3D-2D Anatomical Image Registration Using Neural Initialization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing medical imaging systems face challenges in accurately registering three-dimensional preoperative images with two-dimensional intraoperative images due to differences in image acquisition positions, particularly when using conventional, non-motorized C-arms, which require manual operator input for registration updates and are not cost-effective.
Innovation Solution
A method utilizing neural networks for automatic detection and classification of anatomical structures to determine rigid spatial transformation parameters, enabling registration of three-dimensional and two-dimensional images without manual intervention, employing a first detection neural network for structure identification and a classification neural network for parameter estimation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If a motorized rotational C-arm is used for automatic registration, then registration automation and image fusion accuracy are improved, but device cost and operating room space requirements increase significantly
Solution Approach 1:
The patent replaces the mechanical motorized rotational C-arm system with a software-based neural network registration system that works with conventional mobile C-arms. The neural networks automatically detect anatomical structures and compute transformation parameters, substituting complex mechanical automation with intelligent software processing on standard equipment.
Solution Approach 2:
The system enables the conventional C-arm and registration software to perform automatic registration independently without requiring motorization or specialized hardware. The neural networks self-process the image data and automatically determine spatial transformations, allowing standard equipment to achieve automated functionality.
2Device complexity
If manual operator input is required for registration updates with conventional C-arms, then device cost is reduced, but registration accuracy and time efficiency deteriorate
Solution Approach 1:
The patent substitutes manual operator operations with automated neural network processing. The neural networks automatically detect anatomical structures in real-time images and compute transformation parameters without human intervention, replacing the manual registration process with intelligent software automation while keeping the physical C-arm simple and mobile.
3Measurement precision
If neural networks are trained on patient-specific data, then registration accuracy for that patient is improved, but data processing time and computational resources increase
Solution Approach 1:
The patent performs the training action in advance, before the actual surgical procedure. The neural networks are trained on generic anatomical data sets during system setup or preparation phases, so that when during surgery the pre-trained networks can quickly process patient images without requiring real-time training, thus achieving both accuracy and efficiency.
Solution Approach 2:
The patent uses generic anatomical training data that can be applied universally across different patients rather than patient-specific training. The neural networks learn general anatomical structures and variations that are common to all patients, enabling the same trained models to achieve accurate registration for any patient without requiring individualized training datasets.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The invention relates to a registration method between a first three-dimensional image acquired according to a first acquisition mode, and comprising anatomical structures of a patient and a second two-dimensional image, acquired according to a second acquisition mode by an image acquisition device mounted on a mobile scopy arm in rotation and translation, the second image comprising a part of the anatomical structures of said patient, the registration implementing a rigid spatial transformation defined by rotation and translation parameters.The method includes an automatic detection (52) of anatomical structures in the two-dimensional image by application of a first detection neural network; trained on a generic database, an estimation (54-58), from the anatomical structures automatically detected in said second two-dimensional image, by application of at least one classification neural network previously trained on a generic database, of the rotation and translation parameters of said rigid spatial transformation, and an iconic 3D/2D registration (64) between the first three-dimensional image and the second two-dimensional image starting from an initialization with said rigid spatial transformation.