A method and system for analyzing pituitary adenoma invasion images

Abstract

The present invention discloses a method and system for analyzing pituitary adenoma invasion imaging data, which relates to the technical field of medical prediction. A pituitary adenoma invasion imaging data analysis system includes: an imaging feature extraction module, an auxiliary factor analysis module, and an invasion prediction and judgment module. By combining the imaging data and auxiliary biomarker data of patients and using a cross-attention mechanism and a feature fusion strategy for analysis, the present invention can make full use of information from different data sources to improve the comprehensiveness and accuracy of pituitary adenoma invasion prediction; using a bidirectional propagation network to process pituitary adenoma images can extract more detailed and accurate imaging features; through the pituitary adenoma invasion prediction model, combining invasion data auxiliary prediction factors and patient imaging feature images can perform high-precision invasion prediction, and fusing key features of different modalities helps to improve the accuracy and reliability of invasion prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of medical prediction, and particularly to a method and system for analyzing pituitary adenoma invasion image data. Background Art

[0002] The invasiveness of pituitary adenomas has a great impact on the health of patients. Especially when the tumor invades adjacent structures such as the cavernous sinus, optic chiasm, and brainstem, it may cause severe neurological dysfunctions, vision loss, headache, and hormone secretion disorders, etc. Therefore, early and accurate prediction of the invasiveness of pituitary adenomas is crucial for optimizing treatment plans, reducing complications, and improving the quality of life of patients. Currently, the conventional methods for pituitary adenoma invasion image analysis mainly rely on MRI and CT images, and evaluate the invasiveness by observing the morphology, size, boundary of the tumor, and its relationship with surrounding structures. However, due to the resolution limitation of traditional imaging techniques, the subtle invasion between the tumor and surrounding tissues cannot be clearly shown, thus affecting the accuracy of prediction.

[0003] Therefore, it is necessary to improve the resolution of images and perform high-dimensional feature processing. At the same time, fusing imaging and biomarker data can provide more levels of information, further improving the accuracy and comprehensiveness of predicting the invasion of pituitary adenomas into the cavernous sinus, suprasellar region, or other parts. Summary of the Invention

[0004] The present invention aims to provide a method and system for analyzing pituitary adenoma invasion image data, and uses high-dimensional feature pituitary adenoma image data for invasion prediction.

[0005] A method for analyzing pituitary adenoma invasion image data includes the following steps:

[0006] Obtain a set of pituitary adenoma images of a patient; the set of pituitary adenoma images of the patient contains consecutive N pituitary adenoma images P of the patient with an interval time window, where n = 1, 2,..., N, and the length of the time window is L; based on the set of pituitary adenoma images of the patient and a pituitary adenoma image feature extraction model for processing, obtain a pituitary adenoma invasion feature image; the pituitary adenoma image feature extraction model uses a bidirectional propagation network to perform image processing on the set of pituitary adenoma images of the patient, and is used to obtain a high-quality pituitary adenoma invasion feature image for subsequent analysis; n Obtain the patient's auxiliary biomarker data; based on the patient's auxiliary biomarker data and a pituitary adenoma biomarker prediction and analysis model for analysis, obtain an invasion data auxiliary prediction factor;

[0007]

[0008] ​Predictive analysis is performed based on invasion data-assisted predictors, patient pituitary adenoma invasion feature images, and a pituitary adenoma invasion prediction model to obtain the predictive analysis results of the patient's pituitary adenoma; the pituitary adenoma invasion prediction model uses a cross-attention mechanism and a feature fusion strategy for predictive analysis, and is used for high-precision invasion prediction based on invasion data-assisted predictors and patient pituitary adenoma invasion feature images; subsequent operations are performed on the patient according to the predictive analysis results of the patient's pituitary adenoma.

[0009] As a preferred technical solution of the present invention, the pituitary adenoma image feature extraction model includes an image preprocessing layer, an image feature extraction layer, and an image output layer;

[0010] The image preprocessing layer is used to perform image super-resolution processing on the patient pituitary adenoma image set to obtain a preprocessed patient pituitary adenoma image set; among them, the preprocessed patient pituitary adenoma image set contains N preprocessed patient pituitary adenoma images C n ;

[0011] The image feature extraction layer is used to extract features from the preprocessed patient pituitary adenoma image set to obtain patient pituitary adenoma invasion feature images;

[0012] The image output layer is used to output patient pituitary adenoma invasion feature images.

[0013] As a preferred technical solution of the present invention, the specific steps of performing image super-resolution processing in the image preprocessing layer include:

[0014] The image preprocessing layer contains an image alignment layer and an image reconstruction layer; in the image alignment layer, there are a forward propagation network and a backward propagation network; in the image reconstruction layer, there are a feature extraction network, a residual refinement network, and an image reconstruction network;

[0015] In the backward propagation network, the patient pituitary adenoma images P n+1 、patient pituitary adenoma images P n+2 and patient pituitary adenoma images P n are aligned to obtain a backward patient pituitary adenoma image set F n ; the backward patient pituitary adenoma image set F n is sent into a deformable convolution block for further feature alignment to obtain a backward patient pituitary adenoma image F n ';

[0016] In the forward propagation network, the patient pituitary adenoma images P n-1 、patient pituitary adenoma images P n-2 and patient pituitary adenoma images P n are aligned to obtain a forward patient pituitary adenoma image set Z n; The forward patient pituitary adenoma image set Z n is sent into the deformable convolutional block for further feature alignment to obtain the forward patient pituitary adenoma image Z n ’;

