Two-Stage 2D-3D Medical Image Registration for Pelvic Fractures
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing 2D-3D medical image registration methods for pelvic fracture reduction surgeries face challenges such as non-convex optimization problems, high computational resource requirements, complexity, and low accuracy due to individual specificities, making them unsuitable for efficient and accurate registration during surgeries.
Innovation Solution
A two-stage registration method combining a regression network based on CNN-Transformer for rough estimation and a fine registration using a point-to-plane 2D-3D registration algorithm with global optical flow and PPC to optimize the initial spatial pose, reducing the need for manual labeling and minimizing radiation exposure.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If iterative optimization based on image similarity is used, then registration accuracy can be improved, but the optimization process is non-convex and easily trapped in local optimal solutions
Solution Approach 1:
The registration process is divided into two distinct stages: rough registration using a regression network to provide an initial pose estimate, and fine registration using iterative optimization to refine the result. This segmentation allows each stage to operate under different assumptions and optimization criteria, preventing the fine registration from being trapped in local optima by starting from a good initial guess provided by the rough registration stage.
Solution Approach 2:
The regression network performs preliminary registration to generate an initial spatial pose before the iterative optimization begins. This preliminary action establishes a reliable starting point that is close to the true solution, thereby reducing the search space for the subsequent iterative optimization and preventing convergence to local optima.
2Ease of operation
If 2D-3D registration based on reinforcement learning is used, then registration can be performed through Markov process, but it requires large amount of parameter space and computing resources
Solution Approach 1:
The invention extracts the essential registration function from the complex reinforcement learning framework and implements it through a regression network that directly maps image data to spatial pose parameters. This extraction eliminates the need for large parameter spaces and complex Markov processes while maintaining automated registration capability.
Solution Approach 2:
Instead of using reinforcement learning to learn optimal policies through trial and error, the regression network is trained to directly copy the mapping relationship between image features and spatial pose. This copying approach achieves automated registration without requiring the extensive computing resources needed for reinforcement learning training and execution.
3Measurement precision
If 2D-3D registration based on anatomical landmark detection is used, then spatial pose can be calculated using pnp algorithm, but it requires large number of manual labels and is complex and time consuming
Solution Approach 1:
The regression network performs self-service by automatically detecting and extracting relevant features from the input images and directly regressing the spatial pose parameters without requiring manual landmark annotation. The network processes the images autonomously, eliminating the time-consuming manual labeling process while maintaining accurate spatial pose estimation.
4Productivity
If 2D-3D registration based on convolution neural network is used, then spatial post can be predicted directly, but accuracy is low due to individual specificities of patients
Solution Approach 1:
The registration system is segmented into two stages: a rough registration stage using a regression network that provides a fast initial estimate, and a fine registration stage that refines the accuracy. This segmentation allows the system to achieve both high speed (from the rough registration) and high accuracy (from the fine registration), overcoming the limitation of single-stage CNN approaches.
Solution Approach 2:
The regression network performs preliminary registration to establish an initial spatial pose quickly, leveraging pre-trained knowledge to provide accurate enough initial estimates. This preliminary action enables subsequent fine registration to focus only on refinement, achieving both speed and accuracy that a single-stage approach cannot achieve.
Data Source
AI summary
The present disclosure relates to a 2D-3D medical image registration method. The method includes obtaining a preoperative CT image and an intraoperative Xray image of a target bone block, inputting the preoperative CT image and the intraoperative Xray image into a depth-based learning-based regression network, and roughly estimating an initial spatial posture of the target bone block by using the regression network; and adjusting a projection of the preoperative CT image based on the initial spatial posture to generate a DRR image; and inputting the DRR image, the intraoperative Xray image, and the preoperative CT image into a pre-trained correspondence point relationship estimation network, estimating a feature point corresponding relationship between the DRR image and the Xray image by using the corresponding point relationship estimation network, and optimizing the initial spatial posture according to the feature point corresponding relationship to obtain an optimized spatial posture.