model complexity compared to fixed-capacity architectures
Data Source
AI summary
A method for performing deformable image registration of a first volumetric medical image to a second volumetric medical image may comprise estimating a time varying velocity field between the images by encoding coordinates using a time varying positional embedding and using the encoded coordinates and a Neural Field Ordinary Differential Equation (NFODE) to generate a prediction of the rate of change of the deformation field. The method may further comprise integrating the estimated velocity field to generate a deformation field and applying the generated deformation field to the first volumetric medical image to generate a registered volumetric medical image. The NFODE May comprise a non-stationary Neural ODE parameterized by an Implicit Neural Representation.