A cervical cancer recognition system and method based on magnetic resonance image processing

Abstract

The present application belongs to the technical field of cervical cancer identification, and particularly relates to a cervical cancer identification system and method based on magnetic resonance image processing. Through spatial registration and unified coordinate system processing, the present application enables accurate cross-modality fusion of multi-sequence images. By constructing an anatomical level propagation structure data, the present application combines lesion analysis with real tissue structure, which is conducive to reflecting the actual infiltration path. By combining the blood flow change information of dynamic enhanced images with morphological features, the present application effectively expresses the abnormal blood supply of tumors, improves the identification accuracy, automatically screens key cervical candidate instances under limited labeling conditions, reflects the diagnostic contribution through instance attention weight, which is conducive to improving the identification efficiency, extracts the infiltration propagation parameters based on the anatomical level, quantitatively expresses the lesion expansion trend, provides a basis for clinical judgment, and reduces the influence of incidental noise or artifacts on lesion identification through cross-sequence difference analysis and three-dimensional stability screening.

Description

Technical Field

[0001] This invention belongs to the field of cervical cancer identification technology, specifically relating to a cervical cancer identification system and method based on magnetic resonance imaging processing. Background Technology

[0002] Cervical cancer is one of the most common malignant tumors in women, ranking among the top female reproductive system malignancies globally. Early detection and accurate staging are crucial for developing scientific treatment plans, improving patient survival rates, and enhancing quality of life. Traditional diagnostic methods for cervical cancer primarily rely on cervical cytology, tissue biopsy, and imaging examinations. While cytology and biopsy can directly obtain pathological information, they are invasive procedures and carry risks of uneven sampling or missed lesions, making it difficult to comprehensively reflect the extent of the lesion and changes in tissue structure.

[0003] In contrast, magnetic resonance imaging (MRI) is non-invasive and high-resolution, clearly displaying the anatomical structure and lesion morphology of the cervix and surrounding tissues, especially in terms of soft tissue contrast, significantly superior to CT and ultrasound. T2-weighted images reflect tissue morphology and structure, while dynamic contrast-enhanced (DCE) images provide hemodynamic information about the lesion, showing lesion perfusion and angiogenesis, providing important evidence for differentiating between benign and malignant lesions and assessing the extent of local invasion. Therefore, MRI plays an irreplaceable role in the early diagnosis, lesion localization, and clinical staging of cervical cancer, and has become a commonly used imaging examination method in clinical practice.

[0004] However, despite the advantages of MRI in terms of high resolution and multiple sequences, traditional image-based diagnosis still faces a series of challenges. Clinically acquired MRI data is usually multi-sequence and multi-temporal data. There are spatial and signal characteristics differences between different sequences, making information integration difficult. At the same time, the complex anatomical structure of the cervix and its surrounding tissues means that subtle changes in lesions may be overlooked. Therefore, relying solely on traditional imaging diagnostic methods has certain limitations in early lesion detection, lesion extent assessment, and judgment of local invasion trends. Summary of the Invention

[0005] The purpose of this invention is to provide a cervical cancer identification system and method based on magnetic resonance imaging processing, which can identify key lesion areas and conduct comprehensive evaluation of cervical cancer, thereby improving the accuracy of early diagnosis and clinical decision support capabilities.

[0006] The specific technical solution adopted by this invention is as follows: A method for cervical cancer identification based on magnetic resonance imaging (MRI) image processing includes: Acquire pelvic magnetic resonance imaging data of the object to be detected, perform spatial registration processing on the pelvic magnetic resonance imaging data, and obtain magnetic resonance imaging dataset. Based on a pre-defined pelvic anatomical structure model, the magnetic resonance imaging dataset is divided into anatomical regions to obtain cervical region information. Based on the cervical region information, anatomical hierarchical propagation structure data is constructed. Hemodynamic curves were constructed based on magnetic resonance imaging datasets, and hemodynamic feature maps were generated in a unified coordinate system. A deep neural network was constructed to obtain cervical tissue morphological depth features and hemodynamic depth features based on magnetic resonance imaging datasets, obtain image diagnostic labels, perform weakly supervised learning training on the deep neural network, and obtain key cervical candidate instances and corresponding instance attention weights. Based on instance-based attention weights and hemodynamic feature maps, morphological depth features and hemodynamic depth features are fused to obtain a fusion vector. Infiltration propagation parameters are obtained based on anatomical hierarchical propagation structure data and the fusion vector. Based on the degree of difference between different sequence image signals in the magnetic resonance imaging dataset, signal difference parameters are obtained. When the signal difference parameters meet the preset abnormal conditions, segmentation processing is performed to obtain stable three-dimensional region data of cervical lesions. Based on the three-dimensional regional data of stable cervical lesions and combined with infiltration and propagation parameters, a comprehensive assessment feature set of lesions is constructed, and cervical cancer identification results are obtained based on the comprehensive assessment feature set of lesions.

[0007] In a preferred embodiment, pelvic magnetic resonance imaging (MRI) data of the object to be detected is acquired, and spatial registration processing is performed on the pelvic MRI data to obtain an MRI dataset, including: Acquire pelvic magnetic resonance imaging data of the subject to be examined, including T2-weighted image data and dynamic contrast-enhanced image data; Image quality verification and abnormal image removal were performed on T2-weighted image data and dynamic contrast-enhanced image data to obtain effective magnetic resonance image data. Spatial alignment processing is performed on the effective magnetic resonance image data based on a preset reference sequence to obtain spatially aligned magnetic resonance image data. Spatial registration is performed on spatially aligned magnetic resonance images to obtain spatially registered magnetic resonance image data. The T2-weighted image data and dynamic contrast-enhanced image data from the spatially registered magnetic resonance imaging data are integrated to construct a magnetic resonance imaging dataset.

[0008] In a preferred embodiment, the magnetic resonance imaging dataset is divided into anatomical regions based on a pre-defined pelvic anatomical structure model to obtain cervical region information. Based on this cervical region information, hierarchical anatomical propagation structure data is constructed, including: Obtain a preset pelvic anatomical structure model, which includes cervical epithelial layer structure information, cervical stroma layer structure information, and cervical surrounding tissue structure information. Based on the magnetic resonance imaging dataset and combined with the pelvic anatomical structure model, the magnetic resonance imaging dataset is processed for anatomical structure recognition to obtain the initial anatomical region segmentation results. Cervical region information is extracted based on the initial anatomical region segmentation results. The cervical region information includes the cervical epithelial region, the cervical stroma region, and the pericervical tissue region. Obtain the region boundary data and region location parameters based on cervical region information; Based on the regional boundary data and regional location parameters, a regional connectivity topology table is constructed, and anatomical hierarchical propagation structure data is generated based on the regional connectivity topology table.

[0009] In a preferred embodiment, hemodynamic curves are constructed based on a magnetic resonance imaging dataset, and a hemodynamic feature map is generated in a unified coordinate system, including: Extracting dynamic contrast-enhanced image data from magnetic resonance imaging datasets; The dynamic contrast-enhanced image data is decomposed over time to extract the enhanced signal intensity data of multiple time phases; Based on the enhanced signal intensity data, construct a set of hemodynamic curve parameters corresponding to each pixel; Under a unified spatial coordinate system, the set of hemodynamic curve parameters and T2-weighted image data are spatially mapped and fused. Based on the fusion processing results, a hemodynamic feature map is generated.

[0010] In a preferred embodiment, a deep neural network is constructed. Based on a magnetic resonance imaging dataset, morphological depth features and hemodynamic depth features of cervical tissue are obtained, and image diagnostic labels are acquired. The deep neural network is then trained using weakly supervised learning to obtain key cervical candidate instances and their corresponding instance attention weights, including: Construct a deep neural network, wherein the deep neural network includes a first feature extraction channel and a second feature extraction channel; Based on cervical region information, the cervical region in the magnetic resonance imaging dataset is divided into blocks under a unified spatial coordinate system to generate multiple cervical candidate image blocks, and candidate instances are labeled. Based on the candidate instances, the corresponding T2-weighted image data is extracted, and combined with the first feature extraction channel, cervical tissue morphological depth features are generated. Based on candidate instances, extract corresponding dynamic contrast-enhanced image data and combine it with the second feature extraction channel to generate hemodynamic depth features; The morphological depth features and hemodynamic depth features of cervical tissue are fused to generate instance fusion features; Obtain the image diagnostic labels corresponding to the magnetic resonance imaging dataset, and combine them with instance fusion features to perform weakly supervised learning training on the deep neural network to obtain instance classification response values ​​and instance attention weights. Based on the instance classification response value and instance attention weight, the contribution index of the candidate instance is obtained. Candidate instances with contribution index higher than the preset threshold are selected, marked as key cervical candidate instances, and the corresponding instance attention weight is obtained.

