A method and system for improving quality of hysteroscopic surgery image data

Abstract

The application provides a quality improvement method and system for hysteroscopic surgery image data, and relates to the technical field of image processing. The method comprises the following steps: acquiring a hysteroscope image of a target patient at a specific time point, and acquiring multi-modal physiological time series data of the target patient before and after the specific time point; performing time series feature fusion analysis on the multi-modal physiological time series data to obtain individual physiological cycle phase information corresponding to the specific time point; inputting the individual physiological cycle phase information into a pre-trained healthy uterine cavity image generation intelligent agent to output a standard healthy uterine cavity reference image corresponding to the specific time point; comparing the hysteroscope image with the standard healthy uterine cavity reference image, generating a pathological residual feature map, and enhancing the display of the pathological residual feature map. The technical problem that abnormal features are not prominent in the existing hysteroscope image is solved.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method and system for improving the quality of hysteroscopic surgical image data. Background Technology

[0002] In the clinical diagnosis and treatment of hysteroscopic surgery, hysteroscopic images are the core imaging basis for doctors to carry out the diagnosis of uterine lesions and surgical operations. However, the image quality has obvious limitations and is difficult to meet the actual needs of precise clinical diagnosis and treatment.

[0003] This is because the normal physiological morphology of a woman's uterine cavity undergoes dynamic physiological changes with the individual's menstrual cycle. These physiological changes are easily confused with pathological abnormalities, making it impossible for static hysteroscopic images to accurately distinguish between normal physiological structures and pathological abnormal areas. At the same time, the visual recognition of tiny abnormal areas in hysteroscopic images is low, which in turn makes the abnormal features of hysteroscopic images not stand out. Summary of the Invention

[0004] This invention addresses the technical problems in existing technologies, such as the dynamic physiological changes in the normal physiological morphology of the female uterine cavity with the individual's menstrual cycle, and the low visual recognition of abnormal areas in the original hysteroscopic surgical images, which result in significant physiological interference, blurred abnormal features, insufficient effective detail recognition, and poor overall visualization quality in hysteroscopic images. The invention provides a method and system for improving the quality of hysteroscopic surgical image data.

[0005] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:

[0006] In a first aspect, the present invention provides a method for improving the quality of hysteroscopic surgical image data, comprising:

[0007] Acquire hysteroscopic images of the target patient at a specific time point, and acquire multimodal physiological time-series data of the target patient before and after the specific time point;

[0008] The multimodal physiological time-series data are subjected to time-series feature fusion analysis to obtain individual physiological cycle phase information corresponding to the specific time point;

[0009] The individual physiological cycle phase information is input into a pre-trained healthy uterine cavity image generation agent, and a standard healthy uterine cavity reference image corresponding to the specific time point is output.

[0010] The hysteroscopic images are compared with the standard healthy uterine cavity reference images to generate pathological residual feature maps, and the pathological residual feature maps are enhanced for display.

[0011] Secondly, the present invention provides a system for improving the quality of hysteroscopic surgical image data, comprising:

[0012] The data acquisition module is used to acquire hysteroscopic images of the target patient at a specific time point, and to acquire multimodal physiological time-series data of the target patient before and after the specific time point.

[0013] The phase information calculation module is used to perform temporal feature fusion analysis on the multimodal physiological time series data to obtain the individual physiological cycle phase information corresponding to the specific time point;

[0014] The standard image generation module is used to input the individual's physiological cycle phase information into a pre-trained healthy uterine cavity image generation agent and output a standard healthy uterine cavity reference image corresponding to the specific time point.

[0015] The contrast enhancement module is used to compare the hysteroscopic image with the standard healthy uterine cavity reference image, generate a pathological residual feature map, and enhance the display of the pathological residual feature map.

[0016] The beneficial effects of this invention are:

[0017] Compared to existing technologies, this application first obtains hysteroscopic images of the target patient at a specific time point and multimodal physiological time-series data before and after that specific time point, providing complete and relevant basic data support for subsequent image quality improvement processing, ensuring the targeting and accuracy of subsequent processing; secondly, it performs time-series feature fusion analysis on the multimodal physiological time-series data to obtain individual physiological cycle phase information, which can accurately capture the current intrauterine physiological state of the target patient, providing a basis for avoiding image interference caused by the physiological cycle; thirdly, it inputs the individual physiological cycle phase information into a pre-trained healthy intrauterine image generation agent, outputting a standard healthy intrauterine reference image corresponding to a specific time point, providing a precise individualized healthy control benchmark for hysteroscopic image quality optimization; finally, it generates a pathological residual feature map by comparing the hysteroscopic image with the standard healthy intrauterine reference image, and enhances the display of the pathological residual feature map, which can effectively eliminate normal morphological interference caused by the physiological cycle, clearly highlight the details of abnormal areas in the hysteroscopic image, and improve the visualization and recognition of abnormal features.

[0018] Through the above technical solutions, this application establishes a complete image quality improvement process, from basic data acquisition, physiological interference avoidance, healthy control construction to abnormal feature enhancement. This effectively reduces the interference of dynamic changes in individual physiological cycles on hysteroscopic surgical images, improves the identification of abnormal regions and the overall visualization quality of the images, and optimizes the effectiveness and usability of hysteroscopic surgical image data. This solves the problems of significant physiological interference, blurred abnormal features, and insufficient effective details in existing hysteroscopic surgical images, providing high-quality image data support for hysteroscopic surgical-related image applications. Attached Figure Description

[0019] Figure 1 A flowchart illustrating a method for improving the quality of hysteroscopic surgical image data provided by the present invention;

[0020] Figure 2 This is a schematic diagram of a system for improving the quality of hysteroscopic surgical image data provided by the present invention.

[0021] In the attached diagram, the components represented by each number are as follows:

[0022] Data acquisition module 11, phase information calculation module 12, standard image generation module 13, contrast enhancement module 14. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0025] In the description of this invention, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed herein.

[0026] Example 1, as Figure 1 As shown, this embodiment of the invention provides a method for improving the quality of hysteroscopic surgical image data, including:

[0027] S10: Acquire hysteroscopic images of the target patient at a specific time point, and acquire multimodal physiological time-series data of the target patient before and after the specific time point.

[0028] Hysteroscopic surgery is an important minimally invasive gynecological diagnostic and treatment method. By inserting a scope with a camera through the vagina and cervix into the uterine cavity, it is possible to observe the original surgical image data such as endometrial morphology, blood vessel distribution, and tissue texture in real time and intuitively.

[0029] Hysteroscopic images are one of the important bases for doctors to judge the nature and extent of lesions in the uterine cavity and to formulate surgical plans. Their quality and interpretability are directly related to the early lesion detection rate and surgical accuracy.

[0030] However, the interpretation of hysteroscopic images has long faced an inherent physiological challenge: the endometrium is a target organ of female hormones, and its visual characteristics, such as thickness, echogenicity, glandular opening morphology, and vascular density, undergo regular dynamic changes with the menstrual cycle. During the proliferative phase, the endometrium gradually thickens and angiogenesis occurs; during the secretory phase, the endometrium exhibits a secretory response and significant edema; during menstruation, the endometrium sheds and bleeds. These cyclical physiological fluctuations lead to significant differences in the appearance of hysteroscopic images of the same patient and the same lesion site acquired at different stages of the cycle. For example, normal congestion and edema during ovulation may be misdiagnosed as endometritis, while the physiological thickening of the endometrium before menstruation is easily confused with endometrial hyperplasia.

[0031] Traditional hysteroscopic image analysis methods typically treat images as static, independent samples, ignoring the individual physiological cycle context of the patient at the time of image acquisition. Neither classic methods based on handcrafted features nor modern computer vision models based on deep learning have incorporated the crucial biological variable of the menstrual cycle into their calibration.

[0032] This defect leads to the following technical problems: the same lesion presents different visual features at different stages of the cycle, resulting in semantic drift in hysteroscopic images. For example, normal physiological congestion, thickening, and secretions are frequently misjudged as pathological changes, increasing unnecessary biopsies and psychological burden. Furthermore, the features of small lesions may be masked by strong periodic background noise, making them difficult to identify under conventional white light imaging.

[0033] Based on this, this application first obtains hysteroscopic images of the target patient at a specific time point, and then obtains multimodal physiological time-series data of the target patient before and after the specific time point, which serves as contextual supplementation for the hysteroscopic images and as the data basis for subsequent time-series feature fusion analysis.

[0034] Specifically, step S10 in the method includes:

[0035] Multimodal physiological time-series data include continuously monitored basal body temperature sequences, serum hormone level test data, time-series data of endometrial morphological characteristics extracted from pelvic ultrasound images, and menstrual history information recorded in electronic medical records.

[0036] In this embodiment, the specific time point is the precise time node at which the original images of the uterine cavity are acquired in real time during hysteroscopic surgery. The multimodal physiological time-series data of the target patient before and after the specific time point includes continuously monitored basal body temperature sequences, serum hormone level test data, time-series data of endometrial morphological characteristics extracted from pelvic ultrasound images, and menstrual history information recorded in electronic medical records. These multimodal physiological time-series data are time-series data that reflect the patient's physiological state from different dimensions, especially the cyclical changes in ovarian function and endometrium.

[0037] Among them, the continuous monitoring basal body temperature sequence refers to the oral or axillary temperature record taken every morning upon waking and before any activity during at least one complete menstrual cycle for the target patient. Basal body temperature rises by 0.3-0.5℃ after ovulation due to the effect of progesterone, forming a typical biphasic curve, which is an important indirect indicator for determining whether ovulation has occurred and the function of the corpus luteum.

[0038] For example, basal body temperature data of a target patient for 30 consecutive days can be obtained from a smart thermometer, health app, or electronic medical record to form a continuous monitoring sequence of basal body temperature at one sampling point per day.

[0039] Serum hormone level testing data refers to six sex hormone indicators obtained through venous blood sampling, including at least follicle-stimulating hormone (FSH), luteinizing hormone (LH), estradiol, and progesterone. The cyclical fluctuations in serum hormone levels are the fundamental driving force behind endometrial proliferation, secretion, and shedding.

