A computer-implemented preoperative warning method for sacral tumors and related equipment

Abstract

The present invention provides a computer-implemented preoperative warning method for sacral tumors and related equipment, which are applied to the field of data processing technology. The present application processes a preset three-dimensional densely connected network model based on a target training sample set and a verification sample set to generate a target three-dimensional densely connected network model; processes a target user's preoperative sacral non-contrast computed tomography image to generate a three-dimensional region of interest image of the target user; processes the target user's clinical data information to generate the target user's physiological attribute parameters; and processes the target user's preoperative sacral non-contrast computed tomography image, the target user's three-dimensional region of interest image, the target user's dynamic risk parameters, and the target user's physiological attribute parameters based on the target three-dimensional densely connected network model to generate sacral warning information, wherein the sacral warning information is used to characterize the attribute information of the target user's preoperative sacral lesions.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a computer-implemented preoperative warning method for sacral tumors and related equipment. Background Art

[0002] Bone tumors pose a serious threat to life and health and are the third leading cause of death in cancer patients under the age of 20. They include four types of tumors: primary malignant, benign, metastatic, and caused by local invasion of visceral malignant tumors. They have different biological behaviors. Benign tumors are usually stable, while sacral tumors affect the stability of the lumbosacral joint. Accurate identification of bone tumors is of great significance for clinical decision-making. However, the early symptoms of sacral tumors lack specificity and are mostly manifested as neurological deficits and lower back pain.

[0003] Imaging is crucial for early warning of bone tumors. CT, with its high resolution, can detect lesions as small as 3 mm and provide crucial information, such as tumor matrix mineralization, to aid diagnosis. However, despite the low incidence of bone tumors, inexperienced physicians can easily misdiagnose, leading to unnecessary biopsies, patient discomfort, increased costs, delayed diagnosis, and even death. In this context, the development of sensitive and effective computer-assisted early warning tools is crucial.

[0004] It should be noted that the information disclosed in the above background technology section is only used to enhance the understanding of the background of the present disclosure, and therefore may include information that does not constitute prior art known to ordinary technicians in the field. Summary of the Invention

[0005] The purpose of this application is to provide a computer-implemented preoperative early warning method and related equipment for sacral tumors, which, at least to some extent, overcome the problems of existing technologies. By collecting various data from target users, including clinical, NCCT images, multimodal, and diagnostic interval data, and in terms of feature processing, features are extracted from the multimodal data to obtain dynamic risk parameters, and features are extracted from the clinical data to obtain physiological attribute parameters. After normalizing and standardizing these data and image data, a weighted fusion strategy (with learnable weights assigned to the image, dynamic risk parameter, and physiological attribute parameter channels) is used to generate multi-channel feature data. Ultimately, the multi-channel feature data is input into a model, and early warning information is generated based on preset dynamic thresholds, effectively assisting in the determination of tumor benign or malignant status and improving the accuracy of preoperative diagnosis.

[0006] Other features and advantages of the present application will become apparent from the following detailed description, or may be learned in part by practice of the invention.

[0007] According to one aspect of the present application, a preoperative early warning method for sacral tumors executed by a computer is provided, comprising: obtaining clinical data information of a target user, a preoperative non-contrast computed tomography image of the sacrum of the target user, multimodal data information of the target user, target diagnostic interval data information of the target user, a preset three-dimensional densely connected network model, a training sample set, and a verification sample set, wherein the multimodal data information of the target user includes T1-weighted imaging information of the target user, T2-weighted imaging information of the target user, and positron emission tomography-computed tomography information; preprocessing the training sample set to generate a target training sample set; processing the preset three-dimensional densely connected network model based on the target training sample set and the verification sample set to generate a target three-dimensional densely connected network model. a network model; processing the preoperative non-contrast computed tomography image of the sacrum of the target user to generate a three-dimensional region of interest image of the target user; processing the multimodal data information of the target user and the target diagnostic interval data information of the target user to generate the dynamic risk parameters of the target user; processing the clinical data information of the target user to generate the physiological attribute parameters of the target user; processing the preoperative non-contrast computed tomography image of the sacrum of the target user, the three-dimensional region of interest image of the target user, the dynamic risk parameters of the target user and the physiological attribute parameters of the target user based on the target three-dimensional densely connected network model to generate sacrum warning information, wherein the sacrum warning information is used to characterize the attribute information of the preoperative sacral lesions of the target user.

[0008] Another aspect of the present application is a preoperative warning device for sacral tumors, characterized in that it includes: an acquisition module for acquiring clinical data information of a target user, a preoperative non-contrast computed tomography image of the sacrum of the target user, multimodal data information of the target user, target diagnostic interval data information of the target user, a preset three-dimensional densely connected network model, a training sample set and a verification sample set, wherein the multimodal data information of the target user includes T1-weighted imaging information of the target user, T2-weighted imaging information of the target user and positron emission tomography-computed tomography information; a processing module for preprocessing the training sample set to generate a target training sample set; processing the preset three-dimensional densely connected network model based on the target training sample set and the verification sample set to generate a target A three-dimensional densely connected network model; processing the target user's preoperative sacrum non-contrast computed tomography image to generate a three-dimensional region of interest image of the target user; processing the target user's multimodal data information and the target diagnostic interval data information to generate the target user's dynamic risk parameters; processing the target user's clinical data information to generate the target user's physiological attribute parameters; processing the target user's preoperative sacrum non-contrast computed tomography image, the target user's three-dimensional region of interest image, the target user's dynamic risk parameters and the target user's physiological attribute parameters based on the target three-dimensional densely connected network model to generate sacrum warning information, wherein the sacrum warning information is used to characterize the attribute information of the target user's preoperative sacral lesions.

[0009] According to another aspect of the present application, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a second processor, the computer program implements the above-mentioned preoperative early warning method for sacral tumors.

[0010] The present application provides a computer-implemented preoperative warning method and related equipment for sacral tumors. The server collects various types of data from target users, including clinical, NCCT images, multimodal and diagnostic interval data, and carefully preprocesses the training samples to ensure data quality. When building the model, the three-dimensional densely connected network is optimized with the help of target training and verification samples, and the final model is determined by sampling feature training and verification. At the same time, the NCCT image is processed based on the anatomical structure restriction information to generate the region of interest image. In terms of feature processing, features are extracted from multimodal data to obtain dynamic risk parameters, and features are extracted from clinical data to obtain physiological attribute parameters. After these data and image data are normalized and standardized, a weighted fusion strategy is used to generate multi-channel feature data. Finally, the multi-channel feature data is input into the model, and warning information is generated based on the preset dynamic threshold, which effectively assists in determining the benign or malignant nature of the tumor and improves the accuracy of preoperative diagnosis.

[0011] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 A flowchart illustrating a computer-implemented preoperative early warning method for sacral tumors provided in one embodiment of the present application is shown;

[0013] Figure 2 A schematic structural diagram of a preoperative warning device for sacral tumors provided in one embodiment of the present application is shown. DETAILED DESCRIPTION

[0014] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0015] The following combination Figure 1 The following describes a preoperative early warning method for sacral tumors according to an exemplary embodiment of the present application. In one embodiment, the present application also provides a preoperative early warning method for sacral tumors executed by a computer and related equipment. Figure 1 The flowchart of a preoperative early warning method for sacral tumors executed by a computer according to an embodiment of the present application is schematically shown. Figure 1 As shown, the method is applied to the server and includes:

[0016] S101, obtaining clinical data information of a target user, a preoperative non-contrast computed tomography image of the sacrum of the target user, multimodal data information of the target user, target diagnostic interval data information of the target user, a preset three-dimensional densely connected network model, a training sample set, and a validation sample set.

[0017] In one embodiment, relevant clinical data of the target user is extracted from a hospital information system (HIS), electronic medical record (EMR), or clinical database. This includes basic demographic information, such as name, age, and gender; medical history, such as previous illnesses, surgical history, and allergies; family medical history, particularly information on familial genetic diseases related to tumors; and current symptoms, such as pain intensity, pain location, and the presence of neurological dysfunction. The target user is assigned to undergo a preoperative non-contrast computed tomography (NCCT) scan of the sacrum using standard imaging equipment. During the scan, the equipment operating procedures are strictly followed to ensure high-quality image data, including appropriate scanning angles, slice thicknesses, and resolution. After the scan is completed, the image data is stored in a standardized medical imaging format (such as DICOM) and securely transmitted to a designated data analysis server or workstation to ensure that the data is not damaged or lost during transmission.

