Encoder-Decoder Network for Fast PET Image Reconstruction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current biomedical image reconstruction techniques, such as PET imaging, are time-consuming and require significant computational resources, often taking hours to produce images that may suffer from artifacts due to data/model mismatches and noise, posing challenges in clinical applications.
Innovation Solution
A deep convolutional encoder-decoder network is employed to directly reconstruct PET images from raw projection data, utilizing simulated data to learn geometric, statistical, and regularization models, reducing reconstruction time and improving image quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If conventional reconstruction methods are used, then image quality can be maintained, but reconstruction time becomes excessively long (several hours)
Solution Approach 1:
The patent replaces conventional iterative mechanical reconstruction algorithms with a deep neural network model that performs reconstruction in a single forward pass. The encoder-decoder architecture with convolutional layers substitutes the traditional mechanical iterative optimization process, achieving 17x speedup while maintaining image quality.
Solution Approach 2:
The neural network model is pre-trained on simulated data to learn the reconstruction mapping before actual use. This preliminary training phase allows the model to perform rapid reconstruction during clinical operation without requiring iterative computation at reconstruction time.
2Use of energy by moving object
If conventional reconstruction methods are used, then comprehensive image processing can be performed, but computational resources are excessively consumed
Solution Approach 1:
The patent replaces computationally intensive iterative reconstruction algorithms with a pre-trained deep neural network that performs reconstruction through efficient forward propagation. This substitution dramatically reduces GPU memory usage and computational power requirements while maintaining diagnostic image quality.
Solution Approach 2:
The patent uses simulated data to create a trained neural network model that copies the reconstruction capability without requiring access to the actual raw projection data during inference. The model learns from simulated training data and applies this knowledge to reconstruct clinical images efficiently.
3Manufacturing precision
If conventional reconstruction methods are used, then detailed processing can be performed, but artifacts appear due to data/model mismatches and noise
Solution Approach 1:
The patent replaces deterministic iterative reconstruction algorithms that are sensitive to data-model mismatches with a data-driven neural network. The model learns robust reconstruction patterns from simulated data that accounts for various noise conditions and system imperfections, reducing artifacts in clinical images.
Solution Approach 2:
The patent changes the reconstruction approach from parameter-based iterative optimization to a learned mapping function. The neural network learns optimal reconstruction parameters and patterns during training on simulated data, enabling it to handle noise and mismatches more effectively than conventional methods.
Data Source
AI summary
The present disclosure is directed to systems and methods for reconstructing biomedical images. A projection preparer may identify a projection dataset derived from a tomographic biomedical imaging scan. An encoder-decoder model may reconstruct reconstructing tomographic biomedical images from projection data. The encoder-decoder model may include an encoder. The encoder may include a first series of transform layers to generate first feature maps using the projection dataset. The encoder-decoder may include a decoder. The decoder may include a second series of transform layers to generate second features maps using the first feature maps from the encoder. The system may include a reconstruction engine executable on the one or more processors. The reconstruction engine may apply the encoder-decoder model to the projection dataset to generate a reconstructed tomographic biomedical image based on the second feature maps generated by the decoder of the encoder-decoder model.