Human brain tumor prediction model training system and training method

Abstract

The invention relates to the technical field of tumor prediction, in particular to a human brain tumor prediction model training system and method. The selection analysis unit is used for determining a data source state according to the heterogeneous crossover abundance and the ROI characteristic coefficient and determining a data selection mode according to the data source state; the extension selection unit is used for determining a processing mode of each data source according to a tumor data category in multi-data-source extension selection; the supplementary selection unit is used for determining a representative data source based on a data source evaluation index, determining representative combinations according to ROI similarity and effective similarity and determining a supplementary selection mode of each representative combination according to a combination comparison coefficient in supplementary selection of the representative data source; a model training unit; the stability of a model prediction result can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of tumor prediction, and in particular to a human brain tumor prediction model training system and training method. Background Art

[0002] In oncology research and clinical practice, establishing accurate tumor prediction models is crucial for developing personalized treatment plans and evaluating treatment efficacy. However, due to the complexity and heterogeneity of tumors, a single data source often fails to provide the comprehensive and accurate information required for model training. Therefore, selecting training data to improve the stability of tumor prediction models is a pressing technical challenge for those skilled in the art.

[0003] Chinese patent publication number CN119361146A discloses a tumor recurrence risk assessment and early warning system based on data analysis, including: multiple data collection modules, a big data processing platform, a model architecture system, and an individual data processing system. The present invention solves the problems of the existing technology lacking big data support, limited individualized prediction capabilities, and insufficient real-time and continuous monitoring. The present invention integrates large-scale historical case data and uses these rich data sets for model training, which can significantly improve the accuracy and generalization ability of the prediction model. It also integrates multiple data sources such as genomic information, living habits, and environmental factors to provide each patient with a personalized risk assessment, and a real-time monitoring interface connects to various smart medical devices to collect changes in patients' daily physiological parameters, providing continuous health management recommendations for patients after tumor surgery. It can be seen that the above technical solution has the following problems: the training data is not screened, and when the balance of data in different diagnosis and treatment cycles is poor, the model cannot fully learn the characteristics of minority data, resulting in poor stability of the model prediction results. Summary of the Invention

[0004] To this end, the present invention provides a human brain tumor prediction model training system and training method to overcome the problem in the prior art that when the balance of data in different diagnosis and treatment cycles is poor, the model cannot fully learn the characteristics of minority data, resulting in poor stability of the model prediction results.

[0005] To achieve the above objectives, the present invention provides a human brain tumor prediction model training system, comprising:

[0006] a data acquisition unit, configured to acquire tumor data from a plurality of data sources;

[0007] A selection and analysis unit is connected to the data acquisition unit and is used to determine the data source status according to the heterogeneous cross abundance and the ROI characteristic coefficient, and determine the data selection method according to the data source status, wherein the data selection method is multi-data source extended selection or representative data source supplementary selection;

[0008] an extended selection unit, connected to the data acquisition unit and the selection and analysis unit, respectively, for determining, in the extended selection of multiple data sources, a processing method for each data source according to the tumor data category, the processing method including determining a first screening method according to a training difficulty coefficient and determining a second screening method according to a determination condition;

[0009] The first screening method is to select tumor data based on angiographic characteristic values and rhythm risk, or to select tumor data based on tumor reference values; the second screening method is to select tumor data based on associated ROI characteristic values and mediated diffusion index, or to select tumor data based on data evaluation values;

[0010] a supplementary selection unit, connected to the data acquisition unit and the selection and analysis unit, respectively, for determining, in the supplementary selection of representative data sources, a representative data source based on a data source evaluation index, determining a representative combination based on ROI similarity and effective similarity, and determining a supplementary selection method for each representative combination based on a combination comparison coefficient, wherein the supplementary selection method is to select tumor data based on characteristic keywords or dynamic difference coefficients;

[0011] A model training unit is connected to the extended selection unit and the supplementary selection unit respectively, and is used to train the tumor prediction model using the selected tumor data as training data.

[0012] Furthermore, the selection and analysis unit responds to the data source status to determine the data selection method, including:

[0013] If the data source status of the selected analysis unit response is that the heterogeneous cross abundance is less than the preset heterogeneous cross abundance or the ROI characteristic coefficient is less than the preset ROI characteristic coefficient, then the data selection method is determined to be multi-data source extended selection;

[0014] If the data source status of the selected analysis unit response is that the heterogeneous cross abundance is greater than or equal to the preset heterogeneous cross abundance and the ROI characteristic coefficient is greater than or equal to the preset ROI characteristic coefficient, the data selection method is determined to be representative data source supplementary selection.

[0015] Furthermore, the extended selection unit determines the tumor data category based on the diagnosis and treatment effectiveness and the diagnosis and treatment fluctuation value. The tumor data category includes:

[0016] A type of tumor data with a diagnosis and treatment effectiveness greater than or equal to the preset diagnosis and treatment effectiveness and a diagnosis and treatment fluctuation value greater than or equal to the preset diagnosis and treatment fluctuation value;

[0017] Category II tumor data whose diagnosis and treatment effectiveness is less than the preset diagnosis and treatment effectiveness or whose diagnosis and treatment fluctuation value is less than the preset diagnosis and treatment fluctuation value.

[0018] Furthermore, the extended selection unit determines a processing method for each data source according to the tumor data category, including:

[0019] For each type of tumor data in a single data source, the processing method is to determine the first screening method according to the training difficulty coefficient;

[0020] For each type II tumor data in a single data source, the processing method is to determine the second screening method according to the judgment conditions.

[0021] Furthermore, the extended selection unit determines a first screening method according to the training difficulty coefficient, including:

[0022] If the training difficulty coefficient is greater than or equal to the preset training difficulty coefficient, it is determined that the first screening method is to select tumor data based on the angiographic characteristic value and the rhythm risk;

[0023] If the training difficulty coefficient is less than the preset training difficulty coefficient, it is determined that the first screening method is to select tumor data according to the tumor reference value.

[0024] Furthermore, the method for confirming the training difficulty coefficient includes:

[0025] If the heterogeneous correlation is greater than or equal to the preset heterogeneous correlation, the training difficulty coefficient is determined according to the heterogeneous interaction difficulty;

[0026] If the heterogeneous correlation is less than the preset heterogeneous correlation, the training difficulty coefficient is determined according to the homogeneous correlation coefficient.

