An intelligent system for analyzing anal fistula based on artificial intelligence and a method thereof

Abstract

The application discloses an intelligent anus fistula analysis system and method based on artificial intelligence, and belongs to the technical field of intelligent anus fistula analysis, and comprises a multi-source data collection unit, a data preprocessing unit, a segmentation processing module, a local lesion feature enhancement unit, a multi-modal feature extraction unit, a time-frequency joint processing unit, a phased feature optimization unit and an efficacy evaluation unit.The application can comprehensively depict anus fistula lesion healing characteristics, and improve the objectivity and comprehensiveness of efficacy evaluation.The application can accurately separate and respectively depict long-term healing trend and short-term local stress fluctuation of the lesion, so that the efficacy evaluation is more in line with the real clinical course law.The application introduces a phased feature optimization mechanism in combination with the whole cycle treatment stage of the anus fistula, can perform fine quantitative evaluation at different treatment stages, and realizes phased efficacy analysis on patients.

Description

Technical Field

[0001] This invention relates to the field of intelligent analysis technology for anal fistulas, specifically to an artificial intelligence-based intelligent analysis system and method for anal fistulas. Background Technology

[0002] Anal fistula is a disease characterized by its complex fistula structure, difficulty in treatment, and high recurrence rate, which has a serious negative impact on patients' mental health and quality of life. Symptoms include purulent discharge, pain, and perianal itching, severely affecting patients' quality of life. Long-term untreated anal fistulas can also lead to damage to the anal sphincter function and even increase the risk of cancer.

[0003] Surgical treatment remains the most effective method for treating anal fistulas. However, due to the complexity and tortuous nature of the fistula tract, the presence of branches and deep dead spaces, the unique anatomical and physiological environment of the anus, and the heavy reliance on the surgeon's experience, surgical outcomes vary widely. Therefore, postoperative rehabilitation for anal fistulas requires more precise and personalized treatment plans.

[0004] Currently, there are few clinical assessment methods, resulting in low diagnostic accuracy. Treatment often relies on the doctor's experience, leading to numerous technical shortcomings and pain points.

[0005] (1) The assessment is highly subjective and relies heavily on the subjective experience of clinicians. The data from multiple sources are independent of each other and cannot comprehensively and systematically reflect the healing of lesions.

[0006] (2) The treatment of anal fistula involves multiple treatment nodes and stages, and the conditions of each patient are also different. Existing technologies focus on analyzing temporal characteristics, but do not fully explore time-frequency characteristics, cannot distinguish between long-term healing trends and short-term local fluctuation characteristics, and are difficult to fit the temporal pattern of clinical treatment.

[0007] (3) Without combining the quality stages of anal fistula to construct a stage-by-stage feature optimization mechanism, it is impossible to carry out detailed quantitative assessment for different treatment stages, making it difficult to meet the needs of patients for stage-by-stage analysis.

[0008] Based on this, the present invention designs an artificial intelligence-based intelligent analysis system and method for anal fistula to solve the above problems. Summary of the Invention

[0009] To address the aforementioned shortcomings of existing technologies, this invention provides an artificial intelligence-based intelligent analysis system and method for anal fistula.

[0010] To achieve the above objectives, the present invention provides the following technical solution:

[0011] An artificial intelligence-based intelligent analysis system for anal fistula includes:

[0012] The multi-source data collection unit collects multi-source diagnostic data from patients with anal fistula, including imaging data, pathological data, and medical records.

[0013] The data preprocessing unit is used to preprocess multi-source diagnostic data;

[0014] The segmentation processing module is used to segment image data and pathological data separately to obtain the image lesion segmentation mask and pathological tissue segmentation mask under the corresponding treatment node.

[0015] The local lesion feature enhancement unit uses a sliding window local feature enhancement algorithm to perform targeted enhancement of local lesion features on imaging and pathological lesion feature maps;

[0016] The multimodal feature extraction unit extracts the image temporal features V1(T) at each treatment node based on the enhanced image data and the corresponding image lesion segmentation mask; extracts the pathological temporal features V2(T) at each treatment node based on the preprocessed pathological data and pathological tissue segmentation mask; and extracts the disease temporal features V3(T) at each treatment node based on the disease structured data; where T={T1,T2,...,T...} n}, T n For the nth treatment node; associate the treatment time nodes of V1(T), V2(T), and V3(T) and concatenate them in chronological order to form a temporal feature matrix;

[0017] The time-frequency joint processing unit is used to normalize the time-series feature matrix, remove outliers, and use weighted wavelet transform to perform time-frequency decomposition on the treatment time sequence with non-uniform time intervals, splitting it into low-frequency trend components and high-frequency fluctuation components. Finally, the original time-domain features, low-frequency trend features and high-frequency fluctuation features are adaptively fused through dynamic weights.

[0018] The phased feature optimization unit, using the treatment node as an index, optimizes the time-frequency joint feature X. TF The network is split according to treatment stages, obtaining time-frequency feature subsets for each treatment stage. A DenseNet optimized network consisting of dense blocks and transition layers is constructed, and multi-scale dilated convolutional modules are embedded inside the dense blocks to obtain the depth-optimized feature set O of the time-frequency feature subsets for each treatment stage. s ;

[0019] The efficacy assessment unit is based on a deeply optimized feature set O s To conduct quantitative assessments of the treatment effects at each stage of the patient's treatment.

[0020] Furthermore, the imaging data includes: enhanced perianal MRI images, endorectal ultrasound images, and endoscopic time-series images acquired at different treatment stages; the pathological data includes: pathological sections of continuous anal fistula tracts and pathological annotation data of inflammatory tissue acquired at different treatment stages; and the medical record data includes: electronic medical record texts and clinical symptom scores of anal fistulas corresponding to different treatment stages.

[0021] Furthermore, the clinical symptom scores of anal fistula at each treatment node are obtained based on the electronic medical record text of the corresponding treatment node and the clinical symptom scoring table for anal fistula.

