Early detection method for urinary system tumors based on neural network
By performing historical database retrieval, heterogeneity differentiation, and multi-scale feature analysis in urinary system tumor detection, and combining feature fusion with convolutional neural networks, the problem of insufficient sensitivity and accuracy in early detection of existing methods is solved, achieving more efficient early tumor detection.
Patent Information
- Application Number
- CN202510962955.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-14
- Publication Date
- 2025-11-07
AI Technical Summary
Existing methods for detecting urinary system tumors are insufficient in terms of sensitivity and accuracy in early detection. In particular, traditional imaging examinations have a high rate of missed diagnoses, tumor marker detection lacks specificity, and existing neural network methods are unable to effectively handle diverse and dynamic features, resulting in insufficient feature extraction and fusion.
By acquiring the target sample type, retrieving historical detection databases, differentiating different types, extracting features from feature sets, utilizing multi-scale feature analysis and cross-scale interactive fusion, and combining convolutional neural networks for feature extraction and fusion, comprehensive capture and accurate analysis of multi-scale features are achieved.
It improves the accuracy and reliability of early detection of urinary system tumors, can extract key features more accurately, comprehensively reflect signs of tumor development, and improve the accuracy and generalization ability of detection results.
Smart Images
Figure CN120913875A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of medical detection, in particular to an early detection method for urinary system tumors based on a neural network. BACKGROUND
[0002] As a serious threat to human health, urinary system tumors often have hidden early symptoms, so that patients are usually in the middle and late stages when diagnosed, which greatly affects the treatment effect and the survival rate of patients. Therefore, early detection of urinary system tumors is crucial to improve the prognosis of patients.
[0003] Traditional urinary system tumor detection methods have many limitations. For example, imaging examinations such as ultrasound and CT can provide information on the location and shape of tumors, but their sensitivity to early detection of small tumors is low, and misdiagnosis often occurs. While tumor marker detection has some convenience, its lack of specificity often leads to misdiagnosis, causing unnecessary psychological burden and financial pressure for further examination for patients.
[0004] With the continuous development of artificial intelligence technology, the application of neural networks in the field of medical detection has become a research hotspot. However, existing neural network-based urinary system tumor detection methods still face many challenges in practical application. On the one hand, historical detection data often has diversity and complexity, and different types of tumors may differ in detection data. How to effectively classify and process these data is the key to improving detection accuracy. On the other hand, the occurrence and development of tumors is a dynamic process, and detection data may exhibit different characteristics at different times. Existing methods are difficult to fully capture the relationship between these dynamic characteristics.
[0005] In addition, existing detection methods also have shortcomings in feature extraction and fusion. Traditional feature extraction methods can only extract single-scale features, and cannot fully utilize the complementarity between multi-scale features, resulting in limited accuracy of detection results. In the feature fusion process, how to achieve effective cross-scale interaction and fusion is also a problem to be solved.
[0006] To address the above problems, the present application proposes an early detection method for urinary system tumors based on a neural network, aiming to improve the accuracy and reliability of early detection of urinary system tumors through effective processing of historical detection data, multi-scale feature analysis, and cross-scale interaction and fusion. SUMMARY
[0007] The present application aims to provide an early detection method for urinary system tumors based on a neural network to solve the problems raised in the background.
[0008] To achieve the above object, the application provides a method for early detection of urinary system tumors based on a neural network, which comprises the following steps:
[0009] acquiring a target type of a target sample, searching in a historical detection database based on the target type, and determining a set of historical tumor detection records;
[0010] distinguishing the set of historical tumor detection records, and determining a set of distinguished historical detection records;
[0011] extracting features from the set of distinguished historical detection records, obtaining a set of feature central values and a set of feature correlation time periods;
[0012] using the set of feature correlation time periods as a set of receptive periods of a plurality of detection network layers, performing multi-scale feature analysis on a detection data sequence of the target sample in a preset monitoring window by using the plurality of detection network layers configured, and obtaining a set of detection features;
[0013] performing cross-scale interaction fusion on the set of detection features, obtaining a set of target interaction fusion detection features, and using the set of target interaction fusion detection features as an early detection result of the urinary system tumor.
[0014] Preferably, the step of extracting features from the set of distinguished historical detection records, and obtaining a set of feature central values and a set of feature correlation time periods comprises the following steps:
[0015] extracting detection data from the set of distinguished historical detection records, and obtaining a set of distinguished historical detection data sequences;
[0016] extracting instantaneous detection data from the set of distinguished historical detection data sequences respectively, and determining a set of historical instantaneous detection data, wherein each historical instantaneous detection data is detection data at a time point of maximum data change fluctuation in each distinguished historical detection data sequence;
[0017] extracting features from the set of historical instantaneous detection data, obtaining a set of historical instantaneous detection features, and performing centralized analysis on the set of historical instantaneous detection features, and determining a set of feature central values;
[0018] using the set of historical instantaneous detection features as an index, performing feature correlation time period diffusion identification on the set of distinguished historical detection data sequences, and determining a set of historical instantaneous detection feature correlation time periods, wherein each historical instantaneous detection feature correlation time period reflects a duration of data associated with a historical instantaneous detection feature at a tumor occurrence time in each distinguished historical detection data sequence;
[0019] Calculate mean values of the plurality of historical instantaneous detection feature correlation time periods respectively, to obtain the plurality of feature correlation time periods.
[0020] Preferably, instantaneous feature extraction is performed on the plurality of historical instantaneous detection data sets to obtain a plurality of historical instantaneous detection feature sets, and centralized analysis is performed on the plurality of historical instantaneous detection feature sets to determine a plurality of feature central values, including:
[0021] Instantaneous feature extraction is performed on the plurality of historical instantaneous detection data sets using a feature extraction network layer to obtain the plurality of historical instantaneous detection feature sets;
[0022] Feature mean value calculation is performed on the plurality of historical instantaneous detection feature sets to determine a plurality of historical instantaneous detection feature mean values;
[0023] The plurality of historical instantaneous detection feature mean values are iterated in the plurality of historical instantaneous detection feature sets according to a preset iteration time period to obtain a plurality of iteration historical instantaneous detection features;
[0024] When the aggregation amount of the plurality of iteration historical instantaneous detection features is less than or equal to the aggregation amount of the plurality of historical instantaneous detection feature mean values, the plurality of historical instantaneous detection feature mean values are taken as the plurality of feature central values.
