Aggregation based on deep set learning in magnetic resonance reconstruction

Patent metadata

Field	Value
Publication number	US-12039636-B2
Application number	US-202117473206-A
Country	US
Kind code	B2
Filing date	Sep 13, 2021
Priority date	Sep 13, 2021
Publication date	Jul 16, 2024
Grant date	Jul 16, 2024

Abstract

For reconstruction in medical imaging using a scan protocol with repetition, a machine learning model is trained for reconstruction of an image for each repetition. Rather than using a loss for that repetition in training, the loss based on an aggregation of images reconstructed from multiple repetitions is used to train the machine learning model. This loss for reconstruction of one repetition based on aggregation of reconstructions for multiple repetitions is based on deep set-based deep learning. The resulting machine-learned model may better reconstruct an image from a given repetition and/or a combined image from multiple repetitions than a model learned from a loss per repetition.

First claim

What is claimed is:

1. A method of machine training for magnetic resonance (MR) reconstruction in medical imaging, the method comprising:
acquiring training data for an MR protocol using multiple repetitions;
machine training a neural network for the MR reconstruction using the training data, the neural network training for the MR reconstruction to output an image for each one of the multiple repetitions, a first loss used in the training being based on an aggregation of the images from the multiple repetitions, the first loss comprising a measure of difference between the aggregation of the images and a ground truth image such that the neural network learns for each of the multiple repetitions based on the aggregation across the multiple repetitions; and
storing the neural network as machine trained.

2. The method of claim 1 wherein machine training comprises training the neural network to output the image for each one of the multiple repetitions independently of the output for other ones of the multiple repetitions.

3. The method of claim 1 wherein machine training comprises training the neural network to output the image for at least one of the multiple repetitions in dependence on information from the neural network used to output an image for another one of the multiple repetitions.

4. The method of claim 3 wherein the dependence comprises a global pooling across instances of the neural network for respective ones of the multiple repetitions.

5. The method of claim 1 wherein acquiring the training data comprises acquiring with the MR protocol being a diffusion-weighted, turbo-spin-echo, contrast with different echo times, or contrast with different flip angles protocol.

6. The method of claim 1 wherein machine training comprises machine training the neural network for repetition reconstruction preserving invariance by permutation with respect to the multiple repetitions.

7. The method of claim 1 wherein machine training comprises machine training where the aggregation is an average, a geometric product, or a geometric mean.

8. The method of claim 1 wherein the neural network comprises an input layer configured to accept information for a single one of the multiple repetitions for the output of the image for that single one of the multiple repetitions.

9. The method of claim 1 wherein the neural network comprises an input layer having separate inputs of different directions or contrasts, and wherein the training is to output the image from the different directions or contrasts for each of the multiple repetitions.

10. The method of claim 9 wherein machine training comprises applying a joint loss of the first loss based on the aggregation and a repetition loss for each of the different directions or contrasts of each of the multiple repetitions.

11. The method of claim 1 wherein machine training comprises applying a joint loss of the first loss based on the aggregation and a repetition loss each image for each of the multiple repetitions.

12. A method for reconstruction of a medical image in a medical imaging system, the method comprising:
scanning, by the medical imaging system, a patient, the scanning resulting in measurements over a series of scans of an imaging protocol;
reconstructing, by an image processor applying a machine-learned model, a scan image for each of the scans of the series, the machine-learned model having been trained for use for each scan of the imaging protocol based on a loss function from a combination of training images from different scans for the imaging protocol, the loss function comprising a difference between the combination of the training images and a ground truth image;
combining the scan images into the medical image; and
displaying the medical image.

13. The method of claim 12 wherein scanning comprises magnetic resonance scanning with the imaging protocol comprising a diffusion-weighted imaging protocol or a turbo-spin-echo imaging protocol, and wherein the combination used to train is an average.

14. The method of claim 12 wherein reconstructing the scan image for one of the scans of the series occurs prior to the scanning for another of the scans of the series.

15. A system for reconstruction in medical imaging, the system comprising:
a medical scanner configured to repetitively scan a region of a patient pursuant to a protocol, the scan providing scan data in repetitions of the protocol;
an image processor configured to reconstruct, for each of the repetitions, a representation of the region, the image processor configured to reconstruct by application of a machine-learned model having been trained for the reconstruction for each of the repetitions based on a loss function of a difference between an aggregate of outputs from the repetitions of the protocol and a ground truth, the image processor further configured to combine the representations from the repetitions; and
a display configured to display an image of the region from the combined representations.

16. The system of claim 15 wherein the medical scanner comprises a magnetic resonance scanner, wherein protocol comprises a diffusion-weighted, turbo-spin-echo, or contrast protocol, and wherein the combination of representations comprises a sum, a geometric mean, or a geometric product.

17. The system of claim 15 wherein the machine-learned model comprises a neural network and wherein the neural network is repetitively used for the repetitions as a function in a deep set.

18. The system of claim 15 wherein the loss function comprised a joint loss of first and second losses, the first loss being between the aggregate and the ground truth and the second loss being between one of the representations for one of the repetitions and another ground truth.

Assignees

• Siemens Healthineers Ag

Inventors

• Arberet Simon
• Mailhe Boris
• Benkert Thomas
• Nickel Marcel Dominik
• Mostapha Mahmoud
• Nadar Mariappan S

Classifications

• G06T12/00 Primary

  Tomographic reconstruction from projections · CPC title

• G16H30/20 Primary

  for handling medical images, e.g. DICOM, HL7 or PACS · CPC title

• G06N3/08

  Learning methods · CPC title

• G06N3/0455

  Auto-encoder networks; Encoder-decoder networks · CPC title

• G06N3/084

  Backpropagation, e.g. using gradient descent · CPC title