Deep Neural Network 3D CT Reconstruction from Sparse X-ray Projections
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for reconstructing 3D computed tomography (CT) volumes from X-ray radiographs require numerous images, increasing radiation exposure, time, and cost, and existing approaches for generating 3D segmentation masks from limited X-ray images are either manual, semi-automatic, or rely on statistical shape models, which are inefficient.
Innovation Solution
A method using a trained deep neural network, specifically a deep image-to-image network within a conditional-generative adversarial network, to generate a sparse 3D volume from a small number of X-ray images, which is then decoded into a final reconstructed 3D CT volume and segmentation mask, reducing the need for extensive image data and improving automation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If many X-ray images are taken from various angles for CT reconstruction, then the quality and detail of the 3D CT volume is improved, but patient radiation exposure increases and acquisition time and cost increase
Solution Approach 1:
The patent applies partial action by using only a limited number of X-ray images (e.g., 3-7 images) instead of the many images required by conventional CT, achieving sufficient reconstruction quality for segmentation purposes while reducing radiation exposure. The deep learning model compensates for the limited data through learned priors from training on complete CT scans.
Solution Approach 2:
The patent introduces a deep learning model as an intermediary that transforms a sparse set of X-ray images into a reconstructed 3D volume. This intermediary learns the mapping from limited projections to complete volumetric data, enabling quality reconstruction without requiring numerous input images.
2Measurement precision
If many X-ray images are taken for CT reconstruction, then the accuracy of the 3D CT volume is improved, but acquisition time increases
Solution Approach 1:
The method uses a partial set of X-ray images (minimal subset) rather than complete angular sampling, reducing acquisition time while maintaining sufficient accuracy for anatomical segmentation through deep learning-based reconstruction.
3Measurement precision
If manual or semi-automatic methods are used for generating 3D segmentation masks from limited X-ray images, then some level of accuracy can be achieved, but the process is inefficient and time-consuming
Solution Approach 1:
The system applies self-service by using the reconstructed 3D volume to automatically generate segmentation masks through the deep learning model, eliminating the need for manual or semi-automatic segmentation processes. The model learns anatomical structures from training data and applies this knowledge to automatically segment the reconstructed volumes.
Solution Approach 2:
The patent replaces manual mechanical segmentation processes with an automated deep learning-based system that processes the reconstructed 3D volume to generate segmentation masks, significantly improving efficiency while maintaining or improving accuracy.
4Productivity
If conventional tomographic reconstruction algorithms are used with limited X-ray images, then a sparse 3D volume can be generated, but the quality and completeness of the reconstructed volume is insufficient for clinical use
Solution Approach 1:
The deep learning model serves as an intermediary that enhances the sparse volume generated by conventional algorithms, transforming it into a complete, high-quality 3D CT volume suitable for clinical segmentation tasks.
Solution Approach 2:
The method changes the parameter of image completeness by using the deep learning model to fill in missing information and enhance the sparse reconstruction, transforming it from an incomplete representation to a clinically useful volumetric dataset.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
This approach allows for the automated reconstruction of 3D CT volumes and segmentation masks from a minimal number of X-ray images, reducing patient radiation exposure, time, and costs while providing accurate anatomical reconstructions, enabling efficient pre-operative planning and surgery guidance.
Implementation Method 1
generating a final reconstructed 3D CT volume from the sparse 3D volume using a trained deep neural network
Implementation Method 2
generating the sparse 3D volume from the first X-ray image and the second X-ray image using a tomographic reconstruction algorithm
Data Source
Figure 1
Figure 2~4
Figure 5
AI summary
A method and apparatus for automated reconstruction of a 3D computed tomography (CT) volume from a small number of X-ray images is disclosed. A sparse 3D volume is generated from a small number of x-ray images using a tomographic reconstruction algorithm. A final reconstructed 3D CT volume is generated from the sparse 3D volume using a trained deep neural network. A 3D segmentation mask can also be generated from the sparse 3D volume using the trained deep neural network.