Abdominal ultrasound examination method, system and apparatus

Abstract

The present disclosure provides an abdominal ultrasound examination method, system and device, which can acquire and preprocess abdominal ultrasound images to obtain effective region images, input the effective region images into a pre-trained standard section recognition network to obtain a section type recognition result, perform organ segmentation when the recognition result is a standard section, continuously track the organ segmentation result of the current frame to obtain the organ segmentation result of the subsequent frame, output the frame-by-frame organ segmentation mask image to obtain the final organ segmentation result, schedule a lesion recognition network corresponding to the organ type according to the final organ segmentation result to perform inference, obtain the lesion detection result of the organ in the several organ layers, and finally calculate the size of the several organs and the lesions on the organs to generate an abdominal ultrasound examination data statistical result. In the present disclosure, the organ recognition and lesion detection accuracy is high, the misrecognition and missed detection situations are avoided, the examination efficiency is high, and comprehensive and accurate auxiliary diagnosis support can be provided.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of ultrasound images, and in particular to an abdominal ultrasound examination method, system and device. BACKGROUND

[0002] At present, the application of artificial intelligence in the medical imaging field mainly focuses on single-organ lesion detection, such as thyroid, breast, heart and other specific organ ultrasound image recognition systems. These systems can achieve good automatic recognition and lesion detection effect through deep learning algorithm training on standard sections.

[0003] However, abdominal ultrasound, as a complex examination scenario involving multiple organs and various sections, presents significant challenges to existing artificial intelligence technology. Unlike single-organ examination, abdominal ultrasound requires simultaneous identification and analysis of multiple organs such as liver, gallbladder, pancreas, and spleen, involving numerous sections, far exceeding the complexity of single-organ examination. Existing systems mainly rely on standard section training, with poor adaptability to variant sections and non-standard sections, resulting in low organ recognition accuracy and inaccurate boundary segmentation. In addition, existing technology lacks a dedicated lesion recognition model for abdominal multi-organ characteristics, failing to fully utilize the anatomical and functional features of different organs.

[0004] Therefore, there is currently a lack of a complete abdominal ultrasound artificial intelligence solution that can achieve full-process automation from standard section recognition to multi-organ segmentation to lesion detection. Compared to single-organ examination systems with relatively mature technology, the multi-organ and multi-section characteristics of abdominal ultrasound make it a technical difficulty in the field of artificial intelligence assisted diagnosis. SUMMARY

[0005] The purpose of the present disclosure is to provide an abdominal ultrasound examination method, system and device to solve the above technical problems.

[0006] To achieve the above purpose, in a first aspect, the present disclosure provides an abdominal ultrasound examination method, comprising the following steps:

[0007] Obtaining an abdominal ultrasound image and preprocessing to obtain an effective region image;

[0008] Inputting the effective region image into a pre-trained standard section recognition network to obtain a section type recognition result;

[0009] When the section type recognition result is a standard section, performing organ segmentation to obtain the organ segmentation result of the current frame, continuously tracking the organ segmentation result of the current frame to obtain the organ segmentation result of the subsequent frame until the organ target disappears, and outputting the frame-by-frame organ segmentation mask image to obtain the final organ segmentation result;

[0010] According to the final organ segmentation result, a plurality of organ layers of organ types are dynamically separated, a lesion recognition network corresponding to the organ type is scheduled to perform inference, and lesion detection results of the organs in the plurality of organ layers are obtained.

[0011] Based on the final organ segmentation result, the lesion detection result and the abdominal ultrasound image, the sizes of the plurality of organs and the lesions on the organs are calculated, and an abdominal ultrasound examination data statistical result is generated.

[0012] In a preferred embodiment, the step of obtaining an abdominal ultrasound image and preprocessing to obtain an effective area image comprises:

[0013] An abdominal ultrasound video signal output by an endoscopic device is converted into a sequence of digital image frames by a video decoder, and an RGB format image frame by frame is output.

[0014] The RGB format image frame by frame is subjected to fan-shaped effective area detection to obtain a preprocessed effective area image frame by frame.

[0015] In a preferred embodiment, the step of obtaining an abdominal ultrasound image and preprocessing to obtain an effective area image comprises:

[0016] The input RGB format image is subjected to grayscale processing, and the grayscale value is calculated by a weighted average method:

[0017] G(x,y)=0.299×R(x,y)+0.587×G(x,y)+0.114×B(x,y);

[0018] wherein (x,y) is the coordinate of the pixel in the image, R(x,y), G(x,y), and B(x,y) are the intensity values of the red, green, and blue channels of the pixel at the coordinate (x,y), respectively, and 0.299, 0.587, and 0.114 are the weight coefficients set according to the color sensitivity characteristics of the human eye;

[0019] A radial gradient detection algorithm is used to identify the fan-shaped boundary, and the radial gradient intensity of each pixel point is calculated:

[0020] ;

[0021] wherein r is the polar coordinate radius, θ is the polar coordinate angle, I is the grayscale value G(x,y), represents the rate of change of the grayscale value I in the r direction, and I(r+1,θ)-I(r,θ) is the difference value of the grayscale values of the adjacent two radius pixels at the same angle θ;

[0022] The center position and the boundary arc line of the fan are detected by Hough transformation, and the cumulative function is:

[0023] ;

[0024] wherein (a, b) is the candidate point of the center of the circle, R is the candidate value of the radius of the circle, and δ is the Dirac function;

[0025] According to the detected sector parameter of the radial gradient strength, the center of the circle, and the radius of the circle, a binary mask matrix is generated, pixel values in the sector region are retained, pixel values outside the sector region are set to zero, and a preprocessed effective region image is output.

[0026] In a preferred embodiment, the step of inputting the effective region image into a pre-trained standard section recognition network to obtain a section type recognition result comprises:

[0027] constructing an initial standard section recognition network;

[0028] constructing a data set according to a preset standard section category, training the standard section recognition network, and obtaining a pre-trained standard section recognition network;

[0029] inputting the preprocessed effective region image into the pre-trained standard section recognition network, and obtaining a corresponding section type recognition result through forward propagation calculation, wherein the section type recognition result comprises a section type identifier and a confidence score.

[0030] In a preferred embodiment, the step of inputting the preprocessed effective region image into the pre-trained standard section recognition network and obtaining a corresponding section type recognition result through forward propagation calculation comprises:

[0031] extracting multi-scale feature maps of different scales based on the effective region image;

[0032] obtaining channel attention weights based on the multi-scale feature maps through global average pooling and global maximum pooling operations:

[0033] Att(F) = σ(MLP(AvgPool(F)) + MLP(MaxPool(F)))

[0034] wherein σ is a Sigmoid activation function, MLP is a multi-layer perceptron, F is a feature map, AvgPool(F) represents averaging all pixel values of each channel of the feature map F, and MaxPool(F) represents taking the maximum value of all pixel values of each channel of the feature map F; the channel attention weights are multiplied with the feature map element by element to output a fusion feature map;

[0035] determining whether it is a standard section and a section type based on the fusion feature map.