[0025] Preferably, the method includes:
[0026] When the aggregation amount of the plurality of iteration historical instantaneous detection features is greater than the aggregation amount of the plurality of historical instantaneous detection feature mean values, it is determined whether the aggregation amount difference between the plurality of iteration historical instantaneous detection features and the plurality of historical instantaneous detection feature mean values is greater than or equal to a preset aggregation amount difference threshold value, if yes, iteration is continued based on the plurality of iteration historical instantaneous detection features until a maximum iteration number is met, and the plurality of iteration historical instantaneous detection features obtained in the last iteration are taken as the plurality of feature central values;
[0027] If not, iteration is stopped, and the plurality of iteration historical instantaneous detection features are taken as the plurality of feature central values.
[0028] Preferably, the plurality of historical instantaneous detection feature sets are used as indexes to perform feature correlation time period diffusion identification on the plurality of differentiated historical detection data sequence sets to determine a plurality of historical instantaneous detection feature correlation time period sets, including:
[0029] Detection feature extraction is performed on the plurality of differentiated historical detection data sequence sets using the feature extraction network layer to obtain a plurality of differentiated historical detection feature sequence sets;
[0030] randomly extracting one historical instantaneous detection feature from the plurality of historical instantaneous detection feature sets as a first historical instantaneous detection feature, and matching a corresponding first distinguished historical detection feature sequence from the plurality of distinguished historical detection feature sequence sets;
[0031] performing neighbor search on the first historical instantaneous detection feature in the first distinguished historical detection feature sequence according to a preset approximate correlation scale, to obtain a first historical instantaneous detection feature neighborhood;
[0032] statistically determining a duration of the first historical instantaneous detection feature neighborhood, and taking a statistical result as a first historical instantaneous detection feature correlation period;
[0033] performing feature correlation period diffusion recognition on the plurality of historical instantaneous detection feature sets in the corresponding plurality of distinguished historical detection data sequence sets according to a preset neighbor correlation scale, to determine a plurality of historical instantaneous detection feature correlation period sets.
[0034] Preferably, the plurality of detection feature sets are cross-scale interactively fused to obtain a target interactive fusion detection feature set, including:
[0035] randomly extracting a first detection feature set and a second detection feature set from the plurality of detection feature sets;
[0036] calculating the first detection feature set and the second detection feature set for similarity recognition to determine a first feature similarity set;
[0037] normalizing the first feature similarity set to obtain a first feature similarity normalized value set;
[0038] performing convolution calculation on the first feature similarity normalized value set and the second detection feature set to obtain a first interactive fusion detection feature set;
[0039] randomly extracting a third detection feature set from the plurality of detection feature sets, and cross-scale interactively fusing the third detection feature set with the first interactive fusion detection feature set to obtain a second interactive fusion detection feature set;
[0040] After multiple cross-scale interactive fusions, until the detection features in the plurality of detection feature sets are fused, the target interactive fusion detection feature set is obtained.
[0041] Preferably, the method comprises:
[0042] obtaining a plurality of sample feature similarity normalized value sets, a plurality of sample detection feature sets, and a plurality of sample interactive fusion detection feature sets as training data;
[0043] Supervise training of the network layer constructed based on the convolutional neural network by using the training data until the training converges, and obtain the trained convolutional network layer;
[0044] Perform convolutional calculation on the first feature similarity normalization value set and the second detection feature set by using the convolutional network layer, and obtain the first interaction fusion detection feature set.
[0045] Preferably, the set of historical tumor detection records is distinguished by different types, and a plurality of distinguished historical detection record sets are determined, including:
[0046] Randomly extract a plurality of historical tumor detection records from the set of historical tumor detection records;
[0047] Enumerate the plurality of historical tumor detection records two by two to obtain a plurality of enumeration combinations;
[0048] Determine whether there is an enumeration combination with a record similarity greater than a preset record similarity threshold in the plurality of enumeration combinations, and if not, the plurality of historical tumor detection records are taken as a plurality of distinguished targets;
[0049] Based on the plurality of distinguished targets, the set of historical tumor detection records is distinguished by different types according to the preset record similarity threshold, and the plurality of distinguished historical detection record sets are obtained, wherein each distinguished historical detection record set corresponds to a distinguished target.
[0050] Preferably, it is determined whether there is an enumeration combination with a record similarity greater than a preset record similarity threshold in the plurality of enumeration combinations, and if so, the enumeration combinations with a record similarity greater than a preset record similarity threshold in the plurality of enumeration combinations are merged into the same distinguished target;
[0051] Based on the merged distinguished target, the set of historical tumor detection records is distinguished by different types according to the preset record similarity threshold, and the plurality of distinguished historical detection record sets are obtained.
[0052] Preferably, the plurality of distinguished historical detection record sets are distinguished by different types for verification, including:
[0053] Randomly extract two distinguished historical detection record sets from the plurality of distinguished historical detection record sets;
[0054] Calculate the cross-set record similarity mean of the two distinguished historical detection record sets;
[0055] If the cross-set record similarity mean is less than or equal to a preset cross-set similarity threshold, the plurality of distinguished historical detection record sets are maintained;
[0056] If the cross-set record similarity mean is greater than a preset cross-set similarity threshold, then the heterogeneous classification of the historical tumor detection record set is re-performed until the cross-set similarity threshold requirement is met.
[0057] Compared with the prior art, the present application has the following advantages:
[0058] The early detection method of the urinary system tumor based on the neural network has many significant advantages. First, after obtaining the target type of the target sample, the historical tumor detection record set is determined by searching in the historical detection database based on the type, which can select the historical data similar to the target sample, and provide more valuable reference for subsequent detection. By classifying the historical tumor detection record set, a plurality of classified historical detection record sets are determined, so that the different types of historical data are effectively classified, and the interference between different types of data is avoided, so that the features of various data can be extracted more accurately.