[0011] In a preferred embodiment, based on instance attention weights and hemodynamic feature maps, morphological depth features and hemodynamic depth features are fused to obtain a fused vector. Infiltration propagation parameters are then obtained based on anatomical hierarchical propagation structure data and the fused vector, including: Based on instance attention weights, morphological depth features and hemodynamic depth features are weighted and fused, and feature enhancement encoding is performed by combining hemodynamic feature maps to obtain a fusion vector; Obtain the anatomical hierarchical propagation structure data corresponding to the candidate instances, and perform mapping processing on the fusion vector to generate feature distribution results under different cervical anatomical levels; The fused feature vectors are mapped to the corresponding cervical anatomical level regions to form a feature distribution matrix for each anatomical level. The characteristic distribution differences between each anatomical level are obtained along the cervical canal, a propagation gradient sequence along the cervical canal is generated, and the cervical canal propagation intensity parameter is obtained based on the propagation gradient sequence along the cervical canal. The characteristic distribution differences of adjacent areas are obtained along the direction of the tissue surrounding the cervix, a propagation gradient sequence along the direction of the tissue surrounding the cervix is ​​generated, and the propagation intensity parameter of the tissue surrounding the cervix is ​​obtained based on the propagation gradient sequence along the direction of the tissue surrounding the cervix. By combining the cervical canal transmission intensity parameters and the transmission intensity parameters of the surrounding cervical tissues, lesion infiltration and transmission parameters are formed.

[0012] In a preferred embodiment, signal difference parameters are obtained based on the degree of difference between different sequence image signals in a magnetic resonance imaging dataset. When the signal difference parameters meet preset abnormality conditions, segmentation processing is performed to obtain stable three-dimensional region data of cervical lesions, including: Acquire magnetic resonance imaging datasets and extract the signal intensity values ​​of corresponding pixels in a unified spatial coordinate system from different image sequences. The different image sequences include T2-weighted image data and dynamic contrast-enhanced image data. The signal intensity values ​​of different sequence image data are standardized, and the signal difference metric value of each pixel between different sequences is obtained to construct a signal difference parameter map; Obtain a preset anomaly judgment threshold, compare the signal difference parameter map with the preset anomaly judgment threshold, and obtain candidate anomaly regions that meet the anomaly conditions; Based on the candidate abnormal regions, region segmentation processing is performed in conjunction with the corresponding sequence image data to obtain candidate regions for cervical lesions; The candidate regions of cervical lesions were reconstructed in three dimensions, and the stability of the reconstruction results was screened to obtain stable three-dimensional region data of cervical lesions.

[0013] In a preferred embodiment, a comprehensive assessment feature set for lesions is constructed based on three-dimensional regional data of stable cervical lesions and combined with invasive spread parameters. Cervical cancer identification results are obtained based on this comprehensive assessment feature set, including: Based on stable cervical lesion three-dimensional region data, lesion parameters and three-dimensional lesion regions are obtained. Among them, lesion parameters include lesion volume parameters, maximum diameter parameters and morphological regularity parameters. Texture heterogeneity features are obtained from the three-dimensional lesion area, and enhancement non-uniformity features are extracted from the dynamic contrast-enhanced image data. By integrating lesion parameters, texture heterogeneity features, enhancement non-uniformity features, and infiltration and propagation parameters, a comprehensive lesion assessment feature set is constructed. The system acquires a preset identification strategy and combines it with a comprehensive lesion assessment feature set to obtain cervical cancer identification results. These results include benign / malignant classification results, local invasion risk scores, lesion invasion direction assessment information, and structured assessment data for clinical staging.

[0014] The present invention also provides a cervical cancer identification system based on magnetic resonance image processing, used in the above-mentioned cervical cancer identification method based on magnetic resonance image processing, comprising: The image data module is used to acquire pelvic magnetic resonance imaging data of the object to be detected, and to perform spatial registration processing on the pelvic magnetic resonance imaging data to obtain a magnetic resonance imaging dataset. The propagation structure module divides the magnetic resonance imaging dataset into anatomical regions based on a pre-set pelvic anatomical structure model to obtain cervical region information, and constructs anatomical hierarchical propagation structure data based on the cervical region information. The feature map module constructs hemodynamic curves based on magnetic resonance imaging datasets and generates hemodynamic feature maps in a unified coordinate system; The deep feature module is used to construct a deep neural network. It obtains cervical tissue morphological depth features and hemodynamic depth features based on the magnetic resonance imaging dataset, obtains image diagnostic labels, performs weakly supervised learning training on the deep neural network, and obtains key cervical candidate instances and corresponding instance attention weights. The infiltration propagation module, based on instance attention weights and hemodynamic feature maps, fuses morphological depth features and hemodynamic depth features to obtain a fusion vector, and obtains infiltration propagation parameters based on anatomical hierarchical propagation structure data and the fusion vector; The three-dimensional region module obtains signal difference parameters based on the degree of difference between different sequence image signals in the magnetic resonance imaging dataset. When the signal difference parameters meet the preset abnormal conditions, segmentation processing is performed to obtain stable three-dimensional region data of cervical lesions. The lesion identification module is used to construct a comprehensive lesion assessment feature set based on the three-dimensional regional data of stable cervical lesions and the invasion and spread parameters, and to obtain cervical cancer identification results based on the comprehensive lesion assessment feature set.

[0015] And, a cervical cancer identification terminal based on magnetic resonance imaging processing, comprising: One or more processors; A storage device on which one or more programs are stored; When one or more programs are executed by one or more processors, the one or more processors implement a cervical cancer identification method based on magnetic resonance imaging.

[0016] The technical effects achieved by this invention are as follows: This invention enables accurate cross-modal fusion of multi-sequence images through spatial registration and unified coordinate system processing. By constructing anatomical hierarchical propagation structure data, it combines lesion analysis with real tissue structure, which helps to reflect the actual invasion path. By combining blood flow change information and morphological features of dynamically enhanced images, it effectively expresses abnormal tumor blood supply and improves identification accuracy. Under limited annotation conditions, it automatically screens key cervical candidate instances and reflects diagnostic contribution through instance attention weights, which helps to improve identification efficiency. Based on anatomical hierarchy, it extracts invasion propagation parameters, which can quantify the lesion expansion trend and provide a basis for clinical judgment. Through cross-sequence differential analysis and three-dimensional stability screening, it reduces the impact of occasional noise or artifacts on lesion identification. It outputs structured assessment data containing classification, risk score, and staging information, which can be directly used for clinical decision support. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method provided by the present invention; Figure 2 This is a system module diagram provided by the present invention. Detailed Implementation

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0019] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0020] Secondly, the term "an embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in a preferred embodiment" appearing in different places throughout this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that mutually excludes other embodiments.

[0021] Furthermore, the present invention will be described in detail with reference to the schematic diagrams. When describing the embodiments of the present invention in detail, the schematic diagrams are merely examples for ease of explanation and should not limit the scope of protection of the present invention.