[0040] The temporal data of endometrial morphological characteristics extracted from pelvic ultrasound images refers to a record of the changes over time of a series of quantitative indicators related to the endometrium, measured and obtained through transvaginal or transabdominal ultrasound examination. These include at least endometrial thickness, endometrial type, endometrial volume, and endometrial peristaltic wave frequency. Endometrial thickness is the most direct indicator reflecting the proliferative state of the endometrium, typically increasing from 2-4 mm during menstruation to 8-14 mm during ovulation. Endometrial type can be categorized into different stages, such as the three-line sign, homogeneous hyperechoic area, and blurred borders.

[0041] For example, if the target patient has undergone four ultrasound examinations in the past three months, and the endometrial thickness is recorded as 5mm, 9mm, 12mm and 7mm respectively, and the examination dates are marked accordingly, then a time-series data of endometrial morphological characteristics with a length of 4 is constituted.

[0042] The menstrual history information recorded in electronic medical records refers to the patient's self-reported menstrual history stored in the hospital information system in text or structured data form, including at least the age of menarche, the date of the first day of the last menstrual period, the usual length of the menstrual cycle, the number of days of menstruation, and a self-assessment of regularity. Continuously recorded menstrual history information can be used to extract the historical menstrual cycle length sequence, which is the basic data source for calculating the menstrual cycle regularity score.

[0043] It should be noted that, except for menstrual history information, not all of the above four types of data are required for implementation. In clinical applications, multimodal physiological time-series data can be adaptively fused based on the available data sources. However, to achieve better technical results, it is preferable to collect as diverse and abundant multimodal physiological time-series data as possible to improve the estimation accuracy of individual menstrual cycle phase information, thereby providing more accurate phase condition inputs for the subsequent generation of standard healthy uterine cavity reference images and the decoupling of pathological residual feature maps.

[0044] In summary, compared to existing technologies, this application acquires hysteroscopic images of a target patient at a specific time point and obtains multimodal physiological time-series data of the target patient before and after that specific time point. Thus, by simultaneously acquiring hysteroscopic images and multimodal physiological time-series data before and after their acquisition, a spatiotemporally aligned data foundation is provided for subsequently transforming menstrual cycle physiological fluctuations into individualized, quantifiable image correction benchmarks.

[0045] S20: Perform time-series feature fusion analysis on the multimodal physiological time-series data to obtain individual physiological cycle phase information corresponding to the specific time point.

[0046] Because the morphology and function of the endometrium change regularly with the menstrual cycle, the appearance of normal hysteroscopic images varies significantly at different stages of the cycle. If this cyclical background is ignored and images acquired at different stages of the cycle are directly compared, it is very easy to misinterpret normal physiological thickening or congestion as pathological changes.

[0047] Based on this, this application uses multimodal physiological time-series data to perform time-series feature fusion analysis, thereby capturing the time-series change patterns and feature dimensions of the multimodal physiological time-series data and integrating them. It abandons the one-sidedness of single physiological modality data, eliminates interference information through multi-data collaborative analysis, and obtains individual physiological cycle phase information.

[0048] Among them, individual physiological cycle phase information refers to the quantitative information of physiological cycle stage that is unique to the target patient and accurately corresponds to a specific time point, rather than a general physiological cycle division result, which can uniquely match the patient's current physiological state of the uterine cavity.

[0049] Specifically, step S20 in the method includes:

[0050] Based on the menstrual history information recorded in the electronic medical records of the multimodal physiological time series data, the menstrual cycle regularity score of the target patient is calculated;

[0051] Obtain a pre-constructed phase estimation submodule array, wherein the phase estimation submodule array contains M phase estimation submodules;

[0052] Multiply the menstrual cycle regularity score by M and round down to obtain the number of times the phase estimation submodule is called, N, where N is a positive integer between 1 and M;

[0053] N phase estimation submodules are randomly selected from the phase estimation submodule array. The selected N phase estimation submodules process the multimodal physiological time series data respectively and output N intermediate phase information accordingly.

[0054] The N intermediate phase information are fused and calculated to generate individual physiological cycle phase information.

[0055] In this embodiment, the menstrual cycle regularity score of the target patient is first calculated based on the menstrual history information recorded in the electronic medical record from the multimodal physiological time-series data. Specifically, the menstrual cycle regularity score is obtained through quantitative calculation based on the menstrual history information recorded in the electronic medical record. The essence of the menstrual cycle regularity score is to quantify the stability of the target patient's past menstrual cycle into a continuous value between 0 and 1, which is used to reflect the degree of irregularity of the target patient's menstrual cycle. The higher the menstrual cycle regularity score, the more irregular the menstruation.

[0056] Secondly, a pre-constructed phase estimation submodule array is obtained. This array contains M phase estimation submodules: each submodule is an independent, pre-trained model that takes multimodal physiological time-series data as input and outputs an estimate of the physiological cycle phase corresponding to that data. M is a pre-defined fixed positive integer representing the total number of phase estimation submodules in the array. M can be dynamically set based on computing resources, desired estimation accuracy, and clinical response time requirements, for example, set to 5, 10, or 20.

[0057] For example, a phase estimation submodule array can be constructed using the following technical approach:

[0058] 1. Submodule Construction: Since one of the core objectives of this application is to enable the phase estimation submodule array to cover diverse menstrual cycle patterns, rather than relying on the strong expressive power of a single model, a homogeneous ensemble strategy is preferred to construct M phase estimation submodules. Specifically, M phase estimation submodules are constructed according to the same network structure and initial parameter configuration. The network structure can be a multilayer perceptron, whose model composition includes: an input layer for receiving feature vectors from multimodal physiological time-series data; two or three hidden layers, each containing 64 to 256 neurons, with ReLU activation function; and an output layer that outputs a multidimensional feature vector as the intermediate phase information. Parameter configuration includes: weight initialization using Xavier uniform initialization, bias initialization to 0, Adam optimizer, and a learning rate set to 0.001. It should be noted that the above structure and parameters are only one feasible baseline configuration, and those skilled in the art can adaptively adjust them according to actual application scenarios.

[0059] 2. Submodule Training: Multimodal physiological time-series data are collected from a large-scale female population to form the original training sample dataset. Simultaneously, real menstrual cycle phase information corresponding to each sample is obtained through methods such as LH peak detection, ultrasound ovulation monitoring, or cycle day calibration to form the original sample label dataset. Using random sampling with replacement, M data subset pairs are extracted in parallel from the original training sample dataset and the original sample label dataset. The number of samples in each data subset pair is comparable to the size of the original dataset, thus constructing M training subsets with different distributions. Taking the training process of any phase estimation submodule as an example: one data subset pair is randomly selected from the M data subset pairs and divided into a training set, a validation set, and a test set in a 7:1.5:1.5 ratio. During training, the multimodal physiological time-series data from the training set is used as input features, the corresponding real menstrual cycle phase information is used as supervision labels, mean squared error is used as the loss function, and mini-batch gradient descent is used for parameter updates. The batch size is set to 32 or 64. After each training cycle, the loss value is calculated on the validation set. When the validation set loss no longer decreases for 10 consecutive cycles, the model is considered to have converged, training is stopped, and a trained phase estimation submodule is obtained. Subsequently, the remaining M-1 phase estimation submodules are trained sequentially using the same method.

[0060] 3. Array integration: The trained M phase estimation sub-modules are logically integrated to form a phase estimation sub-module array.

[0061] Preferably, to further improve the accuracy and generalization ability of the phase estimation submodule array, a heterogeneous modeling strategy can be adopted. For example, in the M phase estimation submodules, some use multilayer perceptrons, some use temporal convolutional networks, and some use lightweight Transformers, respectively, to fit the mapping relationship from multimodal physiological time-series data to phase information with different network topologies. In this way, by introducing diversity at the model level, it complements the diversity at the data level, thereby further improving the adaptability of the phase estimation submodule array to complex and irregular periodic patterns.

[0062] It should be noted that the network architecture selection, hyperparameter settings, optimization algorithms, loss functions, training strategies, and ensemble methods involved in the above model construction and training process are all mature existing technologies. Details such as the specific number of network layers, neurons, and training epochs not specified in this application can be fully implemented by those skilled in the art without creative effort, based on common knowledge and conventional experimental capabilities, after reading this specification.

[0063] Next, the menstrual cycle regularity score is multiplied by M and rounded to obtain the number of phase estimation submodule calls, N, where N is a positive integer between 1 and M. Specifically, the menstrual cycle regularity score ranges from 0 to 1. Multiplying it by the total number of phase estimation submodules M and rounding it yields the dynamically changing number of phase estimation submodule calls, N. This is because the menstrual cycle regularity score is positively correlated with the degree of menstrual irregularity; therefore, the more irregular the patient's menstrual cycle, the larger the value of N, and the more phase estimation submodules are called, thus improving the accuracy of phase estimation for irregular cycles.

[0064] The rounding function can be selected as rounding up, rounding down, or rounding to the nearest integer.

[0065] For example, assuming M=10, if a patient's menstrual cycle is extremely regular with a regularity score of 0.2, then 0.2 × 10 = 2, rounded down to obtain the number of phase estimation submodule calls N = 2. Conversely, if a patient's menstrual cycle is extremely irregular with a regularity score of 0.9, then 0.9 × 10 = 9, rounded down to obtain the number of phase estimation submodule calls N = 9. This design reflects the scheduling logic of fewer calls for patients with regular menstrual cycles and more calls for patients with irregular cycles: for patients with regular cycles, their physiological timeline pattern is simple and highly predictable, and a high-confidence phase estimate can be obtained by calling a small number of phase estimation submodules; for patients with irregular cycles, their physiological timeline pattern is complex and highly uncertain, requiring more calls to phase estimation submodules for collective decision-making to reduce single-model bias.

[0066] Furthermore, N phase estimation submodules are randomly selected from the phase estimation submodule array. These N selected submodules process the multimodal physiological time-series data and output N intermediate phase information respectively. Specifically, random selection avoids the analytical bias caused by fixed modules. The N phase estimation submodules process the same set of multimodal physiological time-series data synchronously and independently, and each phase estimation submodule outputs an independent intermediate phase information, ultimately resulting in N different intermediate phase information.

