Pelvic and abdominal dynamics detection method and system based on artificial intelligence

Abstract

The present invention provides a pelvic and abdominal dynamics detection method and system based on artificial intelligence, which relates to the technical field of pelvic floor function detection, including arranging a thermistor sensor array on the surface of a vaginal detection probe, identifying the contact area between the vaginal detection probe and the vaginal wall according to temperature data, and calculating the area of ​​the contact area to obtain an intelligently designed vaginal detection probe; using the intelligently designed vaginal detection probe to obtain time-synchronized detection data, and at the same time constructing a pelvic floor dysfunction disease knowledge graph; inputting the time-synchronized detection data into the pelvic floor dysfunction disease knowledge graph, using a graph reasoning algorithm to match corresponding disease nodes in the pelvic floor dysfunction disease knowledge graph, and calculating the confidence of each disease node, performing differential diagnosis on disease nodes with higher confidence, and recommending personalized rehabilitation plans from the pelvic floor dysfunction disease knowledge graph according to the differential diagnosis results.

Description

Technical Field

[0001] The present invention relates to pelvic floor function detection technology, and in particular to a pelvic and abdominal dynamics detection method and system based on artificial intelligence. Background Art

[0002] At present, the incidence of pelvic floor dysfunction diseases (such as vaginal wall prolapse, pelvic organ prolapse, urinary incontinence, etc.) is high, seriously affecting the quality of life of patients. Traditional pelvic floor function testing methods mainly rely on the subjective judgment of doctors, lack of quantitative objective evaluation methods, and the diagnosis process is subjective and uncertain.

[0003] In the existing technology, although some detection methods have introduced various instruments and equipment for auxiliary detection, most of them are still limited to the measurement of a single physiological parameter and cannot fully reflect the patient's pelvic floor function. At the same time, these detection methods usually only focus on the detection data itself, lack of in-depth analysis and intelligent processing of the data, and it is difficult to provide valuable reference opinions for clinical diagnosis and treatment.

[0004] Therefore, it is necessary to design a pelvic and abdominal dynamics detection method and system that integrates artificial intelligence technology to solve the defects in the existing technology and improve the intelligence and personalization level of pelvic floor dysfunction disease detection and diagnosis. By combining intelligently designed vaginal detection probes, multi-source heterogeneous data fusion processing, knowledge graph reasoning and other technologies, a more scientific and accurate solution is provided for the diagnosis and treatment of pelvic floor dysfunction. Summary of the invention

[0005] The embodiments of the present invention provide a pelvic and abdominal dynamics detection method and system based on artificial intelligence, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] Provided is a pelvic and abdominal dynamics detection method based on artificial intelligence, comprising:

[0008] Arrange a thermistor sensor array on the surface of the vaginal detection probe, collect temperature data in real time during the vaginal detection process, identify the contact area between the vaginal detection probe and the vaginal wall according to the temperature data, and calculate the contact area area, and dynamically adjust the shape and pressure of the vaginal detection probe according to the calculated contact area area to obtain an intelligently designed vaginal detection probe;

[0009] The multi-source heterogeneous data collected during the pelvic floor function test are collected using an intelligently designed vaginal detection probe, and the collected multi-source heterogeneous data are standardized and synchronized to obtain time-synchronized test data. At the same time, relevant knowledge about pelvic floor dysfunction diseases is extracted from medical literature and clinical guidelines to construct a knowledge graph of pelvic floor dysfunction diseases.

[0010] The time-synchronized detection data is input into the pelvic floor dysfunction disease knowledge graph, and the graph reasoning algorithm is used to match the corresponding disease nodes in the pelvic floor dysfunction disease knowledge graph, and the confidence of each disease node is calculated, and the disease nodes with higher confidence are differentially diagnosed, and the differential diagnosis results are output, and personalized rehabilitation plans are recommended from the pelvic floor dysfunction disease knowledge graph according to the differential diagnosis results.

[0011] In an optional embodiment,

[0012] An array of thermistor sensors is arranged on the surface of the vaginal detection probe, and temperature data is collected in real time during the vaginal detection process. The contact area between the vaginal detection probe and the vaginal wall is identified according to the temperature data, and the contact area area is calculated. According to the calculated contact area area, the shape and pressure of the vaginal detection probe are dynamically adjusted. The intelligently designed vaginal detection probe includes:

[0013] A miniature infrared thermal imager is arranged on the vaginal detection probe, and a temperature-radiation compensation function is constructed for the original infrared thermal image collected by the miniature infrared thermal imager using a pre-established coupling model of heat conduction inside the probe and surface radiation, and a corrected infrared thermal image is obtained by pixel-by-pixel correction, wherein the corrected infrared thermal image reflects the temperature distribution on the surface of the vaginal wall;

[0014] Based on the corrected infrared thermal image, a three-dimensional geometric model of the vaginal detection probe is constructed using feature points extracted from the surface of the vaginal detection probe, and the corrected infrared thermal image is aligned with the three-dimensional geometric model through rigid body transformation and non-rigid deformation to obtain a three-dimensional temperature distribution map, which includes temperature, spatial position and morphological information of the contact area;

[0015] Input the three-dimensional temperature distribution map into a pre-trained fully convolutional neural network, extract high-level semantic features of the three-dimensional temperature distribution map through multi-layer convolution and pooling operations, and output the segmentation result of the contact area between the vaginal detection probe and the vaginal wall;

[0016] The segmentation result is synchronized and aligned in time and space with the temperature distribution, the corrected infrared thermal map, the pressure distribution and the contact area of ​​the vaginal detection probe to construct a contact state data set;

[0017] Extract high-level semantic features of infrared thermal images and dynamic change features of temperature distribution and pressure distribution data in the contact state data set, and perform feature fusion to obtain the contact state feature vector;

[0018] The contact state feature vector is input into a pre-trained contact state intelligent evaluation model to predict the contact area between the vaginal detection probe and the vaginal wall. Based on the obtained contact area, combined with the pressure distribution uniformity and temperature distribution consistency, the contact state quality score is calculated. The shape and pressure of the vaginal detection probe are dynamically adjusted according to the contact state quality score to obtain an intelligently designed vaginal detection probe.

[0019] In an optional embodiment,

[0020] The steps before inputting the contact state feature vector into the pre-trained contact state intelligent assessment model include:

[0021] Acquire a feature vector of the contact state between the vaginal detection probe and the vaginal wall during vaginal detection as an original sample, randomly generate a perturbation near the feature space of the original sample, and superimpose the perturbation on the original sample to obtain an adversarial sample;

[0022] The original samples and adversarial samples are input into the contact state intelligent evaluation model based on multi-task learning. The contact state intelligent evaluation model simultaneously performs the contact area regression task and the contact state quality classification task, and obtains the loss functions corresponding to the two tasks respectively. The total loss function is constructed based on the loss functions corresponding to the two tasks.

[0023] The parameters of the contact state intelligent evaluation model are fixed, and the adversarial loss function is calculated using the adversarial sample. The adversarial sample is then fixed, and the parameters of the contact state intelligent evaluation model are optimized to minimize the weighted sum of the total loss function and the adversarial loss function as the optimization goal. The parameters of the intelligent evaluation model are updated by the gradient descent algorithm, and the iteration is repeated until the model converges. In each round of iteration, a disturbance amplitude attenuation factor and a weight attenuation factor are introduced to obtain the loss function value of the contact state intelligent evaluation model on the original sample and the loss function value of the adversarial sample on the verification set. According to the difference between the two loss function values, the intensity of the adversarial training is dynamically adjusted to obtain the finally trained contact state intelligent evaluation model.

