Fracture image intelligent typing and treatment strategy recommendation system

Abstract

The invention relates to the technical field of medical image processing and intelligent decision support, and provides a fracture image intelligent typing and treatment strategy recommendation system, which comprises an image acquisition and preprocessing module, a topological feature analysis module, a similarity calculation module, a treatment strategy recommendation module and a case library management module, the topological feature analysis module captures essential features of a fracture form through a multi-scale topological feature representation unit and a persistent homology feature extraction unit; the similarity calculation module is used for accurately matching historical similar cases based on a topology-clinical mixed domain self-adaptive similarity calculation mechanism; the treatment strategy recommendation module is combined with similar case analysis and clinical guidelines to generate personalized treatment suggestions, the topology theory is introduced to analyze fracture structure features, invariance of image rotation, zooming and brightness change is achieved, and a real fracture structure and imaging noise can be distinguished; the topological features and the clinical features are fused through an adaptive weight mechanism, and the case matching precision is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image processing and intelligent decision support, in particular to a fracture image intelligent typing and treatment strategy recommendation system, especially a fracture image feature extraction and similarity calculation technology based on topology theory, which is used to assist doctors in accurate fracture typing and personalized treatment strategy making. BACKGROUND

[0002] Fracture is a common clinical disease, and accurate fracture typing is crucial for treatment planning. Currently, clinicians mainly rely on X-ray, CT and other imaging examination methods combined with professional experience to make fracture typing and treatment decisions. However, this method has strong subjectivity, large differences in doctor's experience, and insufficient consistency in judgment results, especially for complex fracture treatment, which is more likely to lead to diagnosis and treatment errors.

[0003] Existing computer-aided fracture diagnosis technology is mainly based on image processing and machine learning methods. These methods mostly rely on pixel-level features or deep learning black box features, although they have achieved certain results in the analysis of certain types of fractures, but there are still obvious deficiencies: first, they are sensitive to image quality and imaging conditions, and are easily affected by noise, contrast changes and other factors; second, they lack the ability to capture the essential features of fracture structures, making it difficult to accurately express the morphological features of fractures; third, they usually only focus on image features, ignoring the influence of patient clinical features on treatment decisions; fourth, most systems only provide typing results, lacking personalized recommendation ability for treatment suggestions.

[0004] Therefore, an intelligent system that can accurately capture the essential structural features of fractures, integrate clinical information, and provide personalized treatment suggestions is urgently needed to improve the accuracy and consistency of fracture diagnosis and treatment. SUMMARY

[0005] The purpose of the present application is to provide a fracture image intelligent typing and treatment strategy recommendation system, which realizes multi-scale topological feature extraction and analysis of fracture images by introducing topology theory, and accurately matches similar cases by combining patient clinical features, to provide personalized treatment strategy suggestions for doctors.

[0006] The present application provides a fracture image intelligent typing and treatment strategy recommendation system, which comprises: An image acquisition and preprocessing module for acquiring fracture medical images and performing standardized processing to generate preprocessed fracture region images; A topological feature analysis module connected to the image acquisition and preprocessing module for receiving the preprocessed fracture region images, constructing a multi-scale topological space representation, and extracting topological features of the fracture, the topological feature analysis module comprising a multi-scale topological feature representation unit and a persistent homology feature extraction unit; The similarity calculation module, connected to the topology feature analysis module, is used to receive the topology features output by the topology feature analysis module, obtain the patient's clinical features, and calculate the comprehensive similarity between the current fracture case and historical cases in the case database based on the topology features and clinical features. The treatment strategy recommendation module, connected to the similarity calculation module, is used to receive a list of similar cases output by the similarity calculation module and generate personalized treatment strategy suggestions based on the treatment plans and clinical guidelines of the similar cases.

[0007] Preferably, the multi-scale topological feature representation unit is used for: A distance transformation is performed on the preprocessed fracture region image to generate a distance field; Define a set of distance thresholds, generate a superset for each threshold and construct a simplicial complex to form a topological space representation; Calculate the topological invariants at each scale, including the number of connected components, loops, and holes; Generate multi-scale topological feature vectors.

[0008] Preferably, the persistent homology feature extraction unit is used for: A filtration sequence was constructed based on the density function of the fracture region; Track the birth and death values ​​of topological features in the filter sequence; Generate a persistent graph and extract persistent features, including persistent entropy, average lifetime, and persistent spectrum; Form a persistent feature vector.

[0009] Preferably, the similarity calculation module includes: The topological domain similarity calculation unit is used to calculate the topological feature similarity between the current case and historical cases; The clinical domain similarity calculation unit is used to calculate the clinical feature similarity between the current case and historical cases; An adaptive weight adjustment unit is used to dynamically adjust the weights of topological domain similarity and clinical domain similarity based on fracture type and patient characteristics; The comprehensive similarity calculation unit is used to calculate the final comprehensive similarity based on a weighted fusion method.