[0067] For example, if M=10 and N=6 is calculated, then 6 phase estimation sub-modules are randomly selected from the 10 sub-modules, and 6 intermediate phase information are output after processing the multimodal physiological time series data respectively.

[0068] Finally, the N intermediate phase information pieces are fused and calculated to generate individual physiological cycle phase information. Specifically, the fusion calculation refers to integrating and processing the N independent intermediate phase information pieces to eliminate the analysis errors and abnormal deviations of individual phase estimation sub-modules, ultimately generating accurate, stable individual physiological cycle phase information corresponding to a specific time point.

[0069] Specifically, the step of "calculating the menstrual cycle regularity score of the target patient based on the menstrual history information recorded in the electronic medical records of the multimodal physiological time-series data" includes:

[0070] Extract the sequence of continuous historical menstrual cycle lengths of the target patient from the menstrual history information recorded in the electronic medical record;

[0071] Calculate the standard deviation and mean of the historical menstrual cycle length series, and calculate the ratio of the standard deviation to the mean to obtain the coefficient of variation of cycle length;

[0072] The average absolute rate of change of each period is obtained by calculating the ratio of the sum of the absolute values ​​of the rate of change of each adjacent period in the historical menstrual cycle length sequence to the total number of adjacent period pairs.

[0073] The basic fluctuation sub-score is obtained by dividing the coefficient of variation of the cycle length by a first preset clinical threshold, wherein the value range of the basic fluctuation sub-score is 0~1;

[0074] Divide the mean absolute periodic change rate by the second preset clinical threshold to obtain the drastic change sub-score, wherein the value range of the drastic change sub-score is 0~1;

[0075] The arithmetic mean of the baseline fluctuation sub-score and the dramatic change sub-score is calculated to obtain the menstrual cycle regularity score, wherein the value of the menstrual cycle regularity score is positively correlated with the degree of menstrual irregularity.

[0076] In this embodiment, the first step is to extract a continuous historical menstrual cycle length sequence of the target patient from the menstrual history information recorded in the electronic medical record. Specifically, it is necessary to extract the duration data of multiple consecutive and complete menstrual cycles of the target patient to form a historical menstrual cycle length sequence. This sequence must ensure continuity and completeness, with no missing or abnormally truncated data. The menstrual cycle length refers to the number of days between the first day of the current menstrual period and the first day of the next menstrual period.

[0077] For example, to obtain statistical significance, the historical menstrual cycle length sequence is required to contain at least three consecutive cycle records. For instance, menstrual records from the past six months are extracted from the target patient's electronic medical record, and the historical menstrual cycle length sequence is calculated to be: [28,31,30,29,33,28].

[0078] Secondly, the standard deviation and mean of the historical menstrual cycle length series are calculated, and the ratio of the standard deviation to the mean is calculated to obtain the coefficient of variation (COP) of cycle length. The COP is an indicator used to quantify the overall fluctuation range of the historical menstrual cycle. The larger the COP value, the higher the overall dispersion and the more significant the fluctuation of the target patient's menstrual cycle. The formula is: COP = Standard deviation of historical menstrual cycle length series / Mean of historical menstrual cycle length series. Continuing the previous example, the calculated mean of the historical menstrual cycle length series is 29.83, and the standard deviation is 1.94. Therefore, the COP = 1.94 / 29.83 = 0.065. A larger COP indicates a greater fluctuation range in cycle length relative to its average level, and a poorer regularity.

[0079] Next, the average absolute cycle change rate is calculated by dividing the sum of the absolute values ​​of the change rates of each adjacent cycle in the historical menstrual cycle length sequence by the total number of adjacent cycle pairs. Specifically, the average absolute cycle change rate quantifies the degree of sudden and drastic changes in the menstrual cycle; a higher average absolute cycle change rate indicates more frequent and pronounced cycle abrupt changes. The calculation process is as follows: first, calculate the change rate of each pair of adjacent cycles, i.e., (length of the next cycle - length of the previous cycle) / length of the previous cycle; then, take the absolute values ​​of these change rates; finally, calculate the arithmetic mean of the absolute values ​​of all change rates to obtain the average absolute cycle change rate. Continuing the previous example, we calculate the rate of change for each pair of adjacent cycles in turn. The rates of change for each pair of adjacent cycles are (31-28) / 28=0.107, (30-31) / 31=-0.032, (29-30) / 29=-0.034, (33-29) / 29=0.138, and (28-33) / 33=-0.152, respectively. Then, we calculate the ratio of the sum of the absolute values ​​of the rates of change for each pair of adjacent cycles to the total number of pairs of adjacent cycles: (0.107+0.032+0.034+0.138+0.152) / 5=0.093, which gives us the average absolute rate of change for the cycle. The larger the average absolute rate of change for the cycle, the higher the frequency and amplitude of the sharp jumps in cycle length between adjacent cycles, and the worse the regularity.

[0080] Furthermore, the coefficient of variation of cycle length is divided by a first preset clinical threshold to obtain the baseline volatility sub-score, where the value range of the baseline volatility sub-score is 0 to 1. Specifically, the first preset clinical threshold is a clinically recognized reference value for baseline volatility in a normal menstrual cycle. The coefficient of variation of cycle length is normalized to the 0-1 range through division to form the baseline volatility sub-score, which directly reflects the degree of abnormality in the baseline volatility of the cycle. At the same time, the range of the baseline volatility sub-score is constrained by 0 to 1. First, the ratio of the coefficient of variation to the first preset clinical threshold is calculated. If the ratio of the coefficient of variation to the first preset clinical threshold is less than or equal to 1, this ratio is directly used as the baseline volatility sub-score; if the ratio of the coefficient of variation to the first preset clinical threshold is greater than 1, the baseline volatility sub-score is set to 1.

[0081] Furthermore, the mean absolute cycle change rate is divided by a second preset clinical threshold to obtain a drastic change sub-score, where the value of the drastic change sub-score ranges from 0 to 1. Specifically, the second preset clinical threshold is a clinically recognized reference value for drastic changes in a normal menstrual cycle. Similarly, the mean absolute cycle change rate is normalized to the 0-1 range through division to form the drastic change sub-score, which directly reflects the degree of abnormality in drastic cycle changes.

[0082] Finally, the arithmetic mean of the baseline fluctuation sub-score and the dramatic change sub-score is calculated to obtain the menstrual cycle regularity score. The magnitude of the menstrual cycle regularity score is positively correlated with the degree of menstrual irregularity. Specifically, the arithmetic mean combines the characteristics of both baseline fluctuation and dramatic change in the cycle. The higher the score, the more irregular the menstrual cycle of the target patient; the lower the score, the more regular the cycle.

[0083] In summary, this application, through the combination of two complementary indicators—the coefficient of variation of cycle length (which measures overall dispersion) and the mean absolute rate of change of cycle (which measures local jumps)—can comprehensively and robustly characterize the regularity of the menstrual cycle. Moreover, the calculation method is entirely based on quantifiable objective data, eliminating the bias of patients' subjective recall, and has high repeatability and clinical applicability.

[0084] Furthermore, the process of determining the "first preset clinical threshold and the second preset clinical threshold" includes:

[0085] Collect a set of menstrual cycle data containing multiple samples, where each sample contains at least menstrual history information, age, and body mass index;

[0086] Using age segmentation and body mass index classification as joint classification criteria, the samples in the menstrual cycle data set are divided into multiple group subsets. The age segmentation includes at least puberty, reproductive age, and perimenopause, and the body mass index classification includes at least underweight, normal, overweight, and obese ranges.

[0087] For each of the aforementioned subsets of the population, the arithmetic mean of the coefficients of variation of the period length of all samples within it is calculated and used as the first clinical reference threshold for each subset of the population.

[0088] For each of the aforementioned subsets, the arithmetic mean of the average absolute periodic rate of change of all samples within it is calculated and used as the second clinical reference threshold for each subset.

[0089] Construct a threshold mapping library to store the mapping relationship between each population subset and the corresponding first clinical reference threshold and second clinical reference threshold;

[0090] The target patient's age and body mass index are used to determine the subset of the population to which they belong, and the corresponding first clinical reference threshold and second clinical reference threshold are obtained from the threshold mapping relationship library, which are respectively used as the first preset clinical threshold and the second preset clinical threshold for the target patient.

[0091] In this embodiment, a menstrual cycle data set containing multiple samples is first collected. Each sample includes at least menstrual history information, age, and body mass index. Specifically, a large number of female subjects' clinical data are collected to form a menstrual cycle data set. Each sample fully includes menstrual history information, age, and body mass index, ensuring the clinical representativeness and data integrity of the samples and covering people with different physiological states and physical characteristics.

[0092] Among them, menstrual history information can reflect the basic fluctuation characteristics, cycle change patterns and physiological baseline status of a woman's menstrual cycle; age can reflect the physiological stage a woman is in, and there are significant differences in the function of the hypothalamus, pituitary and ovarian axis at different physiological stages, which directly affect the inherent regularity of the menstrual cycle; body mass index (BMI) can reflect a woman's nutritional metabolic status and body fat distribution characteristics. This index is closely related to the synthesis and metabolism of core hormones such as estrogen and progesterone, and is a key non-physiological factor that leads to menstrual cycle fluctuations.

[0093] For example, anonymized medical records of 100,000 women of childbearing age can be extracted from hospital information systems and clinical research databases to form a menstrual cycle data set.

[0094] Secondly, using age segmentation and body mass index (BMI) classification as joint classification criteria, the samples in the menstrual cycle data set are divided into multiple subsets. The age segmentation includes at least puberty (approximately 10-19 years old), reproductive age (approximately 20-44 years old), and perimenopause (approximately 45-55 years old). The BMI classification includes at least the underweight range (BMI < 18.5), the normal range (18.5 ≤ BMI < 24), and the overweight / obese range (BMI ≥ 24). For example, these two dimensions can be combined using a Cartesian product to obtain 3 × 3 = 9 subsets, such as the puberty-normal range group, the reproductive age-overweight / obese range group, etc.