[0024] In an optional embodiment,

[0025] The intelligently designed vaginal detection probe is used to collect multi-source heterogeneous data during the pelvic floor function detection process, and the collected multi-source heterogeneous data is standardized and synchronized to obtain the time-synchronized detection data including:

[0026] Constructing a pelvic floor function detection data set, wherein the pelvic floor function detection data set includes vaginal pressure signals, electromyographic signals, and ultrasound image sequences of multiple subjects and manually annotating them to obtain standardized parameters;

[0027] Preprocessing the vaginal pressure signal and the electromyographic signal, inputting the preprocessed vaginal pressure signal and the electromyographic signal into a one-dimensional convolutional neural network, respectively extracting multi-scale local features of the vaginal pressure signal and the electromyographic signal, and mapping the multi-scale local features into a pressure signal feature vector and an electromyographic signal feature vector through a fully connected layer; simultaneously inputting the ultrasound image sequence into a three-dimensional convolutional neural network, extracting the spatiotemporal features of the ultrasound image sequence, and mapping the spatiotemporal features into an ultrasound image feature vector through a fully connected layer;

[0028] splicing the pressure signal feature vector, the electromyographic signal feature vector and the ultrasound image feature vector to form a multimodal feature vector, inputting the multimodal feature vector into a long short-term memory network, and extracting cross-modal spatiotemporal features;

[0029] Inputting the cross-modal spatiotemporal features into a first fully connected layer and a second fully connected layer, predicting standardized parameters of vaginal pressure signals, electromyographic signals, and ultrasound image sequences through the first fully connected layer, and predicting synchronization offsets between vaginal pressure signals, electromyographic signals, and ultrasound image sequences through the second fully connected layer;

[0030] Using manually annotated standardized parameters and synchronization offsets as supervisory signals, jointly training the one-dimensional convolutional neural network, the three-dimensional convolutional neural network, the long short-term memory network, the first fully connected layer, and the second fully connected layer, optimizing the network parameters by minimizing the mean square error loss function, and obtaining a trained data processing model;

[0031] The collected multi-source heterogeneous data are input into the trained data processing model, and the standardized parameters and synchronization offset of the vaginal pressure signal, electromyographic signal and ultrasound image sequence are output. The vaginal pressure signal, electromyographic signal and ultrasound image sequence are standardized according to the standardized parameters, and the vaginal pressure signal, electromyographic signal and ultrasound image sequence are synchronized according to the synchronization offset, and finally the detection data after time synchronization is obtained.

[0032] In an optional embodiment,

[0033] The time-synchronized detection data is input into the pelvic floor dysfunction disease knowledge graph, and the corresponding disease nodes are matched in the pelvic floor dysfunction disease knowledge graph using the graph reasoning algorithm, and the confidence of each disease node is calculated. The differential diagnosis of the disease nodes with higher confidence includes:

[0034] A random walk algorithm is used to generate a node sequence on the knowledge graph of pelvic floor dysfunction diseases, and the node sequence is converted into a low-dimensional dense vector through a word embedding model to obtain an embedded representation of each disease node;

[0035] The time-synchronized detection data is represented as a query vector, and the embedded representation of the disease node is used as a key vector and a value vector, and multiple groups of attention weights are calculated in parallel through a pre-trained graph neural network based on a multi-head attention mechanism to obtain a matching score between the query vector and each disease node;

[0036] Based on expert experience and prior knowledge, the evidence theory framework is used to set the initial confidence for each disease node, and the basic probability assignment of evidence is defined to represent the initial confidence level of different diseases.

[0037] Using the matching score as supporting evidence, constructing a new basic probability assignment, and calculating the correlation evidence between disease nodes using the disease relationship edges in the pelvic floor dysfunction disease knowledge graph;

[0038] The initial confidence, matching score and correlation evidence are synthesized using the combination rules in the evidence theory framework, and the basic probability assignments of multiple evidences are combined by calculating the consistency between the evidences to obtain the final confidence of each disease node;

[0039] The disease nodes are sorted according to the final confidence, a differential diagnosis is performed on the disease node with the highest confidence, and a differential diagnosis result is output.

[0040] In an optional embodiment,

[0041] The disease nodes are sorted according to the final confidence, a differential diagnosis is performed on the disease node with the highest confidence, and the differential diagnosis results are output including:

[0042] Building an intelligent diagnosis rule base based on medical expert knowledge, the intelligent diagnosis rule base includes a set of diagnosis constraint rules, each of which consists of a number of diagnosis states and corresponding legal diagnosis actions;

[0043] Divide the differential diagnosis task into multiple levels of Markov decision subprocesses, each Markov decision subprocess corresponds to a diagnostic subtask, the state space of the diagnostic subtask includes the obtained patient symptoms, signs and auxiliary examination results, and the action space includes a set of alternative diagnostic actions at the current diagnostic subtask level;

[0044] A multi-objective reward function is constructed for each Markov decision subprocess, wherein the multi-objective reward function includes a diagnostic accuracy reward, a diagnostic efficiency reward, and a patient comfort reward, and a diagnostic strategy network is trained using a deep reinforcement learning algorithm, wherein the input of the diagnostic strategy network is a state representation of the diagnostic subtask, and the output is a probability distribution of each alternative diagnostic action;

[0045] After each diagnostic strategy network generates the diagnostic action probability distribution, it determines whether the diagnostic action with the highest probability satisfies the diagnostic constraint rules in the intelligent diagnostic rule base. If so, the current diagnostic action is executed, and the obtained diagnostic action and diagnostic result are used as the state input of the next Markov decision subprocess. Otherwise, the current diagnostic strategy network is punished and the diagnostic action is regenerated.

[0046] The Markov decision subprocesses at multiple levels execute the steps of generating diagnostic actions in sequence according to the levels until all diagnostic subtasks are completed, and the results output by each diagnostic subtask are integrated to obtain the final differential diagnosis result.

[0047] In an optional embodiment,

[0048] Training a diagnostic policy network using a deep reinforcement learning algorithm involves:

[0049] Initialize the parameters of the diagnosis strategy network, randomly sample a batch of diagnosis subtask state representations, input them into the diagnosis strategy network, and calculate the probability distribution of each candidate diagnosis action through forward propagation;

[0050] According to the probability distribution of the output of the diagnostic strategy network, a set of diagnostic action samples is generated using the Monte Carlo method, and the comprehensive instant reward of each diagnostic action sample is calculated according to the multi-objective reward function. At the same time, the advantage function value of each diagnostic action sample is estimated using the generalized advantage estimation method. The loss function is constructed based on the advantage function value of the diagnostic action sample, and the parameters of the diagnostic strategy network are updated through the stochastic gradient descent algorithm. The iteration is repeated until the preset number of iterations is reached to obtain the final trained diagnostic strategy network.

[0051] Among them, the loss function calculation formula of the diagnosis strategy network is as follows:

[0052]

[0053] Where L(θ) represents the loss function, θ represents the diagnostic strategy network parameters, E represents the expectation, τ represents the trajectory of the Markov decision process, π represents the diagnostic strategy, T represents the total time step of the trajectory, γ represents the discount factor, R(·) represents the immediate reward function, λ represents the penalty coefficient for constraint violation, C(·) represents the constraint function, and s t represents the state at time step t, a t represents the action at time step t, logπ θ (at|st) indicates that the diagnostic strategy network is in state s t Next select action a t The logarithmic probability of

[0054] Among them, the Markov decision sub-processes of multiple levels execute the steps of generating diagnostic actions in sequence according to the levels until all diagnostic sub-tasks are completed, including:

[0055] Construct a reward transfer mechanism between Markov decision subprocesses to complete the coordinated optimization of multiple Markov decision subprocesses. The formula of the reward transfer mechanism is as follows:

[0056]

[0057] Among them, K l represents the cumulative reward of the Markov decision subprocess at the lth level, T l represents the total time step of the Markov decision subprocess at the lth layer, t represents the time step of the current Markov decision subprocess, η represents the discount coefficient of the immediate reward, and r l (·) represents the immediate reward of the Markov decision subprocess at the lth layer, represents the state of the Markov decision subprocess at the lth level at time step t, represents the action of the Markov decision subprocess at the lth level at time step t, V l+1 (·) represents the state value function of the l+1th layer Markov decision subprocess, and s1 represents the initial state.

[0058] According to a second aspect of the embodiments of the present invention,

[0059] Provided is a pelvic and abdominal dynamics detection system based on artificial intelligence, comprising:

[0060] The first unit is used to arrange a thermistor sensor array on the surface of the vaginal detection probe, collect temperature data in real time during the vaginal detection process, identify the contact area between the vaginal detection probe and the vaginal wall according to the temperature data, and calculate the contact area area, and dynamically adjust the shape and pressure of the vaginal detection probe according to the calculated contact area area to obtain an intelligently designed vaginal detection probe;

[0061] The second unit is used to collect multi-source heterogeneous data in the process of pelvic floor function detection using an intelligently designed vaginal detection probe, and to standardize and synchronize the collected multi-source heterogeneous data to obtain time-synchronized detection data. At the same time, relevant knowledge about pelvic floor dysfunction diseases is extracted from medical literature and clinical guidelines to construct a knowledge graph of pelvic floor dysfunction diseases.

[0062] The third unit is used to input the time-synchronized detection data into the pelvic floor dysfunction disease knowledge graph, use the graph reasoning algorithm to match the corresponding disease nodes in the pelvic floor dysfunction disease knowledge graph, calculate the confidence of each disease node, perform differential diagnosis on the disease nodes with higher confidence, and output the differential diagnosis results, and recommend personalized rehabilitation plans from the pelvic floor dysfunction disease knowledge graph according to the differential diagnosis results.

