A Pelvis and Sacral Tumor Early Warning Method and Device Based on an End-to-End Deep Learning Model for Non-Enhanced MRI

Abstract

The present invention provides a method and device for early warning of pelvic and sacral tumors based on a non-enhanced MRI end-to-end deep learning model, which are applied to the field of data processing technology. This application preprocesses the preoperative magnetic resonance imaging sequence image information of the target user to generate target MRI sequence image information; processes the tumor status influencing factors of the pelvis and sacrum and the tumor risk adjustment factors of the pelvis and sacrum to generate the dynamic risk parameter information of the target user; processes a preset end-to-end deep learning model based on non-enhanced magnetic resonance imaging based on a training sample set and a validation sample set to generate a target end-to-end deep learning model; processes the target MRI sequence image information, the physiological parameter information of the target user, and the dynamic risk parameter information of the target user based on the target end-to-end deep learning model to generate user early warning information.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and particularly to a method and device for early warning of pelvic and sacral tumors based on a non-enhanced MRI end-to-end deep learning model. Background Art

[0002] Pelvic and sacral tumors (PSTs) are relatively rare, and metastatic tumors are the most common type due to the hematopoietic function in this region. Since PSTs are relatively rare and their clinical manifestations and imaging features are similar, radiologists often lack sufficient clinical experience to make a judgment. In the early stage, PSTs are usually small and asymptomatic, and once detected, they usually have grown large and compress surrounding organs, often requiring surgical intervention. For all primary malignant sacral tumors and benign lesions involving the lower segment, when the S3 nerve root can be preserved, radical resection should be selected. However, due to the complex anatomical structure, the influence of various surrounding organs, and the surgical difficulty in this region.

[0003] In recent years, deep learning models have shown great potential in exploring the nature of tumors. However, in the prior art, there is no early warning of benign and malignant bone tumors based on deep learning models. Against this background, developing sensitive and effective computer-aided early warning tools has become the key.

[0004] It should be noted that the information disclosed in the above background art section is only used to enhance the understanding of the background of the present disclosure, and thus includes information that does not constitute the prior art known to those of ordinary skill in the art. Summary of the Invention

[0005] The purpose of the present application is to provide a method and device for early warning of pelvic and sacral tumors based on a non-enhanced MRI end-to-end deep learning model, which at least to some extent overcomes the problems existing in the prior art. By collecting multi-dimensional data of the target user, covering multiple sequences of preoperative magnetic resonance imaging, physiological parameters, serological markers, microbial communities, and lifestyle information, and at the same time preparing a deep learning model and a sample set based on non-enhanced MRI. Then, the information is finely processed. The imaging sequence is normalized and filled to form a three-channel image to obtain the target MRI sequence image information; the serological and microbial community information is extracted and fused into a tumor status influencing factor, and the lifestyle information is refined into a risk adjustment factor, and the two are combined to obtain dynamic risk parameter information. Finally, the preprocessing sample set is used to train the preset model, the segmentation model is trained with the three-channel image, and the diagnostic model is trained in combination with other information, and early warning information is generated according to the threshold.

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

[0007] According to one aspect of the present application, there is provided a method for warning of pelvic and sacral tumors based on a non-enhanced MRI end-to-end deep learning model, including: obtaining the preoperative magnetic resonance imaging sequence image information of the target user, the physiological parameter information of the target user, the serological marker parameter information of the target user, the microbial community parameter information, the living habit information of the target user, a preset end-to-end deep learning model based on non-enhanced magnetic resonance imaging, a training sample set, and a validation sample set, wherein the preset end-to-end deep learning model based on non-enhanced magnetic resonance imaging includes a preset segmentation model and a preset diagnosis model, and the preoperative magnetic resonance imaging sequence image information of the target user includes the T1-weighted imaging information of the target user, the T2-weighted imaging information of the target user, and the diffusion-weighted imaging information of the target user; preprocessing the preoperative magnetic resonance imaging sequence image information of the target user to generate target MRI sequence image information; processing the serological marker parameter information and the microbial community parameter information of the target user to generate a tumor status influencing factor for the pelvis and sacrum; processing the living habit information of the target user to generate a tumor risk adjustment factor for the pelvis and sacrum; processing the tumor status influencing factor for the pelvis and sacrum and the tumor risk adjustment factor for the pelvis and sacrum to generate dynamic risk parameter information of the target user; processing the preset end-to-end deep learning model based on non-enhanced magnetic resonance imaging by using the training sample set and the validation sample set to generate a target end-to-end deep learning model; and processing the target MRI sequence image information, the physiological parameter information of the target user, and the dynamic risk parameter information of the target user by using the target end-to-end deep learning model to generate user warning information, wherein the user warning information is used to characterize the attribute information of the preoperative pelvic and sacral tumors of the target user.

[0008] Another aspect of the present application is a pelvic and sacral tumor warning device based on a non-enhanced MRI end-to-end deep learning model, which is characterized by including: an acquisition module, configured to acquire preoperative magnetic resonance imaging sequence image information of a target user, physiological parameter information of the target user, serological marker parameter information of the target user, microbial community parameter information, lifestyle information of the target user, a preset end-to-end deep learning model based on non-enhanced magnetic resonance imaging, a training sample set, and a validation sample set. Among them, the preset end-to-end deep learning model based on non-enhanced magnetic resonance imaging includes a preset segmentation model and a preset diagnosis model. The preoperative magnetic resonance imaging sequence image information of the target user includes T1-weighted imaging information of the target user, T2-weighted imaging information of the target user, and diffusion-weighted imaging information of the target user; a processing module, configured to preprocess the preoperative magnetic resonance imaging sequence image information of the target user to generate target MRI sequence image information; process the serological marker parameter information and the microbial community parameter information of the target user to generate a tumor status influencing factor for the pelvis and sacrum; process the lifestyle information of the target user to generate a tumor risk adjustment factor for the pelvis and sacrum; process the tumor status influencing factor and the tumor risk adjustment factor for the pelvis and sacrum to generate dynamic risk parameter information of the target user; process the preset end-to-end deep learning model based on non-enhanced magnetic resonance imaging based on the training sample set and the validation sample set to generate a target end-to-end deep learning model; process the target MRI sequence image information, the physiological parameter information of the target user, and the dynamic risk parameter information of the target user based on the target end-to-end deep learning model to generate user warning information, where the user warning information is used to characterize the attribute information of the preoperative pelvic and sacral tumors of the target user.

[0009] According to yet another aspect of the present application, there is provided a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a second processor, it implements the above-mentioned pelvic and sacral tumor warning method based on a non-enhanced MRI end-to-end deep learning model.

[0010] A method and device for early warning of pelvic and sacral tumors based on an end-to-end deep learning model of non-enhanced MRI. The server collects multi-dimensional data of the target user, covering multiple sequences of preoperative magnetic resonance imaging, physiological parameters, serological markers, microbial communities, and lifestyle information. At the same time, a deep learning model and a sample set based on non-enhanced MRI are prepared. Then, the information is finely processed. The imaging sequence is normalized and filled to form a three-channel image, and the target MRI sequence image information is obtained; the serological and microbial community information is extracted and fused into a tumor status influencing factor, and the lifestyle information is refined into a risk adjustment factor. The two are combined to obtain dynamic risk parameter information. Finally, the preprocessing sample set is used to train the preset model. The segmentation model is trained with the three-channel image, and the diagnostic model is trained in combination with other information. Early warning information is generated according to the threshold.

[0011] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present disclosure. Brief Description of the Drawings

[0012] Figure 1 The flowchart of a method for early warning of pelvic and sacral tumors based on an end-to-end deep learning model of non-enhanced MRI provided by an embodiment of the present application is shown;

[0013] Figure 2 The structural schematic diagram of a device for early warning of pelvic and sacral tumors based on an end-to-end deep learning model of non-enhanced MRI provided by an embodiment of the present application is shown. Detailed Embodiments

[0014] The following describes the preferred embodiments of the present invention 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 combines Figure 1 to describe a method for early warning of pelvic and sacral tumors based on an end-to-end deep learning model of non-enhanced MRI according to an exemplary embodiment of the present application. In one embodiment, the present application also proposes a method and device for early warning of pelvic and sacral tumors based on an end-to-end deep learning model of non-enhanced MRI. Figure 1 The flowchart of a method for early warning of pelvic and sacral tumors based on an end-to-end deep learning model of non-enhanced MRI according to an embodiment of the present application is schematically shown. As Figure 1 shown, the method is applied to a server and includes:

[0016] S101, obtain the pre-operative magnetic resonance imaging sequence image information of the target user, the physiological parameter information of the target user, the serological marker parameter information of the target user, the microbial community parameter information, the lifestyle information of the target user, a preset end-to-end deep learning model based on non-enhanced magnetic resonance imaging, a training sample set, and a validation sample set.