[0022] Furthermore, a sliding window is used to perform a region-by-region sliding scan of the input image and pathological lesion feature map, extracting local lesion features within each window; the features within the window are mapped through a 1×1 convolution operation to complete the initial feature enhancement; the initially enhanced local lesion features are input into the convolutional attention fusion module, and the key lesion feature channels are weighted by channel attention and the core area of ​​the lesion is focused by spatial attention to further highlight key information.

[0023] Furthermore, the specific steps of the time-frequency joint processing unit are as follows:

[0024] (1) Outliers were removed using the 3σ criterion, and the time series features were normalized using the Min-Max normalization method to obtain the normalized time series feature matrix X∈R. n×d , where n is the number of treatment nodes, and d is the total dimension of the concatenated three types of temporal features V1(T), V2(T), and V3(T);

[0025] (2) For the normalized time series feature matrix X∈R n×d A weighted wavelet transform is performed column by column. Each column of the matrix corresponds to a one-dimensional time sequence of a single feature across n non-uniform treatment nodes. The actual time interval between each node is used as the timestamp input to WWZ to transform the one-dimensional time-domain evolution feature into a two-dimensional time-frequency amplitude distribution. Amplitude extraction and dimension alignment are performed on the time-frequency results of all features to construct a time-frequency amplitude matrix |F|∈R. e×n×d e represents the number of frequency points;

[0026] (3) The improved Otsu algorithm is used to determine the high-low frequency splitting threshold f0, and the time-frequency amplitude matrix |F| is split into a low-frequency time-frequency matrix F according to f0. L and high-frequency time-frequency matrix F H and respectively for F L F H By performing mean aggregation along the frequency dimension, low-frequency trend features are obtained. ∈R n×d High-frequency fluctuation characteristics ∈R n×dIt enables the separation of slowly changing long-term evolution information into low-frequency trend components and rapidly changing short-term fluctuation information into high-frequency fluctuation components.

[0027] (4) Time-frequency joint feature fusion: Based on dynamic weight allocation, adaptive weighted fusion is used to deeply fuse time-domain features, low-frequency trend features and high-frequency fluctuation features.

[0028] Furthermore, in step (3), the average amplitude of each frequency component along the frequency dimension of the time-frequency amplitude matrix |F| is calculated at all treatment nodes and feature dimensions to obtain a single-frequency amplitude sequence.

[0029] For each frequency k, calculate its arithmetic mean over all treatment nodes i and all feature dimensions j:

[0030] A(k) =

[0031] The effective frequency range is obtained by removing the DC component and the extreme high-frequency noise component. Each frequency in the range is used as a candidate threshold to divide the single-frequency amplitude sequence into a low-frequency amplitude subset and a high-frequency amplitude subset. The inter-class variance of the two amplitude subsets under each candidate threshold is calculated, and the frequency that maximizes the inter-class variance is selected as the optimal splitting threshold f0.

[0032] Furthermore, step (4) specifically involves: for each feature dimension, constructing a feature association graph structure with time-domain, low-frequency, and high-frequency features as nodes; learning the degree of association between the three types of features (time-domain, low-frequency, and high-frequency) under this dimension through a graph attention network; automatically calculating the feature contribution of each type of feature to the healing of anal fistula; normalizing the feature contribution using the Softmax function to obtain dynamic adaptive weights; and performing dimension-wise weighted fusion of the three types of features to obtain a time-frequency joint feature X that combines the original temporal changes, long-term healing trends, and short-term local reactions. TF The specific formula is as follows:

[0033] X TF (i)=ω t (i)·X(i)+ω l (i)· (i)+ω h (i)· (i) (i=1,2,...,d)

[0034] Where, ω t (i), ω l (i), ω h (i) represents the dynamic adaptive weights of time-domain features, low-frequency trend features, and high-frequency fluctuation features under the i-th feature dimension; X(i) represents the time-domain feature value under the i-th feature dimension. (i) represents the low-frequency trend feature under the i-th feature dimension; (i) represents the high-frequency fluctuation feature under the i-th feature dimension;

[0035] The formula for Softmax normalization is:

[0036]

[0037] Among them, c t (i), c l (i), c h (i) represents the feature contribution of time-domain features, low-frequency trend features, and high-frequency fluctuation features to the healing of anal fistula under the i-th feature dimension, which is learned by graph attention network; exp(·) is an exponential function.

[0038] Furthermore, the specific steps of the phased feature optimization unit are as follows: The treatment node set T = {T1, T2, ..., T...} n The treatment is divided into S treatment phases; using the treatment node as an index, the time-frequency joint feature X is analyzed. TF The process involves segmenting the tissue into stages based on nodes to form a feature subset representing the long-term healing stage. Multi-scale dilated convolutional modules are embedded within dense blocks, and dilated convolutional kernels with different dilation rates are used for multi-scale feature extraction to fully capture the healing characteristics and temporal changes of anal fistula lesions at different scales. Dense blocks enable the intensive reuse and deep mining of multi-layer time-frequency features. A transition layer is used to reduce the dimensionality and normalize the features, removing redundant time-frequency components to obtain a depth-optimized feature set O corresponding to the treatment nodes in each treatment stage. s .

