A method and system for intelligent recognition of ultrasound robots based on AI-powered large-scale ultrasound models.

Abstract

This invention relates to the field of ultrasound robot technology, and more particularly to an intelligent recognition method and system for ultrasound robots based on an AI-powered large-scale ultrasound model. The method includes the following steps: acquiring ultrasound video sequences collected during ultrasound robot scanning and labeling them to obtain raw training data; enhancing the raw training data using a difficult-example data augmentation strategy to obtain enhanced training samples, and then evaluating and filtering these samples through a teacher network to obtain augmented training data; constructing a student network and inputting the augmented training data into the student network for training to obtain a student network model; inputting the test ultrasound video sequence into the teacher network, determining the current target organ based on the multi-organ confidence scores output by the teacher network, switching to the corresponding organ's student network model for real-time segmentation and inference, and outputting the segmentation result. This invention achieves lightweight real-time segmentation and recognition in ultrasound robot scanning scenarios, effectively improving the output stability between adjacent frames.

Description

Technical Field

[0001] This invention relates to the field of ultrasonic robot technology, and in particular to an intelligent recognition method and system for ultrasonic robots based on an AI-based large-scale ultrasonic model. Background Technology

[0002] Ultrasound imaging, as a non-invasive, real-time medical imaging technology, is widely used in clinical diagnosis. However, ultrasound images often suffer from inherent problems such as temporal jitter, blurred boundaries, and organ deformation, leading to unclear organ structural boundaries and insufficient cross-frame stability. In recent years, with the rapid development of deep learning technology, neural network-based automatic ultrasound image segmentation methods have made significant progress in accuracy and robustness. However, most high-performance segmentation models have complex network structures and a large number of parameters, making them difficult to deploy in real time on computationally limited embedded devices or ultrasound robot platforms.

[0003] Knowledge distillation, as an effective model compression technique, can significantly reduce model complexity while maintaining high accuracy by transferring knowledge from complex teacher networks to lightweight student networks. Existing knowledge distillation methods primarily focus on the natural image domain, failing to adequately consider the unique properties of medical ultrasound images and the specific needs of ultrasound robot applications. Particularly in ultrasound video sequence processing, existing lightweight models are prone to issues such as boundary leakage, contour distortion, and inter-frame jitter, affecting the reliability and stability of clinical applications.

[0004] Chinese patent CN115908253A discloses a cross-domain medical image segmentation method based on knowledge distillation, comprising: acquiring training data, wherein the training data is labeled medical image data; designing a teacher network and a student network based on the training data; then acquiring the domain-invariant features of the teacher network and the student network respectively; designing a knowledge distillation algorithm based on the domain-invariant features; and finally training the student network again based on the training data using the teacher network and the knowledge distillation algorithm to obtain a target network, wherein the target network is used to achieve cross-domain medical image segmentation. This method extracts the domain-invariant features of the teacher network and the student network, and designs a double-contrast image distillation algorithm and a domain-invariant cross-distillation algorithm, achieving a lightweight student network with cross-domain segmentation capabilities across different imaging devices. This patent mainly addresses the domain difference problem between different imaging devices, improving the model's generalization through domain adaptation technology. However, this method does not fully consider the temporal continuity constraints of ultrasound video sequences, and still suffers from insufficient cross-frame stability and decreased boundary accuracy when handling dynamic scenes during ultrasound robot scanning. Summary of the Invention

[0005] In view of this, the present invention proposes an intelligent recognition method and system for ultrasonic robots based on an AI-based large ultrasonic model, in order to solve the problems of boundary leakage, contour distortion and inter-frame jitter in the deployment of lightweight ultrasonic segmentation models in embedded devices in the prior art.

[0006] The technical solution of this invention is implemented as follows: This invention provides an intelligent recognition method for ultrasound robots based on an AI-powered large-scale ultrasound model, comprising the following steps:

[0007] S1. Obtain the ultrasound video sequence collected during the ultrasound robot scanning process, and annotate the ultrasound video sequence through the teacher network to obtain the original training data with segmentation mask and confidence label.

[0008] S2. The original training data is augmented using a difficult example data augmentation strategy based on the physical characteristics of ultrasound to obtain augmented training samples. The augmented training samples are then evaluated and screened by the teacher network, and difficult example samples that reduce the confidence of the teacher network while maintaining the semantic structure are retained as augmented training data.

[0009] S3. Construct a student network, input the expanded training data into the student network for training, use the knowledge distillation strategy to transfer the knowledge of the teacher network to the student network, and use optical flow-guided temporal consistency constraints to regularize the adjacent frame outputs of the student network to obtain a lightweight student network model.

[0010] S4. Input the test ultrasound video sequence into the teacher network, determine the current target organ for scanning based on the confidence scores of multiple organs output by the teacher network, switch to the student network model of the corresponding organ for real-time segmentation inference, and output the segmentation results.

[0011] Based on the above technical solutions, preferably, in step S1, the teacher network annotates the ultrasound video sequence including:

[0012] The teacher network performs inference on the input ultrasound video frames and outputs multiple organ masks and corresponding logits.

[0013] Based on the current target organ being scanned, the corresponding target organ mask is selected as the foreground mask, and the masks of the remaining organs are merged as the background mask. Two-channel remapping is then performed to obtain binary labels for the foreground and background.

[0014] A morphological dilation operation is performed on the foreground mask to generate a region of interest (ROI), which is used for spatial gating in the subsequent distillation process.

[0015] Temperature softening is applied to the output logits of the teacher network to obtain the soft label probability distributions of the foreground and background, which are used as the supervision signal for output layer distillation.

[0016] The entropy value is calculated based on the probability distribution of the soft label, and the pixel-level uncertainty weight is obtained by normalizing the entropy value. This weight is then used to calculate the weighted distillation loss, thus completing the annotation of the frame.

[0017] Based on the above technical solutions, preferably, in step S2, the difficult case data augmentation strategy based on ultrasound physical characteristics includes boundary weak echo simulation, organ occlusion simulation, cut-in / departure simulation, and probe angle change simulation; wherein, boundary weak echo simulation is achieved by attenuating the pixel intensity of the organ boundary region, organ occlusion simulation is achieved by generating occlusion blocks in the organ region and filling them with acoustic artifacts or reverberation artifacts, cut-in / departure simulation is achieved by cropping the organ edge region, and probe angle change simulation is achieved by performing geometric deformation operations on the original image; for adjacent frames in the video sequence, a frame pair synchronization enhancement strategy is adopted, and the enhancement parameters of adjacent frames are set in association to maintain the continuity of inter-frame enhancement.