[0017] In the image reconstruction layer, the feature extraction network is used to extract the reverse patient pituitary adenoma image F n ’, the forward patient pituitary adenoma image Z n ’ and the patient pituitary adenoma image P n for feature extraction and splicing to obtain the patient pituitary adenoma feature image T n ; The residual refinement network is used to refine the features of the patient pituitary adenoma feature image T n through multiple residual blocks to obtain the refined patient pituitary adenoma feature image T n ’; In the image reconstruction network, the refined patient pituitary adenoma feature image T n ’ is enlarged four times using the pixel rearrangement method to obtain the high-resolution patient pituitary adenoma feature image G n ; The patient pituitary adenoma image P n is enlarged four times using bilinear interpolation to obtain the low-resolution patient pituitary adenoma feature image D n ; The high-resolution patient pituitary adenoma feature image G n and the low-resolution patient pituitary adenoma feature image D n are added together to obtain the preprocessed patient pituitary adenoma image C n .

[0018] As a preferred technical solution of the present invention, the specific steps for feature extraction in the image feature extraction layer include:

[0019] The image feature extraction layer includes a global feature extraction layer, a local feature division layer, and a feature scoring layer;

[0020] In the global feature extraction layer, the preprocessed patient pituitary adenoma image C n is subjected to global feature extraction to obtain the patient pituitary adenoma feature global feature vector Q n ;

[0021] In the local feature division layer, the preprocessed patient pituitary adenoma image C n is divided into feature image blocks using a sliding window to obtain the patient pituitary adenoma feature image block set J n ; The patient pituitary adenoma feature image block set J n contains I patient pituitary adenoma feature image blocks; Based on the patient pituitary adenoma feature image block set J n feature extraction is performed to obtain the patient pituitary adenoma feature local feature vector J ni , i = 1, 2,..., I;

[0022] Specific steps for feature image block extraction:

[0023] Use the formula I = (X + Y - X') / Y, where I is the preprocessed pituitary adenoma image C of the patient n The total number of feature image blocks to be extracted, X is the height of the preprocessed pituitary adenoma image C of the patient n of the patient, X' is the horizontal height of the pituitary adenoma feature image block after division; Y is the step size of the sliding window;

[0024] In the feature scoring layer, for the global feature vector Q of the pituitary adenoma features of the patient n and the local feature vector J of the pituitary adenoma features of the patient ni perform feature scoring to obtain the pituitary adenoma feature vector scoring set U n ; Based on all pituitary adenoma feature vector scoring sets U n perform screening, and fuse the feature vectors corresponding to the screening conditions to obtain the pituitary adenoma invasion feature image;

[0025] The feature scoring layer is trained based on the attention mechanism.

[0026] As a preferred technical solution of the present invention, the pituitary adenoma invasion prediction model includes an invasion prediction layer and a result output layer;

[0027] The invasion prediction layer is used to perform prediction analysis based on the invasion data auxiliary prediction factor and the pituitary adenoma invasion feature image of the patient to obtain the pituitary adenoma prediction analysis result;

[0028] The invasion prediction layer is constructed based on the CNN model and the cross-attention mechanism;

[0029] The result output layer is used to output the pituitary adenoma prediction analysis result.

[0030] As a preferred technical solution of the present invention, the specific steps for performing prediction analysis in the invasion prediction layer include:

[0031] The invasion prediction layer includes M convolutional kernels with gradually increasing sizes, a parallel cross-attention fusion layer, and a fully connected prediction layer;

[0032] Perform high-dimensional feature extraction on the pituitary adenoma invasion feature image in the M convolutional kernels, and at the same time use the parallel cross-attention fusion layer to fuse the invasion data auxiliary prediction factor for feature fusion to obtain the pituitary adenoma invasion feature prediction vector;

[0033] Perform invasion prediction analysis based on the pituitary adenoma invasion feature prediction vector in the fully connected prediction layer to obtain the pituitary adenoma prediction analysis result.

[0034] As a preferred technical solution of the present invention, the pituitary adenoma biomarker prediction and analysis model includes a data preprocessing layer, a feature judgment layer, and a feature output layer;

[0035] The data preprocessing layer is used to preprocess the patient's auxiliary biomarker data to obtain preprocessed patient's auxiliary biomarker data;

[0036] The feature judgment layer is used to judge the features of the preprocessed patient's auxiliary biomarker data to obtain an invasion data auxiliary prediction factor;

[0037] Specific steps for training the feature judgment layer:

[0038] Collect several groups of biomarker level training samples; each group of biomarker level training samples contains verified feature factors and corresponding biomarker data; combine several groups of biomarker level training samples to obtain a biomarker level training set;

[0039] Input the biomarker level training set into a BP neural network for model training to obtain an initial feature judgment layer; evaluate the initial feature judgment layer. If the initial feature judgment layer passes the model evaluation, use the initial feature judgment layer as the feature judgment layer in the pituitary adenoma biomarker prediction and analysis model; otherwise, continue model training using the biomarker level training set;

[0040] The feature output layer is used to output the invasion data auxiliary prediction factor.

