Intelligent tumor recognition system and method based on big data

Abstract

The invention relates to the technical field of intelligent recognition, in particular to an intelligent tumor recognition system and method based on big data. The system comprises a tumor big data processing module, a tumor identification model construction module, a tumor multi-source feature fusion module and a tumor classification identification mapping module, and can acquire corresponding tumor multi-modal medical image data and patient pathological clinical data through butt joint and perform image registration and filtering cleaning processing. Tumor multi-mode standard image data and patient clinical standard data are obtained; constructing a corresponding tumor recognition model; performing feature analysis and big data fusion splicing on the tumor multi-modal standard image data and the clinical standard data of the patient to obtain a multi-source fusion feature vector of the tumor of the patient; and inputting the multi-source fusion feature vector of the tumor of the patient into a constructed tumor identification model to carry out tumor classification and identification, thereby outputting a medical image tumor identification result. According to the invention, accurate recognition of tumors can be realized.

Description

Technical Field

[0001] The present invention relates to the field of intelligent identification technology, and in particular to a tumor intelligent identification system and method based on big data. Background Art

[0002] In recent years, with the continuous advancement of big data technology and artificial intelligence (AI) technology, intelligent tumor recognition methods based on big data analysis and deep learning have gradually become an important development trend in the field of tumor diagnosis. Big data technology can provide more comprehensive and objective data support for accurate diagnosis of tumors by integrating and analyzing a large amount of medical imaging data, genomic data, clinical medical records data and other information. Deep learning algorithms in artificial intelligence, especially convolutional neural networks (CNNs), have made significant breakthroughs in medical image processing, and can automatically extract features from massive imaging data and perform accurate classification, thereby realizing intelligent tumor recognition. However, most existing tumor recognition methods mainly rely on doctors' experience and judgment and imaging examinations, fail to fully utilize the synergistic effect of multi-source data, and are difficult to adapt to the ever-changing data and individual differences of patients in clinical practice, resulting in limited recognition accuracy. Summary of the invention

[0003] Based on this, it is necessary for the present invention to provide a tumor intelligent identification system based on big data to solve at least one of the above technical problems.

[0004] To achieve the above objectives, a tumor intelligent identification system based on big data includes the following modules: The tumor big data processing module is used to obtain the corresponding tumor multimodal medical imaging data by connecting with various medical imaging devices, and to obtain the corresponding patient pathological clinical data in real time by establishing an interface with the hospital information management system and the electronic medical record system; image registration and filtering and cleaning are performed on the tumor multimodal medical imaging data and the patient pathological clinical data to obtain the tumor multimodal standard imaging data and the patient clinical standard data; The tumor recognition model building module is used to obtain the annotated historical image data of tumors corresponding to the annotated tumor type, location, size, boundary and malignancy, and build the corresponding tumor recognition model using the deep learning grid architecture; The tumor multi-source feature fusion module is used to use convolutional neural networks combined with computer vision technology to perform tumor region feature analysis on tumor multi-modal standard image data to obtain tumor region image feature vectors; determine the pathological malignancy vector of the patient's clinical standard data to obtain the patient's clinical pathological malignancy feature vector; perform big data fusion and splicing on the tumor region image feature vector and the patient's clinical pathological malignancy feature vector to obtain the patient's tumor multi-source fusion feature vector; The tumor classification and recognition mapping module is used to input the multi-source fusion feature vector of the patient's tumor into the constructed tumor recognition model to classify and recognize the corresponding tumor area on the corresponding tumor medical image, so as to classify and map the corresponding tumor type, location, size, boundary and malignancy degree to the corresponding tumor medical image in an intuitive way, thereby outputting the corresponding medical image tumor recognition result.

[0005] Furthermore, the tumor big data processing module includes the following functions: By connecting with various medical imaging devices, corresponding multimodal medical imaging data of tumors can be acquired in real time; By establishing an interface with the hospital information management system and the electronic medical record system, the corresponding patient pathology clinical data, including medical history, tumor symptoms, signs and tumor pathology section examination results, can be obtained in real time; Performing image coordinate registration on each corresponding medical image in the tumor multimodal medical image data to generate corresponding tumor multimodal image registration data in the same coordinate system; Performing image pixel fuzziness analysis on each medical image in the corresponding tumor multimodal image registration data in the same coordinate system to obtain the pixel fuzziness corresponding to each medical image in the tumor multimodal image; performing pixel fuzziness filtering, denoising and standardization on each medical image in the corresponding tumor multimodal image registration data in the same coordinate system based on the pixel fuzziness corresponding to each medical image in the tumor multimodal image to obtain tumor multimodal standard image data; Each sub-data within the patient's pathological clinical data is cleaned and standardized to correct and fill in the corresponding errors and missing values ​​in each sub-data, and each sub-data is unified into the same data format to obtain the patient's clinical standard data.

[0006] Furthermore, the tumor multimodal medical imaging data includes tumor CT images, tumor MRI images and tumor PET images.

[0007] Furthermore, the image coordinate registration of each corresponding medical image in the tumor multimodal medical image data includes: Extracting image feature points of each medical image corresponding to the tumor multimodal medical image data to obtain an image feature point set corresponding to each tumor medical image; The spatial relationship between the image feature points in the same tumor medical image is obtained through each feature point in the image feature point set corresponding to each tumor medical image, and similarity quantization calculation is performed between each feature point in the same tumor medical image in the image feature point set corresponding to each tumor medical image to obtain the similarity measurement between the image feature points in the same tumor medical image; According to the spatial relationship and similarity measurement between the image feature points in the same tumor medical image, the corresponding image feature point set is constructed with a feature association topological structure to determine the edge between each feature point, and the edges are weighted according to the corresponding spatial relationship and similarity measurement, so as to construct a topological structure reflecting the global association between the image feature points in the same tumor medical image, so as to generate the feature association topological structure of each tumor medical image; Based on the corresponding feature points and edges in the topological structure of each tumor medical image feature association, the topological structure registration estimation is performed between the feature points of different modal images to solve the registration correspondence between the feature points of different modal images, and the registration estimation is performed by considering the symmetry between the associated topological structures to generate the tumor multimodal image registration transformation matrix; Based on the tumor multimodal image registration transformation matrix, image coordinate registration is performed on each corresponding medical image in the tumor multimodal medical image data to generate corresponding tumor multimodal image registration data in the same coordinate system.

[0008] Furthermore, the image coordinate registration of each corresponding medical image in the tumor multimodal medical image data based on the tumor multimodal image registration transformation matrix includes: Based on the tumor multimodal image registration transformation matrix, the image coordinates of each corresponding medical image in the tumor multimodal medical image data are initially registered and transformed to generate corresponding tumor multimodal initial registration data in the same coordinate system; The mean square error of the characteristic distance between each tumor medical image in the tumor multimodal initial registration data corresponding to the same coordinate system is calculated to obtain the mean square error of the characteristic distance between each tumor medical image after registration; According to the preset error threshold, the mean square error of the characteristic distance between each tumor medical image after registration is compared and judged. If the mean square error of the characteristic distance is less than the preset error threshold, iterative registration is not required; if the mean square error of the characteristic distance is greater than or equal to the preset error threshold, the corresponding registration transformation matrix is ​​iteratively optimized by introducing rigid body transformation constraints and selecting the corresponding anatomical feature points around the tumor in each tumor medical image. At the same time, the optimized registration transformation matrix is ​​used to iteratively register the corresponding medical images to generate corresponding tumor multimodal image registration data in the same coordinate system.

[0009] Furthermore, the tumor recognition model building module includes the following functions: Obtaining annotated historical image data of tumors corresponding to the annotated tumor type, location, size, boundary, and malignancy; The corresponding tumor type, location, size, boundary and degree of malignancy are combined into a tumor feature vector corresponding to the tumor annotated image, and each sub-feature in the corresponding tumor feature vector is input into the corresponding grid units at different levels in the deep learning grid architecture, where each grid unit adopts a different convolution kernel and activation function. At the same time, cross-layer connection and feedback mechanism are introduced between each grid unit, and the tumor annotated image corresponding to the tumor annotated historical image data is input into the deep learning grid architecture for classification mapping recognition training between image and feature, so as to construct the corresponding tumor recognition model.

[0010] Furthermore, the tumor multi-source feature fusion module includes the following functions: A multi-scale convolution module is constructed using a convolutional neural network to analyze the tumor size characteristics of the tumor area in the tumor multimodal standard image data, so as to use the multi-scale convolution module to capture the characteristic information corresponding to tumors at different positions and sizes of the tumor image, so as to obtain the location and size sub-characteristics of the tumor area; The corresponding tumor image in the tumor multimodal standard image data is binarized to generate tumor multimodal binarized image data; the farthest non-zero pixel point of the corresponding binary image in the tumor multimodal binarized image data is determined to obtain the farthest non-zero pixel point corresponding to the tumor area in the binarized image at 0°, 90°, 180° and 270°, and obtain the optimal non-zero pixel point position of the tumor multimodal image at four azimuth angles; Based on the optimal non-zero pixel points of the tumor multimodal image at four azimuth angles, the corresponding maximum circumscribed rectangular boundary of the tumor region is constructed, and the corresponding tumor image in the tumor multimodal standard image data is segmented based on the maximum circumscribed rectangular boundary of the tumor region to generate a tumor multimodal region boundary segmentation image; the corresponding tumor region boundary in the tumor multimodal region boundary segmentation image is optimized and the boundary feature is detected to obtain the tumor region boundary sub-feature; The tumor region position and size sub-features and the tumor region boundary sub-features are combined into corresponding feature vectors to obtain a tumor region image feature vector; Determine the pathological malignancy vector of the patient's clinical standard data to obtain the patient's clinical pathological malignancy feature vector; The image feature vectors of the tumor region and the clinical pathological malignancy feature vectors of the patients are fused and spliced ​​by big data to obtain the multi-source fusion feature vectors of the patient's tumor.