[0010] Preferably, the topological domain similarity calculation unit is used for: For two persistent graphs, calculate the Wasserstein distance; Calculate the weighted Euclidean distance for multi-scale topological feature vectors; The topological domain similarity is generated by fusing the Wasserstein distance and the weighted Euclidean distance.

[0011] Preferably, the clinical domain similarity calculation unit is used for: Construct a clinical feature vector that includes age coefficient, gender marker, fracture type code, and comorbidity status vector; For numerical features, calculate the normalized Euclidean distance; For categorical features, calculate the matching coefficient; For multi-state features, calculate the Jaccard similarity coefficient; By fusing the similarity of various features, a clinical domain similarity is generated.

[0012] Preferably, the adaptive weight adjustment unit is used for: Initial weights are set based on fracture type and severity; The weights are dynamically adjusted based on feature distribution and discriminative power. Construct a weighted decision tree and automatically select the optimal weights based on fracture characteristics; Based on historical matching accuracy, the weight parameters are optimized through feedback learning.

[0013] Preferably, the treatment strategy recommendation module includes: The case study analysis unit is used to analyze the treatment methods and effects of similar cases; Clinical guideline rules unit, used to integrate standard treatment guidelines; The treatment plan generation unit is used to generate treatment suggestions by integrating information from multiple sources; The scheme evaluation and optimization unit is used to evaluate the feasibility and expected effects of the scheme.

[0014] Preferably, the image acquisition and preprocessing module includes: Image acquisition interface, used to support the DICOM standard, to connect to the hospital's PACS system; Image preprocessing engine, used to perform operations such as noise reduction and contrast enhancement; Fracture area segmentation unit, used for automatic or semi-automatic segmentation of fracture areas; A clinical data integrator used to link patient clinical information and imaging data.

[0015] Preferably, the system further includes a case library management module, connected to the similarity calculation module and the treatment strategy recommendation module, for: Store historical cases containing fracture topological features, clinical characteristics, treatment plans, and prognostic information; The case database is updated based on new case data; Build feature indexes to support efficient retrieval; Conduct quality assessments and maintain the case library.

[0016] The present invention has the following beneficial effects: 1. By introducing topological theory, especially multi-scale topological feature representation and persistent homology analysis techniques, the essential structural features of fractures can be captured, unaffected by image rotation, scaling, and brightness changes, thus significantly improving the accuracy and robustness of fracture classification; 2. A topology-clinical hybrid domain adaptive similarity calculation mechanism was designed, which organically integrates topological features and clinical features, achieving more accurate case matching and providing strong support for personalized treatment; 3. By analyzing the treatment plans and prognoses of similar historical cases, and combining them with clinical guidelines, personalized treatment strategy suggestions can be generated for current patients, helping doctors to develop more scientific treatment plans; 4. The system has good interpretability, allowing doctors to understand the reasons for matching similar cases and the basis for treatment recommendations, thus enhancing the credibility and transparency of decision-making; 5. It can still work effectively under small sample conditions, and is applicable to the classification and treatment of various common and rare fractures, with broad clinical application value. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall architecture of the intelligent fracture image classification and treatment strategy recommendation system of the present invention. Figure 2 This is a schematic diagram of the workflow of the topology feature analysis module of the present invention; Figure 3 This is a schematic diagram illustrating the working principle of the multi-scale topological feature representation unit of the present invention; Figure 4 This is a schematic diagram illustrating the working principle of the persistent homology feature extraction unit of the present invention; Figure 5 This is a schematic diagram of the workflow of the similarity calculation module of the present invention; Figure 6 This is a schematic diagram illustrating the workflow of the treatment strategy recommendation module of the present invention. Detailed Implementation

[0018] Please refer to Figure 1 - Figure 6 The present invention will be further described in detail below with reference to the accompanying drawings and embodiments, but the implementation of the present invention is not limited thereto.

[0019] like Figure 1 As shown, the intelligent fracture image classification and treatment strategy recommendation system provided by the present invention includes an image acquisition and preprocessing module 1, a topological feature analysis module 2, a similarity calculation module 3, a treatment strategy recommendation module 4, and a case library management module 5.

[0020] Image acquisition and preprocessing module 1 acquires fracture medical images and performs standardized processing to generate preprocessed fracture region images. Topological feature analysis module 2, connected to image acquisition and preprocessing module 1, receives the preprocessed fracture region images, constructs a multi-scale topological spatial representation, and extracts the topological features of the fracture. Similarity calculation module 3, connected to topological feature analysis module 2, receives the topological features output by topological feature analysis module 2, acquires the patient's clinical characteristics, and calculates the comprehensive similarity between the current fracture case and historical cases in the case library based on the topological and clinical characteristics. Treatment strategy recommendation module 4, connected to similarity calculation module 3, receives the list of similar cases output by similarity calculation module 3 and generates personalized treatment strategy suggestions based on treatment plans and clinical guidelines for similar cases. Case library management module 5, connected to similarity calculation module 3 and treatment strategy recommendation module 4, stores and manages historical case data.