[0041] A pituitary adenoma invasion image data analysis system includes:

[0042] An image feature extraction module, including an image acquisition unit and a feature extraction unit; the image acquisition unit is used to acquire a patient pituitary adenoma image set; the patient pituitary adenoma image set contains N consecutive patient pituitary adenoma images P with an interval time window n , n = 1, 2,..., N, and the length of the time window is L; the feature extraction unit is used to process based on the patient pituitary adenoma image set and the pituitary adenoma image feature extraction model to obtain a patient pituitary adenoma invasion feature image; the pituitary adenoma image feature extraction model uses a bidirectional propagation network to perform image processing on the patient pituitary adenoma image set to obtain a high-quality patient pituitary adenoma invasion feature image for subsequent analysis;

[0043] An auxiliary factor analysis module, including a data acquisition unit and an auxiliary analysis unit; the data acquisition unit is used to acquire patient auxiliary biomarker data; the auxiliary analysis unit is used to analyze based on the patient auxiliary biomarker data and the pituitary adenoma biomarker prediction and analysis model to obtain an invasion data auxiliary prediction factor;

[0044] The invasion prediction and judgment module includes a prediction and judgment unit and a result output unit; the prediction and judgment unit is used to perform prediction analysis based on invasion data auxiliary prediction factors, the patient's pituitary adenoma invasion characteristic image, and the pituitary adenoma invasion prediction model to obtain the patient's pituitary adenoma prediction analysis result; the pituitary adenoma invasion prediction model uses a cross-attention mechanism and a feature fusion strategy for prediction analysis, and is used to perform high-precision invasion prediction based on invasion data auxiliary prediction factors and the patient's pituitary adenoma invasion characteristic image; the result output unit is used to perform subsequent operations on the patient according to the patient's pituitary adenoma prediction analysis result.

[0045] The present invention has the following advantages:

[0046] 1. By combining the patient's imaging data and auxiliary biomarker data, and using a cross-attention mechanism and a feature fusion strategy for analysis, the present invention can make full use of information from different data sources, improve the comprehensiveness and accuracy of pituitary adenoma invasion prediction; using a bidirectional propagation network to process pituitary adenoma images can extract more detailed and accurate imaging features; through the pituitary adenoma invasion prediction model, combining invasion data auxiliary prediction factors and the patient's imaging feature image, high-precision invasion prediction can be performed, and fusing key features of different modalities helps to improve the accuracy and reliability of invasion prediction.

[0047] 2. The present invention performs image registration through forward and backward propagation networks to ensure the consistency of the patient's pituitary adenoma images in the time series, aligns images at different time points, and eliminates the influence caused by shooting angle, position difference, or equipment change; ensures the spatial alignment of time series images, makes subsequent feature extraction and comparison more accurate, and improves the accuracy of model analysis; by aligning multi-period images, the changes and progress of pituitary adenomas can be better captured, helping to analyze the growth pattern and invasion characteristics of tumors. Description of the Drawings

[0048] Figure 1 It is a schematic structural diagram of a pituitary adenoma invasion imaging data analysis system adopted in an embodiment of the present invention. Detailed Embodiments

[0049] In order to enable those skilled in the art to better understand the technical solutions in the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention.

[0050] Embodiment 1, a method for analyzing pituitary adenoma invasion imaging data, includes the following steps:

[0051] Obtain a set of patient pituitary adenoma images; the set of patient pituitary adenoma images contains N consecutive patient pituitary adenoma images P with a time window interval n, where \(n = 1, 2, \ldots, N\), the length of the time window is \(L\); the length of the time window is specifically set by professionals according to the actual situation. \(L\) can be set to several weeks, months, or a year, and the image acquisition interval within each time window is selected; this interval can be fixed, such as once a month, or flexible, determined according to the patient's treatment plan and the availability of medical images; when obtaining the patient's pituitary adenoma image set, ensure the patient's informed consent and allow their image data to be used for research, and obtain it using medical imaging equipment. For each patient, collect multiple images according to the selected time window \(L\) and image interval, and ensure that the imaging method is the same for each image.

[0052] Based on the patient's pituitary adenoma image set and the pituitary adenoma image feature extraction model, processing is carried out to obtain the patient's pituitary adenoma invasion feature image; the pituitary adenoma image feature extraction model uses a bidirectional propagation network to perform image processing on the patient's pituitary adenoma image set to obtain high-quality patient pituitary adenoma invasion feature images for subsequent analysis.

[0053] The pituitary adenoma image feature extraction model includes an image preprocessing layer, an image feature extraction layer, and an image output layer.

[0054] The image preprocessing layer is used to perform image super-resolution processing on the patient's pituitary adenoma image set to obtain a preprocessed patient pituitary adenoma image set; among them, the preprocessed patient pituitary adenoma image set contains \(N\) preprocessed patient pituitary adenoma images \(C\). n ; The image preprocessing layer uses super-resolution processing to enhance the patient's pituitary adenoma images, which helps to improve the spatial resolution of the images, making the lesion area clearer, especially in small-sized, blurred, or low-resolution images; through super-resolution processing, the details of the images are enhanced, and the edges of the lesions are more clearly defined, which helps to more accurately identify the tumor and its invasion area, especially when the tumor invades the surrounding brain tissue.

[0055] The image feature extraction layer is used to extract features from the preprocessed patient pituitary adenoma image set to obtain the patient's pituitary adenoma invasion feature image; the image feature extraction layer can automatically extract biologically significant features from the preprocessed pituitary adenoma images, such as the morphology, texture, edges, density, etc. of the tumor. These features not only include the size and shape of the tumor but also reveal the invasion features of the tumor to adjacent tissues, such as the expansiveness of the tumor, the fuzziness of the boundary, invasive features, etc.

