Multimodal nuclear medicine image segmentation method, system and storage medium

Abstract

The present invention relates to the field of image segmentation, and in particular to a multimodal nuclear medicine image segmentation method, system and storage medium. The method comprises: classifying the image type of multimodal medical examination images, and locating the lesion area, respectively CT tomographic lesion images and PET lesion images; predicting the lesion geometric mutation morphology of the CT tomographic lesion images, and performing multimodal image fusion boundary correction on the CT tomographic lesion images and the PET lesion images to obtain multimodal lesion image fusion boundary correction data; performing multimodal lesion area segmentation combination optimization on the CT tomographic scan images and the PET images according to the multimodal lesion image fusion boundary correction data to obtain a lesion area segmentation combination optimization strategy; constructing an image segmentation model for the lesion area segmentation combination optimization strategy based on a policy gradient algorithm to obtain a lesion image segmentation model. The present invention makes the image segmentation technology more perfect by optimizing the image segmentation technology.

Description

Technical Field

[0001] The present invention relates to the field of image segmentation, and in particular to a multimodal nuclear medicine image segmentation method, system and storage medium. Background Art

[0002] Nuclear medicine imaging techniques, such as positron emission tomography (PET) and single photon emission computed tomography (SPECT), can provide unique metabolic and functional information and are widely used in oncology, neuroscience, cardiovascular disease and other fields. Due to the limitations of single imaging modality in terms of spatial resolution, contrast, noise characteristics, etc., multimodal image fusion has become the key to solving the problem. Combining PET or SPECT with anatomical images such as computed tomography (CT) and magnetic resonance imaging (MRI) can not only improve the diagnostic value of the image, but also more accurately locate the lesion and quantify the metabolic characteristics. However, the complexity, heterogeneity and nonlinear relationship between modalities of multimodal imaging data have brought great challenges to automated segmentation. In order to cope with these challenges, technologies such as multimodal feature alignment, multi-layer information fusion, and cross-modal attention mechanism have been proposed to further optimize the segmentation performance. For example, the geometric mismatch between modalities can be corrected through a specific preprocessing process, or the association between modalities can be dynamically captured using a multimodal collaborative learning strategy. In addition, strategies such as transfer learning and weakly supervised learning have been introduced to improve the generalization ability of the model. Nevertheless, in practical applications, multimodal image segmentation methods still need to find a balance between algorithm optimization and clinical validation to ensure the reliability and clinical usability of the technology. However, a traditional multimodal nuclear medicine image segmentation method has the problem of blurred boundaries caused by inconsistencies between multimodal data and the problem of accurate matching of complex lesion morphology and metabolic characteristics. Summary of the invention

[0003] Based on this, it is necessary to provide a multimodal nuclear medicine image segmentation method, system and storage medium to solve at least one of the above technical problems.

[0004] To achieve the above object, a multimodal nuclear medicine image segmentation method is provided, the method comprising the following steps:

[0005] Step S1: acquiring a multimodal medical examination image; classifying the multimodal medical examination image by image type to obtain a CT tomography image and a PET image respectively;

[0006] Step S2: locating the lesion area on the CT tomography image and the PET image, respectively, to obtain a CT tomography lesion image and a PET lesion image; predicting the lesion geometric mutation morphology on the CT tomography lesion image to obtain lesion geometric mutation morphology prediction data; performing multi-modal image fusion boundary correction on the CT tomography lesion image and the PET lesion image based on the lesion geometric mutation morphology prediction data to obtain multi-modal lesion image fusion boundary correction data;

[0007] Step S3: performing multimodal lesion region segmentation combination optimization on the CT tomography image and the PET image according to the multimodal lesion image fusion boundary correction data to obtain a lesion region segmentation combination optimization strategy;

[0008] Step S4: construct an image segmentation model based on the policy gradient algorithm for the combined optimization strategy of lesion area segmentation, obtain the lesion image segmentation model, and send the lesion image segmentation model to the cloud platform to execute the multimodal nuclear medicine image segmentation method.

[0009] The present invention first needs to collect multiple types of medical examination images, including CT tomography images and PET images. By automatically classifying the types of these images, the system can identify different types of medical images and provide basic data for subsequent analysis. This classification method not only improves work efficiency, but also provides targeted support for different image processing strategies and algorithms. The accuracy of classification is a prerequisite for ensuring the smooth progress of subsequent steps. CT and PET images are respectively used to locate the lesion area and extract potential lesion areas (i.e., lesion areas). Then, for the lesion area in the CT image, the morphological changes of the lesion are predicted by analyzing its geometric mutation morphology. This process helps to deeply understand the characteristics, development trends and relationship of the lesion with other tissues, and provides a more accurate basis for subsequent lesion analysis and processing. Finally, based on the data predicted by geometric morphology, the multimodal fusion boundary correction of CT and PET images is performed to optimize the fusion effect of the image, so that the images of different modes are more consistent in the lesion area. By analyzing the fusion boundary correction data of the multimodal lesion image, the system will combine the lesion area information in the CT and PET images to perform segmentation combination optimization of the multimodal lesion area. The core goal of this optimization strategy is to improve the accuracy and robustness of lesion region segmentation, especially when dealing with complex or fuzzy boundary areas. By fusing information from different modalities, the optimized lesion region can better reflect the true condition of the disease and provide more accurate image data support for further diagnosis. The policy gradient algorithm is used to construct the lesion region segmentation model. The policy gradient algorithm is a reinforcement learning method that can achieve better segmentation results in complex lesion region segmentation tasks by continuously optimizing the segmentation strategy. The model can not only process images of different modalities, but also maintain efficiency and accuracy when dealing with lesions with irregular shapes and fuzzy boundaries. Finally, the constructed segmentation model will be uploaded to the cloud platform for online execution and further automated processing. This process effectively improves the automation of image segmentation, reduces manual intervention, and improves the efficiency and accuracy of diagnosis. Therefore, the present invention is an optimization processing of a traditional multimodal nuclear medicine image segmentation method, which solves the problem of boundary fuzziness caused by inconsistency between multimodal data and the problem of accurate matching of complex lesion morphology and metabolic characteristics in the traditional multimodal nuclear medicine image segmentation method, reduces the boundary fuzziness caused by inconsistency between multimodal data, and improves the accurate matching of complex lesion morphology and metabolic characteristics.

[0010] Preferably, step S1 comprises the following steps:

[0011] Step S11: Acquire multimodal medical examination images;

[0012] Step S12: performing image grayscale optimization on the multimodal medical examination image to obtain a multimodal medical examination grayscale optimized image;

[0013] Step S13: spatially aligning the multimodal medical examination grayscale optimized images to obtain a medical spatially aligned optimized image;

[0014] Step S14: classify the medical space alignment optimization image into image types to obtain a CT tomography image and a PET image respectively.

[0015] The present invention first collects multimodal medical examination images, which usually include CT (computed tomography) images and PET (positron emission tomography) images. Through this process, it is possible to ensure that image data from different sources and technologies are obtained, covering a wider range of lesion information. Multimodal images provide biological information of different angles and properties, which helps to improve the comprehensiveness and accuracy of disease diagnosis. These images will serve as basic data for subsequent analysis and processing to ensure the effectiveness of multimodal medical image fusion. Grayscale optimization is used to improve the quality of images and enhance the detail performance of images. Multimodal medical images are affected by factors such as noise and insufficient contrast during the acquisition process, resulting in unsatisfactory image quality. By optimizing the grayscale of the image, the brightness, contrast and clarity of the image can be improved, making the lesion area more obvious, helping doctors and automated systems to more accurately identify the lesion area. In addition, grayscale optimization can also reduce the impact of noise and improve the reliability and availability of images in subsequent processing. Multimodal medical examination images after grayscale optimization need to be spatially aligned to ensure the consistency of the spatial positions of images of different modalities. Since CT and PET images are obtained under different scanning technologies, spatial deviations will occur. Through spatial alignment technology, the same anatomical area of ​​CT and PET images can be accurately overlapped, so that images from two different sources can be effectively fused and compared. This step is the key to multimodal image fusion and lays a solid foundation for subsequent analysis and diagnosis. Medical images that have been optimized through spatial alignment will be classified according to different image types, usually including CT tomography images and PET images. The key to this step is to identify the types of different images so that appropriate analysis methods can be used in subsequent image processing. Through image type classification, images of different modalities can be processed separately to ensure that each type of image is optimized in a targeted manner, thereby improving the accuracy and reliability of the final analysis results. This classification process provides a clear basis for subsequent lesion localization and image fusion.

[0016] Preferably, step S2 comprises the following steps:

[0017] Step S21: locating the lesion area on the CT tomography image and the PET image, respectively CT tomography lesion image and PET lesion image;

[0018] Step S22: predicting the geometric mutation morphology of the lesion on the CT section lesion image to obtain the prediction data of the geometric mutation morphology of the lesion;

[0019] Step S23: evaluating the metabolic activity of the lesion mutation region on the PET lesion image according to the lesion geometric mutation morphology prediction data to obtain metabolic activity data of the lesion mutation region;

[0020] Step S24: performing multimodal image fusion boundary correction on the CT tomographic lesion image and the PET lesion image based on the metabolic activity data of the lesion mutation area and the lesion geometric mutation morphology prediction data to obtain multimodal lesion image fusion boundary correction data.

