Deep Learning Framework for Anatomical Functional Image Registration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for registering anatomical and functional medical images in hybrid PET imaging systems are inadequate, leading to errors and artifacts due to subject movement and misalignment between CT and PET images.
Innovation Solution
A deep learning framework that uses two trained convolutional neural networks to extract features and estimate a deformation field, allowing for the registration of anatomical images to functional images, thereby correcting spatial misalignments.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If CT-based attenuation correction is used to correct PET images, then correction quality is improved, but spatial misalignment artifacts occur due to subject movement between scans
Solution Approach 1:
The patent applies preliminary action by performing image registration before attenuation correction. The system first registers the CT image to the PET image using rigid and deformable transformation models, then applies the derived transformation parameters to correct the PET image. This preliminary registration step ensures that the anatomical structures in the CT image align with the functional structures in the PET image before correction is applied, preventing misalignment artifacts.
Solution Approach 2:
The patent uses transformation parameters (rigid and deformable) as intermediaries to bridge the misalignment between CT and PET images. These parameters are derived from registration algorithms that compare anatomical features in both images, and they serve as the mediator that transforms one image space to match the other, enabling accurate attenuation correction without artifacts.
2Speed
If rigid registration is used to align CT and PET images, then alignment speed is improved, but accuracy deteriorates due to physiological motion
Solution Approach 1:
The patent segments the registration process into two distinct stages: rigid registration and deformable registration. The rigid registration stage quickly aligns the images using linear transformations (translation, rotation, scaling), providing a good initial alignment. The deformable registration stage then refines this alignment by applying non-linear transformations to account for physiological motion. This segmentation allows the system to benefit from both the speed of rigid registration and the accuracy of deformable registration.
Solution Approach 2:
The rigid registration serves as a preliminary action that prepares the images for the subsequent deformable registration. By first establishing a coarse alignment through rigid transformation, the system reduces the search space and computational burden for the more intensive deformable registration step, thereby maintaining overall speed while improving final accuracy.
3Measurement precision
If deformable registration is used to correct physiological motion, then alignment accuracy is improved, but computational complexity increases
Solution Approach 1:
The patent segments the complex deformable registration problem into manageable components by first performing rigid registration to establish a baseline alignment. This segmentation reduces the complexity of the deformable registration step, as it only needs to correct for physiological motion rather than handling all types of misalignment. The segmentation strategy makes the overall complex process more computationally tractable.
Solution Approach 2:
The rigid registration acts as a preliminary action that simplifies the subsequent deformable registration task. By pre-aligning the images using computationally efficient rigid transformations, the system reduces the computational burden of the deformable registration step, which can then focus on the more subtle physiological motion corrections without dealing with gross misalignments.
4Measurement precision
If multiple registration stages are used to improve alignment accuracy, then registration quality is improved, but processing time increases
Solution Approach 1:
The patent segments the registration process into hierarchical stages (rigid then deformable) where each stage builds on the previous one. This segmentation allows the system to allocate computational resources efficiently, spending less time on the initial rigid alignment and more time on the refined deformable correction. The segmented approach prevents the need for exhaustive searching at each level, reducing overall processing time while maintaining high quality.
Solution Approach 2:
The rigid registration serves as a preliminary action that quickly establishes a good initial alignment, reducing the amount of work needed in subsequent deformable registration. This preliminary step prevents the system from wasting time on unnecessary computations by providing a strong starting point that is already close to the final solution, thereby reducing total processing time despite multiple stages.
Data Source
AI summary
A framework for registering anatomical to functional images using deep learning. In accordance with one aspect, the framework extracts features by applying an anatomical image and a corresponding functional image as input to a first trained convolutional neural network. A deformation field is estimated by applying the extracted features as input to a second trained convolutional neural network. The deformation field may then be applied to the anatomical image to generate a registered anatomical image.