Apparatus and method for medical image reconstruction using deep learning for computed tomography (CT) image noise and artifacts reduction

Patent metadata

Field	Value
Publication number	US-11517197-B2
Application number	US-201715727216-A
Country	US
Kind code	B2
Filing date	Oct 6, 2017
Priority date	Oct 6, 2017
Publication date	Dec 6, 2022
Grant date	Dec 6, 2022

Abstract

A method and apparatus is provided that uses a deep learning (DL) network to reduce noise and artifacts in reconstructed medical images, such as images generated using computed tomography, positron emission tomography, and magnetic resonance imaging. The DL network can operate either on pre-reconstruction data or on a reconstructed image. The DL network can be an artificial neural network or a convolutional neural network (e.g., using a three-channel volumetric kernel architecture). Different neural networks can be trained depending on the noise level, scanning protocol, or the anatomic, diagnostic or clinical objective of the reconstructed image (e.g., by partitioning the training data into noise-level range and training respective DL networks for each range). Further, the DL networks can be trained to mitigate artifacts, such as the cone-beam artifact.

First claim

The invention claimed is:

1. A medical image processing apparatus, comprising:
a memory storing a plurality of trained models corresponding to a plurality of anatomical structures; and
processing circuitry configured to
obtain data acquired by a computed tomography (CT) scan, the CT scan being performed according to a condition for a designated anatomical structure,
reconstruct a CT image based on the obtained data by an analytical reconstruction method,
select a trained model corresponding to the designated anatomical structure, from the plurality of trained models, and
apply the selected trained model to the reconstructed CT image to generate a processed CT image in which noise or an artifact is mitigated,

wherein
each of the plurality of trained models has been subjected to a training process with training pairs of a first CT image and a second CT image, the first CT image obtained by reconstructing scan data acquired through a CT scan on a specific anatomical structure, the second CT image obtained by reconstructing the scan data acquired through the CT scan on the specific anatomical structure, and
each of the plurality of trained models is obtained by updating a weighting coefficient of a machine-learning model based on an error between the second CT image and the processed CT image to which the machine-learning model has been applied.

2. A medical image processing apparatus, comprising:
processing circuitry configured to
receive an image of an object based on radiation data detected at a detector,
select, for reducing noise in the image, a neural network based on an anatomical structure corresponding to the image, from a plurality of neural networks, wherein the plurality of neural networks has been trained for noise reduction with images of corresponding anatomical structures, and the selected neural network has been trained with multiple images including the anatomical structure, and
apply the selected neural network to the image to generate a processed image as an output of the selected neural network,

wherein the processing circuitry is further configured to
train the neural network using training data and a function that represents a disagreement between pairs of data as an error value, the training data including pairs in which a pair includes defect-exhibiting data paired with corresponding defect-minimized data, and the neural network is trained by performing, for each of the pairs, the steps of
applying the neural network to defect-exhibiting data of a pair to generate network processed data,
calculating, using the function, the error value between the network processed data and the defect-minimized data of the pair,
updating, based on the calculated error value, weighting coefficients of the neural network, and
repeating the steps of applying, calculating, and updating using respective pairs of the training data until one or more stopping criteria are satisfied.

3. The apparatus according to claim 2 , wherein
the neural network is a convolutional neural network that includes a first layer of a plurality of layers that is a convolution layer having a kernel that represents weighting coefficients of connections from neuronal nodes of an input layer to the neuronal nodes of the first layer, the kernel having a predefined two-dimensional array size in a plane of slices of the input data and an array length of three in a direction orthogonal to the plane of the slices of the input data, and
the processing circuitry is further configured to map, using the kernel, the input data to the first layer by performing a weighted sum of pixels of a central slice and two adjacent slices of the input data to generate values of the neuronal nodes of the first layer corresponding to the central slice.

4. The apparatus according to claim 2 , wherein the processing circuitry is further configured to perform the step of training the neural network, for a case in which the defect-exhibiting data are reconstructed images exhibiting a cone-beam artifact that is generated when an image is reconstructed from radiation data acquired using a cone-beam scanning protocol, by
acquiring defect-exhibiting data of a pair of the training data by performing a computed tomography scan of an object using a large-angle scan protocol, and
acquiring defect-minimized data of the pair of the training data by performing a computed tomography scan of the object using a small-angle scan protocol, wherein
the large-angle scan protocol uses a beam angle selected to match a beam angle that is used during an acquisition of the acquired radiation data, and the small-angle scan protocol uses a beam angle selected to reduce the cone-beam artifact below a predefined artifact threshold.

5. The apparatus according to claim 2 , wherein the processing circuitry is further configured to perform the step of training the neural network, for a case in which the defect-exhibiting data span a plurality of noise-level ranges, by
generating defect-minimized data in which a noise level is minimized,
pairing the defect-minimized data with one or more defect-exhibiting data having greater noise levels than the defect-minimized data and corresponding to one or more of the plurality of noise-level ranges, and
training respective neural networks of a plurality of neural networks corresponding to the plurality of noise-level ranges by training each neural network of the plurality of neural networks to correspond to a respective noise-level range of the plurality of noise-level ranges, wherein the each neural network is trained using those pairs of the training data for having defect-exhibiting data with a noise level within the respective noise-level range corresponding to the each neural network.

6. The apparatus according to claim 2 , wherein the processing circuitry is further configured to perform the step of training the neural network, for a case in which the defect-exhibiting data are categorized into a plurality of categories according to one or more of a noise level of the radiation data or the image, and an anatomic structure being scanned or for which the image is to be reconstructed from the acquired radiation data, by
generating a plurality of pieces of defect-minimized data in which a defect is minimized, the defect being one or more of noise and an artifact, and each piece of the plurality of pieces of defect-minimized data representing pre-reconstruction data or a reconstructed image,
generating a plurality of pieces of defect-exhibiting data in which the defect is not minimized, respective pieces of the plurality of pieces of defect-exhibiting data corresponding to and being paired with pieces of the plurality of pieces of defect-minimized data, and respective pieces of the plurality of pieces of defect-exhibiting data being categorized into the plurality of categories and spanning the plurality of categories, and
training each of the plurality of neural networks to correspond with a respective category of the plurality of categories by associating each neural network with a corresponding category and each neural network is trained using those pairs of the training data in the corresponding category.

7. A medical image processing apparatus, comprising:
processing circuitry configured to
receive an image of an object based on radiation data detected at a detector,
select, for reducing noise in the image, a neural network based on an anatomical structure corresponding to the image, from a plurality of neural networks, wherein the plurality of neural networks has been trained for noise reduction with images of corresponding anatomical structures, and the selected neural network has been trained with multiple images including the anatomical structure, and
apply the selected neural ne

Assignees

• Toshiba Medical Sys Corp
• Canon Medical Systems Corp

Inventors

• Zhou Jian
• Yu Zhou
• Liu Yan

Classifications

• G06T12/30

  Image post-processing, e.g. metal artefact correction · CPC title

• G06T12/10 Primary

  Image preprocessing, e.g. calibration, positioning of sources or scatter correction · CPC title

• A61B5/0035 Primary

  adapted for acquisition of images from more than one imaging mode, e.g. combining MRI and optical tomography · CPC title

• A61B5/0073

  by tomography, i.e. reconstruction of 3D images from 2D projections (A61B5/0066 takes precedence) · CPC title

• G16H30/40

  for processing medical images, e.g. editing · CPC title