High-contrast minimum variance imaging method based on deep learning

Patent metadata

Field	Value
Publication number	US-12159378-B2
Application number	US-201917626503-A
Country	US
Kind code	B2
Filing date	Oct 25, 2019
Priority date	Jul 12, 2019
Publication date	Dec 3, 2024
Grant date	Dec 3, 2024

Abstract

Disclosed is a high-contrast minimum variance imaging method based on deep learning. For the problem of the poor performance of a traditional minimum variance imaging method in terms of ultrasonic image contrast, a deep neural network is applied in order to suppress an off-axis scattering signal in channel data received by an ultrasonic transducer, and after the deep neural network is combined with a minimum variance beamforming method, an ultrasonic image with a higher contrast can be obtained while the resolution performance of the minimum variance imaging method is maintained. In the present method, compared with the traditional minimum variance imaging method, after an apodization weight is calculated, channel data is first processed by using a deep neural network, and weighted stacking of the channel data is then carried out, so that the pixel value of a target imaging point is obtained, thereby forming a complete ultrasonic image.

First claim

What is claimed is:

1. A high-contrast minimum variance imaging method based on deep learning, comprising the following steps:
S1. scanning a target object for ultrasound imaging, generating a channel data composed of echo signals received by reception channels of an ultrasound transducer, and performing a respective delay operation with regard to different points for imaging to obtain a delay channel data;
S2. calculating, according to principles of a minimum variance beamforming method, an apodization weight vector for channels based on the delay channel data obtained in the S1; and at the same time, performing short-time Fourier transform on the delay channel data obtained in the S1 to obtain a frequency domain delay channel data;
S3. suppressing off-axis scatter signals in the frequency domain delay channel data obtained in the S2 by using a deep neural network, to obtain the frequency domain delay channel data with the off-axis scatter signals, which are suppressed;
S4. performing an inverse short-time Fourier transform on the frequency domain delay channel data with the off-axis scatter signals, which are suppressed, obtained in the S3 to obtain the delay channel data of each channel that have been processed by the deep neural network;
S5. dividing the delay channel data that have been processed by the deep neural network obtained in the S4 into corresponding sub-aperture vectors; and
S6. performing a weighted summation and calculating an average with the apodization weight vector obtained in the S2 and sub-aperture vectors of the delay channel data that have been processed by the deep neural network obtained in the S5 to obtain image pixel values of the target object for ultrasound imaging, which is corresponded, thereby forming a complete ultrasound image.

2. The high-contrast minimum variance imaging method based on deep learning of claim 1 , wherein in the step S1, the scanning the target object for ultrasound imaging, generating the channel data composed of the echo signals received by the reception channels of the ultrasound transducer, and then performing a delay operation on the channel data comprises: calculating a delay time according to a position of each of target points for imaging, a position of each of scan lines and a position of each reception channel, and mapping the delay time into a signal subscript, so as to extract a signal value corresponding to the target points for imaging in the echo signals of a reception channel and obtain the delay channel data; and
let the number of the target points for imaging on one of the scan lines be P and the number of the reception channels be N, so that a P×N delay channel data matrix is obtained after the delay operation, and let the number of the scan lines be L, then a P×N×L delay channel data matrix M1 is needed for imaging each time, and subsequent steps will be performed based on this delay channel data matrix.

3. The high-contrast minimum variance imaging method based on deep learning of claim 1 , wherein in the step S2, for one of target points for imaging, a delay channel data vector of a length N can be extracted from a delay channel data matrix M1; according to the principles of the minimum variance beamforming method, a spatial smoothing technique is used, that is, a full aperture that contains all the reception channels is divided into several overlapping sub-apertures, a covariance matrix of the delay channel data in each of sub-apertures is calculated individually, and then an average of the covariance matrices of all the sub-apertures is calculated; let the number of the channels of each of the sub-apertures be M, then there are N−M+1 sub-apertures in total, and let the sub-aperture vectors of the delay channel data be xi, where i=1, 2, . . . , N−M+1, and xi contains the delay channel data of i-th to (i+M−1)-th reception channels; then, according to the following formula, the covariance matrix of the delay channel data in each of the sub-apertures is calculated and an average is calculated to obtain a final estimated covariance matrix Rcov:

R
cov

=

1

N
-
M
+
1

⁢

∑

i
=
1

N
-
M
+
1

x
i

·

x
i
H

;

where · represents a vector multiplication, and H represents a conjugate transposition; a minimum variance beamforming method aims at minimizing a variance of pixel values of the target points for imaging, and a optimization problem, i.e., to minimize the variance of the pixel values of the target points for imaging, is expressed as the following formula:

min
w

w
H

·

R
cov

·
w

⁢

s
.
t
.

w
H

·
a

=
1

;

where a is an all-ones vector, · represents the vector multiplication, and w is the apodization weight vector of the channels; and the solution to the optimization problem is:

w
=

R
cov

-
1

·
a

a
H

·

R
cov

-
1

·
a

;

where −1 represents a matrix inversion, and · represents the vector multipl

Assignees

• Univ South China Tech

Inventors

• Chen Junying
• Zhuang Renxin

Classifications

• G06N3/0499

  Feedforward networks · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06T2207/20084

  Artificial neural networks [ANN] · CPC title

• G06T2207/20081

  Training; Learning · CPC title

• G06T2207/10132

  Ultrasound image · CPC title