[0021] The image acquisition and preprocessing module 1 includes an image acquisition interface 11, an image preprocessing engine 12, a fracture region segmentation unit 13, and a clinical data integrator 14.

[0022] The image acquisition interface 11 supports the DICOM standard and can be connected to a hospital PACS system to acquire fracture images. In a preferred embodiment of the invention, the image acquisition interface 11 supports various common medical image formats such as X-ray, CT, and MRI, ensuring the system has broad applicability.

[0023] The image preprocessing engine 12 performs operations such as denoising and contrast enhancement to improve image quality. Specifically, the image preprocessing engine 12 first performs Gaussian filtering on the input image for denoising. The filter kernel size is usually set to 3×3 or 5×5, and the σ value is adaptively adjusted according to the noise level, usually between 0.5 and 2.0. Then, contrast enhancement is performed using the adaptive histogram equalization (CLAHE) method, with the contrast threshold usually set to 3.0 to 5.0 to improve the visibility of fracture lines. Finally, the image is normalized, and the pixel values ​​are normalized to the range of [0,1] for easier subsequent processing.

[0024] The fracture region segmentation unit 13 is used for automatic or semi-automatic segmentation of the fracture region. In one embodiment of the invention, the fracture region segmentation is achieved using a threshold-based adaptive segmentation method combined with morphological processing. Specifically, firstly, the initial threshold is determined using the Otsu algorithm to binarize the image; then, opening and closing operations are applied to remove noise and fill small holes, with the structuring element typically being a disk with a radius of 2-3 pixels; finally, the fracture region is extracted through contour detection. In cases of complex fractures, the system also supports interactive correction by doctors, allowing them to adjust the segmentation results by clicking or drawing.

[0025] The clinical data integrator 14 is used to link patient clinical information and imaging data. The clinical data integrator 14 connects to the hospital information system (HIS) and electronic medical record system (EMR) through a standardized interface to obtain basic patient information (such as age, gender), past medical history, comorbidities and other clinical data, and links this data with fracture images to form a complete patient information package.

[0026] like Figure 2 As shown, the topology feature analysis module 2 includes a multi-scale topology feature representation unit 21 and a persistent homology feature extraction unit 22, which are used to analyze fracture structure features from a topological perspective.

[0027] The multi-scale topological feature representation unit 21 is used to perform distance transformation on the preprocessed fracture region image to generate a distance field; define a set of distance thresholds, generate a superset under each threshold and construct a simple complex to form a topological space representation; calculate the topological invariants at each scale, including the number of connected components, the number of loops and the number of holes; and finally generate a multi-scale topological feature vector.

[0028] Specifically, such as Figure 3 As shown, a distance transformation is first performed on the segmented fracture region image to obtain the distance field D(x,y), which represents the distance from each point in the image to the fracture boundary: , in, Point Euclidean distance to the nearest boundary point; Represents the pixel coordinates in the image; , () represents the coordinates of a point on the boundary; B represents the set of fracture boundary points; min represents the minimum value operation, i.e., finding the distance point. The nearest boundary point.

[0029] Then define the distance threshold set. In a preferred embodiment of the present invention, The threshold is typically set to 8-12, covering multiple scales from fine structure to overall morphology. The threshold selection uses a logarithmic scale to ensure that the features cover structural information at different granularities. For example, for a fracture image with a resolution of 512×512, a typical threshold set could be T= Pixel.

[0030] At each threshold Next, generate a superset. = (x,y) D(x,y) Then based on Constructing simple complexes For two-dimensional fracture images, construct 0-simulacra (points), 1-simulacra (edges), and 2-simulacra (triangles). Specifically: 0-simplex is a superset Pixels in; A 1-simplex is a connection between two adjacent 0-simplexes (pixels); A 2-simplex is a triangle formed by three adjacent 0-simplexes (pixels).

[0031] Adjacency is typically defined using 8-connectivity, which considers adjacent pixels in the horizontal, vertical, and diagonal directions.

[0032] Next, we calculate the topological invariants, or Betti numbers, at each scale. The Betti numbers are important topological invariants in topological spaces, and they represent: The number of connected components reflects the number of bone fragments. The number of loops reflects the closed-loop structure of the fracture fissure. : Number of cavities, reflecting the cavity structure in the fracture area.

[0033] The calculation of the Betti number is based on the boundary operator matrix of the simplicial complex. For complexes... Construct the boundary operator matrix (Mapping from 1-simplex to 0-simplex) and (Mapping from 2-simplex to 1-simplex), then: , , , in, This represents the zero-dimensional Betti number (number of connected components). This represents a one-dimensional Betti number (number of loops). This represents the two-dimensional Betti number (hole count). Represents the first-order boundary operator matrix; Let denote the second-order boundary operator matrix; ker denotes the kernel space of the matrix, i.e., satisfying . The set of all vectors x; Im represents the image space of the matrix, i.e., all vectors that can be represented as... The set of vectors; dim represents the dimension of the space.

