Tumor dynamic segmentation method and model training method suitable for BNCT treatment

Abstract

The invention provides a dynamic tumor segmentation method suitable for BNCT treatment and a model training method. The dynamic tumor segmentation method comprises the following steps: acquiring PET-CT multi-modal data; inputting the PET-CT multi-modal data into a trained tumor segmentation model, so as to obtain a tumor segmentation probability graph based on a spatial position dimension and a boron concentration distribution dimension according to the tumor segmentation model; and performing dynamic threshold segmentation on the tumor segmentation probability graph based on boron drug metabolism kinetic characteristics to obtain a tumor segmentation result. According to the method, the PET-CT multi-mode data, the boron sensitive attention mechanism and the BNCT boron drug metabolism kinetic characteristics are combined, high-precision segmentation of the tumor is achieved, the segmentation result reflects the boron drug biological distribution characteristics at the same time, and more accurate input parameters are provided for a BNCT treatment plan.

Description

Technical Field

[0001] This application belongs to the field of radiation technology and relates to a dynamic tumor segmentation method and model training method suitable for BNCT treatment. Background Technology

[0002] Cancer has become one of the leading diseases posing a serious threat to the health and lives of the Chinese people. With the continuous advancement of medical technology worldwide, the treatment methods for cancer have gradually increased, and treatment effects have also significantly improved, encompassing various approaches such as surgical resection, chemotherapy, radiation therapy, immunomodulatory therapy, and hormone therapy. In recent years, with the continuous progress of radiotherapy technology, the concept of radiotherapy has been constantly updated. Precision radiotherapy accurately targets the treatment area, ensuring the accuracy of radiation irradiation and balanced dose distribution, while possessing high adaptability and minimizing damage to surrounding sensitive organs. Thus, it embodies a highly efficient and rapid treatment method that balances safety and excellent efficacy. Among these, boron neutron capture therapy (BNCT), as an innovative cancer treatment technology combining radiotherapy and drug therapy, has advantages such as low targeted cost, low side effects, high safety, and short treatment course, and has therefore gradually become a hot topic in tumor treatment, achieving good therapeutic results in numerous clinical trials.

[0003] Despite the significant advantages of BNCT, its efficacy is highly dependent on precise delineation of the target area and organs at risk before treatment—that is, accurately identifying tumor boundaries, normal tissue locations, and boron drug distribution areas through medical images, thereby guiding the precise control of neutron beam direction and dosage. Medical image segmentation is a key technical step in this process. This involves precisely segmenting the tumor target volume (GTV), clinical target volume (CTV), and organs at risk (OARs, such as the brainstem, optic nerve, and parotid gland) in multimodal medical images (such as computed tomography providing anatomical information, magnetic resonance imaging reflecting soft tissue contrast, and positron emission tomography showing metabolic activity) at the pixel and voxel level to generate a three-dimensional model, providing input parameters for the BNCT treatment planning system (TPS).

[0004] Therefore, how to accurately segment tumors in BNCT treatment scenarios is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] This application provides a dynamic tumor segmentation method and model training method applicable to BNCT treatment, which is used to solve the technical problem of how to accurately segment tumors in the BNCT treatment scenario.

[0006] The first aspect of this application provides a dynamic tumor segmentation method suitable for BNCT treatment, the method comprising: Acquire PET-CT multimodal data; The PET-CT multimodal data is input into the trained tumor segmentation model to obtain a tumor segmentation probability map based on the spatial location dimension and boron concentration distribution dimension of the tumor segmentation model. Dynamic threshold segmentation is performed on the tumor segmentation probability map based on the boron drug metabolism kinetics characteristics to obtain the tumor segmentation results.

[0007] In some implementations of the first aspect, acquiring PET-CT multimodal data includes: Acquire CT images and extract their spatial and density information as first CT modality data; The CT image is segmented into bones and soft tissues to generate a binary mask image, and the binary mask image is used as the second CT modality data; First PET modal data and second PET modal data are acquired, and the first PET modal data and second PET modal data are weighted and fused to obtain PET fused data; the first PET modal data reflects the initial distribution characteristics of boron drug, and the second PET modal data reflects the retention distribution characteristics of boron drug; The first CT modal data, the second CT modal data, and the PET fusion data are registered together to form the PET-CT multimodal data.

[0008] In some implementations of the first aspect, obtaining the tumor segmentation probability map based on the tumor segmentation model according to the spatial location dimension and the boron concentration distribution dimension includes: The PET-CT multimodal data is encoded using the encoder module of the tumor segmentation model to obtain spatial feature maps and concentration feature maps accordingly. The concentration feature map is spatially attention-weighted according to the boron-sensitive attention module of the tumor segmentation model to obtain boron concentration distribution features. The spatial feature map and the boron concentration distribution feature are fused into the decoder module of the tumor segmentation model according to the skip connection path of the tumor segmentation model, so as to obtain the tumor segmentation probability map according to the decoder module.

[0009] In some implementations of the first aspect, obtaining boron concentration distribution characteristics includes: Obtain the distribution gradient map of the PET fusion data; The distribution gradient map is normalized to obtain the spatial attention weight map; The concentration feature map is subjected to spatial attention weighting based on the spatial attention weight map to obtain the boron concentration distribution characteristics.

[0010] In some implementations of the first aspect, dynamic threshold segmentation of the tumor segmentation probability map based on boron drug metabolism kinetics characteristics to obtain tumor segmentation results includes: The tumor segmentation probability map is optimized for connected components, and the segmentation threshold is dynamically adjusted based on the standard uptake value of the second PET modality data. Based on the segmentation threshold, the tumor segmentation probability map is used to label the tumor target region, and the tumor segmentation result is obtained; wherein... When the standard uptake value of the second PET modality data is greater than a predetermined threshold, the segmentation threshold is dynamically reduced to expand the tumor target area; When the standard uptake value of the second PET modality data is less than a predetermined threshold, the segmentation threshold is dynamically increased to reduce the tumor target area.