[0039] Furthermore, the specific steps of the efficacy assessment unit are as follows:

[0040] Extracting the deep optimization feature O1=[z] corresponding to the preoperative baseline stage 11 ,z 12 ,...,z 1d As the initial lesion benchmark before anal fistula treatment, the positive and negative polarities of each dimension of the depth optimization feature are defined;

[0041] For each treatment node n, calculate the standardized rate of change R of the features relative to the preoperative value by dimension. n,i Then, the dimensions are weighted and fused to obtain the single-node comprehensive healing score I for the nth treatment node. n ;

[0042] (n>1)

[0043] Among them, z 1,i Let z represent the i-th eigenvalue of Z1. n,iR represents the i-th eigenvalue of the depth optimization feature of the nth treatment node; ni >0: Feature improvement; R ni =0: Characteristic stability; R ni <0: Characteristic deterioration;

[0044] I n =

[0045] Among them, w i Represents the clinical importance weight of the i-th feature;

[0046] Let I be the comprehensive healing score for a single node in a certain treatment phase. n1 ~I n2 The time-weighted average was used to calculate the comprehensive healing score E for each stage. S :

[0047]

[0048]

[0049] Among them, a n This represents the temporal weight of the nth treatment node within this stage.

[0050] To better achieve the objectives of this invention, this invention also provides an artificial intelligence-based intelligent analysis method for anal fistulas, comprising the following steps:

[0051] I. Collect multi-source diagnostic data from patients with anal fistulas, including imaging data, pathological data, and medical records;

[0052] 2. Perform data anonymization, data cleaning, and time-series alignment preprocessing on imaging and pathological data; perform data anonymization, data cleaning, time-series alignment preprocessing, and structured fusion on medical documents to generate structured medical data.

[0053] 3. The imaging data and pathological data are segmented separately to obtain the imaging lesion segmentation mask and pathological tissue segmentation mask under the corresponding treatment node.

[0054] Fourth, a sliding window local feature enhancement algorithm is used to enhance the local lesion features of the imaging and pathological lesion feature maps, highlighting the key feature information of the anal fistula lesion;

[0055] V. Based on the enhanced image data and corresponding image lesion segmentation masks, extract the image temporal features V1(T) at each treatment node; based on the preprocessed pathological data and pathological tissue segmentation masks, extract the pathological temporal features V2(T) at each treatment node; extract the disease temporal features V3(T) at each treatment node based on the structured disease data; where T={T1,T2,...,T...}n}, T n For the nth treatment node; associate the treatment time nodes of V1(T), V2(T), and V3(T), and concatenate them in chronological order to form a temporal feature matrix;

[0056] VI. The time-frequency joint processing unit is used to normalize the time-series feature matrix, remove outliers, and use weighted wavelet transform to perform time-frequency decomposition on the treatment time sequence with non-uniform time intervals, splitting it into low-frequency trend components and high-frequency fluctuation components. Finally, the original time-domain features, low-frequency trend features and high-frequency fluctuation features are adaptively fused through dynamic weights.

[0057] VII. Using the treatment node as an index, analyze the time-frequency joint feature X. TF The network is split according to treatment stages, obtaining time-frequency feature subsets for each treatment stage. A DenseNet optimized network consisting of dense blocks and transition layers is constructed, and multi-scale dilated convolutional modules are embedded inside the dense blocks to obtain the depth-optimized feature set O of the time-frequency feature subsets for each treatment stage. s ;

[0058] 8. Based on deep optimization of feature set O s To conduct quantitative assessments of the treatment effects at each stage of the patient's treatment.

[0059] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) Based on multi-source data of images, pathology, and disease text, the present invention processes data through a data preprocessing unit, a segmentation processing module, a local lesion feature enhancement unit, a multimodal feature extraction unit, a time-frequency joint processing unit, and a phased feature optimization unit. The results are quantified through the efficacy evaluation unit, which can comprehensively depict the healing characteristics of anal fistula lesions and improve the objectivity and comprehensiveness of efficacy evaluation.

[0060] (2) In view of the non-uniform time sequence characteristics of multiple treatment nodes and multiple treatment stages of anal fistula, the present invention uses a time-frequency joint processing unit to decompose the treatment time sequence of non-uniform time intervals into low-frequency trend components and high-frequency fluctuation components by using weighted wavelet transform (WWZ). This avoids the loss of features and misjudgment of patterns caused by traditional time-domain analysis alone. It can accurately separate and characterize the long-term healing trend of the lesion and the short-term local stress fluctuations, so that the efficacy assessment is more in line with the real clinical course. Through dynamic weighting, the original time-domain features, low-frequency trend features and high-frequency fluctuation features are adaptively fused. The weights can be automatically allocated according to the contribution of different feature dimensions to the healing assessment, avoiding the assessment bias caused by fixed weights.

[0061] (3) This invention combines the whole cycle treatment of anal fistula with the introduction of a phased feature optimization mechanism, which can perform refined quantitative evaluation at different treatment stages and realize phased efficacy analysis of patients. Attached Figure Description

[0062] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0063] Figure 1 This is a diagram illustrating the architecture of an artificial intelligence-based intelligent analysis system for anal fistulas according to the present invention. Figure 2 This is a flowchart of an artificial intelligence-based intelligent analysis method for anal fistula according to the present invention. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0065] Example 1: Please refer to the accompanying drawings in the instruction manual. Figure 1 An artificial intelligence-based intelligent analysis system for anal fistulas, characterized in that it includes:

[0066] The multi-source data collection unit collects multi-source diagnostic data from patients with anal fistula, including imaging data, pathological data, and medical records.

[0067] Imaging data include: perianal enhanced magnetic resonance imaging (DCE-MRI) images acquired at different treatment points, endorectal ultrasound images (ERUS), endoscopic time-series images, etc.

[0068] Pathological data include: pathological sections of continuous anal fistula tracts and pathological annotation data of inflammatory tissues collected at different treatment stages;

[0069] The medical record data includes: electronic medical record texts corresponding to different treatment stages and clinical symptom scores for anal fistula.

[0070] The clinical symptom scores for anal fistula at each treatment node are obtained based on the electronic medical record text and the anal fistula clinical symptom scoring table for the corresponding treatment node. Specifically, four categories of clinical symptom terms are extracted from the electronic medical record: perianal secretions, pain and activity, perianal manifestations, and perianal induration. These extracted clinical symptom terms are then matched against a pre-defined anal fistula clinical symptom scoring table. If a clinical symptom term matches an item in the scoring table, it is automatically mapped to the corresponding scoring level. If no matching item is found, the closest scoring level is determined through synonym matching and semantic similarity calculation. Finally, the four categories of scores are summed to obtain the patient's anal fistula clinical symptom score.