[0056] The image output layer is used to output the patient's pituitary adenoma invasion feature image; through image processing and feature extraction, the generated patient pituitary adenoma invasion feature image can provide doctors with more detailed image information about the tumor invasiveness.

[0057] The specific steps for performing image super-resolution processing in the image preprocessing layer include:

[0058] The image preprocessing layer includes an image alignment layer and an image reconstruction layer; in the image alignment layer, there are a forward propagation network and a backward propagation network; in the image reconstruction layer, there are a feature extraction network, a residual refinement network, and an image reconstruction network;

[0059] In the backward propagation network, the patient pituitary adenoma images P n+1 、the patient pituitary adenoma images P n+2 and the patient pituitary adenoma images P n are aligned to obtain the backward patient pituitary adenoma image set F n ; the backward patient pituitary adenoma image set F n is sent into a deformable convolution block for further feature alignment to obtain the backward patient pituitary adenoma image F n ';

[0060] In the forward propagation network, the patient pituitary adenoma images P n-1 、the patient pituitary adenoma images P n-2 and the patient pituitary adenoma images P n are aligned to obtain the forward patient pituitary adenoma image set Z n ; the forward patient pituitary adenoma image set Z n is sent into a deformable convolution block for further feature alignment to obtain the forward patient pituitary adenoma image Z n ';

[0061] In the image reconstruction layer, the backward patient pituitary adenoma image F n ', the forward patient pituitary adenoma image Z n ', and the patient pituitary adenoma image P n are used for feature extraction and stitching to obtain the patient pituitary adenoma feature image T n ; the residual refinement network is used to refine the features of the patient pituitary adenoma feature image T n through multiple residual blocks to obtain the refined patient pituitary adenoma feature image T n '; in the image reconstruction network, the refined patient pituitary adenoma feature image T n ' is enlarged four times using the pixel rearrangement method to obtain the high-resolution patient pituitary adenoma feature image G n ; the patient pituitary adenoma image P n is enlarged four times using bilinear interpolation to obtain the low-resolution patient pituitary adenoma feature image D n ; the high-resolution patient pituitary adenoma feature image G n and the low-resolution patient pituitary adenoma feature image D n are added together to obtain the preprocessed patient pituitary adenoma image Cn ;

[0062] Image registration is performed through forward and backward propagation networks to ensure the consistency of the patient's pituitary adenoma images in the time series, align the images at different time points, and eliminate the impacts caused by shooting angles, position differences, or equipment changes; ensure the spatial alignment of the time series images, making subsequent feature extraction and comparison more accurate and improving the accuracy of model analysis; by aligning multi-period images, the changes and progression of pituitary adenomas can be better captured, helping to analyze the growth pattern and invasion characteristics of tumors;

[0063] In the backward and forward propagation networks, after image registration, the images will enter the deformable convolution block for further feature alignment. Deformable convolution can perform flexible adjustments in space, handle deformations or transformations in the images, and is especially suitable for cases where the tumor boundary is blurred or the morphology is complex; the deformable convolution block can dynamically adjust according to the content of the images, and is not limited to rigid transformations when aligning images, and can better adapt to the diversity of tumor morphologies;

[0064] In the image reconstruction layer, the feature extraction network extracts key features and stitches these features together to generate the patient's pituitary adenoma feature image T n ; This multi-view and multi-time point feature fusion helps to capture all-round information of the tumor. By fusing the image features of multiple time points, the dynamic changes of the tumor can be comprehensively captured, helping to improve the understanding of complex features such as tumor morphology and invasiveness; the stitched feature image contains information from multiple time points, providing more context and improving the model's ability to recognize subtle changes; through pixel rearrangement and bilinear interpolation, the resolution of the image is significantly improved, making the details of the tumor and its invasion area clearer, facilitating subsequent analysis and diagnosis; by combining the advantages of high-resolution and low-resolution images, the final preprocessed image can provide clear and detailed image data for subsequent feature extraction and analysis;

[0065] The specific steps for feature extraction in the image feature extraction layer include:

[0066] The image feature extraction layer includes a global feature extraction layer, a local feature division layer, and a feature scoring layer;

[0067] In the global feature extraction layer, global features of the preprocessed patient's pituitary adenoma image C n are extracted to obtain the global feature vector Q of the patient's pituitary adenoma features n ;

[0068] In the local feature division layer, the preprocessed patient's pituitary adenoma image C is divided into feature image patches using a sliding window n to obtain the set J of patient's pituitary adenoma feature image patchesn ; the set J of characteristic image blocks of the patient's pituitary adenoma n contains I characteristic image blocks of the patient's pituitary adenoma; based on the set J of characteristic image blocks of the patient's pituitary adenoma n feature extraction is performed to obtain the local characteristic vector J of the patient's pituitary adenoma ni , i = 1, 2,..., I;

[0069] The specific steps for extracting characteristic image blocks:

[0070] Use the formula I = (X + Y - X') / Y, where I is the total number of characteristic image blocks extracted from the preprocessed pituitary adenoma image C of the patient n of the height, X is the height of the preprocessed pituitary adenoma image C of the patient n of, X' is the horizontal height of the characteristic image blocks of the pituitary adenoma of the patient after division; Y is the step size of the sliding window; the specific values of I and Y are set by professional technicians according to the actual situation;