[0034] Finally, the Betti numbers at each scale are used to construct a multi-scale topological feature vector TF: , Where TF represents the multi-scale topological feature vector; Let represent the k-dimensional Betti number at the i-th threshold, where k=0,1,2,i=1,2,…,n.

[0035] To improve the discriminative power of the features, the topological change rate between different scales was also calculated as an auxiliary feature: , in, This represents the rate of change of the k-th dimension Betti number between the i-th and i+1-th thresholds; This represents the k-dimensional Betti number at the i-th threshold. This represents the k-dimensional Betti number at the (i+1)th threshold. This represents the i-th threshold; This represents the (i+1)th threshold; k=0,1,2 represents the dimension of the Betti number; i=1,2,…,n-1 represents the index of the threshold.

[0036] Thus, the enhanced topological feature vector ETF includes the Betti number and the rate of change: , Where ETF represents the enhanced topological feature vector; TF represents the original multi-scale topological feature vector; This represents the rate of change of the k-th dimension Betti number between the i-th and i+1-th thresholds.

[0037] Multi-scale topological feature representation can capture the topological properties of fracture structures at different scales, and is invariant to image rotation, scaling and brightness changes, providing an essential feature representation of fracture morphology.

[0038] The persistent homology feature extraction unit 22 constructs a filter sequence based on the fracture region density function; tracks the birth and death values ​​of topological features in the filter sequence; generates a persistent graph and extracts persistent features, including persistent entropy, average lifetime and persistent spectrum; and forms a persistent feature vector.

[0039] like Figure 4 As shown, density estimation is first performed on the fracture area to generate a density function f. In one embodiment of the invention, the density function employs the kernel density estimation (KDE) method: , in, Point The density value at the location; n represents the total number of samples; h represents the bandwidth parameter (usually set to 3-5 pixels); m represents the number of pixels in the fracture region; Represents the coordinates of a point in an image; The coordinates of the i-th pixel in the fracture region are represented by K; K represents the kernel function, usually a Gaussian kernel is chosen.

[0040] Constructing an ascending filter sequence based on the density function: ,in t is the density threshold. In a preferred embodiment of the invention, the threshold is typically selected from 10 to 15 uniformly distributed values, covering the entire value range of the density function.

[0041] For each complex in the filter sequence, the homology group is calculated and the birth and death of homology classes are tracked. Specifically, when a topological feature (such as a connected component, loop, or cavity) first appears in the filter sequence, its birth value b is recorded; when the feature is subsequently filled or disappears, its death value d is recorded. Each topological feature corresponds to a birth-death pair (b, d).

[0042] Construct a persistence graph based on birth-death pairs. , where each point Represents the birth value of a topological feature. and death value ; m represents the number of points in the persistent graph. Simultaneously, the persistence p = db of each topological feature is calculated, reflecting the feature's stability.

[0043] To distinguish noise from true structural features, persistent images are filtered, retaining persistent points whose lifetime exceeds a threshold. In embodiments of the present invention, the threshold is typically set to 5% to 10% of the density function range, and the specific value can be adjusted according to the noise level of the fracture image.

[0044] Extract the following statistical features from the persistent graph: Persistence Entropy (PE): Measures the uncertainty of the distribution of persistent points. , in, This represents the persistence entropy of dimension k; k=0,1,2 represent persistent graphs of different dimensions. It is the number of persistent points in dimension k; It is the lifetime of the i-th persistence point, calculated as ; denoted as the sum of the lifetimes of all persistent points in dimension k; log represents the natural logarithm with base e, e≈2.718.

[0045] Mean lifetime (AL): reflects the average stability of topological characteristics. , in, Representing dimensions Average lifespan; It is a dimension The number of persistence points; It is the first The lifespan of a single durability point; =0,1,2 represent persistent graphs of different dimensions.

[0046] Persistent Spectrum (PS): The distribution characteristics of persistent point lifetimes, typically represented by dividing the lifetime values ​​into several intervals and calculating the number of points in each interval. In an embodiment of the present invention, the lifetime value range is divided into 10 equal intervals, resulting in a 10-dimensional spectrum vector.

[0047] Ultimately, the persistent feature vector PF contains persistence entropy, average lifetime, and persistence spectrum for each dimension: , in, Represents a persistent feature vector; PE k Represents the persistence entropy of dimension k; AL k This represents the average lifetime of dimension k; Let represent the value of the persistent spectrum of dimension k in the j-th interval, where k=0,1,2 represent different dimensions, and j=1,2,...,10 represent the 10 intervals of the spectrum.

[0048] Persistent homology analysis can distinguish between the true structure and imaging noise in fracture images, extract clinically significant fracture morphological features, and provide a more reliable basis for subsequent classification and similarity calculation.

[0049] like Figure 5 As shown, the similarity calculation module 3 includes a topological domain similarity calculation unit 31, a clinical domain similarity calculation unit 32, an adaptive weight adjustment unit 33, and a comprehensive similarity calculation unit 34.

