Deep learning-based gastric antrum identification and gastric antrum dynamic index calculation method

Abstract

The invention belongs to the field of medical images, and relates to a deep learning-based antral sinus recognition and antral sinus dynamic index calculation method, which comprises the following steps of: acquiring a standard antral sinus contraction ultrasonic video with a preset time length; the standard antrum constriction ultrasonic video is a video of antrum under a standard antrum slice obtained through ultrasonic waves; performing fuzzy removal processing on the standard antral systolic ultrasonic video to obtain a first antral frame image sequence; performing artifact suppression processing on the first antral sinus frame image sequence to obtain a second antral sinus frame image sequence; tracking the antral region in the second antral frame image sequence to obtain an antral binarization segmentation image; and calculating the indexes such as the antral antrum dynamic index and the contraction frequency of the to-be-detected object in the current state under the bedside ultrasonic image based on a deep learning algorithm system.

Description

Technical Field

[0001] The present invention relates to the field of medical imaging, and specifically discloses a method for antrum identification and antrum motility index calculation based on deep learning. Background Art

[0002] The assessment of gastrointestinal function is a key link in the enteral nutrition support for patients. With the development of ultrasound, trained medical staff can qualitatively and quantitatively evaluate gastrointestinal contents, gastric residual volume, and monitor gastrointestinal motility accurately through the single antrum section method. Currently, clinically, the manual method is mostly used to measure the cross-sectional area of the antrum, that is, doctors or nurses manually mark the long and short axes of the antrum area and approximate it with the area of an ellipse, or doctors or nurses manually trace the boundary of the antrum to obtain the antrum area. However, this operation requires relatively high capabilities of doctors or nurses, requires relatively precise marking, is prone to misoperation, and takes a long time. At the same time, this method is too passive to achieve real-time monitoring of patients. Currently, there is a lack of medical staff resources, and many hospitals do not have the conditions to assign a doctor or nurse to conduct examinations on the detection object in real time. In the existing technology, doctors or nurses choose the single antrum section method. The specific method includes: placing the ultrasound probe below the xiphoid process of the patient at an angle perpendicular to the abdomen, and simultaneously examining the antrum, superior mesenteric artery, left lobe of the liver, and abdominal aorta to locate the antrum position. The ultrasound shows the size of the antrum, and the area of the antrum is calculated by measuring the transverse diameter and anteroposterior diameter of the antrum. Therefore, there is an urgent need for a method and system that can perform real-time monitoring and evaluation of the antrum motility index, so that doctors can timely determine the antrum movement and blood flow changes of the measured person.

[0003] In view of this, the present invention discloses a method for antrum identification and antrum motility index calculation based on deep learning, which calculates indicators reflecting gastrointestinal function such as the antrum motility index and the number of contractions of the patient in the current state under the image of bedside ultrasound based on the deep learning algorithm system, and overcomes the problems of insufficient medical staff, inaccurate manual counting, and inability to dynamically measure relevant indicators in the current clinical practice. Summary of the Invention

[0004] The object of the present invention is to provide a method for antrum identification based on deep learning. The specific scheme includes: collecting a standard antrum contraction ultrasound video with a preset time length; the standard antrum contraction ultrasound video refers to a video of the antrum under the standard antrum section obtained by ultrasonic waves; performing a blur removal process on the standard antrum contraction ultrasound video to obtain a first antrum frame image sequence; performing an artifact suppression process on the first antrum frame image sequence to obtain a second antrum frame image sequence; tracking the antrum area in the second antrum frame image sequence to obtain a binarized segmentation image of the antrum.

[0005] Further, the acquisition of the standard antral contraction ultrasound video with a preset time length includes: placing an ultrasound probe under the xiphoid process of the object to be detected, and obtaining antral slice images in the sagittal direction; using a target recognition algorithm to identify multiple key structures in the obtained antral slice images; the multiple key structures are used to identify the relative position of the antrum in the abdominal cavity; based on the multiple key structures, determining whether the obtained antral slice image belongs to the standard antral slice image; when the antral slice image does not belong to the standard antral slice image, issuing an alarm and giving the probe adjustment direction until the obtained antral slice image belongs to the standard antral slice image; repeating the operation of obtaining antral slice images until a video of standard antral slices with a preset time length is obtained.

