Tomographic reconstruction based on deep learning

• US 10,475,214 B2
• Filed: 04/05/2017
• Issued: 11/12/2019
• Est. Priority Date: 04/05/2017
• Status: Active Grant

• Current Assignee: General Electric Company
• Original Assignee: General Electric Company
• Inventors: Fu, Lin, Ramani, Sathish, De Man, Bruno Kristiaan Bernard, Rui, Xue, Ahn, Sangtae
• Primary Examiner(s): Bernardi, Brenda C
• Application Number:
  US15/480,172
• Publication Number:
  US 20180293762A1
• Time in Patent Office: 951 Days
• Field of Search: 706 15- 44
• CPC Class Codes: G06T 11/003 Reconstruction from projections, e.g. tomography
• CPC Class Codes: G06T 11/006 Inverse problem, transformation from projection-space into object-space, e.g. transform methods, back-projection, algebraic methods
• CPC Class Codes: G06T 2210/41 Medical
• CPC Class Codes: G06T 2211/416 Exact reconstruction
• CPC Class Codes: G06T 2211/441 AI-based methods, deep learning or artificial neural networks

Abstract

The present approach relates to the use of machine learning and deep learning systems suitable for solving large-scale, space-variant tomographic reconstruction and/or correction problems. In certain embodiments, a tomographic transform of measured data obtained from a tomography scanner is used as an input to a neural network. In accordance with certain aspects of the present approach, the tomographic transform operation(s) is performed separate from or outside the neural network such that the result of the tomographic transform operation is instead provided as an input to the neural network. In addition, in certain embodiments, one or more layers of the neural network may be provided as wavelet filter banks.

Claims

1. A method, comprising:

obtaining measured data from a tomography scanner;

calculating one or more tomographic transforms of the measured data, wherein the one or more tomographic transforms comprise at least one of a backprojection, a weighted backprojection, a reprojection, a plurality of diagonal elements of a Fisher information matrix, a variance image, a noise correlation image, or a polynomial of the Fisher information matrix;

providing inputs to a trained neural network, wherein the inputs comprise the one or more tomographic transforms and an input reconstructed image, and the input reconstructed image comprises a filtered backprojection image with a special filter kernel, wherein the filter coefficients are selected that preserve the original sinogram after a reprojection; and

obtaining one or more outputs from the trained neural network based on the inputs.

2. The method of claim 1, further comprising:

training a neural network using a plurality of training cases.

3. The method of claim 2, wherein the plurality of training cases comprises one or more computer generated random patterns comprising Gaussian blobs, random noise patterns, or random shapes.

4. The method of claim 1, wherein the tomography scanner comprises one of an X-ray computed tomography (CT) scanner, a positron emission tomography (PET) scanner, a single photon emission computed tomography (SPECT) scanner, a magnetic resonance imaging (MM) scanner, or a C-arm scanner.

5. The method of claim 1, wherein the one or more outputs comprise an output reconstructed image.

6. The method of claim 1, wherein the inputs further comprise at least one of an input variance image or an input noise correlation image in addition to the input reconstructed image.

7. The method of claim 1, wherein the one or more outputs comprise an output variance image.

8. The method of claim 1, wherein the one or more outputs comprise one or more of a partial volume error corrected image, a partial volume error corrected standardized uptake value, or a partial volume error correction factor.

9. The method of claim 1, wherein the one or more outputs comprise one or more of a denoised image or a noise pattern.

10. The method in claim 1, wherein the neural network is space-variant and based on wavelet filter banks, wavelets, wavelet frames, curvelets, or other sparsifying transforms.

11. A method comprising:

obtaining measured data from a tomography scanner;

providing inputs to a trained neural network comprising one or more of the measured data or one or more tomographic transforms of the measured data, wherein the neural network comprises at least one layer based on wavelets, wavelet frames, curvelets, or other sparsifying transforms, wherein the inputs further comprise a reconstructed image in addition to the one or more tomographic transforms or the one or more of the measure data, wherein the reconstructed image comprises at least one filtered backprojection image with a special filter kernel, wherein the filter coefficients are selected so as to preserve the original sinogram after a reprojection; and

obtaining one or more outputs from the trained neural network based on the inputs.

12. The method of claim 11, wherein the at least one layer comprises a wavelet filter bank.

13. The method of claim 11, wherein the one or more tomographic transforms comprise at least one of a backprojection, a weighted backprojection, a reprojection, a variance image, a noise correlation image, a plurality of diagonal elements of a Fisher information matrix, or a polynomial of the Fisher information matrix.

14. The method of claim 11, wherein the neural network is space-variant.

15. An image processing system comprising:

a processing component configured to execute one or more stored processor-executable routines; and

a memory storing the one or more executable-routines, wherein the one or more executable routines, when executed by the processing component, cause acts to be performed comprising;

acquiring or accessing a set of scan data, wherein the set of scan data is initially represented by a set of original measurements;

calculating one or more tomographic transforms of the set of scan data;

providing the one or more tomographic transforms and an input reconstructed image as inputs to a trained neural network, wherein the input reconstructed image comprises a filtered backprojection image with a special filter kernel, wherein the filter coefficients are selected that preserve the original sinogram after a reprojection, and wherein the trained neural network comprises at least one layer based on a wavelet filter bank, wavelets, wavelet frames, curvelets, or other sparsifying transforms; and

obtaining one or more outputs from the trained neural network based on the inputs.

16. The image processing system of claim 15 wherein the set of scan data comprises computed tomography (CT) scan data, C-arm scan data, positron emission tomography (PET) scan data, single photon emission computed tomography (SPECT) scan data, or magnetic resonance imaging (MRI) scan data.

17. The image processing system of claim 15, wherein the one or more tomographic transforms comprise at least one of a backprojection, a weighted backprojection, a reprojection, a variance image, a noise correlation image, a plurality of diagonal elements of a Fisher information matrix, or a polynomial of the Fisher information matrix.

18. The image processing system of claim 15, wherein the one or more outputs comprise one or more of an output reconstructed image, a variance image, a partial volume error corrected image, a partial volume error corrected standardized uptake value, a partial volume error correction factor, or one or more of a denoised image or a noise pattern.