[0018] Use magnetic resonance imaging (MRI) equipment to obtain T1-weighted and T2-weighted imaging data for the target user. Ensure that the MRI equipment's magnetic field strength and scanning sequence parameters are correctly set to obtain clear and accurate images. During imaging, instruct the user to maintain correct body positioning to minimize the impact of motion artifacts on image quality. After acquisition, the images are stored in a standard format and transmitted to the analysis system. Examinations are performed on equipment equipped with positron emission tomography (PET) and computed tomography (CT) capabilities to obtain PET-CT information. The acquired image data undergoes quality control before storage and transmission. Data on the target diagnosis interval are compiled from the user's medical records. This includes the time of first discovery of the suspected sacral lesion (e.g., the time the user first reported symptoms, the time the first abnormality was detected on relevant imaging studies), the time of first diagnosis, the time of each follow-up visit, and relevant quantitative data on the lesion at each examination (e.g., tumor size and morphological changes, which can be measured by imaging; changes in the lesion's metabolic activity, such as changes in the SUV value in PET-CT). Time data must be accurately recorded to the specific date.

[0019] The 3D-DenseNet121 model was selected as the default 3D densely connected network model. This model should have excellent capabilities for processing 3D medical image data. The densely connected layers in its network structure effectively extract image features and offer performance advantages in related medical image analysis tasks. The 3D-DenseNet121 model consists of multiple layers, primarily convolutional layers, densely connected layers, transition layers, and fully connected layers. The initial layer of the model contains a series of convolutional layers to extract low-level image features, such as edges and textures. These convolutional layers use a 3D convolution kernel to convolve the image in three dimensions. For example, the kernel size is 3x3x3, with a stride of 1 and padding of 'same' to ensure that the output feature map has the same spatial dimensions as the input image. By stacking multiple convolutional layers, more complex and abstract features are extracted. Densely connected layers are a core component of the 3D-DenseNet121 architecture. In a densely connected layer, the input of each layer comes from the concatenation of the output feature maps of all previous layers. This connection method allows information to flow directly in the network, avoiding the information loss problem caused by the deepening of layers in traditional networks, facilitating the back propagation of gradients, and making the model easier to train. There are L densely connected layers in the model, layer( =1,2,...,L) is the number of input feature maps (where No. The number of feature maps output by the layer), the number of output feature maps is Each densely connected layer contains multiple convolution operations, which are used to further extract and transform the concatenated feature maps.

[0020] In order to control the complexity and number of parameters of the model, a transition layer is inserted between densely connected layers. The transition layer mainly consists of a 1x1x1 convolution operation to reduce the number of feature maps (which plays the role of dimensionality reduction), and a 2x2x2 average pooling operation with a step size of 2, which is used to reduce the size of the feature map and reduce the amount of calculation while maintaining a certain feature representation ability to prevent overfitting. The model ends with a fully connected layer, which is used to integrate and map the previously extracted features and finally output the prediction results. The number of neurons in the fully connected layer depends on the specific classification task. For example, in the task of distinguishing benign and malignant sacral tumors, it is set to 2 neurons (corresponding to the benign and malignant categories respectively), and the output is converted into a category probability distribution through the softmax activation function.

[0021] If pretrained weights are available, obtain them from a public model repository (such as models pretrained on large-scale medical imaging datasets) or from a self-trained model. Pretrained weights enable the model to have better feature extraction capabilities during initialization and accelerate model convergence on the target dataset. After obtaining pretrained weights, fine-tune or reinitialize the weights of some layers (such as fully connected layers) based on the characteristics of the target task and data distribution to adapt to the specific needs of sacral tumor classification. If pretrained weights are not used, the model needs to be initialized. For the weights of convolutional and fully connected layers, use a random initialization method such as normal distribution initialization (mean 0, standard deviation adjusted based on network depth and number of nodes, such as or) or uniform distribution initialization. Bias parameters are typically initialized to 0 or a small constant (such as 0.1). During initialization, ensure that the initialization parameters of different layers are coordinated to promote model stability and convergence in the early stages of training.

[0022] The collected sample data was divided into a training set and a validation set. Inclusion criteria included histopathologically confirmed sacral tumors and users with preoperative NCCT images of a single sacral tumor. Exclusion criteria included a prior anticancer diagnosis, poor image quality, repeat users for follow-up or surveillance, and postoperative tumor recurrence. In addition, the applicant collected data from an external testing cohort from Centers 2 and 3 specifically for final model evaluation. This cohort included 55 users with sacral tumors, and the inclusion and exclusion criteria were consistent with those described previously. Furthermore, the applicant implemented a 5-fold cross-validation scheme on the data from Cohort 1, in which the dataset was divided into five subsets. Each subset was used as an internal test set in turn, and the remaining four subsets were used for model training and validation. During the training and validation phases, the applicant randomly divided eligible users into the training and validation sets in an 8:2 ratio. This approach ensured that every portion of the dataset was used as a training, validation, and internal test set, thereby maximizing data utilization. When splitting the dataset, ensure that the training and validation sets have similar distributions across various features (such as age distribution, gender ratio, and ratio of benign to malignant tumors) to avoid data bias from impacting model training and evaluation. Furthermore, implement data augmentation techniques (such as random rotation, flipping, and scaling images to increase the diversity of training samples) on the training set to improve the model's generalization capabilities.

[0023] S102: Preprocess the training sample set to generate a target training sample set.

[0024] In one embodiment, the training sample set is subjected to data cleaning to generate an initial training sample set after invalid data is removed. Preoperative sacral non-contrast computed tomography (NCCT) images, T1-weighted imaging, T2-weighted imaging, and positron emission tomography-computed tomography (PET-CT) images are screened from the original training sample set. The image integrity is checked, and samples with missing images, severe damage (such as image files that cannot be opened normally, blurred images that cannot identify the tumor area, etc.), or obvious artifacts (such as motion artifacts caused by user movement during scanning, artifacts caused by equipment failure, etc.) are removed. Ensure that the image format is uniform. If images of different formats exist, convert them to a format suitable for subsequent processing (such as the common DICOM format or a format that can be converted to the format required for analysis).

[0025] Organize clinical data, including user demographics (e.g., age, gender), medical history (e.g., previous illnesses, surgical history, allergies), family medical history (especially information on familial genetic diseases related to tumors), and current symptoms (e.g., pain intensity, location, and presence of neurological dysfunction). Check data for accuracy and consistency. Correct any obvious errors (e.g., age discrepancies, incorrect gender records, etc.) by verifying with the original medical records or communicating with clinicians. Delete duplicate user data to ensure that each user appears only once in the training sample set. Organize target diagnostic interval data, including the time of first discovery of a suspected sacral lesion, the time of first diagnosis, the time of each follow-up visit or follow-up, and quantitative data related to the lesion at each examination (e.g., tumor size and morphological changes, which can be measured by imaging; changes in lesion metabolic activity, such as changes in SUV values ​​in PET-CT). Check the logic of time data to ensure that the first diagnosis is later than the first symptom onset and that the order of each follow-up visit or follow-up is reasonable. Perform rationality checks on lesion quantification data and remove obviously abnormal or erroneous data (such as unreasonable and large changes in tumor size in a short period of time without reasonable explanation).

[0026] The initial training sample set was normalized to generate missing sample information. For each type of retained medical imaging data (NCCT, T1 / T2-weighted imaging, PET-CT), the mean and standard deviation of the image pixel values ​​were calculated. Appropriate normalization methods, such as zero-mean unit variance normalization, were used to subtract the mean from each pixel value and then divide it by the standard deviation. This made the image data numerically comparable and facilitated subsequent model training. Similar normalization was performed on the SUV values ​​in PET-CT images to ensure they were within the same scale range as other imaging data. Furthermore, the images were resized according to the model input requirements, such as by cropping or padding them to a uniform size (e.g., [224, 224, 64]) to accommodate the input layer structure of the 3D-DenseNet121 model.