[0095] Furthermore, the physiological characteristics of menstrual cycles vary significantly among women of different ages and nutritional statuses. For example, physiological irregularities are common during puberty due to the immaturity of the hypothalamus, pituitary gland, and ovarian axis; cycle fluctuations are normal during perimenopause due to declining ovarian function; and obese women are more prone to infrequent or anovulatory cycles due to abnormal estrogen levels caused by increased aromatase activity in adipose tissue. Therefore, a coefficient of variation of 0.20 for cycle length may still be within the physiological range for a 15-year-old underweight girl; however, for a 30-year-old woman of normal weight, the same 0.20 indicates pathological irregularity. Based on this, this application calculates the arithmetic mean of the coefficient of variation of cycle length for all samples in each subset of the population, as the first clinical reference threshold for each subset; and calculates the arithmetic mean of the average absolute cycle change rate for all samples in each subset of the population, as the second clinical reference threshold for each subset.

[0096] Specifically, the arithmetic mean obtained by summing the coefficients of variation of period lengths for all samples within the same subset and dividing by the sample size is the first clinical reference threshold for that subset, representing the normal baseline periodic fluctuation standard for that subset. Similarly, the arithmetic mean of the average absolute periodic variation rate for all samples within the same subset is the second clinical reference threshold, representing the normal standard for dramatic periodic changes for that subset. Thus, obtaining a reference benchmark within a homogeneous population through stratified statistics is key to achieving personalized thresholds.

[0097] Furthermore, a threshold mapping relationship library is constructed to store the mapping relationship between each population subset and its corresponding first clinical reference threshold and second clinical reference threshold. This threshold mapping relationship library is a structured data storage module that internally stores a one-to-one correspondence between population subset, first clinical reference threshold, and second clinical reference threshold, enabling the function of quickly querying matching thresholds based on population subsets.

[0098] Finally, based on the target patient's age and body mass index (BMI), a subset of the target population is determined. The corresponding first and second clinical reference thresholds are then retrieved from the threshold mapping database and used as the first and second preset clinical thresholds, respectively, to adapt to the target patient. Specifically, the target patient's actual age and BMI are matched to a corresponding subset of the target population. The first and second clinical reference thresholds for that subset are then retrieved from the threshold mapping database and used as the first and second preset clinical thresholds to adapt to the target patient. This achieves individualized matching of the first and second preset clinical thresholds, ensuring that the scoring calculation closely reflects the target patient's physiological characteristics.

[0099] For example, a target patient who is 26 years old and has a BMI of 22.1 is first classified into the "fertile age - normal range" subset; then, the first clinical reference threshold corresponding to this subset is obtained from the threshold mapping relation library, which is 0.14 and the second clinical reference threshold is 0.09; these two values ​​are then assigned to the first preset clinical threshold and the second preset clinical threshold for the calculation of the target patient's menstrual cycle regularity score.

[0100] In summary, this application transforms macroscopic population statistical patterns into microscopic individualized clinical parameters through a technical approach of population stratification, statistical modeling, mapping and database construction, and query adaptation, thereby improving the scientific rigor and accuracy of menstrual cycle regularity scoring.

[0101] Furthermore, the step of "fusing and calculating the N intermediate phase information to generate individual physiological cycle phase information" includes:

[0102] Calculate the arithmetic mean of the N intermediate phase information to obtain the average phase vector;

[0103] The average phase vector is input to a pre-trained phase decoder to map individual physiological cycle phase information, wherein the individual physiological cycle phase information is a scalar value used to quantitatively represent the position of a specific time point in the menstrual cycle of the target patient individual.

[0104] In this embodiment, the arithmetic mean of N intermediate phase information is first calculated as the average phase vector. Specifically, each intermediate phase information is a multi-dimensional feature vector, such as a 256-dimensional floating-point vector. This multi-dimensional feature vector does not directly correspond to a certain number of days in a cycle, but rather is a high-dimensional abstract representation of the current physiological time series data by the phase estimation submodule. By calculating the arithmetic mean of the N multi-dimensional feature vectors element-wise in their corresponding dimensions, an average phase vector of the same dimension is obtained. The average phase vector integrates the knowledge of multiple phase estimation submodules, and the averaging operation reduces the random error and model bias of a single phase estimation submodule.

[0105] Secondly, the average phase vector is input into a pre-trained phase decoder to map and obtain individual menstrual cycle phase information. This individual menstrual cycle phase information is a scalar value used to quantitatively represent the position of a specific time point within the target patient's menstrual cycle. Specifically, the phase decoder is an independent, pre-trained neural network module whose function is to decode the average phase vector in the high-dimensional feature space into low-dimensional, directly interpretable individual menstrual cycle phase information. The scalar value is a real number continuously distributed within a preset cycle length, and its value corresponds to the actual number of days in the target patient's cycle or a standardized cycle scale. For example, if a 28-day standard cycle is mapped to a continuous interval from 0.0 to 28.0, then day 1 corresponds to 1.0, day 14 corresponds to 14.0, and day 28 corresponds to 28.0.

[0106] For example, the phase decoder can be built and trained through the following technical path:

[0107] 1. Model Structure and Parameter Configuration: The phase decoder adopts a multilayer perceptron architecture, mainly consisting of an input layer, hidden layers, and an output layer. The dimension of the input layer is consistent with the dimension of the average phase vector; the hidden layers are set to two or three fully connected layers, each containing 64 to 256 neurons, with ReLU activation function; the output layer contains one neuron, with linear activation function, directly outputting a scalar value as the phase information of the individual's physiological cycle. To prevent overfitting, a Dropout layer can be introduced after the hidden layers, with a dropout rate set to 0.2~0.5, weight initialization using Xavier uniform initialization, and bias initialization to 0.

[0108] 2. Training Process: First, a large-scale sample dataset is collected. Each sample contains the average phase vector output by the phase estimation submodule array as input features, and real individual physiological cycle phase information obtained through methods such as LH peak detection, ultrasound ovulation monitoring, or cycle day calibration as supervision labels. The samples are randomly divided into training, validation, and test sets in an 8:1:1 ratio. Mean squared error is used as the loss function, and the Adam optimizer is used for parameter updates, with an initial learning rate set to 1e-3. During training, a cosine annealing learning rate decay strategy is adopted, and an early stopping mechanism is set: training stops when the validation set loss does not decrease for 10-20 consecutive rounds, and the model parameters are restored to the level with the lowest validation set loss. After training, the phase decoder has the ability to map any input average phase vector to a continuous phase scalar.

[0109] It should be noted that the model architecture, parameter configuration, and training algorithm described above are all mature existing technologies in the field of deep learning. Those skilled in the art, after reading this specification and combining it with common knowledge and conventional experimental capabilities, are fully capable of constructing and training the phase decoder without any creative effort.

[0110] For example, the average phase vector of the target patient is input into a pre-trained phase decoder to map the individual's menstrual cycle phase information. Assuming the average phase vector is a 256-dimensional feature vector with values ​​ranging from -2.3 to 1.8, after forward propagation calculation by the phase decoder, a scalar value of 14.37 is output. This scalar value of 14.37 represents the specific position of the current time point in the target patient's menstrual cycle as day 14.37, indicating that they have entered the early luteal phase, approximately two days after ovulation.

[0111] Compared to existing technologies that rely solely on the last menstrual period date to estimate cycle length (e.g., vaguely identifying it as "day 14 of the cycle") or use coarse staging methods like follicular phase or luteal phase, this application achieves continuous and precise quantitative phase representation through high-dimensional feature decoding. This allows for the capture of even subtle physiological differences, sometimes as small as half a day. Particularly during critical physiological windows such as around ovulation and the endometrial implantation window, a half-day phase shift can be accompanied by changes in LH peak initiation, progesterone receptor expression, or significant differences in endometrial receptivity, directly impacting the accurate differentiation between normal physiological states and pathological changes. Thus, this application not only provides high-resolution input for generating a standard healthy uterine cavity reference image that precisely matches the current phase, fundamentally improving the calculation accuracy of pathological residual feature maps, but also represents a paradigm shift from coarse staging to precise localization in clinical practice. This provides quantifiable technical support for individualized surgical timing, early lesion detection, and fertility preservation.

[0112] In summary, compared to existing technologies, this application performs temporal feature fusion analysis on the multimodal physiological time-series data to obtain individual physiological cycle phase information corresponding to the specific time point. This avoids the limitations and interference of single-modal physiological data. By integrating and fusion-analyzing the temporal features of multimodal physiological time-series data, it accurately obtains individual physiological cycle phase information corresponding to a specific time point, providing precise individualized physiological benchmark support for the subsequent generation of standard healthy uterine cavity reference images.

[0113] S30: Input the individual's physiological cycle phase information into a pre-trained healthy uterine cavity image generation agent, and output a standard healthy uterine cavity reference image corresponding to the specific time point.

[0114] Because the normal shape of a woman's uterine cavity undergoes dynamic physiological changes with her menstrual cycle, there are significant differences in the healthy uterine cavity shape among different individuals and at different phases of the menstrual cycle. Traditional hysteroscopic diagnosis and treatment cannot obtain an individualized healthy uterine cavity reference image that accurately matches the patient's current menstrual cycle phase, making it difficult to effectively distinguish between physiological morphological changes and pathological abnormalities.

[0115] To address the aforementioned issues, this application inputs the individual's physiological cycle phase information into a pre-trained healthy uterine cavity image generation agent, and outputs a standard healthy uterine cavity reference image corresponding to the specific time point.

[0116] Specifically, step S30 in the method includes:

[0117] Hysteroscopic images were collected from multiple healthy volunteers at multiple preset time points during their complete menstrual cycle to form a training image set;

[0118] For each hysteroscopic image in the training image set, the corresponding individual physiological cycle phase information is calculated based on the multimodal physiological time series data recorded synchronously during acquisition, and the individual physiological cycle phase information is used as the annotation for this hysteroscopic image.

[0119] Construct a conditional generative adversarial network that includes a generator and a discriminator;

[0120] The generator is input with the individual's physiological cycle phase information as a condition, and the discriminator is input with the corresponding hysteroscopic image as a real sample. Through adversarial training, the generator learns the mapping relationship from the individual's physiological cycle phase information to a healthy hysteroscopic image under the individual's physiological cycle phase information.