[0006] Further, the multiple key structures include the antrum, the liver, and the abdominal aorta. By identifying the features and spatial coordinate relationships of the multiple key structures, it is determined whether the antral slice image is the standard antral slice image.

[0007] Further, the blurring removal process of the standard antral contraction ultrasound video to obtain the first antral frame image sequence includes: decomposing the standard antral contraction ultrasound video frame by frame to obtain an initial antral image sequence; dynamically dividing multiple standard antral slice images in the initial antral image sequence respectively to obtain multiple groups of antral region blocks; the antral region blocks include near-field region blocks and far-field region blocks; the scale of the near-field region blocks is smaller than the scale of the far-field region blocks; calculating the gray level histograms of each block in multiple groups of antral region blocks respectively to obtain multiple groups of gray level histogram blocks; based on the contrast limit threshold, respectively cropping and allocating each block in multiple groups of gray level histogram blocks to obtain multiple groups of cropped and allocated blocks; equalizing each block in multiple groups of cropped and allocated blocks respectively to obtain multiple groups of block cumulative distribution functions; mapping multiple groups of antral region blocks respectively based on multiple groups of block cumulative distribution functions to obtain multiple mapped antral images; interpolating each pixel in the multiple mapped antral images respectively to obtain the first antral frame image sequence.

[0008] Further, the calculation formula for the gray level histogram of each block in multiple groups of antral region blocks is: ; where represents the number of times the pixel with gray level g appears in the gray level histogram of the th antral region block, that is, the value of the gray level histogram block; i represents the abscissa variable of the antral region block; j represents the ordinate variable of the antral region block; N represents the scale of the antral region block; represents the Dirac function; represents the gray level value of the pixel at the th row and the Represents the grayscale value variable; The calculation formula for separately performing cropping and allocation on each block in multiple groups of grayscale histograms is: ; ; ; Among them, Represents the histogram obtained after cropping and allocation of the th gastric antrum region block, that is, the value of the cropped and allocated block; min represents taking the minimum value; Represents the maximum allowed pixel; Represents the cropping amount; L represents the number of gray levels; max represents taking the maximum value; Represents the contrast limit threshold; The expression of the block cumulative distribution function is: ; Among them, Represents the cumulative distribution function of the th gastric antrum region block; m represents the pixel variable of the cropped and allocated block; Represents the th gastric antrum region block's histogram of the cropped and allocated block.

[0009] Furthermore, the artifact suppression process for the first gastric antrum frame image sequence to obtain the second gastric antrum frame image sequence includes: processing the continuous preset frame video of the first gastric antrum frame image sequence using three-dimensional wavelet basis functions to obtain banded gastric antrum frame images in multiple bands; the multiple bands include a low-frequency sub-band and a high-frequency sub-band; locating the artifacts in the main frequency band of gastric antrum contraction to obtain the artifact region; the main frequency band of gastric antrum contraction is characterized by spatial high frequency and temporal medium frequency; based on the artifact region, performing soft threshold processing on the banded gastric antrum frame images in the high-frequency sub-band to obtain the second gastric antrum frame image sequence.

[0010] Furthermore, the expression of the three-dimensional wavelet basis function is: ; , Represents the three-dimensional wavelet transform coefficient; Represents the three-dimensional video signal, that is, the continuous preset frame video of the first gastric antrum frame image sequence; Represents the three-dimensional wavelet basis function; , and respectively represent the horizontal axis displacement parameter, the vertical axis displacement parameter, and the time displacement parameter; x and y represent the x, y spatial domain; t represents the time domain; Represents the integral operation on the three-dimensional space; Indicates the number of decomposition levels; The calculation formula for the soft threshold processing is: ; ; Wherein, Indicates the soft threshold; Indicates the noise standard deviation estimation; Indicates the natural logarithm function; Indicates the median of the absolute value; HHH indicates the highest frequency sub-band.