[0027] For numerical features in clinical data (such as age), normalization is performed to map them to a specific interval (such as [0,1] or [-1,1]). For example, for age data, the formula (The minimum age is 1 year old and the maximum age is 100 years old) is normalized. For categorical variables (such as gender, disease type in medical history, etc.), one-hot encoding is used to convert them into numerical form. For example, gender can be encoded as male [1,0] and female [0,1] so that the model can process these data. The time interval data in the diagnostic interval data (such as the time interval from the first discovery of symptoms to the first diagnosis, the time interval between each follow-up visit, etc.) is normalized to match the other data types in terms of numerical value. The mean and standard deviation of the time interval are calculated, and then processed using a normalization method similar to that of image data. For numerical data such as the growth rate of lesion size, normalization is also performed to ensure that it is distributed within an appropriate numerical range to facilitate fusion analysis with other data. See the following text for details, and we will not repeat them here.

[0028] Process missing sample information to generate missing value interpolation prediction information. Process missing value interpolation prediction information to generate target variable values. Carefully examine the cleaned and normalized initial training sample set to identify samples and features with missing values. For image data, some image pixels may be missing or some image sequences may be incomplete. For clinical data, some user medical history information may be missing, or family medical history may be partially missing. For diagnostic interval data, the time of a follow-up visit or lesion quantification data may be missing. Record the location and type of missing values ​​to prepare for subsequent interpolation predictions. Select an appropriate missing value interpolation method based on the data type and distribution characteristics. Specifically, for missing pixels in image data, use interpolation methods based on image neighborhood information, such as nearest neighbor interpolation, bilinear interpolation, or spline interpolation, to estimate the missing pixel's value based on the values ​​of surrounding pixels. For numerical features in clinical data (such as age), if there are few missing values, consider using mean interpolation (filling missing values ​​with the mean of the feature), median interpolation (filling with the median), or model-based predictive interpolation methods. For example, a simple linear regression model can be built, using other relevant features (such as gender and certain indicators from medical history) as independent variables to predict missing age values. For categorical variables (such as gender and disease type), mode interpolation (filling missing values ​​with the most frequently occurring category) or predictive interpolation methods based on classification models (such as decision trees or naive Bayesian models, which predict missing categorical values ​​based on other features) can be used. For missing time intervals or lesion quantification data in diagnostic interval data, appropriate interpolation methods can be selected based on the data's time series characteristics or its relationship to other relevant data. These methods include linear interpolation based on data from previous and subsequent time points or interpolation based on the average change trend of the user group.

[0029] According to the selected interpolation method, identified missing values ​​are filled to generate missing value interpolation prediction information. When performing interpolation, care must be taken to maintain the rationality and consistency of the data. For example, when using the mean to interpolate age, the actual age range and rationality must be considered to avoid unreasonable interpolation results (such as the imputed age being outside the normal human age range). When interpolating categorical variables, ensure that the imputed categories are consistent with the actual situation and have a certain degree of rationality within the dataset. When interpolating image data, ensure that the interpolated image is visually and numerically continuous and reasonable, so as not to affect subsequent image feature extraction and analysis.

[0030] In the present invention, the target variable is the benign or malignant classification of sacral tumors (e.g., benign is 0 and malignant is 1). This information is usually obtained from the user's histopathological confirmation results. Ensure that the definition of the target variable is clear and accurate, consistent with the clinical diagnostic criteria, and avoid errors or confusion during the data annotation process. Check the integrity and accuracy of the target variable value to ensure that each sample has a corresponding target variable value and is correctly labeled. For existing incorrect labels (such as mislabeling benign tumors as malignant or vice versa), correct them by checking with the original pathology report, clinician review, etc. If there are few missing values, consider deleting the corresponding samples; if there are many missing values ​​and there is a certain pattern or they can be inferred based on other information, try to use appropriate methods to fill them in (such as inferring based on the user's other clinical characteristics, imaging characteristics, etc.). Finally, accurate and complete target variable values ​​are obtained to provide correct supervision information for subsequent model training.

[0031] The target variable value is processed to generate target variable parameter information, and a target training sample set is generated based on the target variable parameter information. A statistical analysis is performed on the distribution of the target variable (benign and malignant tumors), and statistical indicators such as the number and proportion of benign and malignant samples are calculated. Understand whether the distribution of benign and malignant tumors in the data set is balanced. If there is a serious imbalance (such as the number of malignant samples is far greater than that of benign samples or vice versa), adopt corresponding processing strategies (such as oversampling, undersampling, using weighted loss functions, etc.) in subsequent model training to avoid the model's bias towards the majority category and improve the model's prediction ability for the minority category. Generate target variable parameter information based on the distribution characteristics and analysis results of the target variable. These parameter information include the number of categories of the target variable (in this invention, there are 2 categories: benign and malignant), the number and proportion of samples in each category, and the category weight (if different categories need to be weighted during model training), etc. For example, if benign samples account for 30% and malignant samples account for 70%, in order to balance the model's attention to the two categories, the weight of the benign sample is set to , the weight of the malignant sample is During model training, the loss function is adjusted based on the sample weights, so that the model places more emphasis on samples from minority categories. These target variable parameter information will play an important role in the model training process, helping the model better learn and distinguish the characteristics of samples from different categories.

[0032] After data cleaning, normalization, missing value interpolation, target variable value processing, and target variable parameter information generation, the image data, clinical data, diagnostic interval data, and target variable values ​​are integrated according to the sample. This ensures that the data types in each sample accurately correspond to each other, that is, each user's imaging data, clinical data, and diagnostic interval data correspond to the corresponding target variable value for benign or malignant tumors, forming a complete sample data set. The integrated sample data is further verified and adjusted based on the previously segmented results to ensure that the training sample set is representative, covering different user conditions (e.g., age, gender, tumor type, disease stage, etc.), and has a certain degree of diversity and balance in various features (e.g., imaging features, clinical features, diagnostic interval features, etc.). Furthermore, the training sample set is subjected to necessary preprocessing operations, such as data augmentation (e.g., random rotation, flipping, and scaling of images to increase the diversity of the training samples), to improve the model's generalization ability. This ultimately results in the target training sample set for model training, preparing for subsequent training based on the 3D-DenseNet121 model. The detailed process is described below. Through the above steps, the training sample set is processed comprehensively and meticulously to ensure the data quality is reliable, the format is unified, the features are complete, and the target variables are accurate, laying a solid data foundation for building an accurate and effective preoperative warning model for sacral tumors.

[0033] S103 : Processing the preset three-dimensional densely connected network model based on the target training sample set and the verification sample set to generate a target three-dimensional densely connected network model.

[0034] In one embodiment, any number of data features in the target training sample set are obtained, and a sampling ratio is generated based on the number of each data feature in the target training sample set. Various data features are extracted from the target training sample set, including image features (such as texture, shape, and signal intensity features in preoperative non-contrast sacral CT images, T1-weighted imaging, T2-weighted imaging, and PET-CT images), clinical data features (such as user age, gender, medical history, family medical history, etc.), and diagnostic interval data features (such as the time interval from first symptom onset to first diagnosis, lesion size growth rate, etc.). For image features, image processing algorithms and feature extraction techniques are used (such as the convolutional layers in convolutional neural networks, which can automatically extract low-level and high-level image features). For clinical and diagnostic interval data, corresponding numerical or categorical information is directly obtained. The number of occurrences of each data feature in the training sample set is counted, for example, the number of age groups, the number of users of different genders, the frequency of occurrence of various medical histories, and the number of different imaging features.

[0035] Calculate the sampling ratio based on the number of each data feature. There are many ways to calculate the sampling ratio. For example, it can be calculated based on the ratio of the number of features, that is, the sampling ratio of each feature is equal to the ratio of the number of features to the total number of features. Or adjust the sampling ratio based on the importance of the feature, giving a higher sampling ratio to important features to ensure that these key information can be fully retained during the sampling process. There are n data features in the training sample set, and the number of feature i is count. Then its sampling ratio is In this way, the sampling ratio corresponding to each data feature is generated, providing a basis for subsequent sampling operations.

