Method and apparatus for generating x-ray tomographic image data

Patent metadata

Field	Value
Publication number	US-10559101-B2
Application number	US-201816181685-A
Country	US
Kind code	B2
Filing date	Nov 6, 2018
Priority date	Nov 6, 2017
Publication date	Feb 11, 2020
Grant date	Feb 11, 2020

Abstract

Provided are a method and apparatus for interpolating X-ray tomographic image data by using a machine learning model. A method of interpolating an X-ray tomographic image or X-ray tomographic composite image data includes obtaining a trained model parameter via machine learning that uses a sub-sampled sinogram for learning as an input and uses a full-sampled sinogram for learning as a ground truth; radiating X-rays onto an object at a plurality of preset angular locations via an X-ray source, and obtaining a sparsely-sampled sinogram including X-ray projection data obtained via X-rays detected at the plurality of preset angular locations; applying the trained model parameter to the sparsely-sampled sinogram by using the machine learning model; and generating a densely-sampled sinogram by estimating X-ray projection data not obtained with respect to the object on the sparsely-sampled sinogram.

First claim

What is claimed is:

1. A method comprising:
radiating X-rays onto an object at a plurality of preset angular locations via an X-ray source, and obtaining a sparsely-sampled sinogram including X-ray projection data obtained via the X-rays that passed through the object;
applying a trained model parameter to the sparsely-sampled sinogram by using a machine learning model, to thereby generate trained image data, wherein the trained model parameter is obtained via the machine learning model that uses a sub-sampled sinogram for learning as an input and uses a full-sampled sinogram for learning as a ground truth;
estimating, from the sparsely-sampled sinogram, X-ray projection data with respect to the object that is not included in the sparsely-sampled sinogram; and
generating a densely-sampled sinogram using the trained image data and the estimated X-ray projection data.

2. The method of claim 1 , further comprising:
interpolating the sparsely-sampled sinogram via linear interpolation before applying the trained model parameter to the sparsely-sampled sinogram.

3. The method of claim 1 , wherein
the machine learning model includes a plurality of layers, and
the applying of the trained model parameter comprises:
cropping the sparsely-sampled sinogram into at least one image patch, and
applying the trained model parameter to the at least one image patch using the machine learning model including the plurality of layers.

4. The method of claim 3 , further comprising:
obtaining a residual sinogram; and
maintaining a value of trained X-ray projection data at the plurality of preset angular locations included in the residual sinogram to be equal to a value of X-ray projection data at angles corresponding to the plurality of preset angular locations included in the sparsely-sampled sinogram.

5. The method of claim 4 , wherein the residual sinogram is a sinogram indicating a difference between the densely-sampled sinogram and the sparsely-sampled sinogram.

6. The method of claim 5 , wherein
the maintaining comprises:
applying, to the residual sinogram, a base projection data preserving layer that maintains the value of the trained X-ray projection data at the plurality of preset angular locations included in the residual sinogram to be equal to the value of the X-ray projection data at the angles corresponding to the plurality of preset angular locations included in the sparsely-sampled sinogram, and
adding the sparsely-sampled sinogram to the residual sinogram to which the base projection data preserving layer has been applied, and

the base projection data preserving layer processes a pixel value of at least one piece of X-ray projection data at the plurality of preset angular locations to be 0.

7. The method of claim 6 , wherein
the maintaining comprises applying a base projection data maintaining layer that replaces a pixel value of at least one piece of X-ray projection data obtained at the plurality of angular locations with the value of the X-ray projection data obtained at the angles corresponding to the plurality of angular locations included in the sparsely-sampled sinogram initially provided as an input, and
the applying of the base projection data preserving layer is performed after the adding of the sparsely-sampled sinogram initially provided as an input to the residual sinogram.

8. The method of claim 1 , further comprising:
cropping the sparsely-sampled sinogram into one or more image patches each having a length equal to a sensor size of an X-ray detector in a first direction and a preset width in a second direction perpendicular to the first direction,
wherein the applying of the trained model parameter comprises applying the trained model parameter to the one or more image patches by using the machine learning model.

9. The method of claim 1 , further comprising:
equalizing a sum of pieces of X-ray projection data at a same angular location from among one or more pieces of X-ray projection data included in the densely-sampled sinogram to a sum of pieces of X-ray projection data at an angle corresponding to a same angular location included in the sparsely-sampled sinogram.

10. The method of claim 9 , wherein the equalizing comprises:
interpolating values of trained pieces of X-ray projection data, based on a sum of pieces of X-ray projection data at a first angular location included in the densely-sampled sinogram and a sum of pieces of X-ray projection data at a second angular location that is adjacent to the first angular location included in the densely-sampled sinogram.

11. The method of claim 1 , further comprising:
applying the trained model parameter to the densely-sampled sinogram by using the machine learning model, to thereby train the densely-sampled sinogram; and
applying the trained model parameter to the trained densely-sampled sinogram using the machine learning model.

12. An apparatus comprising:
an X-ray source configured to radiate X-rays to an object at a plurality of preset angular locations;
an X-ray detector configured to detect the X-rays radiated by the X-ray source and passed through the object; and
at least one processor configured to cause the following to be performed:
applying, using a machine learning model, a trained model parameter to a sparsely-sampled sinogram obtained via the X-rays detected by the X-ray detector, to thereby generate trained image data, wherein the trained model parameter is obtained via the machine learning model that uses a sub-sampled sinogram for learning as an input and uses a full-sampled sinogram for learning as a ground truth,
estimating, from the sparsely-sampled sinogram, X-ray projection data with respect to the object that is not included in the sparsely-sampled sinogram, and
generating a densely-sampled sinogram using the trained image data and the estimated X-ray projection data.

13. The apparatus of claim 12 , wherein the at least one processor is further configured to cause the following to be performed:
interpolating the sparsely-sampled sinogram via linear interpolation before applying the trained model parameter to the sparsely-sampled sinogram.

14. The apparatus of claim 12 , wherein
the at least one processor is further configured to cause the following to be performed:
cropping the sparsely-sampled sinogram into at least one image patch, and

the applying the trained model parameter comprises applying the trained model parameter to the at least one image patch using the machine learning model.

15. The apparatus of claim 14 , wherein the at least one processor is further configured to cause the following to be performed:
obtaining a residual sinogram; and
maintaining a value of trained X-ray projection data at the plurality of preset angular locations included in the residual sinogram to be equal to a value of X-ray projection data at angles corresponding to the plurality of preset angular locations included in the sparsely-sampled sinogram.

16. The apparatus of claim 15 , wherein the residual sinogram is a sinogram indicating a difference between the densely-sampled sinogram and the sparsely-sampled sinogram.

17. The apparatus of claim 16 , wherein
the maintaining comprises:
applying, to the residual sinogram, a base projection data preserving layer that maintains the value of the trained X-ray projection data at the plurality of preset angular locations included in the residual sinogram to be equal to the value of the X-ray projection data at the angles corresponding to the plurality of preset angular locations included in the sparsely-sampl

Assignees

• Samsung Electronics Co Ltd
• Korea Advanced Inst Sci & Tech
• Korean Advanced Institute Of Science And Tech

Inventors

• Lee Jong Ha
• Lee Ho Yeon
• Cho Seung Ryong

Classifications

• G06T12/10 Primary

  Image preprocessing, e.g. calibration, positioning of sources or scatter correction · CPC title

• G06T12/30 Primary

  Image post-processing, e.g. metal artefact correction · CPC title

• G06T2207/10081

  Computed x-ray tomography [CT] · CPC title

• G06T2207/20081

  Training; Learning · CPC title

• G06T2211/436

  Limited angle · CPC title