US 2023/0067596 A1Published application

Projection based deep learning with frequency splitting for computed tomography

AssigneeCanon Medical Systems Corporation

InventorsQiulin TANG, Ruoqiao ZHANG, Jian ZHOU, Zhou YU

Life of this patent

Estimating the term…

Loading patent…

Abstract

Data acquired from a scan of an object can be decomposed into frequency components. The frequency components can be input into a trained model to obtain processed frequency components. These processed frequency components can be composed and used to generate a final image. The trained model can be trained, independently or dependently, using frequency components covering the same frequencies as the to-be-processed frequency components. In addition, organ specific processing can be enabled by training the trained model using image and/or projection datasets of the specific organ.

Claims20

Dependent claims

A medical imaging system comprising:

processing circuitry configured to
obtain one or more frequency components of a real projection dataset acquired from scanning an object to be examined, the one or more frequency components of the real projection dataset each covering a predetermined frequency band;
apply a trained model for improving image quality to the one or more frequency components of the real projection dataset to obtain one or more processed frequency components;
and
compose the one or more processed frequency components to obtain a set of processed projection data, wherein the trained model is generated by
obtaining one or more frequency components of a lower quality projection dataset, the one or more frequency components of the lower quality projection dataset each covering a same predetermined frequency band as a frequency component from the one or more frequency components of the real projection dataset,
obtaining one or more frequency components of a higher quality projection dataset, the one or more frequency components of the higher quality projection dataset each covering a same predetermined frequency band as a frequency component from the one or more frequency components of the real projection dataset, and
generating the trained model by training one or more neural networks using the one or more frequency components of the lower quality projection dataset as input learning data and the one or more frequency components of the higher quality projection dataset as target learning data.

5 more elements

A processing method comprising:

obtaining one or more frequency components of a real projection dataset acquired from scanning an object to be examined, the one or more frequency components of the real projection dataset each covering a predetermined frequency band;
applying a trained model for improving image quality to the one or more frequency components of the real projection dataset to obtain one or more processed frequency components;
and
composing the one or more processed frequency components to obtain a set of processed projection data, wherein the trained model is generated by
obtaining one or more frequency components of a lower quality projection dataset, the one or more frequency components of the lower quality projection dataset each covering a same predetermined frequency band as a frequency component from the one or more frequency components of the real projection dataset;
obtaining one or more frequency components of a higher quality projection dataset, the one or more frequency components of the higher quality projection dataset each covering a same predetermined frequency band as a frequency component from the one or more frequency components of the real projection dataset;
and
generating the trained model by training one or more neural networks using the one or more frequency components of the lower quality projection dataset as input learning data and the one or more frequency components of the higher quality projection dataset as target learning data.

5 more elements

Dependencies are inferred from the claim text.