[0050] The topological domain similarity calculation unit 31 calculates the Wasserstein distance for two persistent graphs; calculates the weighted Euclidean distance for multi-scale topological feature vectors; and fuses the Wasserstein distance and the weighted Euclidean distance to generate the topological domain similarity.

[0051] Specifically, for two persistent graphs and Calculate the p-th order Wasserstein distance: , Among them, W p Represents the p-th order Wasserstein distance; and This represents two persistent graphs to be compared; γ represents... arrive A matching function; x represents The point in the middle; γ(x) represents the point that matches x. The point in the middle; Point and The infinite norm distance between them is calculated as ; The order of the distance is usually 1 or 2; This represents taking the minimum value among all possible matches. During the calculation, the case of diagonal projection needs to be considered, which allows points in the persistent graph to match points on the diagonal. This corresponds to the situation where the birth and death of topological features occur almost simultaneously.

[0052] In a preferred embodiment of the invention, p is set to 2, i.e., 2-Wasserstein distance is used, which better reflects the distribution differences of points in the persistent graph. Specifically, the Hungarian algorithm is used to solve for the optimal matching, with a computational complexity of O(m+n). 3 ), where m and n are the number of points in the two persistent graphs, respectively. To improve computational efficiency, the persistent graphs are usually preprocessed to filter out points with very short lifespans (such as those with lifespans less than 5% of the total value range), thus reducing the computational load.

[0053] For multi-scale topological feature vectors and Calculate the weighted Euclidean distance: , in, Indicates the weighted Euclidean distance; and This represents two multi-scale topological feature vectors to be compared; and The eigenvector represents the first eigenvector. One component; It is the first The weight coefficients of each component; 3n represents the dimension of the feature vector, n is the number of thresholds, and each threshold has 3 Betti numbers ( , , Weighting coefficients are typically determined through feature importance analysis. In embodiments of this invention, a feature importance assessment method based on mutual information is employed. The weighting coefficients are proportional to the mutual information value between the feature and the fracture type, ensuring that important features have a greater impact on distance calculations.

[0054] Ultimately, topological domain similarity By fusing the Wasserstein distance and the weighted Euclidean distance, we obtain: ,in, This represents the topological similarity, with a value ranging from [0,1]. A larger value indicates a higher similarity. is the fusion weight, representing the importance of Wasserstein distance in the fusion, with a value range of [0,1], typically between 0.4 and 0.6; exp represents the natural exponential function, used to convert distance into similarity. express Wasserstein distance; This represents the weighted Euclidean distance. In a preferred embodiment of the invention, =0.5 indicates that the two distance metrics are equally important.

[0055] The clinical domain similarity calculation unit 32 constructs a clinical feature vector containing age coefficient, gender marker, fracture type code, and comorbidity status vector; calculates the normalized Euclidean distance for numerical features; calculates the matching coefficient for categorical features; calculates the Jaccard similarity coefficient for multi-state features; and fuses the similarities of each feature to generate the clinical domain similarity.

[0056] Specifically, firstly, a clinical feature vector CF = [age coefficient, gender marker, fracture type code, comorbidity status vector] is constructed, where: Age coefficient: The patient's actual age, usually normalized to the [0,1] interval; Gender indicator: a binary value, 0 for male and 1 for female; Fracture type code: The fracture type coded according to clinical classification, such as the code in the AO / OTA classification; Comorbidity status vector: A vector consisting of multiple binary values, indicating whether a patient has a specific comorbidity, such as diabetes or osteoporosis.

[0057] For numerical features such as age, calculate the normalized nonlinear similarity: , in, This represents age similarity, with a value ranging from [0,1]. A larger value indicates higher similarity. and These are the normalized age values ​​of the two patients; This is a scaling parameter that controls the rate at which similarity changes with age differences, typically set to 0.1-0.2; exp represents the natural exponential function. This nonlinear mapping reflects the nonlinear characteristics of the effect of age differences on fracture healing, i.e., patients of similar age have higher similarity; patients with large age differences experience a rapid decrease in similarity.

[0058] For categorization features such as gender and fracture type, calculate the matching coefficient: , in, Indicates the similarity of categorical features; and The similarity is 1 when the two feature values ​​are the same and 0 when the two feature values ​​are different.

[0059] For multi-state features such as comorbidities, the Jaccard similarity coefficient is calculated: , in, Indicates the similarity of comorbidities; and These are the comorbidity status vectors of two patients; This represents the number of comorbidities shared by both, i.e., the number of 1s after a bitwise AND operation on the vector; This represents the total number of comorbidities, which is the number of 1s after a bitwise OR operation on the vector.

[0060] Ultimately, clinical domain similarity The following is obtained by weighted fusion of the similarity of each feature: , in, This represents the clinical domain similarity, with a value ranging from [0,1]. A larger value indicates a higher similarity. , , and These are the weighting coefficients for age, gender, fracture type, and comorbidities, respectively, satisfying... ; , , and These are the similarities in age, sex, fracture type, and comorbidities. In a preferred embodiment of the invention, the weighting is as follows: , , , This reflects the high importance of fracture type and age in the assessment of clinical similarity.