[0071] In the feature scoring layer, for the global characteristic vector Q of the patient's pituitary adenoma n and the local characteristic vector J of the patient's pituitary adenoma ni feature scoring is performed to obtain the set U of feature vector scores of the patient's pituitary adenoma n ; based on the set U of feature vector scores of all patients' pituitary adenomas n screening is performed, and the feature vectors corresponding to the screening conditions are fused to obtain the invasive characteristic image of the patient's pituitary adenoma;

[0072] The feature scoring layer is trained based on the attention mechanism;

[0073] By performing global feature extraction on the preprocessed pituitary adenoma image C of the patient n a global feature vector is obtained; global features usually include information such as the overall shape, structure, and size of the image, and can describe the macroscopic features of the image; using a sliding window to locally divide the preprocessed pituitary adenoma image of the patient, multiple small image blocks are obtained; after feature extraction for each image block, a local feature vector is generated; local features are usually used to describe the detailed parts of the image, such as the local area of the tumor, shape irregularity, etc.; through local feature extraction, local changes and details of the tumor can be captured, such as the tumor edge, shape changes, invasion of blood vessels or surrounding tissues, etc.; local features can reveal the microscopic changes of the tumor and are of great significance for analyzing the invasiveness, expansibility, and boundary irregularity of the tumor; the sliding window method ensures that every part of the image can be fully analyzed, avoiding the situation of ignoring local features of the image;

[0074] Through the scoring mechanism, the most helpful features for diagnosis can be screened out, avoiding the interference of redundant or irrelevant features, thus improving the efficiency and accuracy of analysis; the feature scoring layer uses the attention mechanism to weight-train each feature, and the attention mechanism can adaptively adjust its weight according to the importance of the feature, so that important features receive more attention during the training process;

[0075] Obtain the patient's auxiliary biomarker data; based on the patient's auxiliary biomarker data and the pituitary adenoma biomarker prediction analysis model, perform analysis to obtain an invasion data auxiliary prediction factor;

[0076] The pituitary adenoma biomarker prediction analysis model includes a data preprocessing layer, a feature judgment layer, and a feature output layer;

[0077] The data preprocessing layer is used to preprocess the patient's auxiliary biomarker data to obtain preprocessed patient auxiliary biomarker data;

[0078] The feature judgment layer is used to judge the features of the preprocessed patient auxiliary biomarker data to obtain an invasion data auxiliary prediction factor;

[0079] The specific steps for training the feature judgment layer:

[0080] Collect several groups of biomarker level training samples; each group of biomarker level training samples contains verified feature factors and corresponding biomarker data; combine several groups of biomarker level training samples to obtain a biomarker level training set;

[0081] Input the biomarker level training set into the BP neural network for model training to obtain an initial feature judgment layer; perform model evaluation on the initial feature judgment layer. If the initial feature judgment layer passes the model evaluation, use the initial feature judgment layer as the feature judgment layer in the pituitary adenoma biomarker prediction analysis model; otherwise, continue to perform model training using the biomarker level training set;

[0082] The feature output layer is used to output the invasion data auxiliary prediction factor;

[0083] The feature judgment layer is mainly used to judge which features in the preprocessed biomarker data are related to the invasiveness of the patient, and extract effective predictors. This process analyzes each feature in the biomarker data with the verified invasion factors to determine which biomarkers can best reflect the invasiveness of the tumor. Through the feature judgment layer, the biomarkers most relevant to the invasiveness of pituitary adenoma can be screened out, providing a more accurate input for subsequent analysis and improving the accuracy of prediction. Through the feature judgment layer, the biomarkers most relevant to invasiveness are accurately screened out, providing efficient input features. Through the training and optimization of the BP neural network, the accuracy of the prediction model is continuously improved, providing accurate predictors for clinical diagnosis.

[0084] Based on the invasion data-assisted predictors, the patient's pituitary adenoma invasion feature image, and the pituitary adenoma invasion prediction model, prediction analysis is carried out to obtain the prediction analysis result of the patient's pituitary adenoma. The pituitary adenoma invasion prediction model uses the cross-attention mechanism and the feature fusion strategy for prediction analysis, and is used to perform high-precision invasion prediction based on the invasion data-assisted predictors and the patient's pituitary adenoma invasion feature image. Subsequent operations are performed on the patient according to the prediction analysis result of the patient's pituitary adenoma.

[0085] The pituitary adenoma invasion prediction model includes an invasion prediction layer and a result output layer.

[0086] The invasion prediction layer is used to perform prediction analysis based on the invasion data-assisted predictors and the patient's pituitary adenoma invasion feature image, and obtain the prediction analysis result of the patient's pituitary adenoma.

[0087] The invasion prediction layer is constructed based on the CNN model and the cross-attention mechanism.

[0088] The result output layer is used to output the prediction analysis result of the patient's pituitary adenoma.

[0089] In the invasion prediction layer, the model uses a convolution kernel to process the patient's pituitary adenoma invasion feature image, which can extract rich spatial features from the image and identify information such as the shape, edge, and structure of the tumor. Combining with the pituitary adenoma invasion feature image can help the model capture the important features of tumor invasiveness. The cross-attention mechanism can dynamically adjust the weights of different input data by jointly modeling the invasion data-assisted predictors and the patient's pituitary adenoma invasion feature image. The cross-attention mechanism can enhance the information fusion effect, enabling the model to better integrate information from different modalities and improve the prediction accuracy. The invasion prediction layer not only relies on the pituitary adenoma invasion feature image but also combines the invasion data-assisted predictors for multi-modal joint prediction. This method combines the morphological information of the image and the numerical information of the biomarker data, and can more comprehensively evaluate the invasiveness of the tumor.

