A brain tumor detection method based on attention mechanism and MRI multi-modal fusion

Abstract

The application discloses a brain tumor detection method based on an attention mechanism and MRI multi-modal fusion, relates to the technical field of brain tumor image analysis, and comprises the following steps: acquiring a first tumor heat map and a second tumor heat map; calculating a consistency measurement score of the two heat maps; if the score does not satisfy a preset threshold, starting feedback iteration: merging images to generate an enhanced training set, and performing weighted training on a loss term corresponding to a multi-modal image pair by taking the normalized measurement score as a quality weight, and iteratively optimizing a model until a condition is satisfied; and finally outputting a brain tumor detection result with high confidence. The application introduces an internal closed-loop verification and self-adaptive optimization mechanism, effectively solves the core problem that the confidence of a detection result cannot be verified and guaranteed in the prior art, and finally realizes the accuracy of closed-loop verification and optimization of the confidence of brain tumor detection based on the attention mechanism and the MRI multi-modal fusion.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of brain tumor image analysis, more particularly, to a brain tumor detection method based on attention mechanism and MRI multi-modal fusion. BACKGROUND

[0002] The precise detection of brain tumors is one of the core challenges in the field of medical image analysis, and multi-modal sequences of magnetic resonance imaging provide the key basis. Although detection methods based on deep learning, especially models introducing attention mechanisms, have made significant progress in this field, the existing technology still has inherent defects in multiple key links, which restricts the accuracy and robustness of its clinical application.

[0003] Most existing methods are limited to shallow fusion of multi-modal MRI images themselves, and fail to deeply integrate key clinical metadata such as patient age, pathological type, and treatment history as prior knowledge into the detection model, resulting in a lack of clinical relevance in model decision-making.

[0004] The generative model used by existing methods to expand the data set lacks effective quality evaluation and control mechanisms, and the authenticity and clinical consistency of the generated images cannot be guaranteed. During the training process, all samples are treated equally without differentiated weighting based on their quality or difficulty, so that low-quality synthetic data may introduce noise and affect model performance.

[0005] The detection process of existing methods usually relies on fixed image preprocessing parameters and network structures, which cannot be dynamically adjusted according to the image characteristics of specific cases, resulting in insufficient generalization ability for irregular or complex cases.

[0006] In addition, for specific detection of post-treatment patients, existing technology is difficult to accurately quantify the relationship between the treatment effect area and the tumor activity area, and lacks a mechanism to combine treatment history data such as radiotherapy dose with image features for consistent evaluation. These defects collectively result in the limitations of existing brain tumor detection methods in terms of precision, adaptability, and clinical practicality.

[0007] Due to the lack of internal closed-loop verification mechanisms for the reliability of generated data and the consistency of model output, the system cannot perform self-evaluation and optimization during the detection process, resulting in an inability to guarantee the confidence of its output results. This inherent contradiction of an open-loop system architecture is an insurmountable limitation of existing technology. SUMMARY

[0008] To solve the above technical problems, a brain tumor detection method based on attention mechanism and MRI multi-modal fusion is provided, which solves the problems raised in the background art.

[0009] In a first aspect, embodiments of this application provide a brain tumor detection method based on attention mechanisms and MRI multimodal fusion, comprising the following steps: S1, acquiring raw MRI images and processing them according to a preset segmentation method to obtain MRI segmented images, and generating a first tumor heatmap through processing with a preset attention detection model; S2, acquiring patient metadata and processing it with the MRI segmented images through a preset image synthesis model to generate a multimodal MRI segmented image, and processing the multimodal MRI segmented image through a preset attention detection model to generate a second tumor heatmap; S3, calculating the consistency metric between the first tumor heatmap and the second tumor heatmap to obtain a metric score; S4, determining whether the metric score is greater than a preset metric threshold; if yes, proceed to step S6; otherwise... Then, the following feedback iteration is executed; S5, the specific feedback iteration is as follows: merge the MRI segmentation image and the multimodal MRI segmentation image to generate an enhanced training set, and input it into the preset attention detection model for training. During the training process, the metric score is normalized to obtain the quality weight parameter, and the loss term corresponding to the multimodal MRI segmentation image in the preset attention detection model is weighted accordingly. The loss term is updated according to the weighting result, and S1 to S4 are executed again and the number of feedback iterations is recorded. If the number of feedback iterations reaches the preset upper limit or the improvement of the metric score in the preset number of consecutive iterations is less than the preset threshold, the feedback iteration is exited; S6, the brain tumor detection result is output according to the preset attention detection model after the feedback iteration.

[0010] In a second aspect, the application provides a brain tumor detection system based on an attention mechanism and MRI multi-modal fusion, comprising: a first tumor heat map processing module: configured to obtain an MRI original image and process it according to a preset segmentation method, to obtain an MRI segmentation image and generate a first tumor heat map by processing the MRI segmentation image through a preset attention detection model; a second tumor heat map processing module: configured to obtain patient metadata and generate a multi-modal MRI segmentation image by processing the MRI segmentation image and the patient metadata through a preset image synthesis model, and generate a second tumor heat map by processing the multi-modal MRI segmentation image through the preset attention detection model; a calculation module: configured to calculate a consistency measure of the first tumor heat map and the second tumor heat map, to obtain a measure score; a judgment module: configured to determine whether the measure score is greater than a preset measure threshold, and if yes, enter an output module, and if no, execute the following feedback iteration; a feedback iteration module: configured to perform the following specific feedback iteration: merge the MRI segmentation image and the multi-modal MRI segmentation image to generate an enhanced training set, and input the enhanced training set into the preset attention detection model for training, in the training process, normalize the measure score to obtain a quality weight parameter, and perform weighted operation on a loss term corresponding to the multi-modal MRI segmentation image in the preset attention detection model according to the quality weight parameter, update the loss term according to the weighted operation result, re-execute the first tumor heat map processing module to the judgment module, and record the number of feedback iterations, and if the number of feedback iterations reaches a preset upper limit or the improvement amplitude of the measure score in a preset consecutive number of iterations is less than a preset threshold, exit the feedback iteration; and an output module: configured to output a brain tumor detection result according to the preset attention detection model after feedback iteration.

[0011] The one or more technical solutions provided in the embodiments of the application have at least the following technical effects or advantages:

[0012] 1. The first heat map is based on original image data, and the second heat map introduces a multi-modal synthetic image generated by patient metadata and more clinically relevant. By calculating the consistency measure score between the two, an internal and quantitative confidence indicator can be output for a single detection task. Only when this internal confidence indicator meets the preset standard, the result will be finally output, thereby fundamentally avoiding the use of low-confidence results, and greatly improving the overall reliability of the output result.