[0021] The CT tomographic scan image and PET image of the present invention are respectively positioned in the lesion area to determine the lesion area (i.e., lesion) present in the image. CT images can provide structural information and reveal the anatomical morphology of organs and tissues, while PET images can reflect metabolic and functional activities. This step lays the foundation for subsequent lesion analysis, morphological prediction, and metabolic evaluation by extracting the lesion area from these two different types of images. Accurate lesion area positioning provides necessary support for accurate diagnosis and treatment of diseases. For CT tomographic lesion images, the geometric mutation characteristics of the lesion are analyzed using a geometric morphological prediction model. This prediction process mainly focuses on the morphological changes of the lesion, such as changes in size, shape, and boundaries, to identify the growth trend or deterioration of the lesion area. CT images, due to their superior anatomical imaging capabilities, can clearly display the geometric characteristics of the lesion in space. Therefore, the prediction of lesion geometric mutation morphology has important clinical value for monitoring lesion progression and evaluating treatment effects, and provides strong support for subsequent image fusion and lesion area analysis. Based on the lesion geometric mutation morphology prediction data obtained in the previous step, the lesion mutation area in the PET image is evaluated for metabolic activity. PET images reflect the metabolic state of the lesion area, and metabolic activity assessment helps to understand whether the lesion area is active or proliferating. By combining the geometric morphological change information of CT images, the changes in the metabolic activity of the lesion can be analyzed more accurately, thereby further evaluating the progression of the disease and the response to treatment. This step helps to improve the comprehensive understanding of the lesion mutation area and provide more accurate support for clinical decision-making. Combining the metabolic activity data of the lesion mutation area and the lesion geometric mutation morphological prediction data, multimodal image fusion and boundary correction of CT tomographic lesion images and PET lesion images are performed. By fusing these two different modal images, not only can the image quality be improved, but also the boundary inconsistency and morphological deviation in the image can be corrected, thereby obtaining a more accurate representation of the lesion area. This fusion and boundary correction process can make up for the limitations of single-modality images, such as density inconsistency in CT images and fuzzy boundaries in PET images, making the final image clearer and more accurate, and providing high-quality imaging data support for subsequent lesion analysis and clinical applications.

[0022] Preferably, step S22 includes the following steps:

[0023] Step S221: extracting the lesion contour from the CT section lesion image to obtain lesion contour data;

[0024] Step S222: performing three-dimensional grid division on the CT tomographic lesion image according to the lesion contour data to obtain three-dimensional lesion grid data;

[0025] Step S223: Calculating the boundary curvature distribution of the lesion contour data based on the lesion three-dimensional grid data to obtain lesion boundary distribution curvature data;

[0026] Step S224: Calculate the distribution curvature approximate mean difference of the lesion boundary distribution curvature data to obtain the distribution curvature approximate mean difference data;

[0027] Step S225: performing lesion geometric mutation point evolution deduction on the lesion three-dimensional grid data according to the distribution curvature adjacent mean difference data to obtain lesion geometric mutation point evolution data;

[0028] Step S226: Predict the lesion geometric mutation morphology based on the distribution curvature proximity mean difference data and the lesion geometric mutation point evolution data to obtain lesion geometric mutation morphology prediction data.

[0029] The present invention first extracts the lesion contour from the CT tomographic lesion image. This process is a key image preprocessing step, which is intended to accurately extract the external contour of the lesion from the CT image. The extraction of the lesion contour can effectively identify the boundary of the lesion area, remove irrelevant background information, and provide clear boundary data for subsequent geometric morphological analysis. Through this process, the system can clearly define the position and shape of the lesion, laying the foundation for further geometric analysis and morphological prediction of the lesion. Based on the extracted lesion contour data, the CT tomographic lesion image is subjected to three-dimensional grid division. The purpose of this operation is to perform fine grid decomposition of the lesion area in three-dimensional space, so that the geometric characteristics of the lesion can be analyzed more accurately. The three-dimensional grid division makes the distribution of the lesion in space clearer, and provides more detailed spatial structure data for subsequent geometric analysis and morphological prediction. Through this three-dimensional spatial representation, the accuracy of lesion analysis can be improved, which is particularly important when processing lesions with complex morphology. Using the lesion three-dimensional grid data, the curvature distribution of the boundary of the lesion contour is further calculated. This calculation process is intended to evaluate the curvature change of the lesion boundary and capture the subtle changes of the lesion edge. For example, information such as the degree of curvature and convex-concave changes of the lesion boundary can provide important clues for the morphological evolution of the lesion. By calculating the curvature distribution, the mutation points and abnormal changes of the lesion boundary can be identified, and the future morphological changes of the lesion can be predicted. This step is crucial for accurately describing the lesion morphology, especially for evaluating the progression of lesion areas such as malignant tumors. The distribution curvature data of the lesion boundary are calculated by the adjacent mean difference. The core of this calculation method is to obtain the morphological information of the lesion boundary by measuring the difference between the curvature of the lesion boundary at adjacent points. The distribution curvature adjacent mean difference reflects the smoothness of the lesion boundary, the amplitude of the curvature change, and other characteristics, which can help identify whether the lesion has a mutation or irregular morphology. This step can further improve the accuracy of lesion morphology prediction, especially when the lesion changes are more complex, it helps to more accurately capture the subtle morphological changes of the lesion. Based on the distribution curvature adjacent mean difference data, the evolution of the geometric mutation points of the lesion three-dimensional grid data is further deduced. This deduction process simulates the morphological changes of the lesion at different time points, especially the evolution trend at the lesion geometric mutation points (such as tumor spread, morphological changes, etc.). Through these data deductions, the system can identify the evolution direction and rate of the lesions, providing a basis for predicting the progression of the disease. This not only helps to determine the malignancy of the lesions, but also provides strong support for the formulation of personalized treatment plans. The distribution curvature proximity mean difference data and the lesion geometric mutation point evolution data will be combined to comprehensively predict the geometric mutation morphology of the lesions. This step will provide a detailed prediction of the future morphological changes of the lesions, including the future geometric morphological changes, boundary mutations, and expansion trends of the lesions. This geometric mutation morphological prediction is crucial for the early identification of the disease and the formulation of personalized treatment plans. By accurately predicting the morphological changes of the lesions.

[0030] Preferably, step S24 comprises the following steps:

[0031] Step S241: performing spatial morphological feature decomposition processing on the lesion geometric mutation morphological prediction data to obtain lesion morphological feature decomposition data;

[0032] Step S242: performing metabolic activity correlation analysis on the metabolic activity data of the lesion mutation area according to the lesion morphology feature decomposition data to obtain lesion morphology-metabolic activity correlation data;

[0033] Step S243: performing trend hotspot local enhancement processing on the CT tomographic lesion image and the PET lesion image based on the lesion morphology-metabolic activity correlation data to obtain trend hotspot enhanced image data;

[0034] Step S244: adjusting the center of gravity of the overlapping area of ​​the CT tomographic lesion image and the PET lesion image according to the trend hotspot enhanced image data to obtain the center of gravity adjustment data of the overlapping area;

[0035] Step S245: performing multimodal image fusion boundary correction on the CT tomographic lesion image and the PET lesion image based on the overlapping area gravity center adjustment data to obtain multimodal lesion image fusion boundary correction data.

[0036] The present invention first performs spatial morphological feature decomposition processing on the lesion geometric mutation morphological prediction data. This operation extracts key features in the lesion morphology, such as the geometric structure, morphological mutation and development trend of the lesion, by performing multi-dimensional analysis on the geometric morphology of the lesion. Through this decomposition processing, the changing law of the lesion morphology can be more accurately identified, thereby providing detailed morphological data for subsequent metabolic activity analysis and image processing. This process helps to deeply understand the spatial characteristics of the lesion and provide basic data for the comprehensive evaluation and prediction of the lesion. According to the lesion morphological feature decomposition data obtained in step S241, the metabolic activity data of the lesion mutation area is further analyzed for correlation. This analysis process is intended to reveal the relationship between the lesion morphological change and its metabolic activity. For example, whether the geometric mutation of the lesion is directly related to the increase or decrease of metabolic activity, or whether the morphological change will cause abnormalities in certain metabolic activities. Through this analysis, the correlation data between the lesion morphology and metabolic activity can be obtained, which provides an important basis for accurately evaluating the biological characteristics and disease progression of the lesion. Based on the lesion morphology-metabolic activity correlation data, the trend hotspot local enhancement processing is performed on the CT tomographic lesion image and the PET lesion image. This step transforms the correlation between morphology and metabolic activity into hotspot information in the image data, thereby enhancing the display of important lesion areas in the image. Local enhancement processing can make the lesion mutation area and the metabolically active part more prominent, thereby increasing the doctor's attention to the key lesion area. Through this processing, the visual effect of the image is enhanced, which helps to further improve the recognition accuracy and diagnostic ability of the lesion area. According to the trend hotspot enhanced image data, the center of gravity of the overlapping areas in the CT tomographic lesion image and the PET lesion image is adjusted. The purpose of this step is to accurately align the lesion areas in the CT and PET images, especially those overlapping areas in the two images. By adjusting the image center of gravity, the image deviation caused by different imaging methods or patient positions can be eliminated, so that the location of the lesion is more accurate. This step is crucial for the fine registration of the image and provides accurate data support for the subsequent multimodal image fusion and boundary correction. Based on the overlapping area center of gravity adjustment data obtained in the previous step, the CT tomographic lesion image and the PET lesion image are further corrected for multimodal image fusion. Through this correction, the boundary inconsistency and morphological deviation that occur during the image fusion process can be eliminated, thereby improving the accuracy and consistency of the multimodal image. Especially when the lesion boundary is not clear or there is a difference between the two imaging modes, boundary correction can help provide clearer and more accurate lesion images, further improving the diagnosis and treatment decision support for the lesion area.