[0017] In one implementation, the pre-operative magnetic resonance imaging sequence image information of the target user includes T1-weighted imaging information, T2-weighted imaging information, and diffusion-weighted imaging (DWI) information. T1-weighted imaging shows anatomical structures well and can clearly present basic structural information such as the skeletal morphology of the pelvis and sacrum and the soft tissue contour, which helps to determine the location of the tumor and its anatomical relationship with the surrounding normal tissues; T2-weighted imaging is more sensitive to edema, inflammation, etc. in diseased tissues and can prominently display the tumor tissue and the edema area around it, providing a basis for judging the scope and invasion degree of the tumor; diffusion-weighted imaging (DWI) can reflect the diffusion movement of water molecules in tissues and is of great significance for detecting regions with restricted water molecule diffusion in tumor tissues, helping to identify the active part of the tumor because the region with dense tumor cells usually has restricted water molecule diffusion and appears as a high signal on DWI. The information provided by these different imaging methods complements each other, providing rich data for a comprehensive understanding of the morphology, location, size, and internal characteristics of the tumor.

[0018] The physiological parameter information of the target users includes aspects such as gender, age, tumor location, and maximum tumor size. Gender information has certain significance in tumor research. There are differences in the incidence rate, type of onset, or prognosis of certain tumors in different genders. For example, men are more likely to suffer from certain malignant pelvic and sacral tumors; age is an important risk factor. Generally speaking, as age increases, the risk of developing tumors also increases accordingly, and different types of tumors are more prevalent in different age groups. For example, primary bone tumors are more likely to occur in children and adolescents, while the proportion of metastatic tumors is relatively higher in adults; tumor location is crucial for diagnosis and treatment. Tumors in different locations have different relationships with surrounding organs and tissues, affecting the difficulty of surgical resection and prognosis. For example, tumors located in the sacrum are close to nerve structures, and more caution is needed during surgery to avoid nerve damage; the maximum tumor size directly reflects the growth degree of the tumor. Larger tumors are more likely to cause symptoms, such as compressing surrounding organs and causing dysfunction, and also indicate stronger tumor invasiveness or a longer disease course. These physiological parameter information is usually obtained through clinical examinations, imaging measurements, and medical record keeping during the user's medical treatment. For age, the date of birth or exact age should be recorded. The tumor location should be described using a unified anatomical nomenclature and positioning method. The maximum tumor size requires clear measurement criteria and methods (such as diameter, volume, etc.). During the data collation process, the data is standardized, and gender is encoded as specific numerical values (such as 0 for female and 1 for male) for easy computer processing and model analysis.

[0019] The serological marker parameter information mainly involves specific markers related to pelvic and sacral tumors, such as bone alkaline phosphatase (BALP) and lactate dehydrogenase (LDH). BALP is of great significance in skeletal diseases, especially bone tumors. Its level change is closely related to the activity of osteoblasts. In many bone tumors (such as osteosarcoma, osteoblastoma, etc.), due to the stimulation of osteoblast activity by tumor cells, the BALP level usually increases. Therefore, it can be used as an important indicator to evaluate the bone metabolism status and tumor activity of tumors; LDH is an enzyme widely present in human tissues. In tumor cells, due to abnormal active metabolism, the production and release of LDH increase, resulting in an increase in serum LDH level. In pelvic and sacral tumors (especially malignant tumors such as Ewing's sarcoma, etc.), an increase in LDH is relatively common. Its level is related to tumor burden, proliferation activity, and prognosis, and can be used to monitor tumor progression and treatment response.

[0020] The microbial community parameter information reflects the characteristics of the microbial community in the target user's body related to pelvic and sacral tumors. In the pelvic and sacral regions, the microbial community affects the tumor microenvironment through various pathways, such as regulating local immune responses, influencing cell signaling, and participating in metabolic processes. Some microorganisms produce toxins or metabolites that directly damage cellular DNA, promoting the proliferation and transformation of tumor cells; or by regulating the host immune system, inhibiting the immune surveillance function, creating favorable conditions for the growth and escape of tumor cells. Therefore, analyzing the microbial community parameter information helps to deeply understand the pathogenesis of tumors. High-throughput sequencing technologies, such as 16S rRNA sequencing and metagenomic sequencing, are usually used to obtain the microbial community parameter information. These technologies can comprehensively and deeply analyze the characteristics of the composition structure of the microbial community (including the relative abundances of different microbial species, diversity indices, etc.), gene expression activity (reflecting the functional state of microorganisms), and metabolic pathway activity (revealing the role of microorganisms in metabolic processes), etc.

[0021] The lifestyle information of the target user covers multiple aspects such as dietary intake, exercise type, exercise intensity, and sleep quality, and these factors are closely related to the risk of pelvic and sacral tumors. In terms of dietary intake, a high-sugar, high-fat, and high-salt diet increases the risk of chronic diseases such as obesity and diabetes, which in turn indirectly affects the occurrence and development of tumors. For example, long-term excessive intake of red meat and processed meat is related to the occurrence of certain tumors (such as colorectal cancer), and colorectal cancer can metastasize to the pelvis and sacrum; the type and intensity of exercise have important effects on the body's functional state, hormone regulation, and skeletal muscle health. Long-term lack of exercise leads to problems such as slowed body metabolism, obesity, and decreased immune function, increasing the risk of tumor occurrence, while moderate exercise (such as weight-bearing exercise helps to increase bone density, and aerobic exercise can improve cardiopulmonary function and immunity) has a positive effect on preventing tumors; poor sleep quality (such as problems like long-term insomnia and sleep apnea) affects the body's endocrine system, immune system, and metabolic function, resulting in hormonal imbalance and increased inflammatory responses, thus creating favorable conditions for the occurrence of tumors.

[0022] The pre-set end-to-end deep learning model based on non-enhanced magnetic resonance imaging is the core tool of the entire research, including a pre-set segmentation model and a pre-set diagnosis model. The pre-set segmentation model is mainly used for automatically segmenting the tumor area in preoperative magnetic resonance imaging sequence images. Its structure is usually based on advanced deep learning algorithms, such as the U-Net architecture. U-Net has a unique encoder-decoder structure. The encoder part gradually extracts the feature information of the image through convolutional layers and pooling layers, gradually reducing the resolution of the image while increasing the number of feature channels to obtain the high-level semantic features of the image. The decoder part then gradually restores the feature map to the original image size through upsampling layers and convolutional layers, and fuses it with the feature maps of the corresponding levels in the encoder to obtain a more accurate segmentation result. This structure enables the model to effectively process tumor features of different scales and accurately identify the tumor boundary. The pre-set diagnosis model judges the benign and malignant nature of the tumor based on the segmentation results and other relevant information (such as physiological parameter information, dynamic risk parameter information, etc.). Its structure is based on advanced architectures such as the Transformer network and combines multi-instance learning strategies, which can comprehensively process multi-modal data, learn the complex relationships between data, and thus improve the accuracy of diagnosis.

[0023] The training sample set and the validation sample set are important data resources for constructing and evaluating deep learning models. These sample sets usually come from multiple medical institutions and cover a large amount of user data of pelvic and sacral tumors pathologically diagnosed as benign or malignant, including multi-faceted data such as preoperative magnetic resonance imaging sequence image information, physiological parameter information, serological biomarker parameter information, microbial community parameter information, and lifestyle information. During the construction of the sample set, it is necessary to strictly screen according to the pre-determined inclusion and exclusion criteria to ensure the representativeness of the samples and the data quality. The inclusion criteria include the discovery of a single pelvic and sacral lesion by MRI examination, complete preoperative MRI images (including T1-weighted, T2-weighted, DWI, and enhanced T1-weighted images, excluding incomplete enhanced MRI sequences), and pathological diagnosis of benign or malignant tumors (classified according to the WHO classification standard), etc.; the exclusion criteria involve cases such as multiple lesions, poor image quality (such as severe artifacts), and users with postoperative recurrence. By carefully constructing the sample set, rich, diverse, and high-quality data are provided for model training, enabling the model to learn the characteristic patterns of different types of tumors and the relationships between relevant factors.

[0024] S102, preprocess the preoperative magnetic resonance imaging sequence image information of the target user to generate target MRI sequence image information.

[0025] In one embodiment, the preoperative magnetic resonance imaging sequence image information of the target user is normalized and filled to generate image information of the target scale specification, wherein the image information of the target scale specification is used to characterize the size of the target user's T1-weighted imaging information, the target user's T2-weighted imaging information, and the target user's diffusion-weighted imaging information. The main purpose of the normalization process is to map the pixel values ​​of the target user's T1-weighted imaging, T2-weighted imaging, and diffusion-weighted imaging (DWI) information to a unified numerical range, so that the image data between different scanning sequences and different users are comparable. In MRI images, the range of pixel values ​​varies depending on device settings, scanning sequence parameters, and individual differences between users. For example, different MRI scanners produce different signal intensity values ​​for the same tissue, which interferes with subsequent image feature-based analysis and model training. Through normalization processing, these intensity changes caused by device and individual differences are eliminated, the pathological features in the image are highlighted, and the accuracy and consistency of image analysis are improved. For example, the histogram equalization method adjusts the histogram distribution of the image to make the contrast of the image more uniform and enhance the contrast between different tissues in the image, thereby more clearly showing the difference between the tumor area and the surrounding normal tissues.