[0036] In a preferred embodiment, when the slice type identification result is a standard slice, the organ segmentation result of the current frame is obtained by organ segmentation, the organ segmentation result of the subsequent frame is obtained by continuously tracking the organ segmentation result of the current frame until the organ target disappears, and the final organ segmentation result is obtained by outputting the frame-by-frame organ segmentation mask image, comprising:

[0037] Based on the standard slice type identification result of the current frame and the processing state of the previous frame, a processing mode of the current frame is selected according to a preset state transition rule, and the processing mode includes a segmentation mode and a tracking mode;

[0038] When the segmentation mode is triggered, the effective region image after preprocessing of the current frame is input into the pre-trained segmentation network, and the organ segmentation mask image of the current frame is obtained by using the pre-trained segmentation network for forward operation inference;

[0039] When the tracking mode is received, the organ segmentation mask image of the previous frame and the effective region image after preprocessing of the current frame are input into the tracking network for calculation to obtain the organ mask image and the confidence score of the current frame;

[0040] Based on the standard slice type, the organ mask image obtained by segmentation or tracking is reasonably checked, and abnormal organ types are filtered out to obtain the final organ segmentation result.

[0041] In a preferred embodiment, according to the final organ segmentation result, the organ image layers of several organ types are dynamically separated, the lesion recognition network corresponding to the organ type is scheduled for inference to obtain the lesion detection result of the organ in the several organ image layers.

[0042] Based on the standard slice type and the organ segmentation result, the image layers are separated, and the independent sub-image of each organ is output;

[0043] According to the current independent sub-image, the pre-constructed lesion recognition network corresponding to the organ is scheduled, and the lesion recognition result of several organs is output after inference calculation.

[0044] In a preferred embodiment, based on the final organ segmentation result, the lesion detection result and the abdominal ultrasound image, the size of several organs and lesions on the organs is calculated, and the abdominal ultrasound examination data statistical result is generated, comprising:

[0045] The side scale lines of the abdominal ultrasound image are identified, and the scale interval is calculated;

[0046] The pixel distance of adjacent scale lines is counted, and the single-pixel interval is calculated based on the scale interval and the pixel distance;

[0047] establish a mapping relationship between pixels and physical sizes based on the single-pixel spacing and physical sizes of the organs and lesions on the organs on the abdominal ultrasound image;

[0048] calculate the long diameter and the short diameter of the organs and the lesions on the organs according to the mapping relationship between pixels and physical sizes;

[0049] draw the long axis and the short axis on the contours of the organs and the lesions on the organs in the abdominal ultrasound image, mark the numerical values of the long diameter and the short diameter at the endpoints of the long axis and the short axis respectively, and output the marked image and a measurement data report.

[0050] In a second aspect, the present disclosure provides an abdominal ultrasound examination system, comprising: a preprocessing module configured to obtain an abdominal ultrasound image and perform preprocessing to obtain an effective region image;

[0051] a standard section recognition module configured to input the effective region image into a pre-trained standard section recognition network to obtain a section type recognition result;

[0052] an organ segmentation module configured to perform organ segmentation to obtain an organ segmentation result of a current frame when the section type recognition result is a standard section, continuously track the organ segmentation result of the current frame to obtain an organ segmentation result of a subsequent frame until an organ target disappears, and output a frame-by-frame organ segmentation mask image to obtain a final organ segmentation result;

[0053] a lesion recognition module configured to dynamically separate organ layers of a plurality of organ types according to the final organ segmentation result, and schedule a lesion recognition network corresponding to the organ types to perform inference to obtain lesion detection results of the organs in the organ layers;

[0054] an axis measurement module configured to calculate sizes of the organs and the lesions on the organs based on the final organ segmentation result, the lesion detection results, and the abdominal ultrasound image, and generate an abdominal ultrasound examination data statistical result.

[0055] In a third aspect, the present disclosure further provides an electronic device, comprising: a memory and one or more processors; the memory is configured to store one or more computer programs; when the one or more computer programs are executed by the one or more processors, the abdominal ultrasound examination method of any one of the embodiments of the first aspect of the present disclosure is implemented.

[0056] In a fourth aspect, the present disclosure further provides a computer storage medium, wherein the computer storage medium stores a computer program, and when the computer program is executed by a processor, the abdominal ultrasound examination method of any one of the embodiments of the first aspect of the present disclosure is implemented.

[0057] In a fifth aspect, the present disclosure also provides a computer program product comprising a computer program or instructions, which, when executed by a processor, implement the abdominal ultrasound examination method according to any of the embodiments of the first aspect of the present disclosure.

[0058] Compared with the prior art, the abdominal ultrasound examination method, system and device provided by the present disclosure adopt a complete technical process: first, a deep learning neural network is used to identify standard sections of an ultrasound image, then accurate organ segmentation is performed based on the section classification result, and an accurate organ segmentation result is obtained by combining a target tracking technology. On this basis, the system automatically dispatches the corresponding exclusive lesion recognition model according to the segmentation result of different organs to realize targeted lesion detection.

[0059] The technical scheme integrates the segmentation and lesion recognition model of multiple organs such as liver, gallbladder, pancreas, spleen and kidney, forming a full-process automatic processing system from section recognition to organ segmentation to lesion detection. Compared with the existing single-organ detection system, the present scheme significantly improves the accuracy and real-time performance of multi-organ recognition, effectively avoids misrecognition and missed detection problems, greatly improves the efficiency and accuracy of abdominal ultrasound examination, and can provide comprehensive and reliable auxiliary diagnosis support for clinical practice. BRIEF DESCRIPTION OF DRAWINGS

[0060] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some embodiments of the present disclosure, and therefore should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0061] Figure 1 Flowchart of the abdominal ultrasound examination method provided by the present disclosure;

[0062] Figure 2 For Figure 1 Flowchart of the sub-step of step S10 in the abdominal ultrasound examination method shown in FIG. 10;

[0063] Figure 3 For Figure 1 Flowchart of the sub-step of step S20 in the abdominal ultrasound examination method shown in FIG. 11;

[0064] Figure 4 For Figure 1 Flowchart of the sub-step of step S30 in the abdominal ultrasound examination method shown in FIG. 12;

[0065] Figure 5 For Figure 1 Flowchart of the actual application of step S30 in the abdominal ultrasound examination method shown in FIG. 13;

[0066] Figure 6 For Figure 1 sub-step flowchart of step S40 in the abdominal ultrasound examination method shown in FIG. 4;

[0067] Figure 7 For Figure 1 sub-step flowchart of step S50 in the abdominal ultrasound examination method shown in FIG. 5;

[0068] Figure 8 Flowchart of the abdominal ultrasound examination method provided by the present disclosure in actual application;

[0069] Figure 9 Principle block diagram of the abdominal ultrasound examination system provided by the present disclosure;

[0070] Figure 10 Principle block diagram of the electronic device provided by the present disclosure. DETAILED DESCRIPTION

[0071] The technical solutions in the embodiments of the present disclosure will be clearly and completely described in the embodiments of the present disclosure combined with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present disclosure, rather than all the embodiments. The components of the embodiments of the present disclosure described and shown in the accompanying drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present disclosure provided in the accompanying drawings is not intended to limit the scope of the claimed present disclosure, but only represents selected embodiments of the present disclosure. Based on the embodiments of the present disclosure, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present disclosure.