[0090] The specific steps for predictive analysis in the invasion prediction layer include:

[0091] The invasion prediction layer includes M convolutional kernels with gradually increasing sizes, a parallel cross-attention fusion layer, and a fully connected prediction layer;

[0092] Perform high-dimensional feature extraction on the patient's pituitary adenoma invasion feature image using M convolutional kernels, and at the same time use the parallel cross-attention fusion layer to fuse the invasion data auxiliary prediction factors for feature fusion to obtain the patient's pituitary adenoma invasion feature prediction vector;

[0093] Perform invasion prediction analysis based on the patient's pituitary adenoma invasion feature prediction vector in the fully connected prediction layer to obtain the patient's pituitary adenoma prediction analysis result;

[0094] In the invasion prediction layer, use M convolutional kernels with gradually increasing sizes to perform multi-level and multi-scale high-dimensional feature extraction on the patient's pituitary adenoma invasion feature image. Each convolutional kernel will process image features of different scales, capturing detailed information in the tumor image, such as morphology, edges, texture, etc.; Convolutional kernels of different sizes can extract features of different scales, effectively capturing features from details to the global level, improving the fineness of the model's judgment of tumor invasiveness; On the basis of convolutional feature extraction, a parallel cross-attention fusion layer is adopted to fuse the invasion data auxiliary prediction factors and the features of the patient's pituitary adenoma invasion feature image; The cross-attention mechanism assigns different weights to image features and data features by calculating the correlation between different modalities, thereby focusing on important information and enhancing the comprehensive ability of the model; After feature fusion, the final patient's pituitary adenoma invasion feature prediction vector will be sent to the fully connected prediction layer. The fully connected layer maps the high-dimensional feature vector to a low-dimensional space for final predictive analysis and outputs the invasion prediction result of the patient's pituitary adenoma;

[0095] Pituitary adenoma is a common benign brain tumor that usually originates from the anterior pituitary gland. Depending on the type, size, and growth direction of the tumor, common invasion sites of pituitary adenoma include the optic chiasm, which may cause visual impairment or homonymous hemianopia; adjacent cerebral blood vessels, such as the internal carotid artery, which may lead to hemodynamic problems; brain structures around the sella turcica, such as the hippocampus, which may cause memory disorders; the brainstem, which may cause changes in vital signs in extreme cases; the sphenoid sinus and paranasal sinuses, which may cause headache and nasal congestion; and cranial nerves, especially the trigeminal nerve and the optic nerve, which may cause facial numbness or vision problems; Performing invasion prediction based on pituitary adenoma images can accurately identify the compression or invasion of the tumor on surrounding structures;

[0096] If it is predicted that the pituitary adenoma of the patient is of low invasiveness, the tumor does not invade surrounding important structures such as the optic chiasm, brainstem, etc., and the growth rate is slow, conservative treatment is generally selected to observe whether the tumor is stable or shrinks spontaneously; if it is predicted that the pituitary adenoma of the patient has moderate invasiveness, it may have begun to invade adjacent tissues, and the growth rate of the tumor is relatively fast, then surgical resection of the tumor may be recommended, especially when the patient shows symptoms such as visual impairment, headache, etc.; specific actual analysis is carried out according to the predicted analysis results of the patient's pituitary adenoma.

[0097] Example 2, a system for analyzing imaging data of pituitary adenoma invasion, see Figure 1 shown in the figure, including:

[0098] An imaging feature extraction module, including an image acquisition unit and a feature extraction unit; the image acquisition unit is used to acquire a set of pituitary adenoma images of the patient; the set of pituitary adenoma images of the patient contains N consecutive pituitary adenoma images P of the patient at time windows, n n = 1, 2,..., N, the length of the time window is L; the feature extraction unit is used to process based on the set of pituitary adenoma images of the patient and the pituitary adenoma image feature extraction model to obtain the pituitary adenoma invasion feature image; the pituitary adenoma image feature extraction model uses a bidirectional propagation network to perform image processing on the set of pituitary adenoma images of the patient to obtain a high-quality pituitary adenoma invasion feature image for subsequent analysis;

[0099] An auxiliary factor analysis module, including a data acquisition unit and an auxiliary analysis unit; the data acquisition unit is used to acquire the patient's auxiliary biomarker data; the auxiliary analysis unit is used to analyze based on the patient's auxiliary biomarker data and the pituitary adenoma biomarker prediction analysis model to obtain an invasion data auxiliary prediction factor;

[0100] An invasion prediction and judgment module, including a prediction judgment unit and a result output unit; the prediction judgment unit is used to perform prediction analysis based on the invasion data auxiliary prediction factor, the pituitary adenoma invasion feature image and the pituitary adenoma invasion prediction model to obtain the predicted analysis result of the patient's pituitary adenoma; the pituitary adenoma invasion prediction model uses a cross-attention mechanism and a feature fusion strategy for prediction analysis to perform high-precision invasion prediction based on the invasion data auxiliary prediction factor and the pituitary adenoma invasion feature image; the result output unit is used to perform subsequent operations on the patient according to the predicted analysis result of the patient's pituitary adenoma.