[0026] Padding is mainly performed on images with a width or height less than 224. In deep learning models, input images are usually required to have uniform size specifications to ensure that the model can correctly process image data. For smaller images, if they are directly input into the model without padding, information loss or model errors will occur during processing. For example, in image analysis models based on convolutional neural networks (CNNs), the operations of convolutional layers and pooling layers rely on fixed-size receptive fields. If the image sizes are inconsistent, the consistency of the receptive fields will be destroyed, affecting the model's extraction and analysis of image features. Through padding, the image is expanded to a size that meets the model input requirements while maintaining the integrity of the image content and avoiding errors introduced by size mismatch.

[0027] The filling operation can be carried out in various ways, such as boundary filling (copying the pixel values of the image boundary), symmetric filling (symmetrically copying the pixel values with the image boundary as the axis of symmetry), or mean filling (filling with the average pixel value of the image), etc. When choosing the filling method, the impact on image features needs to be considered. Boundary filling will introduce some unrealistic information at the image edge. If the edge area of the image has little impact on the analysis, boundary filling is a simple and effective method. Symmetric filling maintains the symmetry of the image to a certain extent and is more suitable for some pelvic and sacral regions with symmetric structural features. Mean filling is relatively smoother and will not introduce obvious boundary effects, but it will slightly blur the image edge. For example, if the tumor is mainly located in the central region of the pelvis and sacrum, boundary filling has little impact on the analysis results; while if the tumor is close to the edge or precise analysis of edge features is required, symmetric filling or more complex filling strategies are more appropriate. The size of the filled image reaches the target scale specification (such as 224×224 or other sizes suitable for model input). At this time, the image information of the generated target scale specification can accurately represent the size of the T1-weighted imaging, T2-weighted imaging, and DWI imaging information, laying a foundation for subsequent image analysis and model processing.

[0028] Process the image information of the target scale specification to generate a three-channel image. Among them, the three-channel image is composed of the T1-weighted imaging information of the target user, the T2-weighted imaging information of the target user, and the diffusion-weighted imaging information of the target user. The generation of the three-channel image is to integrate the T1-weighted imaging, T2-weighted imaging, and DWI imaging information of the target scale specification after normalization and filling processing. Each channel corresponds to the information of one imaging modality. The purpose of doing this is to include the features of multiple MRI imaging modalities in one image, enabling the model to obtain information in the image from multiple perspectives, so as to more comprehensively and accurately identify and analyze the features of pelvic and sacral tumors. For example, the T1-weighted imaging channel provides the anatomical structure information of bones and soft tissues. The T2-weighted imaging channel is more sensitive to the edema and inflammation of diseased tissues, etc. The DWI imaging channel can reflect the diffusion movement of water molecules in tissues, which helps to judge the activity of tumor tissues. By integrating these three imaging informations in a three-channel image, the model comprehensively utilizes these complementary informations to improve the detection and diagnosis ability of tumors.

[0029] In the analysis of pelvic and sacral tumors, three-channel images have obvious advantages. Compared with single-channel images, they can provide richer information. For example, it is difficult for a single T1-weighted image to accurately distinguish the subtle differences between tumor tissue and surrounding tissue, especially when the tumor boundary is unclear or there is a mixture of multiple tissue types. After combining the T2-weighted and DWI imaging information, the model can more accurately outline the tumor boundary by comparing the signal intensity, texture features, and water molecule diffusion in the tumor area of different channels, judge the nature of the tumor (such as benign or malignant), evaluate the degree of tumor invasion (such as whether it invades surrounding nerves, blood vessels, etc.), and monitor the treatment effect of the tumor (such as the performance of changes in tumor tissue after treatment in different imaging channels). In addition, three-channel images also enhance the model's learning ability for image features. When dealing with multi-channel images, the deep learning model automatically learns the correlation and complementarity between different channels, extracts more representative feature vectors, thereby improving the diagnostic accuracy and generalization ability of the model. For example, the model learns that areas with low signal in T1-weighted imaging, high signal in T2-weighted imaging, and restricted water molecule diffusion in DWI imaging are more likely to be tumor tissue. This feature learning based on multi-channel information can more accurately identify tumor features and reduce misjudgment.

[0030] The target MRI sequence image information generated based on three-channel images is the final form of image data obtained after a series of preprocessing steps, which will be used as the input data for subsequent deep learning models (such as a preset segmentation model and a preset diagnosis model). These preprocessing steps include normalization, padding, and three-channel image generation, aiming to convert the original preoperative MRI sequence image information into a format more suitable for model processing while retaining the key information in the images. The target MRI sequence image information contains rich tumor-related features, which can provide comprehensive image data support for the model, helping the model accurately identify tumor regions, analyze tumor features, and perform tasks such as judging the benign and malignant nature of tumors. The generated target MRI sequence image information is highly adaptable to the input requirements of the preset end-to-end deep learning model based on non-enhanced magnetic resonance imaging. The model is usually designed to receive image inputs of specific sizes, formats, and data types, and the target MRI sequence image information after preprocessing exactly meets these requirements. For example, the model requires the input image size to be 224×224×3 (width × height × number of channels), and the data type to be normalized values (such as float32). Through the previous normalization, padding, and three-channel image generation operations, it is ensured that the image data input to the model is consistent with the model's expectations in terms of size, format, and numerical range, enabling the model to effectively process the image data, extract valuable feature information, and then make accurate predictions and diagnoses. This adaptability is a key link to ensure the normal operation and accurate output of the entire deep learning model system, which can improve the training efficiency and diagnostic accuracy of the model, providing a reliable basis for the clinical diagnosis and treatment of pelvic and sacral tumors.

[0031] In the diagnosis and analysis of pelvic and sacral tumors, accurate and effective processing of preoperative magnetic resonance imaging (MRI) sequence image information is crucial. MRI images contain rich anatomical and pathological information, but due to factors such as imaging equipment, scanning parameters, and individual differences of users, the original image data varies in terms of size, intensity distribution, etc., which will bring difficulties to subsequent image analysis and model processing. Through operations such as normalization and padding processing and three-channel image generation, the image data is converted into a form more suitable for computer processing and model input, improving the consistency and recognizability of image features, thus providing strong support for accurate tumor diagnosis.

[0032] S103. Process the serological marker parameter information and the microbial community parameter information of the target user to generate an influencing factor for the tumor state of the pelvis and sacrum.

[0033] In one implementation, serological marker parameter information of a target user is subjected to feature extraction processing to generate target serological marker feature information, where the target serological marker feature information includes bone alkaline phosphatase (BALP) and lactate dehydrogenase (LDH). BALP plays a key role in skeletal diseases, especially bone tumors. Its level change is closely related to osteoblast activity. In many bone tumors (such as osteosarcoma, osteoblastoma, etc.), the proliferation and activity of tumor cells will stimulate an increase in osteoblast activity, resulting in an increase in the BALP level. BALP can be used as an important indicator to evaluate the tumor bone metabolism status and tumor activity. LDH is an enzyme widely present in human tissues. In tumor cells, due to abnormal active metabolism, the production and release of LDH increase significantly, leading to an increase in the serum LDH level. In pelvic and sacral tumors (especially malignant tumors such as Ewing sarcoma, etc.), an increase in LDH is relatively common, and its level is closely related to tumor burden, proliferation activity, and prognosis, and can be used to monitor tumor progression and treatment response.

[0034] For the detected BALP and LDH data, feature extraction is required for subsequent analysis. By calculating their absolute values, change rates, and the degree of deviation from the normal reference value range, etc. For example, calculate the concentration change rates of BALP and LDH at different time points to reflect the dynamic change process of the tumor. If the concentration of BALP or LDH gradually increases over time, it indicates that the tumor is in the progressive stage; conversely, if the concentration decreases, it means that the treatment is effective or the tumor is in the remission stage. Also analyze the correlation between BALP and LDH and other clinical indicators (such as certain indicators in blood routine, liver and kidney function indicators, etc.), and extract the correlation coefficient as a feature to more comprehensively reflect the relationship between the user's physiological state and the tumor. Quantify the extracted features, divide the levels of BALP and LDH into different grades, such as low, medium, and high risk grades, in order to intuitively reflect the tumor risk degree. At the same time, also comprehensively quantify and evaluate the serological marker features in combination with other clinical information of the user (such as age, gender, tumor type, etc.) to improve the accuracy and clinical practicality of the feature information. For example, for users with older age and higher BALP level, a higher risk weight is assigned because age itself is also a risk factor for tumor occurrence, and together with the increase in BALP level, it indicates a higher tumor risk.