[0072] Please refer to Figure 1 which is a flowchart of the abdominal ultrasound examination method provided by the embodiments of the present disclosure. It should be noted that the method of the present disclosure is not limited to the order of the following steps, and in other embodiments, the method of the present disclosure can only include part of the following steps, or part of the steps can be deleted.

[0073] The abdominal ultrasound examination method provided by the present disclosure can be applied to an abdominal ultrasound examination system based on artificial intelligence assistance, realizing efficient and automatic identification and lesion detection of multiple organs, and being able to integrate segmentation and lesion identification models of multiple organs such as liver, gallbladder, pancreas, spleen, and kidney, and provide comprehensive and accurate auxiliary diagnosis support. The abdominal ultrasound examination method includes the following steps:

[0074] Step S10: acquiring abdominal ultrasound images and performing preprocessing to obtain effective region images. Specifically, first, an abdominal ultrasound image video stream is acquired and preprocessed to extract an effective imaging region.

[0075] In some embodiments, please refer to Figure 2 , step S10 includes the following sub-steps:

[0076] S11, acquire the abdominal ultrasound video signal output by the endoscope device, convert the abdominal ultrasound video signal into a sequence of digital image frames through a video decoder, and output frame-by-frame RGB format images;

[0077] S12, perform fan-shaped valid area detection on the frame-by-frame RGB format images to obtain frame-by-frame preprocessed valid area images.

[0078] Specifically, first, the input image is subjected to grayscale processing, and the grayscale value is calculated by the weighted average method:

[0079] G(x,y)=0.299×R(x,y)+0.587×G(x,y)+0.114×B(x,y);

[0080] Where (x,y) is the coordinate of the pixel in the image, R(x,y), G(x,y), and B(x,y) are the intensity values of the red, green, and blue channels of the pixel at coordinate (x,y), respectively, and 0.299, 0.587, and 0.114 are the weight coefficients set according to the color sensitivity characteristics of the human eye;

[0081] Then, the radial gradient detection algorithm is used to identify the fan-shaped boundary, and the radial gradient intensity of each pixel point is calculated:

[0082] ;

[0083] Where r is the polar coordinate radius, θ is the polar coordinate angle, I is the grayscale value G(x,y), represents the rate of change of the grayscale value I in the r direction, and I(r+1,θ)-I(r,θ) is the difference between the grayscale values of the adjacent two radius pixels at the same angle θ;

[0084] Next, the center position and boundary arc line of the fan are detected by Hough transform, and the cumulative function is:

[0085] ;

[0086] Where (a,b) is the candidate point of the center coordinate, R is the candidate value of the circle radius, and δ is the Dirac function;

[0087] Finally, according to the detected fan radial gradient intensity, the circle center coordinates, the fan parameters of the circle radius, a binary mask matrix is generated, the pixel values in the fan area are retained, the pixel values outside the area are set to zero, and the preprocessed effective area image is output. That is, the center (a, b) and the radius R of the fan are found by the circle Hough transform, there can be two radii corresponding to the inner arc and the outer arc, after the center is found, the polar coordinates can be converted, the starting angle and the ending angle of the fan are determined by analyzing the gray scale change or the gradient direction on a specific radius, so as to completely define a fan, and the effective area is accurately extracted.

[0088] The above steps realize the preprocessing of the ultrasonic video signal output by the endoscope device, can extract the effective imaging area, extract the image area targeted by the abdominal ultrasound examination, avoid the interference of other invalid areas, the effective imaging area extraction precision is high, and the imaging area is beneficial to subsequent organ recognition, segmentation and the like.

[0089] Step S20: inputting the effective area image into a pre-trained standard section recognition network to obtain a section type recognition result. Specifically, the pre-trained standard section recognition neural network is used, the input is the frame-by-frame preprocessed image, the neural network is forward operated and inferred to obtain the section type recognition result.

[0090] In some embodiments, referring to Figure 3 , step S20 includes the following substeps:

[0091] S21, constructing an initial standard section recognition network;

[0092] S22, constructing a data set according to a preset standard section category, training the standard section recognition network to obtain a pre-trained standard section recognition network; specifically, the standard section recognition neural network is trained according to the preset standard section category to obtain the pre-trained standard section recognition neural network. The cross-entropy loss function is used for network training, the network parameters are iteratively updated through the back propagation algorithm, until the loss function converges or the preset iteration number is reached, and finally the pre-trained standard section recognition neural network is obtained;

[0093] S23, inputting the preprocessed effective area image into the pre-trained standard section recognition network, obtaining the corresponding section type recognition result through forward propagation calculation, and the section type recognition result includes a section type identifier and a confidence score.

[0094] Specifically, S23 includes:

[0095] Based on the effective area image, multi-scale feature maps of different scales are extracted step by step; specifically, the preprocessed image is input into a multi-scale feature extraction network, which internally contains five groups of cascaded feature extraction units, each group of units is composed of three 3x3 convolution layers, batch normalization layers and activation function layers in series, and feature maps of different scales are extracted step by step, and multi-scale feature maps of 1 / 2, 1 / 4, 1 / 8, 1 / 16 and 1 / 32 resolution are obtained respectively.

[0096] Based on the multi-scale feature maps, channel attention weights are obtained through global average pooling and global maximum pooling operations:

[0097] Att(F) = σ(MLP(AvgPool(F)) + MLP(MaxPool(F)));

[0098] Wherein, σ is a Sigmoid activation function, MLP is a multi-layer perceptron, F is a feature map, AvgPool(F) represents averaging all pixel values of each channel of the feature map F, and MaxPool(F) represents taking the maximum value of all pixel values of each channel of the feature map F; the channel attention weights are multiplied with the feature map element by element, and a fusion feature map is output.