[0022] Please see the appendix Figure 1 As shown, a cervical cancer identification method based on magnetic resonance imaging processing is provided, including: S1. Obtain pelvic magnetic resonance imaging data of the object to be detected, and perform spatial registration processing on the pelvic magnetic resonance imaging data to obtain a magnetic resonance imaging dataset. S2. Based on the preset pelvic anatomical structure model, the magnetic resonance imaging dataset is divided into anatomical regions to obtain cervical region information. Based on the cervical region information, anatomical hierarchical propagation structure data is constructed. S3. Construct hemodynamic curves based on magnetic resonance imaging datasets and generate hemodynamic feature maps in a unified coordinate system; S4. Construct a deep neural network, obtain cervical tissue morphological depth features and hemodynamic depth features based on the magnetic resonance imaging dataset, obtain image diagnostic labels, perform weakly supervised learning training on the deep neural network, and obtain key cervical candidate instances and corresponding instance attention weights. S5. Based on instance attention weights and hemodynamic feature maps, morphological depth features and hemodynamic depth features are fused to obtain a fusion vector. Infiltration propagation parameters are obtained based on anatomical hierarchical propagation structure data and the fusion vector. S6. Based on the degree of difference between different sequence image signals in the magnetic resonance imaging dataset, obtain signal difference parameters. When the signal difference parameters meet the preset abnormal conditions, perform segmentation processing to obtain stable three-dimensional region data of cervical lesions. S7. Based on the three-dimensional regional data of stable cervical lesions and combined with the invasion and spread parameters, construct a comprehensive assessment feature set of lesions, and obtain the cervical cancer identification results based on the comprehensive assessment feature set of lesions.

[0023] As described in steps S1 to S7 above, pelvic magnetic resonance imaging (MRI) data of the subject is acquired, and spatial registration processing is performed on different sequence images to form an MRI dataset with T2-weighted image data and dynamic contrast-enhanced image data in a unified spatial coordinate system. This unified spatial representation eliminates spatial offsets caused by different scanning times, postures, or equipment parameters. In this unified coordinate system, a preset pelvic anatomical structure model is used to divide the MRI dataset into anatomical regions, extracting the cervical epithelial region, cervical stroma region, and pericervical tissue region. Based on the region boundaries and spatial locations, anatomical hierarchical propagation structure data is constructed. This structure is used to express the spatial hierarchical distribution of cervical tissue from the inside out. The dynamic contrast-enhanced image data is then subjected to temporal... Sequence analysis was performed to extract enhancement signal changes across multiple time phases, constructing hemodynamic curves. These curves were then fused with T2 imaging data in a unified spatial coordinate system to form a hemodynamic feature atlas. This atlas reflects the blood perfusion changes in the lesion region. Candidate image patches were generated within the cervical region. Morphological depth features and hemodynamic depth features were obtained through a dual-channel depth feature extraction structure. Weakly supervised learning was then combined with image diagnostic labels to obtain key cervical candidate instances and their corresponding attention weights. This weakly supervised approach allows for the identification of key regions with high diagnostic contribution under limited annotation conditions. Instance attention weights reflect the importance of different regions in diagnosis. Based on these instance attention weights and the hemodynamic feature atlas... Morphological depth features and hemodynamic depth features are fused to form a fusion vector. Then, combined with anatomical layer propagation data, the fusion vector is analyzed for directionality at different anatomical layers to obtain infiltration and propagation parameters of the lesion along the cervical canal and surrounding tissues. By combining anatomical structure and imaging features, the lesion expansion trend is given a clear spatial direction. Signal differences between different sequences in the magnetic resonance imaging dataset are analyzed to form a signal difference parameter map. When abnormal conditions are met, region segmentation is triggered. A three-dimensional lesion region is then formed by spatially connecting continuous slices. Stability screening is used to obtain stable cervical lesion three-dimensional region data. Based on the stable cervical lesion three-dimensional region data, lesion volume, maximum diameter, and shape are extracted. The system constructs a comprehensive lesion assessment feature set by combining regularity parameters with texture heterogeneity, enhancement non-uniformity, and invasion propagation parameters. Then, based on a preset identification strategy, it obtains cervical cancer identification results. These results include not only benign / malignant classification but also local invasion risk scores, invasion direction assessment information, and structured data related to clinical staging, thus forming a complete image assessment output. Through spatial registration and unified coordinate system processing, accurate cross-modal fusion of multi-sequence images is achieved. By constructing anatomical hierarchical propagation structure data, lesion analysis is combined with real tissue structure, which helps reflect the actual invasion path. Combining blood flow changes from dynamically enhanced images with morphological features effectively expresses abnormal tumor blood supply, improving identification accuracy.Automatically selecting key cervical candidate instances under limited annotation conditions and reflecting diagnostic contribution through instance attention weights improves identification efficiency. Extracting infiltration and propagation parameters based on anatomical hierarchy quantifies lesion expansion trends, providing a basis for clinical judgment. Cross-sequence differential analysis and three-dimensional stability screening reduce the impact of incidental noise or artifacts on lesion identification. The system outputs structured assessment data containing classification, risk scores, and staging information, which can be directly used for clinical decision support.

[0024] In a preferred embodiment, pelvic magnetic resonance imaging (MRI) data of the object to be detected is acquired, and spatial registration processing is performed on the pelvic MRI data to obtain an MRI dataset, including: S101. Obtain pelvic magnetic resonance imaging data of the object to be tested, wherein the pelvic magnetic resonance imaging data includes T2-weighted image data and dynamic contrast-enhanced image data. S102. Perform image quality verification and abnormal image removal on T2-weighted image data and dynamic contrast-enhanced image data to obtain effective magnetic resonance image data. S103. Perform spatial alignment processing on the effective magnetic resonance image data based on the preset reference sequence to obtain spatially aligned magnetic resonance image data. S104. Perform spatial registration processing on the spatially aligned magnetic resonance images to obtain spatially registered magnetic resonance image data. S105. Integrate the T2-weighted image data and dynamic contrast-enhanced image data from the spatially registered magnetic resonance imaging data to construct a magnetic resonance imaging dataset.

[0025] As described in steps S101 to S105 above, pelvic magnetic resonance imaging (MRI) images containing T2-weighted and dynamic contrast-enhanced image data are acquired. This ensures that the acquired data simultaneously contains information on tissue structure and enhancement changes. T2-weighted images primarily reflect the morphology, boundaries, and hierarchical relationships of cervical tissue, while dynamic contrast-enhanced images reflect blood perfusion and enhancement processes. These two types of images complement each other in terms of information dimensions. By performing quality checks on the original images, images with motion artifacts, signal loss, scan truncation, or enhancement abnormalities are identified and removed. This ensures the stability and consistency of the image data entering the processing flow, reducing the interference of abnormal images on spatial alignment and feature expression at the source. It also guarantees the comparability of different sequences in signal expression. Using a preset reference sequence as a spatial benchmark, spatial alignment processing is performed on the effective MRI image data, ensuring that images from different sequences maintain consistency at the overall structural level. By unifying the spatial positional relationships of the images, the cervical region is aligned across different sequences. By maintaining a relatively consistent position within the image and eliminating macroscopic positional shifts caused by differences in scanning posture or slight patient movement, the images undergo spatial registration processing under a unified coordinate system after initial spatial alignment. This ensures that pixel positions in different image sequences correspond to the same anatomical structure region, thus forming a precise cross-sequence correspondence. This maintains a consistent spatial representation of the same cervical tissue region in different image sequences, enabling morphological and enhancement information to be correlated and analyzed in a unified location. The T2-weighted image data and dynamic contrast-enhanced image data that have undergone spatial registration processing are then integrated to form a magnetic resonance imaging dataset containing multi-sequence information. This dataset maintains consistency in spatial structure and achieves fusion representation in information dimensions. The processes of anatomical region division, candidate instance generation, and lesion analysis are all based on a unified data structure, transforming the original magnetic resonance images from a discrete, multi-source, and inconsistent state into a unified structured dataset, providing a stable data foundation for the cervical lesion identification process.

[0026] In a preferred embodiment, the magnetic resonance imaging dataset is divided into anatomical regions based on a pre-defined pelvic anatomical structure model to obtain cervical region information. Based on the cervical region information, anatomical hierarchical propagation structure data is constructed, including: S201. Obtain a preset pelvic anatomical structure model, wherein the pelvic anatomical structure model includes cervical epithelial layer structure information, cervical stroma layer structure information and cervical surrounding tissue structure information. S202. Based on the magnetic resonance imaging dataset and combined with the pelvic anatomical structure model, perform anatomical structure recognition processing on the magnetic resonance imaging dataset to obtain the initial anatomical region segmentation results. S203. Extract cervical region information based on the initial anatomical region segmentation results. The cervical region information includes the cervical epithelial region, the cervical stroma region, and the pericervical tissue region. S204. Obtain the region boundary data and region location parameters based on the cervical region information; S205. Based on the regional boundary data and regional location parameters, construct a regional connectivity topology table, and generate anatomical hierarchical propagation structure data based on the regional connectivity topology table.