[0027] Furthermore, the extended selection unit responds to the determination condition to determine the second screening method, including:

[0028] The determination condition of the extended selection unit response is that the edge fuzzy coefficient is greater than or equal to the preset edge fuzzy coefficient and the ROI comparison coefficient is less than the preset ROI comparison coefficient, then the second screening method is determined to select tumor data based on the associated ROI characteristic value and the mediated diffusion index;

[0029] The determination condition of the extended selection unit response is that the edge fuzzy coefficient is less than the preset edge fuzzy coefficient or the ROI comparison coefficient is greater than or equal to the preset ROI comparison coefficient, then the second screening method is determined to select tumor data according to the data evaluation value.

[0030] Furthermore, the method for confirming the associated ROI feature value includes:

[0031] Determine a preliminary characterization interval according to the image characterization value, and increase and adjust the area of the preliminary characterization interval according to the specific index of the preliminary characterization interval, and record the increased and adjusted preliminary characterization interval as the characterization interval;

[0032] The associated ROI feature value is determined according to the ratio of the mutation coefficient corresponding to the characterization interval to the area of the characterization interval.

[0033] Furthermore, the supplementary selection unit determines a representative data source based on the data source evaluation index, determines a representative combination based on the ROI similarity and the effective similarity, and determines a supplementary selection method for each representative combination based on the combination comparison coefficient;

[0034] For a single representative combination,

[0035] If the combined comparison coefficient is less than the preset combined comparison coefficient, it is determined that the supplementary selection method is to select tumor data based on the characteristic keywords;

[0036] If the combined comparison coefficient is greater than or equal to the preset combined comparison coefficient, it is determined that the supplementary selection method is to select tumor data according to the dynamic difference coefficient.

[0037] The present invention also provides a method for training a human brain tumor prediction model, comprising:

[0038] Obtain tumor data from multiple data sources, determine the data source status based on heterogeneous cross-abundance and ROI characteristic coefficients, and determine a data selection method based on the data source status, which can be extended selection of multiple data sources or supplementary selection of representative data sources;

[0039] In the extended selection of multiple data sources, the processing method of each data source is determined according to the tumor data category. The processing method includes determining a first screening method based on the training difficulty coefficient and determining a second screening method based on the judgment condition;

[0040] The first screening method is to select tumor data based on angiographic characteristic values and rhythm risk, or to select tumor data based on tumor reference values; the second screening method is to select tumor data based on associated ROI characteristic values and mediated diffusion index, or to select tumor data based on data evaluation values;

[0041] In the supplementary selection of representative data sources, the representative data source is determined based on the data source evaluation index, the representative combination is determined based on the ROI similarity and effective similarity, and the supplementary selection method for each representative combination is determined based on the combination comparison coefficient. The supplementary selection method is to select tumor data based on characteristic keywords or dynamic difference coefficients;

[0042] The selected tumor data is used as training data to train the tumor prediction model.

[0043] Compared with the prior art, the beneficial effect of the present invention lies in that, in the technical solution of the present invention, the data source status is determined according to the heterogeneous cross abundance and the ROI characteristic coefficient, and the richness of the data source data is effectively reflected by the heterogeneous cross abundance and the ROI characteristic coefficient, and then different data selection methods are adaptively selected according to the data source status, so that the selection of data selection method is more in line with the actual application scenario, which can increase the diversity of tumor data and improve the accuracy of model prediction.

[0044] Furthermore, the present invention determines the tumor data category based on the diagnosis and treatment effectiveness and the diagnosis and treatment fluctuation value, and effectively reflects the diagnosis and treatment effect of the diagnosis and treatment cycle through the diagnosis and treatment effectiveness and the diagnosis and treatment fluctuation value, and then determines the processing method of each data source according to the tumor data category. The processing method can be dynamically adjusted to improve the balance of data in different diagnosis and treatment cycles, thereby improving the stability of the tumor prediction model prediction.

[0045] Furthermore, the present invention effectively reflects the training difficulty of tumor data through the training difficulty coefficient, and then determines the first screening method according to the training difficulty coefficient, avoiding the problem in the prior art that when selecting training data, the model prediction results are biased when similar characteristics of tumors with different treatment cycles are not considered. The ability to select tumor data according to angiographic feature values and rhythm risk can reveal hidden differences in tumor data, improve the richness of data selection, and thereby improve the accuracy and generalization ability of the tumor prediction model.

[0046] Furthermore, the present invention effectively reflects the similarity of the two types of tumor data and the degree of fuzziness of the ROI area through the judgment conditions, and then adaptively selects different second screening methods according to the judgment conditions, which can more comprehensively consider the multi-faceted characteristics of the tumor, avoid the problem of unbalanced model training data selection due to insufficient similarity of a single indicator, take into account the impact of individual differences on data selection, optimize the model training effect, and thus improve the prediction accuracy and stability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 A unit connection diagram of the human brain tumor prediction model training system of the present invention;

[0048] Figure 2 This is a flow chart of the present invention for determining a data selection method based on a data source state;

[0049] Figure 3 This is a flow chart of determining a first screening method according to a training difficulty coefficient according to the present invention;

[0050] Figure 4 Schematic diagram of the human brain tumor prediction model training method of the present invention. DETAILED DESCRIPTION

[0051] In order to make the objects and advantages of the present invention more clearly understood, the present invention is further described below in conjunction with embodiments; it should be understood that the specific embodiments described herein are merely used to explain the present invention and are not intended to limit the present invention.

[0052] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood by those skilled in the art that these embodiments are only used to explain the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0053] It should be noted that, in the description of the present invention, terms such as "up", "down", "left", "right", "inside", and "outside" indicating directions or positional relationships are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and does not indicate or imply that the device or element must have a specific orientation, be constructed and operated in a specific orientation. Therefore, it cannot be understood as a limitation on the present invention.