[0011] In some implementations of the first aspect, the training method for the tumor segmentation model includes: Construct the first multimodal dataset for PET-CT; A first training set is read from the PET-CT multimodal first dataset to train the original segmentation model, and a first basic training loss and a first additional training loss are obtained. The parameters of the original segmentation model are jointly optimized based on the first basic training loss and the first additional training loss to obtain the trained tumor segmentation model.

[0012] Among some implementations of the first aspect, constructing the first PET-CT multimodal dataset includes: Multiple CT training images are acquired and their spatial and density information is extracted to serve as the first CT modality training data. Multiple CT training images are segmented into bones and soft tissues to generate multiple binarized mask training images, and these binarized mask training images are used as second CT modality training data. Acquire multiple first PET modality training data and multiple second PET modality training data, and perform standard uptake value correction on all first PET modality training data and second PET modality training data; Weighted fusion is performed on every two corresponding first PET modality training data and second PET modality training data to obtain multiple PET fused first training data; The first CT modality training data and the PET fusion first training data are jointly annotated to obtain a first structure file; After registering the first CT modality training data, the second CT modality training data, and the PET fusion first training data, they are used together with the first structure file as the first PET-CT multimodal dataset.

[0013] In some implementations of the first aspect, obtaining the first basic training loss and the first additional training loss includes: Obtain the first predicted tumor segmentation probability map output by the original segmentation model; A first boundary gradient map is obtained based on the first predicted tumor segmentation probability map; The first basic training loss is obtained based on the first predicted tumor segmentation probability map and the first structure file; The first additional training loss is obtained based on the first boundary gradient map and the distribution gradient map of the PET fused first training data.

[0014] A second aspect of this application provides a model training method, the method comprising: Construct a second multimodal dataset for PET-CT; The second training set is read from the PET-CT multimodal second dataset to train the original segmentation model, and the second basic training loss and the second additional training loss are obtained. The parameters of the original segmentation model are jointly optimized based on the second basic training loss and the second additional training loss to obtain a trained tumor segmentation model; the tumor segmentation model obtains a tumor segmentation probability map based on the spatial location dimension and the boron concentration distribution dimension.

[0015] In some implementations of the second aspect, obtaining the second basic training loss and the second additional training loss includes: Obtain the second predicted tumor segmentation probability map output by the original segmentation model; A second boundary gradient map is obtained based on the second predicted tumor segmentation probability map; The second basic training loss is obtained based on the second predicted tumor segmentation probability map and the second structure file; The second additional training loss is obtained based on the second boundary gradient map and the distribution gradient map of the second training data fused with PET.

[0016] As described above, the tumor dynamic segmentation method and model training method applicable to BNCT treatment described in this application have the following beneficial effects: First, this application innovatively integrates anatomical structure information and boron drug metabolism information into the tumor segmentation model by inputting PET-CT multimodal data into the model. This overcomes the limitations of traditional single-modal segmentation methods that rely solely on anatomical features or single functional indicators, and provides a more comprehensive biological basis for tumor target localization.

[0017] Secondly, this application innovatively proposes a boron-sensitive attention module, which guides the network to dynamically focus on high boron concentration areas by mining the correlation between PET delayed phase data and tumor boundaries. This significantly improves the ability to distinguish tumors from surrounding normal tissues and effectively reduces segmentation errors caused by uneven boron drug distribution. It performs particularly well in areas where traditional methods are prone to errors, such as tumor edges and invasive growth areas.

[0018] Furthermore, this application innovatively proposes a joint optimization of basic training loss and additional training loss. The model not only ensures the statistical accuracy of segmentation but also enhances the anatomical consistency between the segmentation results and the boron drug distribution, providing a more reliable imaging basis for tumor target localization and boron drug dosage planning in BNCT treatment.

[0019] Finally, this application innovatively combines the BNCT boron drug metabolism kinetics characteristics in the post-processing stage, and avoids the "oversegmentation" or "undersegmentation" problems of simple probability segmentation by optimizing connected components and dynamically adjusting the segmentation threshold. The generated segmentation results are closer to the actual distribution boundary of boron drugs, providing a more accurate basis for dose calculation for the treatment planning system (TPS).

[0020] Furthermore, this application achieves intelligent processing of the entire process from multimodal data preprocessing to clinically usable structural output. The operation process is highly standardized, and the segmentation efficiency is high (significantly shortening the time compared to traditional manual segmentation or single-modal segmentation methods). Moreover, the segmentation results combine anatomical accuracy and functional relevance, which can effectively assist clinicians in quickly determining the tumor target area (GTV, CTV) and the range of high-risk organs, providing a more reliable technical tool for personalized treatment of head and neck tumor patients. Attached Figure Description

[0021] Figure 1 A schematic flowchart of the dynamic tumor segmentation method for BNCT treatment described in the embodiments of this application is shown.

[0022] Figure 2 The diagram shown is a schematic representation of the process for acquiring PET-CT multimodal data as described in an embodiment of this application.

[0023] Figure 3 The diagram shown is a schematic representation of the process for obtaining a tumor segmentation probability map as described in an embodiment of this application.

[0024] Figure 4The diagram shown is a structural schematic of a tumor segmentation model according to an embodiment of this application.

[0025] Figure 5 The diagram shown is a schematic representation of the process for dynamically thresholding the tumor segmentation probability map based on the boron drug metabolism kinetics characteristics as described in an embodiment of this application.

[0026] Figure 6 The diagram shown is a flowchart illustrating the training method of the tumor segmentation model described in this application embodiment.

