AI Metal Segmentation in X-ray Projection Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for reducing metal artifacts in x-ray imaging, particularly in computed tomography, face challenges in accurately segmenting metal objects outside the region of interest and achieving robustness in segmentation, leading to poor reconstruction quality.
Innovation Solution
A method using a trained artificial intelligence segmentation algorithm to calculate binary metal masks in the projection domain, combined with a consistency check in a larger three-dimensional reconstruction region, to enhance the robustness and consistency of metal segmentation, allowing for reliable detection of metal objects outside the region of interest and reducing artifacts through inpainting algorithms.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If metal segmentation is performed using a preliminary three-dimensional reconstruction, then the segmentation process is simplified, but metal objects outside the region of interest cannot be reliably detected and segmentation robustness deteriorates
Solution Approach 1:
The patent transitions from performing segmentation solely in the three-dimensional image domain to incorporating a two-dimensional projection domain. By analyzing projection images directly, the system can detect metal objects outside the ROI that would not be visible in the reconstructed 3D volume, thus improving detection reliability while maintaining process feasibility through the use of available projection data.
Solution Approach 2:
The patent performs metal segmentation in the projection domain before the final 3D reconstruction process. By identifying and masking metal artifacts in the projection images beforehand, the subsequent 3D reconstruction is performed on corrected data, improving the reliability of metal object detection including those outside the ROI, while the segmentation workflow remains integrated and manageable.
2Productivity
If simple thresholding is used for metal segmentation in preliminary reconstruction, then the segmentation process is fast and simple, but segmentation accuracy deteriorates due to poor metal artifact representation
Solution Approach 1:
The patent replaces simple thresholding algorithms with a trained artificial intelligence segmentation algorithm. This AI-based approach analyzes the complex patterns in projection images and can accurately distinguish metal artifacts from other structures, significantly improving segmentation accuracy while maintaining computational efficiency through the use of pre-trained models that can be applied rapidly to each projection image.
3Reliability
If artificial intelligence segmentation is applied to projection images, then segmentation robustness is improved, but the approach lacks consistency and reliability as shown in internal studies
Solution Approach 1:
The patent implements an iterative refinement process where the AI segmentation results are used to create an initial 3D reconstruction, which is then forward-projected back to the projection domain. The segmented metal masks from this reconstructed volume are compared with the original AI segmentation of projection images, and discrepancies are resolved through consistency checks. This feedback loop ensures both robustness and consistency in the final metal segmentation results.
Solution Approach 2:
The patent combines multiple segmentation approaches: AI-based segmentation of projection images, thresholding of the 3D reconstructed volume, and forward-projection consistency checks. By merging these different methods and requiring agreement among them, the system achieves both the robustness of AI methods and the consistency of traditional approaches, resolving the limitations of using either method alone.
Data Source
AI summary
A method for segmenting metal objects in projection images acquired using different projection geometries is provided. Each projection image shows a region of interest. A three-dimensional x-ray image is reconstructed from the projection images in the region of interest. A trained artificial intelligence segmentation algorithm is used to calculate first binary metal masks for each projection image. A three-dimensional intermediate data set of a reconstruction region that is larger than the region of interest is reconstructed by determining, for each voxel of the intermediate data set, as a metal value, a number of first binary metal masks showing metal in a pixel associated with a ray crossing the voxel. A three-dimensional binary metal mask is determined. Second binary metal masks are determined for each projection image by forward projecting the three-dimensional binary metal mask using the respective projection geometries.