[0011] Furthermore, determining the pathological malignancy vector of the patient's clinical standard data includes: Obtain the corresponding tumor incidence probability through the corresponding medical history in the patient's clinical standard data; Obtain the corresponding tumor symptom onset time and tumor symptom onset frequency through the corresponding tumor symptoms in the patient's clinical standard data, and perform tumor severity assessment and calculation based on the tumor symptom onset time and tumor symptom onset frequency to obtain the severity of the tumor symptoms; Obtaining corresponding patient physical sign levels through corresponding physical signs in the patient's clinical standard data; The corresponding tumor tissue pathology section length and tumor tissue pathology section area are obtained through the corresponding tumor pathology section examination results in the patient's clinical standard data, and the pathological malignancy calculation formula is used to perform pathological quantitative calculation on the tumor tissue pathology section length and tumor tissue pathology section area based on the probability of tumor incidence, the severity of tumor symptoms, and the patient's physical sign level, so as to obtain the clinical pathological malignancy; The probability of tumor onset, time of tumor symptom onset, frequency of tumor symptom onset, severity of tumor symptoms, patient's physical sign level, length of tumor tissue pathology section, area of ​​tumor tissue pathology section and clinical pathology malignancy are combined into corresponding feature vectors to obtain the patient's clinical pathology malignancy feature vector.

[0012] Furthermore, the calculation formula for the pathological malignancy degree is specifically: ; In the formula, The clinical pathological malignancy is the length of the tumor tissue pathological section, is the weight corresponding to the slice length, is the area of ​​tumor tissue pathological section, is the weight corresponding to the slice area, is an exponential function, is the probability of tumor incidence, The severity of tumor symptoms, The patient's physical sign level.

[0013] Furthermore, the present invention also provides a method for intelligent tumor identification based on big data, which is implemented based on the intelligent tumor identification system based on big data as described above. The method for intelligent tumor identification based on big data includes: Obtain the corresponding multimodal medical imaging data of tumors by connecting with various medical imaging devices, and obtain the corresponding patient pathological clinical data in real time by establishing interfaces with the hospital information management system and the electronic medical record system; perform image registration and filtering and cleaning on the multimodal medical imaging data of tumors and the patient pathological clinical data to obtain the multimodal standard imaging data of tumors and the patient clinical standard data; Obtain historical tumor image data with annotated tumor type, location, size, boundary, and malignancy, and build a corresponding tumor recognition model using a deep learning grid architecture; Convolutional neural network combined with computer vision technology is used to analyze tumor regional features of tumor multimodal standard imaging data to obtain tumor regional image feature vectors; pathological malignancy vectors are determined for patient clinical standard data to obtain patient clinical pathological malignancy feature vectors; big data fusion and splicing of tumor regional image feature vectors and patient clinical pathological malignancy feature vectors are performed to obtain patient tumor multi-source fusion feature vectors; The multi-source fusion feature vector of the patient's tumor is input into the constructed tumor recognition model to classify and identify the corresponding tumor area on the corresponding tumor medical image, so as to map the corresponding tumor type, location, size, boundary and malignancy level to the corresponding tumor medical image in an intuitive way, thereby outputting the corresponding medical image tumor recognition result.

[0014] Beneficial effects of the present invention: The big data-based tumor intelligent recognition system proposed in the present invention is generally composed of a tumor big data processing module, a tumor recognition model construction module, a tumor multi-source feature fusion module and a tumor classification and recognition mapping module. Compared with the prior art, the beneficial effect of the present application is that, firstly, by docking with medical imaging equipment, different types of imaging data (such as CT, MRI, PET, etc.) can be acquired in real time. These imaging data can provide multi-dimensional information on the size, morphology, location, etc. of the tumor, helping doctors to make preliminary assessments. By docking with the hospital information management system (HIS) and the electronic medical record system (EMR), the patient's pathological clinical data can be acquired in a timely manner, including clinical information such as medical history, tumor symptoms, signs, and tumor pathological section examination results. These data provide key background information for the analysis of multimodal images, which can effectively improve the accuracy of tumor recognition. Through image registration technology, image data from different modalities can be accurately aligned, allowing multimodal data to be comprehensively analyzed in the same coordinate system, effectively avoiding erroneous judgments due to image misalignment. The filtering and cleaning processing steps remove noise, enhance effective information, ensure the quality of input data, and improve the training effect of subsequent deep learning models, laying a good foundation for the construction of accurate tumor recognition models. This process greatly improves the accuracy and reliability of the entire tumor recognition process. Secondly, by annotating tumors on existing historical image data and building a tumor recognition model using a deep learning grid architecture, this process helps train the deep learning model and ensure its accuracy. By accurately annotating the tumor type, location, size, boundary, and malignancy, doctors and annotators provide high-quality model training data for the model. The accuracy of annotated historical image data directly determines the performance of the training model. The deep learning network can gradually extract key features in the image and perform classification and recognition through multi-level feature extraction. In addition, the grid architecture helps to improve the generalization ability of the model, enabling it to be effectively applied in new and unseen medical image data. This step greatly improves the accuracy of tumor recognition through the effective application of deep learning technology, which helps to achieve efficient tumor detection and evaluation in clinical practice. Then, the convolutional neural network (CNN) is combined with computer vision technology to perform feature analysis on the tumor area, thereby extracting tumor image feature vectors. This process enables tumor images to extract fine-grained spatial and morphological features from visual data, including the tumor boundary, size, shape, and its relationship with surrounding tissues. In this way, meaningful image features can be automatically identified through the convolution layer, and these features can be used to accurately locate and identify the tumor area.In addition, by extracting pathological malignant features, more clinical information can be provided to the model, and the clinical relevance of imaging data can be enhanced. By fusing imaging features with clinical features, more accurate tumor assessment can be achieved. Ultimately, the generation of this fusion feature not only enhances the accuracy of tumor region identification, but also enables it to adapt to the ever-changing data and individual differences of patients in clinical practice, thereby making full use of the synergistic effect of multi-source data. Finally, by inputting the multi-source fusion feature vector of the patient's tumor into the constructed tumor recognition model for tumor classification and recognition, the key to this process is to achieve automated tumor classification and recognition, and to intuitively map tumor features (type, location, size, boundary, and malignancy) to the corresponding medical images, helping doctors to more intuitively understand the patient's tumor condition. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments thereof made with reference to the following drawings: Figure 1 This is a module schematic diagram of the tumor intelligent identification system based on big data of the present invention; Figure 2 for Figure 1 Schematic diagram of the functional flow of the tumor big data processing module; Figure 3 for Figure 1 Schematic diagram of the functional flow of the tumor multi-source feature fusion module. DETAILED DESCRIPTION

[0016] The technical system of the present invention is described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by technicians in this field without creative work are within the scope of protection of the present invention.

[0017] In addition, the accompanying drawings are only schematic illustrations of the present invention and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and their repeated description will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities. The functional entities can be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and / or processor systems and / or microcontroller systems.

[0018] It should be understood that, although the terms "first", "second", etc. may be used herein to describe various units, these units should not be limited by these terms. These terms are used only to distinguish one unit from another unit. For example, without departing from the scope of the exemplary embodiments, the first unit may be referred to as the second unit, and similarly the second unit may be referred to as the first unit. The term "and / or" used herein includes any and all combinations of one or more of the listed associated items.

[0019] To achieve this, please refer to Figures 1 to 3 The present invention provides a tumor intelligent identification system based on big data, and the system includes the following modules: The tumor big data processing module is used to obtain the corresponding tumor multimodal medical imaging data by connecting with various medical imaging devices, and to obtain the corresponding patient pathological clinical data in real time by establishing an interface with the hospital information management system and the electronic medical record system; image registration and filtering and cleaning are performed on the tumor multimodal medical imaging data and the patient pathological clinical data to obtain the tumor multimodal standard imaging data and the patient clinical standard data; The tumor recognition model building module is used to obtain the annotated historical image data of tumors corresponding to the annotated tumor type, location, size, boundary and malignancy, and build the corresponding tumor recognition model using the deep learning grid architecture; The tumor multi-source feature fusion module is used to use convolutional neural networks combined with computer vision technology to perform tumor region feature analysis on tumor multi-modal standard image data to obtain tumor region image feature vectors; determine the pathological malignancy vector of the patient's clinical standard data to obtain the patient's clinical pathological malignancy feature vector; perform big data fusion and splicing on the tumor region image feature vector and the patient's clinical pathological malignancy feature vector to obtain the patient's tumor multi-source fusion feature vector; The tumor classification and recognition mapping module is used to input the multi-source fusion feature vector of the patient's tumor into the constructed tumor recognition model to classify and recognize the corresponding tumor area on the corresponding tumor medical image, so as to classify and map the corresponding tumor type, location, size, boundary and malignancy degree to the corresponding tumor medical image in an intuitive way, thereby outputting the corresponding medical image tumor recognition result.

[0020] In the embodiment of the present invention, please refer to Figure 1 FIG. 1 is a schematic diagram of the modules of the intelligent tumor identification system based on big data of the present invention. In this example, the intelligent tumor identification system based on big data includes the following modules: S1: Tumor big data processing module, used to obtain corresponding tumor multimodal medical imaging data by connecting with various medical imaging devices, and to obtain corresponding patient pathological clinical data in real time by establishing interfaces with hospital information management systems and electronic medical record systems; image registration and filtering and cleaning of tumor multimodal medical imaging data and patient pathological clinical data are performed to obtain tumor multimodal standard imaging data and patient clinical standard data; In an embodiment of the present invention, by utilizing dedicated data interface software provided by medical imaging equipment manufacturers, docking with various medical imaging equipment such as CT, MRI, and PET is performed to achieve acquisition of multimodal medical imaging data of tumors. For example, for CT equipment, the tomographic image data generated by CT scanning is transmitted to a data acquisition server through the DICOM (Digital Imaging and Communications in Medicine) interface protocol. For MRI equipment, with the help of its supporting network transmission software, the multi-sequence image data generated by MRI is sent to the server in real time in accordance with the DICOM standard. For PET equipment, its own data output module is utilized to collect PET image data into the server in accordance with the DICOM specification. Through middleware technology, such as the HL7 (Health Level 7) interface engine, an interface is established with the hospital information management system and the electronic medical record system to obtain the patient's pathological clinical data in real time. From the hospital information management system, by writing SQL query statements, the patient's medical history, tumor symptoms, signs and other information are extracted from the patient's basic information table, diagnosis record table and other database tables. The tumor pathology section examination results are connected to the digital pathology system of the pathology department, and the key information in the pathology section image is read using image recognition technology. Combined with the diagnosis report of the pathologist, the image coordinates of the multimodal medical imaging data of the tumor are registered using professional medical image registration software, such as Elastix. The registration algorithm based on mutual information is selected to unify the different modal images into the same coordinate system. The Python OpenCV library is used to filter and denoise the image, such as using the Gaussian filter algorithm to remove noise, and generate the tumor multimodal standard image data. The Python pandas library is used to clean the patient's pathological clinical data, correct the error values, fill the missing values, and unify the data into the same format, and finally obtain the patient's clinical standard data.