[0101] It should be understood that for those of ordinary skill in the art, improvements or transformations can be made according to the above description, and all such improvements and transformations should fall within the protection scope of the appended claims of the present invention. The parts not described in detail in this specification belong to the prior art well-known to those of ordinary skill in the art.

Claims

1. A method for analyzing imaging data of pituitary adenoma invasion, characterized in that, Including the following steps: Obtain a set of images of a patient's pituitary adenoma; the set of images of the patient's pituitary adenoma contains N consecutive images P of the patient's pituitary adenoma at interval time windows n , where n = 1, 2, …, N, and the length of the time window is L; process based on the set of images of the patient's pituitary adenoma and the pituitary adenoma image feature extraction model to obtain the patient's pituitary adenoma invasion feature image; the pituitary adenoma image feature extraction model uses a bidirectional propagation network to process the set of images of the patient's pituitary adenoma for obtaining a high-quality patient's pituitary adenoma invasion feature image for subsequent analysis; Obtain patient auxiliary biomarker data; analyze based on the patient auxiliary biomarker data and the pituitary adenoma biomarker prediction analysis model to obtain an invasion data auxiliary prediction factor; Based on the invasion data auxiliary prediction factor, the patient pituitary adenoma invasion characteristic image, and the pituitary adenoma invasion prediction model, conduct a prediction analysis to obtain the patient pituitary adenoma prediction analysis result; the pituitary adenoma invasion prediction model uses a cross-attention mechanism and a feature fusion strategy for prediction analysis, and is used for high-precision invasion prediction based on the invasion data auxiliary prediction factor and the patient pituitary adenoma invasion characteristic image; perform subsequent operations on the patient according to the patient pituitary adenoma prediction analysis result; The pituitary adenoma image feature extraction model includes an image preprocessing layer, an image feature extraction layer, and an image output layer; The image preprocessing layer is used to perform image super-resolution processing on the patient pituitary adenoma image set to obtain a preprocessed patient pituitary adenoma image set; among them, the preprocessed patient pituitary adenoma image set contains N preprocessed patient pituitary adenoma images C n ; The image feature extraction layer is used to extract features from the preprocessed patient pituitary adenoma image set to obtain the patient pituitary adenoma invasion characteristic image; The image output layer is used to output the patient pituitary adenoma invasion characteristic image; The specific steps for image super-resolution processing in the image preprocessing layer include: The image preprocessing layer contains an image alignment layer and an image reconstruction layer; in the image alignment layer, there is a forward propagation network and a backward propagation network; in the image reconstruction layer, there is a feature extraction network, a residual refinement network, and an image reconstruction network; In the backpropagation network, the pituitary adenoma image P of the patient is aligned using the image registration method n+1 , the pituitary adenoma image P of the patient n+2 , and the pituitary adenoma image P of the patient n are aligned to obtain the set of reverse pituitary adenoma images F n ; the set of reverse pituitary adenoma images F n is fed into the deformable convolution block for further feature alignment to obtain the reverse pituitary adenoma image F n '; In the forward propagation network, the pituitary adenoma images P of the patient are aligned using the image registration method n-1 、the pituitary adenoma images P of the patient n-2 and the pituitary adenoma images P of the patient n to obtain the forward pituitary adenoma image set Z n ; The forward pituitary adenoma image set Z n is sent into the deformable convolution block for further feature alignment to obtain the forward pituitary adenoma image Z n ’; In the image reconstruction layer, the feature extraction network is used to extract the reverse pituitary adenoma image F n ’, the forward pituitary adenoma image Z n ’ and the pituitary adenoma image P n for feature extraction and splicing to obtain the pituitary adenoma feature image T n ; The residual refinement network is used to refine the features of the pituitary adenoma feature image T n through multiple residual blocks to obtain the refined pituitary adenoma feature image T n ’; In the image reconstruction network, the refined pituitary adenoma feature image T n ’ is enlarged four times using the pixel rearrangement method to obtain the high-resolution pituitary adenoma feature image G n ; The pituitary adenoma image P n is enlarged four times using bilinear interpolation to obtain the low-resolution pituitary adenoma feature image D n ; The high-resolution pituitary adenoma feature image G n and the low-resolution pituitary adenoma feature image D n are added together to obtain the preprocessed pituitary adenoma image C n ; The specific steps for feature extraction in the image feature extraction layer include: The image feature extraction layer contains a global feature extraction layer, a local feature division layer, and a feature scoring layer; In the global feature extraction layer, the preprocessed pituitary adenoma image C of the patient is n subjected to global feature extraction to obtain the global feature vector Q of the pituitary adenoma features of the patient n ; In the local feature division layer, the preprocessed pituitary adenoma image C of the patient is divided into feature image patches by using a sliding window n to obtain a set J of pituitary adenoma feature image patches of the patient n ; The set J of pituitary adenoma feature image patches of the patient n contains I pituitary adenoma feature image patches of the patient; Based on the set J of pituitary adenoma feature image patches of the patient n feature extraction is performed to obtain a local feature vector J of pituitary adenoma features of the patient ni , i = 1, 2,..., I; The specific steps for extracting feature image blocks: Using the formula I = (X + Y - X') / Y, where I is the total number of feature image patches extracted from the pre-processed pituitary adenoma image C of the patient, X is the height of the pre-processed pituitary adenoma image C of the patient, X' is the horizontal height of the pituitary adenoma feature image patches of the patient after division; Y is the step size of the sliding window; n X is the height of the pre-processed pituitary adenoma image C of the patient n X' is the horizontal height of the pituitary adenoma feature image patches of the patient after division; Y is the step size of the sliding window; In the feature scoring layer, the global feature vector Q of the patient's pituitary adenoma features n and the local feature vector J of the patient's pituitary adenoma features ni are feature scored to obtain the feature vector score set U of the patient's pituitary adenoma features n ; Based on all the feature vector score sets U of the patient's pituitary adenoma features n are screened, and the feature vectors corresponding to the screening conditions are feature fused to obtain the invasion feature image of the patient's pituitary adenoma; The feature scoring layer is trained based on the attention mechanism; The pituitary adenoma biomarker prediction analysis model includes a data preprocessing layer, a feature judgment layer, and a feature output layer; The data preprocessing layer is used to preprocess the patient auxiliary biomarker data to obtain the preprocessed patient auxiliary biomarker data; The feature judgment layer is used to judge the features of the preprocessed patient auxiliary biomarker data to obtain the invasion data auxiliary prediction factor; The feature output layer is used to output the invasion data auxiliary prediction factor.