[0099] Based on the fusion feature map, it is determined whether it is a standard section and a section type. Specifically, the fusion feature map is input into a section type classification unit, and a section type probability distribution vector is output through two fully connected layers and a Softmax activation function, to determine whether it is a standard section and a specific section type (such as gallbladder long axis section, extrahepatic bile duct section, etc.).

[0100] Understandably, the pre-set standard section categories include one or more of transabdominal aorta longitudinal section, transinferior vena cava longitudinal section, transright anterior lobe branch of portal vein and gallbladder neck right intercostal oblique section, transsecond hepatic portal rib margin oblique section, liver-kidney longitudinal section, portal vein left branch sagittal section, gallbladder long axis section, gallbladder short axis section, hepatic portal extrahepatic bile duct section, pancreas long axis section, pancreatic head short axis section, spleen intercostal anterior oblique coronal section, and non-standard section frame. In actual application, the implementer can adjust the pre-set standard section categories according to the specific situation, and the present disclosure is not limited specifically.

[0101] The above steps realize a method of using standard section recognition as a starting point for multi-organ continuous recognition, which improves the recognition accuracy under non-standard sections.

[0102] Step S30: when the slice type recognition result is a standard slice, performing organ segmentation to obtain an organ segmentation result of the current frame, continuously tracking the organ segmentation result of the current frame to obtain an organ segmentation result of a subsequent frame until the organ target disappears, and outputting frame-by-frame organ segmentation mask images to obtain a final organ segmentation result. Specifically, the preprocessed image sequence and the standard slice type recognition result are input into an organ segmentation module, a segmentation or tracking algorithm is intelligently scheduled according to a system state, and frame-by-frame organ segmentation mask images are efficiently output.

[0103] In some embodiments, referring to Figure 4 , step S30 includes the following sub-steps:

[0104] S31, based on the standard slice type recognition result of the current frame and the processing state of the previous frame, selecting a processing mode of the current frame according to a preset state transition rule, the processing mode including a segmentation mode, a tracking mode. Specifically, the processing mode also includes a standby mode, and the state decision rule includes: if the current frame is a standard slice and the previous frame is a non-standard slice or a different type of standard slice, triggering the segmentation mode; if the current frame is a standard slice and the previous frame is a same type of standard slice, and the tracking confidence score > 0.8, continue to maintain the tracking mode; if the current frame is a non-standard slice and the tracking confidence score > 0.6, continue to maintain the tracking mode; if the tracking confidence score < 0.6, reset to the standby mode; if one standard slice changes to another standard slice, trigger the segmentation mode;

[0105] S32, when the segmentation mode is triggered, inputting the effective region image after preprocessing of the current frame into a pre-trained segmentation network, performing forward operation inference using the pre-trained segmentation network, and obtaining an organ segmentation mask image of the current frame.

[0106] Specifically, the pre-trained organ segmentation neural network is obtained through the following steps:

[0107] First, a feature encoding module is constructed, which includes four groups of dense feature convolution calculation modules connected in series, each module including: a multi-scale convolution branch, which uses 1x1, 3x3 and 5x5 convolution kernels to extract features in parallel; a feature re-calibration layer, which adjusts feature weights through inter-channel correlation learning:

[0108] W = σ (FC2 (ReLU (FC1 (GAP (F))) ) ) ;

[0109] F' = F x W;

[0110] Wherein, F is the input feature map, GAP is the global average pooling, FC1 is the first fully connected layer, ReLU is the activation function, FC2 is the second fully connected layer, sigma is the Sigmoid function, F' is the re-calibrated feature map; 1 set of residual connection structure, retaining low-level feature information;

[0111] Then, 4 groups of decoding paths use the progressive feature restoration module, which includes: 1 group of bilinear interpolation up-sampling to restore the spatial resolution; 1 group of boundary enhancement convolution to strengthen the organ edge features: Edge = Conv(Sobel_x(F) + Sobel_y(F)), wherein Sobel_x(F), Sobel_y(F) respectively represent the X, Y direction convolution kernel of the Sobel operator applied to the image F, and Conv represents the convolution operation; 1 group of multi-level feature fusion to splice the encoding features and decoding features to obtain the decoding fusion map;

[0112] Finally, the decoding fusion map output by the last group is up-sampled to the original image size, and the organ segmentation mask map is output.

[0113] S33, when receiving the tracking mode, input the organ segmentation mask map of the previous frame and the effective region image after preprocessing of the current frame into the tracking network to calculate the organ mask map and the confidence score of the current frame.

[0114] Specifically, first, extract the feature descriptors from the previous frame segmentation mask region and the current frame image respectively, and calculate the region similarity matrix:

[0115] S(i,j) = exp(-||Fi - Fj||² / σ²);

[0116] Wherein, Fi is the i-th region feature of the previous frame, Fj is the j-th candidate region feature of the current frame, and sigma is the standard deviation parameter.

[0117] Then, based on the matching region pair with the highest similarity, estimate the local deformation parameters; that is, first extract the contour key points of the matching region, and calculate the corresponding deformation field of the frame through thin plate spline interpolation:

[0118] f(x,y) = a0 + a1x + a2y + Σ wi×K(||(x,y) - (xi,yi)||);

[0119] Wherein K is the radial basis function, wi is the weight coefficient, a0, a1, a2 are linear parameters.

[0120] Then, map the mask of the previous frame to the current frame according to the deformation field; that is, for each pixel point (x, y) in the mask, calculate the new position (x', y') according to the deformation field, generate the deformed mask using bilinear interpolation, and obtain the organ segmentation mask map of the current frame.

[0121] Recalculate the tracking quality score, i.e. the confidence score:

[0122] Conf = λ1×(mask region feature similarity) + λ2×(boundary continuity) + λ3×(shape consistency);

[0123] Wherein, λ1, λ2, λ3 are the weights of each item, which can be set according to actual needs;

[0124] Finally, output the dirty organ mask and the confidence score obtained by tracking.

[0125] S34, based on the standard section type, the segmentation or tracking of the organ mask map is reasonable to check, filter out the abnormal organ type, get the final organ segmentation result. Specifically, the initial mask obtained by segmentation or tracking and the current standard section type are input into the anatomical prior filtering unit, and the segmentation result is checked for rationality according to the standard section type. Specifically, it includes: constructing a section-organ correspondence table, such as the long axis section of the gallbladder only contains gallbladder and liver; according to the current section type, the effective channel is screened from the multi-channel mask output by segmentation or tracking; detect abnormal segmentation results, if there are organ categories that should not exist, set the corresponding channel to zero; morphological post-processing, remove discrete areas smaller than the threshold area: Area_min = 0.001 × W × H, finally output the filtered accurate organ segmentation mask; time sequence integration of segmentation result, input is the filtered mask of current frame, maintain the segmentation result cache queue, store the segmentation mask of the last K frames (K = 10); when the processing mode is switched, smooth transition processing is performed to avoid sudden changes in the segmentation result; output the final organ segmentation result of the current frame, the organ segmentation result contains the binary mask and the category identification of each organ.