[0059] When traversing the plurality of classified historical detection record sets for feature set extraction, the plurality of feature set values and feature association time periods are obtained, which can deeply mine the key features related to tumor occurrence and their time association in the historical data. The feature association time period is used as the receptive period of the detection network layer, and the configured detection network layer is used to perform multi-scale feature analysis on the detection data sequence of the target sample in the preset monitoring window, to obtain a plurality of detection feature sets. This multi-scale feature analysis method can capture features in the detection data from different angles and levels, and fully reflect the potential signs of tumor occurrence.
[0060] The plurality of detection feature sets are cross-scale interactively fused to obtain a target interactive fusion detection feature set as the early detection result of the urinary system tumor. Through this cross-scale interactive fusion, the complementarity between different scale features is fully utilized, so that the detection result is more accurate and reliable. In the feature extraction process, the feature extraction network layer is used to extract instantaneous features from the historical instantaneous detection data set, and the feature set value is determined by iterative calculation, which can accurately extract the key features in the historical data and improve the accuracy of feature extraction.
[0061] The set of classified historical detection data sequence sets is indexed by the set of historical instantaneous detection feature sets, and the feature association time period is diffused and identified to determine the set of historical instantaneous detection feature association time periods, and the mean value is calculated to obtain the feature association time period, so that the time association between features can be accurately grasped, and important time dimension reference is provided for subsequent multi-scale feature analysis. When the detection feature set is cross-scale interactively fused, the feature similarity is calculated, normalized and convoluted, and the effective fusion between different feature sets is realized, which further improves the accuracy of the detection result.
[0062] The network layer constructed based on the convolutional neural network is supervised trained by using the training data until convergence, and a trained convolutional network layer is obtained, so that the network can better adapt to the detection task and improve the detection accuracy and generalization ability. When the heterogeneous division of the historical tumor detection record set is performed, the historical data can be accurately divided into different sets through the steps of random extraction of records, enumeration combination and judgment of similarity, thereby laying a good foundation for subsequent feature extraction and analysis.
[0063] The accuracy and reliability of the division result are ensured by verifying the heterogeneous division of the divided historical detection record set, and the detection error caused by improper division is avoided. In summary, the method of the present application can effectively improve the accuracy and reliability of early detection of urinary system tumors through a series of scientific and reasonable steps and processing methods, and provides strong support for clinical diagnosis, and has important practical application value. BRIEF DESCRIPTION OF DRAWINGS
[0064] Figure 1 The working principle diagram of the early detection method of urinary system tumors based on neural network is provided.
[0065] Figure 2 The flowchart for determining the feature extraction and correlation period is provided.
[0066] Figure 3 The flowchart for determining the instantaneous feature extraction and central value is provided.
[0067] Figure 4 The flowchart for feature correlation period diffusion recognition is provided.
[0068] Figure 5 The flowchart for heterogeneous division is provided. DETAILED DESCRIPTION
[0069] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.
[0070] Please refer to Figures 1-5 The present application provides an early detection method of urinary system tumors based on neural network, and the specific implementation steps are as follows:
[0071] obtaining a target type of the target sample, performing retrieval in the historical detection database based on the target type, and determining a set of historical tumor detection records. The target type of the target sample can be determined through relevant information during sample collection, such as the source site of the sample, the collection method, and the like. The historical detection database stores a large number of past tumor detection records, which include detection data and corresponding detection results of various types of samples.
[0072] performing heterogeneous classification on the set of historical tumor detection records to determine a plurality of classified sets of historical detection records. The process of heterogeneous classification is to divide the records according to their similarity, and records with high similarity are classified into the same set, and records in different sets have low similarity.
[0073] performing feature set extraction on the plurality of classified sets of historical detection records to obtain a plurality of feature set values and a plurality of feature association time periods. In the extraction process, the detection data in each classified set of historical detection records needs to be analyzed to find the key features.
[0074] using the plurality of feature association time periods as a plurality of receptive periods of a plurality of detection network layers, performing multi-scale feature analysis on a detection data sequence of the target sample within a preset monitoring window by using the plurality of detection network layers configured to obtain a plurality of detection feature sets. The detection network layer is based on a neural network, and different detection network layers have different receptive periods and can analyze detection data from different time scales.
[0075] performing cross-scale interactive fusion on the plurality of detection feature sets to obtain a target interactive fusion detection feature set, and using the target interactive fusion detection feature set as an early detection result of the urinary system tumor. Cross-scale interactive fusion is to fuse detection features at different scales to obtain a more comprehensive and accurate detection result.
[0076] Embodiment 1:
[0077] In the operation of heterogeneous classification on the set of historical tumor detection records, a certain number of historical tumor detection records are randomly extracted from the set of historical tumor detection records. The random extraction is performed in a random number generation manner to ensure that each historical tumor detection record has the same extraction probability. After extraction, the extracted historical tumor detection records are enumerated two by two to form a plurality of enumeration combinations. For example, if n historical tumor detection records are extracted, the number of enumeration combinations is n(n-1) / 2.
[0078] It is necessary to determine whether the record similarity of these enumerated combinations is greater than the preset record similarity threshold. The calculation of record similarity uses a specific algorithm that compares the data in each record, including but not limited to detection index data, detection time, sample type, etc., and calculates a similarity value through weighted calculation. The preset record similarity threshold is a standard value set in advance based on a large amount of historical data and actual detection experience.
[0079] If it is determined that the record similarity of all enumerated combinations is not greater than the preset record similarity threshold, then these extracted historical tumor detection records are used as multiple distinguishing targets. Each distinguishing target represents a different category, and subsequent heterogeneous classification of the entire historical tumor detection record set will be based on these distinguishing targets. Specifically, according to the preset record similarity threshold, the similarity of each record in the historical tumor detection record set with each distinguishing target is calculated, and the record is classified into the set corresponding to the distinguishing target with the highest similarity and greater than or equal to the preset record similarity threshold. In this way, multiple distinguished historical detection record sets are obtained, each corresponding to a distinguishing target.
[0080] If it is found during the judgment process that there are enumerated combinations with record similarity greater than the preset record similarity threshold, these enumerated combinations with record similarity greater than the preset record similarity threshold are merged into the same distinguishing target. The merging method is to integrate the historical tumor detection records in these enumerated combinations to form a new distinguishing target. Then, based on the merged distinguishing target, the historical tumor detection record set is classified according to the preset record similarity threshold, and each record is classified into the corresponding distinguishing target set, thereby obtaining multiple distinguished historical detection record sets.