2. The method for analyzing pituitary adenoma invasion image data according to claim 1, wherein The pituitary adenoma invasion prediction model includes an invasion prediction layer and a result output layer; The invasion prediction layer is used to conduct a prediction analysis based on the invasion data auxiliary prediction factor and the patient pituitary adenoma invasion characteristic image to obtain the patient pituitary adenoma prediction analysis result; The invasion prediction layer is constructed based on the CNN model and the cross-attention mechanism; The result output layer is used to output the patient pituitary adenoma prediction analysis result.

3. The method for analyzing pituitary adenoma invasion image data according to claim 2, characterized in that The specific steps for prediction analysis in the invasion prediction layer include: The invasion prediction layer includes M convolutional kernels with gradually increasing sizes, a parallel cross-attention fusion layer, and a fully connected prediction layer; In the M convolutional kernels, perform high-dimensional feature extraction on the patient pituitary adenoma invasion characteristic image, and at the same time use the parallel cross-attention fusion layer to fuse the invasion data auxiliary prediction factor for feature fusion to obtain the patient pituitary adenoma invasion characteristic prediction vector; In the fully connected prediction layer, conduct invasion prediction analysis based on the patient pituitary adenoma invasion characteristic prediction vector to obtain the patient pituitary adenoma prediction analysis result.

4. A method for analyzing pituitary adenoma invasion imaging data according to claim 3, characterized in that The specific steps for training the feature judgment layer: Collect several groups of biomarker level training samples; each group of biomarker level training samples contains verified characteristic factors and corresponding biomarker data; combine several groups of biomarker level training samples to obtain a biomarker level training set; Input the biomarker level training set into a BP neural network for model training to obtain an initial feature judgment layer; Conduct model evaluation on the initial feature judgment layer. If the initial feature judgment layer passes the model evaluation, use the initial feature judgment layer as the feature judgment layer in the pituitary adenoma biomarker prediction analysis model; otherwise, continue model training using the biomarker level training set.

5. A pituitary adenoma invasion imaging data analysis system, characterized in that, The system applies the method for analyzing pituitary adenoma invasion imaging data according to any one of claims 1-4 above, including: The image feature extraction module includes an image acquisition unit and a feature extraction unit; the image acquisition unit is used to acquire a set of pituitary adenoma images of a patient; the set of pituitary adenoma images of the patient contains N consecutive pituitary adenoma images P of the patient with an interval time window, where n = 1, 2,..., N, and the length of the time window is L; the feature extraction unit is used to process based on the set of pituitary adenoma images of the patient and the pituitary adenoma image feature extraction model to obtain a pituitary adenoma invasion feature image; the pituitary adenoma image feature extraction model uses a bidirectional propagation network to perform image processing on the set of pituitary adenoma images of the patient to obtain a high-quality pituitary adenoma invasion feature image for subsequent analysis; n , where n = 1, 2,..., N, and the length of the time window is L; the feature extraction unit is used to process based on the set of pituitary adenoma images of the patient and the pituitary adenoma image feature extraction model to obtain a pituitary adenoma invasion feature image; the pituitary adenoma image feature extraction model uses a bidirectional propagation network to perform image processing on the set of pituitary adenoma images of the patient to obtain a high-quality pituitary adenoma invasion feature image for subsequent analysis; An auxiliary factor analysis module, including a data acquisition unit and an auxiliary analysis unit; the data acquisition unit is used to acquire patient auxiliary biomarker data; the auxiliary analysis unit is used to analyze based on the patient auxiliary biomarker data and the pituitary adenoma biomarker prediction analysis model to obtain an invasion data auxiliary prediction factor; An invasion prediction judgment module, including a prediction judgment unit and a result output unit; the prediction judgment unit is used to conduct prediction analysis based on the invasion data auxiliary prediction factor, the patient's pituitary adenoma invasion characteristic image, and the pituitary adenoma invasion prediction model to obtain the patient's pituitary adenoma prediction analysis result; the pituitary adenoma invasion prediction model conducts prediction analysis using a cross-attention mechanism and a feature fusion strategy, and is used to conduct high-precision invasion prediction based on the invasion data auxiliary prediction factor and the patient's pituitary adenoma invasion characteristic image; the result output unit is used to perform subsequent operations on the patient according to the patient's pituitary adenoma prediction analysis result.

Images