[0121] After training, the generator in the conditional generative adversarial network is defined as a healthy uterine cavity image generating agent.

[0122] In this embodiment, hysteroscopic images of multiple healthy volunteers collected at multiple preset time points throughout their complete menstrual cycle are first collected to form a training image set. Specifically, the healthy volunteers are female subjects whose uterine cavity is confirmed to be free of any pathological lesions and whose menstrual cycles are regular, as confirmed by clinical examination. The collected images need to cover all physiological stages of the healthy volunteers' complete menstrual cycle and be repeatedly collected at multiple fixed preset time points, such as hysteroscopic examinations and images on days 5, 8, 11, 14, 17, 20, 23, and 26 of the cycle, thereby ensuring that the training image set covers the full cycle and multi-phase imaging features of a healthy uterine cavity.

[0123] Secondly, for each hysteroscopic image in the training image set, the corresponding individual menstrual cycle phase information is calculated based on the multimodal physiological time-series data recorded synchronously during acquisition, and this individual menstrual cycle phase information is used as the annotation for that hysteroscopic image. Specifically, multimodal physiological time-series data is synchronously acquired for each training hysteroscopic image at a corresponding time point, and the unique individual menstrual cycle phase information is calculated using the same method as the aforementioned steps, establishing a one-to-one correspondence between hysteroscopic images and individual menstrual cycle phase information, providing accurate annotation data for model training.

[0124] Next, a conditional generative adversarial network (GAN) is constructed, comprising a generator and a discriminator. A GAN is a generative artificial intelligence network that takes specified information as input. The generator is responsible for synthesizing hysteroscopic images from the input physiological cycle phase information, while the discriminator is responsible for determining whether the input image is a real hysteroscopic image or a fake image synthesized by the generator.

[0125] For example, the generator of the conditional generative adversarial network adopts a deep convolutional encoder-decoder architecture, mainly composed of an input layer, a hidden layer, and an output layer. Its input is a concatenation of a Gaussian noise vector and an embedding vector of individual physiological cycle phase information. Specific parameter configurations can be found as follows: 1. Input layer parameters: The noise vector has a dimension of 100 and is sampled from a standard normal distribution N(0,1); the phase embedding layer is a fully connected layer with an input dimension of 1 and an output dimension of 100, and the activation function is linear activation; the concatenated input feature dimension is 200.

[0126] 2. Hidden Layer Parameters (taking an output image size of 256×256 as an example): First layer: Fully connected layer, input 200 dimensions, output 8192 dimensions, reshaped into an 8×8×512 feature map; Second layer: 4×4 deconvolution layer, stride 2, padding 1, input 512 channels, output 256 channels, output feature map size 16×16; Third layer: 4×4 deconvolution layer, stride 2, padding 1, input 256 channels, output 128 channels, output feature map size... The image size is 32×32; the fourth layer is a 4×4 deconvolutional layer with a stride of 2 and padding of 1, 128 input channels, 64 output channels, and an output feature map size of 64×64; the fifth layer is a 4×4 deconvolutional layer with a stride of 2 and padding of 1, 64 input channels, 32 output channels, and an output feature map size of 128×128; the sixth layer is a 4×4 deconvolutional layer with a stride of 2 and padding of 1, 32 input channels, 3 output channels, and an output feature map size of 256×256. After the second to fifth deconvolutional layers, batch normalization layers and ReLU activation functions are sequentially connected, with the momentum parameter of the batch normalization layers set to 0.8.

[0127] 3. Output layer parameters: The activation function is Tanh, and the output pixel value range is [-1, 1].

[0128] For example, the discriminator of the conditional generative adversarial network adopts a deep convolutional neural network architecture. Its input is a feature map concatenated in the channel dimension by embedding vectors of the image to be discriminated and the phase information of the individual's physiological cycle. It mainly consists of an input layer, a hidden layer, and an output layer. Specific parameter configurations can be found as follows: 1. Input layer parameters: Image input: size 256×256×3, pixel value range [-1,1]; Phase embedding layer: fully connected layer, input dimension 1, output dimension 65536, reshaped into a 256×256×1 single-channel feature map; Concatenated input feature map: 256×256×4.

[0129] 2. Hidden Layer Parameters: First Layer: 4×4 convolutional layer, stride 2, padding 1, input 4 channels, output 64 channels, output feature map size 128×128; Second Layer: 4×4 convolutional layer, stride 2, padding 1, input 64 channels, output 128 channels, output feature map size 64×64; Third Layer: 4×4 convolutional layer, stride 2, padding 1, input 128 channels, output 256 channels, output feature map size 32×32; Fourth Layer: 4×4 convolutional layer, stride 2, padding 1, input 256 channels, output 512 channels, output feature map size 16×16; Fifth Layer: 4×4 convolutional layer, stride 2, padding 1, input 512 channels, output 512 channels, output feature map size 8×8; Sixth Layer: Flatten the 8×8×512 feature map into a 2048-dimensional feature vector; Seventh Layer: Fully connected layer, input 2048-dimensional, output 1-dimensional. After the first convolutional layer: LeakyReLU activation function with a negative slope of 0.2 and no batch normalization; after the second to fifth convolutional layers: batch normalization layers are sequentially connected to LeakyReLU activation functions, and the momentum parameter of the batch normalization layers is set to 0.8.

[0130] 3. Output layer parameters: The activation function is Sigmoid, and the output range is [0,1]. It represents the probability that the input image is a real healthy hysteroscopic image and matches the phase information of the individual's menstrual cycle.

[0131] It should be noted that the network structure, parameter configuration and training strategy described above are mature technical solutions for implementing conditional generative adversarial networks by those skilled in the art. The specific number of layers, convolutional kernel size, number of feature map channels and other details can be adaptively adjusted according to the actual task requirements.

[0132] Furthermore, individual menstrual cycle phase information is used as a conditional input to the generator, and the corresponding hysteroscopic image is used as a real sample input to the discriminator. Through adversarial training, the generator learns the mapping relationship from individual menstrual cycle phase information to healthy hysteroscopic images under that phase. Specifically, during adversarial training, the generator generates healthy hysteroscopic images based on individual menstrual cycle phase information, and the discriminator distinguishes between the generated healthy hysteroscopic images and real hysteroscopic images. The two processes alternately optimize, with the generator attempting to synthesize a healthy hysteroscopic image sufficient to deceive the discriminator, while the discriminator continuously improves its ability to distinguish between real and fake images. After multiple rounds of competition, the generator's synthesis ability gradually increases, ultimately enabling it to synthesize photorealistic healthy hysteroscopic images that conform to the physiological characteristics of any input individual menstrual cycle phase information.

[0133] For example, the adversarial training process can be implemented using the following parameter configuration and training procedure:

[0134] 1. Optimizer configuration: The Adam optimizer is used, with the learning rate of both the generator and discriminator set to 0.0002, the first moment decay coefficient to 0.5, the second moment decay coefficient to 0.999, the numerical stability constant to 1e-8, and the batch size set to 16 or 32.

[0135] 2. Loss function configuration: A weighted combination of adversarial loss and L1 reconstruction loss from standard conditional generative adversarial networks is adopted. The adversarial loss uses binary cross-entropy, and the L1 reconstruction loss is used to constrain the consistency between the generated image and the real image at the pixel level.

[0136] 3. Training Process: In each iteration, the generator parameters are first fixed. A batch of real and healthy hysteroscopic images and their corresponding individual menstrual cycle phase information are sampled from the real data distribution and input into the generator to synthesize a batch of images. The real hysteroscopic images and the synthesized healthy hysteroscopic images are paired with their phase conditions. Simultaneously, a batch of noise vectors is sampled from the prior noise distribution and input into the discriminator for discrimination. The discriminator loss is calculated and backpropagation is used to update the discriminator parameters. Subsequently, the discriminator parameters are fixed, and noise vectors are sampled again and input into the generator to synthesize healthy hysteroscopic images. The generator loss is calculated and backpropagation is used to update the generator parameters. The above process is repeated alternately until the preset number of training rounds is reached, such as 100-200 rounds, or the early stopping condition is met, such as the validation set loss not decreasing for 20 consecutive rounds. After training, the generator model parameters with the lowest validation set loss are saved, thus obtaining the healthy hysteroscopic image generation agent.

[0137] It should be noted that the optimizer, loss function, training process and hyperparameters mentioned above are all mature technical solutions for implementing conditional generative adversarial networks in this field. Those skilled in the art can make adaptive adjustments to parameters such as batch size, learning rate and number of training rounds according to actual task requirements.

[0138] Finally, after training, the generator in the conditional generative adversarial network is defined as a healthy uterine cavity image generation agent. Specifically, the trained generator can independently receive individual menstrual cycle phase information and directly output a standard healthy uterine cavity reference image corresponding to the phase, without relying on a discriminator, thus forming a directly applicable healthy uterine cavity image generation agent.

[0139] Furthermore, the individual's physiological cycle phase information is input into a pre-trained intelligent agent for generating healthy uterine cavity images, outputting a standard healthy uterine cavity reference image corresponding to a specific time point. This standard healthy uterine cavity reference image is a virtual hysteroscopic image synthesized by the intelligent agent for generating healthy uterine cavity images based on the target patient's current individual physiological cycle phase information, assuming the patient is in a completely healthy state. It can accurately reflect the normal anatomical features of a healthy uterine cavity at that specific phase, such as endometrial morphology, vascular distribution, and glandular openings. Using this as a personalized health benchmark for subsequent image comparisons, compared to using a fixed template or arbitrarily selected healthy images, it can fundamentally eliminate the visual difference interference caused by periodic physiological fluctuations, making the generation of pathological residual feature maps more accurate.

[0140] It is important to emphasize that the goal of the training phase is to synthesize healthy hysteroscopic images, while the output of the healthy hysteroscopic image generation agent in the application phase is a standard healthy hysteroscopic reference image. The former is a training supervision signal, while the latter is the actual product output. The two are semantically consistent, but they are distinguished by their names to clarify their functional roles in different stages.