[0054] Furthermore, it should be noted that, in the description of the present invention, unless otherwise expressly specified or limited, the terms "mounted," "connected," and "connected" should be understood in a broad sense. For example, they may refer to fixed connections, detachable connections, or integral connections; mechanical connections or electrical connections; direct connections or indirect connections through an intermediate medium; and internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0055] See also Figures 1 to 3 As shown, the present invention provides a human brain tumor prediction model training system, comprising:

[0056] a data acquisition unit, configured to acquire tumor data from a plurality of data sources;

[0057] A selection and analysis unit is connected to the data acquisition unit and is used to determine the data source status according to the heterogeneous cross abundance and the ROI characteristic coefficient, and determine the data selection method according to the data source status, wherein the data selection method is multi-data source extended selection or representative data source supplementary selection;

[0058] an extended selection unit, connected to the data acquisition unit and the selection and analysis unit, respectively, for determining, in the extended selection of multiple data sources, a processing method for each data source according to the tumor data category, the processing method including determining a first screening method according to a training difficulty coefficient and determining a second screening method according to a determination condition;

[0059] The first screening method is to select tumor data based on angiographic characteristic values and rhythm risk, or to select tumor data based on tumor reference values; the second screening method is to select tumor data based on associated ROI characteristic values and mediated diffusion index, or to select tumor data based on data evaluation values;

[0060] a supplementary selection unit, connected to the data acquisition unit and the selection and analysis unit, respectively, for determining, in the supplementary selection of representative data sources, a representative data source based on a data source evaluation index, determining a representative combination based on ROI similarity and effective similarity, and determining a supplementary selection method for each representative combination based on a combination comparison coefficient, wherein the supplementary selection method is to select tumor data based on characteristic keywords or dynamic difference coefficients;

[0061] A model training unit is connected to the extended selection unit and the supplementary selection unit respectively, and is used to train the tumor prediction model using the selected tumor data as training data.

[0062] The application scenario of the present invention is the training of a prognosis prediction model for human brain tumors. The present invention includes several data sources, and a single data source stores several tumor data. The single tumor data includes treatment data corresponding to different consultation times of a single patient. The treatment data corresponding to a single consultation time includes examination information, medication records, condition description, and treatment plan. The examination information includes but is not limited to human brain X-ray images, blood test information, and ultrasound contrast information. The blood test information includes but is not limited to CA15-3 concentration values and red blood cell counts. Each human brain X-ray image corresponds to an ROI area. The ROI area is the tumor area manually circled by the doctor using an outlining tool. The ultrasound contrast information includes but is not limited to TIC images and ultrasound contrast videos. The TIC image is a two-dimensional image with time as the horizontal axis and the signal intensity of the contrast agent as the vertical axis. This is content that is easy for technicians in this field to understand and will not be described in detail.

[0063] The present invention provides several historical records, any of which records the heterogeneous cross abundance, ROI feature coefficient, diagnosis and treatment effectiveness, diagnosis and treatment fluctuation value, training difficulty coefficient, and edge fuzzy coefficient in at least one historical process of training a human brain tumor prediction model. Each historical record corresponds to a qualified mark, which records whether the training process of the human brain tumor prediction model meets user requirements. The qualified mark can be recorded manually. It is understandable that the user can determine whether the training process of the human brain tumor prediction model meets the requirements based on self-set indicators. The self-set indicators can be but are not limited to the misjudgment rate, which will not be elaborated here. The misjudgment rate is the number of times the tumor prediction model is used to incorrectly predict the tumor treatment plan.

[0064] Using the selected tumor data as training data to train the tumor prediction model can make the prediction results of the tumor prediction model more accurate. The specific training process is easy to understand for those skilled in the art and will not be described in detail.

[0065] Specifically, the selection and analysis unit responds to the data source status to determine the data selection method, including:

[0066] If the data source status of the selected analysis unit response is that the heterogeneous cross abundance is less than the preset heterogeneous cross abundance or the ROI characteristic coefficient is less than the preset ROI characteristic coefficient, then the data selection method is determined to be multi-data source extended selection;

[0067] If the data source status of the selected analysis unit response is that the heterogeneous cross abundance is greater than or equal to the preset heterogeneous cross abundance and the ROI characteristic coefficient is greater than or equal to the preset ROI characteristic coefficient, the data selection method is determined to be representative data source supplementary selection.

[0068] The data source state includes a first data source state and a second data source state. The first data source state is that the heterogeneous cross abundance is less than the preset heterogeneous cross abundance or the ROI characteristic coefficient is less than the preset ROI characteristic coefficient. The second data source state is that the heterogeneous cross abundance is greater than or equal to the preset heterogeneous cross abundance and the ROI characteristic coefficient is greater than or equal to the preset ROI characteristic coefficient.

[0069] The heterogeneous cross abundance is the maximum value among the sub-heterogeneous cross abundances corresponding to each data source. For a single data source, this data source is recorded as the target data source, and all data sources other than the target data source are recorded as reference data sources. The sub-heterogeneous cross abundance corresponding to the target data source = the number of identical keywords / the total number of keywords appearing in the treatment regimen corresponding to each tumor data in each data source. The treatment regimen corresponding to each tumor data in each reference data source is recorded as the reference record, and the treatment regimen corresponding to each tumor data in the target data source is recorded as the target record. The number of identical keywords is the number of keywords that appear in both the target record and the reference record.

[0070] The ROI characteristic coefficient is the maximum value among the sub-ROI characteristic coefficients corresponding to each data source. The sub-ROI characteristic coefficient corresponding to a single data source = ROI fluctuation coefficient + diagnosis and treatment fluctuation coefficient;

[0071] The ROI fluctuation coefficient is the standard deviation of the ROI reference values corresponding to each tumor data in a single data source. The ROI reference value corresponding to a single tumor data is the area of the ROI region in the human brain X-ray image taken at the time of the patient's first visit in the tumor data. The diagnosis and treatment fluctuation coefficient is the standard deviation of the diagnosis and treatment effectiveness corresponding to each tumor data in a single data source. The diagnosis and treatment effectiveness corresponding to a single tumor data = the CA15-3 concentration value corresponding to the treatment data with the earliest visit time for the patient in the tumor data - the CA15-3 concentration value corresponding to the treatment data with the latest visit time for the patient in the tumor data. The CA15-3 concentration value is generally measured using an enzyme-linked immunosorbent assay or a chemiluminescence immunoassay. This is easily understood by those skilled in the art and is not described in detail here.