[0018] Based on the above technical solution, preferably, in step S2, the evaluation and screening of the enhanced training samples through the teacher network includes: inputting the enhanced image into the teacher network to obtain the enhanced confidence score and segmentation mask, calculating the enhancement effectiveness evaluation value, and retaining samples with evaluation values ​​higher than a preset threshold as expanded training data; the calculation formula for the enhancement effectiveness evaluation value is:

[0019] ;

[0020] Where the subscript t represents the teacher network, The confidence level of the target organ in the teacher network on the original image; To enhance the confidence of target organs in the teacher network on the image; The teacher segmentation mask corresponding to the original image; To enhance the teacher segmentation mask corresponding to the image; This is the power factor of the difficulty gain; The power coefficients for semantic preservation; The minimum confidence threshold; For indicator functions; For intersection, union, and comparison.

[0021] Based on the above technical solutions, preferably, step S3 specifically includes:

[0022] S31. Construct a student network, wherein the student network adopts a lightweight U-Net architecture;

[0023] S32. Input the expanded training data into the student network and train it using a knowledge distillation strategy to obtain the knowledge distillation loss. The knowledge distillation strategy includes output layer distillation, feature layer distillation, and negative sample suppression. Output layer distillation aligns the output probability distributions of the student network and the teacher network in the region of interest. Feature layer distillation aligns the feature representations of the student network and the teacher network in multiple coding layers. Negative sample suppression suppresses the foreground response of the student network in non-target organ regions.

[0024] S33. The student network is trained using optical flow-guided temporal consistency constraints to constrain the output consistency of the student network in adjacent frames, thus obtaining the temporal consistency loss.

[0025] S34. By jointly optimizing the knowledge distillation loss and the temporal consistency loss, a lightweight student network model is obtained.

[0026] Based on the above technical solutions, the preferred knowledge distillation loss includes output layer distillation loss, feature layer distillation loss, and negative sample suppression loss.

[0027] The output layer distillation loss includes: within the region of interest (ROI), the KL divergence between the output probability distributions of the student network and the teacher network is weighted using uncertainty weights to calculate the output layer distillation loss;

[0028] The feature layer distillation loss includes: setting adapters in multiple coding layers of the student network to map student network features to the same dimension as teacher network features; calculating the distance loss between student network features and teacher network features in the region of interest (ROI) and boundary regions; and weighting the boundary regions through boundary weights to calculate the feature layer distillation loss.

[0029] The negative sample suppression loss includes: aggregating the non-target organ masks output by the teacher network to obtain non-target organ regions, suppressing the foreground response of the student network within the non-target organ regions, and calculating the negative knowledge suppression loss.

[0030] Based on the above technical solution, preferably, in step S33, the temporal consistency loss includes student self-consistency loss and teacher guidance loss; wherein...

[0031] The student self-consistency loss includes: calculating the motion field between adjacent frames through an optical flow estimation algorithm, identifying occlusion regions through forward and backward optical flow consistency checks, mapping the student network's output in the previous frame to the current frame position through the motion field, and constraining the consistency between the student network's current frame output and the mapped output in non-occlusion regions to obtain the student self-consistency loss.

[0032] The teacher guidance loss includes: mapping the output of the teacher network in the previous frame to the current frame position through the motion field, and mapping the output of the teacher network in the current frame to the previous frame position through the inverse motion field. The KL divergence between the student network output and the mapped teacher soft label is constrained in the non-occluded area to obtain the teacher guidance loss.

[0033] Based on the above technical solutions, preferably, in step S34, the formula for calculating the total training loss of joint optimization is as follows:

[0034] ;

[0035] ;

[0036] ;

[0037] ;

[0038] ;

[0039] Where the subscript t represents the teacher network and the subscript s represents the student network. The task loss between the student network and the truth label; For output layer distillation loss, For pixel position, ROI gating weights For uncertain weights, For student networks in pixels The output probability distribution after softening at temperature T. For teacher networks in pixels The output probability distribution after softening at temperature T, where T represents the temperature softening coefficient. Let KL divergence be a metric. To mitigate loss through negative knowledge, For aggregation masks of non-target organ regions, Negative knowledge weight, The non-target organ inhibition weighting coefficient. Output the probability of the foreground of the student network; For characteristic layer distillation loss, For feature layer index, For student networks in the first Features at the layer pixel, For the teacher network in the first Features at the layer pixel, For the first Layer adapter, For boundary weights, This represents the square of the L2 norm, which is the square of the Euclidean distance. For time-series consistency loss, The weight of student self-consistency loss, Weighting of teacher-guided loss, For the students' own losses, This results in losses for teachers' guidance.

[0040] Based on the above technical solutions, preferably, step S4 includes:

[0041] S41. Input the test ultrasound video sequence into the teacher network and obtain the confidence scores of multiple organs output by the teacher network.

[0042] S42. Monitor the confidence level of the multiple organs in real time. When the confidence level of a certain organ exceeds the preset switching threshold in several consecutive frames, determine that the current scanning target is that organ.

[0043] S43. Based on the determined current target organ, switch to the corresponding student network model, and use a lag switching strategy to ensure that the confidence of the new organ remains stable before performing the switching operation.

[0044] S44. Perform segmentation processing on the student network model after inputting the test ultrasound video sequence and output the segmentation results.

[0045] This invention provides an intelligent recognition system for ultrasound robots based on an AI-powered large-scale ultrasound model, comprising:

[0046] Data acquisition and annotation module: used to acquire ultrasound video sequences collected during ultrasound robot scanning, and to annotate the ultrasound video sequences through the teacher network to obtain raw training data;

[0047] Data augmentation module: used to augment the original training data and evaluate and screen the augmented training samples through the teacher network to obtain augmented training data;

[0048] Knowledge distillation training module: used to construct student network, input expanded training data into the student network for training, use knowledge distillation strategy to transfer knowledge from teacher network to student network, and use optical flow-guided temporal consistency constraint to regularize the student network to obtain a lightweight student network model;

[0049] Intelligent recognition and segmentation module: It is used to input the test ultrasound video sequence into the teacher network, determine the current target organ for scanning based on the confidence scores of multiple organs output by the teacher network, switch to the student network model of the corresponding organ for segmentation processing, and output the segmentation results.