[0141] In summary, compared to existing technologies, this application inputs the individual's menstrual cycle phase information into a pre-trained healthy uterine cavity image generation agent, outputting a standard healthy uterine cavity reference image corresponding to the specific time point. Thus, based on accurate individual menstrual cycle phase information, it can generate a standardized healthy uterine cavity reference image that perfectly matches the patient's individual needs at a specific time point, providing a unique and accurate healthy control benchmark for subsequent extraction of pathological residual feature maps.

[0142] S40: Compare the hysteroscopic image with the standard healthy uterine cavity reference image, generate a pathological residual feature map, and enhance the display of the pathological residual feature map.

[0143] In hysteroscopic surgical images, the normal physiological morphology of the uterine cavity changes dynamically with the individual's menstrual cycle, which can be easily confused with pathological abnormalities. Furthermore, small lesions have low visual recognition in the original images, and it is impossible to intuitively and accurately separate and highlight pathological areas based solely on the original images.

[0144] To address the aforementioned issues, this application compares the hysteroscopic images with the standard healthy uterine cavity reference images to generate a pathological residual feature map, and enhances the display of the pathological residual feature map.

[0145] Specifically, step S40 in the method includes:

[0146] An encoder-decoder network structure is adopted, and some network layer parameters are shared with the generator in the healthy uterine cavity image generation agent to construct a pathological feature decoupling device.

[0147] The pathological feature decoupler is trained using the reconstruction loss function and the adversarial difference loss function as joint constraints to obtain the trained pathological feature decoupler. The reconstruction loss function is used to constrain that the hysteroscopic image can be reconstructed by pixel-level addition of the standard healthy uterine cavity reference image and the pathological residual feature map to be generated. The adversarial difference loss function is used to constrain the minimization of the distance between the pathological residual feature map and the residual feature distribution of a set of real pathological samples in the feature space.

[0148] The trained pathological feature decoupling device uses the standard healthy uterine cavity reference image as a health benchmark to perform feature decoupling calculation on the hysteroscopic image, separates cycle-related features and cycle-independent residual features, and maps the cycle-independent residual features to generate a pathological residual feature map.

[0149] In this embodiment, an encoder-decoder network structure is first adopted, and some network layer parameters are shared with the generator in the healthy uterine cavity image generation agent to construct a pathological feature decoupling device. Specifically, the encoder-decoder network structure has the functions of feature extraction and feature decoupling; by sharing some network layer parameters with the generator in the healthy uterine cavity image generation agent, the healthy uterine cavity image features already learned by the generator can be reused, reducing the training cost of the pathological feature decoupling device and improving the decoupling accuracy between healthy and pathological features.

[0150] Specifically, the encoder-decoder network structure consists of two parts: an encoder and a decoder. For example, the encoder uses a multi-layer convolutional network to progressively downsample the input image and extract high-dimensional semantic features; the decoder uses a multi-layer deconvolutional network to progressively upsample the feature map output by the encoder, restoring it to the original image resolution, and outputting a pathological residual feature map with the same size as the input image. For example, the encoder may contain 4 to 6 convolutional layers, each with a kernel size of 3×3 or 4×4, a stride of 2, and the number of channels increasing progressively from 32 to 512. Each layer is followed by a batch normalization layer and a LeakyReLU activation function. The decoder structure is symmetrical to the encoder, using deconvolutional layers for upsampling, followed by a batch normalization layer and a ReLU activation function, and finally outputting a single-channel pathological residual feature map through a 1×1 convolutional layer and a Tanh or Sigmoid activation function.

[0151] For example, sharing some network layer parameters means that the encoder of the pathological feature decoupling agent and the encoder of the generator in the healthy uterine cavity image generation agent share at least the network weights of the first few convolutional layers. The generator in the healthy uterine cavity image generation agent has been trained with a large number of healthy hysteroscopic images, and its encoder has the ability to extract anatomical structural features that are strongly correlated with the phase of the menstrual cycle from hysteroscopic images. By inheriting these pre-trained weights, the pathological feature decoupling agent directly obtains the sensitivity and discriminative power to the visual features of a healthy uterine cavity, thereby using a standard healthy uterine cavity reference image as a health benchmark to more accurately decouple the input hysteroscopic image into cycle-related features and cycle-independent residual features. In this way, not only can the training cost and data dependence of the pathological feature decoupling agent be reduced, but the accuracy and robustness of feature separation can also be improved through knowledge transfer.

[0152] It should be noted that the specific parameters of the encoder-decoder, such as the number of layers, kernel size, and number of channels, can be adaptively adjusted according to the input image resolution, computing resources, and task complexity.

[0153] Secondly, a pathological feature decoupling function is trained using a reconstruction loss function and an adversarial difference loss function as joint constraints, resulting in a trained pathological feature decoupling function. The reconstruction loss function ensures that a hysteroscopic image can be reconstructed by pixel-level addition of a standard healthy uterine cavity reference image and the pathological residual feature map to be generated. The adversarial difference loss function minimizes the distance between the pathological residual feature map and the residual feature distribution of a set of real pathological samples in the feature space. Specifically, the reconstruction loss function ensures that the decoupled healthy and pathological features can reconstruct the original hysteroscopic image, avoiding feature loss; the adversarial difference loss function ensures that the generated pathological residual feature map is consistent with the distribution of clinical real pathological residual features, guaranteeing the authenticity of the pathological features.

[0154] The reconstruction loss function is used to constrain the reconstruction of the hysteroscopic image by pixel-level addition of the standard healthy uterine cavity reference image and the pathological residual feature map to be generated. The mathematical form of the reconstruction loss function can be expressed as: L_recon=||I_real-(I_healthy+I_residual)||, where L_recon is the reconstruction loss function, I_real is the hysteroscopic image, I_healthy is the standard healthy uterine cavity reference image, and I_residual is the pathological residual feature map to be generated. Its physical meaning is that it requires that the sum of the healthy part and the residual part must be able to perfectly restore the original image, and information loss or redundancy is not allowed.

[0155] The adversarial discriminant loss function is used to constrain the distance between the pathological residual feature map generated during training and the residual feature distribution of a set of real pathological samples in the feature space to be minimized. A set of pathologically confirmed hysteroscopic images, including lesion types such as polyps, submucosal fibroids, intrauterine adhesions, and endometrial hyperplasia, are collected in advance. The real pathological residual feature maps relative to the corresponding healthy baseline are extracted by differential extraction to form the real pathological residual feature distribution. The adversarial discriminant loss function, through an additional discriminator network, forces the pathological residual feature map generated by the pathological feature decoupling device to approximate the real distribution at the feature distribution level, rather than just at the pixel level. The mathematical form of the adversarial discriminant loss function can be expressed as: L_adv=E[logD(R_real)]+E[log(1-D(R_fake))], where R_real is the real pathological residual feature map, R_fake is the pathological residual feature map generated by the pathological feature decoupling device, and D is the discriminator network. The physical meaning of this loss function is that it forces the learned residual features to not only reconstruct the original image numerically, but also conform to the statistical regularity of real pathological samples at the semantic feature level, thereby effectively suppressing non-pathological differences, such as changes in lighting, lens angle, and physiological fluctuations, which are misjudged as lesions.

[0156] Through the dual constraints of the reconstruction loss function and the adversarial difference loss function, the pathological feature decoupler simultaneously achieves the completeness and specificity of feature decomposition, which can improve the accuracy, robustness and clinical interpretability of pathological residual feature maps.

[0157] Finally, the trained pathological feature decoupling device uses a standard healthy uterine cavity reference image as the health benchmark to perform feature decoupling calculations on the hysteroscopic image, separating cycle-related features from cycle-independent residual features, and mapping the cycle-independent residual features to generate a pathological residual feature map. Here, cycle-related features refer to normal physiological features of the uterine cavity that change with the menstrual cycle, while cycle-independent residual features refer to pathological abnormal features unrelated to the menstrual cycle. The pathological feature decoupling device completely removes physiological features, retaining only cycle-independent residual features, and mapping them into a visualized pathological residual feature map.

[0158] Among them, the pathological residual feature map is a two-dimensional feature matrix with the same size as the original hysteroscopic image. Each element value in the matrix represents the degree to which the corresponding pixel position deviates from the normal physiological structure.

[0159] Furthermore, step S40 of the method further includes:

[0160] The pathological residual feature map is subjected to pseudo-color mapping to generate a pseudo-color enhanced map.

[0161] The pseudo-color enhanced image and the hysteroscopic image are spatially registered and pixel-level fused to generate a difference-enhanced display image.

[0162] Morphological analysis is performed on the high-response regions in the difference-enhanced display image to extract the boundary contours and area parameters of the high-response regions, and the boundary contours and area parameters are superimposed on the difference-enhanced display image in the form of graphical annotations.

[0163] In this embodiment, the pathological residual feature map is first subjected to pseudo-color mapping to generate a pseudo-color enhanced image. Specifically, the original pathological residual feature map is a single-channel grayscale feature map, and the response intensity difference in the pathological region is not obvious; pseudo-color mapping maps the pixel values ​​with different response intensities in the single-channel grayscale feature map to different colors to form a pseudo-color enhanced image.

[0164] For example, a heatmap color scheme can be used: areas with a residual value of 0 are mapped to transparent or dark blue, areas with a medium residual value are mapped to yellow, and areas with a high residual value are mapped to bright red. Through this mapping, the lesion area appears visually "burning," greatly improving the speed and accuracy of human eye detection of abnormal areas.

[0165] Secondly, the pseudo-color enhanced image and the hysteroscopic image are spatially registered and pixel-level fused to generate a difference-enhanced display image. Spatial registration ensures perfect alignment of pixel coordinates between the pseudo-color enhanced image and the original hysteroscopic image; since they originate from the same image, this can be achieved through simple size alignment. Pixel-level fusion involves weighted overlay of the pseudo-color enhanced image onto the original image as a semi-transparent layer. The fusion method can be 50% original image + 50% pseudo-color or adaptive transparency mixing based on residual values. The generated difference-enhanced display image preserves the anatomical details of the original hysteroscopic image while adding prominent color cues to lesion areas.

