Deep Learning Quasi-Projection Operator for Tomographic Image Reconstruction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image reconstruction methods face challenges in producing accurate images from incomplete or noisy measurement data due to insufficient data sufficiency, leading to artifacts and limited clinical usefulness.
Innovation Solution
A deep-learning-assisted iterative image reconstruction method using a quasi-projection operator, which incorporates a priori information to regularize intermediate images, reducing noise and artifacts by transforming them into more accurate representations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If standard analytic reconstruction algorithms (e.g., filtered back-projection) are used, then the reconstruction process is simple and fast, but the reconstructed images contain conspicuous artifacts when projection data are insufficient
Solution Approach 1:
A deep learning model is introduced as an intermediary component between the analytic reconstruction algorithm and the final image output. The model takes low-quality reconstructed images containing artifacts as input and transforms them into high-quality images by learning the mapping from artifact-prone reconstructions to artifact-free images, effectively mediating between fast but inaccurate reconstruction and accurate but slow iterative methods
Solution Approach 2:
The solution combines two different reconstruction approaches into a composite system: the speed advantage of analytic methods (filtered back-projection) is preserved for initial reconstruction, while the accuracy advantage of iterative methods is incorporated through the deep learning model that learns from training data. This composite approach achieves both fast processing and high image quality
2Manufacturing precision
If iterative image reconstruction methods are used to recover high-quality images from imperfect data, then image accuracy is improved, but the reconstruction process becomes computationally intensive and time-consuming
Solution Approach 1:
A deep learning model is pre-trained offline using pairs of images generated by iterative reconstruction methods. This preliminary training phase captures the complex mapping from incomplete projection data to high-quality images. During actual reconstruction, the pre-trained model rapidly processes new data without requiring iterative computation, thus achieving high accuracy without the computational burden during operation
Solution Approach 2:
The deep learning model learns to copy the high-quality image reconstruction capability from training examples without replicating the computationally intensive iterative process. The model captures the essential features and patterns of accurate reconstructions during training, then reproduces this capability efficiently during inference, avoiding repeated iterative computations
3Manufacturing precision
If sufficient projection data are collected to ensure accurate reconstruction, then image quality is improved, but the imaging time or radiation exposure increases
Solution Approach 1:
The deep learning model serves as an intermediary that compensates for insufficient projection data. It learns the statistical relationships and anatomical constraints from training data, then applies this knowledge to reconstruct accurate images from undersampled projections, effectively mediating between limited data and high-quality reconstruction without requiring increased imaging time or radiation dose
Data Source
AI summary
Systems and methods reconstruct an image of a subject or object by performing an iterative reconstruction method to produce a plurality of intermediate images and the image of the subject, and transforming at least one selected intermediate image from the plurality of intermediate images using a quasi-projection operator. The quasi-projection operator includes a deep-learning model configured to map the at least one selected intermediate image to at least one regularized intermediate image. In addition, systems and methods for training the deep-learning model using a training data set that includes a plurality of training images and a plurality of corresponding training object images is disclosed.