Multichannel deep learning reconstruction of multiple repetitions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing medical imaging techniques, such as MR, CT, PET, and SPECT, often reconstruct images with multiple repetitions and directions separately, missing opportunities to leverage common information across these repetitions, leading to inefficient and suboptimal results due to the need for separate neural networks for each configuration.
Innovation Solution
A single multichannel network architecture is adapted to accommodate varying numbers of repetitions and directions using an adapter that duplicates, dispatches, or employs encoder-decoder networks to ensure all data is utilized, allowing for a unified reconstruction process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If separate neural networks are configured for each configuration of number of directions and repetitions, then the network can be trained on dedicated data for that configuration, but the system becomes inefficient and requires multiple networks with smaller training datasets
Solution Approach 1:
A single multichannel neural network is designed to handle multiple configurations of directions and repetitions universally. The network accepts variable numbers of input channels corresponding to different repetitions and directions, eliminating the need for separate networks for each configuration while maintaining reconstruction quality through shared learned features.
Solution Approach 2:
The input data is segmented into multiple channels representing different repetitions and directions. Each channel is processed independently through the network, allowing the system to handle variable configurations by simply adjusting the number of input channels rather than requiring separate networks.
2Device complexity
If a single network is trained to accommodate a fixed number of repetitions in the channel dimension, then the network structure is simplified, but it cannot handle varying numbers of repetitions and directions
Solution Approach 1:
The network architecture incorporates dynamic adaptability where the number of input channels can vary based on the actual number of repetitions and directions in the input data. This dynamic structure allows the same network to handle different configurations without requiring structural changes or retraining, resolving the contradiction between simplicity and versatility.
Solution Approach 2:
The network is designed to accept variable parameters for the number of repetitions and directions. By changing the input channel count parameter rather than the network architecture itself, the system maintains structural simplicity while achieving high adaptability to different scanning configurations.
3Measurement precision
If multiple separate networks are used for different configurations, then each network can be optimized for its specific configuration, but the training dataset for each network becomes much smaller
Solution Approach 1:
Multiple configurations of data that would otherwise require separate networks are merged into a single multichannel network. This consolidation allows all available data across different configurations to be used for training the same network, significantly increasing the effective training dataset size and improving reconstruction accuracy through more robust feature learning.
Data Source
AI summary
Systems and methods for reconstruction for a medical imaging system. An adapter is used to adapt scan data so that different quantities of repetitions or directions may be used to train and implement a single multichannel backbone network.