CN-122025096-A - Multi-mode intelligent evaluation method and system for pelvic floor dysfunction

CN122025096ACN 122025096 ACN122025096 ACN 122025096ACN-122025096-A

Abstract

The application relates to the technical field of medical treatment, and discloses a multi-mode intelligent evaluation method and system for pelvic floor dysfunction. Three spans from static evaluation to dynamic prediction, single diagnosis to personalized intervention, and offline model to online evolution are provided. The accuracy, the interpretability and the clinical applicability of pelvic floor dysfunction assessment are improved, so that the system can identify the compensation state, predict rehabilitation response and generate a feasible scheme, and finally the advantages of misdiagnosis rate and excessive treatment risk are reduced, and the method is used for solving the problems that the verification gold collection standard is mainly dependent on ultrasound and inquiry results, so that the model is better in performance in verification after learning bias and forms a self-reinforced bias closed loop.

Inventors

ZHANG XIANFENG
LV BO
GUAN LIWEI
LIU DUN
Cao changdong
CHEN SHI
ZHANG NING

Assignees

中国人民解放军联勤保障部队第九六七医院

Dates

Publication Date: 20260512
Application Date: 20260129

Claims (10)

1. A multi-mode intelligent evaluation method for pelvic floor dysfunction is characterized by comprising the following specific steps: Step S1, multi-mode acquisition and credibility quantification are carried out, namely synchronously acquiring an ultrasonic image of the pelvic floor, electromyographic signal data, urine dynamic data, questionnaire data and medical history data, recording acquisition condition parameters, carrying out signal-to-noise ratio analysis and quality grading on the data by adopting a confidence evaluation network, and generating a credibility weight matrix; S2, extracting static structural features, namely performing anatomical structure segmentation on the ultrasonic image, extracting bladder neck position and levator ani muscle split hole area parameters, quantifying parameter uncertainty through a probabilistic deep learning method, and generating a structural feature vector containing a confidence interval; S3, dynamic functional characteristic extraction, namely performing time-frequency domain analysis on the electromyographic signals, extracting basic characteristics of resting tension and maximum contraction force, identifying contraction coordination indexes and fatigue recovery slope time sequence modes, and performing pressure flow rate collaborative analysis on a urine dynamic curve to generate functional characteristic vectors; s4, feature decoupling processing, namely constructing a cross-modal countermeasure learning frame, setting a structure and function decoupling discriminator, restraining ultrasonic features to only contain morphological information and myoelectric features to only contain functional information through a gradient inversion layer, and realizing feature space decoupling; S5, causal constraint fusion, namely, introducing a pelvic floor function causal graph network, defining the pelvic floor function as a causal link from structural damage to function compensation to symptom expression, setting a causal gating unit at a fusion layer, and adopting a structural equation model to constrain fusion weights; S6, designing a loss function, namely constructing a multi-task loss function comprising diagnosis consistency loss, rehabilitation response prediction loss and long-term risk prediction loss, and balancing short-term accuracy and long-term predictability through a learnable weight; Step S7, generating individuation parameters, namely generating executable rehabilitation parameters through a strategy network based on an evaluation output result and combining patient compliance history and living environment constraint conditions, and outputting expected effects and risk prompts; And S8, feeding back the actual execution effect of the rehabilitation training to the model as a new label, adopting online migration learning to update the causal graph structure and the characteristic weight, setting a model performance monitoring module, and triggering local retraining when the evaluation deviation of the specific crowd continuously increases.
2. The method for multi-modal intelligent assessment of pelvic floor dysfunction according to claim 1, wherein in step S1, the specific steps of multi-modal collection and reliability quantification are as follows: S1.1, acquiring pelvic floor anatomical structure image data, pelvic floor surface myoelectric signal data and urodynamic pressure flow rate data under the same time reference by a pelvic floor ultrasonic imaging device, a surface myoelectric acquisition device and a urodynamic detection device, synchronously reading structured symptom questionnaire data and fertility history text data, recording ultrasonic probe space orientation parameters, myoelectric electrode skin contact impedance parameters and bladder filling degree parameters, and marking and aligning various data according to time stamps to form an original multi-mode data set; S1.2, inputting the original multi-modal data set into a pre-training confidence evaluation network, calculating the signal-to-noise ratio metric value of each data sample, dividing the credibility level of the data sample according to a preset quality grading threshold, distributing credibility weight coefficients for each sample, and constructing a credibility weight matrix corresponding to the sample dimension of the original multi-modal data set.
3. The method for multi-modal intelligent assessment of pelvic floor dysfunction according to claim 1, wherein in step S2, the specific steps of static structural feature extraction are as follows: s2.1, performing anatomical structure segmentation on the ultrasonic image through a pre-training segmentation network, identifying the boundary between the bladder neck contour and the levator ani muscle split hole, calculating the bladder neck position coordinate and the levator ani muscle split hole area measurement value, and generating a structure parameter set; S2.2, inputting the structure parameter set into an uncertainty quantization module, calculating posterior distribution of each parameter by adopting a Bayesian inference method, determining upper and lower limits of confidence intervals, and constructing a structure feature vector consisting of parameter measurement values and confidence intervals thereof.
4. The method for multi-modal intelligent assessment of pelvic floor dysfunction according to claim 1, wherein in step S3, the dynamic function feature extraction is specifically as follows: s3.1, processing the electromyographic signals through a time-frequency conversion method, extracting basic characteristic parameters of resting tension and maximum contraction force, identifying a contraction coordination index and a fatigue recovery slope high-order time sequence mode, and generating an electromyographic characteristic parameter set; S3.2, extracting a urodynamics curve, analyzing a pressure and flow rate cooperative mode of the urodynamics curve, extracting cooperative characteristic parameters, combining myoelectricity characteristic parameter sets and urodynamics cooperative characteristic parameters, and constructing a dynamic function characteristic vector.