[0037] Preferably, step S243 includes the following steps:

[0038] Extracting hot spots of regional metabolic abnormalities from the lesion morphology-metabolic activity correlation data to obtain hot spots of metabolic abnormality regions;

[0039] According to the hotspot metabolic abnormality area data, the CT tomographic lesion image is subjected to local morphological multi-scale enhancement processing to obtain CT multi-scale morphological enhancement data;

[0040] Based on CT multi-scale morphological enhancement data and hotspot metabolic abnormality area data, the PET lesion images are processed with metabolic intensity hotspots to obtain PET metabolic hotspot stratification data;

[0041] According to the PET metabolic hotspot stratification data and CT multi-scale morphological enhancement data, trend hotspot local enhancement processing is performed to obtain trend hotspot enhanced image data.

[0042] The present invention extracts hot spots with abnormal metabolic activity in the lesion area by analyzing the correlation between the lesion morphology and metabolic activity. The core goal of this processing step is to identify the areas with abnormal metabolism in the lesion, which are usually related to the growth or metastasis of the tumor. By extracting the metabolic abnormality hot spot data, the metabolically active or abnormal parts in the lesion can be accurately located, thereby providing accurate regional information for subsequent multimodal image processing. This helps to improve the quantitative analysis and diagnostic accuracy of the lesion, especially when dealing with complex or early lesions, it can better reveal the activity of the lesion. By performing local morphological multi-scale enhancement processing on the CT tomographic lesion image according to the extracted hot spot metabolic abnormality area data. The purpose of this operation is to enhance the areas related to metabolic abnormalities in the CT image, so that the morphological characteristics of the lesion are more obvious. Multi-scale enhancement technology can highlight the details of the lesion at different scales, especially when the boundary between complex lesions or lesions and surrounding tissues is blurred, it helps to improve the recognizability of the lesion edge. Through this processing, it is possible to provide clearer and more detailed lesion information for subsequent image analysis, which is particularly important for the early detection of lesions such as tumors. Based on the multi-scale morphological enhancement data of CT images and the hotspot metabolic abnormality area data, the PET lesion images are hierarchically processed for metabolic intensity hotspots. This hierarchical processing can divide the metabolic hotspots in the PET image into multiple different levels according to the intensity of metabolic activity, so as to better reflect the strength and change trend of metabolic activity in the lesion area. By combining the morphological information of CT images and the metabolic data of PET images, the lesion area can be accurately divided at the spatial and metabolic levels, thereby improving the understanding of the nature of the lesion and the course of the disease. This step makes the metabolic information of PET images more targeted and structured, providing a clear reference for further analysis of the lesion. Based on the PET metabolic hotspot hierarchical data and the CT multi-scale morphological enhancement data, the lesion image is locally enhanced for trend hotspots. The purpose of this step is to highlight those hotspot areas with metabolic abnormalities and morphological change trends in the image, especially those with active metabolism and significant morphological changes. Through local enhancement processing, the details of these hotspot areas in the image are further magnified, thereby improving the visualization effect of the key areas of the lesion. This step helps doctors focus more on the core area of ​​the lesion during diagnosis. Especially when multimodal images are fused, the enhanced image can make the spatial information and metabolic characteristics of the lesion more prominent, thus providing a more accurate basis for subsequent treatment decisions.

[0043] Preferably, step S3 comprises the following steps:

[0044] Step S31: normalizing the multimodal lesion image fusion boundary correction data to obtain lesion image fusion boundary correction normalized data;

[0045] Step S32: performing multimodal lesion region segmentation combination optimization on the CT tomography image and the PET image based on the lesion image fusion boundary correction normalization data to obtain a lesion region segmentation combination optimization strategy.

[0046] The present invention first normalizes the multimodal lesion image data after boundary correction. The main purpose of normalization is to eliminate the differences between different imaging modes (such as CT and PET images) due to different scanning methods, image resolutions or grayscale ranges, so that the scales of all data are consistent, which is convenient for subsequent analysis and processing. This operation can ensure effective comparison and fusion between different types of image data, and improve the accuracy and consistency of image processing. The normalized data makes the lesion image more unified in space, providing a standardized data basis for subsequent lesion area segmentation and analysis, ensuring the data quality and accuracy in the processing process. The normalized lesion image is fused with boundary correction data to perform segmentation combination optimization of multimodal lesion areas on CT tomography images and PET images. This process combines the different characteristics of CT and PET images in the lesion area, and optimizes the segmentation results of the lesion area through the complementarity of multimodal information. This optimization strategy can more accurately identify the boundaries and morphology of the lesion, especially when the boundaries of the lesion area are fuzzy or the similarity of adjacent tissues is high, and can effectively distinguish tumors from healthy tissues, thereby improving the accuracy of diagnosis. Ultimately, the segmentation combination optimization strategy can obtain a more accurate and comprehensive lesion area division, providing more valuable image data for clinical use.

[0047] Preferably, step S32 includes the following steps:

[0048] Step S321: performing boundary gradient direction tensor analysis on the lesion image fusion boundary correction normalization data to obtain multi-dimensional gradient direction tensor field data;

[0049] Step S322: performing inter-modality gradient cross-correlation calculation on the CT tomography image and the PET image based on the multi-dimensional gradient direction tensor field data to obtain modality gradient cross-correlation weight data;

[0050] Step S323: performing multi-modal lesion region segmentation combination optimization on the CT tomography image and the PET image according to the modal gradient cross-correlation weight data to obtain a lesion region segmentation combination optimization strategy.

[0051] The present invention can extract the gradient direction information of the lesion area by performing boundary gradient direction tensor analysis on the normalized data after the correction of the lesion image fusion boundary, and form a multi-dimensional gradient direction tensor field. The analysis captures the detailed features of the lesion edge by calculating the gradient direction and change rate of each point in the image. The tensor field can more accurately describe the changes in structure and texture in the image, especially in the area with blurred boundaries, which helps to accurately identify the spatial distribution and morphological characteristics of the lesion. This analysis provides a basis for subsequent image fusion and segmentation, making the outline of the lesion clearer and more accurate, and provides reliable direction information for subsequent multi-modal image fusion and regional optimization processing. Based on the multi-dimensional gradient direction tensor field data, the inter-modal gradient cross-correlation calculation is performed on the CT tomography image and the PET image. The calculation method can measure the similarity of two different modal (CT and PET) images in the gradient direction, and then generate a gradient cross-correlation weight data between modalities. The beneficial effect of this process is that it can effectively integrate the advantages of CT and PET images by measuring the gradient consistency of different modalities in the image, and overcome the limitations of single modality images in terms of details, contrast or noise. The obtained weight data can be used to further adjust the importance of different modalities in multimodal lesion images, making the final fusion result more accurate and robust, especially in the precise positioning of lesion boundaries and metabolic abnormality areas, providing effective support. According to the modal gradient cross-correlation weight data obtained in the previous step, the multimodal lesion area segmentation combination optimization is performed on the CT tomography image and PET image. This optimization strategy maximizes the complementarity of the two modal images when segmenting the lesion area by combining the weight information of CT and PET images. CT images provide precise structural information, while PET images show the intensity of metabolic activity. The combination of the two can more comprehensively reveal the morphology and functional status of the lesion. Through combined optimization, the error in single modality segmentation can be reduced, and the accuracy and precision of segmentation can be improved. Especially in complex lesion areas or when lesions are similar to normal tissues, the optimization strategy can effectively improve the reliability of the segmentation results, and ultimately provide more accurate and comprehensive information for the diagnosis, analysis and treatment of lesions.

[0052] Preferably, the present invention further provides a multimodal nuclear medicine image segmentation system for executing the multimodal nuclear medicine image segmentation method as described above, the multimodal nuclear medicine image segmentation system comprising:

[0053] A medical image type classification module is used to obtain multimodal medical examination images; image type classification is performed on the multimodal medical examination images to obtain CT tomography images and PET images respectively;

[0054] The fusion boundary correction module is used to locate the lesion area on the CT tomography image and the PET image, respectively, to obtain the CT tomography lesion image and the PET lesion image; to predict the lesion geometric mutation morphology on the CT tomography lesion image to obtain the lesion geometric mutation morphology prediction data; to perform multi-modal image fusion boundary correction on the CT tomography lesion image and the PET lesion image based on the lesion geometric mutation morphology prediction data to obtain the multi-modal lesion image fusion boundary correction data;

[0055] A segmentation combination optimization module is used to perform multi-modal lesion region segmentation combination optimization on CT tomography images and PET images according to multi-modal lesion image fusion boundary correction data to obtain a lesion region segmentation combination optimization strategy;

[0056] The segmentation model construction module is used to construct an image segmentation model based on the policy gradient algorithm for the combined optimization strategy of lesion area segmentation, obtain the lesion image segmentation model, and send the lesion image segmentation model to the cloud platform to execute the multimodal nuclear medicine image segmentation method.

[0057] Preferably, the present invention further provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed, implements the multimodal nuclear medicine image segmentation method as described in any one of the above.