[0013] 2. When the consistency measure score does not meet the standard, instead of simply reporting an error or outputting a low-quality result, a built-in feedback iteration loop is started. The data generated in this detection task is merged into an enhanced training set and immediately used to retrain the preset attention detection model, making the model a dynamic optimization adaptive model for the current specific patient. This solves the problem of performance degradation of fixed models when dealing with cases with poor image quality, special tumor morphology, or large differences in distribution from the training set. Through iteration, the model can learn how to better handle this specific case, significantly improving the robustness and generalization ability of the method.

[0014] 3. A low consistency score means that the multi-modal image generated this time may deviate significantly from the true situation. After normalization, the measure score is used as a quality weight parameter, and the loss term corresponding to the multi-modal MRI segmentation image is specially weighted. The corresponding learning goal should be given a lower weight to avoid the model being misled by "bad samples". This ensures that the optimization direction of the model is always to improve the confidence of the final output, thereby accelerating the convergence speed and making the optimization process more efficient and accurate. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 A structural schematic diagram of a brain tumor detection method based on attention mechanism and MRI multi-modal fusion provided by an embodiment of the present application;

[0016] Figure 2 A general logic flowchart of a brain tumor detection method based on attention mechanism and MRI multi-modal fusion provided by an embodiment of the present application;

[0017] Figure 3 A logic flowchart of the dynamic correction step of the quality weight parameter provided by an embodiment of the present application;

[0018] Figure 4 A logic flowchart of the multi-scale segmentation negative feedback step provided by an embodiment of the present application;

[0019] Figure 5 A structural schematic diagram of a brain tumor detection system based on attention mechanism and MRI multi-modal fusion provided by an embodiment of the present application. DETAILED DESCRIPTION

[0020] The brain tumor detection method based on attention mechanism and MRI multi-modal fusion provided by the embodiment of the present application solves the technical problem of insufficient accuracy of brain tumor detection confidence closed-loop verification optimization based on attention mechanism and MRI multi-modal fusion in the prior art.

[0021] To solve the above problems, the inventors noticed that the relevance of clinical metadata and image features was not fully utilized, and considered combining patient information with medical images through feature fusion technology. To address the problem of uncontrollable data quality, a closed-loop verification mechanism was proposed to guide iterative optimization of the model through a hot Figure 1 consistency evaluation. To address the parameter solidification defect, a dynamic adjustment strategy was designed to enable the system to adaptively optimize the processing flow according to the specific case characteristics. These thoughts ultimately formed a technical route combining multi-modal fusion and feedback iteration.

[0022] Therefore, the present application proposes a brain tumor detection method based on attention mechanism and MRI multi-modal fusion, including the following steps: obtaining the MRI original image and processing it according to the preset segmentation method to obtain the MRI segmentation image, and generating the first tumor heat map through the preset attention detection model; obtaining patient metadata and synthesizing multi-modal MRI segmentation images with the MRI segmentation image to generate the second tumor heat map; calculating the consistency measure score of the two heat maps; determining whether to perform feedback iteration training according to the score result, and weighting and optimizing the loss term in the training using the measure score; repeating the iteration until the termination condition is met; and finally outputting the optimized detection result.

[0023] Compared with the prior art, the present scheme integrates clinical information deeply into the detection process through the metadata fusion mechanism, solving the problem of ignoring individualized features in traditional methods. The introduction of the closed-loop verification system realizes the quality self-monitoring of the detection process, overcoming the defect of insufficient reliability of generated data. The dynamic iteration mechanism enables the model to adjust the processing strategy according to the specific case characteristics, improving the adaptability of the system to different clinical scenarios. The quantitative evaluation method of treatment effect area effectively distinguishes the tissue repair response from the tumor active area, making up for the deficiencies of the prior art in this field.

[0024] In order to better understand the above technical solutions, the above technical solutions will be described in detail below in conjunction with the drawings in the specification and specific embodiments.

[0025] As Figure 1As shown, a structure schematic diagram of a brain tumor detection method based on attention mechanism and MRI multi-modal fusion provided by an embodiment of the present application, comprising: S1, acquiring an MRI original image and processing it according to a preset segmentation method to obtain an MRI segmentation image and generate a first tumor heat map through a preset attention detection model; S2, acquiring patient metadata and processing it with the MRI segmentation image through a preset image synthesis model to generate a multi-modal MRI segmentation image, and processing the multi-modal MRI segmentation image through a preset attention detection model to generate a second tumor heat map; S3, calculating the consistency measure of the first tumor heat map and the second tumor heat map to obtain a measure score; S4, judging whether the measure score is greater than a preset measure threshold, yes to step S6, otherwise executing the following feedback iteration; S5, the specific feedback iteration is as follows: merging the MRI segmentation image and the multi-modal MRI segmentation image to generate an enhanced training set, and inputting it into the preset attention detection model for training, in the training process, obtaining a quality weight parameter after normalizing the measure score and weighting the loss term corresponding to the multi-modal MRI segmentation image in the preset attention detection model according to the quality weight parameter, updating the loss term according to the weighting operation result, re-executing S1 to S4 and recording the feedback iteration number, if the feedback iteration number reaches a preset upper limit or the improvement amplitude of the measure score in a preset continuous number of iterations is less than a preset threshold, the feedback iteration is exited; S6, outputting the brain tumor detection result according to the feedback iteration of the preset attention detection model.

[0026] Figure 2 The overall logic flow schematic diagram of a brain tumor detection method based on attention mechanism and MRI multi-modal fusion provided by an embodiment of the present application, wherein the MRI segmentation image refers to a structured image obtained after standardization preprocessing of the original MRI data, which can be specifically realized by a sliding window segmentation combined with morphological filtering method, for eliminating image noise and retaining key anatomical structures.

[0027] The multi-modal MRI segmentation image refers to a composite image fusing patient clinical data and image features, which can be specifically realized by a combination of feature encoding and generative adversarial network, for enhancing the perception ability of the model to individualized features.

[0028] The consistency measure refers to an evaluation index for quantifying the consistency of different modalities, which can be specifically calculated by a weighted combination of Dice coefficient and Hausdorff distance, for comprehensively evaluating the reliability of the model output.

[0029] The feedback iteration refers to a model optimization process based on the quality evaluation result, which can be specifically realized by dynamically adjusting the loss function weight parameter, for improving the attention degree of the model to key features.