[0021] S2: Tumor recognition model building module, used to obtain the annotated historical image data of tumors corresponding to the annotated tumor type, location, size, boundary and malignancy, and build the corresponding tumor recognition model using the deep learning grid architecture; In the embodiment of the present invention, the annotated historical image data of the tumor corresponding to the annotated tumor type, location, size, border and malignancy is obtained from the medical image database of the hospital, and the database is connected and the data is filtered from the relevant tables storing the image data and the annotated information by using SQL query statements. Assuming that there is a "medical_images" table storing the image data path and a "tumor_annotations" table storing the tumor annotation information in the database, the two tables are inner-joined, such as "SELECT medical_images.image_path,tumor_annotations.tumor_type,tumor_annotations.location,tumor_annotations.size,tumor_annotations.border,tumor_annotations.malignancy FROM medical_images INNER JOIN tumor_annotations ON medical_images.image_id = At the same time, a deep learning grid architecture is constructed by using the deep learning framework TensorFlow. In the architecture, each grid unit is defined as a convolution layer. Grid units at different levels use convolution kernels of different sizes and parameters. For example, the first layer of grid units uses 3×3 convolution kernels, and the second layer uses 5×5 convolution kernels. The activation functions are selected from different types such as ReLU and Sigmoid. Cross-layer connections are introduced between each grid unit through the TensorFlow API, and the tf.keras.layers.add() function is used. The outputs of different layers are added and fused; a feedback mechanism is introduced to feed back the output of the previous moment to the input of the current moment through a recurrent neural network unit (such as LSTM), the tumor annotated image corresponding to the tumor annotated historical image data is read into the TensorFlow dataset object, the image size is adjusted to the size required by the model input, such as 224×224 pixels, the image data and the corresponding tumor feature vector are input into the deep learning grid architecture, the cross entropy loss function and the Adam optimizer are used to perform classification mapping recognition training between images and features, and after multiple rounds of training (such as 50 rounds), the corresponding tumor recognition model is constructed.

[0022] S3: Tumor multi-source feature fusion module, used to use convolutional neural network combined with computer vision technology to perform tumor region feature analysis on tumor multi-modal standard image data to obtain tumor region image feature vector; determine the pathological malignancy vector of the patient's clinical standard data to obtain the patient's clinical pathological malignancy feature vector; perform big data fusion and splicing on the tumor region image feature vector and the patient's clinical pathological malignancy feature vector to obtain the patient's tumor multi-source fusion feature vector; In an embodiment of the present invention, a convolutional neural network is constructed by using Python's TensorFlow deep learning framework to perform tumor regional feature analysis on tumor multimodal standard image data, and a multi-scale convolution module is constructed, which includes convolution layers with convolution kernels of different sizes, such as 3×3, 5×5, and 7×7 convolution kernels. The multimodal standard image data of the tumor is input into the module, and each convolution layer performs a convolution operation on the tumor image to capture the feature information corresponding to tumors of different positions and sizes. The output of each convolution layer is subjected to feature extraction, such as using a global average pooling operation to convert the output feature map into a fixed-length vector to obtain a tumor regional image feature vector. For the patient's clinical standard data, the pandas library in Python is used to extract the patient's medical history. The probability of tumor incidence is extracted from the data, the time and frequency of symptom onset are obtained from tumor symptoms and the severity is calculated, the sign level is calculated from the physical signs, the section length and area are extracted from the tumor pathological section examination results, and the clinical pathological malignancy is calculated by combining these data according to the pre-established formula. The probability of tumor incidence, the time of tumor symptom onset, the frequency of tumor symptom onset, the severity of tumor symptoms, the patient's physical sign level, the length of tumor tissue pathological section, the area of ​​tumor tissue pathological section and the clinical pathological malignancy are merged into the patient's clinical pathological malignancy feature vector. The tumor region image feature vector and the patient's clinical pathological malignancy feature vector are fused and spliced ​​using Python's numpy library to perform big data fusion and splicing, and finally the patient's tumor multi-source fusion feature vector is obtained.

[0023] S4: Tumor classification and recognition mapping module is used to input the multi-source fusion feature vector of the patient's tumor into the constructed tumor recognition model to classify and recognize the corresponding tumor area on the corresponding tumor medical image, so as to classify and map the corresponding tumor type, location, size, boundary and malignancy degree to the corresponding tumor medical image in an intuitive way, thereby outputting the corresponding medical image tumor recognition result.

[0024] In an embodiment of the present invention, the multi-source fusion feature vector of the patient's tumor is organized into a data structure consistent with the input format of the tumor recognition model and input into the constructed tumor recognition model. The tumor recognition model is based on a deep learning grid architecture and uses the feature mapping relationship learned during the training process to process the input feature vector. The model outputs the prediction results of the tumor type, position, size, boundary and malignancy corresponding to each medical image. The Python OpenCV library is used to classify and map these prediction results to the corresponding tumor medical images in an intuitive manner. For example, for the tumor position, the predicted tumor area is marked with a rectangular box on the image; for the tumor size, the predicted size value is marked next to the rectangular box; for the tumor type and malignancy, it is displayed in text form in the corner of the image, and the tumor medical image with the recognition result annotation is saved and output to obtain the corresponding medical image tumor recognition result, providing medical personnel with intuitive and accurate tumor recognition analysis information.

[0025] Furthermore, the tumor big data processing module includes the following functions: By connecting with various medical imaging devices, corresponding multimodal medical imaging data of tumors can be acquired in real time; By establishing an interface with the hospital information management system and the electronic medical record system, the corresponding patient pathology clinical data, including medical history, tumor symptoms, signs and tumor pathology section examination results, can be obtained in real time; Performing image coordinate registration on each corresponding medical image in the tumor multimodal medical image data to generate corresponding tumor multimodal image registration data in the same coordinate system; Performing image pixel fuzziness analysis on each medical image in the corresponding tumor multimodal image registration data in the same coordinate system to obtain the pixel fuzziness corresponding to each medical image in the tumor multimodal image; performing pixel fuzziness filtering, denoising and standardization on each medical image in the corresponding tumor multimodal image registration data in the same coordinate system based on the pixel fuzziness corresponding to each medical image in the tumor multimodal image to obtain tumor multimodal standard image data; Each sub-data within the patient's pathological clinical data is cleaned and standardized to correct and fill in the corresponding errors and missing values ​​in each sub-data, and each sub-data is unified into the same data format to obtain the patient's clinical standard data.

[0026] As an embodiment of the present invention, refer to Figure 2 As shown, Figure 1 FIG. 1 is a functional flow chart of a tumor big data processing module in FIG. 1 . In this embodiment, the tumor big data processing module includes the following functions: S11: Acquire corresponding multimodal medical imaging data of tumors in real time by connecting with various medical imaging devices; In an embodiment of the present invention, by connecting to the corresponding dedicated data interface of the medical imaging equipment corresponding to the hospital, connecting to various medical imaging equipment, and realizing the real-time acquisition of multimodal medical imaging data of tumors. For example, for CT equipment, the DICOM (Digital Imaging and Communications in Medicine) interface protocol is used to transmit the tomographic image data obtained by CT scanning to the data acquisition server. In terms of magnetic resonance imaging (MRI) equipment, the multi-sequence image data generated by MRI is sent to the server in real time according to the DICOM standard through the network transmission software supporting the equipment. For positron emission tomography (PET) equipment, with the help of its own data output module, the PET image data is collected into the server in accordance with the DICOM specification. After each scan is completed, the device automatically pushes the image data to the specified storage path, and the server receives and stores these data in real time, ensuring the acquisition of comprehensive multimodal medical imaging data of tumors, providing a basis for subsequent analysis.

[0027] S12: Acquire corresponding patient pathological clinical data in real time by establishing an interface with the hospital information management system and the electronic medical record system, including medical history, tumor symptoms, signs, and tumor pathological biopsy results; In an embodiment of the present invention, by utilizing middleware technology, such as the HL7 (Health Level 7) interface engine, an interface is established with a hospital information management system and an electronic medical record system to obtain patient pathological clinical data in real time. In the hospital information management system, by writing SQL query statements, the patient's medical history information, including past disease experience, is extracted from database tables such as the patient basic information table and the diagnosis record table. For tumor symptom and sign data, relevant information is extracted from the doctor's record documents in accordance with specific text format specifications using natural language processing tools (such as NLTK, Natural Language Toolkit). The tumor pathological section examination results are connected to the digital pathology system of the pathology department, and the key information in the pathological section image is read using image recognition technology. Combined with the pathologist's laboratory examination report, these data are integrated and transmitted to the data repository. In this way, accurate and comprehensive patient pathological clinical data can be obtained in real time.