[0071] The clinical symptom scoring scale for anal fistula is detailed in Table 1, which includes four aspects: perianal discharge, pain and mobility, perianal manifestations, and induration. According to the scoring criteria, individual scores are graded according to severity; the higher the overall score, the more severe the anal fistula.

[0072] Table 1 Clinical Symptom Scoring Table for Anal Fistula

[0073]

[0074] The data preprocessing unit performs preprocessing on multi-source diagnostic data, including data anonymization, data cleaning, and time-series alignment. It also performs format conversion, noise reduction, grayscale normalization, and standardization of image resolution and size for image data; and noise reduction, color correction, distortion repair, and annotation standardization for pathological data. The unit also performs structured fusion of electronic medical record text and anal fistula clinical symptom scores to generate structured disease data.

[0075] The segmentation module is used to segment image data and pathological data separately. For image data, the focus is on segmenting fistulas, lesions and surrounding sphincter structures. For pathological data, the focus is on segmenting inflammatory tissue, granulation tissue, fibrotic areas and normal perianal glands, to obtain the image lesion segmentation mask and pathological tissue segmentation mask under the corresponding treatment node.

[0076] The local lesion feature enhancement unit uses a sliding window local feature enhancement algorithm to perform targeted enhancement of local lesion features on imaging and pathological lesion feature maps, highlighting key feature information of anal fistula lesions.

[0077] Specifically, the sliding window parameters are set to be adjustable from 3×3 to 7×7, with a sliding step size of 1 to 2 pixels. The sliding window is used to perform a region-by-region sliding scan of the input image and pathological lesion feature map, extracting local lesion features within each window. A 1×1 convolution operation is then used to perform feature mapping on the features within the window, enhancing key information such as lesion edge contours, tissue texture differences, and uneven grayscale distribution, thus completing preliminary feature enhancement. The preliminarily enhanced local lesion features are then input into the convolutional attention fusion module. Channel attention weighting of key lesion feature channels and spatial attention focusing on the core area of ​​the lesion further highlight key information such as fistula boundaries, inflammatory tissue, and damaged sphincter muscles.

[0078] The multimodal feature extraction unit extracts the image temporal features V1(T) at each treatment node based on the enhanced image data and the corresponding image lesion segmentation mask; extracts the pathological temporal features V2(T) at each treatment node based on the preprocessed pathological data and pathological tissue segmentation mask; and extracts the disease temporal features V3(T) at each treatment node based on the disease structured data; where T={T1,T2,...,T...} n}, T n For the nth treatment node (e.g., preoperative, 1 week postoperative, 2 weeks postoperative, 1 month postoperative, etc.), by associating the treatment time nodes of V1(T), V2(T), and V3(T), ensure that the three types of features—image, pathology, and disease condition—are aligned on the same temporal dimension and spliced ​​together in chronological order to form a temporal feature matrix.

[0079] Among them, the temporal features of the images are spatial morphological features such as fistula morphology, lesion volume, sphincter integrity, lesion grayscale distribution and texture heterogeneity extracted under the regional constraints of the image lesion segmentation mask, as well as temporal change features such as the size changes of fistulas and lesions between different treatment nodes.

[0080] The pathological temporal features are the area proportions, cell density, and tissue texture features of inflammatory tissue, granulation tissue, fibrotic areas, and normal perianal glands extracted under the constraints of pathological tissue segmentation mask, as well as the temporal change features of various tissue regions between different treatment nodes.

[0081] The temporal characteristics of the disease are based on the clinical symptom scores, secretions, pain and activity, perianal manifestations and induration of anal fistula extracted from structured disease data, as well as the temporal changes in the relief or aggravation trends of the above indicators and the magnitude of score changes between each treatment node.

[0082] The time-frequency joint processing unit is used to normalize the time-series feature matrix, remove outliers, and use weighted wavelet transform (WWZ) to decompose the treatment time series with non-uniform time intervals into low-frequency trend components and high-frequency fluctuation components. Finally, the original time-domain features, low-frequency trend features and high-frequency fluctuation features are adaptively fused through dynamic weights.

[0083] The specific steps are as follows:

[0084] (1) Outliers were removed using the 3σ criterion, and the time series features were normalized using the Min-Max normalization method to obtain the normalized time series feature matrix X∈R. n×d , where n is the number of treatment nodes, and d is the total dimension of the concatenated three types of temporal features V1(T), V2(T), and V3(T);

[0085] (2) For the normalized time series feature matrix X∈R n×d Weighted wavelet transform (WWZ) is performed column by column. Each column of the matrix corresponds to a one-dimensional time sequence of a single feature across n non-uniform treatment nodes. The actual time interval between each node is used as the timestamp input to WWZ to transform the one-dimensional time-domain evolution feature into a two-dimensional time-frequency amplitude distribution. Amplitude extraction and dimension alignment are performed on the time-frequency results of all features to construct a time-frequency amplitude matrix |F|∈R. e×n×d , e represents the number of frequency points.

[0086] (3) The improved Otsu algorithm is used to determine the high- and low-frequency splitting threshold f0, specifically:

[0087] The average amplitude of each frequency component along the frequency dimension of the time-frequency amplitude matrix |F| is calculated at all treatment nodes and feature dimensions to obtain a single-frequency amplitude sequence.

[0088] For each frequency k, calculate its arithmetic mean over all treatment nodes (i=1..n) and all feature dimensions (j=1..d):

[0089] A(k) =

[0090] Based on the sparse and long-cycle treatment time sequence characteristics of anal fistula before surgery, 1 week after surgery, 2 weeks after surgery, and 1 month after surgery, the DC component and extreme high-frequency noise component are removed to obtain the effective frequency range; each frequency in this range is used as a candidate threshold to divide the single-frequency amplitude sequence into a low-frequency amplitude subset and a high-frequency amplitude subset.

