CT Image Reconstruction Using Multiscale CNNs for Incomplete Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing CT image reconstruction methods face limitations in reducing radiation dosage and achieving high image quality, particularly in scenarios with incomplete projection data due to detector under-sampling, sparse-angle scanning, limited-angle scanning, or linear trajectory scanning.
Innovation Solution
A method utilizing a pooling layer-free convolutional neural network to process projection data on multiple scales, performing back-projection operations, and fusing CT images across scales to enhance image quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional analytical reconstruction algorithms (filtered back-projection, FDK) or iterative reconstruction methods (ART, MAP) are used, then the reconstruction process is straightforward and computationally manageable, but the image quality approaches the limit and cannot be further improved to meet diverse imaging demands
Solution Approach 1:
The patent replaces traditional mechanical/mathematical reconstruction algorithms (filtered back-projection, iterative methods) with a deep learning-based convolutional neural network system. The CNN automatically learns optimal reconstruction features from training data, substituting the manual algorithm design process with an automated neural network approach that achieves superior image quality while handling complex scanning trajectories.
Solution Approach 2:
The patent transforms the reconstruction problem by changing the parameter space from traditional algorithmic parameters (filter types, iteration numbers) to neural network parameters (weights, biases, architecture configurations). By training the CNN on diverse scanning data, the system adapts to different scanning trajectories and reconstruction requirements through parameter optimization rather than algorithmic modification.
2Object-affected harmful factors
If radiation dosage is reduced to meet safety requirements, then patient safety improves, but the projection data becomes incomplete leading to degraded image quality
Solution Approach 1:
The patent converts the harmful effect of incomplete projection data (caused by reduced radiation dosage) into a beneficial training scenario for the neural network. The CNN is trained specifically on incomplete or undersampled projection data, learning to recognize and reconstruct meaningful image features even when data is missing. This transforms the previously harmful data deficiency into a useful training condition that enables the network to handle low-dose scenarios effectively.
Solution Approach 2:
The patent performs preliminary training of the convolutional neural network using complete or high-quality projection data before actual low-dose reconstruction. During this pre-training phase, the network learns optimal feature extraction and reconstruction patterns. When actual low-dose scanning occurs, the already-trained network can immediately process incomplete data effectively, having prepared in advance for the data quality challenges.
3Productivity
If detector under-sampling or sparse-angle scanning is used to reduce scanning time and radiation, then scanning efficiency improves, but the projection data becomes incomplete requiring more complex reconstruction
Solution Approach 1:
The patent replaces traditional information-completion methods (iterative algorithms attempting to fill missing data) with a neural network-based information-inference system. The CNN learns to directly infer missing projection information from available undersampled data by recognizing patterns and correlations in the measured projections, substituting complex iterative completion algorithms with a more efficient learned inference process.
Data Source
Figure 1
Figure 2~3
Figure 4
AI summary
A method and device for reconstructing a CT image and a storage medium are disclosed. CT scanning is performed on an object to be inspected to obtain projection data on a first scale. Projection data on a plurality of other scales is generated from the projection data on the first scale. Projection data on each scale is processed on the corresponding scale by using a first convolutional neural network to obtain processed projection data, and a back-projection operation is performed on the processed projection data to obtain a CT image on the corresponding scale. CT images on the plurality of scales are fused to obtain a reconstructed image of the object to be inspected. With the solutions according to the embodiments of the present disclosure, a CT image with a higher quality can be reconstructed.