[0027] As described in steps S201 to S205 above, a preset pelvic anatomical structure model is introduced as an anatomical recognition benchmark template. This model includes information on the cervical epithelial layer, cervical stroma, and surrounding tissue structures. It is used to define the spatial distribution and morphological characteristics of different tissue layers in the image. By using the structural outlines, hierarchical order, and spatial distribution characteristics in the model as reference constraints, the image analysis process can be carried out around the actual anatomical structure, rather than relying solely on image grayscale changes for segmentation. This reduces misjudgments of regions caused by blurred tissue boundaries or signal overlap. Under a unified spatial coordinate system, the pelvic anatomical structure model is mapped to the magnetic resonance imaging dataset, and structural matching processing is performed on different tissue regions in the image. The model matches the structural contours of each anatomical level with the corresponding signal regions in the images. Based on the differences in tissue morphology in T2-weighted images and the enhancement distribution characteristics in dynamic contrast-enhanced images, the model structure is locally adjusted. Through the fusion and recognition of structural contours and image edge information, initial anatomical region segmentation results are generated, dividing the cervical tissue in the images into initial regions with clear anatomical significance. Based on the initial anatomical region segmentation results, regions belonging to the cervical structure are screened. According to the hierarchical identifiers in the model, the screened regions are labeled as the cervical epithelial layer, cervical stroma, and pericervical tissue regions, respectively. Independent spatial region identifier information is generated for each level region, forming a cervical region information set. The anatomical structures in the image are transformed into a hierarchical data representation. After obtaining information about the cervical region, a three-dimensional voxel set is extracted for each anatomical level region. By detecting the boundary positions between the internal and external voxels of the region, a boundary voxel set is obtained. The boundary voxels are then processed continuously to form the outer contour boundary data of the region. The coordinate sequence of the boundary points is recorded in three-dimensional space to form the region boundary description information. The spatial range of the region in a unified coordinate system is extracted, the coordinates of the region's center position are obtained, the spatial extension range of the region in the anterior-posterior, left-right, and up-down directions is obtained, and the spatial offset information of the region relative to the center of the overall pelvic structure is obtained. This gives each anatomical level region a clear spatial location expression and boundary description. Each anatomical level region is then... As topological nodes, based on regional boundary data, it identifies whether two regions have boundary contact or continuous distribution. If two regions have adjacent or overlapping boundaries in space, a node connection record is established. The connection record stores the connection region number and the location of the contact boundary, forming a regional connection topology table containing all regional connection situations. Based on the connection order of each region in the topology table, the cervical epithelium, cervical stroma, and surrounding tissues are arranged hierarchically. According to the direction of the cervical canal and the direction of expansion of surrounding tissues, a hierarchical propagation path sequence between regions is established. The path sequence and regional hierarchical identifiers are integrated to form anatomical hierarchical propagation structure data, which is used to express the continuous distribution characteristics between different anatomical levels and provide a spatial basis for the analysis of lesion invasion direction.

[0028] In a preferred embodiment, hemodynamic curves are constructed based on a magnetic resonance imaging dataset, and a hemodynamic feature map is generated in a unified coordinate system, including: S301. Extracting dynamic contrast-enhanced image data based on magnetic resonance imaging dataset; S302. Perform time-series decomposition processing on the dynamic contrast enhancement image data to extract the enhancement signal intensity data of multiple time phases; S303. Based on the enhanced signal intensity data, construct a set of hemodynamic curve parameters corresponding to each pixel. S304. Under a unified spatial coordinate system, the hemodynamic curve parameter set and T2-weighted image data are spatially mapped and fused. Based on the fusion processing results, a hemodynamic feature map is generated.

[0029] As described in steps S301 to S304 above, dynamic contrast-enhanced image data is extracted from the magnetic resonance image dataset after spatial registration. Based on the extracted enhancement start point, all time frames are uniformly mapped to the standard time axis. Time frames with inconsistent sampling intervals are resampled to form time series data with equal time intervals. The time frame before enhancement is selected as the baseline signal, and baseline subtraction is performed on subsequent time frames to obtain the true enhancement change amplitude. According to the enhancement signal change trend, the time series is divided into the initial enhancement stage (signal rapid rise stage), peak stage (signal reaches maximum stage), and delay stage (signal falls back or stabilizes stage). The division method includes detecting the signal change slope change point, identifying the signal peak time point, and segmenting the stage according to the time interval. The early enhancement signal intensity, peak enhancement signal intensity, and delay period signal intensity are extracted in each stage, ultimately forming enhancement signal intensity data with multiple time phases. After completing the time phase extraction, the time series signal of each spatial pixel is extracted, and the enhancement signals are arranged in chronological order to obtain signal change trend description information, including enhancement start position, signal strength, and signal intensity. The signal rise amplitude, peak duration, and signal attenuation trend are used to structure and organize temporal change information. A hemodynamic curve parameter set is established for each pixel to express the enhancement behavior characteristics of the tissue region. The curve parameter set of each pixel is mapped back to a unified spatial coordinate system and aligned with the corresponding T2 image position to form a spatially consistent parameter distribution matrix. For different enhancement behaviors, the parameters are expressed in layers (e.g., enhancement velocity layer, enhancement amplitude layer, peak duration layer, and delay change layer), with each layer forming an independent spatial distribution map. Different parameter layers are superimposed on a unified spatial structure, combined with the tissue contour in the T2 image for regional restriction, and the parameters within the same region are spatially consistent. Finally, a hemodynamic feature atlas is formed, which includes the enhancement behavior distribution at each spatial location, the spatial changes of different enhancement stages, and the hemodynamic expression corresponding to the tissue structure. The blood flow changes are expressed in atlas form, so that the enhancement trend of the lesion can be continuously distributed in space, which is conducive to the analysis of the direction of propagation. The enhancement behavior of each region can be traced back to a specific spatial location, enhancing the medical interpretability of the results.

[0030] In a preferred embodiment, a deep neural network is constructed. Based on a magnetic resonance imaging dataset, morphological depth features and hemodynamic depth features of cervical tissue are obtained, and image diagnostic labels are acquired. The deep neural network is then trained using weakly supervised learning to obtain key cervical candidate instances and their corresponding instance attention weights, including: S401. Construct a deep neural network, wherein the deep neural network includes a first feature extraction channel and a second feature extraction channel; S402. Based on cervical region information, the cervical region in the magnetic resonance imaging dataset is divided into blocks under a unified spatial coordinate system to generate multiple cervical candidate image blocks and to label candidate instances. S403. Extract the corresponding T2-weighted image data based on the candidate instances, and combine it with the first feature extraction channel to generate cervical tissue morphological depth features; S404. Extract the corresponding dynamic contrast-enhanced image data based on the candidate instance, and combine it with the second feature extraction channel to generate hemodynamic depth features; S405. Fusion processing is performed on the morphological depth features and hemodynamic depth features of cervical tissue to generate instance fusion features; S406. Obtain the image diagnostic labels corresponding to the magnetic resonance imaging dataset, and combine them with instance fusion features to perform weakly supervised learning training on the deep neural network to obtain instance classification response values ​​and instance attention weights. S407. Obtain the contribution index of candidate instances based on the instance classification response value and instance attention weight, filter candidate instances with contribution index higher than the preset threshold, mark them as key cervical candidate instances, and obtain the corresponding instance attention weight.