[0063] According to a third aspect of the embodiments of the present invention,

[0064] An electronic device is provided, comprising:

[0065] processor;

[0066] a memory for storing processor-executable instructions;

[0067] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0068] According to a fourth aspect of the embodiments of the present invention,

[0069] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0070] In this embodiment, the contact area between the probe and the vaginal wall is detected in real time by the thermistor sensor array, and the probe shape and pressure are adjusted dynamically, so that the probe can better fit the shape of the vagina, improve the detection accuracy, reduce discomfort, and improve the detection comfort. By synchronously processing the multi-source heterogeneous data collected during the detection process, more complete and consistent patient detection data can be obtained, providing high-quality data support for subsequent diagnosis and analysis. By constructing medical knowledge into a knowledge graph and combining it with patient detection data for graph reasoning, the diagnosis process can be guided and constrained by domain knowledge, thereby improving the accuracy and interpretability of disease diagnosis. Based on the patient's specific symptoms and diagnosis results, personalized rehabilitation plans are recommended from the knowledge graph, which helps to improve the pertinence and effectiveness of rehabilitation. By applying artificial intelligence technology to all aspects of the diagnosis and treatment of pelvic floor dysfunction, the intelligence and automation level of the diagnosis and treatment process is improved, which helps to reduce the cost of diagnosis and treatment and improve the efficiency of diagnosis and treatment. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 It is a flow chart of a method for detecting pelvic and abdominal dynamics based on artificial intelligence according to an embodiment of the present invention;

[0072] Figure 2 It is a schematic diagram of the structure of a pelvic and abdominal dynamics detection system based on artificial intelligence according to an embodiment of the present invention. DETAILED DESCRIPTION

[0073] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0074] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0075] Figure 1 FIG. 1 is a flow chart of a method for detecting pelvic and abdominal dynamics based on artificial intelligence according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0076] S101. Arrange a thermistor sensor array on the surface of the vaginal detection probe, collect temperature data in real time during the vaginal detection process, identify the contact area between the vaginal detection probe and the vaginal wall based on the temperature data, and calculate the contact area area. Based on the calculated contact area area, dynamically adjust the shape and pressure of the vaginal detection probe to obtain an intelligently designed vaginal detection probe.

[0077] A polymer material with good biocompatibility and hydrophilicity is coated on the probe surface to improve the biocompatibility of the probe with the vaginal mucosa and reduce tissue irritation and inflammatory response. At the same time, an array of thermistor sensors is arranged on the probe surface to collect temperature distribution data on the probe surface in real time. The contact state between the probe and the vaginal wall is determined based on the temperature data, and the area of ​​the identified contact area is calculated. When the contact area is lower than the preset threshold, the probe is controlled to expand or apply pressure toward the vaginal wall; when the contact area exceeds the preset threshold, the probe is controlled to contract or reduce pressure. Through iterative optimization, dynamic matching and adaptive adjustment of the contact area between the probe and the vaginal wall are achieved to ensure the consistency and repeatability of the test data.

[0078] In an optional embodiment,

[0079] An array of thermistor sensors is arranged on the surface of the vaginal detection probe, and temperature data is collected in real time during the vaginal detection process. The contact area between the vaginal detection probe and the vaginal wall is identified according to the temperature data, and the contact area area is calculated. According to the calculated contact area area, the shape and pressure of the vaginal detection probe are dynamically adjusted. The intelligently designed vaginal detection probe includes:

[0080] A miniature infrared thermal imager is arranged on the vaginal detection probe, and a temperature-radiation compensation function is constructed for the original infrared thermal image collected by the miniature infrared thermal imager using a pre-established coupling model of heat conduction inside the probe and surface radiation, and a corrected infrared thermal image is obtained by pixel-by-pixel correction, wherein the corrected infrared thermal image reflects the temperature distribution on the surface of the vaginal wall;

[0081] Based on the corrected infrared thermal image, a three-dimensional geometric model of the vaginal detection probe is constructed using feature points extracted from the surface of the vaginal detection probe, and the corrected infrared thermal image is aligned with the three-dimensional geometric model through rigid body transformation and non-rigid deformation to obtain a three-dimensional temperature distribution map, which includes temperature, spatial position and morphological information of the contact area;

[0082] Input the three-dimensional temperature distribution map into a pre-trained fully convolutional neural network, extract high-level semantic features of the three-dimensional temperature distribution map through multi-layer convolution and pooling operations, and output the segmentation result of the contact area between the vaginal detection probe and the vaginal wall;

[0083] The segmentation result is synchronized and aligned in time and space with the temperature distribution, the corrected infrared thermal map, the pressure distribution and the contact area of ​​the vaginal detection probe to construct a contact state data set;

[0084] Extract high-level semantic features of infrared thermal images and dynamic change features of temperature distribution and pressure distribution data in the contact state data set, and perform feature fusion to obtain the contact state feature vector;

[0085] The contact state feature vector is input into a pre-trained contact state intelligent evaluation model to predict the contact area between the vaginal detection probe and the vaginal wall. Based on the obtained contact area, combined with the pressure distribution uniformity and temperature distribution consistency, the contact state quality score is calculated. The shape and pressure of the vaginal detection probe are dynamically adjusted according to the contact state quality score to obtain an intelligently designed vaginal detection probe.

[0086] Exemplarily, a miniature infrared thermal imager is set on the vaginal detection probe, and the original infrared thermal image collected by the miniature infrared thermal imager is subjected to temperature-radiation compensation and correction. First, a temperature-radiation compensation function is constructed using a pre-established coupling model of heat conduction inside the probe and surface radiation. The coupling model takes into account the heat conduction process inside the probe and the radiation characteristics of the surface, and establishes a quantitative relationship between temperature and radiation by solving the heat conduction equation and the radiation transfer equation. Then, the original infrared thermal image is corrected pixel by pixel using the temperature-radiation compensation function to eliminate the thermal effect of the probe itself and the influence of environmental factors, and obtain a corrected infrared thermal image that accurately reflects the temperature distribution on the surface of the vaginal wall.

[0087] After obtaining the corrected infrared thermal image, the characteristic points on the surface of the vaginal detection probe are used to construct a three-dimensional geometric model of the probe. The characteristic points on the probe surface, such as corner points and edges, are extracted by three-dimensional scanning or structured light photography of the probe surface. Using these characteristic points, a three-dimensional geometric model of the probe is constructed through triangulation or surface reconstruction algorithms. Then, the corrected infrared thermal image is aligned with the three-dimensional geometric model of the probe through rigid body transformation and non-rigid deformation algorithms. Rigid body transformation includes operations such as translation and rotation, which are used to correct the position and direction differences between the infrared thermal image and the geometric model. Non-rigid deformation takes into account the deformation when the probe contacts the vaginal wall, and the infrared thermal image is accurately aligned with the deformed geometric model through algorithms such as thin plate splines. Finally, a three-dimensional temperature distribution map containing the temperature, spatial position and morphological information of the contact area is obtained.

[0088] The three-dimensional temperature distribution map is input into the pre-trained fully convolutional neural network to segment the contact area between the vaginal detection probe and the vaginal wall. The fully convolutional neural network adopts an end-to-end training method, and learns the high-level semantic features in the temperature distribution map through multi-level feature extraction of convolutional layers and pooling layers. The input of the network is the three-dimensional temperature distribution map, and the output is the category to which each pixel belongs (contact area or non-contact area). Through the back propagation algorithm and gradient descent method, the cross entropy loss function is minimized and the network parameters are optimized so that it can accurately segment the contact area.

[0089] The contact area segmentation results are synchronized and aligned in time and space with the temperature distribution of the vaginal detection probe, the corrected infrared thermal map, the pressure distribution, the contact area and other information to construct a contact state data set. Time and space synchronization aims to unify data of different modalities into the same time and space reference system to ensure data consistency and comparability. The alignment process maps data of different resolutions and scales to a unified grid through methods such as image registration and interpolation, which facilitates subsequent feature extraction and fusion.

[0090] The high-level semantic features of the infrared thermal image and the dynamic change features of the temperature distribution and pressure distribution data are extracted from the contact state data set, and the features are fused. The high-level semantic features of the infrared thermal image are extracted through a convolutional neural network to capture the local pattern and texture information of the temperature distribution. The dynamic change features of the temperature distribution and pressure distribution data are extracted through time series analysis methods, such as time domain statistical features, frequency domain features, etc., reflecting the change trend and periodicity of the data over time. The features of these different modes are fused, and a feature vector that comprehensively represents the contact state is obtained through splicing, weighting, etc.