[0027] Figure 7 The diagram shows a flowchart of the process for constructing the first PET-CT multimodal dataset as described in an embodiment of this application.

[0028] Figure 8 The diagram shows a flowchart illustrating the process of obtaining the first basic training loss and the first additional training loss as described in an embodiment of this application.

[0029] Figure 9 The diagram shown is a flowchart illustrating the model training method described in the embodiments of this application. Detailed Implementation

[0030] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0031] It should be noted that in the following description, reference is made to the accompanying drawings, which illustrate several embodiments of this application. It should be understood that other embodiments may also be used, and changes in mechanical composition, structure, electrical system, and operation may be made without departing from the spirit and scope of this application. The following detailed description should not be considered limiting, and the scope of the embodiments of this application is defined only by the claims of the published patent. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application. Spatial terms such as “upper,” “lower,” “left,” “right,” “below,” “below,” “lower part,” “above,” “upper part,” etc., may be used herein to illustrate the relationship between one element or feature shown in the figures and another element or feature.

[0032] Furthermore, as used herein, the singular forms “a,” “an,” and “the” are intended to include the plural forms as well, unless the context indicates otherwise. It should be further understood that the terms “comprising,” “including,” indicate the presence of the stated feature, operation, element, component, item, kind, and / or group, but do not preclude the presence, occurrence, or addition of one or more other features, operations, elements, components, items, kinds, and / or groups. The terms “or” and “and / or” as used herein are to be interpreted inclusively, or mean any one or any combination thereof.

[0033] Before providing a further detailed description of this application, the nouns and terms used in the embodiments of this application are explained, and the nouns and terms used in the embodiments of this application shall be interpreted as follows: CT stands for Computed Tomography, which uses an X-ray beam to perform a rotating scan of a specific part of the human body and receives the attenuated rays through a detector. The computer reconstructs the acquired data into a cross-sectional anatomical image, completely eliminating the overlap interference of traditional X-ray images. Its images can clearly show details of different density structures such as bones, calcifications, soft tissues, and lungs.

[0034] MRI (Magnetic Resonance Imaging) uses a powerful external magnetic field and radio frequency pulses to cause the hydrogen nuclei in the human body to resonate and absorb energy. When the radio frequency pulse stops, the hydrogen atoms release energy and return to their original state; this process is captured by a receiver as a signal. Computer processing of these signals generates extremely high-resolution images of soft tissues, especially clearly showing the brain, spinal cord, muscles, and joints.

[0035] PET: Positron Emission Tomography is a nuclear medicine imaging technique that displays physiological metabolic functions by detecting the distribution of positron-emitting tracers in the body. Patients are injected with drugs containing radioactive isotopes, which accumulate in metabolically abnormal cells (such as tumor cells). The device detects the gamma rays released when the tracer decays and generates a three-dimensional functional image of the body's biochemical activities.

[0036] DICOM-RT: Digital Imaging and Communications in Medicine – Radiotherapy, is a series of extended objects and functional modules within the international standard DICOM protocol specifically designed for the field of radiotherapy. It defines how to store, transmit, and exchange radiotherapy-specific data, such as radiotherapy setups, doses, plans, and treatment records.

[0037] IN regularization: Instance Normalization, which standardizes the feature map of each channel of a single sample to eliminate style differences, stabilize training, and improve the model's generalization ability.

[0038] ReLU function: A commonly used activation function for neural networks. It maps all negative inputs to zero while keeping positive values ​​unchanged, thereby introducing nonlinearity and alleviating the gradient vanishing problem.

[0039] Sobel operator: The Sobel operator is a discrete differential operator used for image processing that quickly detects edges by calculating an approximation of the gradient of image brightness.

[0040] The softmax function is a function that transforms any real vector into a probability distribution. The sum of its output values ​​is 1, and it is often used in the output layer of multi-class classification tasks.

[0041] Dice Loss: A loss function based on the Sørensen-Dice coefficients, which measures the overlap between the predicted segment and the true label, and is often used to deal with class imbalance problems in image segmentation tasks.

[0042] Sørensen-Dice coefficient: A statistical indicator used to measure the similarity between two sample sets.

[0043] Focal Loss: Focal Loss is a loss function designed for class imbalance. By introducing an adjustable focusing parameter, it dynamically reduces the weight of easily classified samples, making the model more focused on training difficult-to-classify samples.

[0044] BNCT is a binary targeted radiotherapy technique based on nuclear reactions. With its unique advantages, it has stood out in the treatment of various types of cancer and is considered a "precision treatment weapon" for head and neck tumors (especially relapsed / refractory cases). Its principle is as follows: first, the patient is injected with a substance containing B... -10 Drugs that can be specifically absorbed by tumor cells (such as BPA or BSH) are selectively enriched in tumor cells, and then the affected area is irradiated with a neutron beam. -10After absorbing neutrons, a nuclear fission reaction occurs, producing α particles and Li ions with high linear energy transfer (LET). These particles have a very short range in tumor cells (only about 5-9 micrometers, which is comparable to the cell diameter), thus enabling precise killing of boron-rich tumor cells with almost no damage to surrounding normal tissues. BNCT avoids the drug resistance problems of chemotherapy, targeted therapy, and immunotherapy, and the efficacy of BNCT can be predicted by confirming the degree of aggregation of boron carriers in tumor cells in advance. Compared with traditional carbon ion therapy or photon radiotherapy, boron neutron capture therapy has at least the following unique advantages: (1) Ultra-high targeting: selective enrichment of tumors dependent on boron drugs, achieving dual precision of "biological targeting + physical targeting"; (2) Short treatment course: treatment can be completed with a single or a few neutron irradiations (traditional radiotherapy requires 20-30 times), significantly improving patient compliance; (3) Protection of normal tissues: the risk of damage to surrounding sensitive organs (such as salivary glands and brainstem) is extremely low, especially suitable for patients with recurrent tumors or contraindications to surgery.