[0028] S13: performing image coordinate registration on each corresponding medical image in the tumor multimodal medical image data to generate corresponding tumor multimodal image registration data in the same coordinate system; In an embodiment of the present invention, by using professional medical image registration software, such as Elastix, image coordinate registration is performed on each medical image in the multimodal medical image data of the tumor. First, the medical image data of different modalities such as CT, MRI, PET, etc. are imported into the Elastix software, and in the software interface, a suitable registration algorithm is selected, such as a registration algorithm based on mutual information. For images of different modalities, one of the images (such as CT images) is set as a reference image, and the other images are set as floating images. The software continuously adjusts the spatial position and direction of the floating image through an optimization algorithm, so that the floating image and the reference image match as much as possible in anatomical structure. For example, for CT and MRI images of brain tumors, the brain tissue, blood vessels and other structures in the MRI image are accurately aligned with the corresponding structures in the CT image in spatial coordinates through the registration algorithm. After registration, the corresponding multimodal image registration data of the tumor is generated in the same coordinate system, which is convenient for subsequent joint analysis of images of different modalities.

[0029] S14: performing image pixel fuzziness analysis on each medical image in the corresponding tumor multimodal image registration data in the same coordinate system to obtain pixel fuzziness corresponding to each medical image in the tumor multimodal image; performing pixel fuzziness filtering, denoising and standardization on each medical image in the corresponding tumor multimodal image registration data in the same coordinate system based on the pixel fuzziness corresponding to each medical image in the tumor multimodal image to obtain tumor multimodal standard image data; In an embodiment of the present invention, the OpenCV library of Python is used to perform image pixel blur analysis on each medical image in the corresponding tumor multimodal image registration data in the same coordinate system, so that the registered medical image data is read into the Python environment in the form of an array, and the image is processed using the Laplace operator or Gaussian difference (DoG) operator of OpenCV. For example, for a CT image, the pixel blur of the image is obtained by calculating the variance of the Laplace operator. If the variance is small, it means that the image details are rich and the pixel blur is low; conversely, if the variance is large, the pixel blur is high. Based on the pixel blur corresponding to each medical image, the Gaussian filtering algorithm is used to perform pixel blur filtering and denoising on the image in the OpenCV library, and the parameters of the Gaussian kernel are adjusted according to the pixel blur. For images with high blur, the Gaussian kernel size is appropriately increased to better remove noise. After denoising, an image normalization method is used, such as normalizing the image pixel value to a range of 0-1, to finally obtain tumor multimodal standard image data, improve image quality, and facilitate subsequent analysis.

[0030] S15: Clean and standardize each sub-data in the patient's pathological clinical data to correct and fill in the corresponding errors and missing values ​​in each sub-data, and unify each sub-data into the same data format to obtain the patient's clinical standard data.

[0031] In an embodiment of the present invention, the various sub-data in the patient's pathological clinical data are cleaned and standardized by using Python's pandas library, so that the patient's pathological clinical data obtained from the hospital information management system and the electronic medical record system are read into the pandas DataFrame object. For text-type data such as medical history, tumor symptoms, and physical signs, regular expressions and string processing functions are used to check and correct misspellings, grammatical errors, etc. For missing values, mean filling, median filling, or missing value prediction algorithms based on machine learning are used to fill them. For example, for the numerical data of patient age, if there are missing values, the mean of the ages of other patients is calculated, and the missing values ​​are filled with the mean. For numerical data such as tumor pathological section examination results, the data range is checked, and obviously unreasonable data is removed. The various sub-data are unified into the same data format, such as unifying all date data into the "YYYY-MM-DD" format. After these processes, the patient's clinical standard data is obtained, which provides high-quality data support for subsequent intelligent identification of tumors based on big data.

[0032] Furthermore, the tumor multimodal medical imaging data includes tumor CT images, tumor MRI images and tumor PET images.

[0033] Furthermore, the image coordinate registration of each corresponding medical image in the tumor multimodal medical image data includes: Extracting image feature points of each medical image corresponding to the tumor multimodal medical image data to obtain an image feature point set corresponding to each tumor medical image; In an embodiment of the present invention, image feature points of each medical image in the multimodal medical image data of tumors are extracted by using the Scikit-Image library of Python, so that medical image data of different modalities such as CT, MRI, and PET are read into the Python environment in the form of arrays. For each medical image, the SIFT (Scale-Invariant Feature Transform) algorithm or the SURF (Speeded-Up Robust Features) algorithm in the Scikit-Image library is used. Taking the SIFT algorithm as an example, the sift.detectAndCompute() function is called, which automatically detects feature points in the image and calculates its descriptor. For example, for a CT image, hundreds of feature points are detected after processing. These feature points have unique position, scale and direction information in the image, and the feature points and descriptors detected for each tumor medical image are organized into a set, and finally the image feature point set corresponding to each tumor medical image is obtained, which provides basic data for the subsequent analysis of the relationship between the image feature points.

[0034] Preferably, the spatial relationship between the image feature points in the same tumor medical image is obtained through each feature point in the image feature point set corresponding to each tumor medical image, and a similarity quantization calculation is performed between each feature point in the same tumor medical image in the image feature point set corresponding to each tumor medical image to obtain a similarity measure between the image feature points in the same tumor medical image; In the embodiment of the present invention, a custom algorithm is written in Python based on the image feature point set corresponding to each tumor medical image, and the spatial relationship between the image feature points in the same tumor medical image is obtained through the coordinate information of each feature point. For example, for feature point A (x1, y1) and feature point B (x2, y2), the Euclidean distance between them is calculated as distance=sqrt( + ), which is used to represent the spatial distance relationship. At the same time, the descriptor of the feature point is used to quantify the similarity between each feature point. Taking SIFT descriptor as an example, the similarity between two feature point descriptors is calculated using Euclidean distance or cosine similarity algorithm. For example, the cosine similarity of the SIFT descriptors of feature point A and feature point B is calculated. If the similarity value is closer to 1, it means that the two feature points are more similar. By performing such calculations on all feature point pairs in the same tumor medical image, the similarity measurement between the image feature points in the same tumor medical image is obtained. These measurement values ​​reflect the similarity between the feature points in terms of features and spatial positions.

[0035] Preferably, a feature association topological structure is constructed for a corresponding set of image feature points according to the spatial relationship and similarity measurement between the image feature points in the same tumor medical image, so as to determine the edge between each feature point, and weights are assigned to the edges therebetween according to the corresponding spatial relationship and similarity measurement, thereby constructing a topological structure reflecting the global association between the image feature points in the same tumor medical image, so as to generate feature association topological structures of each tumor medical image; In an embodiment of the present invention, a feature association topology is constructed using Python's NetworkX library according to the spatial relationship and similarity measurement between image feature points in the same tumor medical image. First, an empty undirected graph object G=nx.Graph() is created. For the image feature point set of the same tumor medical image, each feature point is added to the graph as a node of the graph, such as G.add_nodes_from(feature_points_set). Then, the edge between the nodes is determined according to the spatial relationship and similarity measurement between the feature points. If the spatial distance between the two feature points is within a certain threshold and the similarity measurement exceeds the set similarity threshold (such as the spatial distance is less than 50 pixels and the similarity is greater than 0.8), then the edge between the two nodes is added to the graph, such as G.add_edge(feature_point_A, feature_point_B), weights are assigned to edges according to spatial relationships and similarity metrics. For example, the smaller the spatial distance and the higher the similarity, the greater the weight of the edge. In this way, a topological structure reflecting the global association between image feature points in the same tumor medical image is constructed, and the topological structures corresponding to each tumor medical image are organized into sets, thereby generating feature-related topological structures of each tumor medical image.

[0036] Preferably, a topological structure registration estimation is performed between feature points of different modal images based on the corresponding feature points and edges in the associated topological structure of each tumor medical image feature, so as to solve the registration correspondence between the feature points of different modal images, and the registration estimation is performed by considering the symmetry between the associated topological structures, so as to generate a tumor multimodal image registration transformation matrix; In an embodiment of the present invention, a registration algorithm based on feature point matching in Python's OpenCV library is used to perform topological structure registration estimation between feature points of images of different modalities based on the associated topological structure of each tumor medical image feature. First, for images of two different modalities (such as CT and MRI), their image feature associated topological structures are obtained respectively, and the brute force matching algorithm of OpenCV is used to match the feature points of the two topological structures according to the descriptors of the feature points. For example, in the feature point set of the CT image topological structure, the feature point that is most similar to the feature point descriptor in the MRI image topological structure is found to establish a matching pair. At the same time, the symmetry between the associated topological structures is considered for registration estimation, that is, not only the matching relationship from CT to MRI is found, but also whether the matching relationship from MRI to CT is consistent is verified. Through iterative optimization of the matching process, the registration correspondence between the feature points of images of different modalities is solved. Finally, a tumor multimodal image registration transformation matrix is ​​generated according to the matching results. The matrix contains transformation parameters such as translation, rotation and scaling, which are used to unify the coordinates of images of different modalities into the same system.

[0037] Preferably, image coordinate registration is performed on each corresponding medical image in the tumor multimodal medical image data based on the tumor multimodal image registration transformation matrix to generate corresponding tumor multimodal image registration data in the same coordinate system.

[0038] In an embodiment of the present invention, the image coordinates of each corresponding medical image in the tumor multimodal medical image data are registered by using the Python Scikit-Image library and the generated tumor multimodal image registration transformation matrix, and the medical image data such as CT, MRI, and PET are read into the Python environment. For each medical image, the transform.warp() function in the Scikit-Image library is used, and the tumor multimodal image registration transformation matrix is ​​passed in as a parameter. For example, for an MRI image, transformed_image=transform.warp(mri_image, transformation_matrix), the function performs coordinate transformation on each pixel of the MRI image according to the transformation matrix, and maps it to the same coordinate system. Such processing is performed on all medical images of different modalities in turn to generate corresponding tumor multimodal image registration data in the same coordinate system. These data provide an image data foundation with a unified coordinate reference for subsequent intelligent tumor recognition based on big data. Furthermore, the image coordinate registration of each corresponding medical image in the tumor multimodal medical image data based on the tumor multimodal image registration transformation matrix includes: Based on the tumor multimodal image registration transformation matrix, the image coordinates of each corresponding medical image in the tumor multimodal medical image data are initially registered and transformed to generate corresponding tumor multimodal initial registration data in the same coordinate system; In an embodiment of the present invention, by using Python's Scikit-Image library and a previously generated tumor multimodal image registration transformation matrix, the image coordinates of each corresponding medical image in the tumor multimodal medical image data are initially registered and transformed, and medical image data such as CT, MRI, and PET are read into the Python environment in the form of an array. For each medical image, the transform.warp() function in the Scikit-Image library is called, and the tumor multimodal image registration transformation matrix is ​​passed in as a parameter. For example, for a CT image ct_image, the transformed_ct_image =transform.warp(ct_image, transformation_matrix) operation is executed. The function performs coordinate transformation on each pixel of the CT image according to the transformation matrix and maps it to the same coordinate system. Other modal images such as MRI and PET are processed in the same way to generate corresponding tumor multimodal initial registration data in the same coordinate system. These initial registration data provide a basis for further evaluation and optimization of the registration effect.