[0050] The intelligent recognition method and system for ultrasound robots based on AI large-scale ultrasound models of the present invention have the following advantages over existing technologies:

[0051] (1) This invention constructs a knowledge distillation architecture with a large multi-organ segmentation model as the teacher network and a lightweight U-Net as the student network. It combines a difficult example data expansion strategy based on the physical characteristics of ultrasound and a temporal consistency constraint guided by optical flow to achieve lightweight real-time segmentation and recognition in the ultrasound robot scanning scenario. While significantly reducing the number of model parameters and computational complexity, it maintains segmentation accuracy close to that of the teacher network and effectively improves the output stability between adjacent frames.

[0052] (2) The multi-level knowledge distillation strategy of the present invention, through the synergistic effect of output layer alignment, feature layer alignment and negative sample suppression, enables the student network to acquire intermediate feature representation ability and discrimination ability to suppress false activation of non-target organs while learning the probability distribution of the teacher network. Among them, the boundary-weighted feature distillation mechanism strengthens the detailed representation of organ contours, while the negative sample suppression mechanism reduces the erroneous response of non-target regions in multi-organ scanning scenarios, thereby improving the specialization and segmentation accuracy of the specialized model. At the same time, the optical flow-guided temporal consistency constraint introduced effectively suppresses the inter-frame jump and output jitter phenomenon commonly seen in lightweight models in video sequence processing through the dual constraints of student self-consistency term and teacher guidance term. In the non-occluded area, the optical flow field is used to align the prediction results of the previous and next frames, ensuring the smoothness and continuity of the segmentation mask in the time dimension and improving the visual stability of ultrasound video segmentation.

[0053] (3) The multi-organ intelligent switching mechanism of the present invention monitors the confidence of various organs in real time through the teacher network and combines the lag judgment strategy. It can automatically load the corresponding student network model according to the organ type in the field of view during the ultrasound robot scanning process, avoids the problem of balancing the segmentation accuracy of each organ caused by using a single general model, improves the segmentation performance of each model on the target organ, and avoids frequent model switching during probe movement through lag switching, ensuring the stability of system operation.

[0054] (4) The difficult example data augmentation strategy of the present invention covers typical difficult example scenarios in ultrasound imaging in a targeted manner through enhancement methods such as boundary weak echo simulation, organ occlusion simulation, cutting in and leaving simulation and probe angle change simulation. It also automatically selects high-value samples that moderately reduce the confidence of the teacher network but maintain the semantic structure through the enhancement effectiveness evaluation function, avoiding invalid or excessively distorted samples introduced by simple random enhancement, thereby improving the student network's generalization ability to challenging situations such as blurred boundaries and sound shadow occlusion in ultrasound images. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is a flowchart of the ultrasonic robot intelligent recognition method of the present invention;

[0057] Figure 2 This is an overall architecture diagram of the ultrasonic robot intelligent recognition system of the present invention;

[0058] Figure 3 This is a flowchart for expanding the difficult example data of the present invention. Detailed Implementation

[0059] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0060] like Figure 1 and Figure 2 As shown, this invention provides an intelligent recognition method for ultrasound robots based on an AI-powered large-scale ultrasound model, comprising the following steps:

[0061] S1. Obtain the ultrasound video sequence collected during the ultrasound robot scanning process, and annotate the ultrasound video sequence through the teacher network to obtain the original training data with segmentation mask and confidence label.

[0062] S2. The original training data is augmented using a difficult example data augmentation strategy based on the physical characteristics of ultrasound to obtain augmented training samples. The augmented training samples are then evaluated and screened by the teacher network, and difficult example samples that reduce the confidence of the teacher network while maintaining the semantic structure are retained as augmented training data.

[0063] S3. Construct a student network, input the expanded training data into the student network for training, use the knowledge distillation strategy to transfer the knowledge of the teacher network to the student network, and use optical flow-guided temporal consistency constraints to regularize the adjacent frame outputs of the student network to obtain a lightweight student network model.

[0064] S4. Input the test ultrasound video sequence into the teacher network, determine the current target organ for scanning based on the confidence scores of multiple organs output by the teacher network, switch to the student network model of the corresponding organ for real-time segmentation inference, and output the segmentation results.

[0065] In a specific embodiment of the present invention, step S1, in which the teacher network annotates the ultrasound video sequence, includes:

[0066] The teacher network infers from the input ultrasound video frames and outputs multi-organ masks and corresponding logits. The teacher network uses a pre-trained multi-organ segmentation model, MOFO, which is built on a deep convolutional neural network architecture and supports simultaneous segmentation and recognition of K different organ classes, where K is the total number of organ classes, which can be set to 5-15 classes depending on the application scenario, with a typical value of 8 classes. The teacher network takes a 640×480 pixel ultrasound image as input and outputs a K-channel segmentation mask. And the corresponding unnormalized classification score logits, where the superscripts 1, 2, ..., K represent channel indices and the subscript t represents the teacher network.

[0067] Based on the currently scanned target organ, the corresponding target organ mask is selected as the foreground mask, and the masks of the remaining organs are merged to form the background mask. Two-channel remapping is then performed to obtain binary classification labels for the foreground and background. Specifically, let the currently scanned target organ be... Then the foreground mask Background mask Obtained through two-channel remapping The binary label is used as a supervisory signal for the student network.

[0068] A morphological dilation operation is performed on the foreground mask to generate a region of interest (ROI) for spatial gating in the subsequent distillation process. The dilation operation uses circular or square structuring elements, with a dilation radius r ranging from 5 to 15 pixels, preferably 10 pixels, to maximize the ROI. It can cover the organ itself and its surrounding area to a certain extent, thus focusing on the region containing the target organ information when calculating the distillation loss, reducing the interference of irrelevant background on training.

[0069] Temperature softening is applied to the output logits of the teacher network to obtain the soft label probability distributions for the foreground and background, which are used as the supervision signal for output layer distillation. Temperature softening is achieved by scaling the logits with a temperature parameter T before applying the softmax function. , where T is the temperature coefficient. Higher temperature values ​​result in a smoother probability distribution, reduce overconfident predictions, and enhance the transitivity of relative relationships between categories, thus providing richer supervisory information for student networks.