[0091] The contact state feature vector is input into the pre-trained contact state intelligent evaluation model to predict the contact area between the vaginal detection probe and the vaginal wall. The evaluation model is trained by a supervised learning method, with the contact state feature vector as input and the manually annotated contact area as the supervision signal to learn the mapping relationship between features and area. The model can use machine learning algorithms such as support vector machines and random forests, or deep neural networks such as multi-layer perceptrons. By minimizing the mean square error loss function, the model parameters are optimized to improve the accuracy of contact area prediction. Based on the predicted contact area, the contact state quality score is calculated in combination with the uniformity of pressure distribution and the consistency of temperature distribution. The uniformity and consistency can be quantified by statistical indicators such as variance and entropy to reflect the stability and reliability of the contact state. The area, uniformity, consistency and other indicators are weighted and combined to obtain a score for comprehensive evaluation of the contact state quality.

[0092] According to the contact state quality score, the shape and pressure of the vaginal detection probe are dynamically adjusted to achieve intelligent design. By introducing intelligent materials such as shape memory alloys and electroactive polymers, or adopting flexible driving methods such as pneumatic and hydraulic, the probe can change its shape and pressure distribution in real time according to the score feedback. The optimization goal is to maximize the contact state quality score, and the optimal probe shape and pressure parameters are searched through iterative optimization algorithms such as genetic algorithms and particle swarm optimization. At the same time, factors such as material properties and mechanical constraints are considered to ensure the feasibility and safety of the probe design. Finally, an intelligent vaginal detection probe that can actively adapt to the vaginal morphology and achieve stable, uniform and reliable contact is obtained.

[0093] In this embodiment, infrared thermal imaging technology is used in combination with machine learning and intelligent optimization algorithms to achieve real-time monitoring, evaluation and feedback adjustment of the contact state between the vaginal detection probe and the vaginal wall. Through temperature-radiation compensation and correction, an accurate temperature distribution map of the vaginal wall surface is obtained. The contact area is segmented using a fully convolutional neural network, multimodal contact state features are extracted, and the contact area is predicted. Based on the contact state quality score, the probe shape and pressure are optimized to achieve intelligent adaptation of the probe and the vagina. This method improves the accuracy, reliability and comfort of vaginal detection, and provides a new technical means for the diagnosis and treatment of gynecological diseases.

[0094] In an optional embodiment,

[0095] The steps before inputting the contact state feature vector into the pre-trained contact state intelligent assessment model include:

[0096] Acquire a feature vector of the contact state between the vaginal detection probe and the vaginal wall during vaginal detection as an original sample, randomly generate a perturbation near the feature space of the original sample, and superimpose the perturbation on the original sample to obtain an adversarial sample;

[0097] The original samples and adversarial samples are input into the contact state intelligent evaluation model based on multi-task learning. The contact state intelligent evaluation model simultaneously performs the contact area regression task and the contact state quality classification task, and obtains the loss functions corresponding to the two tasks respectively. The total loss function is constructed based on the loss functions corresponding to the two tasks.

[0098] The parameters of the contact state intelligent evaluation model are fixed, and the adversarial loss function is calculated using the adversarial sample. The adversarial sample is then fixed, and the parameters of the contact state intelligent evaluation model are optimized to minimize the weighted sum of the total loss function and the adversarial loss function as the optimization goal. The parameters of the intelligent evaluation model are updated by the gradient descent algorithm, and the iteration is repeated until the model converges. In each round of iteration, a disturbance amplitude attenuation factor and a weight attenuation factor are introduced to obtain the loss function value of the contact state intelligent evaluation model on the original sample and the loss function value of the adversarial sample on the verification set. According to the difference between the two loss function values, the intensity of the adversarial training is dynamically adjusted to obtain the finally trained contact state intelligent evaluation model.

[0099] For example, first, a perturbation vector is randomly generated near the feature space of the original sample, and each element of the perturbation vector is sampled from a Gaussian distribution with a mean of zero and a small variance. Then, the generated perturbation vector is superimposed on the original sample to obtain an adversarial sample.

[0100] The original samples and the generated adversarial samples are input into the contact state intelligent evaluation model based on multi-task learning for training. The contact state intelligent evaluation model performs two tasks simultaneously: the contact area regression task and the contact state quality classification task. The contact area regression task aims to predict the actual contact area between the vaginal detection probe and the vaginal wall, and the mean square error loss function is used to measure the difference between the predicted value and the true value. The contact state quality classification task divides the contact state into excellent, good, medium, and poor levels, and the cross entropy loss function is used to measure the difference between the predicted probability distribution and the true distribution. By jointly optimizing these two tasks, the model can simultaneously learn the quantitative estimation of the contact area and the qualitative evaluation of the contact state quality, capturing the contact state characteristics of different granularities.

[0101] During the model training process, an adversarial training strategy is used to improve the robustness of the model. First, the parameters of the contact state intelligent evaluation model are fixed, and the adversarial loss function is calculated using adversarial samples. The adversarial loss function measures the performance of the model on adversarial samples, and encourages the model to learn to deal with perturbations and adversarial attacks by maximizing the prediction error of the model on adversarial samples. Then, the adversarial samples are fixed and the parameters of the contact state intelligent evaluation model are optimized. The optimization goal is to minimize the weighted sum of the total loss function and the adversarial loss function. The total loss function includes the loss functions of the contact area regression task and the contact state quality classification task. The parameters of the model are updated through a gradient descent algorithm, such as the Adam optimizer, so that it has good performance on both the original sample and the adversarial sample.

[0102] In each round of iteration, the perturbation amplitude attenuation factor and weight attenuation factor are introduced to dynamically adjust the intensity of adversarial training. The perturbation amplitude attenuation factor controls the difference between the adversarial sample and the original sample. As the training progresses, the perturbation amplitude gradually decreases, making the adversarial sample closer to the original sample. The weight attenuation factor controls the weight of the adversarial loss function in the total loss function. As the training progresses, the weight gradually decreases, making the model pay more attention to the performance of the original sample.

[0103] During the model training process, the performance of the model is regularly evaluated on the validation set to obtain the loss function values ​​of the model on the original sample and the adversarial sample. The difference between the two loss function values ​​is calculated, that is, the difference between the loss function value of the original sample and the loss function value of the adversarial sample. The larger the difference, the worse the performance of the model on the adversarial sample, and the adversarial training needs to be strengthened; the smaller the difference, the performance of the model on the adversarial sample is close to that of the original sample, and the intensity of adversarial training can be appropriately reduced. According to the size of the difference, the perturbation amplitude attenuation factor and the weight attenuation factor are dynamically adjusted to achieve automatic control of the adversarial training intensity, so that the model can achieve a balance between the original sample and the adversarial sample.

[0104] The iterative training process is repeated until the model converges, that is, the loss function value on the validation set stabilizes at a low level. Finally, an intelligent contact state evaluation model enhanced by adversarial training is obtained, which can effectively cope with disturbances and adversarial attacks and has strong robustness and generalization capabilities.

[0105] In this embodiment, adversarial samples are generated in the feature space of the original samples to challenge the decision boundary of the model, so as to encourage the model to learn more generalized and stable feature representations. At the same time, a multi-task learning framework is adopted to jointly optimize the contact area regression task and the contact state quality classification task to capture the contact state features of different granularities and improve the prediction accuracy and interpretability of the model. During the training process, the balance between model performance and robustness is achieved by dynamically adjusting the intensity of adversarial training to avoid overfitting and instability. The final contact state intelligent evaluation model can accurately predict the contact area between the vaginal detection probe and the vaginal wall, evaluate the quality level of the contact state, provide doctors with objective and reliable auxiliary diagnosis information, and improve the accuracy and effectiveness of vaginal detection.

[0106] S102. Use an intelligently designed vaginal detection probe to collect multi-source heterogeneous data during pelvic floor function testing, and standardize and synchronize the collected multi-source heterogeneous data to obtain time-synchronized detection data. At the same time, extract relevant knowledge about pelvic floor dysfunction diseases from medical literature and clinical guidelines to construct a knowledge graph of pelvic floor dysfunction diseases.

[0107] Among them, multi-source heterogeneous data include vaginal pressure signals, electromyographic signals and ultrasound image sequences.