[0031] As described in steps S401 to S407 above, to achieve the synergistic expression of cervical tissue structural features and hemodynamic features, a deep feature extraction network is constructed, comprising a first feature extraction channel and a second feature extraction channel. The first feature extraction channel is used to extract cervical tissue morphological features. Its network structure consists of multi-level convolutional feature extraction modules. Each module includes a convolutional layer, a normalization layer, and an activation layer. The convolutional kernel size can be 1×1, 3×3, or 5×5, the convolutional stride can be 1 or 2, and the number of output channels in the convolutional layer can be progressively increased from 32 to 512. The number of convolutional modules can be set to 3 to 6, and each module contains 2 to 4 convolutional layers. Residual connections or skip connections can be set between modules to enhance the deep structural features. The second feature extraction channel is used to extract hemodynamic features from dynamic contrast-enhanced images. Its input data is an enhanced image sequence containing multiple temporal phases, ranging from 5 to 30. This channel first extracts enhancement signal change features through a temporal feature extraction layer (with a temporal convolution kernel size of 3, 5, or 7 and 16 to 64 convolutional channels). Subsequently, it extracts spatial blood flow distribution features through a spatial convolution layer (with a spatial convolution kernel size of 3×3 or 5×5 and 64 to 256 output channels). Both channels share a unified input coordinate system. Under this unified spatial coordinate system, a three-dimensional cervical region is extracted based on cervical region information. The cervical region is then divided into sliding window blocks according to a preset size, preserving the coverage of cervical tissue. For example, image patches that reach a preset threshold are assigned a unique number to form a candidate instance set. Through instantiation processing, the overall image is transformed into a locally analyzable region. Gray-level normalization is performed on the candidate instances, background suppression is applied to the cervical region, tissue edge information is preserved, tissue texture at different scales is extracted, edge change information is extracted, tissue contour continuity features are extracted, spatial encoding of the internal tissue structure arrangement is performed, tissue density variation distribution is obtained, and the degree of tissue morphological irregularity is obtained. Layer-by-layer encoding is performed to form morphological depth features with spatial hierarchical expression capabilities, used to express abnormalities in cervical tissue structure. Pre-enhancement baseline signal is extracted, early temporal phase signal is obtained, early enhancement amplitude is extracted, and enhancement speed variation is obtained. The enhancement trend is mapped to an early enhancement parameter map, and enhancement distribution patterns, enhancement concentration, enhancement heterogeneity, and enhancement boundary changes are extracted. Hierarchical encoding is then performed to generate hemodynamic depth features to express abnormal tumor blood supply behavior. These two types of features are mapped to a unified feature space, aligning spatial dimensions and feature lengths. Based on the importance distribution of candidate instance regions, different fusion weights are assigned to morphological and blood flow features. Important regions are enhanced, and the two types of features are concatenated. The concatenated features are then cross-encoded to output instance fusion features that simultaneously contain information on structural and blood supply abnormalities. Image-level diagnostic labels are obtained and used as overall supervision signals. The fusion features of each candidate instance are used as training input.During training, the network outputs instance classification response values ​​and instance attention weights. In the weakly supervised learning phase, image-level diagnostic labels are used as overall supervisory signals and bound to the candidate instance set. Instance fusion features are used as input, and an instance-level response generation mechanism outputs instance classification response values ​​to reflect the degree to which candidate instances support the overall diagnostic result. Each pelvic MRI image is considered a whole sample, while multiple candidate regions divided within the cervical region are considered multiple instance regions. Each instance region is processed through morphological feature extraction and hemodynamic feature extraction channels to obtain corresponding instance feature expressions. A category response evaluation is performed on each instance feature to determine the association between the instance and cervical cancer. The response intensity of category relevance is normalized based on the relative differences in response intensity among instances, thus obtaining the attention weight of each instance in the overall image diagnosis. The attention weight reflects the contribution of different instance regions to the image diagnosis result; regions with higher response intensity receive higher attention weights, while regions with lower response intensity receive lower attention weights. After obtaining the attention weights of each instance, all instance features are weighted and fused according to their corresponding weights to form a comprehensive feature representation of the overall image. Based on this comprehensive feature representation, image-level diagnostic results are generated. During training, the generated image-level diagnostic results are compared with real image diagnostic labels to improve the network performance. The parameters are gradually adjusted to increase the response of instances relevant to the actual lesion area and decrease the response of irrelevant areas during the iteration process. This gradually concentrates the instance attention weights on the actual lesion area. Without the need for precise labeling of the lesion area, the network can automatically focus on and identify key lesion areas based solely on image-level labels, obtaining stable instance attention weights. This provides reliable feature basis for lesion invasion and spread analysis and 3D lesion area determination. Through the consistency constraint between the overall response of the instance set and the image-level diagnostic labels, the network gradually learns the feature performance of key cervical areas during training. After weakly supervised learning is completed, the instance classification response values ​​and instance attention weights are adjusted. The system performs joint expression processing to generate candidate instance contribution indicators. All candidate instances are then ranked and filtered according to preset screening rules. Instances with contribution indicators exceeding a threshold are marked as key cervical cancer candidate instances. Corresponding instance attention weights are output for infiltration and spread analysis and comprehensive lesion assessment. Through weak supervision, key region identification can be completed relying solely on image-level labels, reducing manual annotation costs and making cervical cancer identification less dependent on single image features, thus improving accuracy. The contribution screening mechanism enables automatic discovery of key cervical cancer candidate instances, demonstrating stronger identification capabilities for irregular tumors and early small lesions. Instance attention weights directly reflect the system's focus areas, enhancing clinical reliability.

[0032] In a preferred embodiment, based on instance attention weights and hemodynamic feature maps, morphological depth features and hemodynamic depth features are fused to obtain a fused vector. Infiltration propagation parameters are then obtained based on anatomical hierarchical propagation structure data and the fused vector, including: S501. Based on instance attention weights, morphological depth features and hemodynamic depth features are weighted and fused, and feature enhancement encoding is performed in combination with hemodynamic feature maps to obtain a fusion vector. S502. Obtain the anatomical hierarchical propagation structure data corresponding to the candidate instance, and perform mapping processing on the fusion vector to generate feature distribution results under different cervical anatomical levels. S503. Map the fused feature vectors to the corresponding cervical anatomical level regions to form a feature distribution matrix under each anatomical level. S504. Obtain the characteristic distribution differences between each anatomical level along the cervical canal, generate a propagation gradient sequence along the cervical canal, and obtain the cervical canal propagation intensity parameter based on the propagation gradient sequence along the cervical canal. S505. Obtain the characteristic distribution differences of adjacent areas along the direction of the tissue surrounding the cervix, generate a propagation gradient sequence along the direction of the tissue surrounding the cervix, and obtain the propagation intensity parameter of the tissue surrounding the cervix based on the propagation gradient sequence along the direction of the tissue surrounding the cervix. S506. Combine the cervical canal propagation intensity parameters and the propagation intensity parameters of the surrounding cervical tissues to form lesion infiltration and propagation parameters.

[0033] As described in steps S501 to S506 above, morphological depth features and hemodynamic depth features are mapped to a unified feature space. Scale normalization is performed on the two types of features to ensure consistency in dimension and numerical range between features from different sources. Instance attention weights for corresponding candidate instances are obtained, and the expression intensity of the instance during the fusion process is determined based on these attention weights. Weight modulation is applied to morphological depth features to highlight areas with obvious structural abnormalities, and weight modulation is applied to hemodynamic depth features to highlight areas with enhanced abnormalities. The two types of modulated features are combined to form initial fusion features. These initial fusion features are then mapped to their corresponding spatial locations in the hemodynamic feature atlas and introduced into that location. The enhanced behavioral distribution information is used to locally enhance the fusion features, making them simultaneously include tissue structure information, enhanced behavioral trends, and spatial blood supply distribution. The enhanced fusion features are then uniformly expressed to form fusion vectors that simultaneously reflect morphological and blood supply abnormalities. Anatomical hierarchical propagation structure data corresponding to candidate instances are obtained. Based on the instance's spatial location, the fusion vectors are mapped to the corresponding anatomical levels. The distribution of fusion vectors in different levels is recorded, forming a hierarchical feature distribution result, giving the fusion features anatomical hierarchical meaning. The cervical epithelium, cervical stroma, and pericervical tissue layers are divided into regions, and the fusion vectors are assigned to the corresponding anatomical level regions. The spatial distribution of fused features is generated within each hierarchical region. The distribution results are organized into a feature distribution matrix, which is used to express the distribution status of abnormal features in each level. The anatomical levels are arranged sequentially according to the cervical canal direction. The feature distribution of adjacent levels is continuously compared to obtain the feature change trend along the cervical canal direction, forming a propagation gradient sequence along the cervical canal direction. The propagation intensity along the cervical canal direction is expressed based on the gradient sequence, which is used to describe the trend of lesions developing into the deeper parts of the cervix. Adjacent anatomical regions are identified within the pericervical tissue region, and the feature distribution changes between adjacent regions are analyzed to form a propagation gradient sequence along the surrounding tissue direction. The propagation intensity along the surrounding tissue direction is expressed based on the gradient sequence. The intensity of lesion expansion is used to describe the outward spread behavior of the lesion. The intensity of propagation in the cervical canal direction and the intensity of propagation in the surrounding tissues are obtained. The two directional expressions are integrated to form the lesion infiltration and propagation parameters, which are used to describe the direction, trend and extent of lesion expansion. Through anatomical hierarchical mapping, the lesion propagation can be expressed along the real anatomical path, rather than a simple regional classification. The fusion vector unifies the expression of morphological and hemodynamic information, improves the accuracy of lesion expansion identification, and has a stronger ability to identify complex cases that develop towards both the cervical canal and spread to the surrounding tissues. The infiltration and propagation parameters can help assess the degree of local invasion and improve the structured level of imaging assessment.