[0039] Preferably, the mean square error of characteristic distances between the tumor medical images in the corresponding tumor multimodal initial registration data in the same coordinate system is calculated to obtain the mean square error of characteristic distances between the tumor medical images after registration; In an embodiment of the present invention, a custom function is written in Python to calculate the mean square error of feature distances between various tumor medical images in the corresponding tumor multimodal initial registration data in the same coordinate system. First, starting from the image feature point set corresponding to each tumor medical image previously obtained, for each pair of tumor medical images of different modalities (such as CT and MRI images), in the images after initial registration, the Euclidean distance between the corresponding feature points is calculated according to the feature point coordinate information. For example, assuming that there is a feature point A(x1, y1) in the CT image, and the corresponding feature point in the MRI image after registration is A'(x1', y1'), the Euclidean distance between them is calculated as distance=sqrt( + ), perform such distance calculations on all corresponding feature point pairs, and then calculate the average of the sum of the squares of these distances, that is, the mean square error of the feature distances between the two tumor medical images after registration. Perform such calculations on all combinations of different modal images (such as CT and PET, MRI and PET, etc.) to obtain the mean square error of the feature distances between each tumor medical image, so as to quantitatively evaluate the registration effect.

[0040] Preferably, the mean square error of the characteristic distances between the various tumor medical images after registration is compared and judged according to a preset error threshold. If the mean square error of the characteristic distances is less than the preset error threshold, iterative registration is not required. If the mean square error of the characteristic distances is greater than or equal to the preset error threshold, the corresponding registration transformation matrix is ​​iteratively optimized by introducing rigid body transformation constraints and selecting the corresponding anatomical feature points around the tumor in each tumor medical image. At the same time, the optimized registration transformation matrix is ​​used to iteratively register the corresponding medical images to generate corresponding tumor multimodal image registration data in the same coordinate system.

[0041] In an embodiment of the present invention, an error threshold is pre-set, for example, 0.5 (this value can be determined based on a large number of experiments and clinical experience), so as to compare the mean square error of the characteristic distances between the tumor medical images after registration with the preset error threshold by using Python's conditional judgment statement. If the mean square error of the characteristic distances is less than the preset error threshold, such as the mean square error of the characteristic distances between a CT and MRI image is 0.4, which is less than 0.5, then it is considered that the current registration effect is good and there is no need to perform iterative registration. If the mean square error of the characteristic distances is greater than or equal to the preset error threshold, such as the mean square error of the characteristic distances between a CT and PET image is 0.6, which is greater than 0.5, then by introducing a rigid body transformation constraint, a Python optimization algorithm library (such as Sci py's optimization module) selects the anatomical feature points (such as vascular bifurcation points, organ edge points, etc.) corresponding to the tumor in each tumor medical image, and iteratively optimizes the corresponding registration transformation matrix. In each iteration, the translation, rotation and scaling parameters of the transformation matrix are adjusted to gradually reduce the distance between the corresponding feature points. For example, in the first iteration, the translation parameters are adjusted according to the position of the anatomical feature points. In the second iteration, the rotation parameters are adjusted according to the direction of the feature points. The optimized registration transformation matrix is ​​used to perform registration transformation on the corresponding medical images again, such as using the transform.warp() function again. Finally, the corresponding tumor multimodal image registration data in the same coordinate system is generated to improve the registration accuracy and provide more accurate image data for intelligent tumor recognition.

[0042] Furthermore, the tumor recognition model building module includes the following functions: Obtaining annotated historical image data of tumors corresponding to the annotated tumor type, location, size, boundary, and malignancy; In the embodiment of the present invention, the annotated historical image data of the tumor corresponding to the annotated tumor type, location, size, border and malignancy is obtained from the medical image database of the hospital, and the database is connected and the data is filtered from the relevant tables storing the image data and the annotated information by using SQL query statements. Assuming that there is a "medical_images" table storing the image data path and a "tumor_annotations" table storing the tumor annotation information in the database, the two tables are inner-joined, such as "SELECT medical_images.image_path,tumor_annotations.tumor_type,tumor_annotations.location,tumor_annotations.size,tumor_annotations.border,tumor_annotations.malignancy FROM medical_images INNER JOIN tumor_annotations ON medical_images.image_id = "tumor_annotations.image_id", the query results contain the complete annotation information corresponding to each tumor annotated image. For example, the annotation information obtained for a CT image is: the tumor type is adenocarcinoma, the location is in the right lower lobe of the lung, the size is 3 cm in diameter, the boundary is clear, and the malignancy is moderate. These query results are organized into a data set to provide basic data for the subsequent construction of a tumor recognition model.

[0043] Preferably, the corresponding tumor type, location, size, boundary and degree of malignancy are combined into a tumor feature vector corresponding to the tumor annotated image, and each sub-feature in the corresponding tumor feature vector is respectively input into the grid units corresponding to different levels in the deep learning grid architecture, wherein each grid unit adopts a different convolution kernel and activation function, and a cross-layer connection and feedback mechanism is introduced between each grid unit. The tumor annotated image corresponding to the tumor annotated historical image data is input into the deep learning grid architecture for classification mapping recognition training between the image and the feature, so as to construct the corresponding tumor recognition model.

[0044] In an embodiment of the present invention, the corresponding tumor type, position, size, boundary and malignancy are combined into a tumor feature vector corresponding to the tumor annotated image by using Python's numpy library. For example, the tumor type is one-hot encoded. Assuming that there are 5 tumor types, adenocarcinoma is encoded as [1, 0, 0, 0, 0]; the position information is converted into coordinate values ​​according to the image coordinate system, such as (100, 150); the size is represented by a diameter value, such as 3; the boundary can be quantified as a value between 0 and 1 according to the clarity, such as 0.8; the malignancy is also quantified, with a moderate value of 0.6. These values ​​are combined into a feature vector [1, 0, 0, 0,0, 100, 150, 3, 0.8, 0.6], and build a deep learning grid architecture by using the deep learning framework TensorFlow. In the architecture, each grid unit is defined as a convolution layer. Grid units at different levels use convolution kernels of different sizes and parameters. For example, the first layer grid unit uses a 3×3 convolution kernel, and the second layer uses a 5×5 convolution kernel. The activation function selects different types such as ReLU and Sigmoid. Cross-layer connections are introduced between each grid unit through the TensorFlow API, and the outputs of different layers are added and fused using the tf.keras.layers.add() function. A feedback mechanism is introduced to feed back the output of the previous moment to the input of the current moment through a recurrent neural network unit (such as LSTM). The tumor annotated image corresponding to the tumor annotated historical image data is read into the TensorFlow dataset object, and the image size is adjusted to the size required by the model input, such as 224×224 pixels. By inputting the image data and the corresponding tumor feature vector into the deep learning grid architecture, the cross entropy loss function and the Adam optimizer are used to perform classification mapping recognition training between the image and the feature. After multiple rounds of training (such as 50 rounds), the corresponding tumor recognition model is constructed.

[0045] Furthermore, the tumor multi-source feature fusion module includes the following functions: A multi-scale convolution module is constructed using a convolutional neural network to analyze the tumor size characteristics of the tumor area in the tumor multimodal standard image data, so as to use the multi-scale convolution module to capture the characteristic information corresponding to tumors at different positions and sizes of the tumor image, so as to obtain the location and size sub-characteristics of the tumor area; The corresponding tumor image in the tumor multimodal standard image data is binarized to generate tumor multimodal binarized image data; the farthest non-zero pixel point of the corresponding binary image in the tumor multimodal binarized image data is determined to obtain the farthest non-zero pixel point corresponding to the tumor area in the binarized image at 0°, 90°, 180° and 270°, and obtain the optimal non-zero pixel point position of the tumor multimodal image at four azimuth angles; Based on the optimal non-zero pixel points of the tumor multimodal image at four azimuth angles, the corresponding maximum circumscribed rectangular boundary of the tumor region is constructed, and the corresponding tumor image in the tumor multimodal standard image data is segmented based on the maximum circumscribed rectangular boundary of the tumor region to generate a tumor multimodal region boundary segmentation image; the corresponding tumor region boundary in the tumor multimodal region boundary segmentation image is optimized and the boundary feature is detected to obtain the tumor region boundary sub-feature; The tumor region position and size sub-features and the tumor region boundary sub-features are combined into corresponding feature vectors to obtain a tumor region image feature vector; Determine the pathological malignancy vector of the patient's clinical standard data to obtain the patient's clinical pathological malignancy feature vector; The image feature vectors of the tumor region and the clinical pathological malignancy feature vectors of the patients are fused and spliced ​​by big data to obtain the multi-source fusion feature vectors of the patient's tumor.