[0061] The adaptive weight adjustment unit 33 sets initial weights based on fracture type and severity; dynamically adjusts weights according to feature distribution and discriminative power; constructs a weight adjustment decision tree to automatically select the optimal weights based on fracture characteristics; and optimizes weight parameters through feedback learning based on historical matching accuracy.

[0062] Specifically, initial weights are first set based on the fracture type and severity. : , in, This represents the initial weight, with a value range of [0,1]. The larger the value, the higher the importance of the topological features in the similarity calculation. The mapping function is represented, and weights are determined based on the fracture type and severity. In an embodiment of the invention, for structurally complex comminuted fractures, the initial weights are... Setting the initial weights higher (e.g., 0.7-0.8) increases the importance of topological features; for simple transverse fractures, the initial weights... Set it to a lower value (e.g., 0.4-0.5).

[0063] Then, based on feature distribution and discriminative power, the weights are dynamically adjusted: , in, This indicates the adjusted weights; Indicates the initial weights; This represents the variance of the topological similarity between the current case and historical cases. This represents the variance of the clinical domain similarity calculated between the current case and historical cases; a larger variance indicates higher feature discrimination. This adjustment mechanism can adaptively increase the weight of feature domains with high discrimination, thereby improving the accuracy of similarity calculation.

[0064] To more precisely control weight adjustment, a weight adjustment decision tree was constructed, automatically selecting the optimal weights based on fracture characteristics. The branch nodes of the decision tree include fracture type, patient age group, and comorbidity status, while leaf nodes correspond to different weight values. For example, for an elderly patient (>65 years old) with a comminuted fracture and concurrent osteoporosis, the optimal weights might be... =0.65; however, for simple fractures in young patients (30-40 years old) without significant complications, the optimal weight might be... =0.45.

[0065] Finally, based on historical matching accuracy, the weight parameters are optimized through a feedback learning mechanism. Specifically, the system records the correlation between the result of each similarity calculation and the final treatment effect, adjusting the weight parameters by maximizing the treatment success rate. The optimization objective function is: , in, This indicates the operation to find the maximum value; This represents the weight parameter to be optimized; N represents the number of historical cases. This represents the treatment plan for the i-th case; This indicates an indicator function that takes a value of 1 when the treatment is successful and 0 otherwise. This represents the overall similarity of the i-th case calculated based on the weight λ.

[0066] Through the above mechanism, the adaptive weight adjustment unit 33 can dynamically adjust the weights of topological features and clinical features in similarity calculation based on fracture type, patient characteristics and historical experience, so as to achieve more accurate case matching.

[0067] The comprehensive similarity calculation unit 34 calculates the final comprehensive similarity based on a weighted fusion method: , in, This represents the overall similarity score, with a value ranging from [0,1]. A larger value indicates a higher degree of similarity. The weights representing topological features are determined by the adaptive weight adjustment unit 33; Indicates topological similarity; Indicates clinical domain similarity.

[0068] In this way, the system can adaptively adjust the importance of topological and clinical features based on the specific circumstances of the fracture and the patient's characteristics, thereby achieving more accurate similarity calculations.

[0069] After the similarity calculation is completed, the system sorts all historical cases in descending order of similarity and selects the Top-N (usually N=5-10) similar cases as the basis for treatment strategy recommendations.

[0070] like Figure 6 As shown, the treatment strategy recommendation module 4 includes a case treatment analysis unit 41, a clinical guideline rule unit 42, a treatment plan generation unit 43, and a plan evaluation and optimization unit 44.

[0071] Case treatment analysis unit 41 analyzes the treatment methods and effects of similar cases. Specifically, for each similar case, its treatment plan information is extracted, including the treatment method (conservative treatment or surgical treatment), specific treatment methods (such as the specific methods of external fixation and internal fixation), materials used, drug treatment plan, etc., as well as the prognosis after treatment, including fracture healing time, functional recovery, complications, etc.

[0072] In a preferred embodiment of the present invention, the case treatment analysis employs a weighted voting mechanism, meaning that cases with higher similarity receive a greater weight in their treatment plan. For the i-th similar case, the weight of its treatment plan is: , in, Indicates the first The weight of each similar case; Indicates the first The overall similarity of 10 similar cases; N represents the number of similar cases selected. This represents the sum of the overall similarity scores of all similar cases.

[0073] Clinical guideline rule unit 42 integrates standard treatment guidelines, transforming treatment recommendations from orthopedic clinical guidelines into a computer-processable format. In embodiments of this invention, treatment guidelines are stored in a rule base, and the general form of the rules is as follows: IF <fracture type> AND <patient characteristics> THEN <treatment recommendations>;

[0074] For example, a typical rule might be: IF fracture type = femoral neck fracture AND age > 65 AND osteoporosis = THEN recommends total hip replacement; Clinical guideline rule unit 42 uses a rule matching engine to find matching treatment rules based on the current patient's fracture type and clinical characteristics, serving as a reference for generating treatment strategies.

