Image reconstruction using machine learning regularizers
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing tomographic reconstruction methods using machine learning models introduce bias from training datasets, limiting their effectiveness and requiring users to either accept this bias or restrict their use, leading to suboptimal image reconstructions.
Innovation Solution
A machine learning model is trained using multiple learning datasets, including high and low-quality images, to minimize bias, and is integrated into an iterative reconstruction technique with deep learning neural networks as regularizers to refine image features and remove artifacts.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a machine learning model is used in tomographic reconstruction to improve image quality when data is insufficient, then reconstruction artifacts are reduced, but bias is introduced from the training dataset
Solution Approach 1:
The patent segments the training data into multiple diverse datasets representing different objects and conditions. By dividing the training process into multiple independent datasets rather than using a single homogeneous dataset, the system reduces the bias that would arise from any single dataset while maintaining the benefits of machine learning for artifact reduction.
Solution Approach 2:
The patent changes the parameter of training data diversity by using multiple datasets with different characteristics, object types, and acquisition conditions. This parameter change allows the machine learning model to learn generalizable features while minimizing dataset-specific bias, thereby improving both reconstruction quality and objectivity.
2Measurement precision
If the aggressiveness of machine learning is limited to avoid bias, then training data bias is reduced, but reconstruction improvement is compromised
Solution Approach 1:
The patent applies segmentation by using multiple specialized machine learning models, each trained on a specific dataset, rather than one aggressive model trained on a single dataset. This segmentation allows each model to be applied with appropriate aggressiveness for its specific domain while maintaining overall objectivity through diversity.
Solution Approach 2:
The patent creates a composite approach by combining multiple machine learning models trained on different datasets. This composite strategy integrates the strengths of various models while canceling out their individual biases, enabling more aggressive application of machine learning without compromising objectivity.
3Measurement precision
If multiple learning datasets are used to train a machine learning model, then bias is minimized, but training complexity increases
Solution Approach 1:
The patent segments the training process into independent stages, where each machine learning model is trained separately on its own dataset. This segmentation of the training process reduces the immediate complexity of handling multiple datasets simultaneously, while still achieving the benefit of reduced bias through diversity.
Solution Approach 2:
The patent creates universal machine learning models that can be applied across different reconstruction scenarios. By training models on diverse datasets that cover multiple object types and conditions, each model gains multi-functionality, reducing the need for separate specialized models and thereby reducing overall system complexity.
Data Source
Figure 1~2
Figure 3
Figure 4
AI summary
A system and method for reconstructing an image of a target object using an iterative reconstruction technique can include a machine learning model as a regularization filter (100). An image data set for a target object generated using an imaging modality can be received, and an image of the target object can be reconstructed using an iterative reconstruction technique that includes a machine learning model as a regularization filter (100) used in part to reconstruct the image of the target object. The machine learning model can be trained prior to receiving the image data using learning datasets that have image data associated with the target object, where the learning datasets providing objective data for training the machine learning model, and the machine learning model can be included in the iterative reconstruction technique to introduce the object features into the image of the target object being reconstructed.