[0046] As an embodiment of the present invention, refer to Figure 3 As shown, Figure 1 FIG. 1 is a functional flow chart of a tumor multi-source feature fusion module in FIG. 1 . In this embodiment, the tumor multi-source feature fusion module includes the following functions: S31: Use a convolutional neural network to construct a multi-scale convolution module to perform tumor size feature analysis on the tumor area in the tumor multimodal standard image data, so as to use the multi-scale convolution module to capture the feature information corresponding to tumors at different positions and sizes of the tumor image, so as to obtain the position and size sub-features of the tumor area; In an embodiment of the present invention, a multi-scale convolution module is constructed by using Python's TensorFlow deep learning framework. In TensorFlow, multi-scale convolution is achieved by defining a plurality of convolution layers with convolution kernels of different sizes. For example, a multi-scale convolution module including 3×3, 5×5 and 7×7 convolution kernels is constructed, and the tumor multimodal standard image data is input into the module. Each convolution layer performs convolution operations on the tumor image respectively. The 3×3 convolution kernel captures the detailed features of smaller areas in the tumor image, the 5×5 convolution kernel focuses on the features of medium-sized areas, and the 7×7 convolution kernel can obtain the overall features of larger areas. In this way, the multi-scale convolution module can capture the feature information corresponding to tumors of different positions and sizes in the tumor image, and extract and analyze the features of the output of each convolution layer. For example, a global average pooling operation is used to convert the output feature map of each convolution layer into a vector of a fixed length. These vectors are combined to represent the position and size sub-features of the tumor area, providing key feature data for subsequent tumor analysis.

[0047] S32: binarizing the corresponding tumor image in the tumor multimodal standard image data to generate tumor multimodal binarized image data; determining the farthest non-zero pixel point of the corresponding binarized image in the tumor multimodal binarized image data to obtain the farthest non-zero pixel point corresponding to the tumor area in the binarized image at 0°, 90°, 180° and 270°, and obtaining the optimal non-zero pixel point position of the tumor multimodal image at four azimuth angles; In an embodiment of the present invention, the tumor image corresponding to the tumor multimodal standard image data is binarized by using Python's OpenCV library, so that the tumor image is read into the Mat object of OpenCV, and the Otsu threshold method is used to convert the image into a binary image through the cv2.threshold() function, that is, the pixel values ​​are only 0 (background) and 255 (tumor area). For example, for a CT tumor image, after the binarization process, the tumor area is clearly separated from the background, and the tumor multimodal binary image data is generated. Then, for each binary image in the binary image data, the pixel points of the image are traversed to determine the farthest non-zero pixel points in the directions of 0° (horizontally to the right), 90° (vertically upward), 180° (horizontally to the left), and 270° (vertically downward), with 0° as the background. Taking the direction as an example, traverse the pixel rows from the left border of the image to the right, record the position of the first non-zero pixel, and continue traversing until the rightmost non-zero pixel in the row is found. This point is the farthest non-zero pixel in the 0° direction. The same method is used to determine the farthest non-zero pixel in the other three directions, and the optimal non-zero pixel positions of the tumor multimodal image in four azimuth angles are obtained. These positions are of great significance for determining the boundary range of the tumor.

[0048] S33: constructing a corresponding maximum circumscribed rectangular boundary of the tumor region based on the optimal non-zero pixel points of the tumor multimodal image at four azimuth angles, and performing tumor region segmentation on the corresponding tumor image in the tumor multimodal standard image data based on the maximum circumscribed rectangular boundary of the tumor region to generate a tumor multimodal region boundary segmentation image; performing boundary optimization and boundary feature detection on the corresponding tumor region boundary in the tumor multimodal region boundary segmentation image to obtain tumor region boundary sub-features; In an embodiment of the present invention, the maximum circumscribed rectangle boundary of the corresponding tumor area is constructed by using a Python mathematical calculation library (such as numpy) according to the optimal non-zero pixel points of the tumor multimodal image at four azimuth angles. For example, the left and right boundaries of the rectangle are determined according to the horizontal coordinates of the farthest non-zero pixel points in the directions of 0° and 180°, and the upper and lower boundaries of the rectangle are determined according to the vertical coordinates of the farthest non-zero pixel points in the directions of 90° and 270°, thereby obtaining the maximum circumscribed rectangle of the tumor area, and the cv2.rectangle() function of the OpenCV library is used to generate the maximum circumscribed rectangle on the corresponding tumor image in the tumor multimodal standard image data. Draw the rectangle, and use image masking to set the part outside the rectangular area to 0 to achieve tumor region segmentation and generate a tumor multimodal region boundary segmentation image. For the tumor region boundary corresponding to the segmented image, use the edge detection algorithm of the OpenCV library (such as the Canny edge detection algorithm) to perform boundary optimization and boundary feature detection. The Canny algorithm accurately detects the tumor region boundary through Gaussian filtering denoising, calculating gradient amplitude and direction, non-maximum suppression, and dual threshold detection. It also extracts the length, curvature and other features of the boundary to obtain the tumor region boundary sub-features, providing more feature information for comprehensive analysis of the tumor region.

[0049] S34: merging the tumor region position and size sub-features and the tumor region boundary sub-features into corresponding feature vectors to obtain a tumor region image feature vector; In an embodiment of the present invention, the position and size sub-features of the tumor region and the boundary sub-features of the tumor region are merged into corresponding feature vectors by using the numpy library of Python. Assuming that the position and size sub-features of the tumor region are a vector [x1, y1, size1, feature1, ...] with a length of 10, and the boundary sub-features of the tumor region are a vector [length, curvature1, curvature2, ...] with a length of 8, the two vectors are connected in sequence by the numpy.concatenate() function to obtain a new vector [x1, y1, size1, feature1, ...,length, curvature1, curvature2, ...] with a length of 18. This vector is the image feature vector of the tumor region, which integrates various feature information such as the position, size and boundary of the tumor region, and provides important image feature basis for subsequent fusion with clinical data and tumor identification.

[0050] S35: determining a pathological malignancy vector for the patient's clinical standard data to obtain a clinical pathological malignancy feature vector for the patient; In an embodiment of the present invention, a pathological malignancy vector is determined for the patient's clinical standard data, and data related to the malignancy of the tumor, such as the pathological grade of the tumor, the level of tumor markers, etc., are extracted from the patient's clinical standard data. It is assumed that the pathological grade is divided into three levels: high, medium, and low, which are represented by the numbers 3, 2, and 1 respectively; the tumor marker level is normalized to between 0 and 1 according to the specific numerical value, and these data are combined into a vector, for example, [3, 0.6] (indicating that the pathological grade is high, and the tumor marker level is 0.6 after normalization), to obtain the patient's clinical pathological malignancy feature vector, which reflects the malignancy information of the patient's tumor in clinical pathology. After being fused with the tumor region image feature vector, it can more comprehensively describe the patient's tumor characteristics.

[0051] S36: Perform big data fusion and splicing on the tumor region image feature vector and the patient's clinical pathological malignancy feature vector to obtain the patient's tumor multi-source fusion feature vector.

[0052] In an embodiment of the present invention, the image feature vector of the tumor region and the clinical pathology malignant feature vector of the patient are fused and spliced ​​by using the numpy library of Python. Assume that the image feature vector of the tumor region is image_vector = [x1, y1, size1, ...], and the clinical pathology malignant feature vector of the patient is clinical_vector = [grade, marker_level, ...]. The two vectors are connected by the numpy.concatenate() function to obtain a new vector fusion_vector = [x1, y1, size1, ..., grade, marker_level,...], which is the multi-source fusion feature vector of the patient's tumor. The vector integrates the features of tumor images and clinical pathology, provides richer and more comprehensive input features for the intelligent tumor recognition model based on big data, and helps to improve the accuracy of tumor recognition.

[0053] Furthermore, determining the pathological malignancy vector of the patient's clinical standard data includes: Obtain the corresponding tumor incidence probability through the corresponding medical history in the patient's clinical standard data; In an embodiment of the present invention, the medical history corresponding to the clinical standard data of the patient is processed by using the Python data analysis library pandas to obtain the probability of tumor incidence. It is assumed that the clinical standard data of the patient is stored in a pandas DataFrame object, and the medical history information is in the "medical_history" column. Keywords and information related to tumor incidence are extracted from the medical history text by regular expressions and text mining technology. For example, if the medical history mentions information such as "many people in the family have cancer" and "long-term exposure to carcinogens", this information is matched with a pre-established tumor incidence risk knowledge base. Different factors related to the risk are assigned corresponding weights in the risk knowledge base, such as a family cancer history weight of 0.6 and a long-term exposure to carcinogens weight of 0.4. The probability of tumor incidence is calculated by weighted summation based on the matched factors and their weights. If a patient's medical history matches two factors, family cancer history and long-term exposure to carcinogens, the probability of tumor incidence = 0.6×0.4+ 0.4×0.3 = 0.36 (assuming the full score is 1), thereby quantifying the possibility of tumor incidence in the patient and providing data support for subsequent comprehensive analysis.

[0054] Preferably, the corresponding tumor symptom onset time and tumor symptom onset frequency are obtained through the corresponding tumor symptoms in the patient's clinical standard data, and the tumor severity assessment calculation is performed based on the tumor symptom onset time and tumor symptom onset frequency to obtain the severity of the tumor symptoms; In an embodiment of the present invention, the tumor symptoms corresponding to the patient's clinical standard data are analyzed with the help of Python's pandas library, and data is extracted from the column storing the tumor symptom information. Assume that the column is named "tumor_symptoms". For each patient's tumor symptom record, the tumor symptom occurrence time is extracted from the symptom description text through a time parsing function, such as the datetime.strptime() function. For example, if the symptom record is "cough symptoms began to appear on 2023-05-10", the extracted time is May 10, 2023. The frequency of tumor symptom occurrence is calculated by counting the number of occurrences of the same symptom in the entire record. Assuming that the cough symptom appears 5 times in the record and the total number of records is 20, the frequency of cough symptom occurrence = 5 ÷ 20 = 0.25. According to the time and frequency of tumor symptom occurrence, a tumor severity assessment calculation formula is established. For example, tumor symptom severity = symptom occurrence frequency × (current time - earliest symptom occurrence time) The number of days ÷ 30 (convert days into months for unified measurement). If the earliest symptom onset is May 10, 2023, and the current time is August 10, 2023, a total of 92 days, and the frequency of cough symptoms is 0.25, then the severity of tumor symptoms = 0.25×(92÷30)≈0.77, which can quantify the severity of tumor symptoms and reflect the impact of the tumor on the patient.