[0072] The values of the preset heterogeneous cross abundance and the preset ROI characteristic coefficient can be determined by the user according to the actual application scenario. The larger the values of the preset heterogeneous cross abundance and the preset ROI characteristic coefficient, the greater the user's demand for extended selection of multiple data sources. A preset heterogeneous cross abundance and a preset ROI characteristic coefficient are provided, and the historical records for supplementary selection of representative data sources are detected. The average value of the heterogeneous cross abundance corresponding to the historical records that can meet the user's needs is recorded as the preset heterogeneous cross abundance, and the average value of the ROI characteristic coefficient corresponding to the historical records that can meet the user's needs is recorded as the preset ROI characteristic coefficient.

[0073] Specifically, the extended selection unit determines the tumor data category based on the diagnosis and treatment effectiveness and the diagnosis and treatment fluctuation value. The tumor data categories include:

[0074] A type of tumor data with a diagnosis and treatment effectiveness greater than or equal to the preset diagnosis and treatment effectiveness and a diagnosis and treatment fluctuation value greater than or equal to the preset diagnosis and treatment fluctuation value;

[0075] Category II tumor data whose diagnosis and treatment effectiveness is less than the preset diagnosis and treatment effectiveness or whose diagnosis and treatment fluctuation value is less than the preset diagnosis and treatment fluctuation value.

[0076] Among them, the diagnosis and treatment fluctuation value = the standard deviation of the CA15-3 concentration values corresponding to each visit time in the single tumor data / the proportion of the frequency of decrease, the proportion of the frequency of decrease = the number of decrease visits / (the number of visit times in the single tumor data - 1), for each visit time in the single tumor data, all visit times except the earliest visit time are recorded as reference visit times, and the method for confirming the decrease visit time is as follows: for a single reference visit time, if the CA15-3 concentration value of the reference visit time is greater than the CA15-3 concentration value of the adjacent visit time corresponding to the reference visit time, then the reference visit time is recorded as the decrease visit time, and the adjacent visit time corresponding to the single reference visit time is the visit time that is adjacent to the reference visit time and earlier than the reference visit time;

[0077] The values of the preset diagnosis and treatment effectiveness and the preset diagnosis and treatment fluctuation value can be determined by the user according to the actual application scenario. The smaller the values of the preset diagnosis and treatment effectiveness and the preset diagnosis and treatment fluctuation value, the greater the user's need to determine the first screening method according to the training difficulty coefficient. A value of the preset diagnosis and treatment effectiveness and the preset diagnosis and treatment fluctuation value is provided, and the historical records of determining the first screening method according to the training difficulty coefficient are detected. The average value of the diagnosis and treatment effectiveness corresponding to the historical records that can meet the user's needs is recorded as the preset diagnosis and treatment effectiveness, and the average value of the diagnosis and treatment fluctuation values corresponding to the historical records that can meet the user's needs is recorded as the preset diagnosis and treatment fluctuation value.

[0078] Specifically, the extended selection unit determines the processing method of each data source according to the tumor data category, including:

[0079] For each type of tumor data in a single data source, the processing method is to determine the first screening method according to the training difficulty coefficient;

[0080] For each type II tumor data in a single data source, the processing method is to determine the second screening method according to the judgment conditions.

[0081] Specifically, the extended selection unit determines the first screening method according to the training difficulty coefficient, including:

[0082] If the training difficulty coefficient is greater than or equal to the preset training difficulty coefficient, it is determined that the first screening method is to select tumor data based on the angiographic characteristic value and the rhythm risk;

[0083] If the training difficulty coefficient is less than the preset training difficulty coefficient, it is determined that the first screening method is to select tumor data according to the tumor reference value.

[0084] The value of the preset training difficulty coefficient can be determined by the user according to the actual application scenario. The larger the value of the preset training difficulty coefficient, the greater the user's need to select tumor data according to the tumor reference value. A value of the preset training difficulty coefficient is provided, and the historical records of selecting tumor data according to the tumor reference value are detected. The average value of the training difficulty coefficients corresponding to the historical records that can meet the user's needs is recorded as the preset training difficulty coefficient.

[0085] When selecting tumor data based on angiographic feature values or rhythm risk, the first tumor data of the first category in the first reference sequence is used as a starting point, and the tumor data of the first category in the first reference sequence are sequentially selected as the selected training data according to a preset interval. The first preset interval is the number of tumor data of the first category between two adjacent tumor data of the first category in the selected first reference sequence. The first preset interval is positively correlated with the total amount of tumor data of the first category in the data source.

[0086] The first reference sequence is confirmed by:

[0087] When selecting tumor data based on angiographic feature values and rhythm risk, the first reference sequence is a sequence in which each type of tumor data in a single data source is sorted in descending order according to the comprehensive evaluation value, where comprehensive evaluation value = angiographic feature value + rhythm risk;

[0088] When selecting tumor data based on tumor reference values, the first reference sequence is a sequence in which each type of tumor data in a single data source is sorted in descending order according to the tumor reference values corresponding to the time of first visit;

[0089] Rhythm risk = tumor reference value corresponding to the first visit time of a single Class I tumor data / rhythm disorder coefficient. Rhythm disorder coefficient = |peak coefficient corresponding to the first visit time of a single Class I tumor data - preset peak coefficient|. The preset peak coefficient is determined by taking the tumor reference value corresponding to the first visit time of a single Class I tumor data as a1, and recording the average of the peak coefficients corresponding to the Class I tumor data whose first visit time is a1 in the historical records as the preset peak coefficient.