[0058] The beneficial effects of the present invention first require the collection of various types of medical examination images, including CT tomography images and PET images. By automatically classifying the types of these images, the system can identify different types of medical images and provide basic data for subsequent analysis. This classification method not only improves work efficiency, but also provides targeted support for different image processing strategies and algorithms. The accuracy of classification is a prerequisite for ensuring the smooth progress of subsequent steps. CT and PET images are used to locate the lesion area respectively, and potential lesion areas (i.e., lesion areas) are extracted. Then, for the lesion area in the CT image, the morphological changes of the lesion are predicted by analyzing its geometric mutation morphology. This process helps to deeply understand the characteristics, development trends and relationship of the lesion with other tissues, and provides a more accurate basis for subsequent lesion analysis and processing. Finally, based on the data predicted by geometric morphology, the multimodal fusion boundary correction of CT and PET images is performed to optimize the fusion effect of the image, so that the images of different modes are more consistent in the lesion area. By analyzing the fusion boundary correction data of the multimodal lesion image, the system will combine the lesion area information in the CT and PET images to perform segmentation combination optimization of the multimodal lesion area. The core goal of this optimization strategy is to improve the accuracy and robustness of lesion region segmentation, especially when dealing with complex or fuzzy boundary areas. By fusing information from different modalities, the optimized lesion region can better reflect the true condition of the disease and provide more accurate image data support for further diagnosis. The policy gradient algorithm is used to construct the lesion region segmentation model. The policy gradient algorithm is a reinforcement learning method that can achieve better segmentation results in complex lesion region segmentation tasks by continuously optimizing the segmentation strategy. The model can not only process images of different modalities, but also maintain efficiency and accuracy when dealing with lesions with irregular shapes and fuzzy boundaries. Finally, the constructed segmentation model will be uploaded to the cloud platform for online execution and further automated processing. This process effectively improves the automation of image segmentation, reduces manual intervention, and improves the efficiency and accuracy of diagnosis. Therefore, the present invention is an optimization processing of a traditional multimodal nuclear medicine image segmentation method, which solves the problem of boundary fuzziness caused by inconsistency between multimodal data and the problem of accurate matching of complex lesion morphology and metabolic characteristics in the traditional multimodal nuclear medicine image segmentation method, reduces the boundary fuzziness caused by inconsistency between multimodal data, and improves the accurate matching of complex lesion morphology and metabolic characteristics. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 A schematic diagram of the steps of a multimodal nuclear medicine image segmentation method;

[0060] Figure 2 for Figure 1 Detailed implementation steps of step S2 in FIG.

[0061] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0062] The technical method of the present invention is described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by technicians in this field without creative work are within the scope of protection of the present invention.

[0063] In addition, the accompanying drawings are only schematic illustrations of the present invention and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and their repeated description will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities. The functional entities can be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and / or processor methods and / or microcontroller methods.

[0064] It should be understood that, although the terms "first", "second", etc. may be used herein to describe various units, these units should not be limited by these terms. These terms are used only to distinguish one unit from another unit. For example, without departing from the scope of the exemplary embodiments, the first unit may be referred to as the second unit, and similarly the second unit may be referred to as the first unit. The term "and / or" used herein includes any and all combinations of one or more of the listed associated items.

[0065] To achieve this, please refer to Figure 1 to Figure 2 , a multimodal nuclear medicine image segmentation method, the method comprising the following steps:

[0066] Step S1: acquiring a multimodal medical examination image; classifying the multimodal medical examination image by image type to obtain a CT tomography image and a PET image respectively;

[0067] Step S2: locating the lesion area on the CT tomography image and the PET image, respectively, to obtain a CT tomography lesion image and a PET lesion image; predicting the lesion geometric mutation morphology on the CT tomography lesion image to obtain lesion geometric mutation morphology prediction data; performing multi-modal image fusion boundary correction on the CT tomography lesion image and the PET lesion image based on the lesion geometric mutation morphology prediction data to obtain multi-modal lesion image fusion boundary correction data;

[0068] Step S3: performing multimodal lesion region segmentation combination optimization on the CT tomography image and the PET image according to the multimodal lesion image fusion boundary correction data to obtain a lesion region segmentation combination optimization strategy;

[0069] Step S4: construct an image segmentation model based on the policy gradient algorithm for the combined optimization strategy of lesion area segmentation, obtain the lesion image segmentation model, and send the lesion image segmentation model to the cloud platform to execute the multimodal nuclear medicine image segmentation method.

[0070] In the embodiment of the present invention, reference Figure 1 FIG. 1 is a schematic diagram of a multimodal nuclear medicine image segmentation method of the present invention. In this example, the multimodal nuclear medicine image segmentation method includes the following steps:

[0071] Step S1: acquiring a multimodal medical examination image; classifying the multimodal medical examination image by image type to obtain a CT tomography image and a PET image respectively;

[0072] In an embodiment of the present invention, a CT tomographic image and a PET (positron emission tomography) image of a patient are sequentially acquired by a high-resolution imaging device. Specifically, the CT scanning device is adjusted to a mode suitable for whole-body scanning, the target area of ​​the patient's body is scanned in layers, and the tomographic data containing the lesion is acquired. These image data generate an initial grayscale image through the original signal conversion module inside the device. At the same time, after the PET device injects a radioactive tracer into the target area, the distribution of the tracer is recorded to generate corresponding PET imaging data. After the imaging is completed, the multimodal images are classified and processed. First, metadata is extracted from CT and PET images by an image file format parsing module, and the image source is distinguished according to the identification of the scanning device and imaging parameters (such as layer thickness, energy intensity, etc.). Subsequently, an algorithm based on image spectrum analysis is used to use the spectrum characteristics of the two types of images as a basis for distinction, and they are marked as CT images and PET images, respectively. In this process, relying on the texture feature extraction and grayscale distribution statistics of the image, the automatic classification of CT images and PET images is realized, providing a guarantee for the subsequent accurate analysis of the lesion area.

[0073] Step S2: locating the lesion area on the CT tomography image and the PET image, respectively, to obtain a CT tomography lesion image and a PET lesion image; predicting the lesion geometric mutation morphology on the CT tomography lesion image to obtain lesion geometric mutation morphology prediction data; performing multi-modal image fusion boundary correction on the CT tomography lesion image and the PET lesion image based on the lesion geometric mutation morphology prediction data to obtain multi-modal lesion image fusion boundary correction data;

[0074] In an embodiment of the present invention, the lesion region is first located on the CT tomography image and the PET image. The specific operation is to extract the potential lesion region through image edge detection and region growing algorithm. The gradient-based Canny edge detection algorithm is applied to the CT image to generate target edge information; the threshold segmentation algorithm based on brightness gradient is used for the PET image to determine the high radioactive activity area. The lesion regions of the two images generate CT tomography lesion images and PET lesion images respectively. Subsequently, the lesion geometric mutation morphology is predicted for the CT tomography lesion image. The morphological change gradient of the lesion region in multiple directions is calculated by using the boundary change analysis method based on geometric topology theory, and the geometric mutation morphological features of the lesion are calculated in combination with the distribution of morphological features. The analysis data is organized into a multidimensional tensor to generate lesion geometric mutation morphology prediction data. Finally, based on the lesion geometric mutation morphology prediction data, the CT tomography lesion image and the PET lesion image are subjected to multimodal image fusion boundary correction. The two modal data are geometrically corrected by image registration algorithms (such as affine transformation and bilinear interpolation). The mutual information method is used to adjust the boundary of the fused image so that the boundary details of the CT image are consistent with the metabolic hotspot area of ​​the PET image, and finally the multimodal lesion image fusion boundary correction data is obtained.

[0075] Step S3: performing multimodal lesion region segmentation combination optimization on the CT tomography image and the PET image according to the multimodal lesion image fusion boundary correction data to obtain a lesion region segmentation combination optimization strategy;

[0076] In an embodiment of the present invention, a multimodal lesion region segmentation combination optimization is performed on CT images and PET images according to the multimodal lesion image fusion boundary correction data. First, the fusion boundary correction data is normalized, and the image grayscale distribution is adjusted using a histogram equalization algorithm to improve the image contrast and enhance the detail performance of the target area. After processing, the lesion image fusion boundary correction normalized data is generated. Next, based on the normalized data, the CT image and the PET image are segmented and combined and optimized. By combining the boundary gradient direction analysis method and the gradient enhancement algorithm of image segmentation, the direction tensor information of the lesion region boundary is calculated. Subsequently, the inter-modal gradient cross-correlation calculation method is used, and the cross-correlation weights of the CT and PET lesion regions are used as the optimization criterion, and finally the lesion region segmentation combination optimization strategy is obtained.

[0077] Step S4: construct an image segmentation model based on the policy gradient algorithm for the combined optimization strategy of lesion area segmentation, obtain the lesion image segmentation model, and send the lesion image segmentation model to the cloud platform to execute the multimodal nuclear medicine image segmentation method.

[0078] In an embodiment of the present invention, an image segmentation model is constructed using a policy gradient algorithm based on a combined optimization strategy for lesion region segmentation. First, the eigenvalue distribution range of the segmentation model is calculated for the boundary information in the optimization strategy and the gradient cross-correlation weight data between modalities to ensure that the segmentation result has boundary consistency and regional integrity. Subsequently, the Policy Gradient Method is used to optimize the weight distribution of the segmentation model. The weight distribution of the policy function is iteratively adjusted to maximize the segmentation accuracy of the lesion region. After the optimization is completed, a lesion image segmentation model with clear regional division capabilities is generated. Finally, the segmentation model is uploaded to the cloud platform in the form of an image feature parameter set for batch execution and real-time application of subsequent nuclear medicine image segmentation methods. This step ensures the scalability of the system and the real-time nature of the segmentation results.