[0126] Please refer to Figure 5, which is a flowchart when step S30 is actually applied. Specifically, when the standard cross-section recognition result obtained in step S20 is a standard cross-section, the organ segmentation unit is dispatched to perform image target (organ) segmentation, and the segmentation result is post-processed by the anatomical prior filtering unit according to the standard cross-section recognition result; after obtaining the segmentation result, the subsequent video continuation frame continues to use the segmentation algorithm and the prior filtering unit to continuously track the segmentation target. Its beneficial effects are: the traditional scheme uses a lightweight segmentation algorithm for continuous prediction to achieve real-time segmentation. The innovation of this scheme is that a larger and more complex segmentation network can be used to obtain the most reliable s-th frame (i.e. the current frame) recognition result, and then a lightweight tracking algorithm is enabled to continuously track the subsequent frame (s+1, s+2, s+3...s+n) result, and the recognition result is further processed according to the prior filtering, which ensures the accuracy and real-time performance. This is not an enlightenment brought by the traditional technical scheme, but a technical scheme design derived from deep thinking of the ultrasonic examination scene, patient condition, and examination method.

[0127] Step S40: dynamically separate a plurality of organ layers of organ types according to the organ segmentation result, dispatch the lesion recognition network corresponding to the organ type to perform reasoning, and obtain the lesion detection result of the organ in the plurality of organ layers. Specifically, the organ layers are dynamically separated according to the organ segmentation result, and the corresponding lesion recognition network is intelligently dispatched to perform reasoning, and the lesion detection frame and segmentation mask are output.

[0128] In some embodiments, referring to Figure 6 Step S40 includes the following sub-steps:

[0129] S41, based on the standard cross-section type and the organ segmentation result, separate the layers and output independent sub-images of each organ. Specifically, first, perform organ channel analysis and traverse the valid channels in the multi-channel mask; then, extract the organ area, calculate the minimum bounding rectangle for each valid organ mask; then, generate organ sub-images, crop the original image according to the bounding rectangle and appropriately expand the boundary; finally, output the organ type identifier and the corresponding ROI sub-image set.

[0130] S42, according to the current independent sub-image scheduling pre-constructed corresponding organ lesion recognition network, a plurality of organ lesion recognition results are output through inference calculation. Specifically, a pre-set corresponding organ lesion recognition neural network is constructed, including gallbladder, extrahepatic bile duct, liver, intrahepatic bile duct, pancreas, spleen, etc. According to the current true corresponding ROI image set, the corresponding organ lesion recognition neural network is scheduled, and the lesion recognition result is obtained through inference calculation. The pre-constructed corresponding organ lesion recognition neural network is obtained by training the pre-set organ lesion data set. In this embodiment, YOLOV10 network is used, and the training is performed (such as gallbladder lesion recognition neural network, liver lesion recognition neural network, etc.). According to the ROI organ sub-image type, the corresponding organ lesion recognition neural network is called, the neural network forward inference is executed, the lesion detection result is predicted, and a plurality of organ lesion recognition results are obtained.

[0131] Step S50: based on the final organ segmentation result, the lesion detection result and the abdominal ultrasound image, the size of a plurality of organs and lesions on the organs is calculated, and an abdominal ultrasound examination data statistical result is generated. Specifically, the organ segmentation result, the lesion detection result and the original ultrasound image are input, the pixel and physical size mapping relationship is established by recognizing the image side scale line, the target length and short diameter recognition algorithm is used, and the pixel and physical size corresponding relationship is combined, the long and short diameters of the organ and the lesion are output through calculation.

[0132] In some embodiments, referring to Figure 7 , step S50 includes the following sub-steps:

[0133] S51, recognizing the side scale line of the abdominal ultrasound image, and calculating the scale interval;

[0134] S52, counting the pixel distance of adjacent scale lines, and calculating the single-pixel interval based on the scale interval and the pixel distance;

[0135] S53, establishing a pixel and physical size mapping relationship based on the single-pixel interval and the physical size of the organ and the lesion on the organ on the abdominal ultrasound image;

[0136] Specifically, in steps S51-S53, the original ultrasound image is input. First, the edge region extraction is performed, and the non-imaging region around the image is located through the boundary detection algorithm. The specific formula is as follows:

[0137] Edge_mask = (I < threshold) & (gradient(I) < ε)

[0138] Where threshold is the gray threshold, and ε is the gradient threshold.

[0139] Then, the scale line feature detection is performed, the Hough transform is used to detect parallel line segments in the edge area, vertical scale lines (θ≈0° or 180°) and horizontal scale lines (θ≈90° or 270°) are detected;

[0140] Subsequently, the scale interval calculation is performed, the pixel distance of adjacent scale lines is counted, the main interval value is obtained through histogram statistics, and the single grid pixel interval is output;

[0141] Finally, the physical size corresponding to the current scale is obtained from the ultrasonic device, the single grid pixel interval is combined, the mapping coefficient Scale is calculated, and the pixel and physical size mapping relationship is obtained;

[0142] Scale = L_physical / d_pixel

[0143] Where L_physical is the single grid physical length (mm), and d_pixel is the single grid pixel number;

[0144] S54, according to the pixel and physical size mapping relationship, the long diameter and short diameter of a plurality of organs and lesions on the organs are calculated; specifically, the organ / lesion segmentation mask and the conversion coefficient are input, and the contour extraction and smoothing are performed using the opencv preset function to extract the target boundary from the segmentation mask:

[0145] Contour = findContour(Mask);

[0146] Contour_smooth = GaussianFilter(Contour, σ=2);

[0147] The minimum circumscribed ellipse fitting is performed, the least square method is used to fit the ellipse equation, and the major axis parameters a_major and a_minor of the ellipse are extracted;

[0148] The physical size conversion is performed to obtain the measurement result:

[0149] L_major_mm = a_major×Scale;

[0150] L_minor_mm = a_minor×Scale;

[0151] The long diameter and short diameter measurement values (unit: mm) of the organ / lesion are output;

[0152] S55. Draw the major and minor axes on the contours of organs and lesions on organs in abdominal ultrasound images, and label the values ​​of the major and minor axes at their endpoints, respectively. Output the labeled image and measurement data report. Specifically, first, generate image annotations, draw major and minor axes on the contours of organs / lesions, label the measurement values ​​at the endpoints of the axes, and use different colors to distinguish between organ (e.g., green) and lesion (e.g., red) measurements; generate a measurement report and output structured data.

