Medical Image Reconstruction via Non-Cartesian to Cartesian Data Transformation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional medical image reconstruction using deep neural networks is limited by insufficient accuracy and unsatisfactory picture quality due to the challenges of processing non-Cartesian data.
Innovation Solution
A medical image processing apparatus that acquires non-Cartesian sampled data, derives product-sums with multiple coefficient sets to generate equispaced sampled data, and reconstructs medical images using a combination of processors for improved image generation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a deep neural network is used to learn all calculations for reconstruction of non-Cartesian data, then the reconstruction process can be performed, but the accuracy of reconstruction is insufficient and the picture quality is not satisfactory
Solution Approach 1:
The patent segments the reconstruction process into two distinct parts: (1) a deep neural network that learns only the transformation from non-Cartesian to Cartesian data, and (2) a conventional Fourier transform that performs the actual image reconstruction. This segmentation allows the neural network to focus solely on data transformation while the Fourier transform handles reconstruction, thereby improving both accuracy and picture quality without requiring the network to learn the entire reconstruction pipeline.
2Productivity
If non-Cartesian data is processed directly through deep learning for reconstruction, then the processing can be completed, but noise and artifacts are not adequately mitigated
Solution Approach 1:
The patent introduces an intermediary transformation step where non-Cartesian data is converted to Cartesian data through a trained deep neural network before undergoing conventional Fourier transform reconstruction. This intermediary step acts as a mediator that prepares the data in a format suitable for conventional reconstruction algorithms, effectively mitigating noise and artifacts while maintaining processing efficiency.
3Adaptability or versatility
If all reconstruction calculations are learned by a deep neural network, then the system can handle non-Cartesian data, but the device complexity and computational burden increase significantly
Solution Approach 1:
The patent divides the complex reconstruction task into two manageable components: a relatively simple deep neural network that only performs non-Cartesian to Cartesian data transformation, and a conventional Fourier transform that handles the reconstruction. This segmentation dramatically reduces the computational burden and device complexity compared to using a deep neural network to learn the entire reconstruction process, while still maintaining the ability to handle non-Cartesian data.
Data Source
AI summary
According to one embodiment, a medical image processing apparatus includes an acquirer, a first processor and a second processor. The acquirer is configured to acquire nonequispaced sampled data from a test object. The first processor is configured to derive product-sums of the nonequispaced sampled data acquired by the acquirer and a plurality of coefficient sets and generate equispaced sampled data including a plurality of elements with which the product-sums derived for the coefficient sets are associated as element values. The second processor is configured to generate a medical image in which at least part of the test object has been imaged through reconstruction basis on the equispaced sampled data generated by the first processor.