[0036] Based on the sampling ratio, the target training sample set is sampled to generate a preset number of sampling features. Based on any data feature and each sampling feature, multiple data groups are generated, wherein each data group contains a preset number of data samples, and at least one data sample includes identification information. Based on the calculated sampling ratio, the target training sample set is sampled. For each data sample, whether the sample is selected into the sampling set is determined based on the data features it contains and the corresponding sampling ratio. For example, using a random sampling method, for each sample, a random number r ( ),if( (where is the sampling ratio of a feature in the sample), the sample is selected and included in the sampling set. This process is repeated until the preset number of sampled features is reached. This sampling method can preserve the distribution of data features while reducing the number of training samples, improving training efficiency and, to a certain extent, avoiding overfitting. Especially when the training sample set is large, the sampling operation can make the training process more efficient and feasible.

[0037] For the preset number of sampling features generated, any data feature is combined with each sampling feature to generate multiple data groups. The specific operation can be to use each sampling feature as the basis of a group of data, and then add other data features in sequence to form a sample group containing multiple data features. Each data group contains a preset number of data samples, and ensures that at least one data sample includes identification information (such as whether the identification sample is a risk factor affecting sacral lesions, or other identification related to the special properties of the sample). These identification information can play an important role in subsequent model training and evaluation, such as being used to distinguish different types of samples, measure the importance of samples, or serve as a reference for model prediction. By reasonably combining data features to form data groups, the information in the sample set can be fully utilized to provide rich and diverse data inputs for model training, so that the model can learn the relationship and interaction between different features.

[0038] A pre-set 3D densely connected network model is trained based on data samples from multiple datasets to generate a trained 3D densely connected network model. A pre-set 3D densely connected network model (such as the 3D-DenseNet121 model) is trained using data samples from the generated datasets. Prior to training, the data samples are preprocessed to meet the model's input requirements, such as resizing images, normalizing numerical data, and encoding categorical variables (for detailed preprocessing steps, see the previous section). The preprocessed data samples are input into the model, which extracts features and learns from the data based on its internal network structure (including convolutional layers, densely connected layers, transition layers, and fully connected layers). During training, the model's predictions are calculated through forward propagation. A loss function (such as the cross-entropy loss function) is then used to calculate the difference between the predicted results and the true labels (e.g., benign or malignant tumor labels). Next, the gradient of the loss function with respect to the model parameters is calculated using the backpropagation algorithm. Based on the gradient, an optimizer (such as the Adam optimizer) is used to update the model parameters, continuously adjusting the model weights and biases so that the model's predictions gradually approach the true labels. This process is repeated until the predetermined number of training rounds is completed or other training stopping conditions are met (such as the loss value no longer decreases or reaches a set minimum threshold).

[0039] The trained 3D densely connected network model is processed based on the validation set to generate validation results. If the data sample containing identification information in the validation results represents a risk factor for sacral lesions, the trained 3D densely connected network model is used as the target 3D densely connected network model. After model training is complete, the trained 3D densely connected network model is processed using the validation set to generate validation results. The data samples in the validation set undergo the same preprocessing as the training samples and are then input into the trained model to obtain the model's prediction results for the validation samples. Based on the prediction results and the true labels of the validation samples, various evaluation metrics such as precision, recall, F1 score, and area under the curve (AUC) are calculated to comprehensively assess the model's performance on unseen data. These evaluation metrics reflect the model's accuracy, sensitivity, specificity, and overall discriminative ability, helping to determine whether the model is overfitting or underfitting, and its effectiveness in practical applications. For example, precision represents the proportion of samples correctly predicted by the model to the total number of samples, while AUC measures the model's ability to distinguish between positive and negative examples by calculating the area under the receiver operating characteristic (ROC) curve.

[0040] Examine the data samples containing identifiers in the validation results to determine whether they represent risk factors for sacral lesions. Data analysis is used to determine which identifiers are associated with the risk of sacral lesions. For example, specific imaging features, clinical characteristics (such as older age, a specific family history), or diagnostic interval characteristics (such as a faster lesion growth rate) are considered risk factors. If the validation results contain a large number of samples containing these risk factors or the model demonstrates good predictive performance (e.g., high accuracy and large AUC) on samples associated with these risk factors, this indicates that the model is effectively identifying and processing risk-related information. If the validation results demonstrate that the trained model performs well in processing risk factors for sacral lesions, the trained 3D densely connected network model is designated as the target 3D densely connected network model. This target model will be used for subsequent prediction and early warning of preoperative sacral lesions in target patients, providing support for clinical diagnosis. If the validation results are unsatisfactory, adjust the model parameters, reselect features, increase training data, or implement other improvement measures. Retrain and validate again until a satisfactory target model is obtained. Through this model selection process based on validation results, it can be ensured that the final model has good performance and reliability, can accurately predict the attribute information of sacral lesions in practical applications, and provide valuable reference for clinical decision-making.

[0041] S104: Process the preoperative non-contrast computed tomography image of the sacrum of the target user to generate a three-dimensional region-of-interest image of the target user.

[0042] In one embodiment, pre-set anatomical structure constraint information is obtained, where the pre-set anatomical structure constraint information is used to characterize organ information that matches the sacral structure. Detailed anatomical knowledge related to the sacral structure is obtained by consulting professional anatomical atlases, medical imaging anatomy textbooks, or authoritative medical databases. These resources include information on the normal human sacrum's morphology, position, and anatomical relationships with surrounding organs and tissues, such as the proximity of the sacrum to the pelvis, spine, nerves, blood vessels, and surrounding soft tissues. Through in-depth research on this data, the range of organs and tissues closely associated with the sacral structure in both physiological and pathological states is determined, providing an accurate anatomical basis for subsequent treatment. The specific content of the pre-set anatomical structure constraint information is clarified. Organs and tissues that are spatially closely adjacent to the sacrum and may impact tumor assessment and management are identified. For example, pelvic bone structures within a certain range, nearby nerve plexuses (such as the sciatic nerve and cauda equina), important blood vessels (such as branches of the internal iliac artery), and soft tissue areas potentially invaded by the tumor are identified. This information is described and recorded in a standardized manner, for example, by defining the anatomical coordinate range, structure name and relative position relationship, etc., to form preset anatomical structure restriction information that can be used for subsequent image processing, ensuring that these key structural information can be accurately identified and utilized when processing the target user's images.

[0043] The target user's preoperative sacral non-contrast computed tomography (NCCT) images were resampled to generate NCCT images of the target size. The target size parameters were determined based on the input requirements of the pre-set 3D densely connected network model and the convenience of subsequent processing. For example, based on the model's optimal adaptation requirements for the input image size, the image was resampled to a size of [224, 224, 64] (width, height, depth). This size ensured sufficient image detail while achieving a good balance between computational resources and model processing efficiency. Furthermore, considering the need to observe tumors and surrounding structures at different scales, the target size should be chosen to highlight the relationship between the lesion and surrounding anatomical structures, facilitating accurate identification and analysis of tumor features. Based on the image characteristics and target size requirements, an appropriate resampling algorithm was selected, taking into account both image quality and computational efficiency. For NCCT images of the sacrum, the bilinear interpolation algorithm was a suitable choice, generating relatively smooth and accurate images of the target size without excessively increasing the computational burden, thus meeting the image quality requirements of subsequent processing. The selected resampling algorithm was then used to process the target user's preoperative sacral non-contrast NCCT images. The pixel values ​​of the original image are recalculated and distributed according to the algorithm to generate a non-contrast computed tomography image of the sacrum of the target size. During the resampling process, it is important to ensure that the image's anatomical structure is preserved. In particular, the tumor region and its relative positional relationship to surrounding tissue should not be significantly deformed or distorted due to resampling. Resampling converts the original image to a uniform size, making it more suitable for subsequent image analysis and model input requirements, laying the foundation for accurate extraction of the tumor region and other relevant information.