[0075] The treatment plan generation unit 43 integrates multi-source information to generate treatment recommendations. Specifically, firstly, based on treatment plans of similar cases, preliminary treatment recommendations are generated through a weighted voting mechanism; then, the preliminary recommendations are adjusted and optimized in conjunction with the recommendation rules of clinical guidelines; finally, treatment parameters (such as the selection of fixation materials and postoperative rehabilitation plans) are personalized according to the patient's specific circumstances (such as age, comorbidities, etc.).

[0076] In embodiments of the present invention, treatment plan generation employs a hierarchical decision-making process: 1. First, decide on the treatment method: conservative treatment or surgical treatment; 2. For conservative treatment, further determine the specific methods (such as plaster cast immobilization, functional exercises, etc.); 3. For surgical treatment, further determine the surgical method (such as internal fixation, external fixation, joint replacement, etc.) and specific technical parameters; 4. Finally, a complete treatment plan is generated, including treatment methods, material selection, drug treatment, and rehabilitation plan.

[0077] The treatment plan evaluation and optimization unit 44 assesses the feasibility and expected outcomes of the treatment plan. Specifically, it evaluates the generated treatment plan, considering the following factors: 1. Clinical feasibility: Does the treatment plan conform to the actual clinical situation, and has it taken into account the hospital's technical conditions and equipment? 2. Patient suitability: Does the treatment plan take into account the patient's specific circumstances, such as age, physical condition, and economic situation? 3. Expected Outcomes: Based on the treatment outcomes of similar cases, predict the possible effects of the current treatment plan, including healing time, functional recovery, and risk of complications; 4. Cost-effectiveness: Consider the balance between treatment costs and expected results.

[0078] If the evaluation results show that there are problems with the plan, it is returned to the treatment plan generation unit 43 for adjustment and optimization until a satisfactory treatment plan is obtained.

[0079] Finally, the treatment strategy recommendation module 4 outputs complete treatment strategy suggestions, including: 1. Specific treatment plan and parameters; 2. Supporting evidence, including treatment outcomes of similar cases and recommendations from clinical guidelines; 3. Expected results and risk assessment; 4. Alternative options and their comparison.

[0080] This treatment strategy recommendation, based on similar case analysis and integration of clinical guidelines, can provide doctors with personalized, evidence-based treatment suggestions, assisting them in developing more scientific treatment plans.

[0081] The case library management module 5 is connected to the similarity calculation module 3 and the treatment strategy recommendation module 4. It is used to store historical cases containing fracture topological features, clinical features, treatment plans and prognostic information; update the case library according to new case data; build feature indexes to support efficient retrieval; and conduct quality assessment and maintenance of the case library.

[0082] Specifically, each case in the case library contains the following information: 1. Basic Information: Patient ID, age, gender, etc.; 2. Clinical features: fracture type, comorbidities, past medical history, etc.; 3. Imaging data: original fracture images and preprocessed images; 4. Topological features: multi-scale topological feature vectors and persistent feature vectors; 5. Treatment Information: Specific treatment plan, material selection, surgical details, etc.; 6. Prognostic information: fracture healing time, functional recovery, complications, etc.

[0083] To support efficient retrieval, the case library management module 5 has constructed a feature index, employing a multi-level index structure: 1. Primary Index: Quickly filter fractures of the same type based on fracture type; 2. Secondary index: based on clinical characteristics, such as age group, gender, etc.; 3. Three-level index: Approximate nearest neighbor (ANN) index based on topological features, supporting fast similarity calculation.

[0084] In a preferred embodiment of the present invention, the ANN index of topological features employs the Locality Sensitive Hash (LSH) method, which greatly improves the efficiency of similar case retrieval. For a case database containing 10,000 cases, the retrieval time is typically controlled within 100ms, meeting the requirements for real-time response.

[0085] The case library management module 5 is also responsible for updating and maintaining the case library. When a new case completes treatment and has a prognostic outcome, the system adds its characteristics and treatment results to the case library and updates the corresponding indexes. Simultaneously, the system regularly conducts quality assessments of the case library, removing duplicate or low-quality cases to ensure the quality and effectiveness of the case library.

[0086] Through the detailed description of the above embodiments, those skilled in the art can clearly understand the structure and working principle of the intelligent fracture image classification and treatment strategy recommendation system of the present invention. The present invention achieves accurate capture of the essential structural features of fractures by introducing topological theory, particularly multi-scale topological feature representation and persistent homology analysis technology; it achieves more accurate case matching through a topology-clinical hybrid domain adaptive similarity calculation mechanism; and it provides doctors with personalized and evidence-based treatment recommendations based on similar case analysis and clinical guideline integration. These innovations give the system significant advantages in fracture classification accuracy, personalized treatment recommendations, and clinical decision support capabilities, and it has broad clinical application prospects.