[0011] Further, tracking the antrum region in the second antrum frame image sequence to obtain an antrum binary segmentation image includes: segmenting the second antrum frame image sequence through a first time window to obtain multiple groups of antrum frame image subsequences; for the first frame image of each group of antrum frame image subsequences, using an image segmentation network to extract the antrum boundary point set to obtain a network boundary point set; for the remaining frame images of each group of antrum frame image subsequences, using an optical flow field to track the antrum boundary point set to obtain an optical flow boundary point set; fusing the network boundary point set and the optical flow boundary point set through Kalman filtering to obtain an optimal boundary; using a spline curve to fit the optimal boundary, segmenting the antrum region of the current frame, and outputting the antrum binary segmentation image.

[0012] The present invention also provides a method for calculating the antrum motility index of the antrum recognition method based on deep learning according to any one of the above, including: determining the cross-sectional area size of the antrum binary segmentation image, and calculating the actual physical area of the antrum frame by frame to obtain an antrum area sequence; dividing the antrum area sequence through a second time window to obtain multiple antrum area subsequences; determining the minimum value and the maximum value in each antrum area subsequence, and taking the minimum value as the contraction wave peak and the maximum value as the relaxation wave valley to obtain a motion sequence; taking the alternately appearing contraction wave peaks and relaxation wave valleys as one contraction time, performing a period determination on the motion sequence to obtain the total contraction times of the object to be detected within a preset time length; based on the total contraction times, determining the antrum motility index of the contraction times.

[0013] Further, calculating the actual physical area of the antrum includes: based on the antrum binary segmentation image, statistically calculating the area and position of the connected region; taking the connected region with the largest area as the antrum region, and extracting the contour of the antrum region; statistically calculating the total number of pixels in the antrum region to determine the antrum pixel area; based on the antrum pixel area, determining the actual physical area of the antrum.

[0014] The present invention has the following advantages and beneficial effects: The present invention collects the antrum motility index and the number of contractions of the person to be measured by fixing an ultrasonic probe on the person to be measured, and compares the collected ultrasonic images with standard images in real time to correct the position and angle of the ultrasonic probe, so as to obtain a standard short-axis section of the antrum. The collected data is transmitted to the background and various data of the patient are calculated through a deep learning algorithm, which improves the accuracy of data calculation and avoids manual errors.

[0015] The present invention can perform real-time determination and early warning on non-standard antrum video images by judging and identifying the antrum section images obtained by ultrasonic waves.

[0016] The present invention can automatically segment the antrum region and extract the parameter of the cross-sectional area size of the antrum by tracking and identifying the antrum.

[0017] The present invention can identify antrum contractions and relaxations that cannot be recognized by the naked eye through continuous recording of the antrum cross-sectional area and peak detection algorithm in time series. Description of the Drawings

[0018] Figure 1 is an exemplary flowchart of a method for antrum recognition based on deep learning proposed by the present invention; Figure 2 is an exemplary schematic diagram of a standard antrum section in the present invention. Detailed Embodiments

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Usually, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0020] Figure 1 is an exemplary flowchart of a method for antrum recognition based on deep learning proposed by the present invention. As Figure 1 shown, the method for antrum recognition based on deep learning includes: Collecting a standard antrum contraction ultrasonic video with a preset time length; the standard antrum contraction ultrasonic video refers to a video of the antrum under a standard antrum section obtained by ultrasonic waves. The preset time length refers to the time length set in advance for obtaining the standard antrum video. For example, the preset time length can be 6 minutes, that is, collecting a 6-minute standard antrum contraction ultrasonic video. The standard antrum section can refer to using the abdominal aorta and the left lobe of the liver as the antrum section markers, and the antrum is located within the triangle formed by the liver and the abdominal aorta. As Figure 2 shown, I represents the antrum, II represents the left lobe of the liver, III represents the pancreas, IV represents the superior mesenteric artery, and V represents the abdominal aorta.