[0070] Information entropy is calculated based on the probability distribution of soft tags, and pixel-level uncertainty weights are obtained by normalizing the entropy value. These weights are then used to calculate the weighted distillation loss, completing the annotation of the frame. The formula for calculating information entropy is:

[0071] ;

[0072] in The probability value predicted by the teacher network for category c at pixel x is obtained after softening by a temperature coefficient T. For categorical variables, in a set Take the value from the middle. Indicates the foreground (target organ). This represents the background (non-target organ). For a binary classification problem, the maximum entropy value is... The corresponding foreground and background probabilities are both uniformly distributed, with a probability of 0.5.

[0073] The uncertainty weight is defined as:

[0074] ;

[0075] The normalized value range is [0,1], where 0 represents complete uncertainty and 1 represents complete certainty. This weight is used in the subsequent distillation loss calculation to reduce the impact of unreliable predictions by the teacher network at fuzzy or boundary conditions.

[0076] In a specific embodiment of the present invention, such as Figure 3 As shown, in step S2, the difficult case data augmentation strategy based on ultrasound physical characteristics includes boundary weak echo simulation, organ obstruction simulation, incision and departure simulation, and probe angle change simulation.

[0077] Specifically, boundary weak echo simulation is achieved by attenuating the pixel intensity of the organ boundary region. First, a target organ mask is generated using a teacher network. Extracting organ boundary bands:

[0078] ;

[0079] Where d is the boundary width, and dilate and erode are morphological dilation and erosion operations, respectively. Image intensity is reduced within the boundary band to simulate weak echoes, specifically affecting the enhanced image. have .when It is valid if it is valid, otherwise Where I is the original image, The attenuation coefficient is... , Standard deviation The Gaussian weighting function is used to achieve a smooth transition at the boundary, avoiding abrupt strength discontinuities.

[0080] Organ occlusion simulation is achieved by generating occlusion blocks within the organ region and filling them with either acoustic shadowing artifacts or reverberation artifacts. Occlusion regions occupying 10%–40% of the region of interest are randomly generated within the organ's region of interest. The occlusion shape can be rectangular, elliptical, or irregular polygonal, and its center position and size parameters are randomly selected. One of two artifacts is then filled within the occlusion region: acoustic shadowing artifacts are filled with an exponential decay along the depth direction, using a fill function of [function name missing]. ,in Let x be the sound shadow intensity value at depth y at position x. The initial intensity value is the starting point of the sound shadow, usually taken as the intensity at that location in the original image. y is the depth coordinate of the pixel, measured along the direction of ultrasonic wave propagation. The attenuation coefficient simulates the energy loss of sound waves when blocked by strong reflectors. Reverberation artifacts are created by filling local high-intensity regions and superimposing Gaussian noise. To achieve this, the intensity gain factor is set to a value of 1.5 to 2.5, preferably 2.0, to simulate echo enhancement caused by multiple reflections. The standard deviation of Gaussian noise is taken as 5% to 10% of the image dynamic range to simulate the instability of reverberation.

[0081] The entry and exit simulation is achieved by cropping the organ's edge region. The organ is spatially cropped along random directions, retaining 20%–80% of its original size. The cropping direction is randomly selected from four directions: up, down, left, and right, to simulate the dynamic process of the organ entering or leaving the field of view due to probe movement. The cropped image maintains its original resolution, and the cropped portion is filled with texture from the area outside the boundary or with black to maintain image size consistency.

[0082] The simulation of probe angle change is achieved by performing geometric deformation operations on the original image. Rotation is applied through an affine transformation. ~ Scale by 0.85~1.15 times, cut A transformation of ~0.2 is used to simulate small adjustments in probe orientation. Elastic deformation is achieved using a smoothed random displacement field with displacement amplitudes of 5~15 pixels and a smoothing kernel standard deviation. With a resolution of 20-30 pixels, a random displacement field is generated, smoothed using Gaussian filtering, and then applied to the image to achieve local non-rigid deformation, simulating the elastic deformation of tissue. Geometric transformations are applied only within the region of interest (ROI) of the organ, remaining stable outside the ROI to avoid introducing excessive background changes.

[0083] For adjacent frames in a video sequence, a frame-pair synchronization enhancement strategy is employed, where enhancement parameters for adjacent frames are correlated to maintain the continuity of enhancement between frames. For adjacent frame pairs used in training... The same enhancement type is used for the same frame pair, and the enhancement parameters are fine-tuned based on the t-th frame. Specifically, the enhancement parameters are adjusted based on the t-th frame. Frame enhancement parameters ,in Small perturbations are used to simulate natural changes caused by probe movement while maintaining temporal continuity. Continuous values ​​are also used for spatial transformation parameters such as rotation angles to ensure a smooth visual transition between enhanced images of consecutive frames. Enhanced samples are probabilistically mixed: original samples account for 50%, weak echo enhancement for 20%, occlusion enhancement for 15%, cut-in / de-cut enhancement for 10%, and geometric deformation for 5%, ensuring the model sees sufficient original distribution while increasing the diversity of difficult examples.

[0084] The augmented image is re-input into the teacher network for inference, generating updated pseudo-labels, logits, intermediate features, regions of interest, and uncertainty weights, allowing the pseudo-labels to automatically adapt to the augmented image features. The teacher network then evaluates and filters the augmented training samples, including the following steps: [The text then describes the process of using the augmented image...] The enhanced confidence level is obtained by inputting the teacher network. and segmentation mask The confidence level is calculated as the average confidence level of teachers within the region of interest.

[0085] ;

[0086] Teacher confidence scores were also calculated for the enhanced images. .

[0087] Calculate the enhancement effectiveness evaluation value, and retain samples with evaluation values ​​higher than a preset threshold as augmented training data. The formula for calculating the enhancement effectiveness evaluation value is:

[0088] ;

[0089] in The target organ confidence score of the teacher network on the original image. To enhance the confidence of target organs in the teacher network on images, The teacher segmentation mask corresponding to the original image. To enhance the teacher segmentation mask corresponding to the image, This is the power factor of the difficulty gain. The power coefficient for semantic preservation. The minimum confidence threshold, This is an indicator function; it returns 1 if the condition is met, and 0 otherwise. This is the intersection-union ratio (IUGR). The evaluation function optimizes its structure through a three-term multiplication constraint, automatically balancing difficulty gain and semantic fidelity. A higher evaluation value is achieved only when the enhancement moderately decreases teacher confidence, largely preserves the semantic structure, and allows teachers to still reliably identify the information. The evaluation is calculated after generating the enhanced samples. Only keep The sample, of which The threshold value is used to filter low-quality augmented samples, typically ranging from 0.15 to 0.20, preferably 0.18. This evaluation function quantifies the distillability of augmented samples based on the effectiveness of the augmentation step itself, i.e., high-value difficult samples that teachers are capable of guiding. Compared to simple crossover ratio threshold screening, this function is more effective in selecting difficult samples that are beneficial to student learning.