[0034] In a preferred embodiment, signal difference parameters are obtained based on the degree of difference between different sequence image signals in the magnetic resonance imaging dataset. When the signal difference parameters meet preset abnormality conditions, segmentation processing is performed to obtain stable three-dimensional region data of cervical lesions, including: S601. Obtain a magnetic resonance imaging dataset and extract the signal intensity values ​​of corresponding pixels in a unified spatial coordinate system for different sequence image data. The different sequence image data include T2-weighted image data and dynamic contrast-enhanced image data. S602. Standardize the signal intensity values ​​of different sequence image data, obtain the signal difference measurement value of each pixel between different sequences, and construct a signal difference parameter map. S603. Obtain a preset anomaly judgment threshold, compare the signal difference parameter map with the preset anomaly judgment threshold, and obtain candidate anomaly regions that meet the anomaly conditions. S604. Based on the candidate abnormal regions, perform region segmentation processing in conjunction with the corresponding sequence image data to obtain candidate regions for cervical lesions. S605. Perform three-dimensional reconstruction on the candidate regions of cervical lesions, and screen the stability of the reconstruction results to obtain stable three-dimensional region data of cervical lesions.

[0035] As described in steps S601 to S605 above, different sequence image data, including T2-weighted images and dynamic contrast-enhanced images, are extracted from the magnetic resonance imaging dataset. Under a unified spatial coordinate system, the pixel positions corresponding to the different sequence images are aligned to ensure that pixels at the same location represent the same anatomical location. The extracted data includes the signal intensity value of each pixel under different sequences, providing a basis for differential analysis. The signal of each pixel across different sequences is standardized to ensure that the intensity values ​​of each sequence are within the same dimension. Signal difference measurement can be obtained by calculating the absolute difference, relative change, or normalized Euclidean distance between the signals of pixels in different sequences. The difference measurement value of each pixel is used to construct a signal difference parameter map (pixel difference heatmap) to visually display the possible location of the lesion. A predefined abnormality threshold is set (which can be determined through historical data statistics or expert experience). The pixel values ​​in the signal difference parameter map are compared with the threshold, and pixels exceeding the threshold are marked as abnormality candidates. The method can employ pixel-by-pixel thresholding or adaptive thresholding based on local neighborhood averaging to improve noise resistance. Image segmentation is performed on candidate abnormal regions to obtain continuous lesion regions. Threshold-based region growing methods or graph-based segmentation algorithms (such as conditional random fields, level set methods, etc.) can be used to ensure that the lesion boundaries are continuous and conform to the anatomical structure. During the segmentation process, multi-sequence image information is combined, and boundary recognition can be enhanced by fusing T2 and DCE signals. The output cervical lesion candidate regions are a preliminary spatial representation of the lesions in three dimensions. The candidate lesion regions on the two-dimensional slices are stacked and interpolated along the scanning direction to construct a three-dimensional model of the lesions. The stability of the three-dimensional lesions is screened to remove isolated small voxels, edge artifacts, or discontinuous regions to ensure that the reconstructed regions represent the real lesions. Three-dimensional reconstruction can use voxel connectivity analysis and volume continuity checks. Stability screening can be performed based on minimum volume threshold, spatial connectivity, and morphological regularity to obtain stable three-dimensional cervical lesion region data.

[0036] In a preferred embodiment, a comprehensive assessment feature set for lesions is constructed based on three-dimensional regional data of stable cervical lesions and combined with invasion and spread parameters. Cervical cancer identification results are obtained based on this comprehensive assessment feature set, including: S701. Obtain lesion parameters and three-dimensional lesion regions based on stable cervical lesion three-dimensional region data, wherein the lesion parameters include lesion volume parameters, maximum diameter parameters and morphological regularity parameters; S702. Obtain texture heterogeneity features based on the three-dimensional lesion area, and extract enhancement non-uniformity features based on dynamic contrast-enhanced image data. S703. Integrate lesion parameters, texture heterogeneity features, enhancement non-uniformity features, and infiltration and propagation parameters to construct a comprehensive lesion assessment feature set; S704. Obtain the preset identification strategy and combine it with the comprehensive lesion assessment feature set to obtain the cervical cancer identification results. The cervical cancer identification results include benign and malignant classification results, local invasion risk score, lesion invasion direction assessment information, and structured assessment data for clinical staging.

[0037] As described in steps S701 to S704 above, lesion parameters are extracted from the stable cervical lesion 3D region data, including lesion volume parameters (obtained by counting 3D voxels), maximum diameter parameters (measured by the longest connected distance in the 3D lesion region, reflecting the degree of lesion expansion), and morphological regularity parameters (calculated based on the deviation between the 3D lesion surface and the optimal fitted geometry (e.g., an ellipsoid)). A complete 3D lesion region is constructed. Based on the 3D lesion region, a multi-scale gray-level co-occurrence matrix (GLCM) and gray-level co-occurrence matrix are used. Texture features are extracted using gradient co-occurrence or wavelet decomposition methods to reflect the homogeneity or irregularity of the internal structure of lesions. Signal changes in the enhancement curve at different time phases are extracted from dynamic contrast-enhanced image data to quantify differences in blood perfusion within the lesion, reflecting the distribution and activity of microvessels. Lesion morphological parameters, texture heterogeneity features, enhancement heterogeneity features, and infiltration propagation parameters are uniformly fused. The fusion method can employ weighted vector combination, linear or nonlinear mapping after feature normalization, to obtain a comprehensive lesion assessment feature set. A unique comprehensive feature set can be generated for each lesion. The feature vector, combined with a preset identification strategy (including rule-based judgment, threshold setting, risk scoring formula or training-derived reference distribution), evaluates the comprehensive assessment features of lesions. The output identification results include: benign or malignant classification (judging whether the lesion is benign or malignant based on the comprehensive feature set), local invasion risk score (quantifying the risk of local spread by combining infiltration and propagation parameters and lesion morphology), lesion invasion direction assessment information (reflecting the potential expansion direction of the lesion along the cervical canal and surrounding tissues), and structured clinical staging data (converting the identification results into standardized staging indicators to assist clinical decision-making). It can achieve high-accuracy identification without relying on pixel-level annotation. It makes full use of three-dimensional structure and hemodynamic information, and integrates morphological, textural, enhancement, and invasive features to achieve a comprehensive assessment of lesions. Using the comprehensive lesion assessment features, it can accurately distinguish between benign and malignant lesions, improve the accuracy of early cervical cancer detection, and integrate infiltration and propagation parameters to quantify the spread trend of lesions along the cervical canal and surrounding tissues, providing a basis for clinical prediction and intervention. Through three-dimensional lesion reconstruction and feature fusion, it can reduce the risk of misjudgment caused by single-sequence or two-dimensional images and enhance diagnostic stability.