[0045] In the field of medical image segmentation, traditional methods such as thresholding, region growing, and level set methods have long faced bottlenecks, including high reliance on manual intervention, limited ability to recognize complex boundaries, and insufficient generalization performance across datasets. This is especially true for head and neck tumors, whose complex anatomical structures, blurred and difficult-to-distinguish boundaries, and inconsistent resolution, contrast, and feature representation across different modalities make traditional segmentation techniques insufficient for the rapid clinical diagnosis and treatment needs of head and neck tumors (BNCT). With the rapid development of artificial intelligence, deep learning-based intelligent segmentation technology has become a key research direction for overcoming these limitations. It achieves intelligent processing from raw medical images to target region segmentation. This big data-driven learning paradigm enables deep learning to demonstrate unprecedented accuracy and stability in segmentation tasks involving head and neck tumors, which have complex anatomical structures and high morphological variability, providing reliable technical support for precision medicine.

[0046] However, current image segmentation techniques in head and neck tumor radiotherapy are mainly designed for traditional photon radiotherapy, and their core deficiency lies in not considering the unique "biological targeting" mechanism of boron-neutron tumor therapy (BNCT). Existing technologies generally construct segmentation models based on CT or MRI single-modal images, delineating target boundaries solely through tumor anatomical features (such as density differences or soft tissue contrast), while ignoring boron drug concentration information. This may result in segmentation results that cannot accurately match the treatment target requirements of BNCT, especially for marginal regions with insufficient boron concentration, affecting the accuracy of neutron irradiation.

[0047] Furthermore, existing neural network technologies for head and neck tumor image segmentation have not been optimized for the specific needs of BNCT. Although deep learning models such as U-Net and 3D U-Net perform well in single-modal segmentation of CT / MRI in traditional radiotherapy, these general models have significant adaptability issues when directly applied to BNCT scenarios: on the one hand, existing neural network architectures do not use PET images as a core input modality and lack specially designed network modules to extract and analyze boron concentration distribution features; on the other hand, network designs for multimodal (CT / MRI / PET) feature fusion are insufficient, especially in that they do not consider the collaborative learning mechanism of boron drug distribution and anatomical structure information, making it difficult for the network to simultaneously capture the morphological features and biological targeting features of the tumor.

[0048] Currently, neural network models in BNCT scenarios still face the challenge of few-shot learning. Due to the scarcity of specific imaging data on boron drug distribution, existing models struggle to maintain generalization ability under small-sample conditions and lack efficient feature extraction capabilities for complex head and neck anatomical structures, resulting in a gap between segmentation accuracy and clinical needs. More importantly, the segmentation results output by existing systems often lack boron concentration information, causing treatment planning systems to be unable to accurately calculate the actual tumor radiation dose and neutron flux distribution, thus limiting the realization of the precision advantages of BNCT.

[0049] To at least address the aforementioned technical issues, this application provides a tumor dynamic segmentation method and model training method suitable for BNCT treatment. Through multimodal data fusion, boron drug metabolism characteristic guidance, and clinically oriented post-processing optimization, it systematically solves the key problem of accurate tumor target area segmentation in head and neck tumor BNCT, providing intelligent imaging support for the precise implementation of BNCT.

[0050] Figure 1 A schematic flowchart of the dynamic tumor segmentation method for BNCT treatment described in the embodiments of this application is shown, as follows: Figure 1 As shown in the embodiments of this application, the tumor dynamic segmentation method for BNCT treatment includes steps S1 to S3.

[0051] S1. Acquire PET-CT multimodal data.

[0052] In some embodiments, this application provides multi-dimensional information support for subsequent tumor segmentation by acquiring PET-CT multimodal data. Figure 2 The diagram shows a flowchart of the process for acquiring PET-CT multimodal data as described in an embodiment of this application. Figure 2 As shown, acquiring PET-CT multimodal data includes steps S11 to S14.

[0053] S11. Acquire CT images and extract the spatial and density information of the CT images as first CT modality data.

[0054] In some embodiments, after acquiring a patient's CT image and extracting its spatial resolution, voxel spacing, and coordinate information, this is used as the first CT modality data. That is, the first CT modality data is essentially the original CT image that preserves the spatial location and density information of the anatomical structures. Here, density information refers to CT values. Furthermore, a windowing algorithm can be used to dynamically optimize the window width and window level parameters based on the grayscale distribution characteristics of the first CT modality data, focusing on enhancing the contrast between tumor tissue and surrounding normal tissue, and highlighting the detailed features of the lesion area.

[0055] S12. The CT image is segmented into bones and soft tissues to generate a binary mask image, and the binary mask image is used as the second CT modality data.

[0056] In some embodiments, for CT images, a threshold segmentation method is first used for preliminary tissue separation based on differences in Henlein Units (HU values). Typically, a threshold range (e.g., 100-3000 HU values) is set to extract bone structures, and a lower threshold range (e.g., -100-100 HU values) is used to distinguish soft tissues. Subsequently, morphological operations (e.g., erosion, dilation) are used to optimize the segmentation boundaries, eliminate noise and volume separation effects, and finally generate a binary mask image. In this image, bone and soft tissue regions are labeled as foreground (pixel value 255), and the remaining background regions are labeled as background (pixel value 0), serving as the second CT modality data. In practice, the second CT modality data can effectively suppress the interference of bone tissue on subsequent tumor segmentation.

[0057] S13. Obtain first PET modal data and second PET modal data, and perform weighted fusion of the first PET modal data and the second PET modal data to obtain PET fused data.

[0058] The first PET modal data reflects the initial distribution characteristics of boron drug, while the second PET modal data reflects the retention distribution characteristics of boron drug.