[0090] Contrast characteristic value = peak coefficient + diffusion coefficient, peak coefficient = signal intensity corresponding to peak time / time interval from initial time to peak time; for a single tumor data set, the time point corresponding to the maximum signal intensity in the TIC image corresponding to the first visit time of the tumor data is recorded as the peak time, the end time is the maximum time point corresponding to the TIC curve in the TIC image corresponding to the first visit time, and the initial time is the minimum time point corresponding to the TIC curve in the TIC image corresponding to the first visit time. Diffusion coefficient = (area of the first region / time interval from peak time to end time) - (area of the second region / time interval from initial time to peak time);

[0091] The straight line passing through the peak time and perpendicular to the x-axis is recorded as the first straight line, the straight line passing through the end time and perpendicular to the x-axis is recorded as the second straight line, and the straight line passing through the initial time and perpendicular to the x-axis is recorded as the third straight line. The area of the first region is the area of the closed region between the first straight line, the second straight line, the x-axis and the TIC curve. The area of the second region is the area of the closed region between the first straight line, the third straight line, the x-axis and the TIC curve.

[0092] Specifically, the method for determining the training difficulty coefficient includes:

[0093] If the heterogeneous correlation is greater than or equal to the preset heterogeneous correlation, the training difficulty coefficient is determined according to the heterogeneous interaction difficulty;

[0094] If the heterogeneous correlation is less than the preset heterogeneous correlation, the training difficulty coefficient is determined according to the homogeneous correlation coefficient.

[0095] If the heterogeneous correlation is greater than or equal to the preset heterogeneous correlation, the training difficulty coefficient is the average value of the heterogeneous interaction difficulty corresponding to each type of tumor data in a single data source;

[0096] If the heterogeneous correlation is less than the preset heterogeneous correlation, the training difficulty coefficient is positively correlated with the homogeneous correlation coefficient;

[0097] The heterogeneous interaction difficulty corresponding to a single Class I tumor data in a single data source is the maximum value of the difficulty thresholds corresponding to the Class I tumor data and each Class II tumor data in the data source. The difficulty threshold corresponding to a single Class I tumor data and a single Class II tumor data = hidden difference coefficient / ROI similarity. The hidden difference coefficient is the time interval between the target visit time and the first visit time in a single Class I tumor data. For a single Class II tumor data, the tumor reference value corresponding to the earliest visit time of the Class II tumor data is recorded as a, and the tumor reference value of the target visit time corresponding to the single Class I tumor data is recorded as b. ROI similarity = 1 / |ab|. The target visit time is confirmed as follows: Tumor data, detect the deviation coefficient corresponding to each visit time of this type of tumor data, record the visit time with the smallest deviation coefficient as the target visit time, the deviation coefficient corresponding to a single visit time = |tumor reference value corresponding to the visit time - a|, the tumor reference value corresponding to a single visit time = pixel fluctuation coefficient / pixel mean + texture coefficient, texture coefficient = (4π×S) / L2, S is the area of the ROI region in the human brain X-ray image corresponding to a single visit time, L is the perimeter of the ROI region in the human brain X-ray image corresponding to a single visit time, the pixel mean is the average value of the pixel values corresponding to each pixel point in the ROI region, and the pixel fluctuation coefficient is the standard deviation of the pixel values corresponding to each pixel point in the ROI region;

[0098] Intraclass correlation coefficient = 1 / standard deviation of the tumor reference value corresponding to the first visit time of each type of tumor data in a single data source;

[0099] The method for confirming heterogeneous correlation is that, for a single data source, each treatment plan corresponding to each type of tumor data in the data source is recorded as the first record, and each treatment plan corresponding to each type of tumor data in the data source is recorded as the second record. The distribution of the number of keywords in the first record and the second record is recorded as m1 and m2. Heterogeneous correlation = (the number of identical keywords in the first record and the second record) / (the larger value of m1 and m2). The value of the preset heterogeneous correlation can be determined by the user according to the actual application scenario. The smaller the value of the preset heterogeneous correlation, the greater the user's need to determine the training difficulty coefficient based on the difficulty of heterogeneous interaction. A preset heterogeneous correlation value is provided, and the preset heterogeneous correlation is 60%.

[0100] Specifically, the extended selection unit responds to the determination condition to determine the second screening method, including:

[0101] The determination condition of the extended selection unit response is that the edge fuzzy coefficient is greater than or equal to the preset edge fuzzy coefficient and the ROI comparison coefficient is less than the preset ROI comparison coefficient, then the second screening method is determined to select tumor data based on the associated ROI characteristic value and the mediated diffusion index;

[0102] The determination condition of the extended selection unit response is that the edge fuzzy coefficient is less than the preset edge fuzzy coefficient or the ROI comparison coefficient is greater than or equal to the preset ROI comparison coefficient, then the second screening method is determined to select tumor data according to the data evaluation value.

[0103] The determination condition includes a first determination condition and a second determination condition. The first determination condition is that the edge fuzzy coefficient is greater than or equal to a preset edge fuzzy coefficient and the ROI comparison coefficient is less than a preset ROI comparison coefficient. The second determination condition is that the edge fuzzy coefficient is less than the preset edge fuzzy coefficient or the ROI comparison coefficient is greater than or equal to the preset ROI comparison coefficient.

[0104] The edge fuzzy coefficient is the average of the sub-fuzzy coefficients corresponding to each type of tumor data in a single data source. The sub-fuzzy coefficient is determined by recording the human brain X-ray image corresponding to the initial visit time of the type of tumor data as the target image, recording the pixel points corresponding to the edge of the ROI area in the target image as the target pixel points, and recording the pixel points with the same pixel value as the target pixel points and located outside the ROI area as the reference pixel points. The sub-fuzzy coefficient = number of reference pixels / mean distance, where the mean distance is the average of the shortest distances from each reference pixel point to the ROI area.

[0105] The ROI comparison coefficient is the maximum value of the comparison reference values corresponding to each type of tumor data in a single data source, recorded as d1, and the minimum value is recorded as d2. The ROI comparison coefficient = (d1-d2) / d2. The comparison reference value corresponding to a single type of tumor data = the tumor reference value corresponding to the first visit time of the type of tumor data / the ROI reference value corresponding to the type of tumor data;

[0106] The values of the preset edge fuzzy coefficient and the preset ROI comparison coefficient can be determined by the user according to the actual application scenario. The larger the value of the preset edge fuzzy coefficient and the smaller the value of the preset ROI comparison coefficient, the greater the user's need to select tumor data based on the data evaluation value. The values of the preset edge fuzzy coefficient and the preset ROI comparison coefficient are provided, and the historical records of selecting tumor data based on the data evaluation value are detected. The average value of the edge fuzzy coefficients corresponding to the historical records that can meet the user's needs is recorded as the preset edge fuzzy coefficient, and the average value of the ROI comparison coefficients corresponding to the historical records that can meet the user's needs is recorded as the preset ROI comparison coefficient.

