AI Geometric Alignment for Raw CT Image Standardization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Computed tomography (CT) images exhibit geometric variability due to differences in patient skull size and shape, imaging direction, and imaging protocols, which affects diagnostic accuracy and consistency across different imaging systems and protocols.
Innovation Solution
An artificial intelligence-based method and device that performs geometric alignment and preprocessing of raw CT images by using a head segmentation unit to predict bone and brain masks, followed by rigid and non-rigid registration to standardize the images, ensuring consistent transformation and alignment.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If different imaging protocols are used for CT scans, then the imaging can be adapted to different patient conditions and preferences, but the geometric variability of the images increases
Solution Approach 1:
The patent introduces an intermediary registration process that acts as a mediator between diverse imaging protocols and the final analysis system. The method uses protocol identification to detect the imaging protocol used, then applies protocol-specific registration parameters and transformation algorithms to standardize images from different protocols into a common reference frame, thereby resolving the geometric variability caused by protocol diversity
Solution Approach 2:
The patent systematically changes registration parameters based on the identified imaging protocol. Different protocols (e.g., head CT, chest CT, abdominal CT, cone-beam CT) have different optimal registration parameters including transformation type (rigid, affine, or non-rigid), landmark selection criteria, and alignment priorities. The system dynamically adjusts these parameters to maintain geometric consistency across protocols
2Measurement precision
If geometric alignment preprocessing is applied to CT images, then diagnostic accuracy and consistency are improved, but the processing time and computational complexity increase
Solution Approach 1:
The patent performs geometric alignment preprocessing as a preliminary step before diagnostic analysis or deep learning inference. By pre-registering images to a standard anatomical reference frame and pre-identifying protocol-specific characteristics, the system prepares the data in advance, reducing the computational burden during actual diagnostic operations and enabling faster real-time analysis
Solution Approach 2:
The patent segments the geometric alignment process into distinct modular steps: protocol identification, image preprocessing, landmark detection, transformation application, and quality verification. This segmentation allows each step to be optimized independently and enables selective application of processing intensity based on clinical needs, balancing accuracy with processing time
3Adaptability or versatility
If deep learning models are trained on CT images with geometric variability, then the model can handle diverse imaging scenarios, but the model performance and stability decrease
Solution Approach 1:
Instead of training the deep learning model to be invariant to geometric variations (the conventional approach), the patent inverts the strategy by pre-aligning all input images to a standard reference frame before they reach the model. This inversion transforms the problem from teaching the model to handle variability to ensuring variability is removed before inference, resulting in more stable and reliable model performance
Data Source
AI summary
The present disclosure relates to artificial intelligence-based devices and methods for geometric alignment and preprocessing of raw CT images. According to the present disclosure, when sensitive to a local error according to each characteristic, only linear transformation is applied to use the centrally aligned rigid registration result, and when the region of interest between patients is spatially matched or monitoring is required, a non-rigid registration result may be used.