[0059] In some embodiments, the first PET modality data is the early-phase PET data, which refers to PET images 30 minutes after boron injection in the patient, reflecting the initial distribution characteristics of boron; while the second PET modality data is the delayed-phase PET data, which refers to PET images 60 minutes after boron injection in the patient, reflecting the retention characteristics of boron within the tumor. Acquiring dual-phase PET data allows for a comprehensive characterization of the heterogeneity of boron enrichment in tumors. For example, areas with high early uptake but rapid metabolism within 1 hour may indicate a risk of drug loss, while areas with low early uptake but continuous accumulation within 1 hour may conceal subclinical lesions not identified by traditional imaging. Therefore, acquiring both modalities effectively provides information on boron distribution within the patient.

[0060] Furthermore, since images at different time points provide different biological information, especially in the imaging of lesions such as tumors, this embodiment of the application performs a weighted summation of the first PET modal data and the second PET modal data to obtain a comprehensive PET fusion data, thereby effectively integrating biological information from different time points and thus more comprehensively reflecting the metabolic state of the tissue.

[0061] S14. The first CT modal data, the second CT modal data, and the PET fusion data are registered together to form the PET-CT multimodal data.

[0062] In some embodiments, when a threshold segmentation is performed on the original CT image to obtain a binary mask image, the original CT image and the binary mask image are already aligned, meaning that the first CT modal data and the second CT modal data have been registered.

[0063] In some embodiments, pixel values ​​of the PET fusion data are mapped to a standard range consistent with the first CT modality data to complete grayscale normalization, thereby ensuring dimensional consistency of data from different modalities in subsequent processes. Further, a rigid registration algorithm based on mutual information is employed, using the first CT modality data as a reference, to extract the grayscale histograms and gradient features of the first CT modality data and the PET fusion data, thereby calculating the initial mutual information value. Subsequently, the transformation parameters are iteratively adjusted using a gradient descent optimization algorithm to gradually increase the registered mutual information value until a preset convergence threshold is reached, thus accurately mapping the coordinates of each voxel in the PET fusion data to the coordinate system of the first CT modality data through a transformation matrix, thereby completing the registration operation. Finally, the registered first CT modality data, second CT modality data, and PET fusion data are used as PET-CT multimodal data, providing spatially consistent model input data for subsequent tumor segmentation and boron drug distribution analysis based on multimodal information.

[0064] S2. Input the PET-CT multimodal data into the trained tumor segmentation model to obtain a tumor segmentation probability map based on the spatial location dimension and boron concentration distribution dimension of the tumor segmentation model.

[0065] In some embodiments, PET-CT multimodal data is stitched together along the channel dimension and then input into a trained tumor segmentation model to provide multidimensional information support for tumor segmentation, including anatomical localization, tissue inhibition, and dynamic distribution of boron drugs. This enables the model to achieve precise segmentation of the tumor target area based on the input multimodal data, utilizing the spatial location dimension and the boron concentration distribution dimension. Figure 3 The diagram shown is a schematic representation of the process for obtaining a tumor segmentation probability map as described in an embodiment of this application. Figure 3 As shown, obtaining the tumor segmentation probability map includes steps S21 to S23.

[0066] S21. The PET-CT multimodal data is encoded according to the encoder module of the tumor segmentation model to obtain spatial feature maps and concentration feature maps accordingly.

[0067] S22. The concentration feature map is subjected to spatial attention weighting processing according to the boron-sensitive attention module of the tumor segmentation model to obtain boron concentration distribution features.

[0068] S23. The spatial feature map and the boron concentration distribution feature are fused into the decoder module of the tumor segmentation model according to the skip connection path of the tumor segmentation model, so as to obtain the tumor segmentation probability map according to the decoder module.

[0069] Figure 4 The diagram shown is a structural schematic of a tumor segmentation model according to an embodiment of this application. The following will be combined with... Figure 4 Steps S21 to S23 in the above embodiments will be explained. Please refer to [link / reference]. Figure 4 As shown, when PET-CT multimodal data is stitched together along the channel dimension and input into the tumor segmentation model, the model first convolves the input data using a 3x3x3 3D ​​convolution kernel, outputting a new tensor with more feature channels but unchanged spatial size. Then, this new tensor undergoes IN regularization and ReLU activation, and is processed using a 3x3x3 residual module before being input into the encoder path.

[0070] Please continue reading. Figure 4As shown, the encoder path undergoes five stages of contraction to obtain corresponding multi-scale spatial feature maps and density feature maps. After these five stages, highly condensed spatial feature maps and density feature maps are obtained at the bottom bottleneck layer. Each stage utilizes a 3x3x3 downsampling residual module and a 3x3x3 residual module for processing. Each 3x3x3 downsampling residual module includes a 1x1x1 convolutional layer, a 3x3x3 convolutional layer, and a 1x1x1 convolutional layer; each 3x3x3 residual module includes two 3x3x3 convolutional layers.

[0071] It should be noted that the spatial feature map and concentration feature map mentioned in step S21 actually include feature maps of different scales output at each stage in the encoder path. This application does not make any additional distinction between the feature maps output at each stage.

[0072] Please continue reading. Figure 4 As shown in the embodiment of this application, the tumor segmentation model includes skip connection paths. These skip connection paths are used to pass feature maps of the same scale from the encoder path to the decoder, allowing the decoder to focus on spatial details during decoding and thus mitigating the gradient vanishing problem.

[0073] Furthermore, in some embodiments, the skip connection path in the tumor segmentation model described in this application integrates a boron-sensitive attention module. This module is used to enhance the guiding role of boron drug distribution characteristics in PET fusion data on tumor segmentation. Please continue reading. Figure 4 As shown, the concentration feature map extracted at each stage of the encoder path will be passed to the boron-sensitive attention module through a skip connection path, so that the concentration feature map can be spatially attention-weighted to obtain the corresponding boron concentration distribution features.