[0035] Feature extraction is performed on the microbial community parameter information to generate the gene expression activity feature information of the microbial community and the composition structure feature information of the microbial community. In the pelvic and sacral regions, the microbial community can affect the tumor microenvironment through multiple pathways. For example, certain microorganisms can produce toxins or metabolites that directly damage cellular DNA and promote the proliferation and transformation of tumor cells; the microbial community also regulates the host immune system and inhibits the immune surveillance function, thus creating favorable conditions for the growth and escape of tumor cells. In addition, changes in the composition and function of the microbial community affect the metabolic environment of tumors and promote the survival and proliferation of tumor cells. High-throughput sequencing technologies, such as 16S rRNA sequencing and metagenomic sequencing, are usually used to obtain microbial community parameter information. 16S rRNA sequencing is mainly used to analyze the composition structure of the microbial community. By determining the 16S rRNA gene sequence, different microbial species are identified and their relative abundances are determined. Metagenomic sequencing can more comprehensively reveal the gene composition and functional potential of the microbial community, not only identifying known microorganisms but also discovering new microbial species and gene functions. These technologies can provide rich data to deeply analyze the characteristics of the composition structure of the microbial community (including the relative abundances of different microbial species, diversity indices, etc.), gene expression activity (reflecting the functional state of microorganisms), and metabolic pathway activity (revealing the role of microorganisms in the metabolic process), providing strong support for studying the relationship between the microbial community and tumors.

[0036] Feature fusion is performed on the target serological marker feature information, the gene expression activity feature information of the microbial community, and the composition structure feature information of the microbial community to generate a comprehensive feature vector. The serological marker feature information (such as the features related to BALP and LDH), the gene expression activity feature information of the microbial community, and the composition structure feature information each reflect factors related to pelvic and sacral tumors from different perspectives. However, when these information are used alone, they cannot comprehensively and accurately evaluate the tumor state. Through feature fusion, these information from different sources are integrated together, enabling the model to learn the synergistic effects and comprehensive impacts among them, thereby more comprehensively depicting the complex biological characteristics of tumors. For example, an elevated BALP level indicates enhanced osteogenic activity in bone tumors, but combined with the gene expression activity features related to immune regulation in the microbial community, it reveals the association between the immune state changes and abnormal bone metabolism in the tumor microenvironment. This comprehensive information is more crucial for accurately judging the tumor state. PCA is a dimensionality reduction technique that projects high-dimensional features into a low-dimensional space through linear transformation of the original features while maximizing the retention of the variance information of the data. In this process, feature fusion and dimensionality reduction are achieved, reducing data redundancy and improving the model processing efficiency.

[0037] Suppose the serological marker feature vector is (where indicating BALP-related features, indicating LDH-related features), and the gene expression activity feature vector of the microbial community is ( indicating the expression activity feature of the i-th gene), and the composition structure feature vector of the microbial community is ( indicating composition structure features such as the relative abundance of the j-th microorganism). First, determine the weight vector , which respectively correspond to the weights of serological marker features, gene expression activity features, and composition structure features. Then, the comprehensive feature vector F is calculated by the following formula: .

[0038] Process the comprehensive feature vector to generate feature impact relevance information, and generate a tumor status impact factor for the pelvis and sacrum based on the feature impact relevance information. To process the fused comprehensive feature vector to generate feature impact relevance information, data analysis and modeling methods need to be adopted. For example, a prediction model is constructed based on machine learning algorithms such as logistic regression, decision tree, and neural network. The model learns the complex non-linear relationship between the comprehensive feature vector and the tumor status (such as benign or malignant, tumor stage, etc.), thereby revealing the relevance between features and the degree of their impact on the tumor status. For example, the logistic regression model analyzes the contribution size (represented by the regression coefficient) of each feature in predicting the tumor status by establishing a mathematical relationship between the feature and the probability of tumor occurrence, and then determines the feature impact relevance. The decision tree model intuitively shows how features affect the classification of the tumor status through different branch decisions. Each node represents a feature, and the branches represent different value ranges of the feature. By analyzing the structure of the tree and the node splitting rules, the interaction between features and the order of their impact on the tumor status can be understood.

[0039] In the logistic regression model, a positive regression coefficient indicates a positive correlation between the feature and tumor occurrence or progression. The larger the absolute value of the coefficient, the stronger its influence. A negative coefficient indicates a negative correlation. In the decision tree model, the hierarchical position and branching order of features in the tree reflect their importance and sequence for classifying tumor status. Features closer to the root node usually have a greater impact on the classification result. This associative information helps researchers deeply understand how serological markers and microbial community characteristics jointly affect tumor status. For example, it is found that there is a synergistic effect between BALP and specific microbial community composition characteristics. When the BALP level increases and the relative abundance of a certain beneficial microorganism decreases, the risk of tumor occurrence increases significantly. The tumor status influencing factor of the pelvis and sacrum is a comprehensive index used to quantitatively reflect the overall influence degree of serological markers and microbial community-related characteristics on tumor status. It integrates the feature influence associative information obtained from the previous analysis, converting multiple complex feature relationships into a single value or vector for more intuitive assessment of tumor status risk. This factor serves as an important reference basis for judging tumor occurrence, development, treatment response, and prognosis.

[0040] S104, processes the lifestyle information of the target user to generate a tumor risk adjustment factor for the pelvis and sacrum.

[0041] In one implementation, feature extraction processing is performed on the lifestyle information of the target user to generate dietary intake features, exercise type features, exercise intensity features, and sleep quality features. Analyze the impact of the intake balance of nutrients in the diet on the risk of pelvis and sacrum tumors. For example, insufficient intake of calcium and vitamin D affects bone health and is related to the occurrence and development of certain bone tumors. The lack of antioxidant nutrients such as vitamin C, vitamin E, and selenium leads to increased oxidative stress, damages cellular DNA, and increases the risk of tumor occurrence. Weight-bearing exercise helps increase bone density, enhance the strength and stability of bones, and has a positive effect on preventing osteoporosis and related tumors (such as giant cell tumors of bone, etc.) in the pelvis and sacrum. Aerobic exercise improves cardiopulmonary function, promotes blood circulation, and enhances immune system function, helping to reduce the risk of tumor occurrence. Flexibility exercise helps maintain joint mobility, reduce muscle tension, improve the overall flexibility of the body, and is helpful for preventing pelvis and sacrum problems caused by poor posture or musculoskeletal imbalance. When extracting features, record the situation of the target user's participation in different types of exercise, including information such as exercise items, exercise frequency, and duration. Poor long-term sleep quality leads to hormonal imbalance, affects the normal secretion of growth hormone, which plays an important role in bone growth and repair, and its abnormality is related to the occurrence of pelvis and sacrum tumors. At the same time, lack of sleep or poor sleep quality weakens immune system function, reduces the body's immune surveillance and clearance ability of tumor cells, and increases the risk of tumor occurrence.

[0042] Process the dietary intake characteristics, exercise type characteristics, exercise intensity characteristics, and sleep quality characteristics to generate a dietary risk value, an exercise risk value, a sleep quality value, and a dynamic feature weight value. Among them, the dynamic feature weight value is generated based on the lifestyle information of different target users within a preset time period. Classify the calculated dietary risk value, such as into low-risk, medium-risk, and high-risk levels, so as to intuitively evaluate the impact degree of diet on the risk of pelvic and sacral tumors. During the classification process, consider individual difference factors, such as the influence of age, gender, basic health status, etc. on the dietary risk. For example, due to the decline in digestive function in the elderly, their tolerance and nutritional absorption ability for certain foods are different, and the risk under the same dietary pattern is higher than that of the young; the changes in the demand for calcium and vitamin D in women during specific physiological periods (such as menopause) affect the dietary risk assessment. According to these individual differences, appropriately adjust the dietary risk value to make the evaluation result more accurately reflect the actual risk status of different individuals.

[0043] Consider the influence of individual factors on the exercise risk value, such as age, gender, physical condition (whether there are skeletal joint diseases, cardiovascular diseases, etc.), and exercise habit history. The risk of the elderly during high-intensity exercise is relatively high, and the risk score related to exercise intensity should be adjusted according to their age; the exercise risk changes during special periods such as menstruation or pregnancy in women, and the evaluation model needs to be adjusted accordingly; for users with existing skeletal joint diseases (such as osteoporosis, arthritis, etc.) or cardiovascular diseases, certain exercise types or intensities are not suitable, and these limiting factors need to be considered when calculating the exercise risk value to avoid aggravating the condition or increasing the tumor risk due to exercise. By comprehensively considering individual factors, the exercise risk value can more accurately reflect the actual situation of each target user, providing a basis for personalized exercise recommendations. Consider the influence of sleep environment and habit factors on the sleep quality value and tumor risk. Adverse sleep environment factors such as noise, light, and temperature are given a certain risk score according to their interference degree. The sleep quality value of users who have been in an adverse sleep environment for a long time decreases accordingly, and the tumor risk increases. Bad sleep habits such as using electronic devices before going to bed and irregular work and rest also affect sleep quality. By investigating and understanding the sleep habits of target users, appropriately adjust the calculation of the sleep quality value for users with bad habits to more comprehensively evaluate the relationship between sleep quality and the risk of pelvic and sacral tumors.

