Transformer-CNN Reconstruction of Sparse CT Rotational Topograms
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Rotational projection positional imaging (RPLR) in computed tomography faces limitations due to sparse sampling patterns causing data loss, which affects image quality and scan speed, particularly in high stepping scenarios.
Innovation Solution
A method combining a transformer and CNN architecture to rearrange and process rotational topogram data, predicting missing information by considering all available context, followed by upsampling and detail recovery using a guided upsampling network to enhance image quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If high step (pitch) is used in RPLR, then scan speed is improved, but image quality deteriorates due to sparse sampling pattern causing data loss
Solution Approach 1:
The patent introduces a deep learning model (transformer and CNN) as an intermediary to process and complete the sparse projection data. The transformer encodes the input data with positional information, and the CNN decodes it to generate high-quality topogram images, effectively bridging the gap between sparse sampling and complete image reconstruction.
Solution Approach 2:
The patent changes the processing parameters by applying deep learning transformations to the projection data. The model learns optimal parameter transformations during training to reconstruct complete topogram images from sparse inputs, enabling high-speed scanning without quality loss.
2Loss of time
If sparse sampling pattern is used in RPLR, then data acquisition time is reduced, but data completeness deteriorates causing data loss
Solution Approach 1:
The patent replaces traditional mechanical interpolation or reconstruction methods with a deep learning-based computational approach. The transformer-CNN model substitutes conventional signal processing mechanics with learned patterns, enabling accurate reconstruction from sparse data without requiring complete physical sampling.
Solution Approach 2:
The patent performs preliminary training of the deep learning model using complete topogram data before actual scanning. This preliminary action allows the model to learn the relationship between complete and sparse data representations, so that during actual scanning, it can accurately reconstruct images even with minimal input data.
3Device complexity
If conventional interpolation processing is used for sparse data, then computational complexity is reduced, but image fidelity deteriorates
Solution Approach 1:
The patent transforms the problem from simple spatial interpolation to a multi-dimensional learning problem. By adding positional encoding dimensions and using multi-layer neural network transformations, the model processes data in expanded dimensional space, capturing complex relationships that simple 2D interpolation cannot achieve.
Data Source
Figure 1~2
Figure 3~4
Figure 5~6
AI summary
The present invention discloses a method for processing rotational topogram data, a computer program, a computer-readable storage medium and a computing device. The method for processing rotational topogram data for computed tomography comprises: rearranging the rotational topogram data as one-dimensional data; using a transformer to process the one-dimensional data; rearranging the transformer-processed one-dimensional data as two-dimensional data, so as to obtain a recovered topogram. The present invention uses a transformer to predict missing information in the rotational topogram data, and the reconstructed topogram thereof has high image quality, reducing the limitations of the reconstructed topogram in the prior art (such as interpolation processing in the case of high stepping). To reduce the computing power requirements, the rotational topogram data can be compressed, and the lost information can then be recovered by means of a CNN.