[0074] In some embodiments, obtaining boron concentration distribution features includes: obtaining a distribution gradient map of the PET fusion data; normalizing the distribution gradient map to obtain a spatial attention weight map; and performing spatial attention weighting on the concentration feature map based on the spatial attention weight map to obtain the boron concentration distribution features. Specifically, since the PET fusion data is obtained by weighted summation of the first PET modality data and the second PET modality data, each pixel in the first PET modality data and the second PET modality data actually has a corresponding standard uptake value. After fusion, each pixel in the PET fusion data will have a response value, which actually reflects the boron drug concentration in each voxel. At this time, the boron-sensitive attention module first extracts the gradient magnitude and direction of the response value in three-dimensional space using the Sobel operator or a similar edge detection operator to obtain a gradient map. This gradient map intuitively reflects the distribution differences of boron drug concentration in tumor tissue (tumor edge areas often show higher gradient values ​​due to differences in boron drug penetration rates). Then, based on this gradient map, the boron-sensitive attention module generates a spatial attention weight map using a normalized exponential function. In this model, regions with larger gradient magnitudes (i.e., high boron concentrations or tumor boundary regions) have higher weight values, and vice versa. Finally, the spatial attention weight map is multiplied element-wise with the concentration feature map to dynamically weight the concentration feature map, thereby obtaining the corresponding boron concentration distribution characteristics.

[0075] Among them, the boron concentration distribution characteristics can guide the network to prioritize tumor regions associated with high boron concentrations during subsequent feature fusion and decoding processes, significantly improving the distinction between tumor target areas and background tissues.

[0076] It should be noted that, as Figure 4 As shown, each stage of the encoder generates a concentration feature map at one scale. The multi-scale concentration feature maps are input into the boron-sensitive attention module, and the corresponding multi-scale boron concentration distribution features are output according to the above process.

[0077] Finally, please continue reading Figure 4 As shown, the skip connection path transmits multi-scale spatial feature maps and multi-scale boron concentration distribution features to the decoder. The decoder then performs transposed convolutions along the decoder path based on contextual and detail information, ultimately outputting a tumor segmentation probability map through a 1x1x1 3D convolution kernel and a softmax function. The tumor segmentation probability map is a probability distribution map of the same size as the PET-CT multimodal data input to the model. The gray value of each voxel represents the probability that the location belongs to the tumor target area; a higher probability value indicates a stronger confidence in tumor attribution, meaning it is more likely to be a tumor target area.

[0078] S3. Based on the boron drug metabolism kinetics characteristics, the tumor segmentation probability map is dynamically thresholded to obtain the tumor segmentation results.

[0079] In some embodiments, this application performs post-processing on the tumor segmentation probability map output by the tumor segmentation model to obtain tumor segmentation results, thereby making the segmentation results more closely match the actual distribution boundary of boron drugs. Figure 5 The diagram shows a flowchart illustrating the dynamic threshold segmentation of the tumor segmentation probability map based on the boron drug metabolism kinetics characteristics described in this application embodiment. Figure 5 As shown, the dynamic threshold segmentation of the tumor segmentation probability map based on the boron drug metabolism kinetics characteristics includes steps S31 to S32.

[0080] S31. Optimize the connected components of the tumor segmentation probability map and dynamically adjust the segmentation threshold based on the standard uptake value of the second PET modality data.

[0081] S32. Based on the segmentation threshold, the tumor segmentation probability map is labeled with tumor target regions to obtain the tumor segmentation result.

[0082] Specifically, when the standard uptake value of the second PET modality data is greater than a predetermined threshold, the segmentation threshold is dynamically lowered to expand the tumor target area; when the standard uptake value of the second PET modality data is less than the predetermined threshold, the segmentation threshold is dynamically raised to shrink the tumor target area.

[0083] In some embodiments, for the tumor segmentation probability map output by the tumor segmentation model, three-dimensional connected component analysis technology is used to identify and remove noise by setting a voxel adjacency threshold. Specifically, the tumor segmentation probability map is viewed as a grid composed of voxels, and the connectivity between voxels is determined by defining 26-adjacency (or 6-adjacency, 18-adjacency) relationships. Subsequently, the image is traversed, and seed filling or equivalent labeling is performed on each foreground voxel, classifying the sets of voxels that are connected to each other into independent connected components. Finally, a volume threshold is set based on prior knowledge (e.g., only retaining regions with more than 100 voxels), and connected components with a volume smaller than the threshold are identified as noise and their voxel values ​​are moved from the foreground to the background, thereby achieving accurate identification and removal of noise.

[0084] Furthermore, the optimized tumor segmentation probability map is converted into a binary mask, and the segmentation threshold is dynamically adjusted using the standard uptake values ​​of the second PET modality data for segmentation. Specifically, since boron drugs have a longer retention time in tumors during BNCT, the standard uptake values ​​of the second PET modality data better reflect the actual boron drug uptake in tumor tissue. Therefore, a predetermined threshold is set for the second PET modality data. When the standard uptake value is greater than the predetermined threshold, the segmentation threshold is dynamically decreased; when the standard uptake value is less than the predetermined threshold, the segmentation threshold is dynamically increased.

[0085] In some embodiments, a segmentation threshold is set to 0.5, meaning that regions with a probability ≥ 0.5 in the binary mask are marked as tumor target areas. Simultaneously, a predetermined threshold is set to 2.5, allowing regions in the binary mask with a probability between 0.4 and 0.5 but a corresponding standard uptake value > 2.5 to also be marked as tumor target areas, thus ensuring the integrity of tumor boundaries. In other words, the segmentation confidence of regions is increased based on the standard uptake value of the second PET modality data, allowing regions that do not meet the segmentation threshold but have high standard uptake values ​​to be marked as tumor target areas and segmented. This avoids the "oversegmentation" or "undersegmentation" problems that may result from solely relying on probability values ​​in the tumor segmentation probability map. This is particularly suitable for scenarios where the SUV values ​​of the tumor overlap with those of surrounding normal tissue.

