Truncated CT Image Prediction Using Deep Learning Calibration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Computed tomography (CT) systems face challenges in reconstructing accurate images when detected objects exceed the scanning field, leading to data truncation and suboptimal image quality in truncated portions.
Innovation Solution
A method involving preprocessing projection data by padding truncated portions with information from untruncated portions and using a trained learning network to calibrate initial images from polar to rectangular coordinates and back, improving image prediction accuracy for truncated regions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional mathematical models (e.g., water model) are used to predict truncated portion images, then the imaging process can be completed, but the image quality of the truncated portion varies and performance is not optimal
Solution Approach 1:
The patent replaces conventional mathematical models (water model, etc.) with a deep learning network based approach. The neural network is trained on paired data of complete images and their corresponding truncated versions, learning to predict the truncated portion from available projection data. This substitution of mathematical algorithms with learned models achieves superior image quality and stability in reconstructing truncated portions.
2Productivity
If data truncation occurs when detected object exceeds scanning field, then the scanning process can be completed faster, but complete projection data cannot be acquired leading to reconstruction errors
Solution Approach 1:
The patent introduces a deep learning network as an intermediary to recover missing projection data information. The network takes available projection data (including truncated portions padded with boundary values) and learns to predict the missing information by comparing with training data patterns, effectively mediating between incomplete measurements and complete image reconstruction.
Solution Approach 2:
The patent performs preliminary padding of truncated projection data with boundary values before reconstruction, and pre-trains the neural network on complete imaging data. This preliminary preparation enables the system to handle truncated data effectively during actual scanning, allowing faster scanning without complete data acquisition.
3Ease of manufacture
If padding is applied to truncated projection data, then the data can be processed for reconstruction, but the initial reconstructed image of truncated portion contains errors
Solution Approach 1:
The patent implements a feedback mechanism where the initially reconstructed image (with padding errors) is fed into a refinement neural network. This network compares the reconstructed image with the input projection data and iteratively corrects errors, using feedback loops to progressively improve image accuracy and eliminate artifacts from padding.
Data Source
AI summary
The present application provides a method and device for obtaining a predicted image of a truncated portion, an imaging method and system, and a non-transitory computer-readable storage medium. The method for obtaining a predicted image of a truncated portion comprises preprocessing projection data to obtain, by reconstruction, an initial image of the truncated portion; and calibrating the initial image based on a trained learning network to obtain the predicted image of the truncated portion.