[0079] Preferably, step S1 comprises the following steps:

[0080] Step S11: Acquire multimodal medical examination images;

[0081] Step S12: performing image grayscale optimization on the multimodal medical examination image to obtain a multimodal medical examination grayscale optimized image;

[0082] Step S13: spatially aligning the multimodal medical examination grayscale optimized images to obtain a medical spatially aligned optimized image;

[0083] Step S14: classify the medical space alignment optimization image into image types to obtain a CT tomography image and a PET image respectively.

[0084] In the embodiment of the present invention, the operation of acquiring multimodal medical examination images needs to be combined with the working characteristics of different medical imaging devices. First, a tomographic scanning operation is performed by a CT device. The specific method is to set scanning parameters, including layer thickness, scanning speed and tube voltage. After the full range of the target area is scanned layer by layer, multiple tomographic grayscale images are generated. Subsequently, nuclear medical imaging is performed using a PET device, and a radioactive tracer is used to mark the target metabolic area, and the tracer is injected into the patient's body through an intravenous injection. After the tracer is evenly distributed, PET imaging is started, the radioactive distribution signal in the body is recorded, and PET image data is generated. The above process is strictly carried out in accordance with the operating procedures to ensure that the image quality meets the requirements of subsequent analysis. All image data are stored in DICOM format after acquisition, providing unified input data for the next step of processing. Multimodal medical examination images are subjected to grayscale optimization processing. First, the grayscale distribution of the CT image is adjusted using a histogram equalization algorithm to enhance the detail performance of low-contrast areas. The specific operation is to calculate the cumulative distribution function of the image grayscale histogram and redistribute each grayscale value so that the overall brightness range of the image is more uniform. For PET images, a brightness enhancement method based on gamma correction is used to highlight high-radioactive areas by adjusting the gamma coefficient while avoiding excessive brightening of low-radioactive backgrounds. During the processing, optimization parameters are set according to the characteristics of different modalities to ensure that the optimized images can retain modal characteristics. Finally, a set of multimodal medical examination grayscale optimized images are generated to provide high-quality input for the spatial alignment operation in the subsequent steps. Multimodal medical examination grayscale optimized images are spatially aligned to eliminate the perspective offset and displacement errors during the imaging process. The specific implementation method includes the following steps. First, CT images are selected as the reference modality, and the SIFT (scale-invariant feature transform) algorithm is used to extract key points and generate a set of feature descriptors. Subsequently, the same feature extraction process is performed on PET images. Next, the key points of CT and PET images are paired using the nearest neighbor-based feature matching method. After the matching is completed, the affine transformation matrix between the two sets of key points is calculated, and the parameters of the matrix are optimized by the least squares method to ensure the alignment accuracy. Finally, the PET image is geometrically corrected using the affine transformation so that it is completely aligned with the CT image in space, and the medical space alignment optimized image is obtained, laying the foundation for image classification. Image type classification was performed on medical spatial alignment optimization images. First, the metadata of the image was parsed to extract relevant information about the imaging modality, such as the model of the scanning device, parameter settings, and imaging protocols. This information can preliminarily distinguish between CT and PET images. Subsequently, the image type was further confirmed using a spectrum analysis method. The specific method was to perform a fast Fourier transform (FFT) on each image, calculate its frequency domain features, and extract the energy distribution in the low-frequency region. According to the prominence of the high-frequency features of CT images and the low-frequency energy distribution characteristics of PET images, a discrimination threshold was set to divide the images into two categories.Finally, the classification results are stored as CT scan images and PET images in separate folders, providing clear input data for subsequent lesion analysis.

[0085] Preferably, step S2 comprises the following steps:

[0086] Step S21: locating the lesion area on the CT tomography image and the PET image, respectively CT tomography lesion image and PET lesion image;

[0087] Step S22: predicting the geometric mutation morphology of the lesion on the CT section lesion image to obtain the prediction data of the geometric mutation morphology of the lesion;

[0088] Step S23: evaluating the metabolic activity of the lesion mutation region on the PET lesion image according to the lesion geometric mutation morphology prediction data to obtain metabolic activity data of the lesion mutation region;

[0089] Step S24: performing multimodal image fusion boundary correction on the CT tomographic lesion image and the PET lesion image based on the metabolic activity data of the lesion mutation area and the lesion geometric mutation morphology prediction data to obtain multimodal lesion image fusion boundary correction data.

[0090] As an example of the present invention, refer to Figure 2 As shown, in this example, step S2 includes:

[0091] Step S21: locating the lesion area on the CT tomography image and the PET image, respectively CT tomography lesion image and PET lesion image;

[0092] In an embodiment of the present invention, CT tomographic images and PET images are used to locate the lesion area. First, a segmentation method based on a watershed algorithm is applied to the CT image to enhance the significance of the grayscale gradient and accurately separate the lesion area and the background tissue. On this basis, texture features are extracted by calculating the grayscale co-occurrence matrix, including parameters such as energy, contrast and uniformity, to locate the suspected lesion area. In the PET image, an adaptive threshold segmentation method is used to set a dynamic threshold, and a high metabolic area is extracted according to the local peak of the radioactive distribution. Subsequently, the results of the two modalities are combined, and the lesion area is jointly aligned by geometric correction to generate the corresponding CT tomographic lesion images and PET lesion images. The key to this process is to use a modality-specific feature extraction algorithm to ensure the positioning accuracy of the lesion area.

[0093] Step S22: predicting the geometric mutation morphology of the lesion on the CT section lesion image to obtain the prediction data of the geometric mutation morphology of the lesion;

[0094] In an embodiment of the present invention, the CT section lesion image is used to predict the geometric mutation morphology of the lesion. The specific operation first uses the Hough transform algorithm based on edge detection to extract the boundary curve of the lesion area and obtain the geometric contour of the lesion. Subsequently, the segmented spline interpolation method is used to smooth the boundary curve of the lesion to remove artifact interference. On this basis, the geometric center of the lesion and its shape characteristic parameters, including the major axis, minor axis, eccentricity and shape moment, are calculated. Then, through the time series analysis method, the dynamic change trend of the lesion geometry is modeled based on the multi-layer slice data, the occurrence location of the geometric mutation point and its change amplitude are predicted, and finally the lesion geometry mutation morphology prediction data is generated as the input data for subsequent analysis.

[0095] Step S23: evaluating the metabolic activity of the lesion mutation region on the PET lesion image according to the lesion geometric mutation morphology prediction data to obtain metabolic activity data of the lesion mutation region;

[0096] In an embodiment of the present invention, the metabolic activity of the lesion mutation area is evaluated for the PET lesion image according to the lesion geometric mutation morphology prediction data. First, the key area in the PET image is located using the mutation morphology prediction data, and a dynamic radioactivity extraction operation is performed on the area. The specific steps include using the block weighted average method to count the radioactivity counts in the lesion area, and calculating the standardized uptake value (SUV) to quantify the metabolic activity intensity. At the same time, the metabolic activity of adjacent slices is three-dimensionally reconstructed by a spatial transformation method to form a regional metabolic activity heat map. Then, a gradient-based activity difference detection algorithm is used to analyze the metabolic activity difference between the lesion area and the adjacent tissue, and finally the metabolic activity data of the lesion mutation area is generated to provide basic data support for subsequent fusion boundary correction.

[0097] Step S24: performing multimodal image fusion boundary correction on the CT tomographic lesion image and the PET lesion image based on the metabolic activity data of the lesion mutation area and the lesion geometric mutation morphology prediction data to obtain multimodal lesion image fusion boundary correction data.

[0098] In an embodiment of the present invention, based on the metabolic activity data of the lesion mutation area and the lesion geometric mutation morphology prediction data, multimodal image fusion boundary correction is performed on the CT tomographic lesion image and the PET lesion image. First, the boundary control points in the CT image are determined using the lesion geometric mutation morphology prediction data, the metabolic activity distribution of the PET image is registered, and the metabolic activity boundary of the lesion area is preliminarily corrected by the variational method. Then, based on the joint alignment optimization strategy, the Poisson fusion algorithm is used to seamlessly fuse the CT and PET images in the boundary area and smooth the edge transition. Finally, the corrected image is pixel-level refined by local resampling technology to generate multimodal lesion image fusion boundary correction data with high-precision lesion area boundaries, providing accurate input for subsequent image segmentation and clinical analysis.

[0099] Preferably, step S22 includes the following steps:

[0100] Step S221: extracting the lesion contour from the CT section lesion image to obtain lesion contour data;

[0101] Step S222: performing three-dimensional grid division on the CT tomographic lesion image according to the lesion contour data to obtain three-dimensional lesion grid data;

[0102] Step S223: Calculating the boundary curvature distribution of the lesion contour data based on the lesion three-dimensional grid data to obtain lesion boundary distribution curvature data;

[0103] Step S224: Calculate the distribution curvature approximate mean difference of the lesion boundary distribution curvature data to obtain the distribution curvature approximate mean difference data;

[0104] Step S225: performing lesion geometric mutation point evolution deduction on the lesion three-dimensional grid data according to the distribution curvature adjacent mean difference data to obtain lesion geometric mutation point evolution data;

[0105] Step S226: Predict the lesion geometric mutation morphology based on the distribution curvature proximity mean difference data and the lesion geometric mutation point evolution data to obtain lesion geometric mutation morphology prediction data.