[0030] Further, S5 further comprises a quality weight parameter dynamic correction step: a preset basic adjustment coefficient is extracted from a brain tumor detection database; a treatment effect area is output by performing treatment effect area recognition segmentation on the multi-modal MRI segmentation image through a preset recognition segmentation model, the treatment effect area comprising a tumor necrosis area image, an edema zone image and an active tumor area image, the preset recognition segmentation model comprising a first preset recognition segmentation model, a second preset recognition segmentation model and a third preset recognition segmentation model; a spatial overlap degree of the treatment effect area and the original tumor area in the multi-modal MRI segmentation image is calculated to generate a treatment response consistency score; based on the treatment response consistency score, the quality weight parameter is dynamically corrected, and the correction formula is as follows: ; wherein, represents the corrected quality weight parameter, represents the quality weight parameter before correction, represents the treatment response consistency score, is a preset basic adjustment coefficient; the corrected quality weight parameter is used for the next feedback iteration.

[0031] In the present embodiment, Figure 3 is a logical flow diagram of the quality weight parameter dynamic correction step provided by the present embodiment, wherein the preset basic adjustment coefficient refers to a benchmark parameter for balancing the influence degree of the treatment response consistency score on the weight parameter, which can be realized by an empirical value or a value obtained by training historical data, and is used to prevent the weight parameter from fluctuating sharply due to the mutation of the treatment response consistency score.

[0032] The treatment effect area refers to an imaging change area caused by treatment, specifically including a necrotic core area identified by a radiotherapy dose distribution map, an edema zone area identified by morphological features, and an active tumor area identified by dynamic contrast enhancement.

[0033] The treatment response consistency score refers to an index quantifying the spatial matching degree of the treatment effect area and the original tumor area, which is obtained by calculating the spatial overlap degree and combining a treatment time decay factor, and is used to reflect the correlation degree of treatment intervention and imaging performance.

[0034] In the model training process, the treatment effect sub-regions are extracted by multiple dedicated segmentation models, and then the spatial matching degrees of these sub-regions and the original tumor region are calculated. By introducing a treatment time decay factor, the time dimension is incorporated into the scoring system, so that the effects of recent treatment are given higher weight in the scoring calculation. The preset basic adjustment coefficient in the correction formula serves as a balancing term to ensure that some of the original weight parameters retain useful information when the treatment response consistency is low. This dynamic correction mechanism enables the quality weight parameter to be adaptively adjusted according to the true correlation between the treatment effect and the tumor region, thereby optimizing the utilization efficiency of multi-modal data in the training process.

[0035] By introducing the spatial matching degree calculation of the treatment effect region, the accuracy of the model in identifying the active tumor region after treatment is improved. Combined with the time decay factor to dynamically adjust the quality weight parameter, the system's perception of the timeliness of treatment intervention is enhanced, thereby improving the spatiotemporal adaptability and reliability of the brain tumor detection model in complex clinical scenarios.

[0036] Further, the treatment effect region identification and segmentation specifically includes: extracting the radiotherapy dose distribution map in the patient treatment history data from the brain tumor detection database, the radiotherapy dose distribution map being consistent in size with the multi-modal MRI segmentation image; based on the treatment effect region obtained after matching the radiotherapy dose distribution map and the multi-modal MRI segmentation image, performing segmentation identification on the treatment effect region with a radiotherapy dose higher than a preset radiotherapy dose threshold through a first preset identification segmentation model to obtain a necrotic core region and a tumor necrosis zone image; segmenting and identifying an edema zone around the necrotic core region through a second preset identification segmentation model to obtain an edema zone image; and segmenting and identifying the treatment effect region identified through dynamic contrast enhancement through a third preset identification segmentation model to obtain an active tumor zone image.

[0037] In this embodiment, the radiotherapy dose distribution map refers to a spatial distribution map recording the actual radiation dose of each part of the patient during radiotherapy, which can be implemented by extracting the derived DICOM RT dose file format from the brain tumor detection database. The spatial registration with the MRI image is completed through a rigid or non-rigid registration algorithm.

[0038] The treatment effect region refers to the region that has undergone histopathological changes after radiotherapy, which can be defined by imageomic feature extraction combined with morphological analysis to distinguish necrosis, edema, and active tumor tissue.

[0039] The first preset identification segmentation model refers to a deep learning segmentation network specifically for high-dose radiation regions, which can be implemented by a convolutional neural network based on the U-Net architecture, and the training data set contains a large number of labeled samples of necrotic tissue after radiotherapy.

[0040] The second preset recognition segmentation model refers to a special segmentation model for recognizing the edema zone around the necrosis zone, and can be specifically implemented by using a network with an attention mechanism having edge perception capability to segment by strengthening the boundary features of the edema zone and normal tissues.

[0041] The third preset recognition segmentation model refers to a segmentation model for the active tumor region, and can be specifically implemented by using a 3D convolutional neural network trained by a dynamic contrast-enhanced MRI sequence to recognize by capturing the enhanced features caused by tumor angiogenesis.

[0042] First, spatial registration is used to ensure that the radiotherapy dose distribution map and the multi-modal MRI segmentation image are anatomically aligned. In the high-dose radiation region, due to the direct killing effect of radiation on tumor cells, this region mainly contains necrotic tissue, which is accurately segmented by the first preset recognition segmentation model. The edema zone around the necrotic core is formed by the inflammatory response and vascular permeability changes induced by radiotherapy, and the boundary is detected based on the water content difference by using the second preset recognition segmentation model. For the remaining active tumor region, the enhanced features of the newly formed blood vessels are captured by using a dynamic contrast-enhanced MRI sequence, and the metabolically active tumor tissue is recognized by using the third preset recognition segmentation model. The collaborative work of the three special models realizes the fine division of the treatment effect region.

[0043] Traditional methods usually rely only on a single MRI sequence for rough segmentation, which cannot effectively distinguish between necrotic tissue and active tumor caused by radiotherapy, lacks spatial correlation analysis of radiotherapy dose distribution, and leads to segmentation results that are disconnected from treatment history data.

[0044] The present scheme integrates radiotherapy dose maps and multi-modal MRI data to establish a dose-effect correlation model, and uses a dose-dependent segmentation strategy to avoid the risk of misjudgment of post-radiotherapy changes as tumor progression in traditional methods. The multi-model collaborative working mechanism improves the segmentation consistency of complex cases.

[0045] Further, the calculation of the treatment response consistency score specifically includes: calculating the Dice coefficients of the tumor necrosis zone image, the edema zone image, and the active tumor region image with the original MRI image respectively, and weighting and summing to obtain a response consistency coefficient; extracting the treatment history data of the patient from the brain tumor detection database to obtain the treatment days and the preset tissue repair time constant; the specific formula of the treatment response consistency score is as follows: , represents the response consistency coefficient, wherein , wherein represents a treatment time calculation attenuation factor, is the treatment days, is the preset tissue repair time constant.

[0046] In this embodiment, the Dice coefficient refers to a similarity measurement index by calculating the degree of overlap of two sample sets, which can be realized by using the spatial overlap algorithm commonly used in the field of medical image segmentation, and is used to quantify the spatial consistency of the treatment effect region and the original tumor region.