[0091] Calculate the inter-class variance of the two amplitude subsets under each candidate threshold, and select the frequency that maximizes the inter-class variance as the optimal splitting threshold f0; split the time-frequency amplitude matrix |F| into a low-frequency time-frequency matrix F according to f0. L and high-frequency time-frequency matrix F Hand respectively for F L F H By performing mean aggregation along the frequency dimension, low-frequency trend features are obtained. ∈R n×d High-frequency fluctuation characteristics ∈R n ×d It enables the separation of slow-changing long-term evolution information into low-frequency trend components and fast-changing short-term fluctuation information into high-frequency fluctuation components.

[0092] Low frequency trend characteristics Frequency ≤ f0 corresponds to the long-term healing trend of anal fistula (such as the resolution of anal fistula inflammation).

[0093] High-frequency fluctuation characteristics Frequency > f0 corresponds to short-term local reactions of anal fistula (such as local reactions after a single treatment).

[0094] (4) Time-frequency joint feature fusion: Based on dynamic weight allocation, adaptive weighted fusion is used to deeply fuse time-domain features, low-frequency trend features and high-frequency fluctuation features;

[0095] Specifically, for each feature dimension (d dimensions in total), a feature association graph structure is constructed with time-domain, low-frequency, and high-frequency features as nodes. A graph attention network (GAT) is used to learn the degree of association between the three types of features in that dimension, automatically calculating the feature contribution of each type of feature to fistula healing. The feature contribution is normalized using the Softmax function to obtain dynamic adaptive weights. The three types of features are then weighted and fused dimension by dimension to obtain a time-frequency joint feature X that combines the original temporal changes, long-term healing trends, and short-term local responses. TF The specific formula is as follows:

[0096] X TF (i)=ω t (i)·X(i)+ω l (i)· (i)+ω h (i)· (i) (i=1,2,...,d)

[0097] Where, ω t (i), ω l (i), ω h (i) represents the dynamic adaptive weights of time-domain features, low-frequency trend features, and high-frequency fluctuation features under the i-th feature dimension; X(i) represents the time-domain feature value under the i-th feature dimension. (i) represents the low-frequency trend feature under the i-th feature dimension; (i) represents the high-frequency fluctuation feature under the i-th feature dimension;

[0098] The formula for Softmax normalization is:

[0099]

[0100] Among them, c t (i), c l (i), c h (i) represents the feature contribution of time-domain features, low-frequency trend features, and high-frequency fluctuation features to the healing of anal fistula under the i-th feature dimension, which is learned by graph attention network (GAT); exp(·) is an exponential function.

[0101] The phased feature optimization unit, using the treatment node as an index, optimizes the time-frequency joint feature X. TF The network is split according to treatment stages, obtaining time-frequency feature subsets for each treatment stage. A DenseNet optimized network consisting of dense blocks and transition layers is constructed. Multi-scale dilated convolutional modules are embedded within the dense blocks to obtain the depth-optimized feature set O of the time-frequency feature subsets for each treatment stage. s ;

[0102] Specifically, following the treatment process for anal fistula patients, the treatment node set T = {T1, T2, ..., T...} is defined as follows: n The treatment is divided into S treatment phases; using the treatment node as an index, the time-frequency joint feature X is analyzed. TF The process involves a step-by-step splitting based on nodes, specifically: the joint time-frequency feature X. TF Each row in the table corresponds to all features of a treatment node, and each column corresponds to a feature dimension; X is extracted during the preoperative baseline stage. TF The features in the first row of XTF are used to form a subset of features for the preoperative baseline stage; the features in the second and third rows of XTF are extracted for the short-term postoperative repair stage to form a subset of features for the short-term repair stage; and the features in the fourth to nth rows of XTF are extracted for the long-term postoperative healing stage to form a subset of features for the long-term healing stage.

[0103] Furthermore, the treatment stages for each patient can be set to the same default method or set according to the patient's condition, and a patient stage recording unit can be added for doctors to record the patient's stage.

[0104] For example, the default method divides the treatment stages as follows: preoperative baseline stage (T1), postoperative first stage (T2~T3), and postoperative second stage (T4~Tn).

[0105] For example, treatment stages can be reasonably divided according to the patient's condition: Patient A's treatment stages are divided into the first postoperative stage (1 week after surgery), the second postoperative stage (2 weeks after surgery), the third postoperative stage (1 month after surgery), and the fourth postoperative stage (2 months after surgery).

[0106] Multi-scale dilated convolutional modules are embedded within dense blocks, and dilated convolutional kernels with different dilation rates (r=1, 2, 3) are used for multi-scale feature extraction to fully capture the healing characteristics and temporal changes of anal fistula lesions at different scales. Dense blocks enable the dense reuse and deep mining of multi-layer time-frequency features. A transition layer is used to reduce the dimensionality and normalize the features, removing redundant time-frequency components, and obtaining the depth-optimized feature set O corresponding to the treatment nodes in each treatment stage. s For example, denoted as O1=Z1, O2=Z 2~3 O3=Z 4∼n O s To integrate multi-scale, phased, and highly identifiable features of anal fistula healing depth.

[0107] The efficacy assessment unit is based on a deeply optimized feature set O s To conduct quantitative assessments of the treatment effects at each stage of the patient's treatment.

[0108] Specifically, extract the deep optimization feature O1=[z] corresponding to the preoperative baseline stage. 11 ,z 12 ,...,z 1d As the initial lesion benchmark before anal fistula treatment, positive and negative polarities are defined for each dimension of depth optimization features according to the clinical requirements of anal fistula. Negative polarity means the smaller the better: such as fistula volume, inflammatory area, clinical symptom score, etc.; positive polarity means the larger the better: such as wound healing rate, etc.