[0044] Process a non-contrast computed tomography (NCCT) image of the sacrum of the target size to generate information about the cropped target tumor region. Image processing algorithms and techniques are used to analyze the NCCT image of the target size to preliminarily locate the tumor region. Methods based on image grayscale values, texture features, or morphological analysis are employed. For example, threshold segmentation techniques can be used to segment potential tumor regions based on the grayscale difference between tumor tissue and surrounding normal tissue. Alternatively, texture feature analysis algorithms can be used to identify regions with specific texture patterns (such as the uneven texture often exhibited by tumor tissue) as candidate tumor regions. Combining these methods, the approximate location and extent of the tumor in the image can be preliminarily determined, providing a reference for further precise cropping. Based on the preliminarily located tumor region, precise cropping is performed in accordance with pre-set anatomical constraints. This ensures that the cropped target tumor region encompasses not only the tumor tissue itself but also potentially affected tissues and structures within a certain range (determined by the anatomical constraints). During the cropping process, care should be taken to avoid cropping out information that is critical for tumor diagnosis and analysis, such as the relationship between the tumor and surrounding vital blood vessels and nerves. Through precise cropping, a relatively small image portion focused on the tumor and its surrounding key areas is obtained, reducing the interference of irrelevant information, highlighting the tumor characteristics, facilitating subsequent more detailed analysis and processing of the tumor area, and also helping to improve the efficiency and accuracy of subsequent model processing.

[0045] The cropped target tumor region is processed based on pre-set anatomical constraints to generate a 3D region-of-interest image for the target user. The 3D anatomical data (such as the 3D model or spatial coordinate range of surrounding organs and tissues) contained in the pre-set anatomical constraints is integrated with the reconstructed 3D tumor region. This ensures that the 3D region-of-interest image not only captures the tumor itself but also clearly displays its positional relationship with surrounding critical anatomical structures in 3D space, providing richer information for comprehensive tumor assessment. The generated 3D region-of-interest image is post-processed and optimized to enhance its visualization and information presentation. For example, image contrast and brightness can be adjusted to enhance the clarity of the tumor region and surrounding structures. The image can also be smoothed to reduce noise, but care should be taken to avoid over-smoothing that may lead to loss of image detail. Pseudo-color coding techniques can also be applied to assign different colors based on the characteristics of different tissues (such as tumor tissue, normal bone tissue, and soft tissue) to enhance image readability and recognition. Through these post-processing operations, high-quality three-dimensional images of the target user's region of interest are generated, making them more suitable for subsequent clinical analysis and diagnosis, and as input data for the preset three-dimensional densely connected network model, providing strong support for the accurate prediction of sacral lesion properties.

[0046] S105 : Process the multimodal data information of the target user and the target diagnostic interval data information of the target user to generate a dynamic risk parameter of the target user.

[0047] In one embodiment, feature extraction is performed on the target user's T1-weighted imaging information and T2-weighted imaging information to generate signal intensity features, texture features, shape features, and enhancement curve features. Using professional medical image processing software, the target user's T1-weighted and T2-weighted imaging information is loaded to ensure that the image data is complete, formatted uniformly, and the resolution meets analysis requirements. An image segmentation algorithm is used to accurately identify and extract the tumor region, reducing the impact of human error on subsequent analysis. When calculating signal intensity features, robust statistical methods are used to calculate the mean, variance, maximum, and minimum values ​​of the signal intensity within the tumor region. Furthermore, based on statistical analysis results from large-scale clinical sample data, a signal intensity threshold range with diagnostic significance for sacral tumors is determined. For example, a study of thousands of sacral tumor cases found that a mean signal intensity value below 80 (units) and a variance below 50 (units) on T1-weighted images indicates a high likelihood of benign tumors. The calculated signal intensity features are normalized and mapped to the interval [0, 1] to eliminate the numerical differences caused by different imaging devices and scanning parameters, ensuring that the features are comparable when multi-source data are fused.

[0048] Based on the gray-level co-occurrence matrix (GLCM) method and previous research on sacral tumor texture features, suitable combinations of orientations (0°, 45°, 90°, 135°) and distance parameters (1 pixel, 2 pixels) were determined for this method. To improve computational efficiency and accuracy, parallel computing technology was used to accelerate the frequency statistics of pixel pairs at different orientations and distances. After calculating texture feature parameters such as contrast, correlation, energy, and entropy, each texture feature was normalized using a standardization formula (such as z-score normalization) to a mean of 0 and a standard deviation of 1, enhancing the stability and comparability of the features across different cases. Furthermore, spatial frequency analysis of texture features was introduced to supplement the description of tumor texture features, further enhancing the ability of texture features to distinguish benign from malignant tumors.

[0049] Use high-precision edge detection algorithms (such as the improved Canny edge detection algorithm, combined with morphological operations to optimize edge extraction effects) to obtain the precise outline of the tumor area. Calculate the perimeter, area, and circularity of the tumor (Formula: ) and eccentricity. Considering the differences in tumor size and image scale among different users, a normalization method based on image moments is used to normalize shape features, ensuring that they are not affected by the actual tumor size and image resolution. Furthermore, a database of common sacral tumor shape features is constructed. By comparing and matching these features with standard shape patterns in the database, it assists in determining tumor type and benign or malignant propensity, providing an intuitive reference for shape features in clinical diagnosis.

[0050] For T1-weighted contrast-enhanced imaging sequences, serial T1-weighted contrast-enhanced images are acquired at standard clinical time points (e.g., 1, 3, 5, and 7 minutes after contrast agent injection). During image preprocessing, image registration algorithms (e.g., registration based on maximizing mutual information) and motion correction techniques (e.g., optical flow) are employed to ensure precise registration of images at different time points and eliminate interference from user motion on enhancement curve feature calculation. For each tumor pixel, an enhancement curve is plotted showing the signal intensity over time, and its peak time, rising slope, falling slope, and area under the curve are calculated. To improve calculation accuracy and stability, spline interpolation is used to smooth the enhancement curve before feature calculation. Enhancement curve features are normalized to ensure compatibility with other features in terms of numerical range, facilitating subsequent fusion analysis. Furthermore, based on clinical case data, enhancement curve feature templates for different sacral tumor types are established. Similarity comparison with the templates assists in determining the nature and invasiveness of the tumor.

[0051] Feature extraction and processing are performed on positron emission tomography-computed tomography information to generate target uptake value features. PET-CT image fusion technology is used to precisely align PET and CT images and accurately locate the tumor region. An adaptive threshold segmentation algorithm is used to automatically determine the segmentation threshold based on the metabolic activity characteristics of tumor tissue in PET images, and standardized uptake value (SUV) information of the tumor region is extracted. The maximum, minimum, mean, and standard deviation of the SUV are calculated to comprehensively characterize the metabolic activity level of tumor tissue. Taking into account the calibration differences between different PET-CT devices, a normalization method based on phantom calibration is used to correct and normalize the SUV features to ensure comparability between different devices. In addition, SUV histogram analysis is introduced to extract statistical features such as skewness and kurtosis of the SUV distribution, describing the metabolic heterogeneity of tumors from a more comprehensive perspective and providing more valuable information for tumor diagnosis.

[0052] Target diagnosis interval data for the target user is quantified to generate visit interval information and lesion size growth rate values. Target diagnosis interval data is collected from the target user's medical records, including the time of first discovery of the suspected sacral lesion, the time of first diagnosis, the time of each follow-up visit, and relevant quantitative lesion data from each examination (such as tumor size and morphological changes, which can be measured by imaging; changes in lesion metabolic activity, such as changes in SUV values ​​in PET-CT). Time data is carefully verified for accuracy to ensure that the chronological sequence is reasonable and free of logical errors (e.g., the time of first diagnosis must be later than the time of first symptom onset). For lesion quantitative data, quality control measures are implemented to ensure measurement accuracy and reproducibility, such as averaging multiple measurements by experienced radiologists or professional technicians. Visit interval information (i.e., the time difference from first symptom onset to first diagnosis) and the time interval between each follow-up visit or follow-up visit are calculated in days. To calculate the lesion size growth rate, a linear regression model is used to fit the tumor growth curve based on imaging measurements (such as diameter or volume) at different time points. The slope of the growth curve is calculated as the lesion size growth rate value. The access interval information and lesion size growth rate values ​​are normalized and mapped to a specific interval (e.g., [-1, 1]) to match the numerical range of other features and facilitate subsequent fusion analysis. Furthermore, the correlation between access interval and lesion size growth rate and tumor benign or malignant status and prognosis is analyzed to provide a basis for the calculation of dynamic risk parameters.