[0047] Weighted summation refers to linear superposition after different regions are given different weights, which can be realized by using the weight distribution strategy based on the experience of clinical experts, and is used to reflect the difference in clinical importance of different tumor sub-regions in treatment response evaluation.

[0048] Treatment days refer to the time interval from the end of the last treatment to the current MRI examination, which can be automatically extracted from the treatment record data in the hospital information system, and is used to represent the time cumulative effect after treatment.

[0049] Tissue repair time constant refers to a biological parameter describing the self-repair rate of a specific tissue type after radiotherapy, which can be realized by using the tissue-specific repair rate constant obtained from radiobiology experiments, and is used to quantify the influence of time factor on treatment effect evaluation.

[0050] Specifically, after obtaining the segmentation results of the necrotic tumor region, the edema zone and the active tumor region, the Dice coefficients of each region and the corresponding region of the original MRI image are calculated respectively. For example, the Dice coefficient calculation of the necrotic tumor region can reflect the consistency of the volume change of the region before and after treatment. Through the weight coefficient set by the experience of clinical experts, for example, the weight of the necrotic region is 0.5, the weight of the edema zone is 0.3, and the weight of the active region is 0.2. The weighted sum of the Dice coefficients of the three regions is obtained. DY value. At the same time, the actual treatment days T of the patient and the preset τ value are extracted from the database, and the time attenuation factor δ is calculated by the exponential function. Finally, DY is multiplied by δ to obtain the RT score, which contains both the spatial consistency measurement and the influence of time attenuation effect.

[0051] The traditional method only performs static evaluation based on the image features at a single time point, without considering the time dynamic change factor after treatment. However, the present scheme introduces the tissue repair time constant and the treatment days to construct an exponential decay model, so that the treatment response score can dynamically reflect the influence of the biological repair process on the evaluation result. In addition, the existing technology usually only focuses on the consistency evaluation of a single tumor region, while the present scheme realizes the comprehensive evaluation of complex tumor substructure through a multi-region weighting mechanism.

[0052] The present application realizes the spatio-temporal quantitative analysis in treatment response evaluation, solves the evaluation deviation problem caused by ignoring the time cumulative effect of the existing method, and improves the evaluation accuracy of heterogeneous tumors through a multi-region weighting mechanism.

[0053] Further, the loss term corresponding to the multi-modal MRI segmentation image in the preset attention detection model is subjected to weighting operation, and the specific process is as follows: the metric score and the preset normalization parameter under the current feedback iteration are obtained from the brain tumor detection database; the metric score is normalized by the preset normalization parameter to generate a quality weight parameter; the loss term corresponding to the multi-modal MRI segmentation image is subjected to weighting operation calculation based on the quality weight parameter to generate a weighted loss term; and the model parameter corresponding to the preset attention detection model is updated based on the weighted loss term.

[0054] In the embodiment, the preset normalization parameter refers to a scaling factor for mapping the metric score to a weight interval, and can be specifically implemented by using a maximum-minimum value normalization or Z-score standardization method, for eliminating the dimension difference of scores in different iteration stages.

[0055] The quality weight parameter refers to a converted value of the metric score after normalization processing, and can be specifically implemented by using linear transformation or non-linear activation function processing, for representing the credibility weight of the current multi-modal data in model training.

[0056] The weighted loss term refers to a loss calculation result fused with the quality weight parameter, and can be specifically implemented by using the product operation of the cross-entropy loss and the weight parameter, for realizing differentiated gradient update in the back propagation process.

[0057] Specifically, in the model training stage, the metric score generated in the current iteration and the preset normalization parameter are called from the database at each feedback iteration. The metric score is linearly mapped to the 0-1 interval by the preset normalization parameter to form a quality weight parameter, which is subjected to sample-by-sample multiplication operation with the cross-entropy loss term corresponding to the multi-modal MRI segmentation image, so that high-consistency samples obtain greater weight in loss calculation. The weighted loss term updates the convolution kernel parameter and the attention weight matrix of the attention detection model through the back propagation algorithm, optimizes the feature extraction capability of low-consistency samples, and reduces the negative influence of noise data.

[0058] The traditional method uses a fixed weight coefficient for all synthetic data in the training process, and cannot dynamically adjust the learning intensity according to the data reliability. The present scheme constructs a dynamic weight mechanism through real-time metric score, so that the model can autonomously identify high-quality synthetic data and strengthen its feature learning, while weakening the interference of low-quality data, effectively improving the generalization ability of the model in complex cases.

[0059] The application realizes dynamic perception and differentiated management of multi-modal synthetic data quality, automatically strengthens the learning weight of high consistency samples in the model training process, significantly reduces the parameter deviation problem caused by low-quality synthetic data, enables the attention detection model to maintain a stable performance improvement trajectory in the iterative optimization process, and finally improves the clinical credibility of brain tumor detection results.

[0060] Further, the specific process of obtaining the MRI original image and segmenting according to the pre-defined segmentation method is: different size segmentation windows are used to perform parallel segmentation on the original MRI image; different original first tumor heat maps are generated for each parallel segmented original MRI image, and the structural similarity with the preset reference heat map is calculated; the segmentation window with the highest similarity is selected as the pre-defined segmentation method, and the corresponding parallel segmented original MRI image is taken as the MRI segmented image.

[0061] In the embodiment, the segmentation window refers to a local area division unit for image segmentation, which can be implemented by using sliding windows of different sizes, such as 32x32 pixels, 64x64 pixels or 128x128 pixels rectangular windows. By adjusting the window size, different tumor morphological characteristics can be adapted.

[0062] Parallel segmentation refers to using multiple segmentation windows to independently process the same MRI image at the same time, which can be implemented through multi-threaded computing or distributed processing architecture to improve processing efficiency and preserve multi-scale feature information.

[0063] Structural similarity refers to the matching degree of heat maps and reference heat maps in spatial distribution and morphological characteristics, which can be calculated by using a structural similarity index algorithm. By comparing the similarity scores of three dimensions of brightness, contrast and tumor boundary clarity, it is ensured that the selected segmentation method can accurately reflect the true spatial distribution of the tumor.

[0064] Specifically, in the MRI image segmentation stage, the original MRI image is first input into multiple parallel processing channels, and each channel uses a segmentation window of different size to divide the image into regions. Each processing channel corresponding to a segmentation window independently generates a tumor heat map, and the heat map is labeled with a probability of tumor area by a preset attention detection model. Subsequently, the system calculates the structural similarity of the tumor heat maps generated by each channel with the reference heat map labeled by medical experts, and selects the segmentation window with the highest similarity score as the optimal segmentation scheme for the current case by quantitatively comparing the quantization indicators of each heat map in three dimensions of brightness, contrast and tumor boundary clarity. The segmentation result corresponding to the selected segmentation window will be used as the basic data for subsequent multi-modal fusion, ensuring that different cases can adaptively match the best segmentation parameters.