[0081] After completing the heterogeneous classification, the multiple distinguished historical detection record sets need to be verified for heterogeneous classification. The specific steps are to randomly extract two distinguished historical detection record sets from the multiple distinguished historical detection record sets. Here, the random extraction is also determined by the random number generation method to extract the set. After extraction, the cross-set record similarity average of the two distinguished historical detection record sets is calculated. When calculating the cross-set record similarity average, the similarity of each record in the first set with each record in the second set needs to be calculated, and then the average of all similarity values is taken to obtain the cross-set record similarity average.
[0082] The calculated cross-set record similarity average is compared with a preset cross-set similarity threshold. If the cross-set record similarity average is less than or equal to the preset cross-set similarity threshold, it indicates that the difference between the two distinguished historical detection record sets is large, and the result of the heterogeneous distinction is reliable, so the multiple distinguished historical detection record sets remain unchanged. If the cross-set record similarity average is greater than the preset cross-set similarity threshold, it indicates that the record similarity between the two sets is high, and the heterogeneous distinction may not be accurate enough. At this time, the historical tumor detection record set needs to be re-divided into heterogeneous sets. Repeat all the steps from extracting historical tumor detection records to calculating the cross-set record similarity average until the cross-set similarity threshold requirement is met.
[0083] Throughout the implementation process, the operation of each step strictly follows the established rules and algorithms, ensuring accurate and reasonable heterogeneous division of the historical tumor detection record set, and providing a reliable basis for subsequent feature extraction and detection analysis. For example, when randomly extracting historical tumor detection records, the randomness of extraction is ensured, avoiding human interference; when calculating the record similarity, multiple dimensions of data are considered to ensure that the similarity value accurately reflects the similarity between records; when performing heterogeneous division verification, the operation is continuously adjusted and repeated to ensure that the multiple distinguished historical detection record sets obtained finally have sufficient differences and meet the requirements of the detection method.
[0084] Embodiment 2:
[0085] When traversing the multiple distinguished historical detection record sets for feature set extraction, detection data extraction operations are performed on each distinguished historical detection record set. For each distinguished historical detection record set, all detection records contained therein are read in sequence, and the detection data in these detection records are arranged in chronological order to obtain multiple distinguished historical detection data sequence sets. Each distinguished historical detection data sequence set corresponds to a distinguished historical detection record set, and the data sequences therein reflect the change of detection data over time in the set.
[0086] The instantaneous detection data needs to be extracted from each of the differentiated historical detection data sequence sets to determine a plurality of historical instantaneous detection data sets. For each differentiated historical detection data sequence, the data points therein need to be analyzed one by one to find the time at which the maximum data fluctuation occurs. The maximum data fluctuation refers to the maximum magnitude of the change in the value of the detection data before and after the time, among the entire data sequence. After finding the time, the detection data corresponding to the time is extracted as the historical instantaneous detection data. Each differentiated historical detection data sequence corresponds to a historical instantaneous detection data, and the historical instantaneous detection data of all differentiated historical detection data sequences is collected to form a historical instantaneous detection data set. By analogy, the above operation is performed on a plurality of differentiated historical detection data sequence sets to ultimately obtain a plurality of historical instantaneous detection data sets.
[0087] After obtaining a plurality of historical instantaneous detection data sets, the feature extraction network layer is used to extract the instantaneous features of these sets to obtain a plurality of historical instantaneous detection feature sets. The feature extraction network layer is constructed based on a neural network, which can automatically extract the information that best reflects the features of the input detection data. For each historical instantaneous detection data set, it is input into the feature extraction network layer, and the network layer outputs the corresponding historical instantaneous detection feature set after a series of calculations and processing. Each historical instantaneous detection feature set contains the key feature information of the historical instantaneous detection data set.
[0088] The plurality of historical instantaneous detection feature sets need to be analyzed to determine a plurality of feature central values. First, the plurality of historical instantaneous detection feature sets are traversed, and the mean value of the feature values in each set is calculated to obtain a plurality of historical instantaneous detection feature means. For example, for a historical instantaneous detection feature set, which contains a plurality of feature values, the feature values are added and divided by the number of feature values to obtain the historical instantaneous detection feature mean of the set.
[0089] According to the preset iteration period, the plurality of historical instantaneous detection feature means are iterated in the plurality of historical instantaneous detection feature sets to obtain a plurality of iteration historical instantaneous detection features. The preset iteration period is a time interval or iteration number set in advance. During the iteration process, the historical instantaneous detection feature mean is used as the basis to find iteration historical instantaneous detection features that better represent the features in the historical instantaneous detection feature set. Each iteration adjusts and updates the current iteration historical instantaneous detection feature according to certain rules.
[0090] In the iteration process, it is necessary to constantly judge the relationship between the aggregation of the plurality of iteration historical instantaneous detection features and the aggregation of the plurality of historical instantaneous detection feature mean values. The aggregation can be measured by calculating the distance or variance between the feature values, which reflects the concentration degree of the feature values. When the aggregation of the plurality of iteration historical instantaneous detection features is less than or equal to the aggregation of the plurality of historical instantaneous detection feature mean values, it means that the iteration historical instantaneous detection features at this time can better represent the concentration trend of the features, and therefore the plurality of historical instantaneous detection feature mean values is used as the plurality of feature central values.
[0091] If in the iteration process, the aggregation of the plurality of iteration historical instantaneous detection features is greater than the aggregation of the plurality of historical instantaneous detection feature mean values, it is necessary to further judge whether the aggregation difference between the plurality of iteration historical instantaneous detection features and the plurality of historical instantaneous detection feature mean values is greater than or equal to the preset aggregation difference threshold value. The preset aggregation difference threshold value is a standard value set in advance according to actual needs and experience.
