Apparatus and method for medical image reconstruction using deep learning to improve image quality in positron emission tomography (PET)

Patent metadata

Field	Value
Publication number	US-11801029-B2
Application number	US-202117554019-A
Country	US
Kind code	B2
Filing date	Dec 17, 2021
Priority date	May 31, 2018
Publication date	Oct 31, 2023
Grant date	Oct 31, 2023

Abstract

A deep learning (DL) convolution neural network (CNN) reduces noise in positron emission tomography (PET) images, and is trained using a range of noise levels for the low-quality images having high noise in the training dataset to produce uniform high-quality images having low noise, independently of the noise level of the input image. The DL-CNN network can be implemented by slicing a three-dimensional (3D) PET image into 2D slices along transaxial, coronal, and sagittal planes, using three separate 2D CNN networks for each respective plane, and averaging the outputs from these three separate 2D CNN networks. Feature-oriented training can be implemented by segmenting each training image into lesion and background regions, and, in the loss function, applying greater weights to voxels in the lesion region. Other medical images (e.g. MRI and CT) can be used to enhance resolution of the PET images and provide partial volume corrections.

First claim

The invention claimed is:

1. An apparatus, comprising:
processing circuitry configured to
acquire a reconstructed positron emission tomography (PET) image, the reconstructed PET image being reconstructed from emission data representing coincidence counts of respective pairs of gamma rays arising from electron-positron annihilation events, the coincidence counts being detected at a plurality of detector elements,
acquire a neural network including weighting coefficients of connections between neuronal nodes of respective layers of a plurality of layers between an input layer and an output layer of the neural network, the neural network having been trained using a training dataset that, for a given noise-minimized reconstructed data, includes two or more noise-exhibiting reconstructed data having greater noise levels than in the noise-minimized reconstructed data and having different noise levels, the two or more noise-exhibiting reconstructed data being reconstructed using subsets of a full PET dataset used to reconstruct the given noise-minimized reconstructed data, the training of the neural network including optimizing a loss function representing respective differences between the given noise-minimized reconstructed data and each of the two or more noise-exhibiting reconstructed data, and
apply the reconstructed PET image to the acquired neural network to generate a noise reduced image.

2. The apparatus according to claim 1 , wherein the processing circuitry is further configured to
reconstruct the PET image using a same reconstruction method that was used to reconstruct the two or more noise-exhibiting reconstructed data and the given noise-minimized reconstructed data of the training dataset used to train the acquired neural network, and
acquire the neural network wherein the neural network has been trained using the two or more noise-exhibiting reconstructed data that are reconstructed using subsets of a full PET dataset used to reconstruct the given noise-minimized reconstructed data, each of the subsets including a different predefined amount or percentage of the full PET dataset.

3. The apparatus according to claim 1 , wherein the processing circuitry is further configured to apply the reconstructed PET image to the acquired neural network without adjustments based on statistical properties of the reconstructed PET image.

4. The apparatus according to claim 1 , wherein the processing circuitry is further configured to acquire the neural network wherein the neural network has been trained to provide a same image quality for the noise reduced image independently of a noise level of the reconstructed PET image by the acquired neural network being trained to optimize the loss function simultaneously for the two or more noise-exhibiting reconstructed data having different noise levels that spans a predefined range of noise levels corresponding to images reconstructed based on PET scans using the plurality of detector elements.

5. The apparatus according to claim 1 , wherein the processing circuitry is further configured to train the neural network in advance of the emission data being detected at the plurality of detector elements by
obtaining the training dataset for training a neural network, the training dataset including a plurality of noise-minimized PET images respectively paired with two or more corresponding noise-exhibiting images of a plurality of noise-exhibiting PET images having various noise levels that are greater than a noise level of a corresponding noise-minimized image, wherein, for each of the plurality of noise-minimized PET images, the noise-minimized PET image is reconstructed using a respective full PET emission dataset, and the two or more corresponding noise-exhibiting images are reconstructed from subsets of the respective full PET emission dataset that are selected to provide a range of noise levels among the two or more corresponding noise-exhibiting images, and
training the neural network by iteratively adjusting tunable parameters of the neural network to minimize a loss function representing a difference between a respective noise-minimized image and an output when a noise-exhibiting image of the training dataset is applied to the neural network, the tunable parameters being adjusted to simultaneously minimize the loss function for noise-exhibiting images having noise levels throughout the range of noise levels among the two or more noise-exhibiting images of the training dataset.

6. The apparatus according to claim 5 , wherein the processing circuitry is further configured to train the neural network using the training dataset that includes another medical image corresponding to the respective noise-minimized image, the another medical image being one of a magnetic resonance image and an X-ray computed tomography image.

7. The apparatus according to claim 1 , wherein the subsets of the full PET dataset used to reconstruct the given noise-minimized reconstructed data are subsets of different sizes.

8. An apparatus, comprising:
processing circuitry configured to
acquire a reconstructed positron emission tomography (PET) image, the reconstructed PET image being reconstructed from emission data representing coincidence counts of respective pairs of gamma rays arising from electron-positron annihilation events, the coincidence counts being detected at a plurality of detector elements,
acquire a neural network including weighting coefficients of connections between neuronal nodes of respective layers of a plurality of layers between an input layer and an output layer of the neural network, the neural network having been trained using a training dataset that, for a given noise-minimized reconstructed data, includes two or more noise-exhibiting reconstructed data having greater noise levels than in the noise-minimized reconstructed data, the two or more noise-exhibiting reconstructed data being reconstructed using subsets of a full PET emission dataset used to reconstruct the given noise-minimized reconstructed data, the training of the neural network including optimizing a loss function representing respective differences between the given noise-minimized reconstructed data and each of the two or more noise-exhibiting reconstructed data, and
apply the reconstructed PET image to the acquired neural network to generate a noise reduced image,

wherein the processing circuitry is further configured to acquire the neural network wherein the neural network has been trained using a plurality of noise-minimized reconstructed data, each of the plurality of noise-minimized reconstructed data being paired with each of two or more noise-exhibiting reconstructed data that are respectively generated by reconstructing PET images using subsets of the full PET dataset that is used to reconstruct the each of plurality of noise-minimized reconstructed data,
the subsets are selected to produce the two or more noise-exhibiting reconstructed data that span a predefined range of statistical properties corresponding to images reconstructed based on PET scans performed using the plurality of detector elements, and,
when the reconstructed image to which the acquired neural network is applied is within the predefined range of statistical properties, an image quality of the noise reduced image is less affected by statistical properties of the reconstructed PET image than if the noise reduced image were generated using a neural network trained using noise-exhibiting reconstructed data that all had a same value for the statistical properties.

9. A method, comprising:
acquiring a reconstructed positron emission tomography (PET) image, the reconstructed PET image reconstructed from emission data representing coincidence counts of respecti

Assignees

• Canon Medical Systems Corp

Inventors

• Chan Chung
• Zhou Jian
• Asma Evren

Classifications

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/09

  Supervised learning · CPC title

• A61B6/5258 Primary

  involving detection or reduction of artifacts or noise · CPC title

• A61B6/032

  Transmission computed tomography [CT] · CPC title

• A61B6/037

  Emission tomography · CPC title