[0090] In a specific embodiment of the present invention, step S3 specifically includes:

[0091] S31. Construct a student network using a lightweight U-Net architecture. The U-Net consists of an encoder and a decoder. The encoder comprises four convolutional downsampling layers, each containing two 3×3 convolutional layers and one 2×2 max-pooling layer, with feature channels of 32, 64, 128, and 256 respectively. The decoder uses a symmetrical upsampling structure, performing upsampling through 2×2 transposed convolutions and making skip connections with the features of the corresponding layers in the encoder. Finally, a 1×1 convolution outputs the probability predictions for K organ classes. Compared to the teacher network, the student network has fewer parameters and faster inference speed, meeting the real-time deployment requirements of embedded devices.

[0092] S32. Input the expanded training data into the student network and train it using a knowledge distillation strategy to obtain the knowledge distillation loss. The knowledge distillation strategy includes output layer distillation, feature layer distillation, and negative sample suppression. Output layer distillation aligns the output probability distributions of the student network and the teacher network in the region of interest. Feature layer distillation aligns the feature representations of the student network and the teacher network in multiple coding layers. Negative sample suppression suppresses the foreground response of the student network in non-target organ regions.

[0093] Specifically, the output layer distillation loss includes: within the Region of Interest (ROI), weighting the KL divergence between the output probability distributions of the student and teacher networks using uncertainty weights to calculate the output layer distillation loss. The student network also performs two-channel remapping, assuming the currently scanned target organ is... Then the student's prospect output Student background output ,in Let s represent the output probability of the student network for organ o. The subscript s denotes the student network. Temperature softening is applied to the student output to obtain... Where T uses the same temperature coefficient as the teacher network. Within the Region of Interest (ROI), the KL divergence between the output probability distributions of the student and teacher networks is weighted using uncertainty weights to obtain the output layer distillation loss, calculated as follows:

[0094] ;

[0095] in, Indicates whether the pixel is located within the teacher's ROI. For uncertainty weights, the normalized entropy value ranges from [0, 1], where 0 represents complete certainty and 1 represents complete uncertainty, ensuring the numerical stability of uncertainty weights; For student networks in pixels The output probability distribution after softening at temperature T. For teacher networks in pixels The output probability distribution after softening at temperature T. Let KL divergence be the distance between K and L; where .

[0096] Feature layer distillation loss includes: setting adapters in multiple coding layers of the student network to map student network features to the same dimension as teacher network features. The adapters use 1×1 convolutions to align the number of channels, denoted as . ,in For feature layer indexing, layers 2, 3, and 4 of the encoder are selected for feature distillation, corresponding to feature resolutions of 1 / 4, 1 / 8, and 1 / 16 of the original image, respectively. The distance loss between student and teacher network features is calculated within the Region of Interest (ROI) and boundary regions. Boundary regions are weighted using boundary weights to calculate the output feature layer distillation loss, calculated using the following formula:

[0097] ;

[0098] in, , For student networks in the first Features at the layer pixel, For the teacher network in the first Features at the layer pixel, This represents the square of the L2 norm, which is the square of the Euclidean distance. For the first Layer adapter, The boundary weights are obtained by normalizing the gradient magnitude of the teacher mask. Specifically, the teacher mask is first normalized. Calculating gradient magnitude using the Sobel operator Then, normalize to the [0, 1] interval to obtain the boundary weights, thereby highlighting the discriminative feature alignment at the contour; when the teacher and student resolutions are inconsistent, bilinear interpolation is used for alignment first.

[0099] The negative sample suppression loss includes: aggregating the non-target organ masks output by the teacher network to obtain non-target organ regions; suppressing the foreground response of the student network within these non-target organ regions; and calculating the output negative knowledge suppression loss. The aggregation mask of the non-target organ regions output by the teacher network is then obtained. Calculate the foreground output probability of the student network within this region. An inhibition loss is applied to reduce erroneous activations of the student network in non-target regions, resulting in an output negative knowledge inhibition loss:

[0100] ;

[0101] in, , The non-target organ inhibition weighting coefficient. The negative knowledge weight is dynamically determined based on the area ratio of the target organ to the non-target organ, and the formula is as follows;

[0102] ;

[0103] in To suppress the intensity coefficient, The total number of pixels in the target organ mask. This represents the total number of pixels in the aggregate mask for non-target organ regions. This is a limiting function. This dynamic weighting mechanism ensures that the inhibitory force is strengthened when non-target organs account for a large proportion, and appropriately relaxed when the target organs account for a large proportion, avoiding over-constraint.

[0104] S33. The student network is trained using optical flow-guided temporal consistency constraints to constrain the output consistency of the student network in adjacent frames, thus obtaining the temporal consistency loss.

[0105] Temporal consistency loss includes student self-consistency loss and teacher guidance loss;

[0106] The student self-consistency loss includes: calculating the motion field between adjacent frames using an optical flow estimation algorithm, identifying occlusion regions through forward and backward optical flow consistency checks, mapping the student network's output in the previous frame to the current frame position through the motion field, and constraining the consistency between the student network's current frame output and the mapped output within the non-occlusion region to obtain the student self-consistency loss.