[0092] If the aggregation difference is greater than or equal to the preset aggregation difference threshold value, it means that the difference between the current iteration historical instantaneous detection features and the historical instantaneous detection feature mean values is large, and further iteration is needed to find more suitable feature central values. At this time, the iteration is continued based on the plurality of iteration historical instantaneous detection features, and the iteration historical instantaneous detection features are constantly updated according to the established iteration rules until the maximum iteration number is reached. When the maximum iteration number is reached, the plurality of iteration historical instantaneous detection features obtained by the last iteration is used as the plurality of feature central values.
[0093] If the aggregation difference is less than the preset aggregation difference threshold value, it means that although the aggregation of the iteration historical instantaneous detection features is greater than the aggregation of the historical instantaneous detection feature mean values, the difference between them is small enough, and further iteration may not significantly improve the accuracy of the feature central values, so the iteration can be stopped, and the current plurality of iteration historical instantaneous detection features is used as the plurality of feature central values.
[0094] Embodiment 3:
[0095] In the feature correlation period diffusion identification of the plurality of distinguished historical detection data sequence sets indexed by the plurality of historical instantaneous detection feature sets, the detection feature extraction network layer is used to perform the detection feature extraction operation on the plurality of distinguished historical detection data sequence sets. Specifically, for each distinguished historical detection data sequence set, the detection data sequence arranged in time sequence in the set is input into the feature extraction network layer, and the network layer extracts the detection features capable of representing the data features from the data sequence through the calculation of multiple layers of neurons, such as convolution, activation and other operations, to form a plurality of distinguished historical detection feature sequence sets. Each distinguished historical detection feature sequence set corresponds to a distinguished historical detection data sequence set, and the feature sequence in the set retains the time sequence information of the original data sequence.
[0096] After the detection feature extraction is completed, one historical instantaneous detection feature is randomly extracted from the plurality of historical instantaneous detection feature sets as a first historical instantaneous detection feature. The random extraction can be realized by generating a random index, so as to ensure that each historical instantaneous detection feature is extracted with equal probability. After the extraction is completed, the corresponding first distinguished historical detection feature sequence needs to be matched from the plurality of distinguished historical detection feature sequence sets. The matching basis is the corresponding relationship between the historical instantaneous detection feature set and the distinguished historical detection feature sequence set, that is, each historical instantaneous detection feature set corresponds to a distinguished historical detection data sequence set, and further corresponds to a distinguished historical detection feature sequence set.
[0097] According to the preset approximate correlation scale, the first historical instantaneous detection feature is subjected to the neighbor search in the first distinguished historical detection feature sequence to obtain the first historical instantaneous detection feature neighborhood. The preset approximate correlation scale is a preset standard for measuring the feature similarity, which can be represented as a distance threshold d th . The process of the neighbor search can be described as follows: for each feature point f i in the first distinguished historical detection feature sequence, the distance d(f first ,f i ) between the feature point and the first historical instantaneous detection feature f first is calculated, and if d(f i ,f first )≤d th , the feature point is included in the neighborhood range. The distance calculation can use the Euclidean distance formula, that is, for two feature vectors and , the Euclidean distance is:
[0098]
[0099] wherein, denotes the feature vector of the feature point f i in the first distinguished historical detection feature sequence, The first historical instantaneous detection feature f represents first The feature vector, where n is the dimension of the feature vector, a k and b k These are the values of the k-th dimension of the eigenvector.
[0100] After obtaining the neighborhood of the first historical instantaneous detection feature, it is necessary to calculate the duration of this neighborhood and use the statistical result as the associated time period of the first historical instantaneous detection feature. The duration of the neighborhood refers to the span of the feature points within the neighborhood on the time axis, which can be specifically determined by the timestamp t of the earliest feature point in the neighborhood. start and the timestamp t of the latest feature point end The calculation shows that the duration is t. end -t start .
[0101] After determining the associated time period for a single historical instantaneous detection feature, it is necessary to perform feature association time period diffusion identification on multiple historical instantaneous detection feature sets within corresponding sets of multiple distinguishable historical detection data sequences, according to a preset nearest neighbor association scale. Specifically, for each feature in each historical instantaneous detection feature set, the steps from feature extraction to statistical neighborhood duration are repeated to determine the corresponding associated time period for each historical instantaneous detection feature, ultimately forming multiple sets of associated time periods for historical instantaneous detection features. Each associated time period for a historical instantaneous detection feature reflects the duration of data associated with a historical instantaneous detection feature at the time of tumor occurrence within a distinguishable historical detection data sequence.
[0102] After obtaining multiple sets of historical instantaneous detection feature association time periods, it is necessary to calculate the mean of each set to obtain multiple feature association time periods. For a certain set of historical instantaneous detection feature association time periods S = {t1, t2, ..., t...} m The mean of} is calculated as follows:
[0103]
[0104] Where m is the number of associated time periods in the set, t i For the i-th associated time period in the set, This represents the mean of the set, i.e., the time period associated with the features.
[0105] Throughout the implementation process, each step must strictly adhere to the established rules and methods. For example, the structure of the feature extraction network layer needs to be designed based on the characteristics of the detection data to ensure that the extracted features accurately reflect the essence of the data; a preset approximate correlation scale d... th The settings need to be combined with historical data and actual detection requirements. If d th If d is too small, the neighborhood may be too small, making it impossible to capture enough related data; if dth If too large, irrelevant feature points may be included in the neighborhood, affecting the accuracy of the correlation period. The timestamp needs to be recorded accurately to a sufficient time unit to ensure the accuracy of the duration calculation. In addition, when performing diffusion recognition on multiple historical instantaneous feature sets, it is necessary to ensure that the correspondence between each feature set and the corresponding set of differentiated historical detection data sequences is correct, so as to avoid deviations in the determination of feature correlation periods due to correspondence errors.
[0106] Embodiment 4:
[0107] When performing cross-scale interactive fusion on multiple detection feature sets, first, a first detection feature set and a second detection feature set are randomly extracted from the multiple detection feature sets. For example, assuming that the multiple detection feature sets are F1, F2, F3, and F4, F1 is determined as the first detection feature set and F2 is determined as the second detection feature set through random number generation. The random extraction here ensures that each detection feature set has the same probability of being extracted, avoiding the bias caused by human selection.