[0086] Furthermore, after labeling, a set of ROI (Region of Interest) structures conforming to the DICOM-RT standard is automatically generated based on a binary mask. Morphological operations (such as dilation and erosion) are used to correct boundary burrs, and anatomical prior knowledge (such as organ location and tumor growth patterns) is combined to perform topological verification of the structures. This ultimately generates a multi-level set of structures including GTV (Gross Tumor Volume), CTV (Clinical Target Volume), and high-risk organs (such as the spinal cord and kidneys). The definition of GTV directly relates to the high-value region of the standard uptake value in the second PET modality data (e.g., standard uptake value > 2.5). CTV is a safe boundary extended by 5-10 mm from the GTV (the extension range can be adjusted according to tumor invasion depth and boron drug diffusion characteristics). High-risk organs are generated by excluding normal tissue with low standard uptake values ​​(e.g., standard uptake value < 1.0) and referencing anatomical contours.

[0087] Ultimately, the RTSTRUCT file output based on the embodiments of this application strictly follows the DICOM standard, including the name, color, boundary coordinates and pixel value mapping relationship of each ROI, and can be directly imported into the treatment planning system (TPS) for dose planning and efficacy evaluation of BNCT.

[0088] Figure 6 The diagram shown is a flowchart illustrating the training method of the tumor segmentation model described in this application. Figure 6As shown, the training method for the tumor segmentation model includes steps S61 to S63.

[0089] S61. Construct the first PET-CT multimodal dataset.

[0090] In some embodiments, the tumor segmentation model described in this application needs to be trained so that it can accurately segment tumors based on input in practical applications. Figure 7 The diagram shows a flowchart illustrating the process of constructing the first multimodal PET-CT dataset as described in an embodiment of this application. Figure 7 As shown, constructing the first PET-CT multimodal dataset includes steps S611 to S616.

[0091] S611. Acquire multiple CT training images and extract the spatial and density information of the CT training images to serve as the first CT modality training data.

[0092] In some embodiments, enhanced CT images of multiple patients are acquired, and information such as spatial resolution, voxel spacing, and coordinates within the CT images are extracted using medical image processing tools to serve as CT training images. Furthermore, the CT training images from multiple patients may originate from different devices. In this case, to address the issue of uneven spatial resolution caused by differences in device slice thickness, the actual size of the CT training image can be used as a reference. Voxels of a fixed size are selected, and images with different voxel spacings are selected and resampled using interpolation methods to serve as the first CT modality training data. This ensures consistent resolution in all directions within the first CT modality training data.

[0093] Furthermore, windowing algorithms can be employed to dynamically optimize window width and level parameters based on the grayscale distribution characteristics of the first CT modality training data, focusing on enhancing the contrast between tumor tissue and surrounding normal tissue to highlight the detailed features of the lesion area. In addition, gantry artifact removal, noise reduction filtering, and contrast enhancement can be performed simultaneously to further improve the signal-to-noise ratio and clarity of the first CT modality training data. Finally, grayscale normalization is performed to eliminate the impact of grayscale differences between devices on model training.

[0094] S612. Perform bone and soft tissue segmentation on the multiple CT training images to generate multiple binarized mask training images, and use the multiple binarized mask training images as the second CT modality training data.

[0095] In some embodiments, for CT training images, after resampling, windowing, and enhancement operations are completed, a threshold segmentation method can be used to perform preliminary bone and soft tissue separation based on differences in Henle units (HU values). The bone and soft tissue regions are labeled as foreground (pixel value 255), and the remaining background regions are labeled as background (pixel value 0), thus serving as the second CT modality training data. Furthermore, the second CT modality training data can be corrected through manual annotation to reduce errors caused by blurred tissue boundaries.

[0096] S613. Acquire multiple first PET modality training data and multiple second PET modality training data, and perform standard uptake value correction on all first PET modality training data and second PET modality training data.

[0097] In some embodiments, patients with head and neck tumors (such as squamous cell carcinoma, adenoid cystic carcinoma, etc.) who have been pathologically diagnosed and meet the requirements for BNCT treatment are selected, and data are collected on their boron drug (e.g., boron injection) administered within 72 hours prior to treatment. 10 PET images were acquired 30 minutes and 60 minutes after B-BPA injection, respectively, to serve as the first and second PET modality training data. Subsequently, relevant data such as injection time, dosage, patient weight, and scanning parameters were read from both the first and second PET modality training data. The standard uptake values ​​(SUVs) of the two data sets were corrected using formula (1) to eliminate numerical fluctuations caused by differences in injection dosage, asynchronous scanning time, and patient body surface area, ensuring the comparability of PET data from different patients.

[0098] (1)

[0099] S614. Weighted fusion of every two corresponding first PET modality training data and second PET modality training data to obtain multiple PET fusion first training data.

[0100] Once calibration is complete, the first and second PET modal training data from the same patient are weighted and summed to obtain the PET fusion first training data. This fusion data comprehensively reflects the distribution information of boron drugs and the metabolic state of tissues.

[0101] S615. Jointly annotate the first CT modality training data and the PET fusion first training data to obtain the first structure file.