[0055] Preferably, the corresponding patient's vital sign level is obtained through the corresponding vital sign in the patient's clinical standard data; In the embodiment of the present invention, the corresponding physical signs in the clinical standard data of the patient are processed by using the data analysis tool of Python to obtain the level of the patient's physical signs. It is assumed that the physical sign data is stored in the "physical_signs" column, and the physical sign data includes information such as body temperature, blood pressure, and heart rate. For body temperature, its value is compared with the normal body temperature range (such as 36.0℃-37.2℃). If the body temperature is 37.5℃, the proportion exceeding the normal range is (37.5-37.2)÷(37.2-36.0)≈0.25. For blood pressure, the deviation ratios of systolic pressure and diastolic pressure from the normal range are calculated respectively. It is assumed that the normal systolic pressure range is 90-140mmHg, and the systolic pressure of a patient is 150mmHg, then the systolic pressure deviation ratio = (150-140)÷(140-90) =0.2; the normal diastolic pressure range is 60-90mmHg, and the diastolic pressure of the patient is 95mmHg, then the diastolic pressure deviation ratio = (95-90)÷(90-60)≈0.17. The deviation ratios of these signs are weighted and summed (weights are set according to the importance of different signs to tumor judgment, such as body temperature weight 0.3, systolic blood pressure weight 0.4, and diastolic blood pressure weight 0.3) to obtain the patient's physical sign level. For example, the patient's physical sign level = 0.3×0.25 + 0.4×0.2 + 0.3×0.17≈0.21. This is used to quantify the degree of deviation of the patient's physical signs from the normal state and assist in tumor-related analysis.

[0056] Preferably, the corresponding tumor tissue pathology section length and tumor tissue pathology section area are obtained through the corresponding tumor pathology section examination results in the patient's clinical standard data, and the tumor tissue pathology section length and tumor tissue pathology section area are pathologically quantified based on the probability of tumor onset, the severity of tumor symptoms, and the patient's physical sign level using a pathological malignancy calculation formula to obtain the clinical pathological malignancy; In an embodiment of the present invention, the length of the tumor tissue pathology section and the area of ​​the tumor tissue pathology section are extracted from the corresponding tumor pathology section examination results in the patient's clinical standard data through the Python pandas library. It is assumed that the tumor pathology section examination results are stored in the "pathological_slice_results" column. The length and area values ​​are extracted from the text through a regular expression. For example, the result description is "the length of the tumor tissue pathology section is 3 cm and the area is 5 cm²", then the length of 3 and the area of ​​5 are extracted. According to the pre-established pathological malignancy calculation formula, the length of the tumor tissue pathology section and the area of ​​the tumor tissue pathology section are pathologically quantified and calculated in combination with the probability of tumor incidence, the severity of tumor symptoms, and the patient's physical sign level to obtain the weights corresponding to the section length and the section area, such as 0.6 and 0.4, to obtain a quantitative value reflecting the malignancy of the tumor, and provide key clinical pathology data for intelligent identification of tumors. In addition, the pathological malignancy calculation formula can also use any weighted measurement method in the art to replace the pathological quantification calculation process, and is not limited to the pathological malignancy calculation formula.

[0057] Preferably, the probability of tumor onset, time of tumor symptom onset, frequency of tumor symptom onset, severity of tumor symptoms, patient's physical sign level, length of tumor tissue pathology section, area of ​​tumor tissue pathology section and clinical pathological malignancy are combined into a corresponding feature vector to obtain the patient's clinical pathological malignancy feature vector.

[0058] In an embodiment of the present invention, the probability of tumor onset, the time of tumor symptom onset, the frequency of tumor symptom onset, the severity of tumor symptoms, the level of patient physical signs, the length of tumor tissue pathology section, the area of ​​tumor tissue pathology section and the clinical pathological malignancy are merged into a corresponding feature vector by using Python's numpy library. Assuming that the probability of tumor onset is 0.7, the time of tumor symptom onset is "3 days", which is converted into a timestamp value of 3 (assuming that the unit is day), the frequency of tumor symptoms is 0.3, the severity of tumor symptoms is 0.5, the level of patient physical signs is 0.2, the length of tumor tissue pathology section is 4, the area of ​​tumor tissue pathology section is 6, and the clinical pathological malignancy is 0.8, these data are organized into a vector [0.7, 3, 0.3, 0.5, 0.2, 4, 6, 0.8], which is the patient's clinical pathological malignancy feature vector. The vector integrates various aspects of clinical pathological information, comprehensively reflects the relevant characteristics of the patient's tumor, provides rich input features for the subsequent tumor intelligent recognition model based on big data, and helps to improve the accuracy of the model in judging the malignancy of the tumor.

[0059] Furthermore, the calculation formula for the pathological malignancy degree is specifically: ; In the formula, The clinical pathological malignancy is the length of the tumor tissue pathological section, is the weight corresponding to the slice length, is the area of ​​tumor tissue pathological section, is the weight corresponding to the slice area, is an exponential function, is the probability of tumor incidence, The severity of tumor symptoms, The patient's physical sign level.

[0060] The present invention obtains a pathological malignancy calculation formula by using a specific mathematical model and verifying it, which is used to perform pathological quantitative calculation on the length and area of ​​tumor tissue pathological sections. The formula fully considers the clinical pathological malignancy. , length of tumor tissue pathology section , the weight corresponding to the slice length , tumor tissue pathological section area , the weight corresponding to the slice area , exponential function , the probability of tumor incidence , severity of tumor symptoms , patient's physical sign level According to the clinical pathological malignancy The correlation between the above parameters constitutes a functional relationship , this formula can realize the pathological quantitative calculation process of the length and area of ​​tumor tissue pathological sections. At the same time, by comprehensively considering multiple aspects of the tumor (such as tissue section length, area, probability of onset, severity of symptoms, level of physical signs, etc.), the formula can more comprehensively reflect the clinical pathological malignancy of the patient. This multi-factor combination method helps to improve the accurate assessment of the tumor and avoid the deviation caused by a single indicator. The formula quantifies the tumor pathological data (length, area, etc.) and combines other clinical data (such as probability of onset, symptoms, signs, etc.) to make the pathological malignancy have a quantitative standard. Through this comprehensive calculation, the malignancy of the patient's tumor in clinical pathology can be more accurately determined, thereby providing doctors with more scientific data support. By using a comprehensive analysis of factors such as the probability of onset, the time and frequency of symptom onset, the formula helps to identify the potential malignant risk of the tumor early. Early identification can promote early intervention, reduce the chance of tumor development, and improve the effect of tumor identification. In addition, this formula not only focuses on basic data such as the length and area of ​​the tumor slice, but also fully considers factors such as the patient's clinical history, symptoms and signs, making the use of pathological data more comprehensive and effective. This method can enhance the accuracy of clinical decision-making and provide a more accurate basis for subsequent research and clinical trials.

[0061] Furthermore, the present invention also provides a method for intelligent tumor identification based on big data, which is implemented based on the intelligent tumor identification system based on big data as described above. The method for intelligent tumor identification based on big data includes: Obtain the corresponding multimodal medical imaging data of tumors by connecting with various medical imaging devices, and obtain the corresponding patient pathological clinical data in real time by establishing interfaces with the hospital information management system and the electronic medical record system; perform image registration and filtering and cleaning on the multimodal medical imaging data of tumors and the patient pathological clinical data to obtain the multimodal standard imaging data of tumors and the patient clinical standard data; Obtain historical tumor image data with annotated tumor type, location, size, boundary, and malignancy, and build a corresponding tumor recognition model using a deep learning grid architecture; Convolutional neural network combined with computer vision technology is used to analyze tumor regional features of tumor multimodal standard imaging data to obtain tumor regional image feature vectors; pathological malignancy vectors are determined for patient clinical standard data to obtain patient clinical pathological malignancy feature vectors; big data fusion and splicing of tumor regional image feature vectors and patient clinical pathological malignancy feature vectors are performed to obtain patient tumor multi-source fusion feature vectors; The multi-source fusion feature vector of the patient's tumor is input into the constructed tumor recognition model to classify and identify the corresponding tumor area on the corresponding tumor medical image, so as to map the corresponding tumor type, location, size, boundary and malignancy level to the corresponding tumor medical image in an intuitive way, thereby outputting the corresponding medical image tumor recognition result.

[0062] Therefore, the embodiments should be regarded as illustrative and non-restrictive from all points, and the scope of the present invention is limited by the appended claims rather than the above description, and it is therefore intended that all changes falling within the meaning and range of equivalent elements of the application documents are included in the present invention.

[0063] The above description is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but should conform to the widest scope consistent with the principles and novel features invented herein.

Claims

1. A tumor intelligent identification system based on big data, characterized in that: Includes the following modules: The tumor big data processing module is used to obtain the corresponding tumor multimodal medical imaging data by connecting with various medical imaging devices, and to obtain the corresponding patient pathological clinical data in real time by establishing an interface with the hospital information management system and the electronic medical record system; image registration and filtering and cleaning are performed on the tumor multimodal medical imaging data and the patient pathological clinical data to obtain the tumor multimodal standard imaging data and the patient clinical standard data; The tumor recognition model building module is used to obtain the annotated historical image data of tumors corresponding to the annotated tumor type, location, size, boundary and malignancy, and build the corresponding tumor recognition model using the deep learning grid architecture; The tumor multi-source feature fusion module is used to use convolutional neural networks combined with computer vision technology to perform tumor region feature analysis on tumor multi-modal standard image data to obtain tumor region image feature vectors; determine the pathological malignancy vector of the patient's clinical standard data to obtain the patient's clinical pathological malignancy feature vector; perform big data fusion and splicing on the tumor region image feature vector and the patient's clinical pathological malignancy feature vector to obtain the patient's tumor multi-source fusion feature vector; The tumor classification and recognition mapping module is used to input the multi-source fusion feature vector of the patient's tumor into the constructed tumor recognition model to classify and recognize the corresponding tumor area on the corresponding tumor medical image, so as to classify and map the corresponding tumor type, location, size, boundary and malignancy degree to the corresponding tumor medical image in an intuitive way, thereby outputting the corresponding medical image tumor recognition result.