[0153] Report = {"organ": {"type": "gallbladder", "long_axis": 85.3, "short_axis":32.1},

[0154] "lesion": {"type": "placeholder", "long_axis": 12.5, "short_axis": 8.2}};

[0155] That is, the optimal view is obtained based on the maximum length of the organ and the maximum length of the lesion.

[0156] Finally, it outputs annotated images and measurement report data, enabling simultaneous processing of segmentation and lesion detection of multiple organs, providing comprehensive diagnostic support, and is suitable for complex abdominal ultrasound examination scenarios.

[0157] Please see Figure 8The flowchart of the abdominal ultrasound examination method provided by the present disclosure is actually applied, specifically, the video reading and preprocessing module extracts the original ultrasound image, and the effective sector region is obtained after cropping, the effective region image is input into the pre-trained standard section recognition network, the standard section recognition module judges whether it is a standard section, if yes, the organ segmentation module performs organ segmentation, and the organ segmentation result is, for example, liver / bile / pancreas, etc., at the same time, the video continues, and the organ tracking module performs tracking, if it is judged that it is not a standard section, or if it is not tracked, it returns to the standard section recognition module, the occupying lesion scheduling recognition module identifies lesions of a plurality of organs based on the segmentation result, wherein each organ has a corresponding lesion recognition module, the segmentation sub-image of each organ is separated based on the segmentation result, for example, including gallbladder segmentation sub-image, extrahepatic bile duct segmentation sub-image, liver segmentation sub-image, intrahepatic bile duct segmentation sub-image, pancreas segmentation sub-image, and spleen segmentation sub-image, etc., and then the corresponding lesion recognition module is scheduled for lesion recognition, for example, the gallbladder segmentation sub-image is input into the gallbladder occupying lesion recognition module, the extrahepatic bile duct segmentation sub-image is input into the extrahepatic bile duct occupying lesion recognition module, the liver segmentation sub-image is input into the liver occupying lesion recognition module, the intrahepatic bile duct segmentation sub-image is input into the intrahepatic bile duct occupying lesion recognition module, the pancreas segmentation sub-image is input into the pancreas occupying lesion recognition module, and the spleen segmentation sub-image is input into the spleen occupying lesion recognition module, etc., to obtain the lesion detection result, that is, the lesion condition on a plurality of types of organs, and finally the axis measurement module obtains the optimal view according to the occupying lesion measurement maximum and obtains the optimal view according to the organ measurement maximum, calculates the size of a plurality of organs and lesions on the organs, and performs size labeling on the abdominal ultrasound image to generate an abdominal examination report, that is, the data statistical result of the abdominal ultrasound examination, which can provide comprehensive and accurate auxiliary diagnosis support.

[0158] The abdominal ultrasound examination method provided by the present disclosure is a multi-organ continuous segmentation and recognition method based on a standard section, which uses standard section recognition as a starting point to perform multi-organ continuous segmentation and recognition, and can improve the recognition accuracy under a non-standard section. The present disclosure adopts a lesion recognition scheduling mechanism driven by organ segmentation results, automatically identifies the organ type according to the organ segmentation results, and schedules the corresponding occupying lesion recognition model, realizes the mechanism of targeted lesion detection, and improves the detection accuracy. The present disclosure is a multi-organ comprehensive artificial intelligence auxiliary diagnosis system, which integrates a comprehensive AI auxiliary diagnosis system of standard section recognition, organ segmentation, organ recognition and scheduling, lesion recognition and result display, and is suitable for abdominal ultrasound examination.

[0159] The disclosure significantly reduces misidentification and missed detection under non-standard sections, improves the accuracy of organ identification, and improves the precision of lesion detection through a standard section-guided segmentation method. The lesion recognition scheduling mechanism driven by the organ segmentation result realizes targeted and accurate lesion detection, and improves the accuracy of overall diagnosis. The disclosure enhances the real-time performance and stability of the system, optimizes the algorithm and model design, ensures real-time response and stable operation of the system in clinical application, meets the actual operation requirements, and automatically identifies organs and detects lesions to assist doctors in quickly completing ultrasound examination and preliminary diagnosis, which can reduce the workload of doctors and improve work efficiency. Support for comprehensive analysis of multiple organs, the system can simultaneously process segmentation and lesion detection of multiple organs, provide comprehensive diagnostic support, and be suitable for complex abdominal ultrasound examination scenarios.

[0160] Understandably, several deep learning modules in the disclosure can automatically schedule the switches of various algorithms through reasonable logic, for example, intelligently scheduling a segmentation network or a tracking network based on an effective area image and a section type identification result, intelligently scheduling a lesion recognition network corresponding to an organ type according to an organ segmentation result, and the like. Reasonable allocation of computing power can be achieved, so that the number of algorithms simultaneously turned on during the entire examination process does not exceed a preset number (for example, 3), a relatively small and stable computing power can be used, and the screening of multiple organs and the diagnosis of multiple diseases can be solved.

[0161] In addition, in the examination and diagnosis scene, doctors often lack scanning order and completeness specifications, and there is no mandatory procedure for accurate positioning and then finding lesions and measuring. Misdiagnosis and missed diagnosis of lesions often occur. However, the disclosure uses the innovative logic of the decision tree model, the decision tree combines accurate positioning of the section, organ range ruling, lesion recognition and diagnosis sign sample ruling, and strict selection of the measurement axis, which not only conforms to the use habits of doctors, but also implements a pre-navigation and quality control for the entire diagnosis and treatment process, effectively managing the diagnosis and treatment behavior of doctors. The disclosure provides a better demonstration for the research and development of medical imaging artificial intelligence, not only looking at pictures but also looking at videos, not only looking at videos but also disassembling videos, multi-semantic splitting, not only serving doctors but also communicating with doctors through semantics, and providing timely reminders and help for doctors.

[0162] In summary, the abdominal ultrasound examination method provided by the disclosure discloses an overall idea of artificial intelligence-assisted abdominal ultrasound examination, which can improve algorithm efficiency, recognition efficiency, reduce misidentification and missed detection, use standard section classification identification as a technical starting point, combine organ segmentation results, and schedule the corresponding space-occupying lesion recognition model to improve the precision of lesion detection, realize efficient and automatic identification and lesion detection of multiple organs, and integrate segmentation and lesion recognition models of multiple organs such as liver, gallbladder, pancreas, spleen, and kidney to provide comprehensive and accurate auxiliary diagnosis support.

[0163] Referring to Figure 9 FIG. 1 is a schematic diagram of an abdominal ultrasound examination system according to an embodiment of the present disclosure. The abdominal ultrasound examination system 100 according to the present disclosure comprises a preprocessing module 10, a standard section identification module 20, an organ segmentation module 30, a lesion identification module 40, and an axis measurement module 50.