[0166] Finally, morphological analysis is performed on the high-response regions in the difference-enhanced image to extract their boundary contours and area parameters. These parameters are then overlaid onto the difference-enhanced image using graphical annotations. Specifically, morphological analysis is a classic technique in digital image processing, including binarization, connected component labeling, edge detection, and contour extraction. High-response regions refer to continuous pixel areas in the pseudo-color enhanced image where the color intensity exceeds a preset threshold (which can be dynamically set according to actual conditions). These regions correspond to locations with high deviations in the pathological residual feature map. Through morphological analysis, the closed boundary contours of each high-response region can be automatically identified, and the pixel area of ​​that region can be calculated.

[0167] Specifically, the boundary contours and area parameters are overlaid on the difference enhancement display image in the form of graphical annotations: the precise boundary contours of each high-response area are drawn with high-brightness solid lines, such as bright green or white, and the area number and calculated area parameters are marked in the form of text boxes near the contours or on the side of the image, providing doctors with dual diagnostic basis of qualitative visualization and quantitative numerical data.

[0168] In summary, compared to existing technologies, this application compares the hysteroscopic image with the standard healthy uterine cavity reference image to generate a pathological residual feature map, and enhances the display of the pathological residual feature map. Thus, by generating a pathological residual feature map by comparing the hysteroscopic image with the standard healthy uterine cavity reference image and enhancing its display, interference from normal morphology caused by the menstrual cycle can be effectively eliminated, abnormal areas can be accurately highlighted, thereby improving the pathological identification accuracy of hysteroscopic surgical images.

[0169] In summary, the embodiments of this application have at least the following technical effects:

[0170] Compared to existing technologies, this application first acquires hysteroscopic images of the target patient at a specific time point, and then acquires multimodal physiological time-series data of the target patient before and after that specific time point. Thus, by simultaneously acquiring hysteroscopic images and multimodal physiological time-series data before and after their acquisition, precise spatiotemporal alignment of image data and physiological data is achieved. This provides a complete, correlated, and spatiotemporally matched data foundation for subsequently transforming menstrual cycle physiological fluctuations into individualized, quantifiable image correction benchmarks, ensuring the targetedness and accuracy of subsequent image quality enhancement processing.

[0171] Secondly, this application performs temporal feature fusion analysis on the multimodal physiological time-series data to obtain individual physiological cycle phase information corresponding to the specific time point. This effectively avoids the limitations and interference of single-modal physiological data. By systematically integrating the temporal change patterns and feature dimensions of multimodal physiological time-series data and performing in-depth fusion analysis, the physiological state of the target patient at a specific time point is accurately captured. This allows for the acquisition of individual physiological cycle phase information corresponding to that specific time point, providing precise and personalized physiological benchmark support for the subsequent generation of standard healthy uterine cavity reference images.

[0172] Furthermore, this application inputs the individual's menstrual cycle phase information into a pre-trained healthy uterine cavity image generation agent, outputting a standard healthy uterine cavity reference image corresponding to the specific time point. In this way, using precise individual menstrual cycle phase information as the core input condition, and relying on a pre-trained healthy uterine cavity image generation agent, an individualized standard healthy uterine cavity reference image that perfectly matches and adapts to the target patient's physiological characteristics at a specific time point can be generated. This provides a unique, reliable, and individualized healthy control benchmark for the accurate extraction of subsequent pathological residual feature maps.

[0173] Finally, this application compares the hysteroscopic image with the standard healthy uterine cavity reference image to generate a pathological residual feature map, and enhances the display of the pathological residual feature map. Thus, by performing pixel-level and feature-level dual comparative analysis on the hysteroscopic image and the standard healthy uterine cavity reference image, a pathological residual feature map containing only abnormal features is generated. Through targeted enhancement display processing, interference from the normal uterine cavity morphology caused by dynamic changes in the menstrual cycle can be effectively eliminated, accurately highlighting abnormal areas and detailed features in the hysteroscopic image, thereby significantly improving the identification accuracy and visualization quality of abnormal features in hysteroscopic surgical images.

[0174] Through the above technical solutions, this application constructs a complete system for improving the quality of hysteroscopic surgical image data, from basic data acquisition, physiological time series analysis, health reference generation to abnormal feature enhancement. It effectively solves the problems of significant physiological interference, blurred abnormal features, and insufficient recognition in existing technologies for hysteroscopic surgical images. It can comprehensively optimize the overall quality of hysteroscopic surgical image data, improve the utilization rate and usability of effective image information, and provide high-quality and highly reliable image data support for image processing and applications related to hysteroscopic surgery.

[0175] Example 2, as Figure 2 As shown, based on the same inventive concept as the method for improving the quality of hysteroscopic surgical image data provided in Embodiment 1, this embodiment of the invention also provides a system for improving the quality of hysteroscopic surgical image data, including:

[0176] The data acquisition module 11 is used to acquire hysteroscopic images of the target patient at a specific time point, and to acquire multimodal physiological time-series data of the target patient before and after the specific time point.

[0177] Phase information calculation module 12 is used to perform temporal feature fusion analysis on the multimodal physiological time series data to obtain individual physiological cycle phase information corresponding to the specific time point;

[0178] The standard image generation module 13 is used to input the individual physiological cycle phase information into a pre-trained healthy uterine cavity image generation agent and output a standard healthy uterine cavity reference image corresponding to the specific time point.

[0179] The contrast enhancement module 14 is used to compare the hysteroscopic image with the standard healthy uterine cavity reference image, generate a pathological residual feature map, and enhance the display of the pathological residual feature map.

[0180] Specifically, the data acquisition module 11 is used for:

[0181] Multimodal physiological time-series data include continuously monitored basal body temperature sequences, serum hormone level test data, time-series data of endometrial morphological characteristics extracted from pelvic ultrasound images, and menstrual history information recorded in electronic medical records.

[0182] Specifically, the phase information calculation module 12 is used for:

[0183] Based on the menstrual history information recorded in the electronic medical records of the multimodal physiological time series data, the menstrual cycle regularity score of the target patient is calculated;

[0184] Obtain a pre-constructed phase estimation submodule array, wherein the phase estimation submodule array contains M phase estimation submodules;

[0185] Multiply the menstrual cycle regularity score by M and round down to obtain the number of times the phase estimation submodule is called, N, where N is a positive integer between 1 and M;

[0186] N phase estimation submodules are randomly selected from the phase estimation submodule array. The selected N phase estimation submodules process the multimodal physiological time series data respectively and output N intermediate phase information accordingly.

[0187] The N intermediate phase information are fused and calculated to generate individual physiological cycle phase information.

[0188] Furthermore, the step of "calculating the menstrual cycle regularity score of the target patient based on the menstrual history information recorded in the electronic medical record in the multimodal physiological time-series data" includes:

[0189] Extract the sequence of continuous historical menstrual cycle lengths of the target patient from the menstrual history information recorded in the electronic medical record;

[0190] Calculate the standard deviation and mean of the historical menstrual cycle length series, and calculate the ratio of the standard deviation to the mean to obtain the coefficient of variation of cycle length;

[0191] The average absolute rate of change of each period is obtained by calculating the ratio of the sum of the absolute values ​​of the rate of change of each adjacent period in the historical menstrual cycle length sequence to the total number of adjacent period pairs.

[0192] The basic fluctuation sub-score is obtained by dividing the coefficient of variation of the cycle length by a first preset clinical threshold, wherein the value range of the basic fluctuation sub-score is 0~1;

[0193] Divide the mean absolute periodic change rate by the second preset clinical threshold to obtain the drastic change sub-score, wherein the value range of the drastic change sub-score is 0~1;

[0194] The arithmetic mean of the baseline fluctuation sub-score and the dramatic change sub-score is calculated to obtain the menstrual cycle regularity score, wherein the value of the menstrual cycle regularity score is positively correlated with the degree of menstrual irregularity.

[0195] Furthermore, the process of determining the "first preset clinical threshold and the second preset clinical threshold" includes:

[0196] Collect a set of menstrual cycle data containing multiple samples, where each sample contains at least menstrual history information, age, and body mass index;

[0197] Using age segmentation and body mass index classification as joint classification criteria, the samples in the menstrual cycle data set are divided into multiple group subsets. The age segmentation includes at least puberty, reproductive age, and perimenopause, and the body mass index classification includes at least underweight, normal, overweight, and obese ranges.

[0198] For each of the aforementioned subsets of the population, the arithmetic mean of the coefficients of variation of the period length of all samples within it is calculated and used as the first clinical reference threshold for each subset of the population.

[0199] For each of the aforementioned subsets, the arithmetic mean of the average absolute periodic rate of change of all samples within it is calculated and used as the second clinical reference threshold for each subset.

[0200] Construct a threshold mapping library to store the mapping relationship between each population subset and the corresponding first clinical reference threshold and second clinical reference threshold;

[0201] The target patient's age and body mass index are used to determine the subset of the population to which they belong, and the corresponding first clinical reference threshold and second clinical reference threshold are obtained from the threshold mapping relationship library, which are respectively used as the first preset clinical threshold and the second preset clinical threshold for the target patient.

[0202] Furthermore, the step of "fusing and calculating the N intermediate phase information to generate individual physiological cycle phase information" includes:

[0203] Calculate the arithmetic mean of the N intermediate phase information to obtain the average phase vector;

[0204] The average phase vector is input to a pre-trained phase decoder to map individual physiological cycle phase information, wherein the individual physiological cycle phase information is a scalar value used to quantitatively represent the position of a specific time point in the menstrual cycle of the target patient individual.

[0205] The standard image generation module 13 is specifically used for:

[0206] Hysteroscopic images were collected from multiple healthy volunteers at multiple preset time points during their complete menstrual cycle to form a training image set;

[0207] For each hysteroscopic image in the training image set, the corresponding individual physiological cycle phase information is calculated based on the multimodal physiological time series data recorded synchronously during acquisition, and the individual physiological cycle phase information is used as the annotation for this hysteroscopic image.

