Neural Network Image Reconstruction Using Masked Sinogram Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional nuclear imaging systems face high memory and computational requirements, leading to inefficient image reconstruction and lower quality medical images due to approximations used to balance image quality with efficiency.
Innovation Solution
A computer-implemented method and system employing a neural network with a Radon inversion layer, applying masks to sinogram data and fully connected layers to generate image patches, which are then combined to produce high-quality medical images with reduced computational costs and memory requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional mathematical algorithms (analytic or iterative) are used for image reconstruction, then image quality can be maintained, but memory and computational requirements become excessively high
Solution Approach 1:
The patent replaces conventional mathematical algorithms (analytic or iterative reconstruction methods) with a deep learning-based neural network system. The neural network is trained offline to learn the mapping from raw sinogram data to reconstructed images, substituting the computational heavy lifting of traditional algorithms with a pre-trained model that can perform reconstruction with lower online computational requirements while maintaining or improving image quality.
Solution Approach 2:
The patent performs preliminary training of the neural network offline before actual image reconstruction. During this preliminary phase, the network learns optimal reconstruction patterns from training data. When actual reconstruction is needed, the pre-trained network can quickly process new sinogram data without requiring the intensive computational resources that conventional iterative algorithms would demand at reconstruction time.
2Productivity
If approximations are used in image formation processes to balance efficiency with image quality, then computational efficiency improves, but image quality deteriorates
Solution Approach 1:
The patent changes the fundamental parameters of the reconstruction process by transitioning from deterministic mathematical algorithms with built-in approximations to a data-driven neural network approach. The neural network learns optimal transformation parameters from training data, enabling it to achieve high image quality without relying on the approximations that limit conventional methods. The network can capture complex non-linear relationships in the data that traditional approximated algorithms cannot represent accurately.
Data Source
AI summary
Systems and methods for reconstructing medical images are disclosed. Measurement data, such as sinogram data, is received from an image scanning system. A plurality of masks are applied to corresponding portions of the measurement data to generate a plurality of masked measurement data portions. In some examples, the measurement data is encoded before the plurality of masks are applied. A neural network including a plurality of fully connected layers is applied to the plurality of masked measurement data portions to generate a plurality of image patches. The plurality of image patches are then combined to generate an initial image. In some examples, refinement and scaling operations are applied to the initial image and corresponding attenuation maps to generate a final image. In some examples, the final image is stored in a database. In some examples, the final image is displayed for diagnosis.