[0108] After extraction, the similarity between the first detection feature set and the second detection feature set needs to be calculated to determine the first feature similarity set. When calculating the similarity, the features in the two sets are compared one by one. For example, the first detection feature set contains features {f 11 ,f 12 ,f 13}, and the second detection feature set contains features {f 21 ,f 22 ,f 23}, the similarity between all feature pairs f 11 and f 21 , f 11 and f 22 , f 11 and f 23 , f 12 and f 21 , etc. is calculated, and these similarity values are collected to form the first feature similarity set. The similarity calculation can use the cosine similarity algorithm, which measures the similarity of two feature vectors by calculating the cosine of the angle between them.
[0109] After obtaining the first feature similarity set, it needs to be normalized to obtain the first feature similarity normalized value set. The purpose of normalization is to map the similarity values to a unified range, facilitating subsequent calculations. For example, assuming that the value range in the first feature similarity set is between 0.2 and 0.8, it is converted to a value between 0 and 1 through the normalization formula. The specific normalization method can be linear normalization, that is, for each similarity value s, the normalized value is (s-s min) / (s max -s min ), wherein s min is the minimum value in the set of similarity values, s max is the maximum value.
[0110] Next, the first interaction fusion detection feature set is obtained by performing convolution calculation on the first feature similarity normalization value set and the second detection feature set using the trained convolution network layer. To train the convolution network layer, training data is needed, which includes multiple sample feature similarity normalization value sets, multiple sample detection feature sets, and multiple sample interaction fusion detection feature sets. These sample data come from historical detection cases, for example, 1000 samples are selected from past urinary system tumor detection data, each sample contains a corresponding feature similarity normalization value set, a detection feature set, and an interaction fusion detection feature set labeled by experts.
[0111] The network layer based on the convolutional neural network is supervised trained using these training data. During the training process, the sample feature similarity normalization value set and the sample detection feature set are input, the network layer outputs the predicted interaction fusion detection feature set, then the prediction result is compared with the sample interaction fusion detection feature set, the loss function is calculated, and the parameters of the network layer are adjusted through the back propagation algorithm to reduce the loss. This process is repeated until the training converges, that is, the loss function no longer decreases significantly, at this time the trained convolution network layer is obtained.
[0112] Suppose the trained convolution network layer has specific convolution kernel parameters, when the first feature similarity normalization value set and the second detection feature set are input into the network layer, the network layer will extract the key features and fuse them through convolution operation of the convolution kernel and the input data, and output the first interaction fusion detection feature set. For example, the input first feature similarity normalization value set can be represented as a matrix S, the second detection feature set is represented as a matrix F, the convolution kernel K in the convolution network layer is convolved with the matrix S and F to obtain the fused feature matrix C, which is the first interaction fusion detection feature set.
[0113] After obtaining the first interaction fusion detection feature set, a third detection feature set is randomly selected from the multiple detection feature sets, assuming that F3 is selected, and it is cross-scale interaction fused with the first interaction fusion detection feature set to obtain the second interaction fusion detection feature set. The fusion process is similar to the previous one, first calculate the similarity between the third detection feature set and the first interaction fusion detection feature set to obtain a new feature similarity set, then normalize it, and then use the trained convolution network layer to perform convolution calculation to obtain the fused second interaction fusion detection feature set.
[0114] In this way, through multiple cross-scale interactive fusion, until all the detection features in the multiple detection feature sets are fused, the final target interactive fusion detection feature set is obtained. For example, after F1, F2, and F3 are fused, F4 is extracted and fused with the previously fused feature set, and so on, until all the detection feature sets participate in the fusion process.
[0115] When randomly extracting the detection feature set, the randomness of the extraction must be ensured; when calculating the similarity, a suitable algorithm must be selected and the accuracy of the calculation must be ensured; the normalization processing must be performed according to the established method to ensure the uniformity of the data; when training the convolution network layer, enough sample data must be used and the sufficiency of the training must be ensured to enable the network layer to accurately fuse the features. Through a series of operations, different scale features in multiple detection feature sets can be effectively interactively fused, thereby obtaining a more comprehensive and accurate target interactive fusion detection feature set, providing a reliable basis for early detection of urinary system tumors. For example, different detection feature sets may have captured different features of early tumors, such as short-term indicator fluctuation features and long-term trend features. Through cross-scale interactive fusion, these different scale features can be combined to more comprehensively reflect early signs of tumors and improve detection accuracy.
[0116] Embodiment 5:
[0117] When using the configured multiple detection network layers to perform multi-scale feature analysis on the detection data sequence of the target sample within the preset monitoring window, it is first necessary to determine that the receptive period of the multiple detection network layers is determined by multiple feature correlation time periods. The feature correlation time periods are derived from the analysis of historical detection data, for example, the multiple feature correlation time periods obtained by calculating the mean value of the historical instantaneous detection feature correlation time period set in Embodiment 3, which reflect the duration of features related to tumor occurrence in historical data. Each detection network layer is configured to correspond to a different feature correlation time period, thereby forming different receptive periods, so that each network layer can analyze the detection data from a specific time scale.
[0118] Suppose the detection data sequence of the target sample within the preset monitoring window is a series of data points collected in chronological order, for example, in a one-week monitoring window, a sample is collected once a day, and each time a set of values containing multiple detection indicators such as protein content, cell morphology parameters, etc. are obtained. These data are arranged in chronological order to form a detection data sequence. When processing this data sequence, the multiple detection network layers each scan and extract features based on the corresponding receptive period.
[0119] For example, the first detection network layer corresponds to a short feature correlation period, such as 1 day. This network layer analyzes the data features in each time window with a time span of 1 day in the detection data sequence, capturing rapid changes in a short period of time. For example, when the detection data sequence contains 7 days of data, the network layer analyzes the single-day data of day 1, day 2, …, day 7 in turn, and extracts instantaneous features such as the maximum value, minimum value, or change amplitude of a certain indicator in a single day.
[0120] The second detection network layer may correspond to a longer receptive period, such as 3 days. It analyzes the data features of consecutive 3 days in the detection data sequence with a time window of 3 days, and extracts trend features such as the upward or downward trend, fluctuation range, etc. of a certain indicator in 3 days.