[0107] Specifically, for adjacent frames The forward optical flow is calculated using an optical flow estimation network. and backward optical flow Optical flow estimation is implemented using algorithms such as RAFT or PWC-Net. During the training phase, the optical flow field can be pre-calculated offline to improve training efficiency. Occlusion regions are identified by checking the consistency of forward and backward optical flow for each pixel. x ,when Located within the image and satisfying The time marker is the non-occluded pixel, where The distance threshold for forward and backward optical flow consistency checks, where W and H are the image width and height. This represents the relative threshold coefficient. For a typical ultrasound image size of 640×480, the diagonal length is approximately 800 pixels, corresponding to an absolute threshold of approximately 2-4 pixels. This normalization scheme ensures a consistent judgment standard across different resolutions. The student network's output from the previous frame is mapped to the current frame position via a motion field. This mapping operation is implemented using the Warp operator, defined as... Where I is the input image or feature map, and F is the optical flow field. and These are the horizontal and vertical components of the optical flow. This is the output of the previous frame of the student network. Mapped output obtained by applying forward optical flow Constrain the current frame output of the student network within the non-occluded area. After mapping, the output To assess consistency, the L1 distance metric is used to measure the difference, resulting in the student self-consistency loss:

[0108] ;

[0109] in, , It is a time-gated mask, formed by the intersection of the deformed results of the ROI in the current frame and the ROI in the previous frame. The distance metric function is L1 distance; The output of the student network at pixel x in frame t; This describes the result of an output frame on the student network being mapped to the current frame via optical flow.

[0110] The teacher guidance loss includes: mapping the output of the teacher network in the previous frame to the current frame position through the motion field, and mapping the output of the teacher network in the current frame to the previous frame position through the inverse motion field. The KL divergence between the student network output and the mapped teacher soft label is constrained in the non-occluded area to obtain the teacher guidance loss.

[0111] Specifically, the teacher-guided item uses the deformation result output by the teacher in the previous frame as a cross-frame reference and applies bidirectional consistency constraints:

[0112] ;

[0113] Simultaneously, the output of the teacher network in the current frame will be... By mapping the inverse motion field to the previous frame position, the inverse mapping result is obtained. Constraining the KL divergence between the student network output and the mapped teacher soft label within the non-occluded region yields:

[0114] Forward teacher guidance loss:

[0115] ;

[0116] Losses from backward teacher guidance:

[0117] ;

[0118] In summary, the teacher's guidance loss is:

[0119] ;

[0120] Robustness of timing consistency is enhanced through bidirectional constraints.

[0121] S34. By jointly optimizing the knowledge distillation loss and the temporal consistency loss, a lightweight student network model is obtained.

[0122] The temporal consistency loss is obtained by combining the student self-consistency loss and the teacher guidance loss:

[0123] ;

[0124] in, and These are the loss weights;

[0125] In summary, the formula for calculating the total training loss of joint optimization is:

[0126] ;

[0127] in The task loss between the student network and the ground truth labels is implemented using a weighted combination of Dice loss and cross-entropy loss, with a weight ratio of 1:1. The weight coefficients of each loss term were determined through experimental tuning. The Adam optimizer was used during training, with an initial learning rate of 0.001, which decayed to 0.5 times the original rate every 30 epochs, for a total of 100 to 200 epochs, and a batch size of 8 to 16. Through joint optimization, the student network achieved segmentation accuracy close to that of the teacher network while maintaining its lightweight characteristics, and also exhibited better inter-frame stability.

[0128] This invention effectively addresses the common issues of boundary blurring and inter-frame jitter in lightweight models for ultrasound video segmentation through multi-level knowledge transfer involving output layer alignment, feature layer alignment, and negative sample suppression, combined with optical flow-guided temporal constraints. Compared to traditional static image distillation methods, the introduced temporal consistency constraint significantly reduces jumps in output between adjacent frames, improving the stability of video segmentation. The negative sample suppression mechanism, specifically designed for multi-organ scanning scenarios in ultrasound robots, effectively reduces false activations of non-target organs, enhancing the specialization of the model.

[0129] In a specific embodiment of the present invention, step S4 includes:

[0130] S41. Input the test ultrasound video sequence into the teacher network and obtain the confidence scores of multiple organ classes output by the teacher network. The teacher network outputs the segmentation mask of K organ classes and the corresponding confidence score for each frame. The confidence score is calculated as the average predicted probability within the mask region of each organ class.

[0131] S42. Real-time monitoring of the confidence levels of the multiple organ types. When the confidence level of a certain organ exceeds a preset switching threshold for several consecutive frames, the current scanning target is determined to be that organ. Specifically, real-time monitoring of the confidence levels of K types of organs output by the teacher network. When the confidence level of a certain organ For N consecutive frames, the threshold is exceeded When the target of the current scan is determined to be the organ, the time window length N is 3 to 5 frames, preferably 3 frames, and the switching threshold is... The value is set between 0.5 and 0.7, with 0.6 being the preferred value. This hysteresis determination mechanism avoids frequent switching caused by single-frame noise, thus improving the stability of organ recognition.

[0132] S43. Based on the determined current target organ, switch to the corresponding organ's student network model. A lag-based switching strategy is employed to ensure the new organ's confidence level remains stable before performing the switching operation. The system pre-trains a dedicated student network model for each organ category. When a target organ switch is detected, the system loads the corresponding organ's student network weights and switches to that model for inference. The lag-based switching strategy requires the new organ's confidence level to remain stable for at least N frames above a threshold before switching, avoiding instantaneous misjudgments during probe movement that could lead to frequent model switching and affect system stability.

[0133] S44. The student network model, after switching the input of the test ultrasound video sequence, performs segmentation processing and outputs the segmentation results. The student network performs real-time inference on the input ultrasound image and outputs the organ segmentation mask for that frame. The inference speed can reach 15-30 frames per second, meeting the real-time scanning requirements of the ultrasound robot. The output segmentation results are used to guide the path planning and scanning control of the ultrasound robot, realizing intelligent autonomous scanning.

[0134] During the inference phase, to ensure real-time performance, a simplified motion estimation based on inter-frame difference is used instead of full optical flow calculation. Motion vectors are estimated by pixel-wise correlation matching on the intermediate feature maps of the student network encoder. The computational complexity is approximately 1 / 5 that of the full optical flow network, providing basic temporal stability while maintaining a processing speed of 15-20 frames per second. Specifically, the cross-correlation of features from adjacent frames is calculated on the second or third layer feature maps of the student network encoder. Pixel displacements are estimated by searching for the position of maximum correlation, thus obtaining a simplified motion field for temporal constraints. This simplified scheme significantly improves inference speed with a slight decrease in accuracy, enabling real-time deployment of the lightweight student network on embedded platforms.

[0135] Through the aforementioned multi-organ intelligent switching mechanism, the system can automatically select the most suitable specialist model for segmentation based on the organ type in the current field of view during ultrasound robot scanning. Compared with using a single general model or teacher network, the specialist student model has higher segmentation accuracy and stronger anti-interference ability on its respective target organs, while maintaining the high inference speed advantage brought by lightweight design.