[0053] Signal intensity, texture, shape, enhancement curve, and target uptake value features are processed to generate a target shape factor. A multimodal feature vector is constructed by integrating these features. Dimensionality reduction algorithms such as principal component analysis (PCA) or independent component analysis (ICA) are used to reduce the multimodal feature vector and extract the principal components as the target shape factor. During the dimensionality reduction process, cross-validation is used to determine the optimal number of principal components or independent components to reduce data dimensionality and computational complexity while preserving feature information. The generated target shape factor is normalized to conform to a standard normal distribution, facilitating subsequent fusion analysis with other parameters. Furthermore, visualization techniques (such as scatter plots of the target shape factor distribution across different cases) are used to visually demonstrate the relationship between the target shape factor and tumor benign or malignant characteristics and other clinical characteristics, providing clinicians with a more intuitive diagnostic reference.

[0054] The target shape factor is processed based on the access time interval information and the lesion size growth rate value to generate the target user's dynamic risk parameter. Based on the access time interval information and the lesion size growth rate value, a linear weighted model is constructed to calculate the target user's dynamic risk parameter. For example, let the dynamic risk parameter (in represents the dynamic risk parameter, represents the target shape factor, Indicates access time interval information, Indicates the growth rate of lesion size, 、 and The weight coefficients are obtained by training a large amount of clinical sample data through machine learning algorithms (such as random forest regression, support vector machine regression, etc.). During the training of the weight coefficients, the leave-one-out cross-validation or k-fold cross-validation (such as k=10) method is used to evaluate the model performance and select the model parameters with the best performance. Based on the calculated dynamic risk parameters, a risk classification standard is established, such as classifying the dynamic risk parameters into low risk ( ), medium risk ( ) and high risk ( ) three levels, providing a clear risk assessment basis for clinical decision-making.

[0055] S106: Process the clinical data information of the target user to generate physiological attribute parameters of the target user.

[0056] In one embodiment, feature extraction is performed on the target user's clinical data information to generate demographic features, tumor location features, target user's genetic features, and target user's comorbidity features. Accurately extract the target user's basic demographic information, including age, gender, height, weight, and ethnicity, from the hospital information system (HIS) or electronic medical record (EMR). Ensure the integrity and accuracy of the data. For missing or unclear data, supplement and verify it through communication with the user or family members, and review other relevant records. For example, if age data is missing, it can be estimated based on information such as the user's first visit to the doctor and date of birth; if gender information is unclear, it can be determined by asking the user or referring to relevant statements in other medical records. These demographic characteristics are organized into a standardized data format for subsequent analysis and processing.

[0057] Based on the target user's preoperative sacral non-contrast computed tomography (NCCT) images, magnetic resonance imaging (MRI) images, or other relevant imaging data, an experienced radiologist or professionally trained medical imaging analyst determines the specific location of the tumor in the sacrum. Standard anatomical positioning systems (such as internationally recognized anatomical nomenclature) are used to accurately describe the tumor's location, recording its positional relationship relative to various sacral anatomical landmarks (such as the sacral promontory and sacral foramen), as well as the tumor's distribution in the anterior-posterior, lateral, and superior-inferior directions. Furthermore, the tumor's proximity to surrounding critical structures (such as nerves, blood vessels, and pelvic organs) is assessed, and this locational information and proximity are converted into quantifiable or categorized feature data, such as the distance between the tumor and nerves and whether the tumor has invaded surrounding blood vessels.

[0058] Obtain the target user's genetic testing report and extract genetic signature information relevant to sacral tumors. This includes the mutation status of specific genes (e.g., detecting the presence of known mutations in genes associated with tumor development and progression, such as TP53 and RB1), gene expression levels (e.g., mRNA expression levels of certain tumor-related genes), and gene copy number variations (e.g., amplifications or deletions of certain gene regions). Gene signature data is normalized to allow analysis within the same framework as other clinical data. For example, gene expression levels are log-transformed to make the data distribution closer to a normal distribution, facilitating subsequent calculations and comparisons. A comprehensive review of the target user's medical history and current health status is conducted to determine the presence of other comorbidities. Comorbidity information can be obtained from diagnostic records, hospitalization records, outpatient visits, and user self-reports in electronic medical records. Common comorbidities associated with sacral tumors include diabetes, cardiovascular diseases (e.g., hypertension, coronary artery disease), and pulmonary diseases (e.g., chronic obstructive pulmonary disease). Each comorbidity is treated as a feature and coded into two categories according to its presence or absence (e.g., presence of comorbidity is recorded as 1, absence as 0), or graded according to the severity of the comorbidity (e.g., mild, moderate, and severe are coded as 1, 2, and 3, respectively) to reflect the degree of impact of the comorbidity on the user's physiological state.

[0059] Demographic characteristics, tumor location characteristics, target user genetic characteristics, and target user comorbidity characteristics are processed separately to generate importance scores and weights for each clinical data point. To calculate the importance scores for each clinical data point, a gradient boosting decision tree (GBDT) model is used. The GBDT model consists of multiple decision trees, each of which learns the characteristic relationships of the data by optimizing an objective function during construction. A GBDT model consisting of decision trees (base learners) is constructed. When constructing the mth decision tree (m = 1, 2, ..., M), a certain percentage (e.g., 70%) of the training data is randomly sampled with replacement from the training data as the training set for the decision tree, and the remaining samples serve as the validation set.

[0060] For each decision tree, the splitting of its leaf nodes is based on information gain or other appropriate splitting criteria. For example, when calculating the splitting criterion of the mth decision tree, the contribution of each feature to reducing the sample impurity (such as Gini impurity or information entropy) will be considered. At a certain node, there are n samples. According to the different values ​​of a certain feature V, the samples are divided into different subsets, and the reduction of sample impurity before and after the division is calculated. In the calculation process, the characteristic vector of the sample is involved (including demographic characteristics, tumor location characteristics, genetic characteristics, and comorbidity characteristics, etc.) and true labels (For example, the benign or malignant label of a tumor, where benign is recorded as 0 and malignant is recorded as 1). The number of leaf nodes in each decision tree It will vary depending on the complexity of the data and the settings of the model. For example, for complex data, more leaf nodes may be generated to better fit the data. In each decision tree, each leaf node has a sample weight sum, which represents the sum of the weights of the samples that reach the leaf node. , which is adjusted during model training based on the importance and misclassification of the samples.

[0061] The method also includes a calculation formula for obtaining the importance score value, which is: ;in, represents the importance score of the jth feature; M is the number of base learners (decision trees) in the gradient boosting decision tree model; is the number of leaf nodes of the mth decision tree; is the sum of the sample weights of the t-th leaf node in the m-th decision tree; n is the number of samples; Represents the features used when splitting the t-th leaf node in the m-th decision tree; is an indicator function, when hour, ,otherwise ; Indicates the reduction in impurity of the i-th sample before and after the t-th split in the m-th decision tree; is the feature vector of the i-th sample at the t-th split in the m-th decision tree, is the true label of the i-th sample (e.g., benign, malignant). For each of the M decision trees, traverse its leaf nodes. When the t-th leaf node of the m-th decision tree is split, the feature used when the node is split It happens to be the age characteristic (i.e. ),but ,otherwise At this point, calculate the reduction in impurity of sample i before and after the node split ,in is the feature vector of sample i at the t-th split of the m-th decision tree (including age features and other related features), is the true label of sample i (tumor benign or malignant). Then, the contribution value of age feature at this node is calculated according to the formula The contribution value of the age feature in each leaf node of all decision trees is summed to obtain the importance score of the age feature. The same method can be applied to calculate the importance scores of other clinical data, such as tumor location features, genetic features, and comorbidity features. In this way, the importance of each feature in distinguishing benign and malignant tumors or other clinical decisions can be quantified, providing a basis for subsequent weight calculations.

