CT Image Reconstruction via Hybrid Analytical-Deep Learning Network
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current X-ray CT imaging reconstruction methods require large amounts of supervised data for training convolutional neural networks, limiting their ability to reduce radiation dosage effectively and improve image quality, especially in cases of incomplete projection data.
Innovation Solution
An image processing method using a convolutional neural network comprising a projection domain network, an analytical reconstruction network layer, and an image domain network, where the network is trained using an a priori model cost function and consistency cost function, allowing for unsupervised training and adjustment of parameters to enhance image quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If supervised training with large amounts of marked data is used to train convolutional neural networks for CT reconstruction, then the neural network can be trained effectively, but the requirement for large amounts of marked data increases the complexity and cost of the training process
Solution Approach 1:
The system performs self-training by automatically generating training data through simulation without requiring manual marking. The neural network is trained using simulated projection data and corresponding images generated by the CT scanning system itself, eliminating the need for external marked datasets.
Solution Approach 2:
A simulation module acts as an intermediary between the CT scanning system and the neural network training process. This module generates synthetic training data by simulating the CT scanning process, providing the neural network with training samples without requiring actual marked medical images.
2Productivity
If traditional analytical reconstruction algorithms are used, then the reconstruction process is straightforward and fast, but the image quality is insufficient especially in cases of incomplete projection data or reduced radiation dosage
Solution Approach 1:
The system merges traditional analytical reconstruction algorithms with deep learning techniques in a hybrid architecture. The analytical reconstruction provides a preliminary result that is then refined by the neural network, combining the speed of traditional methods with the quality enhancement of AI.
Solution Approach 2:
The neural network learns optimal reconstruction parameters and patterns from simulated data, enabling it to adjust and optimize the reconstruction process dynamically. This allows the system to maintain high image quality even when projection data is incomplete or acquired with reduced radiation dosage.
3Object-affected harmful factors
If the radiation dosage is reduced to meet safety requirements, then the safety of patients and operators is improved, but the quality of the reconstructed image deteriorates due to insufficient projection data
Solution Approach 1:
The neural network is trained to recognize and compensate for the characteristics of low-dose projection data through self-generated training examples. By learning from simulated low-dose scenarios, the network can effectively reconstruct high-quality images from actual low-dose measurements without requiring manual intervention or additional marked data.
4Reliability
If supervised training methods are used with real marked data, then the neural network can learn from actual medical images, but the availability of such marked data is limited and the marking process is time-consuming
Solution Approach 1:
The system generates its own training data automatically through simulation, eliminating the time-consuming manual marking process. The simulation module creates synthetic projection data and corresponding ground truth images, providing unlimited training samples without requiring expert annotation time.
Solution Approach 2:
The training data is prepared in advance through simulation before the actual CT scanning and reconstruction process. This preliminary generation of training data allows the neural network to be pre-trained offline, so that during actual operation, the reconstruction can proceed quickly without requiring real-time data marking or preparation.
Data Source
Figure 1~2
Figure 3~4
Figure 5~6
AI summary
An image processing method, an image processing device and a storage medium are disclosed. The image processing method includes processing input projection domain data using a convolutional neural network. The neural network includes a projection domain network, an analytical reconstruction network layer and an image domain network. The projection domain network may process the projection domain data to recover possible missing data, thereby improving quality of the projection. The image domain network processes a reconstructed image to obtain an estimated image. The image domain network is adjusted by using an a priori model cost function based on the estimated image, and back propagation of a gradient is performed through the analytical reconstruction network layer to adjust parameters of a convolutional kernel of the projection domain network. A projection operation is performed on the estimated image by using a system projection matrix of the CT scanning system to obtain a projection result.