GPU 2D-3D Deformable Registration for Lung Fluoroscopy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current 2D-3D rigid registration techniques for medical imaging are computationally intensive and suffer from reduced accuracy due to patient motion, particularly in lung procedures, as they fail to accurately capture 3D deformation and handle overlapping or occluded anatomical structures, and do not effectively utilize multiple 2D views for consistent deformation field determination.
Innovation Solution
Implementing a 2D-3D deformable registration method using multiple 2D fluoroscopic views to recover a fully 3D deformation field, which allows for accurate handling of overlapping structures and occlusions, and utilizing a GPU-based multi-core processing framework with a cost pipeline and gradient pipeline to significantly reduce computational time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If 2D-3D rigid registration is performed to align fluoroscopic images with preoperative CT, then registration can be completed with standard computational resources, but accuracy is reduced due to patient motion and non-rigid deformation
Solution Approach 1:
The patent segments the complex 3D deformation field into multiple 2D deformation fields, each corresponding to a specific 2D fluoroscopic view. By dividing the 3D space into view-specific 2D planes and computing deformations independently for each view, the system achieves accurate 3D registration while maintaining computational feasibility through parallel processing of simpler 2D problems.
Solution Approach 2:
The patent transforms the 3D deformation problem into multiple 2D deformation problems by projecting the 3D CT volume onto 2D fluoroscopic views. Each 2D view undergoes separate deformable registration, and the results are integrated to reconstruct the full 3D deformation field, effectively solving the 3D problem through 2D analysis.
2Measurement precision
If deformable registration is performed to capture 3D deformation from patient motion, then registration accuracy improves, but computational time increases significantly
Solution Approach 1:
The patent divides the computationally intensive 3D deformable registration into multiple independent 2D deformable registrations, one for each fluoroscopic view. This segmentation allows parallel computation of simpler 2D problems, significantly reducing total computational time while maintaining accurate capture of 3D deformation through integration of view-specific results.
Solution Approach 2:
The patent computes deformation fields for only the necessary 2D views rather than attempting exhaustive 3D analysis. By performing deformable registration on a subset of 2D projections that capture the essential deformation, the system achieves sufficient accuracy without the prohibitive computational cost of complete 3D deformable registration.
3Measurement precision
If multiple 2D fluoroscopic views are used for registration, then deformation field determination accuracy improves, but system complexity and computational load increase
Solution Approach 1:
The patent merges multiple 2D deformation fields from different fluoroscopic views into a unified 3D deformation field. By integrating the deformation information from multiple 2D views through a consistent 3D reconstruction process, the system achieves accurate and consistent deformation field determination while managing system complexity through modular architecture.
Solution Approach 2:
The patent creates a universal framework that processes multiple 2D fluoroscopic views through a common deformable registration algorithm. The same 2D-to-3D transformation methodology is applied consistently across all views, providing a multi-functional solution that handles various viewing angles and patient positions with a single unified approach.
Data Source
AI summary
Systems and methods for assisting a physician in a medical intervention comprises performing a 2D-3D deformable registration, and more particularly, performing a 2D-3D registration based on multiple live 2D fluoroscopic views, and implemented on a multi-core processing framework such as a Graphics Processing Unit.