[0038] Please see the appendix Figure 2 As shown, the present invention also provides a cervical cancer identification system based on magnetic resonance image processing, used in the above-mentioned cervical cancer identification method based on magnetic resonance image processing, comprising: The image data module is used to acquire pelvic magnetic resonance imaging data of the object to be detected, and to perform spatial registration processing on the pelvic magnetic resonance imaging data to obtain a magnetic resonance imaging dataset. The propagation structure module divides the magnetic resonance imaging dataset into anatomical regions based on a pre-set pelvic anatomical structure model to obtain cervical region information, and constructs anatomical hierarchical propagation structure data based on the cervical region information. The feature map module constructs hemodynamic curves based on magnetic resonance imaging datasets and generates hemodynamic feature maps in a unified coordinate system; The deep feature module is used to construct a deep neural network. It obtains cervical tissue morphological depth features and hemodynamic depth features based on the magnetic resonance imaging dataset, obtains image diagnostic labels, performs weakly supervised learning training on the deep neural network, and obtains key cervical candidate instances and corresponding instance attention weights. The infiltration propagation module, based on instance attention weights and hemodynamic feature maps, fuses morphological depth features and hemodynamic depth features to obtain a fusion vector, and obtains infiltration propagation parameters based on anatomical hierarchical propagation structure data and the fusion vector; The three-dimensional region module obtains signal difference parameters based on the degree of difference between different sequence image signals in the magnetic resonance imaging dataset. When the signal difference parameters meet the preset abnormal conditions, segmentation processing is performed to obtain stable three-dimensional region data of cervical lesions. The lesion identification module is used to construct a comprehensive lesion assessment feature set based on the three-dimensional regional data of stable cervical lesions and the invasion and spread parameters, and to obtain cervical cancer identification results based on the comprehensive lesion assessment feature set.

[0039] The aforementioned imaging data module acquires pelvic magnetic resonance imaging (MRI) data of the subject, including T2-weighted images and dynamic contrast-enhanced images. It performs spatial registration of the image data to ensure that different sequences correspond to the same anatomical location in a unified coordinate system, thereby eliminating spatial deviations between sequences and constructing a unified MRI dataset. The propagation structure module, based on a pre-defined pelvic anatomical structure model, divides the MRI dataset into anatomical regions, acquiring cervical region information, including the cervical epithelium, stroma, and surrounding tissues. Using the boundary data and spatial location parameters of the cervical region, it constructs a regional connectivity topology table, generating anatomical hierarchical propagation structure data to characterize the spatial propagation paths and hierarchical relationships between different cervical anatomical levels. This data is used for continuous medical imaging. In imaging, a discretized layer definition is used. First, along the cervical canal axis and surrounding tissue direction, the image is divided into equidistant layers of fixed thickness or voxel count, each layer constituting a layer level. When layer boundaries are blurred, the lesion characteristics of each layer are statistically analyzed and weighted averaged with adjacent layers to extract the feature change sequence between adjacent layers as a propagation gradient. This ensures that even with unclear lesion boundaries, the infiltration trend of lesions along different anatomical directions can be stably and reproducibly quantified. The feature atlas module extracts time-series signals from dynamic contrast-enhanced images, constructs a hemodynamic curve for each pixel, and maps and fuses the hemodynamic curves with T2-weighted images in a unified coordinate system to generate a hemodynamic feature atlas. The depth feature module constructs a deep neural network, including cervical tissue morphology... The morphological and hemodynamic feature channels are used to segment the magnetic resonance imaging dataset, extracting morphological and hemodynamic depth features from candidate image blocks. Through weakly supervised learning, combined with image-level diagnostic labels, the network is trained to generate instance classification response values ​​and instance attention weights, thereby selecting key cervical candidate instances. The invasion and propagation module combines morphological and hemodynamic depth features with instance attention weights for weighted fusion, forming a fusion vector. This fusion vector is mapped to anatomical-level propagation structure data, generating feature distribution matrices at different cervical anatomical levels. The feature distribution gradient is analyzed along the cervical canal and surrounding tissues to generate invasion and propagation parameters, characterizing the potential spread trend and intensity of lesions in different directions. (Three-dimensional region...) The domain module analyzes the differences between image signals from different sequences, generates a signal difference parameter map, compares the signal difference parameters with preset abnormality conditions, filters abnormal areas, and performs 3D region segmentation and reconstruction to obtain stable 3D region data of cervical lesions. The lesion identification module integrates lesion volume, maximum diameter, morphological regularity, texture heterogeneity, enhancement inhomogeneity, and invasive propagation parameters to form a comprehensive lesion assessment feature set. Combined with preset identification strategies, it performs benign / malignant classification, local invasion risk scoring, invasion direction assessment, and structured clinical staging assessment of lesions. It combines anatomical structure, hemodynamics, morphology, and invasive features for joint analysis, improving identification accuracy. The depth feature module can automatically mark candidate instances with high contribution, reducing manual intervention.To improve diagnostic efficiency, the infiltration and spread module provides the diffusion trend and intensity of lesions along the cervical canal and surrounding tissues, offering a quantitative basis for predicting local invasion risk. The 3D region module enables spatial visualization and stability analysis of lesions, improving the reliability of lesion identification. The lesion identification module outputs structured and quantitative benign / malignant classification and staging information, assisting doctors in developing precise treatment plans.

[0040] And, a cervical cancer identification terminal based on magnetic resonance imaging processing, comprising: One or more processors; A storage device on which one or more programs are stored; When one or more programs are executed by one or more processors, the one or more processors implement a cervical cancer identification method based on magnetic resonance imaging.

[0041] The above description is merely a preferred embodiment of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention. Structures, devices, and operating methods not specifically described or explained in this invention are implemented according to conventional methods in the art unless otherwise specified or limited.

Claims

1. A method for cervical cancer identification based on magnetic resonance imaging processing, characterized in that, include: Acquire pelvic magnetic resonance imaging data of the object to be detected, perform spatial registration processing on the pelvic magnetic resonance imaging data, and obtain magnetic resonance imaging dataset. Based on a pre-defined pelvic anatomical structure model, the magnetic resonance imaging dataset is divided into anatomical regions to obtain cervical region information. Based on the cervical region information, anatomical hierarchical propagation structure data is constructed. Hemodynamic curves were constructed based on magnetic resonance imaging datasets, and hemodynamic feature maps were generated in a unified coordinate system. A deep neural network was constructed to obtain cervical tissue morphological depth features and hemodynamic depth features based on magnetic resonance imaging datasets, obtain image diagnostic labels, perform weakly supervised learning training on the deep neural network, and obtain key cervical candidate instances and corresponding instance attention weights. Based on instance-based attention weights and hemodynamic feature maps, morphological depth features and hemodynamic depth features are fused to obtain a fusion vector. Infiltration propagation parameters are obtained based on anatomical hierarchical propagation structure data and the fusion vector. Based on the degree of difference between different sequence image signals in the magnetic resonance imaging dataset, signal difference parameters are obtained. When the signal difference parameters meet the preset abnormal conditions, segmentation processing is performed to obtain stable three-dimensional region data of cervical lesions. Based on the three-dimensional regional data of stable cervical lesions and combined with infiltration and propagation parameters, a comprehensive assessment feature set of lesions is constructed, and cervical cancer identification results are obtained based on the comprehensive assessment feature set of lesions.

2. The cervical cancer identification method based on magnetic resonance image processing according to claim 1, characterized in that, Acquire pelvic magnetic resonance imaging (MRI) data of the subject to be examined, perform spatial registration processing on the pelvic MRI data to obtain an MRI dataset, including: Acquire pelvic magnetic resonance imaging data of the subject to be examined, including T2-weighted image data and dynamic contrast-enhanced image data; Image quality verification and abnormal image removal were performed on T2-weighted image data and dynamic contrast-enhanced image data to obtain effective magnetic resonance image data. Spatial alignment processing is performed on the effective magnetic resonance image data based on a preset reference sequence to obtain spatially aligned magnetic resonance image data. Spatial registration is performed on spatially aligned magnetic resonance images to obtain spatially registered magnetic resonance image data. The T2-weighted image data and dynamic contrast-enhanced image data from the spatially registered magnetic resonance imaging data are integrated to construct a magnetic resonance imaging dataset.

3. The cervical cancer identification method based on magnetic resonance image processing according to claim 1, characterized in that, Based on a pre-defined pelvic anatomical structure model, the magnetic resonance imaging dataset is divided into anatomical regions to obtain cervical region information. Based on this cervical region information, anatomical hierarchical propagation structure data is constructed, including: Obtain a preset pelvic anatomical structure model, which includes cervical epithelial layer structure information, cervical stroma layer structure information, and cervical surrounding tissue structure information. Based on the magnetic resonance imaging dataset and combined with the pelvic anatomical structure model, the magnetic resonance imaging dataset is processed for anatomical structure recognition to obtain the initial anatomical region segmentation results. Cervical region information is extracted based on the initial anatomical region segmentation results. The cervical region information includes the cervical epithelial region, the cervical stroma region, and the pericervical tissue region. Obtain the region boundary data and region location parameters based on cervical region information; Based on the regional boundary data and regional location parameters, a regional connectivity topology table is constructed, and anatomical hierarchical propagation structure data is generated based on the regional connectivity topology table.