[0021] In some embodiments, collecting the standard antral contraction ultrasound video of a preset time length includes: Place the ultrasound probe under the xiphoid process of the object to be detected and obtain antral slice images in the sagittal direction. The object to be detected can refer to humans who come to detect gastrointestinal functions. The antral slice direction refers to the image of the antrum obtained by the ultrasound probe. Use a target recognition algorithm to identify multiple key structures in the obtained antral slice images; the multiple key structures are used to identify the relative position of the antrum in the abdominal cavity. The target recognition algorithm can refer to the YOLOv8s deep learning model. The multiple key structures include the antrum, the liver, and the abdominal aorta. By the feature recognition and spatial coordinate relationship of the multiple key structures, determine whether the antral slice image is the standard antral slice image. For example, for the anatomical structure recognition of the standard section, use the YOLOv8s deep learning model to detect three key structures in the B-ultrasound image: the antrum (a hypoechoic tubular structure located below the liver margin); the liver (a medium to high echo homogeneous area, used as a reference for antrum positioning); the abdominal aorta (an anechoic circular / tubular structure). By the feature recognition and spatial coordinate relationship of the three key structures, determine whether the image is the standard section. Based on the multiple key structures, determine whether the obtained antral slice image belongs to the standard antral slice image. When the antral slice image does not belong to the standard antral slice image, issue an alarm and give the probe adjustment direction until the obtained antral slice image belongs to the standard antral slice image. For example, use an edge computing module, a probe posture sensor, and an audible and visual alarm unit to achieve that when an unqualified image is collected, a red frame alarm appears on the screen, and specific prompts for the probe adjustment direction and angle are given to ensure that each frame of the antral contraction video collected in 6 minutes is qualified. Repeat the operation of obtaining the antral slice image until a video of the standard antral slice of a preset time length is obtained.

[0022] Perform blurring removal processing on the standard antral contraction ultrasound video to obtain the first antral frame image sequence. The first antral frame image sequence refers to a sequence composed of multiple ultrasound video frame images after blurring removal. For example, the CLAHE algorithm can be used to enhance the contrast and solve the local blurring problem of the ultrasound image.

[0023] In some embodiments, the blurring removal processing includes: decomposing the standard antral contraction ultrasound video frame by frame to obtain an initial antral image sequence.

[0024] Dynamically partition multiple standard gastric antrum slice images in the initial gastric antrum image sequence respectively to obtain multiple groups of gastric antrum region blocks; the gastric antrum region blocks include near-field region blocks and far-field region blocks; the scale of the near-field region blocks is smaller than that of the far-field region blocks. For example, decompose the gastric antrum ultrasound video frame by frame into static images, process each frame separately, dynamically partition the standard gastric antrum image, for the near-field region (depth < 6 cm), use 8x8 small partitions to enhance the fine structure of the gastric wall; for the far-field region (depth ≥ 6 cm), switch to 16x16 partitions to suppress deep noise.

[0025] Calculate each block in multiple groups of gastric antrum region blocks respectively of the value of the grayscale histogram , where , to obtain multiple groups of grayscale histogram blocks. The calculation formula for the grayscale histogram of each block in multiple groups of gastric antrum region blocks is: ; where represents the number of times the pixel with grayscale value g appears in the grayscale histogram of the th gastric antrum region block, that is, the value of the grayscale histogram block; i represents the abscissa variable of the gastric antrum region block; j represents the ordinate variable of the gastric antrum region block; N represents the scale of the gastric antrum region block; represents the Dirac function, (if , then δ = 1, otherwise δ = 0); represents the grayscale value of the pixel at the th row and jth column of the gastric antrum region block; represents the grayscale value variable.

[0026] Based on the contrast limit threshold, perform clipping and distribution on each block in multiple groups of grayscale histogram blocks respectively to obtain multiple groups of clipped and distributed blocks. The contrast limit threshold refers to a pre-set value for performing clipping and distribution, usually = 2.0 ∼ 3.0. Clipping and distribution means clipping the histogram value of the pixels exceeding the allowed maximum and evenly distributing the clipped amount to all gray levels. The calculation formula for performing clipping and distribution on each block in multiple groups of grayscale histogram blocks respectively is: ; ; ; where represents the histogram obtained after performing clipping and distribution on the th gastric antrum region block, that is, the value of the clipped and distributed block; min represents taking the minimum value; represents the allowed maximum pixel; represents the cropping amount; L represents the number of gray levels; max represents taking the maximum value; represents the contrast limit threshold.