[0136] In another embodiment of the present invention, an intelligent recognition system for ultrasound robots based on an AI-powered large-scale ultrasound model is provided, including a data acquisition and annotation module, a difficult example data expansion module, a knowledge distillation training module, and an intelligent recognition and segmentation module.

[0137] The data acquisition and annotation module is used to acquire ultrasound video sequences collected during ultrasound robot scanning and to annotate these sequences using a teacher network, resulting in raw training data with segmentation masks and confidence labels. This module integrates the teacher network's inference interface, receiving real-time or offline ultrasound video streams, performing multi-organ segmentation inference on each frame, generating annotation information such as segmentation masks, confidence scores, soft label probability distributions, and uncertainty weights for K types of organs, and storing the annotation results as the training dataset.

[0138] The difficult example data augmentation module uses a difficult example data augmentation strategy based on ultrasound physical characteristics to enhance the original training data, resulting in enhanced training samples. These enhanced training samples are then evaluated and filtered by a teacher network, retaining difficult examples that decrease the teacher network's confidence while preserving the semantic structure as augmented training data. This module implements four augmentation strategies: boundary weak echo simulation, organ occlusion simulation, incision / departure simulation, and probe angle change simulation. Each augmentation strategy is encapsulated as an independent processing function, allowing flexible adjustment of augmentation parameters and mixing ratios according to configuration files. The enhanced samples are re-inferred by the teacher network to obtain updated pseudo-labels, and then filtered using an augmentation effectiveness evaluation function. Only samples with evaluation values ​​exceeding a threshold are added to the augmented training dataset.

[0139] The knowledge distillation training module is used to construct the student network. Expanded training data is input into the student network for training. A knowledge distillation strategy is employed to transfer knowledge from the teacher network to the student network. Knowledge distillation includes calculating the output layer distillation loss, feature layer distillation loss, and negative sample suppression loss. Optical flow-guided temporal consistency constraints are used to regularize the training of the student network, resulting in a lightweight student network model. This module includes functional components such as student network architecture definition, training process control, loss function calculation, and model parameter optimization. During training, changes in various loss values ​​are monitored in real time, and model performance is evaluated on the validation set. An early stopping mechanism is triggered when the validation loss no longer decreases for several consecutive epochs. After training, corresponding student network weight files are saved for each organ category for loading and switching during subsequent inference stages.

[0140] The intelligent recognition and segmentation module inputs the test ultrasound video sequence into the teacher network. Based on the multi-organ confidence scores output by the teacher network, it determines the current target organ for scanning, switches to the corresponding student network model for segmentation, and outputs the segmentation result. This module maintains a time-series buffer of organ confidence scores, tracking the confidence scores of various organs in the most recent N frames in real time. When the confidence score of a certain organ consistently and stably exceeds a threshold, it is determined that the target organ is switched. The module has built-in weight files for K organ-specific student networks, and achieves fast switching through a model loader, with the switching latency controlled within 100 to 200 milliseconds. The output segmentation mask undergoes morphological post-processing operations such as hole filling and edge smoothing to obtain the final segmentation result, which is then superimposed on the original ultrasound image, providing real-time feedback for the navigation and control of the ultrasound robot.

[0141] Through the modular design described above, this system achieves a complete workflow from data annotation, difficult example expansion, model training to real-time inference. The modules interact through standardized data interfaces, facilitating system deployment and maintenance. The system supports large-scale training on cloud servers. After training, the lightweight student network model is deployed to the embedded controller of the ultrasonic robot, enabling real-time intelligent segmentation and recognition at the edge.

[0142] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent recognition of ultrasound robots based on an AI-powered large-scale ultrasound model, characterized in that: Includes the following steps: S1. Obtain the ultrasound video sequence collected during the ultrasound robot scanning process, and annotate the ultrasound video sequence through the teacher network to obtain the original training data with segmentation mask and confidence label. S2. The original training data is augmented using a difficult example data augmentation strategy based on the physical characteristics of ultrasound to obtain augmented training samples. The augmented training samples are then evaluated and screened by the teacher network, and difficult example samples that reduce the confidence of the teacher network while maintaining the semantic structure are retained as augmented training data. S3. Construct a student network, input the expanded training data into the student network for training, use the knowledge distillation strategy to transfer the knowledge of the teacher network to the student network, and use optical flow-guided temporal consistency constraints to regularize the adjacent frame outputs of the student network to obtain a lightweight student network model. S4. Input the test ultrasound video sequence into the teacher network, determine the current target organ for scanning based on the multi-organ confidence scores output by the teacher network, switch to the student network model of the corresponding organ for real-time segmentation inference, and output the segmentation results. Step S3 specifically includes: S31. Construct a student network, wherein the student network adopts a lightweight U-Net architecture; S32. Input the expanded training data into the student network and train it using a knowledge distillation strategy to obtain the knowledge distillation loss. The knowledge distillation strategy includes output layer distillation, feature layer distillation, and negative sample suppression. Output layer distillation aligns the output probability distributions of the student network and the teacher network in the region of interest. Feature layer distillation aligns the feature representations of the student network and the teacher network in multiple coding layers. Negative sample suppression suppresses the foreground response of the student network in non-target organ regions. S33. The student network is trained using optical flow-guided temporal consistency constraints to regularize the training. This constrains the output consistency of the student network in adjacent frames, resulting in a temporal consistency loss. The temporal consistency loss includes student self-consistency loss and teacher-guided loss. The student self-consistency loss includes: calculating the motion field between adjacent frames using an optical flow estimation algorithm; identifying occlusion regions through forward and backward optical flow consistency checks; mapping the student network's output from the previous frame to the current frame position using the motion field; and constraining the consistency between the student network's current frame output and the mapped output within non-occlusion regions. The teacher-guided loss includes: mapping the teacher network's output from the previous frame to the current frame position using the motion field; simultaneously mapping the teacher network's output from the current frame to the previous frame position using the reverse motion field; and constraining the KL divergence between the student network output and the mapped teacher soft label within non-occlusion regions. S34. By jointly optimizing the knowledge distillation loss and the temporal consistency loss, a lightweight student network model is obtained.