2. The tumor intelligent identification system based on big data according to claim 1 is characterized in that: The tumor big data processing module includes the following functions: By connecting with various medical imaging devices, corresponding multimodal medical imaging data of tumors can be acquired in real time; By establishing an interface with the hospital information management system and the electronic medical record system, the corresponding patient pathology clinical data, including medical history, tumor symptoms, signs and tumor pathology section examination results, can be obtained in real time; Performing image coordinate registration on each corresponding medical image in the tumor multimodal medical image data to generate corresponding tumor multimodal image registration data in the same coordinate system; Performing image pixel fuzziness analysis on each medical image in the corresponding tumor multimodal image registration data in the same coordinate system to obtain the pixel fuzziness corresponding to each medical image in the tumor multimodal image; performing pixel fuzziness filtering, denoising and standardization on each medical image in the corresponding tumor multimodal image registration data in the same coordinate system based on the pixel fuzziness corresponding to each medical image in the tumor multimodal image to obtain tumor multimodal standard image data; Each sub-data within the patient's pathological clinical data is cleaned and standardized to correct and fill in the corresponding errors and missing values ​​in each sub-data, and each sub-data is unified into the same data format to obtain the patient's clinical standard data.

3. The tumor intelligent identification system based on big data according to claim 2 is characterized in that: The tumor multimodal medical image data includes tumor CT images, tumor MRI images and tumor PET images.

4. The tumor intelligent identification system based on big data according to claim 2 is characterized in that: The image coordinate registration of each medical image corresponding to the tumor multimodal medical image data includes: Extracting image feature points of each medical image corresponding to the tumor multimodal medical image data to obtain an image feature point set corresponding to each tumor medical image; The spatial relationship between the image feature points in the same tumor medical image is obtained through each feature point in the image feature point set corresponding to each tumor medical image, and similarity quantization calculation is performed between each feature point in the same tumor medical image in the image feature point set corresponding to each tumor medical image to obtain the similarity measurement between the image feature points in the same tumor medical image; According to the spatial relationship and similarity measurement between the image feature points in the same tumor medical image, the corresponding image feature point set is constructed with a feature association topological structure to determine the edge between each feature point, and the edges are weighted according to the corresponding spatial relationship and similarity measurement, so as to construct a topological structure reflecting the global association between the image feature points in the same tumor medical image, so as to generate the feature association topological structure of each tumor medical image; Based on the corresponding feature points and edges in the topological structure of each tumor medical image feature association, the topological structure registration estimation is performed between the feature points of different modal images to solve the registration correspondence between the feature points of different modal images, and the registration estimation is performed by considering the symmetry between the associated topological structures to generate the tumor multimodal image registration transformation matrix; Based on the tumor multimodal image registration transformation matrix, image coordinate registration is performed on each corresponding medical image in the tumor multimodal medical image data to generate corresponding tumor multimodal image registration data in the same coordinate system.

5. The big data-based tumor intelligent identification system according to claim 4 is characterized in that: The image coordinate registration of each corresponding medical image in the tumor multimodal medical image data based on the tumor multimodal image registration transformation matrix includes: Based on the tumor multimodal image registration transformation matrix, the image coordinates of each corresponding medical image in the tumor multimodal medical image data are initially registered and transformed to generate corresponding tumor multimodal initial registration data in the same coordinate system; The mean square error of characteristic distances between various tumor medical images in the corresponding tumor multimodal initial registration data in the same coordinate system is calculated to obtain the mean square error of characteristic distances between various tumor medical images after registration; According to the preset error threshold, the mean square error of the characteristic distance between each tumor medical image after registration is compared and judged. If the mean square error of the characteristic distance is less than the preset error threshold, iterative registration is not required; if the mean square error of the characteristic distance is greater than or equal to the preset error threshold, the corresponding registration transformation matrix is ​​iteratively optimized by introducing rigid body transformation constraints and selecting the corresponding anatomical feature points around the tumor in each tumor medical image. At the same time, the optimized registration transformation matrix is ​​used to iteratively register the corresponding medical images to generate corresponding tumor multimodal image registration data in the same coordinate system.

6. The tumor intelligent identification system based on big data according to claim 1 is characterized in that: The tumor recognition model building module includes the following functions: Obtaining annotated historical image data of tumors corresponding to the annotated tumor type, location, size, boundary, and malignancy; The corresponding tumor type, location, size, boundary and degree of malignancy are combined into a tumor feature vector corresponding to the tumor annotated image, and each sub-feature in the corresponding tumor feature vector is input into the corresponding grid units at different levels in the deep learning grid architecture, where each grid unit adopts a different convolution kernel and activation function. At the same time, cross-layer connection and feedback mechanism are introduced between each grid unit, and the tumor annotated image corresponding to the tumor annotated historical image data is input into the deep learning grid architecture for classification mapping recognition training between image and feature, so as to construct the corresponding tumor recognition model.

7. The tumor intelligent identification system based on big data according to claim 1 is characterized in that: The tumor multi-source feature fusion module includes the following functions: A multi-scale convolution module is constructed using a convolutional neural network to analyze the tumor size characteristics of the tumor area in the tumor multimodal standard image data, so as to use the multi-scale convolution module to capture the characteristic information corresponding to tumors at different positions and sizes of the tumor image, so as to obtain the location and size sub-characteristics of the tumor area; The corresponding tumor image in the tumor multimodal standard image data is binarized to generate tumor multimodal binarized image data; the farthest non-zero pixel point of the corresponding binary image in the tumor multimodal binarized image data is determined to obtain the farthest non-zero pixel point corresponding to the tumor area in the binarized image at 0°, 90°, 180° and 270°, and obtain the optimal non-zero pixel point position of the tumor multimodal image at four azimuth angles; Based on the optimal non-zero pixel points of the tumor multimodal image at four azimuth angles, the corresponding maximum circumscribed rectangular boundary of the tumor region is constructed, and the corresponding tumor image in the tumor multimodal standard image data is segmented based on the maximum circumscribed rectangular boundary of the tumor region to generate a tumor multimodal region boundary segmentation image; the corresponding tumor region boundary in the tumor multimodal region boundary segmentation image is optimized and the boundary feature is detected to obtain the tumor region boundary sub-feature; The tumor region position and size sub-features and the tumor region boundary sub-features are combined into corresponding feature vectors to obtain a tumor region image feature vector; Determine the pathological malignancy vector of the patient's clinical standard data to obtain the patient's clinical pathological malignancy feature vector; The image feature vectors of the tumor region and the clinical pathological malignancy feature vectors of the patients are fused and spliced ​​by big data to obtain the multi-source fusion feature vectors of the patient's tumor.

8. The big data-based tumor intelligent identification system according to claim 7, characterized in that: Determining the pathological malignancy vector based on the patient's clinical standard data includes: Obtain the corresponding tumor incidence probability through the corresponding medical history in the patient's clinical standard data; Obtain the corresponding tumor symptom onset time and tumor symptom onset frequency through the corresponding tumor symptoms in the patient's clinical standard data, and perform tumor severity assessment and calculation based on the tumor symptom onset time and tumor symptom onset frequency to obtain the severity of the tumor symptoms; Obtaining corresponding patient physical sign levels through corresponding physical signs in the patient's clinical standard data; The corresponding tumor tissue pathology section length and tumor tissue pathology section area are obtained through the corresponding tumor pathology section examination results in the patient's clinical standard data, and the pathological malignancy calculation formula is used to perform pathological quantitative calculation on the tumor tissue pathology section length and tumor tissue pathology section area based on the probability of tumor incidence, the severity of tumor symptoms, and the patient's physical sign level, so as to obtain the clinical pathological malignancy; The probability of tumor onset, time of tumor symptom onset, frequency of tumor symptom onset, severity of tumor symptoms, patient's physical sign level, length of tumor tissue pathology section, area of ​​tumor tissue pathology section and clinical pathology malignancy are combined into corresponding feature vectors to obtain the patient's clinical pathology malignancy feature vector.

9. The big data-based tumor intelligent identification system according to claim 8, characterized in that: The formula for calculating the pathological malignancy is specifically: ； In the formula, The clinical pathological malignancy is the length of the tumor tissue pathological section, is the weight corresponding to the slice length, is the area of ​​tumor tissue pathological section, is the weight corresponding to the slice area, is an exponential function, is the probability of tumor incidence, The severity of tumor symptoms, The patient's physical sign level.

10. A tumor intelligent identification method based on big data, characterized in that: The method is implemented based on the big data-based tumor intelligent identification system according to claim 1. The big data-based tumor intelligent identification method includes: Obtain the corresponding multimodal medical imaging data of tumors by connecting with various medical imaging devices, and obtain the corresponding patient pathological clinical data in real time by establishing interfaces with the hospital information management system and the electronic medical record system; perform image registration and filtering and cleaning on the multimodal medical imaging data of tumors and the patient pathological clinical data to obtain the multimodal standard imaging data of tumors and the patient clinical standard data; Obtain historical tumor image data with annotated tumor type, location, size, boundary, and malignancy, and build a corresponding tumor recognition model using a deep learning grid architecture; Convolutional neural network combined with computer vision technology is used to analyze tumor regional features of tumor multimodal standard imaging data to obtain tumor regional image feature vectors; pathological malignancy vectors are determined for patient clinical standard data to obtain patient clinical pathological malignancy feature vectors; big data fusion and splicing of tumor regional image feature vectors and patient clinical pathological malignancy feature vectors are performed to obtain patient tumor multi-source fusion feature vectors; The multi-source fusion feature vector of the patient's tumor is input into the constructed tumor recognition model to classify and identify the corresponding tumor area on the corresponding tumor medical image, so as to map the corresponding tumor type, location, size, boundary and malignancy level to the corresponding tumor medical image in an intuitive way, thereby outputting the corresponding medical image tumor recognition result.

Images