[0107] The mediated diffusion index and data evaluation value are confirmed as follows: for a single Class II tumor data set, the CA15-3 concentration value at the time of the first visit of the Class II tumor data set is recorded as k, and the Class II tumor data set with the CA15-3 concentration value k at the time of the first visit selected from the historical records is recorded as the analysis data. The mediated diffusion index = 1 / (the average value of the diagnosis and treatment effectiveness corresponding to each analysis data set), and the data evaluation value = the comparison reference value corresponding to the Class II tumor data set + the sub-fuzzy coefficient corresponding to the Class II tumor data set.

[0108] When selecting tumor data based on the associated ROI feature value and the mediated diffusion index or selecting tumor data based on the data evaluation value, the first second-class tumor data in the second reference sequence is used as the starting point, and the second-class tumor data in the second reference sequence are sequentially selected as the selected training data according to a second preset interval. The second preset interval is the number of second-class tumor data between two adjacent second-class tumor data in the selected second reference sequence, and the second preset interval is positively correlated with the total amount of second-class tumor data in the data source;

[0109] The second reference sequence is confirmed by:

[0110] When selecting tumor data based on the associated ROI characteristic value and the mediated diffusion index, the second reference sequence is a sequence obtained by sorting the second-class tumor data in a single data source in descending order according to the characteristic index, where characteristic index = associated ROI characteristic value + mediated diffusion index;

[0111] When tumor data is selected according to the data evaluation value, the second reference sequence is a sequence obtained by sorting the second type of tumor data in a single data source in descending order according to the data evaluation value.

[0112] Specifically, the method for confirming the associated ROI feature value includes:

[0113] Determine a preliminary characterization interval according to the image characterization value, and increase and adjust the area of the preliminary characterization interval according to the specific index of the preliminary characterization interval, and record the increased and adjusted preliminary characterization interval as the characterization interval;

[0114] The associated ROI feature value is determined according to the ratio of the mutation coefficient corresponding to the characterization interval to the area of the characterization interval.

[0115] Wherein, the image representation value = sub-fuzzy coefficient / preset sub-fuzzy coefficient + ROI reference value / preset ROI reference value; the preset sub-fuzzy coefficient and the preset ROI reference value are confirmed by, for a single second-class tumor data, recording the average of the sub-fuzzy coefficients corresponding to each second-class tumor data in the data source where the second-class tumor data is located as the preset sub-fuzzy coefficient, and recording the average of the ROI reference values corresponding to each second-class tumor data in the data source where the second-class tumor data is located as the preset ROI reference value;

[0116] Determining a preliminary characterization interval based on the image characterization value includes: recording a human brain X-ray image corresponding to the first visit time of a single tumor data set as an analysis image; the preliminary characterization interval is a region whose area is larger than and similar to the ROI region in the analysis image; the area of the preliminary characterization interval is positively correlated with the image characterization value; and it should be noted that any point on the contour of the preliminary characterization interval has the same reference length.

[0117] For a single point on the contour of the preselected representation interval, the point is recorded as the target point. The reference length corresponding to the target point is the length of the line connecting the target point and the reference intersection point. The reference intersection point is the intersection of the reference line segment and the ROI area. The reference line segment is the line connecting the target point and the reference point. The reference point is the center of the circumscribed circle of the ROI area in the analysis image.

[0118] The increase in the area of a single primary representation interval is negatively correlated with the specific index of the primary representation interval;

[0119] For a single tumor data set, the specific index of the preliminary characterization interval is the standard deviation of the adjustment region coefficients corresponding to each Class II tumor data set in the data source where the tumor data is located; the adjustment region coefficient corresponding to a single Class II tumor data set is the average value of the pixel values corresponding to each pixel point in the preliminary characterization interval corresponding to the Class II tumor data set;

[0120] For a single tumor data, the associated ROI feature value = the mutation coefficient corresponding to the characterization interval of the tumor data / the area of the characterization interval of the tumor data, and the mutation coefficient corresponding to the characterization interval = the average value of the pixel values corresponding to each pixel point in the characterization interval / the standard deviation of the pixel values corresponding to each pixel point in the characterization interval.

[0121] Specifically, the supplementary selection unit determines a representative data source based on a data source evaluation index, determines a representative combination based on ROI similarity and effective similarity, and determines a supplementary selection method for each representative combination based on a combination comparison coefficient;

[0122] For a single representative combination,

[0123] If the combined comparison coefficient is less than the preset combined comparison coefficient, it is determined that the supplementary selection method is to select tumor data based on the characteristic keywords;

[0124] If the combined comparison coefficient is greater than or equal to the preset combined comparison coefficient, it is determined that the supplementary selection method is to select tumor data according to the dynamic difference coefficient.

[0125] Determining the representative data source based on the data source evaluation index includes: selecting the data source with the largest data evaluation index as the representative data source; data source evaluation index = sub-ROI characteristic coefficient corresponding to a single data source + sub-heterogeneous cross abundance corresponding to a single data source;

[0126] Determining a representative combination based on ROI similarity and effective similarity includes: performing association analysis on each tumor data in the representative data source, when performing association analysis on a single tumor data, recording the tumor data as target data, recording other tumor data in the representative data source other than the target data as reference data, recording a set of reference data and target data having an ROI similarity with the target data greater than a preset ROI similarity and an effective similarity greater than a preset effective similarity as a representative combination, and continuing to perform association analysis on tumor data not recorded in the representative combination until all tumor data are recorded in the representative combination;

[0127] The confirmation method of ROI similarity and effective similarity is as follows: for any two tumor data, the larger value of the ROI reference values corresponding to the two tumor data is recorded as r1, and the smaller value is recorded as r2. ROI similarity = 1-(r1-r2) / r1; the larger value of the diagnosis and treatment effectiveness corresponding to the two tumor data is recorded as e1, and the smaller value is recorded as e2. Effective similarity = 1-(e1-e2) / e1;