2. The intelligent recognition method for ultrasound robots based on an AI-powered large-scale ultrasound model as described in claim 1, characterized in that: In step S1, the teacher network annotates the ultrasound video sequence, including: The teacher network performs inference on the input ultrasound video frames and outputs multiple organ masks and corresponding logits. Based on the current target organ being scanned, the corresponding target organ mask is selected as the foreground mask, and the masks of the remaining organs are merged as the background mask. Two-channel remapping is then performed to obtain binary labels for the foreground and background. A morphological dilation operation is performed on the foreground mask to generate a region of interest (ROI), which is used for spatial gating in the subsequent distillation process. Temperature softening is applied to the output logits of the teacher network to obtain the soft label probability distributions of the foreground and background, which are used as the supervision signal for output layer distillation. The entropy value is calculated based on the probability distribution of the soft label, and the pixel-level uncertainty weight is obtained by normalizing the entropy value. This weight is then used to calculate the weighted distillation loss, thus completing the annotation of the frame.

3. The intelligent recognition method for ultrasound robots based on an AI-powered large-scale ultrasound model as described in claim 1, characterized in that: In step S2, the hard case data augmentation strategy based on ultrasound physical characteristics includes boundary weak echo simulation, organ occlusion simulation, cut-in / departure simulation, and probe angle change simulation. Among them, boundary weak echo simulation is achieved by attenuating the pixel intensity of the organ boundary region; organ occlusion simulation is achieved by generating occlusion blocks in the organ region and filling them with acoustic shadowing artifacts or reverberation artifacts; cut-in / departure simulation is achieved by cropping the organ edge region; and probe angle change simulation is achieved by performing geometric deformation operations on the original image. For adjacent frames in the video sequence, a frame-pair synchronization enhancement strategy is adopted, and the enhancement parameters of adjacent frames are set in association to maintain the continuity of inter-frame enhancement.

4. The intelligent recognition method for ultrasound robots based on an AI-powered large-scale ultrasound model as described in claim 1, characterized in that: In step S2, the evaluation and selection of enhanced training samples through the teacher network includes: inputting the enhanced image into the teacher network to obtain the enhanced confidence score and segmentation mask, calculating the enhancement effectiveness evaluation value, and retaining samples with evaluation values ​​higher than a preset threshold as expanded training data; the calculation formula for the enhancement effectiveness evaluation value is: ； Where the subscript t represents the teacher network, The confidence level of the target organ in the teacher network on the original image; To enhance the confidence of target organs in the teacher network on the image; The teacher segmentation mask corresponding to the original image; To enhance the teacher segmentation mask corresponding to the image; This is the power factor of the difficulty gain; The power coefficients for semantic preservation; The minimum confidence threshold; For indicator functions; For intersection, union, and comparison.

5. The intelligent recognition method for ultrasound robots based on an AI-powered large-scale ultrasound model as described in claim 1, characterized in that: Knowledge distillation loss includes output layer distillation loss, feature layer distillation loss, and negative sample suppression loss. The output layer distillation loss includes: within the region of interest (ROI), the KL divergence between the output probability distributions of the student network and the teacher network is weighted using uncertainty weights to calculate the output layer distillation loss; The feature layer distillation loss includes: setting adapters in multiple coding layers of the student network to map student network features to the same dimension as teacher network features; calculating the distance loss between student network features and teacher network features in the region of interest (ROI) and boundary regions; and weighting the boundary regions through boundary weights to calculate the feature layer distillation loss. The negative sample suppression loss includes: aggregating the non-target organ masks output by the teacher network to obtain non-target organ regions, suppressing the foreground response of the student network within the non-target organ regions, and calculating the negative knowledge suppression loss.

6. The intelligent recognition method for ultrasound robots based on an AI-powered large-scale ultrasound model as described in claim 1, characterized in that: In step S34, the formula for calculating the total training loss of joint optimization is: ； ； ； ； ； Where the subscript t represents the teacher network and the subscript s represents the student network. The task loss between the student network and the truth label; For output layer distillation loss, For pixel position, ROI gating weights For uncertain weights, For student networks in pixels The output probability distribution after softening at temperature T. For teacher networks in pixels The output probability distribution after softening at temperature T, where T represents the temperature softening coefficient. Let KL divergence be a metric. To mitigate loss through negative knowledge, For aggregation masks of non-target organ regions, Negative knowledge weight, The non-target organ inhibition weighting coefficient. Output the probability of the prospect of the student network; For characteristic layer distillation loss, For feature layer index, For student networks in the first Features at the layer pixel, For the teacher network in the first Features at the layer pixel, For the first Layer adapter, For boundary weights, This represents the square of the L2 norm, which is the square of the Euclidean distance. For time-series consistency loss, The weight of student self-consistency loss, Weighting of teacher-guided loss, For the students' own losses, This results in losses for teachers' guidance.

7. The intelligent recognition method for ultrasound robots based on an AI-powered large-scale ultrasound model as described in claim 1, characterized in that: Step S4 includes: S41. Input the test ultrasound video sequence into the teacher network and obtain the confidence scores of multiple organs output by the teacher network. S42. Monitor the confidence level of the multiple organs in real time. When the confidence level of a certain organ exceeds the preset switching threshold in several consecutive frames, determine that the current scanning target is that organ. S43. Based on the determined current target organ, switch to the corresponding student network model, and use a lag switching strategy to ensure that the confidence of the new organ remains stable before performing the switching operation. S44. Perform segmentation processing on the student network model after inputting the test ultrasound video sequence and output the segmentation results.

8. An intelligent recognition system for ultrasound robots based on an AI-powered large-scale ultrasound model, characterized in that: The system is used to perform the method as described in any one of claims 1-7, comprising: Data acquisition and annotation module: used to acquire ultrasound video sequences collected during ultrasound robot scanning, and to annotate the ultrasound video sequences through the teacher network to obtain raw training data; Data augmentation module: used to augment the original training data and evaluate and screen the augmented training samples through the teacher network to obtain augmented training data; Knowledge distillation training module: used to construct student network, input expanded training data into the student network for training, use knowledge distillation strategy to transfer knowledge from teacher network to student network, and use optical flow-guided temporal consistency constraint to regularize the student network to obtain a lightweight student network model; Intelligent recognition and segmentation module: It is used to input the test ultrasound video sequence into the teacher network, determine the current target organ for scanning based on the confidence scores of multiple organs output by the teacher network, switch to the student network model of the corresponding organ for segmentation processing, and output the segmentation results.

Images