5. The method for multi-modal intelligent assessment of pelvic floor dysfunction according to claim 1, wherein in step S4, the characteristic decoupling process specifically comprises the following steps: S4.1, constructing a cross-modal countermeasure training architecture, configuring a structural feature discriminator and a functional feature discriminator, distinguishing whether an input feature is derived from an ultrasonic mode or a myoelectric mode, and judging whether the feature contains cross-modal redundant information; S4.2, restraining ultrasonic characteristics through a gradient inversion mechanism only maintains morphological information, restraining myoelectric characteristics only maintains functional information, enabling the two types of characteristics to realize information decoupling in a characteristic space, and eliminating hidden interference of a static structure on dynamic functions.
6. The method for multi-modal intelligent assessment of pelvic floor dysfunction according to claim 1, wherein in step S5, the causal constraint fusion is specifically as follows: s5.1, constructing a basin bottom function causal graph network, characterizing the basin bottom function state as a causal conduction path of structural damage, function compensation and symptom expression, defining condition dependency relations among nodes, and generating causal structure priori knowledge; And S5.2, configuring a causal gating unit to be deployed in a feature fusion layer, calculating the contribution degree of each modal feature to causal nodes by adopting a structural equation model, and restricting fusion weight distribution according to prior knowledge of the causal structure.
7. The method for multi-modal intelligent assessment of pelvic floor dysfunction according to claim 1, wherein in step S6, the loss function design is specifically as follows: s6.1, constructing a first sub-loss function for diagnosing consistency constraint, constructing a second sub-loss function for rehabilitation response prediction constraint, constructing a third sub-loss function for long-term risk prediction constraint, and combining the three sub-loss functions in a weighting manner to form a multi-task loss function; S6.2, setting a leachable weight parameter, dynamically adjusting the weight distribution of the first sub-loss function, the second sub-loss function and the third sub-loss function, and balancing the optimization target between the short-term diagnosis accuracy and the long-term predictability.
8. The method for multi-modal intelligent assessment of pelvic floor dysfunction according to claim 1, wherein in step S7, the individualizing parameters are generated as follows: S7.1, performing association mapping on the evaluation output result, the patient compliance historical data and the living environment constraint parameters to construct a strategy network input constraint condition set; S7.2, inputting the constraint condition set into a strategy network to generate rehabilitation training parameters, and synchronously outputting expected effect evaluation data and potential risk warning data corresponding to the parameters.
9. The method for multi-modal intelligent assessment of pelvic floor dysfunction according to claim 1, wherein in step S8, the closed-loop feedback mechanism specifically comprises the following steps: S8.1, collecting actual execution effect data of rehabilitation training, feeding the actual execution effect data back to a model as a new label, and updating a causal graph network structure and characteristic weight parameters by adopting online transfer learning; and S8.2, configuring a model performance monitoring module, continuously monitoring the evaluation deviation change trend of the specific crowd, and automatically triggering the model local retraining process when the deviation continuously increases and exceeds a preset threshold value.
10. The multi-mode intelligent evaluation system for pelvic floor dysfunction is characterized by being used for executing the multi-mode intelligent evaluation method for pelvic floor dysfunction according to any one of the claims 1-9, and comprises a data acquisition module, a structure extraction module, a function extraction module, a characteristic decoupling module, a causal fusion module, a loss optimization module, a parameter generation module and a closed loop evolution module; the data acquisition module synchronously acquires pelvic floor ultrasound, electromyographic signal data, urine dynamic data, questionnaire data and medical history data, records acquisition condition parameters, adopts a confidence evaluation network to perform signal-to-noise ratio analysis and quality classification, and generates a confidence weight matrix; The structure extraction module is used for carrying out anatomical structure segmentation on the ultrasonic image, extracting bladder neck position and levator ani muscle split hole area parameters, quantifying parameter uncertainty through a probabilistic deep learning method and generating a structural feature vector containing a confidence interval; The function extraction module is used for carrying out time-frequency domain analysis on the electromyographic signals, extracting basic characteristics of resting tension and maximum contraction force, identifying a contraction coordination index and a fatigue recovery slope time sequence mode, carrying out pressure flow rate collaborative analysis on a urine dynamic curve, and generating a function feature vector; the feature decoupling module is used for constructing a cross-modal countermeasure learning frame, arranging a structure and function decoupling discriminator, restraining ultrasonic features to only contain morphological information and myoelectric features to only contain functional information through the gradient inversion layer, and realizing feature space decoupling; The causal fusion module is used for introducing a pelvic floor function causal graph network, defining the pelvic floor function as a causal link from structural damage to function compensation to symptom expression, arranging a causal gating unit at a fusion layer, and adopting a structural equation model to restrict fusion weight; The loss optimization module is used for constructing a multi-task loss function comprising diagnostic consistency loss, rehabilitation response prediction loss and long-term risk prediction loss, and balancing short-term accuracy and long-term predictability through a learnable weight; The parameter generation module is used for generating executable rehabilitation parameters through a strategy network based on the evaluation output result and combining the patient compliance history and the living environment constraint condition, and outputting expected effects and risk prompts; and the closed-loop evolution module is used for feeding back the actual execution effect of the rehabilitation training to the model as a new label, updating the causal graph structure and the characteristic weight by adopting online migration learning, setting the model performance monitoring module, and triggering local retraining when the evaluation deviation of the specific crowd continuously increases.

