Systems and methods for deep learning-based image reconstruction

Patent metadata

Field	Value
Publication number	US-10679384-B2
Application number	US-201715720632-A
Country	US
Kind code	B2
Filing date	Sep 29, 2017
Priority date	Sep 29, 2017
Publication date	Jun 9, 2020
Grant date	Jun 9, 2020

Abstract

Methods and systems for deep learning based image reconstruction are disclosed herein. An example method includes receiving a set of imaging projections data, identifying a voxel to reconstruct, receiving a trained regression model, and reconstructing the voxel. The voxel is reconstructed by: projecting the voxel on each imaging projection in the set of imaging projections according to an acquisition geometry, extracting adjacent pixels around each projected voxel, feeding the regression model with the extracted adjacent pixel data to produce a reconstructed value of the voxel, and repeating the reconstruction for each voxel to be reconstructed to produce a reconstructed image.

First claim

What is claimed is:

1. A method comprising:
receiving a set of digital breast tomosynthesis imaging projections;
identifying a voxel to reconstruct;
receiving a trained regression model that maps a set of pixel values to a voxel value, the regression model trained by a regression model trainer based on at least one of acquired projection data or simulated projection data and deployed to map the set of pixel values to the voxel value, wherein the regression model trainer includes at least one of: a) a Digital Anthropomorphic Phantom (DAP) Modeler including an acquisition simulator, an algorithm creator, and a DAP database; b) a Computed Tomography (CT) Modeler including an acquisition simulator, an algorithm creator, and a CT database; or c) an Algorithm Modifier including an acquisition reconstructor and an algorithm database; and
reconstructing the voxel by:
projecting the voxel onto each digital breast tomosynthesis imaging projection in the set of digital breast tomosynthesis imaging projections according to an acquisition geometry associated with acquisition of the set of digital breast tomosynthesis imaging projections;
extracting adjacent pixels around each projected voxel;
feeding the regression model with data for the extracted adjacent pixels data to produce a reconstructed value for the voxel; and
repeating the reconstruction for each voxel to be reconstructed to produce a reconstructed image volume.

2. The method of claim 1 , further including training the regression model on at least one of:
a database including acquired projection data and a 2D mammogram acquired under a same compression, the regression model trained to output a 2D image approximately identical to the 2D mammogram when fed with the acquired projection data; or
a database including simulated projection data and a simulated 2D mammogram acquired under the same compression from a digital anthropomorphic phantom, the regression model trained to output a 2D image approximately identical to the simulated 2D mammogram when fed with the simulated projection data.

3. The method of claim 1 , further including training the regression model on at least one of:
a database including a digital anthropomorphic phantom and simulated projection data obtained from the digital anthropomorphic phantom for a given acquisition geometry, the regression model trained to output a volume approximately identical to the digital anthropomorphic phantom when fed with the simulated projection data;
a database including computed tomography (CT) reconstructed data and simulated projection data obtained from the CT reconstruction data, the regression model trained to output a volume approximately identical to the CT reconstructed data when fed with the simulated projection data; or
a database including acquired projection data and reconstructed data from these projection data with a given reconstruction algorithm, the regression model trained to output a volume approximately identical to the reconstructed data when fed with the acquired projection data.

4. The method of claim 1 , wherein the voxel is part of a slab to be reconstructed.

5. The method of claim 1 , further including producing a reconstructed image using the reconstructed values and displaying the reconstructed image on a user interface.

6. The method of claim 1 , further including receiving a volume, the voxel projected onto voxels from the volume according to an acquisition geometry.

7. A system comprising:
a voxel identifier to identify a voxel to be reconstructed;
an imaging projections data receiver to receive a set of digital breast tomosynthesis imaging projections; and
a voxel reconstructor including:
a voxel projector to project the voxel onto each digital breast tomosynthesis imaging projection in the set of digital breast tomosynthesis imaging projections according to an acquisition geometry associated with acquisition of the set of digital breast tomosynthesis imaging projections;
an adjacent pixel extractor to extract adjacent pixels around each projected voxel; and
a regression model feeder to feed a trained regression model with data for the extracted adjacent pixels to produce a reconstructed value of the voxel, the regression model trained by a regression model trainer based on at least one of acquired projection data or simulated projection data and deployed to map a set of pixel values to a voxel value, wherein the regression model trainer includes at least one of: a) a Digital Anthropomorphic Phantom (DAP) Modeler including an acquisition simulator, an algorithm creator, and a DAP database; b) a Computed Tomography (CT) Modeler including an acquisition simulator, an algorithm creator, and a CT database; or c) an Algorithm Modifier including an acquisition reconstructor and an algorithm database, the regression model feeder to repeat the reconstruction for each voxel to be reconstructed to produce a reconstructed image volume.

8. The system of claim 7 , wherein the regression model trainer includes:
a database including acquired projection data and a 2D mammogram acquired under the a same compression, the regression model trained to output a 2D image approximately identical to the 2D mammogram when fed with the projection data; or
a database including simulated projection data and a simulated 2D mammogram acquired under the same compression from a digital anthropomorphic phantom, the regression model trained to output a 2D image approximately identical to the simulated 2D mammogram when fed with the simulated projection data.

9. The system of claim 7 , further including a feedback generator to identify if a mistake has made on the reconstructed image volume and communicate to the regression model trainer to re-train the regression model.

10. The system of claim 7 , wherein the voxel is part of a slab to be reconstructed.

11. The system of claim 7 , further including a reconstructed value producer to produce a reconstructed value for each reconstructed pixel or voxel, the reconstructed values used to produce the reconstructed image volume.

12. The system of claim 7 , further including a user interface, the user interface to display the reconstructed image volume.

13. A non-transitory computer readable storage medium comprising instructions which, when executed, cause a machine to at least:
receive a set of digital breast tomosynthesis imaging projections;
identify a voxel to reconstruct;
receive a trained regression model that maps a set of pixel values to a voxel value, the regression model trained by a regression model trainer based on at least one of acquired projection data or simulated projection data and deployed to map the set of pixel values to the voxel value, wherein the regression model trainer includes at least one of: a) a Digital Anthropomorphic Phantom (DAP) Modeler including an acquisition simulator, an algorithm creator, and a DAP database; b) a Computed Tomography (CT) Modeler including an acquisition simulator, an algorithm creator, and a CT database; or c) an Algorithm Modifier including an acquisition reconstructor and an algorithm database; and
reconstruct the voxel by:
projecting the voxel onto each digital breast tomosynthesis imaging projection in the set of digital breast tomosynthesis imaging projections according to an acquisition geometry associated with acquisition of the set of digital breast tomosynthesis imaging projections;
extracting adjacent pixels around each projected voxel;
feeding the regression model with data for the extracted adjacent pixels to produce a reconstructed value for the voxel; and
repeating the reconstruction for each voxel to be reconstructed to pr

Assignees

• Gen Electric

Inventors

• Palma Giovanni John Jacques
• Iordache Razvan

Classifications

• G06T11/005 Primary

  Physics · mapped topic

• G06T11/006

  Physics · mapped topic

• A61B6/5205

  involving processing of raw data to produce diagnostic data · CPC title

• G06N3/04

  Architecture, e.g. interconnection topology · CPC title

• G06T7/33

  using feature-based methods · CPC title