[0027] Equalize each block in multiple groups of cropping assignment blocks respectively to obtain multiple groups of block cumulative distribution functions. The expression of the block cumulative distribution function is: ; where represents the cumulative distribution function of the th gastric antrum region block; m represents the pixel variable of the cropping assignment block; represents the th histogram of the cropping assignment block of the gastric antrum region block.

[0028] Map multiple groups of gastric antrum region blocks respectively based on multiple groups of block cumulative distribution functions to obtain multiple mapped gastric antrum images. The calculation formula of the mapping is: ; where represents the mapping formula; represents rounding the input value to obtain the closest integer.

[0029] Interpolate each pixel in the multiple mapped gastric antrum images respectively to obtain the first gastric antrum frame image sequence, so as to avoid discontinuity between blocks and retain the continuity of the gastric antrum edge.

[0030] Perform artifact suppression processing on the first gastric antrum frame image sequence to obtain a second gastric antrum frame image sequence. The second gastric antrum frame image sequence refers to a sequence composed of images obtained by performing artifact suppression processing on the images in the first gastric antrum frame image sequence. In some embodiments, performing the artifact suppression processing on the first gastric antrum frame image sequence to obtain the second gastric antrum frame image sequence includes: processing the continuous preset frame video of the first gastric antrum frame image sequence by using a three-dimensional wavelet basis function to obtain band gastric antrum frame images in multiple bands; the multiple bands include a low-frequency sub-band and high-frequency sub-bands. The continuous preset frame video refers to the time length of the frame video to be processed set in advance. For example, 15 consecutive frame videos can be regarded as a three-dimensional signal (space (x, y) + time t) and decomposed into a low-frequency (LLL) and high-frequency sub-bands (HLL, LHL, LLH, HHH). The expression of the three-dimensional wavelet basis function is: ; where represents the three-dimensional wavelet transform coefficient; represents the three-dimensional video signal, that is, the continuous preset frame video of the first gastric antrum frame image sequence; represents the three-dimensional wavelet basis function; 、 and respectively represent the horizontal axis displacement parameter, the vertical axis displacement parameter, and the time displacement parameter; x and y represent the x, y spatial domain; t represents the time domain; represents the integral operation on the three-dimensional space; represents the number of decomposition levels.

[0031] Locate the artifacts in the main frequency band of gastric antrum contraction to obtain the artifact region; the characteristics of the main frequency band of gastric antrum contraction are high spatial frequency and medium time frequency (e.g., HHL / LHL / HLH).

[0032] Perform soft threshold processing on the frequency band gastric antrum frame images of the high-frequency sub-band based on the artifact region to obtain the second gastric antrum frame image sequence. The calculation formula for the soft threshold processing is: ; ; wherein, represents the soft threshold; represents the noise standard deviation estimation; represents the natural logarithm function; represents the median of the absolute value; HHH represents the highest frequency sub-band.

[0033] Track the gastric antrum region in the second gastric antrum frame image sequence to obtain the binary segmentation image of the gastric antrum. The binary segmentation image of the gastric antrum refers to an image in which the gastric antrum region is white and the background is black. In some embodiments, the tracking of the gastric antrum region in the second gastric antrum frame image sequence to obtain the binary segmentation image of the gastric antrum includes: segmenting the second gastric antrum frame image sequence through a first time window to obtain multiple groups of gastric antrum frame image subsequences. The first time window refers to the time length for segmenting the second gastric antrum frame image sequence. The gastric antrum frame image subsequence refers to an image sequence of multiple second time window lengths obtained by segmenting the second gastric antrum frame image according to the first time window. For example, the first time window can be 6 frames of images.

[0034] For the first frame image of each group of gastric antrum frame image subsequences, use an image segmentation network to extract the gastric antrum boundary point set to obtain the network boundary point set. The first frame image refers to the first frame image of the subsequence. For example, for the first subsequence, its first frame image can be the first frame image of the ultrasound video. The network boundary point set refers to the set of boundary points obtained by segmenting the gastric antrum image through the segmentation network U-Net and extracting the boundary points of the gastric antrum.