[0128] The values of the preset ROI similarity and the preset effective similarity can be determined by the user according to the actual application scenario. The greater the user's similarity requirement for the tumor data in a single representative combination, the larger the values of the preset ROI similarity and the preset effective similarity. The values of the preset ROI similarity and the preset effective similarity are provided as follows: the preset ROI similarity is 75%, and the preset effective similarity is 70%;

[0129] The combination comparison coefficient is determined by recording a single representative combination as the target representative combination, recording the treatment regimen corresponding to each tumor data in the target representative combination as the first record, and recording the treatment regimen corresponding to each tumor data corresponding to each representative combination other than the target representative combination as the second record. The combination comparison coefficient = (the number of keywords that appear in both the first record and the second record) / (the number of different keywords in the first record);

[0130] Recording tumor data in data sources other than the representative data source as to-be-selected data, and recording to-be-selected data whose similarity threshold with the target representative combination is greater than a preset similarity threshold as matching data;

[0131] The similarity threshold between a single data to be selected and the target representative combination = 1 / |ROI reference value corresponding to the data to be selected - ROI mean| + 1 / |diagnosis and treatment effectiveness corresponding to the data to be selected - diagnosis and treatment effectiveness mean|. The average of the ROI reference values corresponding to each tumor data in the target representative combination is recorded as the ROI mean, and the average of the diagnosis and treatment effectiveness corresponding to each tumor data in the target representative combination is recorded as the diagnosis and treatment effectiveness mean.

[0132] The value of the preset combination comparison coefficient can be determined by the user according to the actual application scenario. The smaller the value of the preset combination comparison coefficient, the greater the user's need to select tumor data based on the dynamic difference coefficient. A preset combination comparison coefficient value is provided, and the preset combination comparison coefficient is 50%;

[0133] Selecting tumor data based on characteristic keywords includes: for a single representative combination, recording matching data containing characteristic keywords in the treatment plan as a class of selected data, and selecting a class of selected data in descending order of the number of characteristic keywords until the number of selected class of selected data reaches a preset number corresponding to the representative combination; the number of characteristic keywords is the total number of characteristic keywords contained in each treatment plan corresponding to the single class of selected data;

[0134] Selecting tumor data according to the dynamic difference coefficient includes: recording matching data whose dynamic difference coefficient with the target representative combination is greater than a preset dynamic difference coefficient as second-class selected data, and selecting the second-class selected data in descending order of the dynamic difference coefficient until the number of the selected second-class selected data reaches a preset number corresponding to the representative combination;

[0135] The preset number corresponding to a single representative combination = the pre-selected number corresponding to the representative combination - the number of tumor data in the representative combination. The pre-selected number corresponding to a single representative combination is positively correlated with the corresponding proportion coefficient of the representative combination. The proportion coefficient corresponding to a single representative combination = the number of tumor data in the representative combination / the total amount of tumor data in the representative data source;

[0136] The dynamic difference coefficient between a single matching data set and the target representative combination is the average of the fluctuation differences between the matching data set and each tumor data set in the target representative combination. The fluctuation difference between a single matching data set and a single tumor data set in the target representative combination = |the diagnosis and treatment fluctuation value corresponding to the matching data set - the diagnosis and treatment fluctuation value corresponding to the single tumor data set in the target representative combination|.

[0137] The values of the preset similarity threshold and the preset dynamic difference coefficient can be determined by the user according to the actual application scenario. The greater the user's demand for improving the correlation degree of data selection, the larger the value of the preset similarity threshold and the smaller the value of the preset dynamic difference coefficient. A preset similarity threshold and a preset dynamic difference coefficient are provided. The preset similarity threshold is 85%. The historical records of tumor data selected according to the dynamic difference coefficient are detected, and the average value of the dynamic difference coefficients corresponding to each second category of selected data in the historical records that can meet the user's needs is recorded as the preset dynamic difference coefficient.

[0138] See also Figure 4 , which is a schematic diagram of a method for training a human brain tumor prediction model according to the present invention. The present invention also provides a method for training a human brain tumor prediction model, comprising:

[0139] Obtain tumor data from multiple data sources, determine the data source status based on heterogeneous cross-abundance and ROI characteristic coefficients, and determine a data selection method based on the data source status, which can be extended selection of multiple data sources or supplementary selection of representative data sources;

[0140] In the extended selection of multiple data sources, the processing method of each data source is determined according to the tumor data category. The processing method includes determining a first screening method based on the training difficulty coefficient and determining a second screening method based on the judgment condition;

[0141] The first screening method is to select tumor data based on angiographic characteristic values and rhythm risk, or to select tumor data based on tumor reference values; the second screening method is to select tumor data based on associated ROI characteristic values and mediated diffusion index, or to select tumor data based on data evaluation values;

[0142] In the supplementary selection of representative data sources, the representative data source is determined based on the data source evaluation index, the representative combination is determined based on the ROI similarity and effective similarity, and the supplementary selection method for each representative combination is determined based on the combination comparison coefficient. The supplementary selection method is to select tumor data based on characteristic keywords or dynamic difference coefficients;

[0143] The selected tumor data is used as training data to train the tumor prediction model.

[0144] Thus far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art may make equivalent changes or substitutions to the relevant technical features, and the technical solutions after such changes or substitutions will fall within the scope of protection of the present invention.

[0145] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that the present invention is susceptible to various modifications and variations. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A human brain tumor prediction model training system, characterized in that: include: a data acquisition unit, configured to acquire tumor data from a plurality of data sources; A selection and analysis unit is connected to the data acquisition unit and is used to determine the data source status according to the heterogeneous cross abundance and the ROI characteristic coefficient, and determine the data selection method according to the data source status, wherein the data selection method is multi-data source extended selection or representative data source supplementary selection; an extended selection unit, connected to the data acquisition unit and the selection and analysis unit, respectively, for determining, in the extended selection of multiple data sources, a processing method for each data source according to the tumor data category, the processing method including determining a first screening method according to a training difficulty coefficient and determining a second screening method according to a determination condition; The first screening method is to select tumor data based on angiographic characteristic values and rhythm risk, or to select tumor data based on tumor reference values; the second screening method is to select tumor data based on associated ROI characteristic values and mediated diffusion index, or to select tumor data based on data evaluation values; a supplementary selection unit, connected to the data acquisition unit and the selection and analysis unit, respectively, for determining, in the supplementary selection of representative data sources, a representative data source based on a data source evaluation index, determining a representative combination based on ROI similarity and effective similarity, and determining a supplementary selection method for each representative combination based on a combination comparison coefficient, wherein the supplementary selection method is to select tumor data based on characteristic keywords or dynamic difference coefficients; A model training unit is connected to the extended selection unit and the supplementary selection unit respectively, and is used to train the tumor prediction model using the selected tumor data as training data.