[0109] For each treatment node n, calculate the standardized rate of change R of the features relative to the preoperative value by dimension. n,i Then, the dimensions are weighted and fused to obtain the single-node comprehensive healing score I for the nth treatment node. n ;

[0110] (n>1)

[0111] Among them, z 1,i Let z represent the i-th eigenvalue of Z1. n,i R represents the i-th eigenvalue of the depth optimization feature of the nth treatment node; ni >0: Feature improvement; R ni ≈0: Characteristic stability; R ni <0: Characteristic deterioration;

[0112] I n =

[0113] Among them, w i The clinical importance weights of the i-th feature are learned by a modal attention network; nA higher value indicates better healing.

[0114] At this point, the preoperative single-node comprehensive healing score I1, the short-term single-node comprehensive healing scores I2 and I3, and the long-term single-node comprehensive healing scores I4, ... I are obtained. n ;

[0115] Since the number of treatment nodes varies in each treatment phase, the comprehensive healing score for a single node in a given treatment phase is set as I. n1 ~I n2 The time-weighted average was used to calculate the comprehensive healing score E for each stage. S :

[0116]

[0117]

[0118] Among them, a n This represents the temporal weight of the nth treatment node within this stage.

[0119] Example 2: Building upon Example 1, to further achieve a fair and quantitative comparison of treatment effects among patients with different disease severity, the efficacy evaluation unit also pre-constructs standardized disease baselines:

[0120] Based on statistical results from a large number of clinical data on anal fistula patients, standardized mild depth optimization features O were set. 轻 ∈R 1×d =[z 轻1 , z 轻2 , ...z 轻d Standardized medium-depth optimization feature O 中 ∈R 1×d =[z 中1 , z 中2 , ...z 中d Standardized, heavily optimized features O 重 ∈R 1×d =[z 重1 , z 重2 , ...z 重d [This serves as a unified benchmark for evaluating the efficacy of treatment for different stages of anal fistula; if a patient's clinical symptom score for anal fistula is 0-4, it is considered mild; 5-8, it is considered moderate; and 9-12, it is considered severe.]

[0121] During the calculation, based on whether the patient's clinical symptom score for anal fistula is classified as mild, moderate, or severe, the depth optimization feature O1 corresponding to the patient's preoperative baseline stage is replaced with O. 轻 ∈R 1×d =[z 轻1 , z 轻2 , ...z轻d ]、O 中 ∈R 1×d =[z 中1 , z 中2 , ...z 中d [Or O] 重 ∈R 1×d =[z 重1 , z 重2 , ...z 重d As a standardized benchmark for disease progression;

[0122] Following the method described in Example 1, for each treatment node n, the standardized rate of change R of the features relative to the standardized disease baseline is calculated by dimension. * n,i Then, the dimensions are weighted and fused to obtain the single-node comprehensive healing score I for the nth treatment node. * n Then, a time-weighted average was used to calculate the comprehensive healing score E for each stage. * S .

[0123] Example 3: Please refer to the accompanying drawings in the instruction manual. Figure 2 An artificial intelligence-based intelligent analysis method for anal fistula includes the following steps:

[0124] I. Collect multi-source diagnostic data from patients with anal fistulas, including imaging data, pathological data, and medical records;

[0125] Second, preprocessing of multi-source diagnostic data includes data anonymization, data cleaning, and time-series alignment; image data processing includes format conversion, noise reduction, grayscale normalization, and standardization of image resolution and size; pathological data processing includes noise reduction, color correction, distortion repair, and annotation standardization. Structured fusion of medical document data generates structured medical data.

[0126] Third, the imaging data and pathological data are segmented separately. The imaging data focuses on segmenting the fistula, lesion and surrounding sphincter structure, while the pathological data focuses on segmenting the inflammatory tissue, granulation tissue, fibrotic area and normal perianal glands, to obtain the imaging lesion segmentation mask and pathological tissue segmentation mask under the corresponding treatment node.

[0127] Fourth, a sliding window local feature enhancement algorithm is adopted to enhance the local lesion features of the imaging and pathological lesion feature maps, highlighting the key feature information of the anal fistula lesion.

[0128] V. Based on the enhanced image data and corresponding image lesion segmentation masks, extract the image temporal features V1(T) at each treatment node; based on the preprocessed pathological data and pathological tissue segmentation masks, extract the pathological temporal features V2(T) at each treatment node; extract the disease temporal features V3(T) at each treatment node based on the structured disease data; where T={T1,T2,...,T...} n}, T n For the nth treatment node (e.g., preoperative, 1 week postoperative, 2 weeks postoperative, 1 month postoperative, etc.), by associating the treatment time nodes of V1(T), V2(T), and V3(T), ensure that the three types of features—image, pathology, and disease condition—are aligned on the same temporal dimension and spliced ​​together in chronological order to form a temporal feature matrix.

[0129] VI. The time-frequency joint processing unit is used to normalize the time-series feature matrix, remove outliers, and use weighted wavelet transform (WWZ) to decompose the treatment time series with non-uniform time intervals into low-frequency trend components and high-frequency fluctuation components. Finally, the original time-domain features, low-frequency trend features and high-frequency fluctuation features are adaptively fused through dynamic weights.

[0130] VII. Using the treatment node as an index, analyze the time-frequency joint feature X. TF The network is split according to treatment stages, obtaining time-frequency feature subsets for each treatment stage. A DenseNet optimized network consisting of dense blocks and transition layers is constructed. Multi-scale dilated convolutional modules are embedded within the dense blocks to obtain the depth-optimized feature set O of the time-frequency feature subsets for each treatment stage. s .

[0131] 8. Based on deep optimization of feature set O s To conduct quantitative assessments of the treatment effects at each stage of the patient's treatment.