[0108] In an optional embodiment,

[0109] The intelligently designed vaginal detection probe is used to collect multi-source heterogeneous data during the pelvic floor function detection process, and the collected multi-source heterogeneous data is standardized and synchronized to obtain the time-synchronized detection data including:

[0110] Constructing a pelvic floor function detection data set, wherein the pelvic floor function detection data set includes vaginal pressure signals, electromyographic signals, and ultrasound image sequences of multiple subjects and manually annotating them to obtain standardized parameters;

[0111] Preprocessing the vaginal pressure signal and the electromyographic signal, inputting the preprocessed vaginal pressure signal and the electromyographic signal into a one-dimensional convolutional neural network, respectively extracting multi-scale local features of the vaginal pressure signal and the electromyographic signal, and mapping the multi-scale local features into a pressure signal feature vector and an electromyographic signal feature vector through a fully connected layer; simultaneously inputting the ultrasound image sequence into a three-dimensional convolutional neural network, extracting the spatiotemporal features of the ultrasound image sequence, and mapping the spatiotemporal features into an ultrasound image feature vector through a fully connected layer;

[0112] splicing the pressure signal feature vector, the electromyographic signal feature vector and the ultrasound image feature vector to form a multimodal feature vector, inputting the multimodal feature vector into a long short-term memory network, and extracting cross-modal spatiotemporal features;

[0113] Inputting the cross-modal spatiotemporal features into a first fully connected layer and a second fully connected layer, predicting standardized parameters of vaginal pressure signals, electromyographic signals, and ultrasound image sequences through the first fully connected layer, and predicting synchronization offsets between vaginal pressure signals, electromyographic signals, and ultrasound image sequences through the second fully connected layer;

[0114] Using manually annotated standardized parameters and synchronization offsets as supervisory signals, jointly training the one-dimensional convolutional neural network, the three-dimensional convolutional neural network, the long short-term memory network, the first fully connected layer, and the second fully connected layer, optimizing the network parameters by minimizing the mean square error loss function, and obtaining a trained data processing model;

[0115] The collected multi-source heterogeneous data are input into the trained data processing model, and the standardized parameters and synchronization offset of the vaginal pressure signal, electromyographic signal and ultrasound image sequence are output. The vaginal pressure signal, electromyographic signal and ultrasound image sequence are standardized according to the standardized parameters, and the vaginal pressure signal, electromyographic signal and ultrasound image sequence are synchronized according to the synchronization offset, and finally the detection data after time synchronization is obtained.

[0116] Exemplarily, in order to construct a pelvic floor function detection data processing model, it is first necessary to construct a pelvic floor function detection data set. The data set contains vaginal pressure signals, electromyographic signals and ultrasound image sequences of multiple subjects. The vaginal pressure signal is collected by an intravaginal pressure sensor to reflect the pressure changes of the vaginal wall; the electromyographic signal is collected by a surface electromyographic electrode to reflect the contraction and relaxation of the pelvic floor muscles; the ultrasound image sequence is obtained by transvaginal ultrasound imaging to show the morphology and movement of the pelvic floor organs. The collected data is manually annotated to mark key events such as pressure peaks and electromyographic burst starting points, and standardized parameters such as maximum pressure and electromyographic integral values ​​are calculated as supervisory signals for model training.

[0117] In the data preprocessing stage, the vaginal pressure signal and electromyographic signal are filtered, denoised, normalized, and other operations are performed to remove interference and noise in the signal, making the signal smoother and more standardized. Then, the preprocessed vaginal pressure signal and electromyographic signal are input into the one-dimensional convolutional neural network respectively. The one-dimensional convolutional neural network extracts the multi-scale local features of the signal through multi-layer convolution and pooling operations. Convolution kernels of different scales can capture the patterns and changes of the signal at different time scales, such as short-term oscillations and long-term trends. Through the fully connected layer, the extracted multi-scale local features are mapped to fixed-length pressure signal feature vectors and electromyographic signal feature vectors as compact representations of the signal.

[0118] For ultrasound image sequences, a three-dimensional convolutional neural network is used for feature extraction. The three-dimensional convolutional neural network can simultaneously mine the features of the image sequence in the spatial and temporal dimensions and capture the morphological and motion information of the pelvic floor organs. Through multi-layer three-dimensional convolution and pooling operations, the network gradually extracts high-level semantic features of the image sequence, such as organ edges, motion trajectories, etc. Finally, the extracted spatiotemporal features are mapped to ultrasound image feature vectors through a fully connected layer as a compact representation of the image sequence.

[0119] The pressure signal feature vector, electromyographic signal feature vector and ultrasound image feature vector are concatenated to form a multimodal feature vector as a fusion representation of different modal information. The multimodal feature vector is input into the long short-term memory network to learn the temporal dependencies and interactions between different modalities. The long short-term memory network can selectively retain and update long-term and short-term state information through the gating mechanism, and capture the evolution law of cross-modal features in the time dimension. The output of the network is a cross-modal spatiotemporal feature that integrates the information of pressure, electromyography and ultrasound images.

[0120] In order to predict the standardized parameters of vaginal pressure signals, electromyographic signals and ultrasound image sequences and the synchronization offset between them, the cross-modal spatiotemporal features are input into two parallel fully connected layers. The first fully connected layer maps the cross-modal spatiotemporal features into standardized parameters of each modality, such as maximum pressure and electromyographic integral value, through linear transformation and activation function. The second fully connected layer predicts the synchronization offset between different modal signals, that is, their relative offset on the time axis, through regression analysis.

[0121] In the model training phase, the manually labeled normalized parameters and synchronization offsets are used as supervisory signals to jointly train the one-dimensional convolutional neural network, the three-dimensional convolutional neural network, the long short-term memory network, the first fully connected layer, and the second fully connected layer. By minimizing the mean square error loss function between the predicted value and the true value, the parameters of the network are optimized so that it can accurately predict the normalized parameters and synchronization offsets. The back propagation algorithm and the gradient descent method are used to iteratively update the weights and biases of the network, and finally a trained data processing model is obtained.

[0122] In the model application stage, newly collected multi-source heterogeneous data such as vaginal pressure signals, electromyographic signals, and ultrasound image sequences are input into the trained data processing model. Through forward propagation, the model automatically extracts the features of each modality, fuses cross-modal information, and predicts standardized parameters and synchronization offsets. According to the predicted standardized parameters, the input signals and image sequences are standardized to normalize their amplitudes, scales, etc. According to the predicted synchronization offset, the signals and image sequences of different modalities are aligned on the time axis, the time offset between them is corrected, and the time synchronization of multimodal data is achieved. Finally, standardized and synchronized vaginal pressure signals, electromyographic signals, and ultrasound image sequences are obtained, providing high-quality data support for subsequent pelvic floor function analysis and diagnosis.

[0123] In this embodiment, a multimodal pelvic floor function detection data processing model based on deep learning is constructed. The model uses a one-dimensional convolutional neural network, a three-dimensional convolutional neural network, and a long short-term memory network to extract multi-scale local features, spatiotemporal features, and cross-modal temporal dependencies of vaginal pressure signals, electromyographic signals, and ultrasound image sequences, respectively, to achieve feature learning and fusion of multi-source heterogeneous data. At the same time, through two parallel fully connected layers, the standardized parameters and synchronization offsets are predicted to achieve data standardization and synchronization. The model can automatically and efficiently process pelvic floor function detection data, improve data quality and consistency, and provide a more accurate and reliable basis for the diagnosis and treatment of pelvic floor dysfunction.

[0124] S103. Input the time-synchronized detection data into the pelvic floor dysfunction disease knowledge graph, use the graph reasoning algorithm to match the corresponding disease nodes in the pelvic floor dysfunction disease knowledge graph, calculate the confidence of each disease node, perform differential diagnosis on the disease nodes with higher confidence, and output the differential diagnosis results, and recommend personalized rehabilitation plans from the pelvic floor dysfunction disease knowledge graph based on the differential diagnosis results.

[0125] In an optional embodiment,

[0126] The time-synchronized detection data is input into the pelvic floor dysfunction disease knowledge graph, and the corresponding disease nodes are matched in the pelvic floor dysfunction disease knowledge graph using the graph reasoning algorithm, and the confidence of each disease node is calculated. The differential diagnosis of the disease nodes with higher confidence includes:

[0127] A random walk algorithm is used to generate a node sequence on the knowledge graph of pelvic floor dysfunction diseases, and the node sequence is converted into a low-dimensional dense vector through a word embedding model to obtain an embedded representation of each disease node;

[0128] The time-synchronized detection data is represented as a query vector, and the embedded representation of the disease node is used as a key vector and a value vector, and multiple groups of attention weights are calculated in parallel through a pre-trained graph neural network based on a multi-head attention mechanism to obtain a matching score between the query vector and each disease node;

[0129] Based on expert experience and prior knowledge, the evidence theory framework is used to set the initial confidence for each disease node, and the basic probability assignment of evidence is defined to represent the initial confidence level of different diseases.