[0208] Construct a conditional generative adversarial network that includes a generator and a discriminator;

[0209] The generator is input with the individual's physiological cycle phase information as a condition, and the discriminator is input with the corresponding hysteroscopic image as a real sample. Through adversarial training, the generator learns the mapping relationship from the individual's physiological cycle phase information to a healthy hysteroscopic image under the individual's physiological cycle phase information.

[0210] After training, the generator in the conditional generative adversarial network is defined as a healthy uterine cavity image generating agent.

[0211] The contrast enhancement module 14 is specifically used for:

[0212] An encoder-decoder network structure is adopted, and some network layer parameters are shared with the generator in the healthy uterine cavity image generation agent to construct a pathological feature decoupling device.

[0213] The pathological feature decoupler is trained using the reconstruction loss function and the adversarial difference loss function as joint constraints to obtain the trained pathological feature decoupler. The reconstruction loss function is used to constrain that the hysteroscopic image can be reconstructed by pixel-level addition of the standard healthy uterine cavity reference image and the pathological residual feature map to be generated. The adversarial difference loss function is used to constrain the minimization of the distance between the pathological residual feature map and the residual feature distribution of a set of real pathological samples in the feature space.

[0214] The trained pathological feature decoupling device uses the standard healthy uterine cavity reference image as a health benchmark to perform feature decoupling calculation on the hysteroscopic image, separates cycle-related features and cycle-independent residual features, and maps the cycle-independent residual features to generate a pathological residual feature map.

[0215] Furthermore, the "enhanced display of the pathological residual feature map" includes:

[0216] The pathological residual feature map is subjected to pseudo-color mapping to generate a pseudo-color enhanced map.

[0217] The pseudo-color enhanced image and the hysteroscopic image are spatially registered and pixel-level fused to generate a difference-enhanced display image.

[0218] Morphological analysis is performed on the high-response regions in the difference-enhanced display image to extract the boundary contours and area parameters of the high-response regions, and the boundary contours and area parameters are superimposed on the difference-enhanced display image in the form of graphical annotations.

[0219] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0220] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0221] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0222] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1The function specified in one or more boxes.

[0223] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0224] Although preferred embodiments of the invention have been described, those skilled in the art, once they have learned the basic inventive concept, can make other changes and modifications to these embodiments.

[0225] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for improving the quality of hysteroscopic surgical image data, characterized in that, The method includes: Acquire hysteroscopic images of the target patient at a specific time point, and acquire multimodal physiological time-series data of the target patient before and after the specific time point; The multimodal physiological time-series data are subjected to time-series feature fusion analysis to obtain individual physiological cycle phase information corresponding to the specific time point; The individual physiological cycle phase information is input into a pre-trained healthy uterine cavity image generation agent, and a standard healthy uterine cavity reference image corresponding to the specific time point is output. The hysteroscopic images are compared with the standard healthy uterine cavity reference images to generate a pathological residual feature map, and the pathological residual feature map is enhanced for display. The process involves performing temporal feature fusion analysis on the multimodal physiological time-series data to obtain individual physiological cycle phase information corresponding to the specific time point, including: Based on the menstrual history information recorded in the electronic medical records of the multimodal physiological time series data, the menstrual cycle regularity score of the target patient is calculated; Obtain a pre-constructed phase estimation submodule array, wherein the phase estimation submodule array contains M phase estimation submodules; Multiply the menstrual cycle regularity score by M and round down to obtain the number of times the phase estimation submodule is called, N, where N is a positive integer between 1 and M; N phase estimation submodules are randomly selected from the phase estimation submodule array. The selected N phase estimation submodules process the multimodal physiological time series data respectively and output N intermediate phase information accordingly. Calculate the arithmetic mean of the N intermediate phase information to obtain the average phase vector; The average phase vector is input to a pre-trained phase decoder to map individual physiological cycle phase information, wherein the individual physiological cycle phase information is a scalar value used to quantitatively represent the position of a specific time point in the menstrual cycle of the target patient individual.

2. The method for improving the quality of hysteroscopic surgical image data according to claim 1, characterized in that, Multimodal physiological time-series data include continuously monitored basal body temperature sequences, serum hormone level test data, time-series data of endometrial morphological characteristics extracted from pelvic ultrasound images, and menstrual history information recorded in electronic medical records.

3. The method for improving the quality of hysteroscopic surgical image data according to claim 1, characterized in that, Based on the menstrual history information recorded in the electronic medical records of the multimodal physiological time-series data, the menstrual cycle regularity score of the target patient is calculated, including: Extract the sequence of continuous historical menstrual cycle lengths of the target patient from the menstrual history information recorded in the electronic medical record; Calculate the standard deviation and mean of the historical menstrual cycle length series, and calculate the ratio of the standard deviation to the mean to obtain the coefficient of variation of cycle length; The average absolute rate of change of each period is obtained by calculating the ratio of the sum of the absolute values ​​of the rate of change of each adjacent period in the historical menstrual cycle length sequence to the total number of adjacent period pairs. The basic fluctuation sub-score is obtained by dividing the coefficient of variation of the cycle length by a first preset clinical threshold, wherein the value range of the basic fluctuation sub-score is 0~1; Divide the mean absolute periodic change rate by the second preset clinical threshold to obtain the drastic change sub-score, wherein the value range of the drastic change sub-score is 0~1; The arithmetic mean of the baseline fluctuation sub-score and the dramatic change sub-score is calculated to obtain the menstrual cycle regularity score, wherein the value of the menstrual cycle regularity score is positively correlated with the degree of menstrual irregularity.

4. The method for improving the quality of hysteroscopic surgical image data according to claim 3, characterized in that, The process of determining the first and second preset clinical thresholds includes: Collect a set of menstrual cycle data containing multiple samples, where each sample contains at least menstrual history information, age, and body mass index; Using age segmentation and body mass index classification as joint classification criteria, the samples in the menstrual cycle data set are divided into multiple group subsets. The age segmentation includes at least puberty, reproductive age, and perimenopause, and the body mass index classification includes at least underweight, normal, overweight, and obese ranges. For each of the aforementioned subsets of the population, the arithmetic mean of the coefficients of variation of the period length of all samples within it is calculated and used as the first clinical reference threshold for each subset of the population. For each of the aforementioned subsets, the arithmetic mean of the average absolute periodic rate of change of all samples within it is calculated and used as the second clinical reference threshold for each subset. Construct a threshold mapping library to store the mapping relationship between each population subset and the corresponding first clinical reference threshold and second clinical reference threshold; The target patient's age and body mass index are used to determine the subset of the population to which they belong, and the corresponding first clinical reference threshold and second clinical reference threshold are obtained from the threshold mapping relationship library, which are respectively used as the first preset clinical threshold and the second preset clinical threshold for the target patient.

5. The method for improving the quality of hysteroscopic surgical image data according to claim 1, characterized in that, The process of constructing an intelligent agent from healthy uterine cavity images includes: Hysteroscopic images were collected from multiple healthy volunteers at multiple preset time points during their complete menstrual cycle to form a training image set; For each hysteroscopic image in the training image set, the corresponding individual physiological cycle phase information is calculated based on the multimodal physiological time series data recorded synchronously during acquisition, and the individual physiological cycle phase information is used as the annotation for this hysteroscopic image. Construct a conditional generative adversarial network that includes a generator and a discriminator; The generator is input with the individual's physiological cycle phase information as a condition, and the discriminator is input with the corresponding hysteroscopic image as a real sample. Through adversarial training, the generator learns the mapping relationship from the individual's physiological cycle phase information to a healthy hysteroscopic image under the individual's physiological cycle phase information. After training, the generator in the conditional generative adversarial network is defined as a healthy uterine cavity image generating agent.

6. The method for improving the quality of hysteroscopic surgical image data according to claim 1, characterized in that, Comparing the hysteroscopic images with the standard healthy uterine cavity reference image, a pathological residual feature map is generated to enhance the display of regions in the hysteroscopic images that deviate from normal physiological structures, including: An encoder-decoder network structure is adopted, and some network layer parameters are shared with the generator in the healthy uterine cavity image generation agent to construct a pathological feature decoupling device. The pathological feature decoupler is trained using the reconstruction loss function and the adversarial difference loss function as joint constraints to obtain the trained pathological feature decoupler. The reconstruction loss function is used to constrain that the hysteroscopic image can be reconstructed by pixel-level addition of the standard healthy uterine cavity reference image and the pathological residual feature map to be generated. The adversarial difference loss function is used to constrain the minimization of the distance between the pathological residual feature map and the residual feature distribution of a set of real pathological samples in the feature space. The trained pathological feature decoupling device uses the standard healthy uterine cavity reference image as a health benchmark to perform feature decoupling calculation on the hysteroscopic image, separates cycle-related features and cycle-independent residual features, and maps the cycle-independent residual features to generate a pathological residual feature map.

7. The method for improving the quality of hysteroscopic surgical image data according to claim 1, characterized in that, Enhanced display of the pathological residual feature map includes: The pathological residual feature map is subjected to pseudo-color mapping to generate a pseudo-color enhanced map. The pseudo-color enhanced image and the hysteroscopic image are spatially registered and pixel-level fused to generate a difference-enhanced display image. Morphological analysis is performed on the high-response regions in the difference-enhanced display image to extract the boundary contours and area parameters of the high-response regions, and the boundary contours and area parameters are superimposed on the difference-enhanced display image in the form of graphical annotations.

8. A system for improving the quality of hysteroscopic surgical image data, characterized in that, A method for improving the quality of hysteroscopic surgical image data according to any one of claims 1-7 includes: The data acquisition module is used to acquire hysteroscopic images of the target patient at a specific time point, and to acquire multimodal physiological time-series data of the target patient before and after the specific time point. The phase information calculation module is used to perform temporal feature fusion analysis on the multimodal physiological time series data to obtain the individual physiological cycle phase information corresponding to the specific time point; The standard image generation module is used to input the individual's physiological cycle phase information into a pre-trained healthy uterine cavity image generation agent and output a standard healthy uterine cavity reference image corresponding to the specific time point. The contrast enhancement module is used to compare the hysteroscopic image with the standard healthy uterine cavity reference image, generate a pathological residual feature map, and enhance the display of the pathological residual feature map.

Images