Synthetic Image Registration for Low-Dimensional Landmark Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for registering low-dimensional images with high-dimensional images are either dependent on good initialization, which can fail if the initial pose is incorrect, or require invasive marker implantation, and lack fully automatic solutions.
Innovation Solution
A computer-implemented method that simulates synthetic low-dimensional images from high-dimensional images using learning algorithms to determine landmark positions, allowing for automatic registration without manual initialization or marker implantation, by projecting these positions back into the high-dimensional space and using them to align low-dimensional images with high-dimensional images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If iterative optimization methods with image metrics are used for registration, then registration accuracy can be improved, but the method requires good initialization and fails when the initial pose is incorrect
Solution Approach 1:
The patent applies preliminary action by using a pre-trained neural network to automatically detect landmarks and estimate the initial pose before the iterative optimization process. This preliminary estimation provides a reliable starting point that guides the iterative algorithm, enabling it to converge to the correct registration even when the initial pose is incorrect, thus resolving the contradiction between accuracy and robustness to poor initialization.
2Ease of manufacture
If marker-based registration methods are used, then registration can be performed without iterative optimization, but it requires invasive marker implantation into the object or living being
Solution Approach 1:
The patent uses copying by creating synthetic low-dimensional images from the high-dimensional scan data. These synthetic images serve as virtual copies that contain the necessary information for landmark detection and pose estimation. By working with these copied synthetic images rather than requiring physical markers on the patient, the system achieves simple registration without invasive procedures.
Solution Approach 2:
The patent replaces the mechanical system of physical marker implantation with a computational system using neural networks and image simulation. Instead of mechanically placing markers on the patient's body, the system uses software-based synthetic image generation and automated landmark detection to achieve the same registration function, eliminating the harmful invasive procedure.
3Object-affected harmful factors
If anatomical landmarks are used instead of markers, then no invasive procedure is needed, but it requires sufficient landmarks to be accurately recognized in both modalities and correctly assigned to each other
Solution Approach 1:
The patent creates synthetic low-dimensional images from the high-dimensional scan, providing a controlled copy that contains clearly defined anatomical landmarks. This synthetic copy serves as a ground truth for training the neural network to accurately detect and recognize landmarks. By using this copied data for training, the system achieves high landmark recognition accuracy without requiring invasive markers.
Solution Approach 2:
The patent implements feedback through the training process where synthetic images with known landmark positions are used to train the neural network. The network learns from this feedback loop, continuously improving its ability to accurately detect and assign landmarks in real low-dimensional images. This feedback mechanism ensures high recognition accuracy while avoiding invasive procedures.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A computer-implemented method for registering low dimensional images with a high dimensional image, the method comprising the steps of: a) receiving a high dimensional image (1) of a region of interest, b)simulating synthetic low dimensional images (2) of the region of interest from a number of poses of a virtual low dimensional imaging device (6), from the high dimensional image (1), c) determining positions of landmarks (5) within the low dimensional images (2) by applying a first learning algorithm (C1) to the low dimensional images (2), d) back projecting of the positions of the determined landmarks (5) into the high dimensional image space, to thereby obtain the positions of the landmarks (5;7) determined in step c) in the high dimensional image (1), e) receiving low dimensional images (3) acquired with a low dimensional imaging device of the region of interest, f) determining positions of landmarks (5) within the low dimensional images (3) by applying the first or a second learning algorithm (C1,C2) to the low dimensional images (3), and g) registering the low dimensional images (3) with the high dimensional image (1) based on the positions of the landmarks (5). Further, a method for training an artificial neural network useful in finding landmarks in low dimensional images, a computer program and a system for registering low dimensional images with a high dimensional image is provided.