[0035] For the remaining frame images of each subsequence of gastric antrum frame images, the optical flow field is used to track the set of gastric antrum boundary points to obtain an optical flow boundary point set. The remaining frame images refer to the frame images in the subsequence except the first frame image. For example, for a subsequence with a first time window of 6 frames, the remaining frame images of the first subsequence are the second to sixth frame images. The optical flow boundary point set refers to the set of boundary points obtained by calculating and updating the boundary points through the optical flow field.

[0036] The network boundary point set and the optical flow boundary point set are fused through Kalman filtering to obtain an optimal boundary. The optimal boundary points refer to the fused boundary points.

[0037] The optimal boundary is fitted using a spline curve to segment the gastric antrum region of the current frame in real time, and the binary segmentation image of the gastric antrum is output.

[0038] The present invention also provides a method for calculating the gastric antrum motility index, including: determining the cross-sectional area size of the binary segmentation image of the gastric antrum, and calculating the actual physical area of the gastric antrum frame by frame to obtain a gastric antrum area sequence , where S represents the gastric antrum area sequence, and its elements include multiple gastric antrum areas, the first gastric antrum area , the second gastric antrum area and the Tth gastric antrum area etc., and T represents the total number of frames. The actual physical area refers to the true area of the gastric antrum in the actual situation, and the gastric antrum area sequence refers to the sequence obtained by sorting the gastric antrum areas by time. In some embodiments, calculating the actual physical area of the gastric antrum includes: based on the binary segmentation image of the gastric antrum, statistically calculating the area and position of the connected regions. The connected region refers to a set of pixels in the binary image that have the same pixel value and are connected through adjacent pixels. The connected region with the largest area is taken as the gastric antrum region, and the contour of the gastric antrum region is extracted.

[0039] Statistically calculate the total number of pixels (white pixels in the mask image, with a value of 255) in the gastric antrum region to determine the gastric antrum pixel area, and use the gastric antrum pixel area as a preliminary measure of the gastric antrum area. The calculation formula for the gastric antrum pixel area is: ; where represents the gastric antrum pixel area; x and y respectively represent the horizontal axis variable and the vertical axis variable of the binary segmentation image; W and H respectively represent the width and height of the binary segmentation image; mask represents the mask region, with a value of 255.

[0040] Based on the gastric antrum pixel area, determine the actual physical area of the gastric antrum. The calculation formula for the actual physical area is: ; Among them, actual_area represents the actual physical area; pixel_area represents the pixel area of the gastric antrum; represents the standard physical area; represents the standard pixel area, and the division of the two gives the pixel-square conversion coefficient.

[0041] The gastric antrum area sequence is divided through the second time window to obtain multiple gastric antrum area subsequences. The second time window refers to the time period for dividing the gastric antrum area sequence, which can be divided according to experience or actual situation. The gastric antrum area subsequence refers to the sequence composed of the gastric antrum areas in multiple time periods of the length of the second time window obtained by dividing the gastric antrum area sequence.

[0042] Determine the minimum value and the maximum value in each gastric antrum area subsequence, and take the minimum value as the contraction peak and the maximum value as the relaxation trough to obtain the motion sequence. The contraction peak refers to the minimum value of contraction in each gastric antrum area subsequence during gastric antrum movement. The relaxation trough refers to the maximum value of relaxation in each gastric antrum area subsequence during gastric antrum movement.

[0043] Take the alternately appearing contraction peaks and relaxation troughs as the number of contractions once, and perform cycle determination on the motion sequence to obtain the total number of contractions of the object to be detected within the preset time length: ; Among them, represents the number of gastric antrum contractions within the preset time length (e.g., 6 minutes); min represents taking the minimum value; represents the contraction peak; represents the relaxation trough.

[0044] Based on the total number of contractions, determine the gastric antrum motility index. Taking 6 minutes as an example, calculating the gastric antrum motility index (MI) requires splitting the 6-minute gastric antrum contraction video into 3 2-minute videos, and taking the number of gastric antrum contractions per 2 minutes as the gastric antrum contraction frequency (antral contraction frequency, ACF); continuously measuring the maximum relaxation ( ) and the minimum contraction ( ) areas of the gastric antrum 3 times, and the gastric antrum contraction amplitude (antral contraction amplitude) .