[0044] The calculated dynamic feature weight values are applied to the generation process of the pelvic and sacral tumor risk adjustment factor. Multiply the diet risk value, exercise risk value, and sleep quality value by their corresponding dynamic feature weight values respectively, and then perform weighted summation to obtain the final tumor risk adjustment factor. In this way, the risk adjustment factor can more accurately reflect the comprehensive impact of different lifestyle characteristics on tumor risk at different time points. For example, if a user's recent exercise risk value increases due to a change in exercise habits and the dynamic weight value of the exercise feature is relatively high, then when calculating the risk adjustment factor, the exercise risk value will contribute more to the overall risk, thus highlighting the impact of the recent change in exercise habits on the risk of pelvic and sacral tumors and providing a more targeted basis for personalized tumor risk assessment and intervention.

[0045] Process the diet risk value, exercise risk value, sleep quality value, and dynamic feature weight values to generate the tumor risk adjustment factor for the pelvis and sacrum. The tumor risk adjustment factor for the pelvis and sacrum is calculated by comprehensively considering the diet risk value, exercise risk value, sleep quality value, and their respective corresponding dynamic feature weight values. Assume the diet risk value is , the exercise risk value is , the sleep quality value is , the dynamic feature weight value of the diet risk is , the dynamic feature weight value of the exercise risk is , and the dynamic feature weight value of the sleep quality risk is . Then the calculation formula for the tumor risk adjustment factor is expressed as: . This formula integrates the risk information of different aspects of lifestyle, where the weight value reflects the relative importance of each factor to tumor risk within a specific time period. For example, if within a certain time period, the change in exercise habits is considered to have a more critical impact on tumor risk (i.e., is relatively large), then the exercise risk value will account for a larger proportion when calculating the risk adjustment factor.

[0046] S105. Process the tumor status impact factor for the pelvis and sacrum and the tumor risk adjustment factor for the pelvis and sacrum to generate the dynamic risk parameter information of the target user.

[0047] In one implementation, process the tumor status impact factor for the pelvis and sacrum and the tumor risk adjustment factor for the pelvis and sacrum to generate the tumor morphology change value within a preset time interval. This application includes a calculation formula for obtaining the tumor morphology change value, and the calculation formula is: . Among them, represents the tumor morphology change value, represents the volume of the tumor at time t, represents the initial tumor volume. represents the surface area of the tumor at time t, , represents the morphological irregularity of the tumor at time t, represents the initial morphological irregularity of the tumor, , , are the weight coefficients of volume change, surface area change, and morphological irregularity change, respectively.

[0048] Suppose a user's initial tumor volume is , the tumor volume at the end of the 1st month is , the tumor volume at the end of the 2nd month is , and the tumor volume at the end of the 3rd month is . The tumor size change rate in the 1st visit cycle (0 - 1 month) is . The tumor size change rate in the 2nd visit cycle (1 - 2 months) is . The tumor size change rate in the 3rd visit cycle (2 - 3 months) is . By calculating the tumor size change rates in different visit cycles, it is found that the tumor growth rate of this user accelerates in the 2nd visit cycle, and then slows down slightly in the 3rd visit cycle, but it is still in a growth state overall.

[0049] In the 1st visit cycle, through medical image analysis (such as MRI images), it is found that the tumor shape is approximately spherical and the surface is relatively smooth. By the 2nd visit cycle, a slight bulge appears on one side of the tumor, the overall shape changes to an oval, and the surface is no longer so smooth, with some small depressions. In the 3rd visit cycle, the tumor deforms further, the bulge becomes more obvious, forming an irregular shape similar to a dumbbell, and the depressions on the surface increase and deepen. To quantify these changes, shape descriptors are used, such as circularity (assuming the initial circularity is 0.9, it becomes 0.75 in the 2nd cycle, and drops to 0.6 in the 3rd cycle). The continuous decrease in circularity indicates that the tumor shape gradually deviates from a circle, becomes more complex and irregular, which means that the proliferation rate of tumor cells is inconsistent in local areas, the invasiveness increases, and it is more likely to infiltrate into surrounding tissues.

[0050] In the first visit cycle, the tumor boundary was clear, with neat edges and a distinct demarcation from the surrounding tissues. In the second visit cycle, the tumor boundary began to become blurred in some areas, showing some small burr-like protrusions, indicating that tumor cells began to infiltrate the surrounding tissues. In the third visit cycle, the blurred boundary area expanded, the burrs increased and became longer, and the demarcation between some areas and the surrounding tissues was difficult to clearly define. Calculate the boundary blurriness index (assuming the initial boundary blurriness was 0.15, which became 0.3 in the second cycle and rose to 0.45 in the third cycle). The increase in boundary blurriness indicates a decrease in the stability of the tumor boundary change, an increase in the invasiveness of tumor cells into the surrounding tissues, an increase in the difficulty of completely removing the tumor during surgery, and also an increase in the risk of tumor recurrence and metastasis.

[0051] In another implementation, the initial tumor surface area , the initial tumor morphological irregularity . After 3 months of observation, at time t, the measured tumor volume , the tumor surface area , the tumor morphological irregularity . At the same time, determine the weight coefficients , , according to the statistical analysis of a large amount of clinical data. First, calculate the volume change rate: ; the surface area change rate: ; the morphological irregularity change rate: . Then the tumor morphological change value . This value reflects the comprehensive degree of change in the tumor morphology of this user within these 3 months. A higher value indicates that the tumor has obvious changes in terms of volume, surface area, and morphological irregularity, and closer monitoring is required.

[0052] Changes in tumor volume are directly related to the growth or shrinkage of the tumor. In many pelvic and sacral tumors, rapid volume growth means that tumor cells proliferate actively and are more invasive. For example, malignant tumors (such as osteosarcoma, Ewing sarcoma, etc.) usually have a faster growth rate, while benign tumors (such as osteoid osteoma, osteochondroma, etc.) grow relatively slowly. By monitoring the volume change rate, the growth trend of the tumor can be discovered early and the treatment strategy can be adjusted in time. For example, more active treatment interventions (such as surgery, chemotherapy, radiotherapy, etc.) are required for rapidly growing tumors. Changes in tumor surface area reflect changes in tumor morphology. When the tumor surface becomes more irregular or diffuse, it indicates that the infiltration of tumor cells into surrounding tissues increases, which is of great significance for determining the stage and prognosis of the tumor. For example, in sacral tumors, if the tumor surface area increases rapidly and the morphology becomes complex, it will increase the difficulty of surgical resection and also indicate a poor prognosis for the user. Changes in morphological irregularity are closely related to the malignancy of the tumor. The more irregular the morphology, the higher the heterogeneity of the tumor cells and the more aggressive their biological behavior. For example, during the development of some highly malignant pelvic tumors, the morphological irregularity will gradually increase, showing features such as blurred boundaries, lobed or burr-like shapes, which makes the tumor more likely to invade surrounding important structures such as nerves and blood vessels.

[0053] For the tumor morphology change values ​​within the preset time interval information, characteristic parameter information of different access cycles and dynamic pelvic and sacral structural information are generated, among which the characteristic parameter information of different access cycles includes the rate of change of tumor size, the characteristics of tumor shape change, and the stability characteristics of tumor boundary change. At the initial examination, the bone structure of the pelvis and sacrum was intact, and no obvious bone destruction was found. The surrounding soft tissue morphology was normal, and the nerves and blood vessels were clear, without compression or displacement. The MRI images of the second access cycle showed that there was slight bone absorption in the contact area between the tumor and the sacral bone, which was manifested as a slight decrease in bone density, indicating that the tumor began to affect the bone. In terms of soft tissue, the muscle tissue around the tumor showed mild edema, and the nerves and blood vessels were not obviously compressed, but the vascular blood flow signal was slightly changed, which was a hemodynamic change caused by local metabolic changes of the tumor. In the third access cycle, the scope of sacral bone destruction expanded, the bone cortex became thinner, and small bone defects appeared locally. The nerves were compressed and deformed, the space around the nerves narrowed, and the signal intensity of some nerve fibers changed, indicating that the nerve function was affected. The blood vessels are compressed and displaced significantly, the blood vessel diameter becomes thinner, and the local blood flow rate slows down, which affects the blood supply to the tumor and the nutrient supply to the surrounding tissues, further affecting the growth of the tumor and the function of the surrounding tissues.

[0054] These dynamically changing pelvic and sacral structural information is very important for evaluating the progression of tumors and predicting the prognosis of users. For example, the aggravation of bone destruction affects the stability of the pelvis and increases the risk of fractures; nerve compression and functional changes cause symptoms such as lower limb pain, numbness, and weakness in users, seriously affecting the quality of life; blood vessel compression affects the uptake of chemotherapy drugs by tumors and the treatment effect, and also affects the metabolism and repair of surrounding normal tissues.