[0130] Using the matching score as supporting evidence, constructing a new basic probability assignment, and calculating the correlation evidence between disease nodes using the disease relationship edges in the pelvic floor dysfunction disease knowledge graph;

[0131] The initial confidence, matching score and correlation evidence are synthesized using the combination rules in the evidence theory framework, and the basic probability assignments of multiple evidences are combined by calculating the consistency between the evidences to obtain the final confidence of each disease node;

[0132] The disease nodes are sorted according to the final confidence, a differential diagnosis is performed on the disease node with the highest confidence, and a differential diagnosis result is output.

[0133] In an optional embodiment,

[0134] The disease nodes are sorted according to the final confidence, a differential diagnosis is performed on the disease node with the highest confidence, and the differential diagnosis results are output including:

[0135] Building an intelligent diagnosis rule base based on medical expert knowledge, the intelligent diagnosis rule base includes a set of diagnosis constraint rules, each of which consists of a number of diagnosis states and corresponding legal diagnosis actions;

[0136] Divide the differential diagnosis task into multiple levels of Markov decision subprocesses, each Markov decision subprocess corresponds to a diagnostic subtask, the state space of the diagnostic subtask includes the obtained patient symptoms, signs and auxiliary examination results, and the action space includes a set of alternative diagnostic actions at the current diagnostic subtask level;

[0137] A multi-objective reward function is constructed for each Markov decision subprocess, wherein the multi-objective reward function includes a diagnostic accuracy reward, a diagnostic efficiency reward, and a patient comfort reward, and a diagnostic strategy network is trained using a deep reinforcement learning algorithm, wherein the input of the diagnostic strategy network is a state representation of the diagnostic subtask, and the output is a probability distribution of each alternative diagnostic action;

[0138] After each diagnostic strategy network generates the diagnostic action probability distribution, it determines whether the diagnostic action with the highest probability satisfies the diagnostic constraint rules in the intelligent diagnostic rule base. If so, the current diagnostic action is executed, and the obtained diagnostic action and diagnostic result are used as the state input of the next Markov decision subprocess. Otherwise, the current diagnostic strategy network is punished and the diagnostic action is regenerated.

[0139] The Markov decision subprocesses at multiple levels execute the steps of generating diagnostic actions in sequence according to the levels until all diagnostic subtasks are completed, and the results output by each diagnostic subtask are integrated to obtain the final differential diagnosis result.

[0140] For example, we first need to construct a knowledge graph of pelvic floor dysfunction diseases. This knowledge graph uses medical concepts such as diseases, symptoms, signs, and test results as nodes, and concurrent relationships between diseases, manifestation relationships between symptoms and diseases, etc. as edges to form a multi-relation heterogeneous network. On this knowledge graph, a random walk algorithm is used to generate a node sequence. The random walk algorithm generates a sequence that reflects the semantic relationship between nodes by simulating the walk of random particles on the graph. Each node sequence consists of several disease nodes, which represent the correlation and co-occurrence pattern of these diseases in medical knowledge.

[0141] In order to map discrete disease nodes to continuous vector space, the word embedding model is introduced. The word embedding model represents each disease node as a low-dimensional dense vector by learning the statistical laws of node co-occurrence, which is called the embedding representation of the disease node. In the embedding space, semantically related disease nodes will have similar vector representations, while semantically unrelated disease nodes will have very different vector representations. Through the embedding representation, the semantic similarity and correlation between disease nodes can be characterized, providing a basis for subsequent diagnostic reasoning.

[0142] After obtaining the embedded representation of the disease node, the time-synchronized pelvic floor function detection data is represented as a query vector. The query vector forms a vector representation that comprehensively describes the patient's pelvic floor function status by fusing the features of the vaginal pressure signal, electromyographic signal, and ultrasound image sequence. The embedded representation of the disease node is used as the key vector and value vector, and the matching score between the query vector and each disease node is calculated through a pre-trained graph neural network based on a multi-head attention mechanism. The multi-head attention mechanism captures the correlation between the query vector and the disease node from different semantic subspaces by calculating multiple sets of attention weights in parallel, thereby improving the matching accuracy. The higher the matching score, the stronger the correlation between the query vector and the disease node, and the greater the possibility that the patient suffers from the disease.

[0143] In order to introduce expert experience and prior knowledge, the evidence theory framework is used to set the initial confidence for each disease node. The evidence theory framework is an uncertainty reasoning method that represents the initial confidence level of different diseases by defining the basic probability assignment of evidence. The basic probability assignment is a value between 0 and 1, which represents the confidence level of the expert in a disease node. The initial confidence level can be set based on factors such as the epidemiological data of the disease and clinical experience, reflecting the guiding role of prior knowledge in diagnostic reasoning.

[0144] After obtaining the matching score between the query vector and the disease node, the matching score is used as supporting evidence to construct a new basic probability assignment. The higher the matching score, the more relevant the disease node is to the patient's test data, and the stronger the supporting evidence for the disease. At the same time, the disease relationship edges in the knowledge graph of pelvic floor dysfunction diseases are used to calculate the correlation evidence between disease nodes. The correlation evidence takes into account medical knowledge such as concurrent relationships and causal relationships between diseases, providing additional constraints and guidance for diagnostic reasoning.

[0145] Finally, the combination rules in the evidence theory framework are used to combine the initial confidence, matching score, and correlation evidence to obtain the final confidence of each disease node. The combination rules combine the basic probability assignments of multiple evidences by calculating the degree of consistency between different evidences. The higher the degree of consistency, the more consistent the support of different evidences for the disease node, and the higher the final confidence. Through evidence combination, we can comprehensively consider prior knowledge, data-driven matching scores, and correlation evidence of medical knowledge graphs to improve the accuracy and interpretability of diagnostic reasoning.

[0146] The disease nodes are sorted according to the final confidence level, and the disease node with the highest confidence level is considered to be the disease that the patient is most likely to have. The disease node with the highest confidence level is differentially diagnosed, and a comprehensive analysis is performed on the patient's clinical manifestations, examination results, etc. to determine the final diagnosis result. Finally, the differential diagnosis result is output, including the disease name, confidence level, and related medical explanations and suggestions.

[0147] In an optional embodiment,

[0148] Training a diagnostic policy network using a deep reinforcement learning algorithm involves:

[0149] Initialize the parameters of the diagnosis strategy network, randomly sample a batch of diagnosis subtask state representations, input them into the diagnosis strategy network, and calculate the probability distribution of each candidate diagnosis action through forward propagation;

[0150] According to the probability distribution of the output of the diagnostic strategy network, a set of diagnostic action samples is generated using the Monte Carlo method, and the comprehensive instant reward of each diagnostic action sample is calculated according to the multi-objective reward function. At the same time, the advantage function value of each diagnostic action sample is estimated using the generalized advantage estimation method. The loss function is constructed based on the advantage function value of the diagnostic action sample, and the parameters of the diagnostic strategy network are updated through the stochastic gradient descent algorithm. The iteration is repeated until the preset number of iterations is reached to obtain the final trained diagnostic strategy network.

[0151] Among them, the loss function calculation formula of the diagnosis strategy network is as follows:

[0152]

[0153] Where L(θ) represents the loss function, θ represents the diagnostic strategy network parameters, E represents the expectation, τ represents the trajectory of the Markov decision process, π represents the diagnostic strategy, T represents the total time step of the trajectory, γ represents the discount factor, R(·) represents the immediate reward function, λ represents the penalty coefficient for constraint violation, C(·) represents the constraint function, and s t represents the state at time step t, a t represents the action at time step t, logπ θ (at|st) indicates that the diagnostic strategy network is in state s t Next select action a t The logarithmic probability of

[0154] Among them, the Markov decision sub-processes of multiple levels execute the steps of generating diagnostic actions in sequence according to the levels until all diagnostic sub-tasks are completed, including:

[0155] Construct a reward transfer mechanism between Markov decision subprocesses to complete the coordinated optimization of multiple Markov decision subprocesses. The formula of the reward transfer mechanism is as follows:

[0156]

[0157] Among them, K l represents the cumulative reward of the Markov decision subprocess at the lth level, T l represents the total time step of the Markov decision subprocess at the lth layer, t represents the time step of the current Markov decision subprocess, η represents the discount coefficient of the immediate reward, and r l (·) represents the immediate reward of the Markov decision subprocess at the lth layer, represents the state of the Markov decision subprocess at the lth level at time step t, represents the action of the Markov decision subprocess at the lth level at time step t, V l+1 (·) represents the state value function of the l+1th layer Markov decision subprocess, and s1 represents the initial state.