Description

Multi-mode intelligent evaluation method and system for pelvic floor dysfunction Technical Field The application relates to the technical field of a multi-mode intelligent evaluation method for pelvic floor dysfunction, in particular to a multi-mode intelligent evaluation method and system for pelvic floor dysfunction. Background The multi-mode intelligent evaluation of pelvic floor dysfunction is mainly applied to the crossing field of gynaecology urology and rehabilitation medicine, and aims at the problems of pelvic floor support structure damage and functional degradation caused by factors such as gestational delivery, aging and the like of postpartum and middle-aged and elderly women, and the cooperative quantitative analysis is carried out on the pelvic floor structure form and the functional state through a deep learning algorithm so as to realize the accurate typing and severity grading of the pelvic floor dysfunction. However, in the prior art, time sequence dynamic data such as myoelectricity and urine dynamic can be synchronously acquired, but the systematic defect that a dynamic function is implicitly hijacked by a static structure exists in the model training process. Because the ultrasonic morphological parameters have high signal-to-noise ratio and strong stability, the deep learning model spontaneously gives an excessively high decision weight in the gradient optimization process, and the myoelectricity functional characteristics are secondary factors due to large individual differences and strong signal fluctuation. In addition, the verification collection standard is mainly dependent on ultrasonic and inquiry results, so that the model is better in performance in verification after 'learning bias', and a self-strengthening bias closed loop is formed. Disclosure of Invention The application provides a multi-mode intelligent evaluation method and a system for pelvic floor dysfunction, which have triple spans from static evaluation to dynamic prediction, single diagnosis to individual intervention and offline model to online evolution. The accuracy, the interpretability and the clinical applicability of pelvic floor dysfunction assessment are improved, so that the system can identify the compensation state, predict rehabilitation response and generate a feasible scheme, and finally the advantages of misdiagnosis rate and excessive treatment risk are reduced, and the method is used for solving the problems that the verification gold collection standard is mainly dependent on ultrasound and inquiry results, so that the model is better in performance in verification after learning bias and forms a self-reinforced bias closed loop. In order to achieve the purpose, the application adopts the following technical scheme that the multi-mode intelligent evaluation method for pelvic floor dysfunction comprises the following specific steps: Step S1, multi-mode acquisition and credibility quantification are carried out, namely synchronously acquiring pelvic floor ultrasound, electromyographic signal data, urine dynamic data, questionnaire data and medical history data, recording acquisition condition parameters, carrying out signal-to-noise ratio analysis and quality grading on the data by adopting a confidence evaluation network, and generating a credibility weight matrix; S2, extracting static structural features, namely performing anatomical structure segmentation on the ultrasonic image, extracting bladder neck position and levator ani muscle split hole area parameters, quantifying parameter uncertainty through a probabilistic deep learning method, and generating a structural feature vector containing a confidence interval; S3, dynamic functional characteristic extraction, namely performing time-frequency domain analysis on the electromyographic signals, extracting basic characteristics of resting tension and maximum contraction force, identifying contraction coordination indexes and fatigue recovery slope time sequence modes, and performing pressure flow rate collaborative analysis on a urine dynamic curve to generate functional characteristic vectors; s4, feature decoupling processing, namely constructing a cross-modal countermeasure learning frame, setting a structure and function decoupling discriminator, restraining ultrasonic features to only contain morphological information and myoelectric features to only contain functional information through a gradient inversion layer, and realizing feature space decoupling; S5, causal constraint fusion, namely, introducing a pelvic floor function causal graph network, defining the pelvic floor function as a causal link from structural damage to function compensation to symptom expression, setting a causal gating unit at a fusion layer, and adopting a structural equation model to constrain fusion weights; S6, designing a loss function, namely constructing a multi-task loss function comprising diagnosis consistency loss, rehabilitation response prediction l