[0121] The receptive period of the third detection network layer may be longer, such as 7 days, i.e. covering the entire preset monitoring window. This network layer analyzes the 7-day detection data as a whole, and extracts long-term features such as the average value, overall trend, etc. of the indicator in the entire monitoring period.
[0122] The structure of each detection network layer is designed based on a neural network and includes multiple neuron layers such as convolution layers and activation layers. When the detection data sequence is input into the network layer, the network layer automatically learns and extracts features matching the receptive period through the operation of the weight matrix and the input data. For example, a network layer with a short receptive period may focus more on local mutations of the data, and its convolution kernel size and weight settings are biased towards capturing high-frequency changes; while a network layer with a long receptive period has a larger convolution kernel and weight settings that focus more on the extraction of low-frequency trends.
[0123] In the multi-scale feature analysis process, each detection network layer independently processes the detection data sequence and outputs a corresponding set of detection features. For example, the short-period network layer outputs a set of multiple single-day features, the medium-period network layer outputs a set of multiple 3-day window features, and the long-period network layer outputs a set of overall features for the entire period. Each feature in these sets of detection features corresponds to an abstract representation of the original data at a specific time scale, such as the fluctuation feature of a certain indicator in a short period of time, the trend feature in a medium period, and the overall feature in a long period.
[0124] The cooperative work of multiple detection network layers is reflected in the parallel extraction of features of different scales. For example, when the target sample may have early signs of tumor, the short-period network layer may capture the feature of a sudden increase in a certain indicator in a single day, the medium-period network layer may find the trend of continuous rise of the indicator in nearly 3 days, and the long-period network layer may show the slight upward trend of the indicator compared with historical data in the entire monitoring period. These three types of features reflect the changes in the data from different time dimensions and together constitute multiple detection feature sets.
[0125] In the processing of a specific target sample, assuming that in the detection data sequence of the sample, a tumor marker indicator suddenly rises on the 3rd day, maintains a high level on the 4th-5th day, and slightly decreases on the 6th-7th day but is still higher than the normal range. The short-perception-period detection network layer will extract the feature of sudden increase of the indicator in the 3-day window, the medium-perception-period network layer will extract the feature of continuous high level in the 3-5-day window, and the long-perception-period network layer will extract the feature of overall high level in the 7-day window. These three feature sets respectively describe the changes of the indicator from different scales, providing multi-dimensional information for subsequent cross-scale fusion.
[0126] In the feature extraction process, the parameters of each detection network layer have been optimized through historical data in the training stage. For example, by using historical detection data containing known tumor cases and normal cases, the network layer is supervised trained, and the weight parameters are adjusted so that the network layer can output more discriminative features when facing tumor samples and output more normal pattern features when facing normal samples. After training, the parameters of the network layer are fixed and used for feature analysis of actual target samples.
[0127] The whole multi-scale feature analysis process strictly processes data according to the perception period of each detection network layer, ensuring that each network layer focuses on feature extraction of a specific time scale. Different network layers are independent of each other and complementary to each other, independent in that they process different time window data, and complementary in that the output feature sets describe the sample from different dimensions, together providing comprehensive feature support for early detection of tumors. Through this multi-scale analysis method, abnormal changes in different time scales in the sample can be captured, avoiding the early signals that may be missed by single-scale analysis, thereby improving the comprehensiveness and accuracy of detection. For example, the abnormality of some tumor markers may first appear as short-term dramatic fluctuations, then as medium-term trend changes, and finally as overall level changes in the long-term scale. Multi-scale feature analysis can cover these different stages of change features and provide more abundant information for early detection.
[0128] It is to be understood that the terminology used herein such as first and second, and the like, is only used to distinguish one entity or action from another entity or action, and does not necessarily require or imply any such actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.
[0129] While embodiments of the present application have been shown and described with reference to particular embodiments thereof, it will be understood by those skilled in the art that various changes in form and details can be made therein without departing from the spirit and scope of the application. The scope of the application is defined by the appended claims and their equivalents.
Claims
1. A method for early detection of urinary system tumors based on neural networks, characterized in that, The method comprises: acquiring a target type of a target sample, searching in a historical detection database based on the target type, and determining a set of historical tumor detection records; distinguishing the set of historical tumor detection records, and determining a plurality of sets of distinguished historical detection records; traversing the plurality of sets of distinguished historical detection records to extract feature central values and a plurality of feature correlation time periods; using the plurality of feature correlation time periods as a plurality of receptive periods of a plurality of detection network layers, performing multi-scale feature analysis on a detection data sequence of the target sample within a preset monitoring window by using the plurality of detection network layers configured to obtain a plurality of sets of detection features; performing cross-scale interactive fusion on the plurality of sets of detection features to obtain a set of target interactive fusion detection features, and using the set of target interactive fusion detection features as an early detection result of a urinary system tumor.
2. The method for early detection of urinary system tumors based on neural networks according to claim 1, characterized in that, traversing the plurality of sets of distinguished historical detection records to extract feature central values and a plurality of feature correlation time periods, comprising: traversing the plurality of sets of distinguished historical detection records to extract a plurality of sets of distinguished historical detection data sequences; extracting instantaneous detection data from the plurality of sets of distinguished historical detection data sequences respectively to determine a plurality of sets of historical instantaneous detection data, wherein each historical instantaneous detection data is detection data at a time point of maximum data change fluctuation in each distinguished historical detection data sequence; extracting features from the plurality of sets of historical instantaneous detection data to obtain a plurality of sets of historical instantaneous detection features, and performing centralized analysis on the plurality of sets of historical instantaneous detection features to determine a plurality of feature central values; using the plurality of sets of historical instantaneous detection features as an index, performing feature correlation time period diffusion identification on the plurality of sets of distinguished historical detection data sequences to determine a plurality of sets of historical instantaneous detection feature correlation time periods, wherein each historical instantaneous detection feature correlation time period reflects a duration of data associated with a historical instantaneous detection feature at a tumor occurrence time point in a distinguished historical detection data sequence; calculating a mean value of the plurality of sets of historical instantaneous detection feature correlation time periods to obtain the plurality of feature correlation time periods.