[0106] In an embodiment of the present invention, lesion contour extraction is performed on a CT tomographic lesion image. First, the grayscale gradient of the CT tomographic lesion image is enhanced by using the Canny edge detection algorithm to identify potential edge areas. Subsequently, isolated noise points are removed by morphological operations, including expansion and corrosion operations, to ensure edge continuity. On the basis of obtaining preliminary edges, a chain code encoding algorithm based on boundary tracking is applied to extract the complete closed contour of the lesion area. Finally, the lesion contour data is stored as a lesion contour data file by vectorization to provide input data for subsequent analysis. The key operation in this step is to use a method combining edge detection and morphological tools to ensure accurate extraction of the lesion contour. The CT tomographic lesion image is divided into three-dimensional grids according to the lesion contour data. Using the lesion contour data as input, a two-dimensional grid is generated by the Delaunay triangulation algorithm. Subsequently, the multi-layer slices of the CT tomographic lesion image are superimposed, and three-dimensional lesion point cloud data is generated by an interpolation method. The point cloud data is extracted by using the Marching Cubes algorithm to construct a regularized three-dimensional grid model to form a three-dimensional grid data of the lesion. During the entire meshing process, the volume and surface area of ​​each mesh unit are calculated to ensure that the mesh units are evenly distributed. The mesh data is stored as a structured three-dimensional coordinate data file to support subsequent geometric analysis and deduction. The boundary curvature distribution of the lesion contour data is calculated based on the three-dimensional mesh data of the lesion. First, the surface vertices and their adjacency relationships of the three-dimensional mesh data of the lesion are extracted, and the local surface of each vertex is fitted using the least squares method. Then, the principal curvature and mean curvature of each vertex are calculated to generate a complete curvature distribution matrix. In order to more intuitively display the distribution, the curvature value is visualized by pseudo-color mapping technology, and a curvature distribution image file is generated. The focus of this step is to use the curvature analysis method to quantify the surface morphological characteristics of the lesion and provide an accurate curvature distribution curve for subsequent data analysis. The distribution curvature adjacent mean difference is calculated for the lesion boundary distribution curvature data. For the curvature distribution matrix, the difference between the mean curvature of each vertex and the mean curvature of its neighboring vertices is calculated to obtain the curvature adjacent mean difference. In the specific operation, the weighted moving average method is used to calculate the mean of the neighborhood curvature, and based on this, the curvature change characteristics of the local area are normalized to ensure that the mean difference values ​​of different areas are comparable. Finally, the mean difference results are stored as the distribution curvature neighborhood mean difference data file to provide basic data for the evolution deduction of geometric mutation points. The evolution of the lesion geometric mutation points is deduced based on the distribution curvature neighborhood mean difference data. First, the distribution curvature neighborhood mean difference data is superimposed on the lesion three-dimensional grid model, and the high-change area is marked as the initial position of the geometric mutation point. The dynamic evolution trajectory of the lesion geometry between adjacent slices is simulated through the time step iteration method, and the displacement change of each geometric mutation point in three-dimensional space is recorded.The evolution path of each mutation point is fitted using Bezier curves to generate geometric mutation point evolution data to support the prediction and visualization of geometric morphology. The lesion geometric mutation morphology is predicted based on the distribution curvature proximity mean difference data and the lesion geometric mutation point evolution data. First, the distribution curvature proximity mean difference data and the geometric mutation point evolution data are multi-dimensionally fused to extract the key factors affecting the geometric morphology mutation. Then, the stress field and deformation field of the lesion area are calculated using the finite element analysis method to predict the evolution trend of the mutation morphology. Combining the evolution path and mutation morphology prediction, the prediction results are generated and stored as lesion geometric mutation morphology prediction data files to support subsequent multimodal image processing related operations.

[0107] Preferably, step S24 comprises the following steps:

[0108] Step S241: performing spatial morphological feature decomposition processing on the lesion geometric mutation morphological prediction data to obtain lesion morphological feature decomposition data;

[0109] Step S242: performing metabolic activity correlation analysis on the metabolic activity data of the lesion mutation area according to the lesion morphology feature decomposition data to obtain lesion morphology-metabolic activity correlation data;

[0110] Step S243: performing trend hotspot local enhancement processing on the CT tomographic lesion image and the PET lesion image based on the lesion morphology-metabolic activity correlation data to obtain trend hotspot enhanced image data;

[0111] Step S244: adjusting the center of gravity of the overlapping area of ​​the CT tomographic lesion image and the PET lesion image according to the trend hotspot enhanced image data to obtain the center of gravity adjustment data of the overlapping area;

[0112] Step S245: performing multimodal image fusion boundary correction on the CT tomographic lesion image and the PET lesion image based on the overlapping area gravity center adjustment data to obtain multimodal lesion image fusion boundary correction data.

[0113] In an embodiment of the present invention, the spatial morphological feature decomposition processing is performed on the lesion geometric mutation morphological prediction data. First, spatial geometric features are extracted from the lesion geometric mutation morphological prediction data, including indicators such as three-dimensional contours, boundary curvature distribution, and surface roughness. Then, the singular value decomposition (SVD) algorithm is applied to perform principal component analysis on these features to reduce the high-dimensional feature data to a specific set of main feature vectors. The reduced feature data is divided into feature groups of different scales, which respectively represent the global morphology and local geometric changes of the lesion. Subsequently, the multi-scale decomposition algorithm is used to recombine each feature group to form lesion morphological feature decomposition data, which is stored as a structured data file for subsequent analysis. Metabolic activity correlation analysis is performed on the lesion mutation area metabolic activity data according to the lesion morphological feature decomposition data. First, the lesion morphological feature decomposition data and the metabolic activity data are spatially aligned to ensure that the two are aligned in the same spatial reference frame. Subsequently, the correlation between the morphological features and metabolic activity of each spatial point is calculated based on the Pearson correlation coefficient, and a correlation matrix is ​​generated. For the highly correlated regions in the correlation matrix, statistical significance tests (such as t-tests) were used to screen out key morphological features and metabolic activity-related regions, and the lesion morphology-metabolic activity correlation data were obtained and stored as a structured file in the form of a matrix. Trend hotspot local enhancement processing was performed on CT tomographic lesion images and PET lesion images based on the lesion morphology-metabolic activity correlation data. First, the hotspot regions with high correlation in the morphology-metabolic activity correlation data were located. Then, the Gaussian enhancement filtering method was used to refine the edges of the hotspot regions, and the contrast and detail level of the hotspot regions were enhanced by histogram equalization. On this basis, adaptive sharpening technology was used to further enhance the significance of the hotspot regions to ensure the clear presentation of local features in multimodal images. The processed image data was stored as trend hotspot enhanced image data to support subsequent image boundary optimization. The center of gravity of the overlapping regions of CT tomographic lesion images and PET lesion images were adjusted according to the trend hotspot enhanced image data. First, the overlapping regions of CT images and PET images were extracted from the trend hotspot enhanced image data, and their center of gravity positions were calculated. Next, the weighted centroid of the pixel values ​​in the overlapping area is calculated based on the centroid method, and the CT and PET images are geometrically deformed to make the center of gravity positions completely aligned. During the adjustment process, the bilinear interpolation method is used to spatially resample the image to ensure that the pixel information in the overlapping area is not distorted. The adjusted image data is stored as the center of gravity adjustment data of the overlapping area to ensure the registration accuracy of the multimodal image. The multimodal image fusion boundary correction is performed on the CT tomographic lesion image and the PET lesion image based on the center of gravity adjustment data of the overlapping area. First, a weight-based image fusion algorithm is used to achieve efficient fusion of multimodal information by assigning weights to the edge sharpness of the CT image and the metabolic intensity of the PET image. Subsequently, the region growing method is used to correct the continuity of the boundary of the fused image to ensure that the boundary line is smooth and accurate.Finally, the fusion result is refined by combining the image smoothing algorithm to remove potential artifacts and boundary errors. The processed image is stored as multimodal lesion image fusion boundary correction data for further segmentation and analysis operations.

[0114] Preferably, step S243 includes the following steps:

[0115] Extracting hot spots of regional metabolic abnormalities from the lesion morphology-metabolic activity correlation data to obtain hot spots of metabolic abnormality regions;

[0116] According to the hotspot metabolic abnormality area data, the CT tomographic lesion image is subjected to local morphological multi-scale enhancement processing to obtain CT multi-scale morphological enhancement data;

[0117] Based on CT multi-scale morphological enhancement data and hotspot metabolic abnormality area data, the PET lesion images are processed with metabolic intensity hotspots to obtain PET metabolic hotspot stratification data;

[0118] According to the PET metabolic hotspot stratification data and CT multi-scale morphological enhancement data, trend hotspot local enhancement processing is performed to obtain trend hotspot enhanced image data.

