CT Kernel Synthesis for High-Resolution Low-Noise Imaging
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current CT imaging systems require multiple image series reconstructed with different kernels for diverse diagnostic tasks, leading to increased burden on technologists, archival systems, and reconstruction time, while archiving projection data is infeasible due to large file sizes and proprietary formats, and existing methods introduce artifacts or require excessive computation.
Innovation Solution
A deep-learning based method using a convolutional neural network with a task-based loss function synthesizes multiple image series into a single image series, combining high spatial resolution and low noise levels, optimized for various clinical tasks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple image series are reconstructed with different kernels for diverse diagnostic tasks, then diagnostic accuracy is improved, but the burden on technologists, archival systems, and reconstruction time increases
Solution Approach 1:
The patent combines multiple kernel images (sharp and smooth) into a single synthesized image that integrates the high spatial resolution of sharp kernels with the low noise levels of smooth kernels. This merging approach eliminates the need to manage multiple separate image series while maintaining diagnostic accuracy across different clinical tasks.
Solution Approach 2:
The synthesized single image series serves multiple diagnostic purposes that previously required different kernel-specific images. The image can be used for both high-contrast bone imaging and low-contrast soft tissue evaluation, making a single image universally applicable to diverse diagnostic tasks.
2Adaptability or versatility
If multiple image series are reconstructed with different kernels, then comprehensive diagnostic coverage is improved, but reconstruction time increases
Solution Approach 1:
The system performs kernel synthesis as a preliminary processing step that converts multiple kernel images into a single synthesized image early in the workflow. This preliminary action eliminates the need for subsequent manual selection and comparison of multiple images, saving time during diagnostic interpretation.
3Adaptability or versatility
If projection data are archived for later reconstruction, then flexibility for future reconstructions is improved, but storage requirements and data transfer difficulty increase
Solution Approach 1:
The patent extracts the essential diagnostic information from multiple kernel images and consolidates it into a single synthesized image series. This extraction approach retains the ability to obtain comprehensive diagnostic information without requiring storage of the original large-volume projection data or multiple separate reconstructed series.
4Manufacturing precision
If a sharp kernel is used for high spatial resolution, then edge definition is improved, but image noise increases
Solution Approach 1:
The synthesis algorithm acts as an intermediary that processes both sharp and smooth kernel images to produce a final image with reduced noise. The algorithm mediates between the high-resolution sharp kernel image and the low-noise smooth kernel image to create a synthesized image that achieves both goals simultaneously.
5Object-generated harmful factors
If a smooth kernel is used for low noise levels, then noise reduction is improved, but spatial resolution deteriorates
Solution Approach 1:
The synthesized image combines the beneficial properties of both sharp and smooth kernels, creating a composite image that exhibits both low noise levels and high spatial resolution. This composite approach integrates the noise-reduction characteristics of smooth kernels with the edge-definition capabilities of sharp kernels.
Data Source
Figure 1A~1B
Figure 2A
Figure 2B
AI summary
Systems and methods are provided for synthesizing information from multiple image series of different kernels into a single image series using deep-learning based methods trained using a task-based loss function that includes a sharp loss term and a smooth loss term that parameterize training. For multi-kernel synthesis, a single set of images with desired high spatial resolution and low image noise can be synthesized from multiple image series of different kernels. The synthesized kernel is sufficient for a wide variety of clinical tasks, even in circumstances that would otherwise require many separate image sets. Kernel conversion may be configured to generate images with arbitrary reconstruction kernels from a single baseline kernel. This would reduce the burden on the CT scanner and the archival system, and greatly simplify the clinical workflow.