[0055] Process the characteristic parameter information of different access cycles and the structural information of the dynamic pelvis and sacrum to generate the dynamic risk parameter information of the target user. The weighted summation method is used to calculate the dynamic risk parameter information. Assume that the weight of the tumor size change rate is 0.4, the weight of the tumor shape change characteristic is 0.3, the weight of the tumor boundary change stability characteristic is 0.2, and the weight of the pelvic and sacral structure change is 0.1. Taking the 3rd access cycle as an example, the tumor size change rate in this cycle is , the shape change characteristic (circularity is 0.6), the boundary change stability characteristic (fuzziness is 0.45), and the pelvic and sacral structure change (the comprehensive score of bone destruction and soft tissue invasion is assumed to be 0.35). Then the dynamic risk parameter information . The larger the value of the dynamic risk parameter information, the higher the dynamic risk of the tumor in the access cycle of the user. By regularly calculating (such as calculating once a month or every two months) and monitoring the dynamic risk parameter information, the changing trend of the tumor risk can be grasped in real time. If the dynamic risk parameter information continues to rise, high vigilance should be exercised, such as increasing the dose of chemotherapy drugs, replacing more effective targeted drugs, or performing surgical intervention in a timely manner; if the dynamic risk parameter information remains relatively stable or decreases, the treatment can be continued and closely observed. At the same time, stratify and manage users according to the dynamic risk parameter information, and increase the follow-up frequency for high-risk users (such as changing from once every three months to once a month).

[0056] S106, process the preset end-to-end deep learning model based on non-enhanced magnetic resonance imaging by using the training sample set and the validation sample set to generate a target end-to-end deep learning model.

[0057] In one implementation, the training sample set is preprocessed to generate a training sample set in a target format, where the training sample set in the target format includes three-channel image information of other users, physiological parameter information of other users, and dynamic risk parameter information of other users; the collected training sample set is subjected to data cleaning to remove data records with errors or missing values. For example, if it is found that the age data is missing in the physiological parameter information of a certain user, or the serological marker test result is significantly abnormal and the accuracy cannot be verified, then this sample is excluded. For image data, check whether there are problems such as image damage and too low resolution to ensure that the image quality meets the requirements. Then, convert various types of data into a format suitable for model processing. For the physiological parameter information of users (such as gender, age, tumor location, maximum tumor size, etc.), it is numerically encoded. For example, 0 is used to represent female and 1 is used to represent male; the age is represented by the actual age value; the tumor location is encoded according to a predefined anatomical coding system; the maximum tumor size is uniformly converted into a value in cubic centimeters. For serological marker parameter information (such as bone alkaline phosphatase and lactate dehydrogenase, etc.), the detected concentration values are standardized so that their numerical ranges are within a reasonable interval for subsequent calculation and comparison. The microbial community parameter information (such as the microbial community composition structure and gene expression activity characteristics obtained by high-throughput sequencing, etc.) is converted into a vector form after data parsing and feature extraction. The lifestyle information (such as diet intake, exercise type, exercise intensity, sleep quality, etc.) is also quantified.

[0058] For example, the diet intake is converted into the nutrient intake value according to the food type and intake; the exercise type is assigned corresponding values according to the intensity of different exercises and the impact on the body; the exercise intensity is comprehensively calculated to obtain a quantified value through exercise duration, exercise frequency, and exercise intensity indicators (such as heart rate change, energy consumption, etc.); the sleep quality is comprehensively evaluated to obtain a quantified score according to factors such as sleep duration, sleep efficiency, sleep stage distribution, and sleep disorder conditions. Finally, the processed user data is combined into a training sample set in the target format according to the preset rules, ensuring that each sample contains three-channel image information of other users (integrated by T1-weighted imaging, T2-weighted imaging, and diffusion-weighted imaging information), physiological parameter information, and dynamic risk parameter information.

[0059] To avoid the bias caused by sample imbalance during model training, the training sample set is subjected to sample balancing processing. The number of samples of different categories (such as benign tumor and malignant tumor samples) is counted. If the number of samples of a certain category is found to be too small, oversampling techniques (such as the SMOTE algorithm) are used to increase the number of samples of that category, making the number of samples of each category relatively balanced. At the same time, to improve the generalization ability of the model, the training samples are enhanced. For image data, operations such as random rotation, flipping, and scaling are used to increase the diversity of the images. For example, the pelvic and sacral MRI images of a certain user are randomly rotated by 10 - 20 degrees, or flipped horizontally or vertically, while keeping the tumor feature information in the image unchanged, thus generating new training samples. For physiological parameter information and dynamic risk parameter information, the data changes can be simulated by adding random noise within a certain range, enabling the model to learn more robust feature patterns. The training sample set processed in this way can better represent the actual data distribution, which helps to train a deep learning model with better performance.

[0060] Based on the three-channel image information of other users, the preset segmentation model is processed to generate the segmentation result of the preset tumor lesion area. The preset segmentation model adopts the U-Net structure. Its encoder part extracts features from the input three-channel image information through a series of convolutional layers and pooling layers. The convolutional layers are used to extract local features in the image. For example, through a 3×3 convolutional kernel, the edge details of the tumor area can be detected, and a 5×5 convolutional kernel can capture the texture features of a wider area, etc. The pooling layer gradually reduces the image resolution in a 2×2 max-pooling manner, while expanding the receptive field, enabling the model to obtain more global image feature information. In each layer of the encoder, the number of feature channels gradually increases, so that more advanced and abstract image semantic features can be learned. For example, after convolution operation on the three-channel image input to the first layer of the encoder, 16 feature channels are generated, and the number of feature channels in subsequent layers becomes 32, 64, 128, etc. The decoder part gradually restores the low-resolution feature map to the original image size through upsampling layers and convolutional layers. The upsampling layer uses bilinear interpolation to increase the image resolution, and at the same time, the feature maps of the corresponding levels in the encoder are fused with the feature maps in the decoder through skip connections, so that the model can better utilize feature information of different scales and accurately identify the boundary of the tumor lesion area. For example, in a certain layer of the decoder, the upsampled feature map is first concatenated with the feature map of the corresponding resolution in the encoder, and then the features are further refined through a 3×3 convolutional layer, gradually restoring the accurate segmentation result of the tumor area.

[0061] Suppose the pelvic and sacral MRI images of a certain user are processed, and the segmentation result of the tumor lesion area is obtained after passing through the segmentation model. It is shown in a visual way of the image. On the original MRI image, the tumor area is marked out. For example, the general shape and scope of the tumor are clearly outlined with a red contour line. At the same time, evaluation indicators such as the Dice coefficient and the intersection over union (IoU) are calculated to quantify the accuracy of the segmentation result. If the actual area of the tumor area is 100 square pixels, the intersection of the tumor area segmented by the model and the actual area is 80 square pixels, and the union is 120 square pixels, then the Dice coefficient is 2×80 / (100 + 120)≈0.727, and the IoU is 80 / 120≈0.667. Generally speaking, the higher the Dice coefficient and the IoU value, the more accurate the segmentation result, and the stronger the model's ability to identify the tumor area.

[0062] Based on the segmentation result of the preset tumor lesion area, the physiological parameter information of other users, and the dynamic risk parameter information of other users, the preset diagnosis model is processed to generate a training result. Based on the validation sample set, the preset segmentation model and the preset diagnosis model are processed respectively to generate a validation result. If the identification information included in the validation result and the identification information included in the training result are risk factors characterizing the impact of sacral lesions, then the trained preset segmentation model and preset diagnosis model are used as the target end-to-end deep learning model. The validation sample set also undergoes similar preprocessing steps as the training sample set to ensure that the data format and quality meet the requirements. Then, the three-channel image information of the validation sample set is input into the preset segmentation model respectively to obtain the segmentation result of the tumor lesion area; the segmentation result, together with the physiological parameter information and dynamic risk parameter information in the validation sample set, is input into the preset diagnosis model to obtain a prediction result. For example, for a certain user in the validation set, the segmentation model accurately outlines the boundary of the tumor, and the diagnosis model predicts that the probability of the tumor being benign is 0.7 and the probability of being malignant is 0.3, and it is judged as a benign tumor user according to the threshold.

[0063] Calculate performance metrics such as accuracy, AUC (area under the receiver operating characteristic curve), sensitivity, and specificity in the verification results. Assume there are 80 users in the verification set, with 30 benign tumor users and 50 malignant tumor users in reality. The model correctly predicted 25 benign tumor users and 40 malignant tumor users. Then the accuracy is (25 + 40) / 80 = 0.8125. Calculate the AUC value by plotting the ROC curve. If the AUC is 0.85, it indicates that the model has good performance in differentiating between benign and malignant tumors. The sensitivity is 40 / 50 = 0.8, and the specificity is 25 / 30 ≈ 0.83. If these metrics in the verification results meet the expected requirements, and the identification information included in the verification results (such as the benign or malignant judgment result of the tumor, the risk level classification, etc.) is consistent with the identification information included in the training results in terms of characterizing the risk factors affecting sacral lesions, that is, the model performs stably and meets expectations on both the verification set and the training set, then the trained preset segmentation model and preset diagnosis model are used as the target end-to-end deep learning model. If the verification results are not ideal, for example, the accuracy is lower than the expected threshold, or there is overfitting (such as high accuracy on the training set but low accuracy on the verification set) or underfitting (such as low accuracy on both the training set and the verification set) on the verification set, then the model needs to be adjusted, such as adjusting the hyperparameters of the model (learning rate, number of iterations, network structure, etc.), optimizing the data augmentation method, increasing the amount of training data, etc., and then retraining and validating until a satisfactory target end-to-end deep learning model is obtained.