[0065] The optimal segmentation window can be dynamically selected according to the tumor characteristics of individual cases, the accuracy and adaptability of MRI image segmentation are improved, the tumor region missed or missegmented due to fixed segmentation parameters is reduced, and the reliability of subsequent multi-modal fusion and tumor detection is improved.

[0066] Further, S5 further comprises a multi-scale segmentation negative feedback step: obtaining a preset fluctuation penalty factor and a preset window data reward factor from a brain tumor detection database; recording the segmentation window size , the corresponding metric score and the score change rate in each feedback iteration , wherein indicates the current number of feedback iterations, and a historical data set is constructed ; the average metric score and the score fluctuation coefficient of each segmentation window size are calculated, and the specific average metric score formula is: , wherein is the set of iteration numbers of all segmentation windows in the historical data set, is the metric score in the historical data set; the specific score fluctuation coefficient formula is: ; in the preset consecutive number of feedback iterations, if all exist, the current score fluctuation coefficient is recorded, if the current score fluctuation coefficient is greater than the fluctuation coefficient threshold, S1 to S4 are re-executed and the number of re-executions is recorded, if the number of re-executions is greater than the preset upper limit, the expected utility score of each segmentation window size is calculated, and the specific calculation formula is as follows: ; wherein indicates the preset fluctuation penalty factor, indicates the preset window data reward factor; the corresponding segmentation window size with the highest expected utility score is taken as the new segmentation window size; the expected utility scores are sorted in size, the difference between the maximum expected utility score and the minimum expected utility score is obtained, and if the expected utility score difference is less than the preset differentiation threshold and all expected utility scores are lower than the preset rollback threshold, the MRI original image is rolled back, the multi-scale image block is generated by using the preset pyramid segmentation strategy, and S1 to S4 are re-executed.

[0067] In this embodiment, Figure 4 is the logical flowchart of the multi-scale segmentation negative feedback step provided by the embodiment, wherein the preset fluctuation penalty factor is an adjustment parameter for punishing the segmentation window size with excessive score fluctuation, which can be realized by using an exponential decay coefficient, and the utility score of the high fluctuation window is reduced to improve stability.

[0068] The preset window data reward factor refers to encouraging the adjustment parameter of the segmentation window with more historical data support, which can be realized by a linear growth coefficient, and the reliability is enhanced by increasing the utility score of the data-rich window.

[0069] The score fluctuation coefficient refers to a quantitative index reflecting the dispersion degree of the measurement score of a specific segmentation window in multiple iterations, which can be calculated by the ratio of the standard deviation to the average value, and is used to evaluate the stability of the segmentation window.

[0070] The expected utility score refers to a comprehensive evaluation index considering the accuracy stability and data support of the segmentation window, which can be calculated by an exponential weighted average formula combined with the data quantity ratio, and is used to guide the selection of the optimal segmentation window.

[0071] The pyramid segmentation strategy refers to an image processing method for hierarchical feature extraction through multi-scale image block generation, which can be realized by combining Gaussian pyramid decomposition and Laplace reconstruction, and is used to cope with the segmentation requirements of complex cases.

[0072] Specifically, in the feedback iteration process, the performance of the segmentation window is continuously monitored and a historical data set is constructed. When it is detected that a specific window has insufficient performance improvement in consecutive iterations, the fluctuation coefficient analysis mechanism is triggered. For windows with excessive fluctuations, re-segmentation operation is preferred; when repeated optimization is ineffective, the window optimization mechanism based on the expected utility score is started. The expected utility score balances the average accuracy, fluctuation amplitude and data support degree to intelligently select the most reliable segmentation scale. When all candidate windows perform poorly, the system automatically falls back to multi-scale pyramid segmentation, reconstructs the segmentation basis by hierarchical feature fusion, and thus breaks out of the local optimal dilemma.

[0073] In the preset number of consecutive feedback iterations, if all of , represents the preset score continuous fluctuation threshold, and represents the ratio of the score change rate to the measurement score of the last time, which is still above the threshold with the number of feedback iterations. In this case, the score fluctuation coefficient is introduced to evaluate the window stability. If the current score fluctuation coefficient is greater than the fluctuation coefficient threshold, it indicates that the current window size is sensitive to image content, i.e. unstable response to noise or artifacts, and needs to switch to a more robust size and re-execute S1 to S4.

[0074] The expected utility score comprehensively considers the average performance, stability and data quantity, for example, when the utility score of the 160x160 window is 0.15, it will be about 18% lower than that of the 128x128 window with =0.05.

[0075] ​When all the history windows are invalid, the system ensures multi-scale segmentation through the pyramid segmentation fallback strategy, and covers the possible missed tumor area, such as the tiny paravascular lesion, through different size image blocks.

[0076] The segmentation window adjustment in the specific example of glioma detection is as follows:

[0077] The iteration uses the window 64*64, the metric score sequence is 0.68, 0.69, 0.70, and the change rate is 0.014, 0.014.

[0078] In the fourth iteration, the change rate is lower than 0.015 for three times in succession, and the preset =0.01.

[0079] The =0.75, =0.06, and the utility score =0.71 of the 128*128 window in the history data, and the =0.65 of the 256*256 window. Therefore, the 128*128 window is selected.

[0080] After switching, the score of the fifth iteration is improved to 0.74, and the fluctuation coefficient is reduced to 0.05.

[0081] The scheme dynamically adapts the segmentation window to the image features through data association and quantitative threshold, and improves the robustness of tumor detection.

[0082] The traditional method uses a fixed segmentation window size, which cannot be dynamically adjusted according to the characteristics of the case, resulting in insufficient adaptability to heterogeneous tumors or complex boundaries. The scheme realizes intelligent optimization selection of segmentation scale by establishing a dynamic feedback mechanism of segmentation performance, and simultaneously introduces pyramid segmentation as a bottom line strategy, effectively solving the generalization ability limitation problem caused by fixed parameters.

[0083] Further, S5 further includes an image quality perception adjustment step: obtaining quality evaluation data of the MRI segmentation image and a preset adjustment sensitivity coefficient from a medical image database; the quality evaluation data includes one or more quantitative indicators of noise level, contrast, definition and signal-to-noise ratio of the image; based on the quality evaluation data, the image quality coefficient is calculated through the preset image quality evaluation algorithm; according to the image quality coefficient, the fluctuation coefficient threshold is dynamically adjusted, and the dynamic adjustment is as follows: , wherein, is the adjusted fluctuation coefficient threshold, represents the fluctuation coefficient threshold before adjustment, is a preset adjustment sensitivity coefficient, The image quality coefficient Q is used to represent the image quality; the dynamic adjusted fluctuation coefficient threshold is used as a new fluctuation coefficient threshold to replace the original one in the multi-scale segmentation negative feedback step.