4. The cervical cancer identification method based on magnetic resonance image processing according to claim 1, characterized in that, Hemodynamic curves were constructed based on magnetic resonance imaging datasets, and hemodynamic feature maps were generated in a unified coordinate system, including: Extracting dynamic contrast-enhanced image data from magnetic resonance imaging datasets; The dynamic contrast-enhanced image data is decomposed over time to extract the enhanced signal intensity data of multiple time phases; Based on the enhanced signal intensity data, construct a set of hemodynamic curve parameters corresponding to each pixel; Under a unified spatial coordinate system, the set of hemodynamic curve parameters and T2-weighted image data are spatially mapped and fused. Based on the fusion processing results, a hemodynamic feature map is generated.

5. The cervical cancer identification method based on magnetic resonance image processing according to claim 1, characterized in that, A deep neural network was constructed to obtain cervical tissue morphological depth features and hemodynamic depth features from a magnetic resonance imaging dataset. Image diagnostic labels were obtained, and the deep neural network was trained using weakly supervised learning to obtain key cervical candidate instances and their corresponding instance attention weights, including: Construct a deep neural network, wherein the deep neural network includes a first feature extraction channel and a second feature extraction channel; Based on cervical region information, the cervical region in the magnetic resonance imaging dataset is divided into blocks under a unified spatial coordinate system to generate multiple cervical candidate image blocks, and candidate instances are labeled. Based on the candidate instances, the corresponding T2-weighted image data is extracted, and combined with the first feature extraction channel, cervical tissue morphological depth features are generated. Based on candidate instances, extract corresponding dynamic contrast-enhanced image data and combine it with the second feature extraction channel to generate hemodynamic depth features; The morphological depth features and hemodynamic depth features of cervical tissue are fused to generate instance fusion features; Obtain the image diagnostic labels corresponding to the magnetic resonance imaging dataset, and combine them with instance fusion features to perform weakly supervised learning training on the deep neural network to obtain instance classification response values ​​and instance attention weights. Based on the instance classification response value and instance attention weight, the contribution index of the candidate instance is obtained. Candidate instances with contribution index higher than the preset threshold are selected, marked as key cervical candidate instances, and the corresponding instance attention weight is obtained.

6. The cervical cancer identification method based on magnetic resonance image processing according to claim 1, characterized in that, Based on instance-based attention weights and hemodynamic feature maps, morphological depth features and hemodynamic depth features are fused to obtain a fused vector. Infiltration propagation parameters are then obtained from anatomical hierarchical propagation structure data and the fused vector, including: Based on instance attention weights, morphological depth features and hemodynamic depth features are weighted and fused, and feature enhancement encoding is performed by combining hemodynamic feature maps to obtain a fusion vector; Obtain the anatomical hierarchical propagation structure data corresponding to the candidate instances, and perform mapping processing on the fusion vector to generate feature distribution results under different cervical anatomical levels; The fused feature vectors are mapped to the corresponding cervical anatomical level regions to form a feature distribution matrix for each anatomical level. The characteristic distribution differences between each anatomical level are obtained along the cervical canal, a propagation gradient sequence along the cervical canal is generated, and the cervical canal propagation intensity parameter is obtained based on the propagation gradient sequence along the cervical canal. The characteristic distribution differences of adjacent areas are obtained along the direction of the tissue surrounding the cervix, a propagation gradient sequence along the direction of the tissue surrounding the cervix is ​​generated, and the propagation intensity parameter of the tissue surrounding the cervix is ​​obtained based on the propagation gradient sequence along the direction of the tissue surrounding the cervix. By combining the cervical canal transmission intensity parameters and the transmission intensity parameters of the surrounding cervical tissues, lesion infiltration and transmission parameters are formed.

7. The cervical cancer identification method based on magnetic resonance image processing according to claim 1, characterized in that, Signal difference parameters are obtained based on the degree of difference between different sequence image signals in the magnetic resonance imaging dataset. When the signal difference parameters meet preset abnormality conditions, segmentation processing is performed to obtain stable three-dimensional region data of cervical lesions, including: Acquire magnetic resonance imaging datasets and extract the signal intensity values ​​of corresponding pixels in a unified spatial coordinate system from different image sequences. The different image sequences include T2-weighted image data and dynamic contrast-enhanced image data. The signal intensity values ​​of different sequence image data are standardized, and the signal difference metric value of each pixel between different sequences is obtained to construct a signal difference parameter map; Obtain a preset anomaly judgment threshold, compare the signal difference parameter map with the preset anomaly judgment threshold, and obtain candidate anomaly regions that meet the anomaly conditions; Based on the candidate abnormal regions, region segmentation processing is performed in conjunction with the corresponding sequence image data to obtain candidate regions for cervical lesions; The candidate regions of cervical lesions were reconstructed in three dimensions, and the stability of the reconstruction results was screened to obtain stable three-dimensional region data of cervical lesions.

8. The cervical cancer identification method based on magnetic resonance image processing according to claim 1, characterized in that, Based on three-dimensional regional data of stable cervical lesions and combined with invasive spread parameters, a comprehensive lesion assessment feature set is constructed. Cervical cancer identification results are obtained based on this comprehensive lesion assessment feature set, including: Based on stable cervical lesion three-dimensional region data, lesion parameters and three-dimensional lesion regions are obtained. Among them, lesion parameters include lesion volume parameters, maximum diameter parameters and morphological regularity parameters. Texture heterogeneity features are obtained from the three-dimensional lesion area, and enhancement non-uniformity features are extracted from the dynamic contrast-enhanced image data. By integrating lesion parameters, texture heterogeneity features, enhancement non-uniformity features, and infiltration and propagation parameters, a comprehensive lesion assessment feature set is constructed. The system acquires a preset identification strategy and combines it with a comprehensive lesion assessment feature set to obtain cervical cancer identification results. These results include benign / malignant classification results, local invasion risk scores, lesion invasion direction assessment information, and structured assessment data for clinical staging.

9. A cervical cancer identification system based on magnetic resonance image processing, applied to the cervical cancer identification method based on magnetic resonance image processing according to any one of claims 1 to 8, characterized in that, include: The image data module is used to acquire pelvic magnetic resonance imaging data of the object to be detected, and to perform spatial registration processing on the pelvic magnetic resonance imaging data to obtain a magnetic resonance imaging dataset. The propagation structure module divides the magnetic resonance imaging dataset into anatomical regions based on a pre-set pelvic anatomical structure model to obtain cervical region information, and constructs anatomical hierarchical propagation structure data based on the cervical region information. The feature map module constructs hemodynamic curves based on magnetic resonance imaging datasets and generates hemodynamic feature maps in a unified coordinate system; The deep feature module is used to construct a deep neural network. It obtains cervical tissue morphological depth features and hemodynamic depth features based on the magnetic resonance imaging dataset, obtains image diagnostic labels, performs weakly supervised learning training on the deep neural network, and obtains key cervical candidate instances and corresponding instance attention weights. The infiltration propagation module, based on instance attention weights and hemodynamic feature maps, fuses morphological depth features and hemodynamic depth features to obtain a fusion vector, and obtains infiltration propagation parameters based on anatomical hierarchical propagation structure data and the fusion vector; The three-dimensional region module obtains signal difference parameters based on the degree of difference between different sequence image signals in the magnetic resonance imaging dataset. When the signal difference parameters meet the preset abnormal conditions, segmentation processing is performed to obtain stable three-dimensional region data of cervical lesions. The lesion identification module is used to construct a comprehensive lesion assessment feature set based on the three-dimensional regional data of stable cervical lesions and the invasion and spread parameters, and to obtain cervical cancer identification results based on the comprehensive lesion assessment feature set.

10. A cervical cancer identification terminal based on magnetic resonance imaging processing, characterized in that, include: One or more processors; A storage device on which one or more programs are stored; When one or more programs are executed by one or more processors, the one or more processors implement the cervical cancer identification method based on magnetic resonance image processing as described in any one of claims 1 to 8.

Images