[0164] The preprocessing module 10 is configured to acquire an abdominal ultrasound image and perform preprocessing to obtain an effective region image.

[0165] The standard section identification module 20 is configured to input the effective region image into a pre-trained standard section identification network to obtain a section type identification result.

[0166] The organ segmentation module 30 is configured to perform organ segmentation to obtain an organ segmentation result of a current frame when the section type identification result is a standard section, continuously track the organ segmentation result of the current frame to obtain an organ segmentation result of a subsequent frame until an organ target disappears, and output a frame-by-frame organ segmentation mask image to obtain a final organ segmentation result.

[0167] The lesion identification module 40 is configured to dynamically separate a plurality of organ layers of different organ types according to the organ segmentation result, schedule a lesion identification network corresponding to the organ type to perform inference, and obtain a lesion detection result of the organ in the plurality of organ layers.

[0168] The axis measurement module 50 is configured to calculate the size of a plurality of organs and lesions on the organs based on the final organ segmentation result, the lesion detection result, and the abdominal ultrasound image, and generate an abdominal ultrasound examination data statistical result.

[0169] It can be understood that each of the above functional modules can be stored in the memory in the form of a software program and executed by the processor. In alternative embodiments, each of the above functional modules can be hardware with specific functions, such as a chip with a specific software program burned in.

[0170] It should be noted that in implementation, the abdominal ultrasound examination method can be implemented by the above-mentioned abdominal ultrasound examination system 100, and the abdominal ultrasound examination system 100 can implement the abdominal ultrasound examination of efficient and automatic identification and lesion detection of multiple organs by using one or more specific embodiments of the abdominal ultrasound examination method described in the above embodiments. That is, all embodiments of the abdominal ultrasound examination method provided by the present disclosure are applicable to the abdominal ultrasound examination system 100 provided by the present disclosure, and can achieve the same or similar beneficial effects, which will not be described here.

[0171] Referring to Figure 10 The present disclosure also provides an electronic device, comprising a memory 210 and one or more processors 220.

[0172] Specifically, the memory 210 is configured to store one or more computer programs, and the one or more computer programs, when executed by the one or more processors 220, implement the abdominal ultrasound examination method according to any one of the preceding embodiments.

[0173] The memory 210 can be, but is not limited to, a random access memory (RAM), a read only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), etc. The memory 210 is configured to store programs, and the processor 220 executes the programs after receiving an execution instruction, so as to implement the abdominal ultrasound examination method according to any one of the preceding embodiments. It can be understood that the processor 220 and other possible components access the memory 210 under the control of a memory controller.

[0174] The processor 220 can be an integrated circuit chip having a signal processing capability. The processor 220 can be a general purpose processor, including a central processing unit (CPU), a network processor (NP), etc., and can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, and can implement or execute the methods, steps and logical blocks disclosed in any one of the embodiments of the present disclosure.

[0175] The present disclosure further provides a computer storage medium, and the computer storage medium stores a computer program. The computer program, when executed by a processor, implements the abdominal ultrasound examination method according to any one of the preceding embodiments.

[0176] The present disclosure further provides a computer program product, and the computer program product includes a computer program or instructions. The computer program or instructions, when executed by a processor, implement the abdominal ultrasound examination method according to any one of the preceding embodiments.

[0177] In summary, the abdominal ultrasound examination method, system and device provided by the present disclosure utilize standard section classification recognition as a basis, combine organ segmentation results, and dispatch corresponding space-occupying lesion recognition models, thereby improving the accuracy of lesion detection, achieving efficient and automated recognition and lesion detection of multiple organs, integrating segmentation and lesion recognition models of multiple organs such as the liver, gallbladder, pancreas, spleen, and kidneys, and achieving high organ recognition and lesion detection accuracy, thereby avoiding misrecognition and missed detection, improving examination efficiency, and providing comprehensive and accurate diagnostic support for abdominal ultrasound examination.

[0178] The above merely describes the embodiments of the present disclosure and does not limit the patent scope of the present disclosure. Any equivalent structure or equivalent process transformation using the content of the present disclosure specification and drawings, or direct or indirect application in other related technical fields, is also included in the patent protection scope of the present disclosure.

Claims

1. A method of abdominal ultrasound examination, characterized in that, The method comprises the following steps: acquiring an abdominal ultrasound image and pre-processing to obtain an effective region image; inputting the effective region image into a pre-trained standard section recognition network to obtain a section type recognition result; specifically comprising: constructing an initial standard section recognition network; constructing a data set according to a preset standard section category, training the standard section recognition network to obtain a pre-trained standard section recognition network; inputting the pre-processed effective region image into the pre-trained standard section recognition network to obtain a corresponding section type recognition result through forward propagation calculation, wherein the section type recognition result comprises a section type identifier and a confidence score; wherein the step of inputting the pre-processed effective region image into the pre-trained standard section recognition network to obtain the corresponding section type recognition result through forward propagation calculation comprises: extracting multi-scale feature maps of different scales based on the effective region image; obtaining channel attention weights based on the multi-scale feature maps through global average pooling and global maximum pooling operations: Att(F) = σ(MLP(AvgPool(F)) + MLP(MaxPool(F))) wherein σ is a Sigmoid activation function, MLP is a multi-layer perceptron, F is a feature map, AvgPool(F) represents averaging all pixel values of each channel of the feature map F, and MaxPool(F) represents taking the maximum value of all pixel values of each channel of the feature map F; multiplying the channel attention weights and the feature map element by element to output a fusion feature map; determining whether it is a standard section and a section type based on the fusion feature map; when the section type recognition result is a standard section, performing organ segmentation to obtain an organ segmentation result of the current frame, continuously tracking the organ segmentation result of the current frame to obtain an organ segmentation result of a subsequent frame until the organ target disappears, and outputting a frame-by-frame organ segmentation mask image to obtain a final organ segmentation result; specifically comprising: based on the standard section type recognition result of the current frame and the processing state of the previous frame, selecting a processing mode of the current frame according to a preset state transition rule, wherein the processing mode comprises a segmentation mode and a tracking mode; when the segmentation mode is triggered, inputting the pre-processed effective region image of the current frame into a pre-trained segmentation network, performing forward operation inference using the pre-trained segmentation network to obtain an organ segmentation mask image of the current frame; when the tracking mode is received, inputting the organ segmentation mask image of the previous frame and the pre-processed effective region image of the current frame into a tracking network to obtain an organ segmentation mask image of the current frame and a confidence score; based on the standard section type, performing reasonableness verification on the organ segmentation mask image obtained by segmentation or tracking to filter out abnormal organ types to obtain a final organ segmentation result; dynamically separating a plurality of organ layers of organ types according to the final organ segmentation result, scheduling a lesion recognition network corresponding to the organ type to perform inference to obtain lesion detection results of organs in the plurality of organ layers; Based on the final organ segmentation result, the lesion detection result and the abdominal ultrasound image, the sizes of the organs and the lesions on the organs are calculated, and an abdominal ultrasound examination data statistical result is generated.

