Noise to noise ensemble learning for pet and CT image and data denoising

Patent metadata

Field	Value
Publication number	US-11961209-B2
Application number	US-202017013104-A
Country	US
Kind code	B2
Filing date	Sep 4, 2020
Priority date	Oct 20, 2019
Publication date	Apr 16, 2024
Grant date	Apr 16, 2024

Abstract

A system and method for training a neural network to denoise images. One noise realization is paired to an ensemble of training-ready noise realizations, and fed into a neural network for training. Training datasets can also be retrospectively generated based on existing patient studies to increase the number of training datasets.

First claim

The invention claimed is:

1. A method of generating an image denoising system, comprising:
obtaining imaging data from a set of N studies;
dividing each of the N studies into at least K training-ready noise realizations representing K subsets of the imaging data from each of the N studies; and
training a machine learning-based system, on a study-by study-basis for each study of the set of N studies by a noise-to-noise-ensemble (N2NEN) training method, based on (1) a first noise realization of the at least K training-ready noise realizations only as training data for each study, and (2) a remaining K−1 training-ready noise realizations of the at least K training-ready noise realizations other than the first noise realization for each study only as label data to produce a trained machine learning-based system,
wherein the first noise realization and the remaining K−1 training-ready noise realizations contain noise.

2. The method as claimed in claim 1 , further comprising repeating for at least one study of the set of N studies, training the machine learning-based system, based on (1) a second noise realization of the at least K training-ready noise realizations other than the first noise realization as training data for the at least one study and (2) a remaining K−1 training-ready noise realizations of the at least K training-ready noise realizations other than the second noise realization for the at least one study as label data to produce the trained machine learning-based system.

3. The method as claimed in claim 1 , further comprising repeating on a study-by-study basis for the set of N studies, training the machine learning-based system, based on (1) a second noise realization of the at least K training-ready noise realizations other than the first noise realization as training data and (2) a remaining K−1 training-ready noise realizations of the at least K training-ready noise realizations other than the second noise realization as label data to produce the trained machine learning-based system.

4. The method as claimed in claim 1 , wherein the at least K training-ready noise realizations for each study of the set of N studies are training-ready noise realizations of a first count level, the method further comprising:
obtaining at least K training-ready noise realizations at a second count level for each study of the set of N studies; and
training the machine learning-based system, on a study-by study-basis for each study of the set of N studies, based on (1) a first noise realization of the at least K training-ready noise realizations at the second count level as training data for the each study and (2) a remaining K−1 training-ready noise realizations of the at least K training-ready noise realizations at the second count level other than the first noise realization at the second count level for each study as label data to produce the trained machine learning-based system.

5. The method as claimed in claim 4 , further comprising repeating on a study-by-study basis for the set of N studies, training the machine learning-based system, based on (1) a second noise realization of the at least K training-ready noise realizations other than the first noise realization, but separately at the first and second count levels, as training data, and (2) a remaining K−1 training-ready noise realizations of the at least K training-ready noise realizations other than the second noise realization, but separately at the first and second count levels, as label data to produce the trained machine learning-based system.

6. The method as claimed in claim 1 , wherein the machine learning-based system comprises a neural network and the trained machine learning-based system comprises a trained neural network.

7. The method as claimed in claim 6 , wherein the neural network comprises a deep neural network.

8. The method of claim 1 , wherein the at least K training-ready noise realizations are at least K substantially independent noise realizations.

9. A trained machine learning-based system produced according to the method of claim 1 .

10. The method of claim 1 , wherein N is an integer, N≥2, K is an integer, and K is greater than or equal to three so that the machine-learning system is trained on K−1 pairs of training/label data.

11. A system, comprising:
processing circuitry configured to
obtain imaging data from a set of N studies;
divide each of the N studies into at least K training-ready noise realizations representing K subsets of the imaging data from each of the N studies; and
train a machine learning-based system, on a study-by study-basis for each study of the set of N studies by a noise-to-noise-ensemble (N2NEN) training method, based on (1) a first noise realization of the at least K training-ready noise realizations only as training data for the each study, and (2) a remaining K−1 training-ready noise realizations of the at least K training-ready noise realizations other than the first noise realization for each study only as label data to produce a trained machine learning-based system,

wherein the first noise realization and the remaining K−1 training-ready noise realizations contain noise.

12. The system as claimed in claim 11 , wherein the processing circuitry is further configured to repeat for at least one study of the set of N studies, training the machine learning-based system, based on (1) a second noise realization of the at least K training-ready noise realizations other than the first noise realization as training data for the at least one study, and (2) a remaining K−1 training-ready noise realizations of the at least K training-ready noise realizations other than the second noise realization for the at least one study as label data to produce the trained machine learning-based system.

13. The system as claimed in claim 11 , wherein the processing circuitry is further configured to repeat on a study-by-study basis for the set of N studies, training the machine learning-based system, based on (1) a second noise realization of the at least K training-ready noise realizations other than the first noise realization as training data, and (2) a remaining K−1 training-ready noise realizations of the at least K training-ready noise realizations other than the second noise realization as label data to produce the trained machine learning-based system.

14. The system as claimed in claim 11 , wherein the at least K training-ready noise realizations for each study of the set of N studies are training-ready noise realizations of a first count level, and the processing circuitry is further configured to:
obtain at least K training-ready noise realizations at a second count level for each study of the set of N studies; and
train the machine learning-based system, on a study-by study-basis for each study of the set of N studies, based on (1) a first noise realization of the at least K training-ready noise realizations at the second count level as training data for the each study, and (2) a remaining K−1 training-ready noise realizations of the at least K training-ready noise realizations at the second count level other than the first noise realization at the second count level for each study as label data to produce the trained machine learning-based system.

15. The system as claimed in claim 14 , wherein the processing circuitry is further configured to repeat on a study-by-study basis for the set of N studies, training the machine learning-based system, based on (1) a second noise realization of the at least K training-ready noise realizations other than the first noise realization, but separately at the first and second cou

Assignees

• Canon Medical Systems Corp

Inventors

• Chan Chung
• Zhou Jian
• Asma Evren

Classifications

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/0895

  Weakly supervised learning, e.g. semi-supervised or self-supervised learning · CPC title

• G06N3/0455

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

• G06T5/002 Primary

  Physics · mapped topic

• G06N3/08

  Learning methods · CPC title