[0084] In this embodiment, the quality evaluation data refers to a set of data obtained by objectively evaluating the MRI segmentation image through quantitative indicators. Specifically, image processing algorithms can be used to measure parameters such as noise level, contrast, definition, and signal-to-noise ratio to achieve this purpose. These parameters collectively reflect the feature distinguishability of the image.

[0085] The image quality coefficient Q, where 0≤Q≤1, the larger the value, the higher the image quality.

[0086] The image quality coefficient refers to a numerical indicator for comprehensive evaluation of image quality. Specifically, it can be achieved by inputting various quality evaluation parameters into a pre-set weighted calculation model. This coefficient is used to quantify the availability of images in tumor detection tasks.

[0087] The adjustment sensitivity coefficient is a pre-set parameter that controls the adjustment range of the threshold. It can be obtained through experience or optimized through historical data, and is used to balance the relationship between image quality fluctuation and segmentation window adjustment frequency.

[0088] In the feedback iteration process, when it is necessary to determine whether to adjust the segmentation window, first obtain the quality evaluation data of the current MRI segmentation image, and calculate the coefficient reflecting the overall quality of the image through a pre-set algorithm. According to this coefficient, the fluctuation coefficient threshold is dynamically adjusted. When the image quality is low, the fluctuation threshold is appropriately relaxed to avoid unnecessary window adjustment due to image quality problems; when the image quality is high, a stricter threshold is used to ensure that the segmentation window optimization can effectively improve the detection accuracy.

[0089] The adjusted threshold directly participates in the decision-making process of the segmentation window size, so that the window adjustment strategy can adapt to different quality input images, reducing invalid adjustment operations while ensuring the accuracy of tumor region segmentation.

[0090] Traditional methods usually use a fixed threshold when adjusting the segmentation window, which cannot adapt to the characteristic differences of different quality image data, and is prone to over-adjustment on low-quality images or insufficient adjustment on high-quality images. This scheme introduces an image quality perception mechanism, combines objective quality evaluation results with dynamic threshold adjustment, and thus improves the stability and reliability of the brain tumor detection system in complex clinical environments.

[0091] Further, patient metadata is acquired and processed with a preset image synthesis model to generate multi-modal MRI segmentation images with MRI segmentation images, the specific process being: acquiring patient metadata from a brain tumor detection database; converting the patient metadata into a fixed-dimension metadata feature vector through a metadata encoder; fusing the MRI segmentation images and the metadata feature vector through a feature fusion algorithm to generate a fusion condition vector; after splicing the fusion condition vector and a random noise vector, inputting the generator of the preset image synthesis model based on the conditional generative adversarial network to generate a multi-modal MRI segmentation image containing simulated tumor features.

[0092] In the present embodiment, the patient metadata refers to clinical attribute data related to brain tumor detection, and specifically can include one or more of age, pathological type, treatment stage, and radiotherapy dose history, which are stored and managed through a structured database.

[0093] The metadata encoder refers to a processing module for converting unstructured or semi-structured metadata into a numerical vector, and specifically can be implemented by a multilayer perception or a bidirectional long short-term memory network, which functions to convert heterogeneous clinical data into computable feature expressions.

[0094] The feature fusion module refers to a neural network component that realizes the interaction of image features and metadata features, and specifically can be implemented by a convolution layer after channel splicing or a cross-attention mechanism, which is used to establish the associated mapping of image features and clinical attributes.

[0095] The generator of the conditional generative adversarial network refers to a deep neural network capable of generating synthetic images according to the input condition vector, and specifically can be constructed by combining a U-Net architecture with a residual connection, which functions to convert the fused clinical-image features into synthetic images with pathological relevance.

[0096] The encoder extracts the latent features of the patient metadata, converting discrete clinical information into computable expressions in a continuous vector space. The feature fusion module interacts with image features and metadata features across modalities, enabling the generator to understand the association between the clinical background and image features of a specific patient.

[0097] The introduction of the random noise vector provides diversity assurance for the generation process, avoiding mode collapse.

[0098] The adversarial training mechanism of the generative adversarial network ensures that the synthetic images have tumor feature performances comparable to real cases while maintaining anatomical reasonableness. This generation process effectively controls the clinical relevance of synthetic images through conditional constraint mechanisms while ensuring data augmentation effects.

[0099] The metadata encoder can process non-sequence type metadata based on a transformer structure based on a self-attention mechanism, for example, for mixed data containing radiotherapy dose distribution and pathological typing, the cross-field association is captured through a multi-head attention mechanism. The feature fusion module can optionally use a gated attention mechanism to dynamically adjust the contribution ratio of image features and metadata features. The generator can use a progressive growth architecture, first generating a low-resolution image block, and then gradually refining to the target size, for example, from 64x64 to 256x256 resolution.

[0100] The discriminator adopts a multi-scale architecture, respectively discriminates the generated image at the original resolution, 1 / 2 resolution and 1 / 4 resolution. This multi-scale discrimination strategy can consider both local details and global structure of the image. In the training process, the generator and the discriminator are alternately optimized through the adversarial learning process, constantly improving the quality and diversity of the generated image.

[0101] Traditional methods often use geometric transformation or simple image mixing techniques for data augmentation, lacking the ability to model individual patient characteristics. Existing generative adversarial networks usually only take images as input conditions, and cannot effectively fuse structured clinical data.

[0102] Through the above technical solutions, multi-modal MRI segmentation images highly related to patient clinical characteristics can be generated, effectively improving the clinical value of synthetic data. The scheme enhances the pathological reasonableness of the generated image through the encoding fusion of structured metadata, providing more representative training samples for subsequent tumor detection models. At the same time, the conditional generation mechanism avoids the semantic distortion problem that may be introduced by traditional data augmentation methods, ensuring that the synthetic image conforms to medical prior knowledge in terms of tumor morphology, location and surrounding tissue relationship.

[0103] Further, the output brain tumor detection result is immersed and visualized through VR virtual reality technology, the specific process includes:

[0104] Based on the MRI segmentation image, the multi-modal MRI segmentation image, and the tumor heat map generated after feedback iteration, a three-dimensional grid model of brain tissue is generated through a three-dimensional reconstruction algorithm. At the same time, the tumor detection result is converted into three-dimensional volume data with different color coding and transparency, among which the active tumor area is highlighted in red, the tumor necrosis area is displayed in gray, and the edema zone is covered with semi-transparent blue, to intuitively distinguish different tissue types.