[0132] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An artificial intelligence-based intelligent analysis system for anal fistulas, characterized in that, include: The multi-source data collection unit collects multi-source diagnostic data from patients with anal fistula, including imaging data, pathological data, and medical records. The data preprocessing unit is used to preprocess multi-source diagnostic data; The segmentation processing module is used to segment image data and pathological data separately to obtain the image lesion segmentation mask and pathological tissue segmentation mask under the corresponding treatment node. The local lesion feature enhancement unit uses a sliding window local feature enhancement algorithm to perform targeted enhancement of local lesion features on imaging and pathological lesion feature maps; The multimodal feature extraction unit extracts the image temporal features V1(T) at each treatment node based on the enhanced image data and the corresponding image lesion segmentation mask; extracts the pathological temporal features V2(T) at each treatment node based on the preprocessed pathological data and pathological tissue segmentation mask; and extracts the disease temporal features V3(T) at each treatment node based on the disease structured data; where T={T1,T2,...,T...} n }, T n For the nth treatment node; associate the treatment time nodes of V1(T), V2(T), and V3(T) and concatenate them in chronological order to form a temporal feature matrix; The time-frequency joint processing unit is used to normalize the time-series feature matrix, remove outliers, and use weighted wavelet transform to perform time-frequency decomposition on the treatment time sequence with non-uniform time intervals, splitting it into low-frequency trend components and high-frequency fluctuation components. Finally, the original time-domain features, low-frequency trend features and high-frequency fluctuation features are adaptively fused through dynamic weights. The phased feature optimization unit, using the treatment node as an index, optimizes the time-frequency joint feature X. TF The network is split according to treatment stages, obtaining time-frequency feature subsets for each treatment stage. A DenseNet optimized network consisting of dense blocks and transition layers is constructed, and multi-scale dilated convolutional modules are embedded inside the dense blocks to obtain the depth-optimized feature set O of the time-frequency feature subsets for each treatment stage. s ; The efficacy assessment unit is based on a deeply optimized feature set O s To conduct quantitative assessments of the treatment effects at each stage of the patient's treatment.

2. The artificial intelligence-based intelligent analysis system for anal fistulas according to claim 1, characterized in that, Imaging data includes: enhanced perianal MRI images, endorectal ultrasound images, and endoscopic time-series images acquired at different treatment stages; pathological data includes: pathological sections of continuous anal fistula tracts and pathological annotation data of inflammatory tissues acquired at different treatment stages; medical record data includes: electronic medical record texts and clinical symptom scores of anal fistulas corresponding to different treatment stages.

3. The artificial intelligence-based intelligent analysis system for anal fistulas according to claim 2, characterized in that, The clinical symptom scores of anal fistula at each treatment node were obtained based on the electronic medical record text of the corresponding treatment node and the clinical symptom score table for anal fistula.

4. The artificial intelligence-based intelligent analysis system for anal fistulas according to claim 3, characterized in that, A sliding window is used to perform a region-by-region sliding scan of the input image and pathological lesion feature map, extracting local lesion features within each window; the features within the window are mapped through a 1×1 convolution operation to complete the initial feature enhancement; the initially enhanced local lesion features are input into the convolutional attention fusion module, and the key lesion feature channels are weighted by channel attention and the core area of ​​the lesion is focused by spatial attention to further highlight key information.

5. The artificial intelligence-based intelligent analysis system for anal fistulas according to claim 4, characterized in that, The specific steps of the time-frequency joint processing unit are as follows: (1) Outliers were removed using the 3σ criterion, and the time series features were normalized using the Min-Max normalization method to obtain the normalized time series feature matrix X∈R. n×d , where n is the number of treatment nodes, and d is the total dimension of the concatenated three types of temporal features V1(T), V2(T), and V3(T); (2) For the normalized time series feature matrix X∈R n×d A weighted wavelet transform is performed column by column. Each column of the matrix corresponds to a one-dimensional time sequence of a single feature across n non-uniform treatment nodes. The actual time interval between each node is used as the timestamp input to WWZ to transform the one-dimensional time-domain evolution feature into a two-dimensional time-frequency amplitude distribution. Amplitude extraction and dimension alignment are performed on the time-frequency results of all features to construct a time-frequency amplitude matrix |F|∈R. e×n×d e represents the number of frequency points; (3) The improved Otsu algorithm is used to determine the high-low frequency splitting threshold f0, and the time-frequency amplitude matrix |F| is split into a low-frequency time-frequency matrix F according to f0. L and high-frequency time-frequency matrix F H and respectively for F L F H By performing mean aggregation along the frequency dimension, low-frequency trend features are obtained. ∈R n×d High-frequency fluctuation characteristics ∈R n×d ; It enables the separation of slow-changing long-term evolution information into low-frequency trend components and fast-changing short-term fluctuation information into high-frequency fluctuation components. (4) Time-frequency joint feature fusion: Based on dynamic weight allocation, adaptive weighted fusion is used to deeply fuse time-domain features, low-frequency trend features and high-frequency fluctuation features.

6. The artificial intelligence-based intelligent analysis system for anal fistulas according to claim 5, characterized in that, In step (3), the average amplitude of each frequency component along the frequency dimension of the time-frequency amplitude matrix |F| is calculated at all treatment nodes and feature dimensions to obtain a single-frequency amplitude sequence; For each frequency k, calculate its arithmetic mean over all treatment nodes i and all feature dimensions j: A(k)= ; The effective frequency range is obtained by removing the DC component and the extreme high-frequency noise component. Each frequency in the range is used as a candidate threshold to divide the single-frequency amplitude sequence into a low-frequency amplitude subset and a high-frequency amplitude subset. The inter-class variance of the two amplitude subsets under each candidate threshold is calculated, and the frequency that maximizes the inter-class variance is selected as the optimal splitting threshold f0.