[0045] = ACF x ACA; 1. Calculate the gastric antrum contraction frequency ( ) segment by segment: For each 2-minute segment U (U = 1, 2, 3), = the number of contractions within the Uth 2-minute segment.

[0046] 2. Segmentally calculate the antral contraction amplitude ( ): ; is the antral contraction amplitude in the U-th 2-minute period, is the number of complete cycles within the U-th 2-minute period, is the cycle serial number, is the maximum diastolic area in the -th cycle, is the minimum contraction area in the

[0047] 3. Calculate the antral motility index MI: ; MI is the final antral motility index.

[0048] The above is only the preferred embodiment of the present invention and is not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A gastric antrum recognition method based on deep learning, characterized in that, Including: Collecting a standard antrum contraction ultrasound video of a preset time length; the standard antrum contraction ultrasound video refers to a video of the antrum under a standard antrum slice obtained by ultrasound. Performing a blurring removal process on the standard antrum contraction ultrasound video to obtain a first antrum frame image sequence. Performing an artifact suppression process on the first antrum frame image sequence to obtain a second antrum frame image sequence. Tracking the antrum region in the second antrum frame image sequence to obtain an antrum binary segmentation image.

2. The gastric antrum recognition method based on deep learning according to claim 1, wherein The collecting of the standard antrum contraction ultrasound video of a preset time length includes: Placing an ultrasound probe under the xiphoid process of the object to be detected and obtaining an antrum slice image in the sagittal direction. Using a target recognition algorithm to identify multiple key structures in the obtained antrum slice image; the multiple key structures are used to identify the relative position of the antrum in the abdominal cavity. Based on the multiple key structures, determining whether the obtained antrum slice image belongs to the standard antrum slice image. When the antrum slice image does not belong to the standard antrum slice image, issuing an alarm and giving a probe adjustment direction until the obtained antrum slice image belongs to the standard antrum slice image. Repeating the operation of obtaining the antrum slice image until a video of the standard antrum slice of a preset time length is obtained.

3. The gastric antrum recognition method based on deep learning according to claim 1, characterized in that, The multiple key structures include the antrum, the liver, and the abdominal aorta. By the feature recognition and spatial coordinate relationship of the multiple key structures, it is judged whether the antrum slice image is the standard antrum slice image.

4. The gastric antrum recognition method based on deep learning according to claim 1, characterized in that The performing of the blurring removal process on the standard antrum contraction ultrasound video to obtain a first antrum frame image sequence includes: Frame-by-frame decomposing the standard antrum contraction ultrasound video to obtain an initial antrum image sequence. Performing dynamic block division on multiple standard antrum slice images in the initial antrum image sequence respectively to obtain multiple groups of antrum region blocks; the antrum region blocks include a near-field region block and a far-field region block; the scale of the near-field region block is smaller than the scale of the far-field region block. Calculating the gray-level histogram of each block in multiple groups of antrum region blocks respectively to obtain multiple groups of gray-level histogram blocks. Based on a contrast-limiting threshold, performing clipping and allocation on each block in multiple groups of gray-level histogram blocks respectively to obtain multiple groups of clipped and allocated blocks. Performing equalization on each block in multiple groups of clipped and allocated blocks respectively to obtain multiple groups of block cumulative distribution functions. Based on multiple groups of block cumulative distribution functions, performing mapping on multiple groups of antrum region blocks respectively to obtain multiple mapped antrum images. Performing interpolation on each pixel in the multiple mapped antrum images respectively to obtain the first antrum frame image sequence.

