US 10,832,451 B2Granted

Machine learning in iterative image reconstruction

AssigneesSiemens Healthcare GmbH, Centre Hospitalier Universitaire Vaudois

InventorsRobin Demesmaeker, Tobias Kober, Davide Piccini, Jérôme Yerly

Life of this patent

Estimating the term…

Loading patent…

Abstract

In order to reduce the time and effort required to generate high-quality image reconstructions, a machine-trained neural network may assign a quality score to an image at each iteration of a reconstruction. The neural network may confirm that the iterative reconstruction process increases image quality as each iteration converges to the solution of an optimization problem. The image quality score generated by the neural network may drive the reconstruction toward better image quality by contributing to a regularization term of a cost function minimized by the optimization problem. The neural network may allow for multiple reconstruction of image data to be performed rapidly and for the highest image quality reconstruction to be identified. Additionally, the neural network may provide exit criteria of the iterative reconstruction or may contribute to the optimization problem.

Claims16

Dependent claims

A method for reconstructing image data, the method comprising:

acquiring medical image data;
generating, by a processor, a first medical image, the generating of the first medical image comprising:
performing, by the processor, a first reconstruction iteration on at least a portion of the medical image data;
generating, by the processor, a first quality score of the first medical image, the generating of the first quality score comprising:
applying, by the processor, a machine-trained neural network to the first medical image, the neural network being trained on a plurality of images with a plurality of associated image quality scores;
outputting, by the processor, the first medical image or generating, by the processor, a second medical image based on the first quality score, wherein the generating of the second medical image comprises:
producing a second medical image, the producing of the second medical image comprising:
performing, by the processor, a second reconstruction iteration on the first medical image;
and
generating a second quality score of the second medical image, the generating of the second quality score comprising:
applying, by the processor, the machine-trained neural network to the second medical image;
comparing, by the processor, the first quality score and the second quality score;
determining, by the processor, a change in the quality score based on the comparison of the first quality score and the second quality score;
comparing, by the processor, the change in the quality score to a threshold;
and
performing, by the processor, a third reconstruction iteration when the change in the quality score exceeds the threshold.

13 more elements

A system for reconstruction of image data, the system comprising:

an imaging device configured to acquire medical image data;
an image reconstruction engine configured to reconstruct a medical image based on the medical image data;
a quality scorer configured to apply a trained neutral network to the medical image to generate a quality score of the medical image;
and
a display configured to output the medical image,
wherein the reconstruction of the medical image comprises determination of a value of an optimization problem having a cost function with a first weight for a regularization term and a second weight for a data fidelity term,
wherein the image reconstruction engine is further configured to perform a first reconstruction and a second reconstruction using different values for the first weight and second weight for the first reconstruction and the second reconstruction,
wherein the quality scorer is further configured to generate a first quality score for the first reconstruction and a second quality score for the second reconstruction, and
wherein the display is furthered configured to display a medical image produced by a reconstruction with a highest image quality of the first reconstruction with the first image quality and the second reconstruction with the second image quality.

6 more elements

Dependencies are inferred from the claim text.

Facts

Application: 16/038,855Filed 18 Jul 2018
Published: US 2020/0027251 A123 Jan 2020
Granted: 10 Nov 2020
Term ends: Not estimatedThe source records don’t support an estimate.

Classification19 codes

G06T 12/20
Tomographic reconstruction from projections: inverse problem, i.e. transformations from projection space into object space
G06N 3/0464
Architecture, e.g. interconnection topology: convolutional networks [CNN, ConvNet]
G06N 3/08
Neural networks: learning methods
G06N 3/09
Learning methods: supervised learning