[0064] S107. Process the target MRI sequence image information, the physiological parameter information of the target user, and the dynamic risk parameter information of the target user based on the target end-to-end deep learning model to generate user warning information.

[0065] In one implementation, perform feature extraction processing on the physiological parameter information of the target user to generate the physiological feature information of the target user. The physiological feature information of the target user includes the gender feature information, age feature information, tumor location feature information, and maximum tumor size feature information of the target user. For the physiological parameter information of the target user, first perform feature extraction processing. Assume the target user is a 55-year-old male, with the tumor located on the right side of the sacrum, and the maximum tumor diameter is measured by imaging as 5 cm (the volume is approximately 65.45 cubic centimeters after calculation). The gender feature information is encoded as 1 (indicating male), and the age feature information directly uses the actual age of 55 years as the value. The tumor location feature information divides the sacrum into several regions according to a pre-set anatomical coding system. For example, the right sacrum is encoded as 3. The maximum tumor size feature information is in cubic centimeters, that is, 65.45. In this way, the physiological feature information vector [1, 55, 3, 65.45] of the target user is generated, and these features will be important inputs for subsequent model processing.

[0066] Process the target MRI sequence image information based on the target segmentation model to generate the segmentation result of the tumor lesion area of the target user. Process the three-channel MRI sequence image information of the target user based on the target segmentation model (such as a model using the U-Net architecture). The encoder part of the model gradually extracts image features through convolutional layers and pooling layers. For example, in the convolutional layer, different-sized convolutional kernels (such as 3×3, 5×5, etc.) are used to detect details such as the edges and textures of the tumor area. The pooling layer downsamples in a 2×2 manner to reduce the image resolution while expanding the receptive field to obtain more global image feature information. The decoder part then restores the low-resolution feature map to the original image size through upsampling layers and convolutional layers, and fuses feature maps at different levels through skip connections to accurately identify the boundary of the tumor lesion area.

[0067] After being processed by the segmentation model, the segmentation result of the tumor lesion area of the target user is obtained. Visualized as an image, on the original MRI image, the tumor area is clearly marked. For example, the shape and scope of the tumor are outlined with a green contour line. Assuming the size of the original MRI image is 512×512 pixels, information such as the position, shape, and size of the segmented tumor area in the image can be intuitively presented through the segmentation result. At the same time, evaluation metrics are calculated to quantify the accuracy of the segmentation result, such as the Dice coefficient and the intersection over union (IoU). If the actual area of the tumor area is 80 square pixels, the intersection of the tumor area segmented by the model and the actual area is 65 square pixels, and the union is 90 square pixels, then the Dice coefficient is and the IoU is . A higher Dice coefficient and IoU value indicate that the segmentation model accurately identifies the tumor area of this target user.

[0068] The tumor lesion area segmentation result, physiological characteristic information, and dynamic risk parameter information of the target user are processed based on the target diagnosis model to generate a target prediction value. Based on a preset threshold, the target prediction value is processed to generate a user warning message, which is used to characterize the attribute information of the target user's pre-operative pelvic and sacral tumors. The target diagnosis model receives the tumor lesion area segmentation result, physiological characteristic information, and dynamic risk parameter information as inputs. For the segmentation result, the model further extracts geometric features such as the area, perimeter, and shape complexity of the tumor area, as well as the position features of the tumor area in the image, and converts these features into vector form. The physiological characteristic information [1, 55, 3, 65.45] and the dynamic risk parameter information (assumed to be 0.45, which is obtained based on previous calculations and represents the dynamic risk level of the current tumor) are also input into the model. The model automatically learns the associations and importance weights between these different sources of information through the self-attention mechanism of the Transformer network, so as to comprehensively judge the malignancy of the tumor.

[0069] Assume that the influence weights of the tumor lesion area segmentation result features, physiological characteristic information, dynamic risk parameter information, and their interaction features on the prediction result in the model are , 2, , (These weights are obtained through model training). After model calculation, the target prediction value P is obtained. According to the preset calculation formula (assuming , , , represent specific function transformations of the segmentation result features, physiological characteristic information, dynamic risk parameter information, and interaction features respectively): Assume that the values after function transformation are , , , , is the sigmoid function, then .

[0070] Set the preset threshold to 0.5. Since , according to the model prediction, the tumor of the target user is more likely to be malignant. Therefore, the generated user warning message is "The pre-operative pelvic and sacral tumors may be malignant. Please further examine and evaluate." This warning message can help doctors and users understand the potential risks of tumors in advance so as to take corresponding treatment measures or adjust the treatment plan in a timely manner. If the predicted value is less than the preset threshold, the warning message may be "The pre-operative pelvic and sacral tumors tend to be benign, but regular reexaminations are still required." In this way, based on the processing results of the target end-to-end deep learning model, targeted pre-operative tumor attribute warning information is provided for the target user.

[0071] The server collects multi-dimensional information of the target user, including pre-operative magnetic resonance imaging sequence images (T1, T2, diffusion-weighted imaging), physiological parameters, serological markers, microbial communities, and living habits, etc., and prepares a preset end-to-end deep learning model and a sample set based on non-enhanced MRI. Subsequently, targeted processing is performed on each piece of information. The imaging sequence images are normalized, filled, and converted into three-channel images to generate target MRI sequence image information; features are extracted and fused from the serological and microbial community information to obtain tumor status influencing factors; features are extracted from the living habit information to calculate relevant risk values and weight values, generating tumor risk adjustment factors, and the two are integrated to obtain dynamic risk parameter information.

[0072] Next, the preset model is trained with the pre-processed training sample set. The segmentation model is trained with the three-channel images to obtain the segmentation results of the tumor lesion area. The diagnostic model is trained in combination with other information. After being verified by the verification sample set, the model that meets the risk factor identification of the sacral lesion becomes the target model. The prediction value is calculated by using the target model in combination with various types of information, and the user warning message is generated according to the preset threshold.

[0073] In one implementation, as Figure 2 shown, the present application also provides a pelvic and sacral tumor warning device based on a non-enhanced MRI end-to-end deep learning model, including:

[0074] An acquisition module 201, configured to acquire pre-operative magnetic resonance imaging sequence image information of the target user, physiological parameter information of the target user, serological marker parameter information of the target user, microbial community parameter information, living habit information of the target user, a preset end-to-end deep learning model based on non-enhanced magnetic resonance imaging, a training sample set, and a verification sample set. Among them, the preset end-to-end deep learning model based on non-enhanced magnetic resonance imaging includes a preset segmentation model and a preset diagnostic model. The pre-operative magnetic resonance imaging sequence image information of the target user includes T1-weighted imaging information of the target user, T2-weighted imaging information of the target user, and diffusion-weighted imaging information of the target user;

[0075] A processing module 202 is configured to preprocess the preoperative magnetic resonance imaging sequence image information of the target user to generate target MRI sequence image information; process the serological marker parameter information and the microbial community parameter information of the target user to generate a tumor status influencing factor for the pelvis and sacrum; process the lifestyle information of the target user to generate a tumor risk adjustment factor for the pelvis and sacrum; process the tumor status influencing factor for the pelvis and sacrum and the tumor risk adjustment factor for the pelvis and sacrum to generate dynamic risk parameter information of the target user; process the preset end-to-end deep learning model based on non-enhanced magnetic resonance imaging based on the training sample set and the validation sample set to generate a target end-to-end deep learning model; process the target MRI sequence image information, the physiological parameter information of the target user, and the dynamic risk parameter information of the target user based on the target end-to-end deep learning model to generate a user warning information, wherein the user warning information is used to characterize the attribute information of the preoperative pelvis and sacrum tumors of the target user.

[0076] Each embodiment in this application is described in a related manner. For the same and similar parts among the embodiments, reference can be made to each other. Each embodiment focuses on the differences from other embodiments. In particular, for the embodiments of the pelvis and sacrum tumor warning method, electronic device, electronic equipment, and readable storage medium for evaluating the end-to-end deep learning model based on non-enhanced MRI, since they are basically similar to the embodiments of the pelvis and sacrum tumor warning method based on the end-to-end deep learning model of non-enhanced MRI described above, the description is relatively simple. For the relevant parts, reference can be made to the partial description of the embodiments of the pelvis and sacrum tumor warning method based on the end-to-end deep learning model of non-enhanced MRI described above.