2. The human brain tumor prediction model training system according to claim 1, characterized in that: The selection and analysis unit responds to the data source status to determine the data selection method, including: If the data source status of the selected analysis unit response is that the heterogeneous cross abundance is less than the preset heterogeneous cross abundance or the ROI characteristic coefficient is less than the preset ROI characteristic coefficient, then the data selection method is determined to be multi-data source extended selection; If the data source status of the selected analysis unit response is that the heterogeneous cross abundance is greater than or equal to the preset heterogeneous cross abundance and the ROI characteristic coefficient is greater than or equal to the preset ROI characteristic coefficient, the data selection method is determined to be representative data source supplementary selection.

3. The human brain tumor prediction model training system according to claim 2, characterized in that: The extended selection unit determines the tumor data category according to the diagnosis and treatment effectiveness and the diagnosis and treatment fluctuation value. The tumor data category includes: A type of tumor data with a diagnosis and treatment effectiveness greater than or equal to the preset diagnosis and treatment effectiveness and a diagnosis and treatment fluctuation value greater than or equal to the preset diagnosis and treatment fluctuation value; Category II tumor data whose diagnosis and treatment effectiveness is less than the preset diagnosis and treatment effectiveness or whose diagnosis and treatment fluctuation value is less than the preset diagnosis and treatment fluctuation value.

4. The human brain tumor prediction model training system according to claim 3, characterized in that: The extended selection unit determines a processing method for each data source according to the tumor data category, including: For each type of tumor data in a single data source, the processing method is to determine the first screening method according to the training difficulty coefficient; For each type II tumor data in a single data source, the processing method is to determine the second screening method according to the judgment conditions.

5. The human brain tumor prediction model training system according to claim 4, characterized in that: The extended selection unit determines a first screening method according to the training difficulty coefficient, including: If the training difficulty coefficient is greater than or equal to the preset training difficulty coefficient, it is determined that the first screening method is to select tumor data based on the angiographic characteristic value and the rhythm risk; If the training difficulty coefficient is less than the preset training difficulty coefficient, it is determined that the first screening method is to select tumor data according to the tumor reference value.

6. The human brain tumor prediction model training system according to claim 5, characterized in that: The method for confirming the training difficulty coefficient includes: If the heterogeneous correlation is greater than or equal to the preset heterogeneous correlation, the training difficulty coefficient is determined according to the heterogeneous interaction difficulty; If the heterogeneous correlation is less than the preset heterogeneous correlation, the training difficulty coefficient is determined according to the homogeneous correlation coefficient.

7. The human brain tumor prediction model training system according to claim 4, characterized in that: The extended selection unit determines the second screening method in response to the judgment condition, including: The determination condition of the extended selection unit response is that the edge fuzzy coefficient is greater than or equal to the preset edge fuzzy coefficient and the ROI comparison coefficient is less than the preset ROI comparison coefficient, then the second screening method is determined to select tumor data based on the associated ROI characteristic value and the mediated diffusion index; The determination condition of the extended selection unit response is that the edge fuzzy coefficient is less than the preset edge fuzzy coefficient or the ROI comparison coefficient is greater than or equal to the preset ROI comparison coefficient, then the second screening method is determined to select tumor data according to the data evaluation value.

8. The human brain tumor prediction model training system according to claim 7, characterized in that: The method for confirming the associated ROI feature value includes: Determine a preliminary characterization interval according to the image characterization value, and increase and adjust the area of the preliminary characterization interval according to the specific index of the preliminary characterization interval, and record the increased and adjusted preliminary characterization interval as the characterization interval; The associated ROI feature value is determined according to the ratio of the mutation coefficient corresponding to the characterization interval to the area of the characterization interval.

9. The human brain tumor prediction model training system according to claim 2, characterized in that: The supplementary selection unit determines a representative data source based on the data source evaluation index, determines a representative combination based on the ROI similarity and the effective similarity, and determines a supplementary selection method for each representative combination based on the combination comparison coefficient; For a single representative combination, If the combined comparison coefficient is less than the preset combined comparison coefficient, it is determined that the supplementary selection method is to select tumor data based on the characteristic keywords; If the combined comparison coefficient is greater than or equal to the preset combined comparison coefficient, it is determined that the supplementary selection method is to select tumor data according to the dynamic difference coefficient.

10. A training method using the human brain tumor prediction model training system according to any one of claims 1 to 9, characterized in that: include: Obtain tumor data from multiple data sources, determine the data source status based on heterogeneous cross-abundance and ROI characteristic coefficients, and determine a data selection method based on the data source status, which can be extended selection of multiple data sources or supplementary selection of representative data sources; In the extended selection of multiple data sources, the processing method of each data source is determined according to the tumor data category. The processing method includes determining a first screening method based on the training difficulty coefficient and determining a second screening method based on the judgment condition; The first screening method is to select tumor data based on angiographic characteristic values and rhythm risk, or to select tumor data based on tumor reference values; the second screening method is to select tumor data based on associated ROI characteristic values and mediated diffusion index, or to select tumor data based on data evaluation values; In the supplementary selection of representative data sources, the representative data source is determined based on the data source evaluation index, the representative combination is determined based on the ROI similarity and effective similarity, and the supplementary selection method for each representative combination is determined based on the combination comparison coefficient. The supplementary selection method is to select tumor data based on characteristic keywords or dynamic difference coefficients; The selected tumor data is used as training data to train the tumor prediction model.

Images