[0119] In an embodiment of the present invention, in the extraction process of regional metabolic abnormality hotspots, based on the lesion morphology-metabolic activity correlation data, a threshold segmentation method is used to extract metabolic abnormality hotspot areas. In the specific operation, the metabolic activity data is first normalized, and the data range is standardized to the [0,1] interval. Subsequently, the global mean and standard deviation of the data are calculated according to a statistical method, and the extraction threshold is set to the range of the mean plus two times the standard deviation to ensure that only areas with significant metabolic abnormality are extracted. For each candidate hotspot area, the connected domain analysis method is applied to eliminate small-area noise areas, and finally form the spatial index data of the hotspot metabolic abnormality area. This processing result is stored as hotspot metabolic abnormality area data for subsequent enhancement and hierarchical analysis. Based on the hotspot metabolic abnormality area data, the CT tomographic lesion image is subjected to local morphological multi-scale enhancement processing. In the specific operation, the spatial coordinate information of the hotspot metabolic abnormality area is first extracted, and it is mapped to the CT tomographic image to lock the target area for local enhancement. Subsequently, the multi-scale wavelet transform is used to decompose the detail information of the target area, and the high-frequency component is used for enhancement processing. In order to avoid the amplification of high-frequency noise, the bilateral filtering technology is used to smooth the high-frequency component, and the texture noise is eliminated while maintaining the edge information. The enhanced high-frequency information is resynthesized with the low-frequency component to form an enhanced CT tomographic lesion image. The generated CT multi-scale morphological enhancement data is stored in matrix form for use in subsequent steps. Based on the CT multi-scale morphological enhancement data and the hotspot metabolic abnormality area data, the PET lesion image is subjected to metabolic intensity hotspot stratification processing. First, the CT multi-scale morphological enhancement data and the PET image are aligned by spatial registration method to ensure the consistency of the hotspot area. Then, according to the distribution of metabolic intensity values, the metabolic hotspot area of ​​the PET image is divided into three layers: high metabolism, medium metabolism and low metabolism. The stratification operation uses a piecewise linear normalization method to map the metabolic intensity range of each layer to the [0,1] interval and assign different hierarchical labels. After processing, the metabolic hotspots of each layer are stored as three-dimensional image slices, and the PET metabolic hotspot stratification data is formed and saved in a three-dimensional array format for subsequent trend hotspot enhancement processing. According to the PET metabolic hotspot stratification data and the CT multi-scale morphological enhancement data, the trend hotspot local enhancement processing is performed. First, the two data are fused by mutual information registration technology, and the correlation distribution with the CT morphological enhancement data is calculated for each metabolic hotspot layer. Subsequently, an adaptive contrast enhancement method was used to enhance the edges and details of the high metabolic layer, while texture enhancement technology was used to improve the structural clarity of the medium metabolic layer. For the low metabolic layer, brightness equalization was used to maintain the consistency of the overall metabolic trend. The fused image was stored in a multi-channel format as trend hotspot enhanced image data, providing complete and optimized multimodal information for subsequent steps.

[0120] Preferably, step S3 comprises the following steps:

[0121] Step S31: normalizing the multimodal lesion image fusion boundary correction data to obtain lesion image fusion boundary correction normalized data;

[0122] Step S32: performing multimodal lesion region segmentation combination optimization on the CT tomography image and the PET image based on the lesion image fusion boundary correction normalization data to obtain a lesion region segmentation combination optimization strategy.

[0123] In an embodiment of the present invention, in the process of normalizing the multimodal lesion image fusion boundary correction data, a linear mapping method is used to unify the data into a standardized numerical range, so as to facilitate the unified operation of subsequent processing, extract the global maximum and minimum values ​​of all pixel values ​​in the multimodal lesion image fusion boundary correction data, and then optimize the brightness distribution of the normalized result by a histogram equalization method to improve the local contrast of the details. Finally, the lesion image fusion boundary correction normalized data is generated, and it is saved in a multidimensional array format to support subsequent calculations to ensure the compatibility of data between different segmentation algorithms. In the process of combining and optimizing the multimodal lesion area segmentation of CT tomography images and PET images based on the lesion image fusion boundary correction normalized data, a multi-stage segmentation and optimization strategy is used to achieve accurate lesion area identification. The specific steps are as follows: the normalized data is input into the regional growing algorithm, the high-density lesion boundary area of ​​the CT image is used as the initial seed point, and the high metabolic area of ​​the PET image is combined to expand the joint area according to spatial adjacency and metabolic activity. By setting a spatial consistency threshold, the non-lesion area is eliminated and a preliminary segmentation result is generated. For the boundary area of ​​the preliminary segmentation result, the edge detection algorithm based on gradient amplitude is used to refine the boundary of the CT image, and the metabolic intensity gradient data of the PET image is used to enhance the edge contrast. After the two data are fused, the boundary correction is performed by the multi-layer gradient synthesis method. The optimized boundary area is subjected to morphological analysis, and the area, volume, shape factor and other indicators of the area are calculated, and compared with the statistical characteristics of the lesions extracted from the normalized data. The morphologically abnormal area is readjusted using the dynamic fuzzy segmentation method. The corrected lesion area is aligned with the CT and PET dual modalities, and the cross entropy optimization algorithm is used to fuse the segmentation results of the two modalities to ensure the consistency of the boundary and the coordination of the distribution of metabolic activity.

[0124] Preferably, step S32 includes the following steps:

[0125] Step S321: performing boundary gradient direction tensor analysis on the lesion image fusion boundary correction normalization data to obtain multi-dimensional gradient direction tensor field data;

[0126] Step S322: performing inter-modality gradient cross-correlation calculation on the CT tomography image and the PET image based on the multi-dimensional gradient direction tensor field data to obtain modality gradient cross-correlation weight data;

[0127] Step S323: performing multi-modal lesion region segmentation combination optimization on the CT tomography image and the PET image according to the modal gradient cross-correlation weight data to obtain a lesion region segmentation combination optimization strategy.

[0128] In an embodiment of the present invention, when the boundary gradient direction tensor analysis processing is performed on the normalized data of the lesion image fusion boundary correction, the boundary characteristics of the lesion need to be fully described from the perspective of spatial gradient distribution. The specific operation is as follows: First, the normalized lesion image data is subjected to spatial gradient calculation, and the gradient direction and gradient amplitude of each pixel point are extracted by the Sobolev gradient operator to obtain a two-dimensional or three-dimensional gradient vector field. Subsequently, the gradient vector field is converted into a gradient direction tensor by an extended tensor construction method. Each tensor is composed of a gradient direction, an amplitude and a spatial coordinate. Next, the tensor field is subjected to eigenvalue decomposition, the eigenvalues ​​and eigenvectors of the main gradient directions are extracted, and a directional intensity matrix is ​​constructed to quantify the directional distribution characteristics of the lesion boundary. On this basis, the tensor field is further smoothed by a symmetric filtering technique to remove high-frequency noise and enhance the continuity of the boundary direction. After the processing is completed, the result is stored as multidimensional gradient direction tensor field data for subsequent modal cross-correlation analysis. Based on the multidimensional gradient direction tensor field data, inter-modal gradient cross-correlation calculation is performed on CT tomography images and PET images to quantify the gradient direction similarity between the two modalities. The specific operation is as follows: First, the CT and PET images are spatially registered to the same reference coordinate system, and the gradient fields of the two modalities are calculated respectively to form the CT gradient tensor field and the PET gradient tensor field. Then, the angle between the gradient vectors of each modality is calculated by the direction cosine. Next, the weighted cross-correlation value of the gradient amplitude is calculated, and the direction similarity and gradient intensity distribution are comprehensively considered. The result is expressed as the modal gradient cross-correlation weight matrix. The weight matrix is ​​processed by spatial domain smoothing filtering to suppress isolated outliers and ensure spatial continuity. Finally, the processing result is saved in the form of a matrix as the modal gradient cross-correlation weight data, which is used as an important weight factor for multi-modal segmentation combination optimization. According to the modal gradient cross-correlation weight data, the multi-modal lesion area segmentation combination optimization is performed on the CT tomography image and the PET image to obtain the optimal segmentation result. The specific operation is as follows: First, the preliminary segmentation area of ​​CT and PET is jointly optimized with the modal gradient cross-correlation weight as a guide. For each pixel point, the maximum weight strategy is used to select the segmentation boundary point of the corresponding modality to ensure the consistency of the segmentation area in morphology and metabolic activity. Then, the boundary area with the highest weight is locally refined, and the boundary of the lesion area is dynamically adjusted by using an adaptive region expansion algorithm, guided by the boundary gradient information of the CT modality and combined with the metabolic hotspot distribution of the PET modality. In order to improve the spatial smoothness of the boundary, the morphological reconstruction method is used to remove pseudo-boundaries or discontinuous areas. Finally, the optimized segmentation results of CT and PET are fused, and a global optimization method based on graph cut theory is used to establish a segmentation energy function. The weight data is used as the constraint condition of the energy term to complete the global consistency optimization of the segmentation results. Finally, the combined optimization strategy data of the lesion area segmentation is output for subsequent multimodal nuclear medicine image segmentation and analysis.