Claims

1. A method for warning of pelvic and sacral tumors based on an end-to-end deep learning model of non-enhanced MRI, characterized in that, Including: Obtaining the preoperative magnetic resonance imaging (MRI) sequence image information of the target user, the physiological parameter information of the target user, the serological biomarker parameter information of the target user, the microbial community parameter information, the living habit information of the target user, a preset end-to-end deep learning model based on non-enhanced MRI, a training sample set, and a validation sample set; Preprocessing the preoperative MRI sequence image information of the target user to generate target MRI sequence image information; Processing the serological biomarker parameter information and the microbial community parameter information of the target user to generate a tumor status influencing factor for the pelvis and sacrum; Processing the living habit information of the target user to generate a tumor risk adjustment factor for the pelvis and sacrum; Processing the tumor status influencing factor for the pelvis and sacrum and the tumor risk adjustment factor for the pelvis and sacrum to generate dynamic risk parameter information of the target user; Processing the preset end-to-end deep learning model based on non-enhanced MRI using the training sample set and the validation sample set to generate a target end-to-end deep learning model; Processing the target MRI sequence image information, the physiological parameter information of the target user, and the dynamic risk parameter information of the target user based on the target end-to-end deep learning model to generate user warning information, where the user warning information is used to characterize the attribute information of the preoperative pelvis and sacrum tumors of the target user.

2. The method according to claim 1, wherein The preoperative MRI sequence image information of the target user includes the T1-weighted imaging information of the target user, the T2-weighted imaging information of the target user, and the diffusion-weighted imaging information of the target user; Preprocessing the preoperative MRI sequence image information of the target user to generate target MRI sequence image information, including: Performing normalization and filling processing on the preoperative MRI sequence image information of the target user to generate image information of a target scale specification, where the image information of the target scale specification is used to characterize the size of the T1-weighted imaging information of the target user, the T2-weighted imaging information of the target user, and the diffusion-weighted imaging information of the target user; Processing the image information of the target scale specification to generate a three-channel image, where the three-channel image is composed of the T1-weighted imaging information of the target user, the T2-weighted imaging information of the target user, and the diffusion-weighted imaging information of the target user; Generating target MRI sequence image information based on the three-channel image.

3. The method according to claim 1, wherein Processing the serological biomarker parameter information and the microbial community parameter information of the target user to generate a tumor status influencing factor for the pelvis and sacrum, including: Performing feature extraction processing on the serological biomarker parameter information of the target user to generate target serological biomarker feature information, where the target serological biomarker feature information includes bone alkaline phosphatase and lactate dehydrogenase; Performing feature extraction processing on the microbial community parameter information to generate gene expression activity feature information of the microbial community and compositional structure feature information of the microbial community; Performing feature fusion processing on the target serological biomarker feature information, the gene expression activity feature information of the microbial community, and the compositional structure feature information of the microbial community to generate a comprehensive feature vector; Process the comprehensive feature vector to generate feature influence correlation information; Generate the tumor state influence factor of the pelvis and sacrum based on the feature influence correlation information.

4. The method according to claim 3, characterized in that Process the lifestyle information of the target user to generate the tumor risk adjustment factor of the pelvis and sacrum, including: Perform feature extraction processing on the lifestyle information of the target user to generate dietary intake features, exercise type features, exercise intensity features, and sleep quality features; Process the dietary intake features, exercise type features, exercise intensity features, and sleep quality features to generate dietary risk values, exercise risk values, sleep quality values, and dynamic feature weight values, where the dynamic feature weight values are generated based on the lifestyle information of different target users within a preset time period; Process the dietary risk values, exercise risk values, sleep quality values, and dynamic feature weight values to generate the tumor risk adjustment factor of the pelvis and sacrum.

5. The method according to claim 4, characterized in that, Process the tumor state influence factor of the pelvis and sacrum and the tumor risk adjustment factor of the pelvis and sacrum to generate the dynamic risk parameter information of the target user, including: Process the tumor state influence factor of the pelvis and sacrum and the tumor risk adjustment factor of the pelvis and sacrum to generate the tumor morphology change value within the preset time interval information; Process the tumor morphology change value within the preset time interval information to generate the feature parameter information of different access cycles and the dynamic structure information of the pelvis and sacrum, where the feature parameter information of different access cycles includes the change rate of tumor size, the change characteristics of tumor shape, and the stability characteristics of tumor boundary change; Process the feature parameter information of different access cycles and the dynamic structure information of the pelvis and sacrum to generate the dynamic risk parameter information of the target user; The method further includes a calculation formula for obtaining the tumor morphology change value, and the calculation formula is: ； Among them, represents the tumor morphological change value, represents the tumor volume at time t, represents the initial tumor volume, represents the tumor surface area at time t, , represents the tumor morphological irregularity at time t, represents the initial tumor morphological irregularity, , , are the weight coefficients of volume change, surface area change and morphological irregularity change respectively.

6. The method according to claim 1, wherein The preset end-to-end deep learning model based on non-enhanced magnetic resonance imaging includes a preset segmentation model and a preset diagnosis model; process the preset end-to-end deep learning model based on non-enhanced magnetic resonance imaging using a training sample set and a validation sample set to generate a target end-to-end deep learning model, including: Preprocess the training sample set to generate a training sample set in a target format, where the training sample set in the target format includes the three-channel image information of other users, the physiological parameter information of other users, and the dynamic risk parameter information of other users; Process the preset segmentation model based on the three-channel image information of other users to generate the segmentation result of the preset tumor lesion area; Process the preset diagnosis model based on the segmentation result of the preset tumor lesion area, the physiological parameter information of other users, and the dynamic risk parameter information of other users to generate a training result; Process the preset segmentation model and the preset diagnosis model respectively based on the validation sample set to generate a validation result; If the identification information included in the verification result and the identification information included in the training result are risk factors characterizing the risk of sacral lesions, then the trained preset segmentation model and preset diagnosis model are used as the target end-to-end deep learning model.

7. The method according to claim 5, wherein Based on the target end-to-end deep learning model, process the target MRI sequence image information, the physiological parameter information of the target user, and the dynamic risk parameter information of the target user to generate user warning information, including: Perform feature extraction processing on the physiological parameter information of the target user to generate the physiological feature information of the target user. Among them, the physiological feature information of the target user includes the gender feature information, age feature information, tumor location feature information, and maximum tumor size feature information of the target user; Based on the target segmentation model, process the target MRI sequence image information to generate the tumor lesion area segmentation result of the target user; Based on the target diagnosis model, process the tumor lesion area segmentation result of the target user, the physiological feature information of the target user, and the dynamic risk parameter information of the target user to generate a target prediction value; Based on a preset threshold, process the target prediction value to generate user warning information; The method includes a calculation formula for obtaining the target prediction value, and the calculation formula is: ; among them, is the sigmoid function, , , , are weight coefficients, respectively representing the characteristics of the tumor lesion area segmentation result , physiological characteristic information , dynamic risk parameter information and the interaction characteristics between them on the influence weight of the prediction result.

8. A pelvic and sacral tumor warning device based on an end-to-end deep learning model of non-enhanced MRI, which is used to implement the method described in claim 1, and is characterized in that, The device includes: An acquisition module, configured to acquire the preoperative magnetic resonance imaging sequence image information of the target user, the physiological parameter information of the target user, the serological biomarker parameter information of the target user, the microbial community parameter information, the living habit information of the target user, a preset end-to-end deep learning model based on non-enhanced magnetic resonance imaging, a training sample set, and a verification sample set. Among them, the preset end-to-end deep learning model based on non-enhanced magnetic resonance imaging includes a preset segmentation model and a preset diagnosis model. The preoperative magnetic resonance imaging sequence image information of the target user includes the T1-weighted imaging information of the target user, the T2-weighted imaging information of the target user, and the diffusion-weighted imaging information of the target user; A processing module, configured to preprocess the preoperative magnetic resonance imaging sequence image information of the target user to generate target MRI sequence image information; process the serological biomarker parameter information and microbial community parameter information of the target user to generate tumor state influencing factors for the pelvis and sacrum; process the living habit information of the target user to generate tumor risk adjustment factors for the pelvis and sacrum; process the tumor state influencing factors for the pelvis and sacrum and the tumor risk adjustment factors for the pelvis and sacrum to generate the dynamic risk parameter information of the target user; process the preset end-to-end deep learning model based on non-enhanced magnetic resonance imaging based on the training sample set and the verification sample set to generate a target end-to-end deep learning model; process the target MRI sequence image information, the physiological parameter information of the target user, and the dynamic risk parameter information of the target user based on the target end-to-end deep learning model to generate user warning information, where the user warning information is used to characterize the attribute information of the preoperative pelvis and sacrum tumors of the target user.

9. An electronic device, characterized in that, Including: A first processor; And a memory, configured to store the executable instructions of the first processor; Among them, the first processor is configured to execute the pelvic and sacral tumor warning method based on the non-enhanced MRI end-to-end deep learning model according to any one of claims 1 to 7 by executing the executable instructions.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a second processor, it implements the pelvic and sacral tumor warning method based on the non-enhanced MRI end-to-end deep learning model according to any one of claims 1 to 7.

Images