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-11786205-B2
Application number	US-202117554032-A
Country	US
Kind code	B2
Filing date	Dec 17, 2021
Priority date	May 31, 2018
Publication date	Oct 17, 2023
Grant date	Oct 17, 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 produceuniform 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 image,
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 low-noise image, includes a corresponding weight map that assigns greater weight values to voxels of the given low-noise image that are identified as being within a region of interest relative to weight values assigned to voxels of the given low-noise image that are identified as being background, the training of the neural network including optimizing a loss function that applies the weight map to a measure of differences between the given low-noise image and results of applying the neural network to a corresponding high-noise image, the high-noise image having a noise level that is greater than a noise level of the corresponding low-noise image, and
apply the reconstructed image to the acquired neural network to generate a denoised reconstructed image,

wherein the processing circuitry is further configured to acquire the neural network including the weight map, wherein the weight map was generated by segmenting one of the given low-noise image and the corresponding high-noise image to generate a region-of-interest mask that identifies the region of interest from the background, and assigning, based on the region-of-interest mask, weight values to voxel positions within the given low-noise image and the corresponding high-noise image.

2. The apparatus according to claim 1 , wherein the processing circuitry is further configured to
calculate the loss function for a given voxel by multiplying the weight value at the given voxel by the measure of the difference at the given voxel between the given low-noise image and the corresponding high-noise image.

3. The apparatus according to claim 1 , wherein the processing circuitry is further configured to acquire the neural network, wherein the neural network was trained such that the assigning of the weight values to the voxel positions includes applying a smoothing filter to the region-of-interest mask to cause the weight values to, as a function of position, smoothly transition from a maximum value in the region of interest to a minimum value in the background.

4. The apparatus according to claim 1 , wherein the processing circuitry is further configured to acquire the neural network, wherein the neural network was trained such that the segmenting of the one of the given low-noise image and the corresponding high-noise image further includes identifying a segmented region as either the region of interest or the background based on a received user input.

5. The apparatus according to claim 1 , wherein the processing circuitry is further configured to train the neural network in advance of emission data being detected at the plurality of detector elements by
obtaining the training dataset for training the neural network, the training dataset including a plurality of low-noise images that are each respectively paired with a corresponding high-noise image of a plurality of high-noise images,
determining the weight map that assigns greater weight values to the voxels within the region of interest than to the voxels in the background, and
training the neural network by iteratively adjusting tunable parameters of the neural network to minimize the loss function representing the differences between a respective low-noise image and a corresponding high-noise image of the training dataset, the high-noise image having a noise level that is greater than a noise level of the corresponding low-noise image, the loss function being weighted according to the weight map.

6. The apparatus according to claim 5 , wherein the processing circuitry is further configured to
train the neural network by calculating the loss function for each given voxel by multiplying the weight value at the given voxel by the measure of the difference at the given voxel between the given low-noise image and the corresponding high-noise image.

7. The apparatus according to claim 6 , wherein
the assigning the weight values to the voxel positions by the processing circuitry includes applying a smoothing filter to the region-of-interest mask to cause the weight values to smoothly transition from a maximum value in the region of interest to a minimum value in the background, and
the segmenting the one of the given low-noise image and the corresponding high-noise image by the processing circuitry further includes identifying a segmented region as either the region of interest or the background based on a received user input.

8. The apparatus according to claim 1 , wherein the processing circuitry is further configured to obtain 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, and the processing circuitry is further configured to reconstruct, from the emission data, a positron emission tomography (PET) image as the reconstructed image.

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