CNN-Based PET Reconstruction with MRI Guidance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current PET imaging technologies face limitations due to high Poisson noise and spatial resolution constraints, requiring significant reconstruction time and proprietary raw data, which hinders widespread clinical adoption of Bayesian reconstruction methods.
Innovation Solution
A convolutional neural network (CNN) system that combines PET and MRI information to generate high-resolution PET images, using a maximum likelihood estimation procedure and concatenating MLE and MPRAGE images, allowing for fast computation and standard DICOM image input, thereby approximating model-based iterative reconstruction with anatomical guidance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If Bayesian reconstruction methods with anatomical prior are used, then spatial resolution and signal-to-noise ratio are improved, but reconstruction time increases significantly
Solution Approach 1:
The patent uses a pre-trained CNN model that has learned the transformation from standard PET/MRI images to high-resolution PET images during training. During inference, the model quickly copies this learned transformation to generate high-resolution images without performing the computationally intensive iterative Bayesian reconstruction process, thus achieving high spatial resolution with significantly reduced reconstruction time
Solution Approach 2:
The CNN model is pre-trained offline using pairs of standard PET images and corresponding high-resolution PET images generated through model-based iterative reconstruction with anatomical priors. This preliminary training phase allows the model to learn the complex transformation patterns, so that during actual clinical use, high-resolution images can be generated quickly without repeating the lengthy iterative reconstruction process
2Reliability
If model-based iterative reconstruction with anatomical prior is used, then image quality is improved, but computational complexity and data requirements increase
Solution Approach 1:
The patent replaces the complex mechanical iterative reconstruction process with a neural network-based system. The CNN model, once trained, performs the image enhancement through simple forward propagation operations, substituting the computationally intensive iterative optimization process with a much faster neural network inference process that achieves similar or better image quality
Solution Approach 2:
The patent uses standard clinical PET and MRI images as inputs, which are readily available and require no special processing. The CNN model processes these standard images to generate high-resolution PET images, eliminating the need for proprietary raw PET data and complex reconstruction pipelines, thus simplifying the system while maintaining image quality
Data Source
AI summary
An exemplary system, method and computer-accessible medium for generating an image(s) of a portion(s) of a patient(s) can be provided, which can include, for example, receiving first information associated with a combination of positron emission tomography (PET) information and magnetic resonance imaging (MRI) information, generating second information by applying a convolutional neural network(s) (CNN) to the first information, and generating the image(s) based on the second information. The PET information can be fluorodeoxyglucose PET information. The CNN(s) can include a plurality of convolution layers and a plurality of parametric activation functions. The parametric activation functions can include, e.g., a plurality of parametric rectified linear units. Each of the convolution layers can include, e.g., a plurality of filter kernels. The PET information can be reconstructed using a maximum likelihood estimation (MLE) procedure to generate a MLE image.