[0158] In this embodiment, the random walk algorithm and word embedding model are used to learn the semantic representation of disease nodes on the knowledge graph of pelvic floor dysfunction diseases, and to characterize the correlation and co-occurrence pattern between diseases. Through the graph neural network based on the multi-head attention mechanism, the matching score of the time-synchronized detection data and the disease node is calculated to capture the correlation between the patient's pelvic floor function status and the disease. The evidence theory framework is introduced to comprehensively consider the expert experience, data-driven matching scores, and the correlation evidence of the knowledge graph, and the final confidence of each disease node is obtained through the combination of evidence. Diseases are sorted and differentially diagnosed according to the confidence, and the diagnosis results and medical explanations are output. This method combines medical knowledge, deep learning, and reasoning strategies to provide an intelligent and personalized diagnostic solution to help the early diagnosis and precise treatment of pelvic floor dysfunction diseases.

[0159] Figure 2 FIG. 1 is a schematic diagram of the structure of a pelvic and abdominal dynamics detection system based on artificial intelligence according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0160] The first unit is used to arrange a thermistor sensor array on the surface of the vaginal detection probe, collect temperature data in real time during the vaginal detection process, identify the contact area between the vaginal detection probe and the vaginal wall according to the temperature data, and calculate the contact area area, and dynamically adjust the shape and pressure of the vaginal detection probe according to the calculated contact area area to obtain an intelligently designed vaginal detection probe;

[0161] The second unit is used to collect multi-source heterogeneous data in the process of pelvic floor function detection using an intelligently designed vaginal detection probe, and to standardize and synchronize the collected multi-source heterogeneous data to obtain time-synchronized detection data. At the same time, relevant knowledge about pelvic floor dysfunction diseases is extracted from medical literature and clinical guidelines to construct a knowledge graph of pelvic floor dysfunction diseases.

[0162] The third unit is used to input the time-synchronized detection data into the pelvic floor dysfunction disease knowledge graph, use the graph reasoning algorithm to match the corresponding disease nodes in the pelvic floor dysfunction disease knowledge graph, calculate the confidence of each disease node, perform differential diagnosis on the disease nodes with higher confidence, and output the differential diagnosis results, and recommend personalized rehabilitation plans from the pelvic floor dysfunction disease knowledge graph according to the differential diagnosis results.

[0163] According to a third aspect of the embodiments of the present invention,

[0164] An electronic device is provided, comprising:

[0165] processor;

[0166] a memory for storing processor-executable instructions;

[0167] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0168] According to a fourth aspect of the embodiments of the present invention,

[0169] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0170] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0171] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting pelvic and abdominal dynamics based on artificial intelligence, characterized in that: include: Arrange a thermistor sensor array on the surface of the vaginal detection probe, collect temperature data in real time during the vaginal detection process, identify the contact area between the vaginal detection probe and the vaginal wall according to the temperature data, and calculate the contact area area, and dynamically adjust the shape and pressure of the vaginal detection probe according to the calculated contact area area to obtain an intelligently designed vaginal detection probe; The multi-source heterogeneous data collected during the pelvic floor function test are collected using an intelligently designed vaginal detection probe, and the collected multi-source heterogeneous data are standardized and synchronized to obtain time-synchronized test data. At the same time, relevant knowledge about pelvic floor dysfunction diseases is extracted from medical literature and clinical guidelines to construct a knowledge graph of pelvic floor dysfunction diseases. Input the time-synchronized detection data into the pelvic floor dysfunction disease knowledge graph, use the graph reasoning algorithm to match the corresponding disease nodes in the pelvic floor dysfunction disease knowledge graph, calculate the confidence of each disease node, perform differential diagnosis on the disease nodes with higher confidence, output the differential diagnosis results, and recommend personalized rehabilitation plans from the pelvic floor dysfunction disease knowledge graph according to the differential diagnosis results; The intelligently designed vaginal detection probe is used to collect multi-source heterogeneous data during the pelvic floor function detection process, and the collected multi-source heterogeneous data is standardized and synchronized to obtain the time-synchronized detection data including: Constructing a pelvic floor function detection data set, wherein the pelvic floor function detection data set includes vaginal pressure signals, electromyographic signals, and ultrasound image sequences of multiple subjects and manually annotating them to obtain standardized parameters; Preprocessing the vaginal pressure signal and the electromyographic signal, inputting the preprocessed vaginal pressure signal and the electromyographic signal into a one-dimensional convolutional neural network, respectively extracting multi-scale local features of the vaginal pressure signal and the electromyographic signal, and mapping the multi-scale local features into a pressure signal feature vector and an electromyographic signal feature vector through a fully connected layer; simultaneously inputting the ultrasound image sequence into a three-dimensional convolutional neural network, extracting the spatiotemporal features of the ultrasound image sequence, and mapping the spatiotemporal features into an ultrasound image feature vector through a fully connected layer; splicing the pressure signal feature vector, the electromyographic signal feature vector and the ultrasound image feature vector to form a multimodal feature vector, inputting the multimodal feature vector into a long short-term memory network, and extracting cross-modal spatiotemporal features; Inputting the cross-modal spatiotemporal features into a first fully connected layer and a second fully connected layer, predicting standardized parameters of vaginal pressure signals, electromyographic signals, and ultrasound image sequences through the first fully connected layer, and predicting synchronization offsets between vaginal pressure signals, electromyographic signals, and ultrasound image sequences through the second fully connected layer; Using manually annotated standardized parameters and synchronization offsets as supervisory signals, jointly training the one-dimensional convolutional neural network, the three-dimensional convolutional neural network, the long short-term memory network, the first fully connected layer, and the second fully connected layer, optimizing the network parameters by minimizing the mean square error loss function, and obtaining a trained data processing model; The collected multi-source heterogeneous data are input into the trained data processing model, and the standardized parameters and synchronization offset of the vaginal pressure signal, electromyographic signal and ultrasound image sequence are output. The vaginal pressure signal, electromyographic signal and ultrasound image sequence are standardized according to the standardized parameters, and the vaginal pressure signal, electromyographic signal and ultrasound image sequence are synchronized according to the synchronization offset, and finally the detection data after time synchronization is obtained.

2. The method according to claim 1, characterized in that An array of thermistor sensors is arranged on the surface of the vaginal detection probe, and temperature data is collected in real time during the vaginal detection process. The contact area between the vaginal detection probe and the vaginal wall is identified according to the temperature data, and the contact area area is calculated. According to the calculated contact area area, the shape and pressure of the vaginal detection probe are dynamically adjusted. The intelligently designed vaginal detection probe includes: A miniature infrared thermal imager is arranged on the vaginal detection probe, and a temperature-radiation compensation function is constructed for the original infrared thermal image collected by the miniature infrared thermal imager using a pre-established coupling model of heat conduction inside the probe and surface radiation, and a corrected infrared thermal image is obtained by pixel-by-pixel correction, wherein the corrected infrared thermal image reflects the temperature distribution on the surface of the vaginal wall; Based on the corrected infrared thermal image, a three-dimensional geometric model of the vaginal detection probe is constructed using feature points extracted from the surface of the vaginal detection probe, and the corrected infrared thermal image is aligned with the three-dimensional geometric model through rigid body transformation and non-rigid deformation to obtain a three-dimensional temperature distribution map, which includes temperature, spatial position and morphological information of the contact area; Input the three-dimensional temperature distribution map into a pre-trained fully convolutional neural network, extract high-level semantic features of the three-dimensional temperature distribution map through multi-layer convolution and pooling operations, and output the segmentation result of the contact area between the vaginal detection probe and the vaginal wall; The segmentation result is synchronized and aligned in time and space with the temperature distribution, the corrected infrared thermal map, the pressure distribution and the contact area of ​​the vaginal detection probe to construct a contact state data set; Extract high-level semantic features of infrared thermal images and dynamic change features of temperature distribution and pressure distribution data in the contact state data set, and perform feature fusion to obtain the contact state feature vector; The contact state feature vector is input into a pre-trained contact state intelligent evaluation model to predict the contact area between the vaginal detection probe and the vaginal wall. Based on the obtained contact area, combined with the pressure distribution uniformity and temperature distribution consistency, the contact state quality score is calculated. The shape and pressure of the vaginal detection probe are dynamically adjusted according to the contact state quality score to obtain an intelligently designed vaginal detection probe.