[0105] The generated three-dimensional model and tumor volume data are imported into the VR visualization platform, rendered through the VR headset, and the model details are optimized using real-time rendering technology to ensure high frame rate and low latency in the VR environment. At the same time, dynamic lighting and shadow effects are supported to enhance the depth perception of anatomical structures.

[0106] In the VR environment, the user can interact through the handle controller, including rotating, scaling, translating the brain model, and switching the view of different modalities, integrating patient metadata and treatment response consistency scores, and automatically displaying relevant clinical information in the VR space when the user focuses on a specific tumor area to assist decision-making.

[0107] Multiple users can access the VR environment simultaneously to realize remote collaborative consultation. Each user can mark suspicious areas, add annotations, and discuss detection results in real time in the shared virtual space. The system records all interactions and generates visual reports for subsequent review or training.

[0108] Through the VR visualization step, abstract image data is converted into intuitive and interactive three-dimensional scenes, significantly improving the interpretation efficiency and clinical applicability of brain tumor detection results, especially for preoperative planning and postoperative evaluation of complex cases.

[0109] As shown in Figure 5 the structure diagram of a brain tumor detection system based on attention mechanism and MRI multi-modal fusion provided by the embodiment of the application, comprising: a first tumor heat map processing module: for acquiring MRI original images and processing them according to a preset segmentation method to obtain MRI segmentation images and generate a first tumor heat map through a preset attention detection model; a second tumor heat map processing module: for acquiring patient metadata and processing them with the MRI segmentation images through a preset image synthesis model to generate multi-modal MRI segmentation images, and generating a second tumor heat map through a preset attention detection model; a calculation module: for calculating the consistency measure of the first tumor heat map and the second tumor heat map to obtain a measure score; a judgment module: for judging whether the measure score is greater than a preset measure threshold, yes to enter the output module, otherwise to execute the following feedback iteration; a feedback iteration module: for specific feedback iteration as follows: merging the MRI segmentation images and the multi-modal MRI segmentation images to generate an enhanced training set, and inputting the enhanced training set into the preset attention detection model for training, in the training process, obtaining a quality weight parameter after normalizing the measure score and weighting the loss term corresponding to the multi-modal MRI segmentation images in the preset attention detection model according to the quality weight parameter, updating the loss term according to the weighting operation result, re-executing the first tumor heat map processing module to the judgment module and recording the number of feedback iterations, if the number of feedback iterations reaches a preset upper limit or the improvement amplitude of the measure score in a preset consecutive number of iterations is less than a preset threshold, then exit the feedback iteration; an output module: for outputting the brain tumor detection result according to the preset attention detection model after feedback iteration.

[0110] Those skilled in the art will appreciate that embodiments of the present application can be devised for a variety of applications. It is therefore intended that the present application cover all such modifications and variations of the application disclosed herein provided they come within the scope of the appended claims and their equivalents. It is intended to

[0111] The present application is described in reference to the drawings using a flowchart and / or a block diagram of the method, apparatus (system) and computer program product according to embodiments of the application. It will be understood that each block of the flowchart and / or block diagram, and combinations of blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processing device or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.

[0112] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.

[0113] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.

[0114] While the preferred embodiments of the application have been described, additional variations and modifications can be employed by those skilled in the art. Therefore, the appended claims are intended to cover all such modifications and variations as fall within the scope of the present application.

[0115] Obviously, a person skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application belong to the scope of the claims of the present application and their equivalent technologies, the present application also intends to include these modifications and variations.

Claims

1. A brain tumor detection method based on attention mechanisms and MRI multimodal fusion, characterized in that, Includes the following steps: S1. Acquire the original MRI image and process it according to the preset segmentation method to obtain the MRI segmented image and generate the first tumor heatmap by processing it through the preset attention detection model; S2. Obtain patient metadata and process it with the MRI segmentation image through a preset image synthesis model to generate a multimodal MRI segmentation image. Process the multimodal MRI segmentation image through a preset attention detection model to generate a second tumor heatmap. S3. Calculate the consistency measure between the first tumor heatmap and the second tumor heatmap, and obtain the measure score; S4. Determine whether the measurement score is greater than the preset measurement threshold. If yes, proceed to step S6; otherwise, execute the following feedback iteration. S5. The specific feedback iterations are as follows: The MRI segmentation image and the multimodal MRI segmentation image are merged to generate an enhanced training set, which is then input into a preset attention detection model for training. During the training process, the metric score is normalized to obtain a quality weight parameter, which is then used to perform a weighted operation on the loss term corresponding to the multimodal MRI segmentation image in the preset attention detection model. The loss term is updated based on the weighted operation result, and S1 to S4 are re-executed while recording the number of feedback iterations. If the number of feedback iterations reaches a preset upper limit or the improvement of the metric score in a preset number of consecutive iterations is less than a preset threshold, the feedback iteration is exited. S5 further includes a dynamic correction step for the mass weight parameter: Preset baseline adjustment coefficients were extracted from the brain tumor detection database; The treatment effect region is identified and segmented in the multimodal MRI segmentation image by a preset recognition and segmentation model, and the treatment effect region is output. The treatment effect region includes images of tumor necrosis area, edema zone and active tumor area. The preset recognition and segmentation model includes a first preset recognition and segmentation model, a second preset recognition and segmentation model and a third preset recognition and segmentation model. Calculate the spatial overlap between the treatment effect region and the original tumor region in the multimodal MRI segmentation image to generate a treatment response consistency score; Based on the treatment response consistency score, the quality weight parameter is dynamically corrected using the following formula: ; in, This represents the corrected mass weight parameter. This represents the mass weight parameter before correction. Indicates the consistency score of treatment response. The preset basic adjustment coefficient; The corrected quality weight parameters will be used in the next feedback iteration; S6. Process and output brain tumor detection results based on the preset attention detection model after feedback iteration.

2. The brain tumor detection method based on attention mechanism and MRI multimodal fusion according to claim 1, characterized in that, The identification and segmentation of the therapeutic effect region specifically includes: The radiotherapy dose distribution map was extracted from the patient's treatment history data from the brain tumor detection database. The radiotherapy dose distribution map was consistent with the image size of the multimodal MRI segmentation image. The treatment effect region is obtained by matching the radiotherapy dose distribution map with the multimodal MRI segmentation image. In the treatment effect region where the radiotherapy dose extracted from the radiotherapy dose distribution map is higher than the preset radiotherapy dose threshold, the first preset recognition and segmentation model is used to segment and identify the necrotic core region and the tumor necrosis area image. The edema zone around the necrotic core area is segmented and identified using a second preset recognition and segmentation model to obtain an image of the edema zone. The therapeutic effect region identified by dynamic contrast enhancement is segmented and identified by a third preset recognition and segmentation model to obtain an image of the active tumor region.

