Registration Grid Detection in Low-Dose Surgical Scanner Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for identifying and determining the 3D transformation of a predefined object, such as a registration grid with fiducial markers, in images from an intraoperative scanner are cumbersome, complex, and require significant human intervention, leading to inaccuracies and increased patient risk.
Innovation Solution
A neural network-based method using convolutional neural networks (CNNs) for automatic identification and tracking of registration grids, invariant to grid type and patient anatomy, which includes a denoising process for low-dose medical images to enhance accuracy and reduce radiation exposure.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If manual methods are used to identify and track fiducial markers in medical images, then the registration process can be performed, but the process becomes cumbersome, complex, and requires significant human intervention
Solution Approach 1:
The patent replaces manual mechanical identification and tracking methods with an automated neural network-based computer vision system. The neural network automatically detects fiducial markers in medical images and computes transformation matrices, eliminating the need for manual marker identification and reducing human intervention in the registration process.
Solution Approach 2:
The system enables self-service automation where the neural network independently performs marker detection, image registration, and transformation computation without requiring human operators to manually identify markers or calculate transformations. The system serves itself by automatically processing images and generating registration results.
2Object-affected harmful factors
If low-dose scanning is used to reduce radiation exposure, then patient radiation exposure decreases, but image quality and signal-to-noise ratio deteriorate
Solution Approach 1:
The patent converts the harmful noise present in low-dose images into a beneficial opportunity for demonstrating neural network robustness. Instead of treating noise as a problem to be eliminated, the system uses denoising autoencoders that are specifically trained to recognize and extract fiducial markers from noisy low-dose images, thereby maintaining registration accuracy while minimizing radiation exposure.
Solution Approach 2:
The system performs preliminary denoising and feature extraction using pre-trained neural networks before the actual registration process. By pre-processing low-dose images through denoising autoencoders and using data augmentation during training, the system prepares enhanced image representations that maintain measurement precision while working with low-dose input data.
3Loss of information
If complex grid designs with multiple marker types are used, then more information can be captured, but the design and manufacturing complexity increases
Solution Approach 1:
The patent employs a universal neural network architecture that can handle multiple types of fiducial markers (spheres, cubes, planar markers) and various grid configurations through a single unified system. The neural network is trained on diverse marker types and grid designs, enabling it to automatically adapt to different marker geometries and arrangements without requiring separate specialized systems for each marker type.
Solution Approach 2:
The system uses parameter changes in the neural network training process to adapt to different grid designs and marker types. By varying training parameters, data augmentation techniques, and network architecture configurations, the system can effectively process diverse marker geometries and grid arrangements while maintaining a relatively simple physical grid design.
Data Source
Figure 1A~1C
Figure 1D~1F
Figure 1G~1H
AI summary
A computer-implemented system, comprising: at least one nontransitory processor-readable storage medium that stores at least one of processor-executable instructions or data; and at least one processor communicably coupled to at least one nontransitory processor-readable storage medium, wherein at least one processor: reads (811) the set of images, each comprising an image of a registration grid (181) comprising an array of markers (181B); detects (812) the markers on each of the images by using a marker detection convolutional neural network (CNN); filters (821) the detection results to remove false positives; and determines (831) the 3D location and 3D orientation of the registration grid (181) based on the detected markers (181B).