3. The method for early detection of urinary system tumors based on neural networks according to claim 2, characterized in that, extracting features from the plurality of sets of historical instantaneous detection data to obtain a plurality of sets of historical instantaneous detection features, and performing centralized analysis on the plurality of sets of historical instantaneous detection features to determine a plurality of feature central values, comprising: extracting features from the plurality of sets of historical instantaneous detection data by using a feature extraction network layer to obtain the plurality of sets of historical instantaneous detection features; traversing the plurality of sets of historical instantaneous detection features to determine a plurality of historical instantaneous detection feature mean values; iterating the plurality of historical instantaneous detection feature mean values in the plurality of sets of historical instantaneous detection features according to a preset iteration time period to obtain a plurality of iteration historical instantaneous detection features; when an aggregation amount of the plurality of iteration historical instantaneous detection features is less than or equal to an aggregation amount of the plurality of historical instantaneous detection feature mean values, using the plurality of historical instantaneous detection feature mean values as the plurality of feature central values.
4. The method for early detection of urinary system tumors based on neural networks according to claim 3, characterized in that, When the aggregation quantity of the plurality of iteration history instantaneous detection features is greater than the aggregation quantity of the plurality of history instantaneous detection feature means, it is judged whether the aggregation quantity difference between the plurality of iteration history instantaneous detection features and the plurality of history instantaneous detection feature means is greater than or equal to a preset aggregation quantity difference threshold value, if yes, the plurality of iteration history instantaneous detection features are continued to be iterated based on the plurality of iteration history instantaneous detection features until a maximum iteration number is satisfied, and the plurality of iteration history instantaneous detection features obtained in the last iteration are taken as the plurality of feature central values; If not, iteration is stopped, and the plurality of iteration history instantaneous detection features are taken as the plurality of feature central values.
5. The method for early detection of urinary system tumors based on neural networks according to claim 4, characterized in that, The plurality of history instantaneous detection feature sets are taken as indexes, and feature correlation time period diffusion recognition is performed on the plurality of differentiated history detection data sequence sets to determine a plurality of history instantaneous detection feature correlation time period sets, including: The feature extraction network layer is used to extract detection features from the plurality of differentiated history detection data sequence sets to obtain a plurality of differentiated history detection feature sequence sets; A history instantaneous detection feature is randomly extracted from the plurality of history instantaneous detection feature sets as a first history instantaneous detection feature, and a corresponding first differentiated history detection feature sequence is matched from the plurality of differentiated history detection feature sequence sets; According to a preset approximate correlation scale, a neighbor search is performed on the first history instantaneous detection feature in the first differentiated history detection feature sequence to obtain a first history instantaneous detection feature neighborhood; The duration of the first history instantaneous detection feature neighborhood is counted, and the counting result is taken as a first history instantaneous detection feature correlation time period; According to a preset neighbor correlation scale, the plurality of history instantaneous detection feature sets are subjected to feature correlation time period diffusion recognition in the corresponding plurality of differentiated history detection data sequence sets to determine a plurality of history instantaneous detection feature correlation time period sets.
6. The method for early detection of urinary system tumors based on neural networks according to claim 1, characterized in that, Cross-scale interactive fusion is performed on the plurality of detection feature sets to obtain a target interactive fusion detection feature set, including: A first detection feature set and a second detection feature set are randomly extracted from the plurality of detection feature sets; Similarity recognition is performed on the first detection feature set and the second detection feature set to determine a first feature similarity set; The first feature similarity set is normalized to obtain a first feature similarity normalized value set; The first feature similarity normalized value set and the second detection feature set are subjected to convolution calculation to obtain a first interactive fusion detection feature set; A third detection feature set is randomly extracted from the plurality of detection feature sets, and cross-scale interactive fusion is performed on the third detection feature set and the first interactive fusion detection feature set to obtain a second interactive fusion detection feature set; After multiple cross-scale interactive fusions, the detection features in the plurality of detection feature sets are fused to obtain the target interactive fusion detection feature set.
7. The method for early detection of urinary system tumors based on neural networks according to claim 6, characterized in that, Including: A plurality of sample feature similarity normalized value sets, a plurality of sample detection feature sets, and a plurality of sample interactive fusion detection feature sets are taken as training data; The network layer constructed based on the convolutional neural network is supervised trained by using the training data until the training converges, and a trained convolutional network layer is obtained; The first feature similarity normalization value set and the second detection feature set are subjected to convolutional calculation by using the convolutional network layer, and the first interaction fusion detection feature set is obtained.
8. The method for early detection of urinary system tumors based on neural networks as claimed in claim 1, wherein, The heterogeneous classification of the historical tumor detection record set is performed, and a plurality of classified historical detection record sets are determined, including: A plurality of historical tumor detection records are randomly extracted from the historical tumor detection record set; The plurality of historical tumor detection records are enumerated two by two, and a plurality of enumeration combinations are obtained; It is judged whether there is an enumeration combination with a record similarity greater than a preset record similarity threshold in the plurality of enumeration combinations, and if not, the plurality of historical tumor detection records are taken as a plurality of classification targets; Based on the plurality of classification targets, the heterogeneous classification of the historical tumor detection record set is performed according to the preset record similarity threshold, and the plurality of classified historical detection record sets are obtained, wherein each classified historical detection record set corresponds to a classification target.
9. The method for early detection of urinary system tumors based on neural networks according to claim 8, characterized in that, It is judged whether there is an enumeration combination with a record similarity greater than a preset record similarity threshold in the plurality of enumeration combinations, and if not, the plurality of enumeration combinations are combined into the same classification target; Based on the combined classification target, the heterogeneous classification of the historical tumor detection record set is performed according to the preset record similarity threshold, and the plurality of classified historical detection record sets are obtained.
10. The method for early detection of urinary system tumors based on neural networks according to claim 9, characterized in that, The heterogeneous classification of the plurality of classified historical detection record sets is verified, including: Two classified historical detection record sets are randomly extracted from the plurality of classified historical detection record sets; The cross-set record similarity average of the two classified historical detection record sets is calculated; If the cross-set record similarity average is less than or equal to a preset cross-set similarity threshold, the plurality of classified historical detection record sets are maintained; If the cross-set record similarity average is greater than the preset cross-set similarity threshold, the heterogeneous classification of the historical tumor detection record set is performed again until the cross-set similarity threshold requirement is met.