3. The brain tumor detection method based on attention mechanism and MRI multimodal fusion according to claim 1, characterized in that, The calculation of the treatment response consistency score specifically includes: The Dice coefficients of the tumor necrosis area image, edema zone image, and active tumor area image with the original MRI image were calculated separately, and the weighted summation was used to obtain the response consistency coefficient. The number of treatment days and the preset tissue repair time constant were extracted from the patient's treatment history data from the brain tumor detection database. The specific formula for the treatment response consistency score is as follows: , Represents the response consistency coefficient, where, ,in, This indicates the calculation of the decay factor during treatment time. For the number of days of treatment, This is the preset tissue repair time constant.

4. The brain tumor detection method based on attention mechanism and MRI multimodal fusion according to claim 1, characterized in that, The specific process of weighting the loss terms corresponding to the multimodal MRI segmentation images in the preset attention detection model is as follows: Obtain the metric score and preset normalization parameters under the current feedback iteration from the brain tumor detection database; The metric score is normalized using preset normalization parameters to generate quality weight parameters; The loss terms corresponding to the multimodal MRI segmentation images are weighted and calculated based on the quality weight parameters to generate weighted loss terms. Update the model parameters corresponding to the preset attention detection model based on the weighted loss term.

5. The brain tumor detection method based on attention mechanisms and MRI multimodal fusion according to claim 1, characterized in that, The specific process of acquiring the original MRI image and segmenting it according to a predefined segmentation method is as follows: The original MRI images were segmented in parallel using segmentation windows of different sizes; For each parallel segmented raw MRI image, a different original first tumor thermogram is generated, and its structural similarity to a preset baseline thermogram is calculated. The segmentation window with the highest similarity is selected as the predefined segmentation method, and the original MRI image obtained by parallel segmentation is used as the MRI segmentation image.

6. The brain tumor detection method based on attention mechanism and MRI multimodal fusion according to claim 5, characterized in that, S5 further includes a multi-scale segmentation negative feedback step: A historical dataset is constructed based on the segmentation window size, metric score, and score change rate recorded in the historical feedback iteration. Calculate the average metric score and score fluctuation coefficient for each segmentation window size; If, in a preset number of consecutive feedback iterations, the ratio of the rate of change of the score to the measured score of the previous feedback iteration is greater than the preset continuous fluctuation threshold of the score, and the score fluctuation coefficient is greater than the fluctuation coefficient threshold, then S1 to S4 are executed again. If the number of re-executions exceeds the preset limit, the expected utility score of each segmentation window size is calculated, and the segmentation window size with the highest expected utility score is selected as the new segmentation window size. Sort by expected utility scores, and calculate the difference between the largest and smallest expected utility scores to obtain the expected utility score difference value. If the expected utility score difference value is less than the preset differentiation threshold and all expected utility scores are lower than the preset backoff threshold, then backoff to the original MRI image, use the preset pyramid segmentation strategy to generate multi-scale image blocks, re-execute S1 to S4 and reset the number of re-executions.

7. The brain tumor detection method based on attention mechanism and MRI multimodal fusion according to claim 6, characterized in that, S5 further includes an image quality perception adjustment step: Obtain quality assessment data and preset adjustment sensitivity coefficients for MRI segmentation images from medical imaging databases; The quality assessment data includes one or more quantitative indicators among the image's noise level, contrast, sharpness, and signal-to-noise ratio. Based on the quality assessment data, the image quality coefficient is calculated using a preset image quality assessment algorithm; Based on the image quality coefficient, the fluctuation coefficient threshold is dynamically adjusted as follows: ,in, The adjusted volatility coefficient threshold. This represents the threshold value for the volatility coefficient before adjustment. The preset sensitivity adjustment coefficient, Indicates the image quality coefficient; The dynamically adjusted volatility coefficient threshold is used as the new volatility coefficient threshold and substituted into the multi-scale segmentation negative feedback step.

8. The brain tumor detection method based on attention mechanism and MRI multimodal fusion according to claim 1, characterized in that, The specific process of obtaining patient metadata and processing it with MRI segmentation images using a preset image synthesis model to generate multimodal MRI segmentation images is as follows: Obtain the patient's metadata from the brain tumor detection database; The patient metadata is converted into a fixed-dimensional metadata feature vector using a metadata encoder; The MRI segmentation image and the metadata feature vector are fused using a feature fusion algorithm to generate a fusion condition vector; The fusion condition vector is concatenated with a random noise vector and then input into the generator of the preset image synthesis model based on a conditional generative adversarial network to generate the multimodal MRI segmentation image containing simulated tumor features.

9. A brain tumor detection system based on attention mechanisms and MRI multimodal fusion, characterized in that, include: First tumor heatmap processing module: used to acquire the original MRI image and process it according to the preset segmentation method to obtain the MRI segmented image and generate the first tumor heatmap through the preset attention detection model; The second tumor heatmap processing module is used to acquire patient metadata and process it with MRI segmentation images through a preset image synthesis model to generate multimodal MRI segmentation images. The multimodal MRI segmentation images are then processed through a preset attention detection model to generate a second tumor heatmap. Calculation module: Used to calculate the consistency metric between the first tumor heatmap and the second tumor heatmap, and obtain the metric score; Judgment module: Used to determine whether the measurement score is greater than the preset measurement threshold. If yes, it enters the output module; otherwise, it executes the following feedback iteration. Feedback Iteration Module: Used to perform feedback iteration: merge MRI segmentation images with multimodal MRI segmentation images to generate an enhanced training set, perform weighted training on a preset attention detection model, and exit feedback iteration according to the iteration termination condition; The feedback iteration module also includes a dynamic correction unit for the quality weight parameter: Preset baseline adjustment coefficients were extracted from the brain tumor detection database; The treatment effect region is identified and segmented in the multimodal MRI segmentation image by a preset recognition and segmentation model, and the treatment effect region is output. The treatment effect region includes images of tumor necrosis area, edema zone and active tumor area. The preset recognition and segmentation model includes a first preset recognition and segmentation model, a second preset recognition and segmentation model and a third preset recognition and segmentation model. Calculate the spatial overlap between the treatment effect region and the original tumor region in the multimodal MRI segmentation image to generate a treatment response consistency score; Based on the treatment response consistency score, the quality weight parameter is dynamically corrected using the following formula: ; in, This represents the corrected mass weight parameter. This represents the mass weight parameter before correction. Indicates the consistency score of treatment response. The preset basic adjustment coefficient; The corrected quality weight parameters will be used in the next feedback iteration; Output module: Used to process and output brain tumor detection results based on the preset attention detection model after feedback iteration.

Images