[0087] The embodiments described above are merely illustrative of specific implementations of the present invention, and while the descriptions are detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A fracture image intelligent classification and treatment strategy recommendation system, characterized in that, include: The image acquisition and preprocessing module is used to acquire fracture medical images and perform standardized processing to generate preprocessed fracture area images. The topological feature analysis module, connected to the image acquisition and preprocessing module, is used to receive the preprocessed fracture region image, construct a multi-scale topological space representation, and extract the topological features of the fracture. The topological feature analysis module includes a multi-scale topological feature representation unit and a persistent homology feature extraction unit. The similarity calculation module, connected to the topology feature analysis module, is used to receive the topology features output by the topology feature analysis module, obtain the patient's clinical features, and calculate the comprehensive similarity between the current fracture case and historical cases in the case database based on the topology features and clinical features. The treatment strategy recommendation module, connected to the similarity calculation module, is used to receive a list of similar cases output by the similarity calculation module and generate personalized treatment strategy suggestions based on the treatment plans and clinical guidelines of the similar cases.

2. The intelligent fracture image classification and treatment strategy recommendation system according to claim 1, characterized in that, The multi-scale topological feature representation unit is used for: A distance transformation is performed on the preprocessed fracture region image to generate a distance field; Define a set of distance thresholds, generate a superset for each threshold and construct a simplicial complex to form a topological space representation; Calculate the topological invariants at each scale, including the number of connected components, loops, and holes; Generate multi-scale topological feature vectors.

3. The intelligent fracture image classification and treatment strategy recommendation system according to claim 1, characterized in that, The persistent homology feature extraction unit is used for: A filtration sequence was constructed based on the density function of the fracture region; Track the birth and death values ​​of topological features in the filter sequence; Generate a persistent graph and extract persistent features, including persistent entropy, average lifetime, and persistent spectrum; Form a persistent feature vector.

4. The intelligent fracture image classification and treatment strategy recommendation system according to claim 1, characterized in that, The similarity calculation module includes: The topological domain similarity calculation unit is used to calculate the topological feature similarity between the current case and historical cases; The clinical domain similarity calculation unit is used to calculate the clinical feature similarity between the current case and historical cases; An adaptive weight adjustment unit is used to dynamically adjust the weights of topological domain similarity and clinical domain similarity based on fracture type and patient characteristics; The comprehensive similarity calculation unit is used to calculate the final comprehensive similarity based on a weighted fusion method.

5. The intelligent fracture image classification and treatment strategy recommendation system according to claim 4, characterized in that, The topological domain similarity calculation unit is used for: For two persistent graphs, calculate the Wasserstein distance; Calculate the weighted Euclidean distance for multi-scale topological feature vectors; The topological domain similarity is generated by fusing the Wasserstein distance and the weighted Euclidean distance.

6. The intelligent fracture image classification and treatment strategy recommendation system according to claim 4, characterized in that, The clinical domain similarity calculation unit is used for: Construct a clinical feature vector that includes age coefficient, gender marker, fracture type code, and comorbidity status vector; For numerical features, calculate the normalized Euclidean distance; For categorical features, calculate the matching coefficient; For multi-state features, calculate the Jaccard similarity coefficient; By fusing the similarity of various features, a clinical domain similarity is generated.

7. The intelligent fracture image classification and treatment strategy recommendation system according to claim 4, characterized in that, The adaptive weight adjustment unit is used for: Initial weights are set based on fracture type and severity; The weights are dynamically adjusted based on feature distribution and discriminative power. Construct a weighted decision tree and automatically select the optimal weights based on fracture characteristics; Based on historical matching accuracy, the weight parameters are optimized through feedback learning.

8. The intelligent fracture image classification and treatment strategy recommendation system according to claim 1, characterized in that, The treatment strategy recommendation module includes: The case study analysis unit is used to analyze the treatment methods and effects of similar cases; Clinical guideline rules unit, used to integrate standard treatment guidelines; The treatment plan generation unit is used to generate treatment suggestions by integrating information from multiple sources; The scheme evaluation and optimization unit is used to evaluate the feasibility and expected effects of the scheme.

9. The intelligent fracture image classification and treatment strategy recommendation system according to claim 1, characterized in that, The image acquisition and preprocessing module includes: Image acquisition interface, used to support the DICOM standard, to connect to the hospital's PACS system; Image preprocessing engine, used to perform noise reduction and contrast enhancement operations; Fracture area segmentation unit, used for automatic or semi-automatic segmentation of fracture areas; A clinical data integrator used to link patient clinical information and imaging data.

10. The intelligent fracture image classification and treatment strategy recommendation system according to claim 1, characterized in that, The system also includes a case library management module, which is connected to the similarity calculation module and the treatment strategy recommendation module, and is used to: store historical cases containing fracture topological features, clinical features, treatment plans and prognostic information; and update the case library based on new case data. Build a feature index to support efficient retrieval; conduct quality assessment and maintenance of the case library.

Images