[0062] After obtaining the importance score of each clinical data information, calculate its weight value. Normalize the importance score of each feature so that its sum is 1, and obtain the weight value of each feature. Suppose the importance score of age feature is , the importance score of the tumor location feature is , the importance score of the gene feature is , the importance score of comorbidity characteristics is , then the weight value of the age feature , the weight values ​​of other features can be calculated in the same way. In this way, the weight value of each feature reflects its relative importance in the overall clinical data information. The larger the weight value, the greater the impact of the feature on the physiological attribute parameters of the target user. The importance score value of each clinical parameter information and the weight value of each clinical data information are processed separately to generate the physiological attribute parameters of the target user. According to the calculated weight value of each clinical data information, the corresponding features are weighted and summed to generate the physiological attribute parameters of the target user. After the demographic features are processed, the feature vector (in represents the specific characteristic values ​​in demographic characteristics, such as age, gender, etc. after encoding or numerical conversion), and its corresponding weight vector is ; The tumor location feature vector is , the weight vector is ; The gene feature vector is , the weight vector is ; The comorbidity feature vector is , the weight vector is .

[0063] The physiological attribute parameters of the target user can be calculated using the following formula: (in (represents the transpose of vector X). This weighted summation approach integrates demographic characteristics, tumor location characteristics, genetic characteristics, and comorbidity characteristics to generate a parameter that comprehensively reflects the relationship between the target user's physiological attributes and tumors. This physiological attribute parameter can be used as part of subsequent model input, along with other data (such as imaging data and dynamic risk parameters), to provide more comprehensive information support for accurately predicting sacral tumor attributes.

[0064] S107: Processing the target user's preoperative sacrum non-contrast computed tomography image, the target user's three-dimensional region of interest image, the target user's dynamic risk parameters, and the target user's physiological attribute parameters based on the target three-dimensional densely connected network model to generate sacrum warning information.

[0065] In one embodiment, the image data normalization processing is performed on the target user's preoperative sacral non-contrast CT image and the target user's 3D region of interest image to generate the mean value and standard deviation of the image pixel value. We have a set of preoperative sacral non-contrast CT image datasets. and 3D region of interest image datasets .calculate and The mean value of each pixel (or voxel) in 、 and standard deviation 、 .right Each pixel value in , perform normalization: .right Each voxel value in , perform normalization: .

[0066] The dynamic risk parameters and physiological attribute parameters of the target user are normalized to generate the mean value of the dynamic risk parameter, the standard deviation of the dynamic risk parameter, the mean value of the physiological attribute parameter and the standard deviation of the physiological attribute parameter. Let the dynamic risk parameter be , the physiological attribute parameters are .calculate The mean and standard deviation ,as well as The mean and standard deviation . Normalize the dynamic risk parameters: . Standardize physiological attribute parameters: The mean value of image pixel values, the standard deviation of image pixel values, the mean value of dynamic risk parameters, the standard deviation of dynamic risk parameters, the mean value of physiological attribute parameters and the standard deviation of physiological attribute parameters are fused to generate multi-channel feature data. Assume that the mean value and standard deviation of the normalized image data are and (including fusion results of CT and 3D), the mean and standard deviation of the parameters after standardization are and (Including the fusion results of dynamic risk and physiological attribute parameters). Fuse these data to generate multi-channel feature data , concatenate these data together in order: .

[0067] The multi-channel feature data is processed based on the target three-dimensional densely connected network model to generate a target prediction value. The method includes a calculation formula for obtaining the target prediction value, which is: ; Where L represents the total number of layers of the three-dimensional densely connected network model (including convolutional layers, fully connected layers, etc.), Represents the weight parameter of the lth layer, whose dimension is determined by the size of the input and output feature maps and the number of neurons in the lth layer; represents the learning parameter vector of layer l, Represents the feature transformation operation performed by the lth layer on the input multi-channel feature data F; Represents the bias parameter of layer l, whose dimension is consistent with the output feature map or number of neurons in layer l; Represents the activation function of the output layer. Input into the 3D densely connected network model, which has L layers, and performs feature transformation and propagation layer by layer. ), according to the input multi-channel feature data , combined with the weight parameter , learning parameter vector and bias parameters Perform calculations.

[0068] Specifically, the 3D densely connected network model has L=3 layers. In the first layer: weight parameters is 0.5 (randomly initialized example value), learning parameter vector is [0.2, 0.3] (randomly initialized sample values), the bias parameter is 0.1 (randomly initialized sample value). Input multi-channel feature data After the first layer of feature transformation (The feature transformation function is a linear transformation plus a nonlinear activation function, for example , where relu is a linear rectifier function). The output of the first layer is In layer 2: weight parameters is 0.3, learning parameter vector is [0.1,0.4], the bias parameter is 0.2. The input is the output of layer 1 , after the feature transformation of the second layer The output of the second layer is . In layer 3: weight parameters is 0.2, learning parameter vector is [0.3,0.2], the bias parameter is 0.3. The input is the output of layer 2 , after the feature transformation of the third layer The final target prediction value (here is the activation function of the output layer, and the sigmoid function is used to map the result to between 0 and 1).

[0069] The target prediction value is processed based on the preset threshold to generate sacral warning information, which is used to characterize the attribute information of the target user's preoperative sacral lesions. Assume that the preset threshold is T=0.7. When the target prediction value y meets different conditions, different sacral warning information is generated: if , a high-risk sacral warning message is generated. This means that the target user's preoperative sacral lesion is likely to be of a more serious nature, such as malignant, extensive, or aggressive. , a low-risk sacral warning message is generated. This indicates that the target user's preoperative sacral lesion may have relatively mild attributes, such as benign lesions, small lesion area, and relatively stable lesions.

[0070] The server first collects a wide range of data from target users, including clinical, NCCT images, multimodal, and diagnostic interval data. Training samples are meticulously preprocessed to ensure data quality. When constructing the model, a densely connected three-dimensional network is optimized using target training and validation samples. The final model is determined through training and validation of sampled features. NCCT images are then processed based on anatomical constraints to generate region-of-interest images. For feature processing, dynamic risk parameters are derived from multimodal data, while physiological attribute parameters are derived from clinical data. After normalizing and standardizing these and image data, a weighted fusion strategy (with learnable weights assigned to image, dynamic risk parameter, and physiological attribute parameter channels) is used to generate multichannel feature data. Finally, this multichannel feature data is input into the model, and warning information is generated based on preset dynamic thresholds, effectively assisting in the determination of tumor benign or malignant status and improving the accuracy of preoperative diagnosis.

[0071] In one embodiment, Figure 2 As shown, the present application also provides a preoperative warning device for sacral tumors, comprising:

[0072] an acquisition module 201 for acquiring clinical data information of a target user, a preoperative non-contrast computed tomography image of the sacrum of the target user, multimodal data information of the target user, target diagnostic interval data information of the target user, a preset three-dimensional densely connected network model, a training sample set, and a validation sample set, wherein the multimodal data information of the target user includes T1-weighted imaging information of the target user, T2-weighted imaging information of the target user, and positron emission tomography-computed tomography information of the target user;

[0073] The processing module 202 is used to preprocess the training sample set to generate a target training sample set; process the preset three-dimensional densely connected network model based on the target training sample set and the verification sample set to generate a target three-dimensional densely connected network model; process the preoperative non-contrast computed tomography image of the sacrum of the target user to generate a three-dimensional region of interest image of the target user; process the multimodal data information of the target user and the target diagnostic interval data information of the target user to generate the dynamic risk parameters of the target user; process the clinical data information of the target user to generate the physiological attribute parameters of the target user; process the preoperative non-contrast computed tomography image of the sacrum of the target user, the three-dimensional region of interest image of the target user, the dynamic risk parameters of the target user and the physiological attribute parameters of the target user based on the target three-dimensional densely connected network model to generate sacrum warning information, wherein the sacrum warning information is used to characterize the attribute information of the preoperative sacral lesions of the target user.

[0074] Each embodiment of this application is described in a related manner. Similar portions between the various embodiments can be referenced to each other. Each embodiment focuses on the differences from other embodiments. In particular, the embodiments of the preoperative warning method, electronic device, electronic device, and readable storage medium for evaluating sacral tumors are generally similar to the embodiments of the preoperative warning method for sacral tumors described above, so the description is relatively simple. For related portions, refer to the partial description of the embodiments of the preoperative warning method for sacral tumors described above.

