GPU Tomographic Image Reconstruction Using 3D Blurring Kernels
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current tomographic image reconstruction methods on CPUs are computationally intensive due to large memory requirements and slow processing times, especially for list-mode 3D reconstruction, which limits practical applications to Filtered Back projection or rebinned 2D-OSEM, and existing GPU implementations are limited to fan-beam and parallel beam X-ray CT and do not support line forward- and back-projection or blurring kernels.
Innovation Solution
A method using a graphics processing unit (GPU) for forward- and back-projection of tomographic images, employing a 3-dimensional blurring kernel and trilinear interpolation, to reconstruct images efficiently, allowing for the implementation of line projection algorithms and reducing computational costs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If list-mode 3D-OSEM reconstruction is performed on a CPU, then image reconstruction can be performed with high flexibility and adaptability, but the computation time and processing speed become excessively long
Solution Approach 1:
The patent replaces the CPU-based mechanical computation system with a GPU-based parallel processing system. The GPU's architecture with thousands of cores enables simultaneous execution of multiple projection operations, substituting the sequential processing of CPUs with parallel processing, thereby dramatically reducing computation time while maintaining reconstruction flexibility through programmable shader languages.
Solution Approach 2:
The patent segments the image reconstruction computation into independent projection operations that can be executed in parallel on GPU cores. By dividing the computational task into discrete, concurrent operations (forward projections, back projections, and updates), the system achieves both speedup and adaptability through the GPU's parallel architecture.
2Measurement precision
If the number of detector elements is increased to improve resolution and sensitivity, then image quality improves, but the memory requirements for storing the system matrix become excessively large
Solution Approach 1:
The patent extracts and eliminates the system matrix from the reconstruction process entirely. Instead of storing the complete system matrix in memory, the method computes projections and back-projections on-the-fly using analytical geometry and ray-tracing algorithms, thereby removing the memory bottleneck while preserving image resolution through accurate projection modeling.
Solution Approach 2:
The patent performs preliminary computation of projection geometries and ray paths before reconstruction, storing only essential geometric parameters rather than the full system matrix. This preliminary preparation enables efficient on-the-fly projection calculations during iterative reconstruction, reducing memory requirements while maintaining computational accuracy.
3Productivity
If a large cluster of computers is used to perform 3D list-mode reconstruction, then computation speed improves, but the device complexity and cost increase significantly
Solution Approach 1:
The patent merges multiple computation functions (forward projection, back projection, image update, and filtering) into a single integrated GPU system. By combining these previously separate computational tasks that would require multiple computers into one unified parallel processing architecture, the system achieves cluster-level performance with a single device, thereby reducing complexity while maintaining high productivity.
Solution Approach 2:
The patent creates a universal GPU-based reconstruction platform that can handle multiple reconstruction algorithms (OSEM, MLEM, FBP), different imaging modalities (PET, SPECT, CT), and various projection geometries through programmable shaders. This multi-functional capability replaces the need for specialized computer clusters for different reconstruction tasks with a single versatile system.
4Ease of manufacture
If existing GPU implementations use only 2D texture-mapping capability for fan-beam and parallel beam reconstruction, then implementation is simpler, but the adaptability to handle line projection and blurring kernels is limited
Solution Approach 1:
The patent transitions from 2D texture-mapping to 3D volumetric rendering capabilities of the GPU. By utilizing the GPU's 3D texture memory and volumetric rendering pipelines, the system can handle 3D list-mode projection data and apply 3D blurring kernels, thereby extending adaptability from 2D to 3D while leveraging the same parallel processing architecture.
Solution Approach 2:
The patent changes the fundamental parameters of GPU utilization from 2D texture coordinates to 3D volumetric coordinates and ray-tracing parameters. By modifying how the GPU processes geometric data (from pixel-based 2D mapping to ray-based 3D projection), the system achieves versatility in handling different projection geometries and blurring operations while maintaining implementation efficiency through shader programming.
Data Source
AI summary
The present invention provides a method of reconstructing a tomographic image. In a first step, a tomographic image is forward-projected along a list of geometrical lines in a GPU. This list of geometrical lines may be list-mode event data acquired from a tomographic scanner. Alternatively, the list may be a list of weighted lines derived from a sinogram, a histogram, or a timogram acquired from a tomographic scanner. Next, the list of geometrical lines is back-projected into a 3-dimensional volume using the GPU. The results of the forward- and back-projection are then used to reconstruct the tomographic image, which is then provided as an output, e.g. to make the image available for further processing. Examples of output include storage on a storage medium and display on a display device.