3. The method according to claim 2, characterized in that The steps before inputting the contact state feature vector into the pre-trained contact state intelligent assessment model include: Acquire a feature vector of the contact state between the vaginal detection probe and the vaginal wall during vaginal detection as an original sample, randomly generate a perturbation near the feature space of the original sample, and superimpose the perturbation on the original sample to obtain an adversarial sample; The original samples and adversarial samples are input into the contact state intelligent evaluation model based on multi-task learning. The contact state intelligent evaluation model simultaneously performs the contact area regression task and the contact state quality classification task, and obtains the loss functions corresponding to the two tasks respectively. The total loss function is constructed based on the loss functions corresponding to the two tasks. The parameters of the contact state intelligent evaluation model are fixed, and the adversarial loss function is calculated using the adversarial sample. The adversarial sample is then fixed, and the parameters of the contact state intelligent evaluation model are optimized to minimize the weighted sum of the total loss function and the adversarial loss function as the optimization goal. The parameters of the intelligent evaluation model are updated by the gradient descent algorithm, and the iteration is repeated until the model converges. In each round of iteration, a disturbance amplitude attenuation factor and a weight attenuation factor are introduced to obtain the loss function value of the contact state intelligent evaluation model on the original sample and the loss function value of the adversarial sample on the verification set. According to the difference between the two loss function values, the intensity of the adversarial training is dynamically adjusted to obtain the finally trained contact state intelligent evaluation model.

4. The method according to claim 1, characterized in that The time-synchronized detection data is input into the pelvic floor dysfunction disease knowledge graph, and the corresponding disease nodes are matched in the pelvic floor dysfunction disease knowledge graph using the graph reasoning algorithm, and the confidence of each disease node is calculated. The differential diagnosis of the disease nodes with higher confidence includes: A random walk algorithm is used to generate a node sequence on the knowledge graph of pelvic floor dysfunction diseases, and the node sequence is converted into a low-dimensional dense vector through a word embedding model to obtain an embedded representation of each disease node; The time-synchronized detection data is represented as a query vector, and the embedded representation of the disease node is used as a key vector and a value vector, and multiple groups of attention weights are calculated in parallel through a pre-trained graph neural network based on a multi-head attention mechanism to obtain a matching score between the query vector and each disease node; Based on expert experience and prior knowledge, the evidence theory framework is used to set the initial confidence for each disease node, and the basic probability assignment of evidence is defined to represent the initial confidence level of different diseases. Using the matching score as supporting evidence, constructing a new basic probability assignment, and calculating the correlation evidence between disease nodes using the disease relationship edges in the pelvic floor dysfunction disease knowledge graph; The initial confidence, matching score and correlation evidence are synthesized using the combination rules in the evidence theory framework, and the basic probability assignments of multiple evidences are combined by calculating the consistency between the evidences to obtain the final confidence of each disease node; The disease nodes are sorted according to the final confidence, a differential diagnosis is performed on the disease node with the highest confidence, and a differential diagnosis result is output.

5. The method according to claim 4, characterized in that The disease nodes are sorted according to the final confidence, a differential diagnosis is performed on the disease node with the highest confidence, and the differential diagnosis results are output including: Building an intelligent diagnosis rule base based on medical expert knowledge, the intelligent diagnosis rule base includes a set of diagnosis constraint rules, each of which consists of a number of diagnosis states and corresponding legal diagnosis actions; Divide the differential diagnosis task into multiple levels of Markov decision subprocesses, each Markov decision subprocess corresponds to a diagnostic subtask, the state space of the diagnostic subtask includes the obtained patient symptoms, signs and auxiliary examination results, and the action space includes a set of alternative diagnostic actions at the current diagnostic subtask level; A multi-objective reward function is constructed for each Markov decision subprocess, wherein the multi-objective reward function includes a diagnostic accuracy reward, a diagnostic efficiency reward, and a patient comfort reward, and a diagnostic strategy network is trained using a deep reinforcement learning algorithm, wherein the input of the diagnostic strategy network is a state representation of the diagnostic subtask, and the output is a probability distribution of each alternative diagnostic action; After each diagnostic strategy network generates the diagnostic action probability distribution, it determines whether the diagnostic action with the highest probability satisfies the diagnostic constraint rules in the intelligent diagnostic rule base. If so, the current diagnostic action is executed, and the obtained diagnostic action and diagnostic result are used as the state input of the next Markov decision subprocess. Otherwise, the current diagnostic strategy network is punished and the diagnostic action is regenerated. The Markov decision subprocesses at multiple levels execute the steps of generating diagnostic actions in sequence according to the levels until all diagnostic subtasks are completed, and the results output by each diagnostic subtask are integrated to obtain the final differential diagnosis result.

6. The method according to claim 5, characterized in that Training a diagnostic policy network using a deep reinforcement learning algorithm involves: Initialize the parameters of the diagnosis strategy network, randomly sample a batch of diagnosis subtask state representations, input them into the diagnosis strategy network, and calculate the probability distribution of each candidate diagnosis action through forward propagation; According to the probability distribution of the output of the diagnostic strategy network, a set of diagnostic action samples is generated using the Monte Carlo method, and the comprehensive instant reward of each diagnostic action sample is calculated according to the multi-objective reward function. At the same time, the advantage function value of each diagnostic action sample is estimated using the generalized advantage estimation method. The loss function is constructed based on the advantage function value of the diagnostic action sample, and the parameters of the diagnostic strategy network are updated through the stochastic gradient descent algorithm. The iteration is repeated until the preset number of iterations is reached to obtain the final trained diagnostic strategy network. Among them, the loss function calculation formula of the diagnosis strategy network is as follows: Where L(θ) represents the loss function, θ represents the diagnostic strategy network parameters, E represents the expectation, τ represents the trajectory of the Markov decision process, π represents the diagnostic strategy, T represents the total time step of the trajectory, γ represents the discount factor, R(·) represents the immediate reward function, λ represents the penalty coefficient for constraint violation, C(·) represents the constraint function, and s t represents the state at time step t, a t represents the action at time step t, logπ θ (at|st) indicates that the diagnostic strategy network is in state s t Next select action a t The logarithmic probability of Among them, the Markov decision sub-processes of multiple levels execute the steps of generating diagnostic actions in sequence according to the levels until all diagnostic sub-tasks are completed, including: Construct a reward transfer mechanism between Markov decision subprocesses to complete the coordinated optimization of multiple Markov decision subprocesses. The formula of the reward transfer mechanism is as follows: Among them, K l represents the cumulative reward of the Markov decision subprocess at the lth level, T l represents the total time step of the Markov decision subprocess at the lth layer, t represents the time step of the current Markov decision subprocess, η represents the discount coefficient of the immediate reward, and r l (·) represents the immediate reward of the Markov decision subprocess at the lth layer, represents the state of the Markov decision subprocess at the lth level at time step t, represents the action of the Markov decision subprocess at the lth level at time step t, V l+1 (·) represents the state value function of the l+1th layer Markov decision subprocess, and s1 represents the initial state.

7. An artificial intelligence-based pelvic and abdominal dynamics detection system, used to implement the method described in any one of claims 1 to 6, characterized in that: include: The first unit is used to arrange a thermistor sensor array on the surface of the vaginal detection probe, collect temperature data in real time during the vaginal detection process, identify the contact area between the vaginal detection probe and the vaginal wall according to the temperature data, and calculate the contact area area, and dynamically adjust the shape and pressure of the vaginal detection probe according to the calculated contact area area to obtain an intelligently designed vaginal detection probe; The second unit is used to collect multi-source heterogeneous data in the process of pelvic floor function detection using an intelligently designed vaginal detection probe, and to standardize and synchronize the collected multi-source heterogeneous data to obtain time-synchronized detection data. At the same time, relevant knowledge about pelvic floor dysfunction diseases is extracted from medical literature and clinical guidelines to construct a knowledge graph of pelvic floor dysfunction diseases. The third unit is used to input the time-synchronized detection data into the pelvic floor dysfunction disease knowledge graph, use the graph reasoning algorithm to match the corresponding disease nodes in the pelvic floor dysfunction disease knowledge graph, calculate the confidence of each disease node, perform differential diagnosis on the disease nodes with higher confidence, and output the differential diagnosis results, and recommend personalized rehabilitation plans from the pelvic floor dysfunction disease knowledge graph according to the differential diagnosis results.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images