2. The abdominal ultrasound examination method of claim 1, wherein, The step of obtaining the abdominal ultrasound image and pre-processing to obtain the effective region image comprises: An abdominal ultrasound video signal output by the endoscopic device is converted into a sequence of digital image frames by a video decoder, and an RGB format image frame by frame is output. The step of performing fan-shaped effective region detection on the RGB format image frame by frame to obtain the pre-processed effective region image comprises:

3. The abdominal ultrasound examination method of claim 2, wherein, The input RGB format image is subjected to grayscale processing, and the grayscale value is calculated by a weighted average method: G(x,y)=0.299×R(x,y)+0.587×G(x,y)+0.114×B(x,y); where (x,y) is the coordinate of a pixel in the image, R(x,y), G(x,y) and B(x,y) are the intensity values of the red, green and blue channels of the pixel at the coordinate (x,y), and 0.299, 0.587 and 0.114 are the weight coefficients set according to the color sensitivity characteristics of the human eye; A radial gradient detection algorithm is used to identify the fan-shaped boundary, and the radial gradient intensity of each pixel point is calculated: The center position and boundary arc line of the fan are detected by Hough transformation, and the cumulative function is: ； where r is a polar coordinate radius, θ is a polar coordinate angle, I is a gray value G(x, y), represents a rate of change of the gray value I in the r direction, I(r + 1, θ) - I(r, θ) is a difference value of the pixel gray values of two adjacent radii at the same angle θ; where (a,b) is the center coordinate candidate point, R is the candidate value of the circle radius, and δ is the Dirac function; ； According to the detected fan-shaped parameters of the radial gradient intensity, the center coordinate and the circle radius, a binary mask matrix is generated, the pixel values in the fan-shaped region are retained, the pixel values outside the fan-shaped region are set to zero, and the pre-processed effective region image is output. The step of dynamically separating organ layers of several organ types according to the final organ segmentation result, scheduling a lesion recognition network corresponding to the organ type for inference, and obtaining the lesion detection result of the organ in the several organ layers comprises:

4. The abdominal ultrasound examination method of claim 1, wherein, Based on the standard section type and the organ segmentation result, the layer separation is performed, and independent sub-images of each organ are output. According to the current independent sub-image, a pre-constructed lesion recognition network corresponding to the organ is scheduled, and the lesion detection result of the several organs is output after inference calculation. The step of calculating the sizes of the organs and the lesions on the organs based on the final organ segmentation result, the lesion detection result and the abdominal ultrasound image, and generating an abdominal ultrasound examination data statistical result comprises:

5. The abdominal ultrasound examination method of claim 1, wherein, The side scale lines of the abdominal ultrasound image are recognized, and the scale interval is calculated. The pixel distance of adjacent scale lines is counted, and the single-pixel interval is calculated based on the scale interval and the pixel distance. A pixel-physical size mapping relationship is established based on the single-pixel interval and the physical size of the organs and the lesions on the organs on the abdominal ultrasound image. The long diameter and the short diameter of the several organs and the lesions on the organs are calculated according to the pixel-physical size mapping relationship. ​ Long and short axes are drawn on the contours of organs and lesions on organs in an abdominal ultrasound image, and the values of the long and short diameters are marked at the endpoints of the long and short axes, and an image after marking and a measurement data report are output.

6. An abdominal ultrasound examination system, characterized by The method comprises the following steps: a preprocessing module is configured to acquire an abdominal ultrasound image and perform preprocessing to obtain an effective region image; a standard section recognition module is configured to input the effective region image into a pre-trained standard section recognition network to obtain a section type recognition result; specifically, an initial standard section recognition network is constructed; a data set is constructed according to a preset standard section category, and the standard section recognition network is trained to obtain the pre-trained standard section recognition network; the preprocessed effective region image is input into the pre-trained standard section recognition network, and a corresponding section type recognition result is obtained through forward propagation calculation, wherein the section type recognition result comprises a section type identifier and a confidence score; wherein the preprocessed effective region image is input into the pre-trained standard section recognition network, and a corresponding section type recognition result is obtained through forward propagation calculation, comprising: extracting multi-scale feature maps of different scales based on the effective region image; obtaining channel attention weights based on the multi-scale feature maps through global average pooling and global maximum pooling operations: Att(F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F))) wherein σ is a Sigmoid activation function, MLP is a multi-layer perceptron, F is a feature map, AvgPool(F) represents averaging all pixel values of each channel of the feature map F, and MaxPool(F) represents taking the maximum value of all pixel values of each channel of the feature map F; the channel attention weights are multiplied with the feature map element by element to output a fusion feature map; determining whether it is a standard section and a section type based on the fusion feature map; an organ segmentation module is configured to perform organ segmentation when the section type recognition result is a standard section to obtain an organ segmentation result of a current frame, continuously track the organ segmentation result of the current frame to obtain an organ segmentation result of a subsequent frame until the organ target disappears, and output frame-by-frame organ segmentation mask images to obtain a final organ segmentation result; specifically, based on the standard section type recognition result of the current frame and the processing state of the previous frame, a processing mode of the current frame is selected according to a preset state transition rule, wherein the processing mode comprises a segmentation mode and a tracking mode; when the segmentation mode is triggered, the preprocessed effective region image of the current frame is input into a pre-trained segmentation network, and the pre-trained segmentation network is used for forward operation reasoning to obtain an organ segmentation mask image of the current frame; when the tracking mode is received, the organ segmentation mask image of the previous frame and the preprocessed effective region image of the current frame are input into a tracking network to obtain an organ segmentation mask image of the current frame and a confidence score; the organ segmentation mask image obtained by segmentation or tracking is reasonably checked based on the standard section type, and abnormal organ types are filtered out to obtain a final organ segmentation result. The lesion recognition module is configured to dynamically separate a plurality of organ layers of organ types according to the final organ segmentation result, and schedule a lesion recognition network corresponding to the organ type to perform inference to obtain a lesion detection result of the organ in the plurality of organ layers. The fixed-axis measurement module is configured to calculate sizes of the plurality of organs and lesions on the organs based on the final organ segmentation result, the lesion detection result and the abdominal ultrasound image, and generate an abdominal ultrasound examination data statistical result.

7. An electronic device, comprising: The memory and the one or more processors; The memory is configured to store one or more computer programs; and the one or more processors are configured to implement the abdominal ultrasound examination method according to any one of claims 1-5 when the one or more computer programs are executed.

Images