[0129] Preferably, the present invention further provides a multimodal nuclear medicine image segmentation system for executing the multimodal nuclear medicine image segmentation method as described above, the multimodal nuclear medicine image segmentation system comprising:

[0130] A medical image type classification module is used to obtain multimodal medical examination images; image type classification is performed on the multimodal medical examination images to obtain CT tomography images and PET images respectively;

[0131] The fusion boundary correction module is used to locate the lesion area on the CT tomography image and the PET image, respectively, to obtain the CT tomography lesion image and the PET lesion image; to predict the lesion geometric mutation morphology on the CT tomography lesion image to obtain the lesion geometric mutation morphology prediction data; to perform multi-modal image fusion boundary correction on the CT tomography lesion image and the PET lesion image based on the lesion geometric mutation morphology prediction data to obtain the multi-modal lesion image fusion boundary correction data;

[0132] A segmentation combination optimization module is used to perform multi-modal lesion region segmentation combination optimization on CT tomography images and PET images according to multi-modal lesion image fusion boundary correction data to obtain a lesion region segmentation combination optimization strategy;

[0133] The segmentation model construction module is used to construct an image segmentation model based on the policy gradient algorithm for the combined optimization strategy of lesion area segmentation, obtain the lesion image segmentation model, and send the lesion image segmentation model to the cloud platform to execute the multimodal nuclear medicine image segmentation method.

[0134] Preferably, the present invention further provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed, implements the multimodal nuclear medicine image segmentation method as described in any one of the above.

[0135] Therefore, the embodiments should be regarded as illustrative and non-restrictive from all points, and the scope of the present invention is limited by the appended claims rather than the above description, and it is therefore intended that all changes falling within the meaning and range of equivalent elements of the application documents are included in the present invention.

[0136] The above description is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but should conform to the widest scope consistent with the principles and novel features invented herein.

Claims

1. A multimodal nuclear medicine image segmentation method, characterized in that: The following steps are involved: Step S1: acquiring a multimodal medical examination image; classifying the multimodal medical examination image by image type to obtain a CT tomography image and a PET image respectively; Step S2: locating the lesion area on the CT tomography image and the PET image, respectively, to obtain a CT tomography lesion image and a PET lesion image; predicting the lesion geometric mutation morphology on the CT tomography lesion image to obtain lesion geometric mutation morphology prediction data; performing multimodal image fusion boundary correction on the CT tomography lesion image and the PET lesion image based on the lesion geometric mutation morphology prediction data to obtain multimodal lesion image fusion boundary correction data; Step S2 is specifically as follows: Step S21: locating the lesion area on the CT tomography image and the PET image, respectively CT tomography lesion image and PET lesion image; Step S22: predicting the lesion geometric mutation morphology on the CT section lesion image to obtain lesion geometric mutation morphology prediction data; Step S22 is specifically as follows: Step S221: extracting the lesion contour from the CT section lesion image to obtain lesion contour data; Step S222: performing three-dimensional grid division on the CT tomographic lesion image according to the lesion contour data to obtain three-dimensional lesion grid data; Step S223: performing boundary curvature distribution calculation on the lesion contour data based on the lesion three-dimensional mesh data to obtain lesion boundary distribution curvature data; including: extracting surface vertices and their adjacency relationships of the lesion three-dimensional mesh data, fitting the local surface of each vertex using the least squares method, calculating the principal curvature and average curvature of each vertex, and generating a complete curvature distribution matrix; Step S224: Calculate the distribution curvature proximity mean difference of the lesion boundary distribution curvature data to obtain the distribution curvature proximity mean difference data; including: for the curvature distribution matrix, calculate the difference between the curvature mean of each vertex and the curvature mean of its adjacent vertices to obtain the curvature proximity mean difference value; in the specific operation, use the weighted moving average method to calculate the neighborhood curvature mean, and use this as a benchmark to normalize the curvature change characteristics of the local area to ensure that the curvature proximity mean difference values ​​of different areas are comparable; store the result of the curvature proximity mean difference value as the distribution curvature proximity mean difference data; Step S225: performing lesion geometric mutation point evolution deduction on the lesion three-dimensional grid data according to the distribution curvature adjacent mean difference data to obtain lesion geometric mutation point evolution data; Step S226: predicting the geometric mutation morphology of the lesion according to the distribution curvature proximity mean difference data and the lesion geometric mutation point evolution data, to obtain the lesion geometric mutation morphology prediction data; Step S23: evaluating the metabolic activity of the lesion mutation region on the PET lesion image according to the lesion geometric mutation morphology prediction data to obtain metabolic activity data of the lesion mutation region; Step S24: performing multimodal image fusion boundary correction on the CT tomographic lesion image and the PET lesion image based on the metabolic activity data of the lesion mutation area and the lesion geometric mutation morphology prediction data to obtain multimodal lesion image fusion boundary correction data; Step S3: performing multimodal lesion region segmentation combination optimization on the CT tomography image and the PET image according to the multimodal lesion image fusion boundary correction data to obtain a lesion region segmentation combination optimization strategy; Step S4: construct an image segmentation model based on the policy gradient algorithm for the combined optimization strategy of lesion area segmentation, obtain the lesion image segmentation model, and send the lesion image segmentation model to the cloud platform to execute the multimodal nuclear medicine image segmentation method.

2. The multimodal nuclear medicine image segmentation method according to claim 1, characterized in that: Step S1 includes the following steps: Step S11: Acquire multimodal medical examination images; Step S12: performing image grayscale optimization on the multimodal medical examination image to obtain a multimodal medical examination grayscale optimized image; Step S13: spatially aligning the multimodal medical examination grayscale optimized images to obtain a medical spatially aligned optimized image; Step S14: classify the medical space alignment optimization image into image types to obtain a CT tomography image and a PET image respectively.

3. The multimodal nuclear medicine image segmentation method according to claim 1, characterized in that: Step S24 includes the following steps: Step S241: performing spatial morphological feature decomposition processing on the lesion geometric mutation morphological prediction data to obtain lesion morphological feature decomposition data; Step S242: performing metabolic activity correlation analysis on the metabolic activity data of the lesion mutation area according to the lesion morphology feature decomposition data to obtain lesion morphology-metabolic activity correlation data; Step S243: performing trend hotspot local enhancement processing on the CT tomographic lesion image and the PET lesion image based on the lesion morphology-metabolic activity correlation data to obtain trend hotspot enhanced image data; Step S244: adjusting the center of gravity of the overlapping area of ​​the CT tomographic lesion image and the PET lesion image according to the trend hotspot enhanced image data to obtain the center of gravity adjustment data of the overlapping area; Step S245: performing multimodal image fusion boundary correction on the CT tomographic lesion image and the PET lesion image based on the overlapping area gravity center adjustment data to obtain multimodal lesion image fusion boundary correction data.

4. The multimodal nuclear medicine image segmentation method according to claim 3, characterized in that: Step S243 includes the following steps: Extracting hot spots of regional metabolic abnormalities from the lesion morphology-metabolic activity correlation data to obtain hot spots of metabolic abnormality regions; According to the hotspot metabolic abnormality area data, the CT tomographic lesion image is subjected to local morphological multi-scale enhancement processing to obtain CT multi-scale morphological enhancement data; Based on CT multi-scale morphological enhancement data and hotspot metabolic abnormality area data, the PET lesion images are processed with metabolic intensity hotspots to obtain PET metabolic hotspot stratification data; According to the PET metabolic hotspot stratification data and CT multi-scale morphological enhancement data, trend hotspot local enhancement processing is performed to obtain trend hotspot enhanced image data.

5. The multimodal nuclear medicine image segmentation method according to claim 1, characterized in that: Step S3 includes the following steps: Step S31: normalizing the multimodal lesion image fusion boundary correction data to obtain lesion image fusion boundary correction normalized data; Step S32: performing multimodal lesion region segmentation combination optimization on the CT tomography image and the PET image based on the lesion image fusion boundary correction normalization data to obtain a lesion region segmentation combination optimization strategy.

6. The multimodal nuclear medicine image segmentation method according to claim 5, characterized in that: Step S32 includes the following steps: Step S321: performing boundary gradient direction tensor analysis on the lesion image fusion boundary correction normalization data to obtain multi-dimensional gradient direction tensor field data; Step S322: performing inter-modality gradient cross-correlation calculation on the CT tomography image and the PET image based on the multi-dimensional gradient direction tensor field data to obtain modality gradient cross-correlation weight data; Step S323: performing multi-modal lesion region segmentation combination optimization on the CT tomography image and the PET image according to the modal gradient cross-correlation weight data to obtain a lesion region segmentation combination optimization strategy.

7. A multimodal nuclear medicine image segmentation system, characterized in that: For executing the multimodal nuclear medicine image segmentation method according to claim 1, the multimodal nuclear medicine image segmentation system comprises: A medical image type classification module is used to obtain multimodal medical examination images; image type classification is performed on the multimodal medical examination images to obtain CT tomography images and PET images respectively; The fusion boundary correction module is used to locate the lesion area on the CT tomography image and the PET image, respectively, to obtain the CT tomography lesion image and the PET lesion image; to predict the lesion geometric mutation morphology on the CT tomography lesion image to obtain the lesion geometric mutation morphology prediction data; to perform multi-modal image fusion boundary correction on the CT tomography lesion image and the PET lesion image based on the lesion geometric mutation morphology prediction data to obtain the multi-modal lesion image fusion boundary correction data; A segmentation combination optimization module is used to perform multi-modal lesion region segmentation combination optimization on CT tomography images and PET images according to multi-modal lesion image fusion boundary correction data to obtain a lesion region segmentation combination optimization strategy; The segmentation model construction module is used to construct an image segmentation model based on the policy gradient algorithm for the combined optimization strategy of lesion area segmentation, obtain the lesion image segmentation model, and send the lesion image segmentation model to the cloud platform to execute the multimodal nuclear medicine image segmentation method.

8. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the computer program is executed, the multimodal nuclear medicine image segmentation method as described in any one of claims 1 to 6 is implemented.

Images