Claims

1. A computer-implemented preoperative early warning method for sacral tumors, characterized in that: include: According to the target user's clinical data information, the target user's preoperative sacral non-contrast computed tomography image, the target user's multimodal data information, the target user's target diagnostic interval data information, a preset three-dimensional densely connected network model, a training sample set, and a validation sample set, wherein the target user's multimodal data information includes the target user's T1-weighted imaging information, the target user's T2-weighted imaging information, and positron emission tomography-computed tomography information; Preprocessing the training sample set to generate a target training sample set; Processing the preset three-dimensional densely connected network model based on the target training sample set and the verification sample set to generate a target three-dimensional densely connected network model; processing a preoperative non-contrast computed tomography image of the sacrum of the target user to generate a three-dimensional region of interest image of the target user; Processing the multimodal data information of the target user and the target diagnostic interval data information of the target user to generate a dynamic risk parameter of the target user; The clinical data information of the target user is processed to generate physiological attribute parameters of the target user, including performing feature extraction processing on the clinical data information of the target user to generate demographic characteristics, tumor location characteristics, genetic characteristics of the target user, and comorbidity characteristics of the target user; the demographic characteristics, tumor location characteristics, genetic characteristics of the target user, and comorbidity characteristics of the target user are processed separately to generate an importance score value for each clinical data information and a weight value for each clinical data information; the importance score value for each clinical parameter information and the weight value for each clinical data information are processed separately to generate the physiological attribute parameters of the target user; the calculation formula of the importance score value is: ; represents the importance score of the jth feature; M is the number of decision trees in the gradient boosting decision tree model; is the number of leaf nodes of the mth decision tree; is the sum of the sample weights of the t-th leaf node in the m-th decision tree; n is the number of samples; Represents the features used when splitting the t-th leaf node in the m-th decision tree; is an indicator function, when hour, ,otherwise ; Indicates the reduction in impurity of the i-th sample before and after the t-th split in the m-th decision tree; is the feature vector of the i-th sample at the t-th split in the m-th decision tree, is the true label of the i-th sample; Based on the target three-dimensional densely connected network model, the preoperative sacrum non-contrast computed tomography image of the target user, the three-dimensional region of interest image of the target user, the dynamic risk parameter of the target user, and the physiological attribute parameter of the target user are processed to generate sacrum warning information, including performing image data normalization processing on the preoperative sacrum non-contrast computed tomography image of the target user and the three-dimensional region of interest image of the target user to generate the mean of the image pixel value and the standard deviation of the image pixel value; performing parameter normalization processing on the dynamic risk parameter of the target user and the physiological attribute parameter of the target user to generate the mean of the dynamic risk parameter, the standard deviation of the dynamic risk parameter, the mean of the physiological attribute parameter, and the standard deviation of the physiological attribute parameter; performing fusion processing on the mean of the image pixel value, the standard deviation of the image pixel value, the mean of the dynamic risk parameter, the standard deviation of the dynamic risk parameter, the mean of the physiological attribute parameter, and the standard deviation of the physiological attribute parameter to generate multi-channel feature data; processing the multi-channel feature data based on the target three-dimensional densely connected network model to generate a target prediction value; processing the target prediction value based on a preset threshold to generate sacrum warning information; the calculation formula of the target prediction value is: ; L represents the total number of layers of the three-dimensional densely connected network model, Represents the weight parameter of the lth layer; represents the learning parameter vector of layer l, Represents the feature transformation operation performed by the lth layer on the input multi-channel feature data F; Represents the bias parameter of the lth layer; Represents the activation function of the output layer, wherein the sacrum warning information is used to characterize the attribute information of the target user's preoperative sacrum lesion.

2. The method according to claim 1, wherein Preprocessing the training sample set to generate a target training sample set includes: Performing data cleaning on the training sample set to generate an initial training sample set after removing invalid data; performing normalization on the initial training sample set to generate missing sample information; The missing sample information is processed to generate missing value interpolation prediction information; the missing value interpolation prediction information is processed to generate a target variable value; the target variable value is processed to generate target variable parameter information; and a target training sample set is generated based on the target variable parameter information.

3. The method according to claim 2, wherein The preset three-dimensional densely connected network model is processed based on the target training sample set and the verification sample set to generate a target three-dimensional densely connected network model, including: Obtain any number of data features in the target training sample set; generating a sampling ratio based on the number of each data feature in the target training sample set; Sampling the target training sample set based on the sampling ratio to generate a preset number of sampling features; Processing any data feature and each sampling feature to generate multiple data groups, wherein each data group includes a preset number of data samples, and at least one data sample includes identification information; Training the preset three-dimensional densely connected network model based on data samples in multiple data groups to generate a trained three-dimensional densely connected network model; Processing the trained three-dimensional densely connected network model based on the verification sample set to generate a verification result; If the data sample containing identification information in the verification result is a risk factor representing a risk factor affecting sacral lesions, the trained three-dimensional densely connected network model is used as the target three-dimensional densely connected network model.

4. The method according to claim 1, wherein Processing a preoperative non-contrast computed tomography image of the sacrum of the target user to generate a three-dimensional region of interest image of the target user includes: Acquiring preset anatomical structure restriction information, wherein the preset anatomical structure restriction information is used to represent organ information matching the sacrum structure; performing image resampling processing on the preoperative sacrum non-contrast computed tomography image of the target user to generate a sacrum non-contrast computed tomography image of a target size; processing the sacrum non-contrast computed tomography image of the target size to generate cropped target tumor region information; The cropped target tumor region information is processed based on the preset anatomical structure restriction information to generate a three-dimensional region-of-interest image of the target user.

5. The method according to claim 1, wherein Processing the multimodal data information of the target user and the target diagnostic interval data information of the target user to generate a dynamic risk parameter of the target user includes: performing feature extraction processing on the T1-weighted imaging information and the T2-weighted imaging information of the target user to generate signal intensity features, texture features, shape features, and enhancement curve features; performing feature extraction processing on the PET-CT information to generate target uptake value features; Performing feature quantification processing on the target diagnosis interval data information of the target user to generate access time interval information and lesion size growth rate value; Processing the signal intensity feature, the texture feature, the shape feature, the enhancement curve feature, and the target uptake value feature to generate a target shape factor; The target shape factor is processed based on the access time interval information and the lesion size growth rate value to generate a dynamic risk parameter of the target user.

6. A preoperative warning device for sacral tumors, characterized in that: For implementing the method of claim 1, the apparatus comprises: an acquisition module, configured to acquire clinical data information of a target user, a preoperative non-contrast computed tomography image of the sacrum of the target user, multimodal data information of the target user, target diagnostic interval data information of the target user, a preset three-dimensional densely connected network model, a training sample set, and a validation sample set, wherein the multimodal data information of the target user includes T1-weighted imaging information of the target user, T2-weighted imaging information of the target user, and positron emission tomography-computed tomography information; A processing module is used to preprocess the training sample set to generate a target training sample set; process the preset three-dimensional densely connected network model based on the target training sample set and the verification sample set to generate a target three-dimensional densely connected network model; process the preoperative non-contrast computed tomography image of the sacrum of the target user to generate a three-dimensional region of interest image of the target user; process the multimodal data information of the target user and the target diagnostic interval data information of the target user to generate the dynamic risk parameters of the target user; process the clinical data information of the target user to generate the physiological attribute parameters of the target user; process the preoperative non-contrast computed tomography image of the sacrum of the target user, the three-dimensional region of interest image of the target user, the dynamic risk parameters of the target user and the physiological attribute parameters of the target user based on the target three-dimensional densely connected network model to generate sacrum warning information, wherein the sacrum warning information is used to characterize the attribute information of the preoperative sacral lesions of the target user.

7. An electronic device, characterized in that: include: a first processor; and a memory for storing executable instructions of the first processor; Wherein, the first processor is configured to execute the preoperative warning method for sacral tumors according to any one of claims 1 to 5 by executing the executable instructions.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the second processor, the preoperative early warning method for sacral tumors according to any one of claims 1 to 5 is implemented.

Images