Feature-Map Image Registration for Lower 3D U-Net Compute
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image registration methods for dynamic contrast examinations in medical imaging, such as CT and MRI, face challenges in efficiently handling large amounts of data and require significant computational resources due to the need for 3D U-net calculations and storage of deformation vector fields for multiple image combinations, especially when body movement causes positional deviations between images.
Innovation Solution
An image processing method using a two-stage neural network architecture, where a first neural network extracts feature maps from individual images and a second neural network calculates deformation vector fields from these feature maps, reducing the computational burden by allowing feature maps to be calculated once and reused for multiple image registrations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a 3D U-net architecture is used for image registration, then registration accuracy is improved, but computational resources and processing time increase significantly
Solution Approach 1:
The patent segments the registration process into two distinct stages: a coarse registration stage using a simplified model to align images broadly, followed by a fine registration stage using the 3D U-net architecture to refine alignment precision. This segmentation allows the computationally intensive high-accuracy model to operate only on pre-aligned images, reducing overall processing time while maintaining registration accuracy.
Solution Approach 2:
The patent performs preliminary coarse registration using a computationally efficient model before applying the 3D U-net architecture. This preliminary alignment reduces the search space and initial misalignment, allowing the high-accuracy model to converge faster and requiring fewer computational resources to achieve the same registration quality.
2Measurement precision
If deep learning-based registration is applied to multiple image combinations, then registration quality is improved, but calculation resources and storage requirements increase
Solution Approach 1:
The patent merges multiple registration results by selecting the optimal transformation parameters from different image combinations. Instead of storing all intermediate deformation vector fields from multiple deep learning registrations, the system combines results by comparing transformation matrices and selecting the best alignment, significantly reducing storage requirements while maintaining registration quality.
Solution Approach 2:
The patent discards intermediate deformation vector fields and transformation parameters after they are used to generate the final registered images. Only the essential final transformation parameters and registered images are retained, allowing the system to recover the necessary information for clinical use while discarding redundant computational data that would consume excessive storage capacity.
3Measurement precision
If comprehensive image registration is performed for all time phases, then diagnostic accuracy is improved, but processing time increases
Solution Approach 1:
The patent applies local quality optimization by focusing comprehensive registration only on regions containing lesions or areas of diagnostic interest, while using simplified registration for homogeneous background regions. This allows the system to maintain high diagnostic accuracy for critical areas while reducing processing time for less critical regions through the two-stage approach.
Data Source
AI summary
Provided are an image processing method, an image processing device, a program, and a trained model that can suppress calculation resources required for registration between a plurality of images.An image processing method that is executed by one or more processors includes: allowing the one or more processors to acquire feature maps of each of a plurality of images; and allowing the one or more processors to calculate a deformation vector field from a combination of the feature maps of each image. The one or more processors generate the feature maps respectively from the plurality of images using a first neural network and input the generated combination of the feature maps to a second neural network to calculate the deformation vector field.