7. The artificial intelligence-based intelligent analysis system for anal fistulas according to claim 6, characterized in that, Step (4) specifically involves: for each feature dimension, constructing a feature association graph structure with time-domain, low-frequency, and high-frequency features as nodes; learning the degree of association between the three types of features (time-domain, low-frequency, and high-frequency) under this dimension through a graph attention network; automatically calculating the feature contribution of each type of feature to the healing of anal fistula; normalizing the feature contribution using the Softmax function to obtain dynamic adaptive weights; and performing dimension-wise weighted fusion of the three types of features to obtain a time-frequency joint feature X that combines the original temporal changes, long-term healing trends, and short-term local reactions. TF The specific formula is as follows: X TF (i)=ω t (i)·X(i)+ω l (i)· (i)+ω h (i)· (i) (i=1,2,...,d) Where, ω t (i), ω l (i), ω h (i) represents the dynamic adaptive weights of time-domain features, low-frequency trend features, and high-frequency fluctuation features under the i-th feature dimension; X(i) represents the time-domain feature value under the i-th feature dimension. (i) represents the low-frequency trend feature under the i-th feature dimension; (i) represents the high-frequency fluctuation feature under the i-th feature dimension; The formula for Softmax normalization is: ； Among them, c t (i), c l (i), c h (i) represents the feature contribution of time-domain features, low-frequency trend features, and high-frequency fluctuation features to the healing of anal fistula under the i-th feature dimension, which is learned by graph attention network; exp(·) is an exponential function.

8. The artificial intelligence-based intelligent analysis system for anal fistulas according to claim 7, characterized in that, The specific steps of the phased feature optimization unit are as follows: The treatment node set T = {T1, T2, ..., T...} is used to optimize the feature unit. n The treatment is divided into S treatment phases; using the treatment node as an index, the time-frequency joint feature X is analyzed. TF The process involves segmenting the tissue into stages based on nodes to form a feature subset representing the long-term healing stage. Multi-scale dilated convolutional modules are embedded within dense blocks, and dilated convolutional kernels with different dilation rates are used for multi-scale feature extraction to fully capture the healing characteristics and temporal changes of anal fistula lesions at different scales. Dense blocks enable the intensive reuse and deep mining of multi-layer time-frequency features. A transition layer is used to reduce the dimensionality and normalize the features, removing redundant time-frequency components to obtain a depth-optimized feature set O corresponding to the treatment nodes in each treatment stage. s .

9. The artificial intelligence-based intelligent analysis system for anal fistulas according to claim 8, characterized in that, The specific steps of the efficacy assessment unit are as follows: Extracting the deep optimization feature O1=[z] corresponding to the preoperative baseline stage 11 ,z 12 ,...,z 1d As the initial lesion benchmark before anal fistula treatment, the positive and negative polarities of each dimension of the depth optimization feature are defined; For each treatment node n, calculate the standardized rate of change R of the features relative to the preoperative value by dimension. n,i Then, the dimensions are weighted and fused to obtain the single-node comprehensive healing score I for the nth treatment node. n ; （n＞1） Among them, z 1,i Let z represent the i-th eigenvalue of Z1. n,i R represents the i-th eigenvalue of the depth optimization feature of the nth treatment node; ni >0: Feature improvement; R ni =0: Characteristic stability; R ni <0: Characteristic deterioration; I n = ； Among them, w i Represents the clinical importance weight of the i-th feature; Let I be the comprehensive healing score for a single node in a certain treatment phase. n1 ~I n2 The time-weighted average was used to calculate the comprehensive healing score E for each stage. S : ； ； Among them, a n This represents the temporal weight of the nth treatment node within this stage.

10. An artificial intelligence-based intelligent analysis method for anal fistula, utilizing the artificial intelligence-based intelligent analysis system for anal fistula as described in claim 9, characterized in that, Includes the following steps: I. Collect multi-source diagnostic data from patients with anal fistulas, including imaging data, pathological data, and medical records; 2. Perform data anonymization, data cleaning, and time-series alignment preprocessing on imaging and pathological data; The medical records are desensitized, cleaned, preprocessed with time-series alignment, and fused with structured data to generate structured medical records.

3. The imaging data and pathological data are segmented separately to obtain the imaging lesion segmentation mask and pathological tissue segmentation mask under the corresponding treatment node. Fourth, a sliding window local feature enhancement algorithm is used to enhance the local lesion features of the imaging and pathological lesion feature maps, highlighting the key feature information of the anal fistula lesion; V. Based on the enhanced image data and corresponding image lesion segmentation masks, extract the image temporal features V1(T) at each treatment node; based on the preprocessed pathological data and pathological tissue segmentation masks, extract the pathological temporal features V2(T) at each treatment node; extract the disease temporal features V3(T) at each treatment node based on the structured disease data; where T={T1,T2,...,T...} n }, T n For the nth treatment node; associate the treatment time nodes of V1(T), V2(T), and V3(T), and concatenate them in chronological order to form a temporal feature matrix; VI. The time-frequency joint processing unit is used to normalize the time-series feature matrix, remove outliers, and use weighted wavelet transform to perform time-frequency decomposition on the treatment time sequence with non-uniform time intervals, splitting it into low-frequency trend components and high-frequency fluctuation components. Finally, the original time-domain features, low-frequency trend features and high-frequency fluctuation features are adaptively fused through dynamic weights. VII. Using the treatment node as an index, analyze the time-frequency joint feature X. TF The network is split according to treatment stages, obtaining time-frequency feature subsets for each treatment stage. A DenseNet optimized network consisting of dense blocks and transition layers is constructed, and multi-scale dilated convolutional modules are embedded inside the dense blocks to obtain the depth-optimized feature set O of the time-frequency feature subsets for each treatment stage. s ; 8. Based on deep optimization of feature set O s To conduct quantitative assessments of the treatment effects at each stage of the patient's treatment.

Images