5. The gastric antrum recognition method based on deep learning according to claim 4, characterized in that The calculation formula for the gray-level histogram of each block in multiple groups of antrum region blocks is: ； Among them, represents the number of occurrences of pixels with gray value g in the gray-level histogram of the th gastric antrum region block, that is, the value of the gray-level histogram block; i represents the abscissa variable of the gastric antrum region block; j represents the ordinate variable of the gastric antrum region block; N represents the scale of the gastric antrum region block; represents the Dirac function; represents the gastric antrum region block at the gray value of the pixel in the i-th row and j-th column; represents the gray value variable; The calculation formula for performing clipping and allocation on each block in multiple groups of gray-level histogram blocks respectively is: ； ； ； Among them, represents the histogram obtained after cropping and allocating the th gastric antrum region block, that is, the value of the cropped and allocated block; min represents taking the minimum value; represents the maximum allowed pixel; represents the cropping amount; L represents the number of gray levels; max represents taking the maximum value; represents the contrast limit threshold; The expression of the block cumulative distribution function is: ； Among them, represents the cumulative distribution function of the th gastric antrum region block; m represents the pixel variable of the cropping assignment block; represents the histogram of the cropping assignment block of the th gastric antrum region block.

6. The gastric antrum recognition method based on deep learning according to claim 1, characterized in that, The performing of the artifact suppression process on the first antrum frame image sequence to obtain a second antrum frame image sequence includes: Processing the continuous preset-frame video of the first antrum frame image sequence by using a three-dimensional wavelet basis function to obtain band antrum frame images of multiple frequency bands; the multiple frequency bands include a low-frequency sub-band and a high-frequency sub-band. Locate the artifacts in the main frequency band of antral contraction to obtain the artifact region; Perform soft threshold processing on the frequency-band antral frame images of the high-frequency sub-band based on the artifact region to obtain the second antral frame image sequence.

7. The gastric antrum recognition method based on deep learning according to claim 6, wherein, The expression of the three-dimensional wavelet basis function is: ； Among them, represents the three-dimensional wavelet transform coefficient; represents the three-dimensional video signal, that is, the continuous preset frame video of the first gastric antrum frame image sequence; represents the three-dimensional wavelet basis function; , and respectively represent the horizontal axis displacement parameter, the vertical axis displacement parameter and the time displacement parameter; x and y represent the x, y spatial domain; t represents the time domain; represents the integral operation of the three-dimensional space; represents the decomposition level; The calculation formula for the soft threshold processing is: ； ； Among them, represents the soft threshold; represents the noise standard deviation estimation; represents the natural logarithm function; represents the median of the absolute value; HHH represents the highest frequency sub-band.

8. The gastric antrum recognition method based on deep learning according to claim 1, characterized in that Tracking the antral region in the second antral frame image sequence to obtain the antral binary segmentation image, including: Segment the second antral frame image sequence through the first time window to obtain multiple groups of antral frame image subsequences; For the first frame image of each group of antral frame image subsequences, use an image segmentation network to extract the antral boundary point set to obtain the network boundary point set; For the remaining frame images of each group of antral frame image subsequences, use the optical flow field to track the antral boundary point set to obtain the optical flow boundary point set; Fuse the network boundary point set and the optical flow boundary point set through Kalman filtering to obtain the optimal boundary; Use a spline curve to fit the optimal boundary, segment the antral region of the current frame, and output the antral binary segmentation image.

9. The antral motility index calculation method of the antrum recognition method based on deep learning according to any one of claims 1-8, characterized in that Including: Determine the cross-sectional area size of the antral binary segmentation image, and calculate the actual physical area of the antrum frame by frame to obtain the antral area sequence; Divide the antral area sequence through the second time window to obtain multiple antral area subsequences; Determine the minimum value and the maximum value in each antral area subsequence, and use the minimum value as the contraction peak and the maximum value as the relaxation trough to obtain the motion sequence; Use the alternately appearing contraction peaks and relaxation troughs as the number of contractions once, and perform period determination on the motion sequence to obtain the total number of contractions of the object to be detected within the preset time length; Based on the total number of contractions, determine the antral motility index of the number of contractions.

10. The method for calculating the antral motility index according to claim 9, wherein The calculation of the actual physical area of the antrum includes: Based on the antral binary segmentation image, count the area and position of the connected regions; Take the connected region with the largest area as the antral region, and extract the contour of the antral region; Count the total number of pixels in the antral region to determine the antral pixel area; Based on the antral pixel area, determine the actual physical area of the antrum.

Images