[0102] In some embodiments, the first CT modality training data and the PET fusion first training data are jointly annotated to generate a DICOM-RT structure file as the first structure file. This structure file is generated by fusing CT anatomical boundaries and PET boron concentration distribution, covering both the active core area of ​​the tumor and avoiding the inclusion of normal tissue with insufficient boron concentration, thus providing a clear supervision signal for tumor segmentation model training. Further, the contour and contour point coordinates of the region of interest in the DICOM-RT structure file are parsed, and the contour point coordinates are mapped to the pixel coordinate space of the first CT modality training data and rasterized to generate corresponding binarized image labels.

[0103] S616. After registering the first CT modality training data, the second CT modality training data, and the PET fusion first training data, they are used together with the first structure file as the first PET-CT multimodal dataset.

[0104] In some embodiments, since the second CT modality training data is acquired using threshold segmentation, it is already registered with the first CT modality training data. Furthermore, grayscale normalization and rigid registration are performed on the first CT modality training data and the PET fusion first training data to resolve the positional offset problem caused by differences in imaging principles (functional imaging and anatomical imaging) between the CT image modality and the PET image modality, providing spatially consistent input data for the model. Finally, the registered first CT modality training data, the second CT modality training data, the PET fusion first training data, and the first structure file are used as the first PET-CT multimodal dataset.

[0105] S62. Read the first training set from the PET-CT multimodal first dataset to train the original segmentation model and obtain the first basic training loss and the first additional training loss.

[0106] S63. Optimize the parameters of the original segmentation model based on the first basic training loss and the first additional training loss to obtain the trained tumor segmentation model.

[0107] In some embodiments, to ensure the effectiveness and generalization ability of model training, the first PET-CT multimodal dataset is divided into a first training set, a first validation set, and a first test set in a 7:2:1 ratio. The training set is mainly used for iterative optimization of model parameters; the validation set is used to adjust hyperparameters (such as learning rate and loss function weights) during training and to monitor whether the model is overfitting; the test set is only used for objective evaluation of the final model performance to ensure that the results are not affected by the training process. Specifically, the original segmentation model is trained by reading the first training set, that is, the training data in the first training set is input into the original segmentation model to obtain the first predicted tumor segmentation probability map, thereby obtaining the first basic training loss and the first additional training loss, and then jointly optimizing the original segmentation model based on the training loss. In fact, this application adopts a multi-task loss function collaborative optimization strategy to improve segmentation accuracy and clinical applicability.

[0108] Figure 8 The diagram shows a flowchart illustrating the process of obtaining the first basic training loss and the first additional training loss as described in an embodiment of this application. Figure 8 As shown, obtaining the first basic training loss and the first additional training loss includes steps S631 to S634.

[0109] S631. Obtain the first predicted tumor segmentation probability map output by the original segmentation model.

[0110] S632. Obtain a first boundary gradient map based on the first predicted tumor segmentation probability map.

[0111] S633. Obtain the first basic training loss based on the first predicted tumor segmentation probability map and the first structure file.

[0112] S634. Obtain the first additional training loss based on the first boundary gradient map and the distribution gradient map of the PET fused first training data.

[0113] In some embodiments, the base loss function is a combination of Dice Loss and Focal Loss, and the base training loss can be calculated based on Dice Loss and Focal Loss. The Dice Loss function obtains the intersection-over-union ratio (IoU) of the first predicted tumor segmentation probability map and the binarized image labels, thereby optimizing for the problems of small tumor target area proportion and positive-negative sample imbalance in medical images, and narrowing the gap between the first predicted tumor segmentation probability map and the binarized image labels. Meanwhile, Focal Loss reduces the loss weight of easily classified background regions, focusing the optimization on difficult-to-classify tumor boundary regions, further refining the segmentation contour.

[0114] In some embodiments, considering that the tumor target boundary should exhibit a high degree of alignment with the boron drug concentration gradient abrupt change region, this application introduces an additional training loss. Specifically, a first boundary gradient map is obtained based on the first predicted tumor segmentation probability map, and the cosine similarity between the first boundary gradient map and the distribution gradient map of the PET-fused first training data is used as the additional training loss. By minimizing the cosine similarity, the model is forced to learn a tumor boundary consistent with the boron drug distribution gradient, avoiding segmentation boundary shifts caused by solely relying on image grayscale. Therefore, through joint optimization of the basic loss and the additional loss, this application ensures both the statistical accuracy of segmentation and enhances the anatomical consistency between the segmentation results and the boron drug distribution, providing a more reliable imaging basis for tumor target localization and boron drug dosage planning in BNCT treatment.

[0115] This application also provides a model training method, which trains a tumor segmentation model that can obtain a tumor segmentation probability map based on the spatial location dimension and the boron concentration distribution dimension. Figure 9 The diagram shown is a flowchart illustrating the model training method described in the embodiments of this application. Figure 9 As shown, the model training method includes steps S91 to S93.

[0116] S91. Construct the second PET-CT multimodal dataset.

[0117] S92. Read the second training set from the PET-CT multimodal second dataset to train the original segmentation model and obtain the second basic training loss and the second additional training loss.

[0118] S93. Optimize the parameters of the original segmentation model based on the second basic training loss and the second additional training loss to obtain a trained tumor segmentation model.

[0119] Specifically, obtaining the second basic training loss and the second additional training loss includes: obtaining the second predicted tumor segmentation probability map output by the original segmentation model; obtaining the second boundary gradient map based on the second predicted tumor segmentation probability map; obtaining the second basic training loss based on the second predicted tumor segmentation probability map and the second structure file; and obtaining the second additional training loss based on the second boundary gradient map and the distribution gradient map of the PET fused second training data.

[0120] The specific steps of the model training method described in the embodiments of this application can be referred to the content of the above embodiments, and will not be repeated here.

[0121] Therefore, this application innovatively integrates anatomical structural information and boron drug metabolism information into the tumor segmentation model by inputting PET-CT multimodal data. This overcomes the limitations of traditional single-modal segmentation methods that rely solely on anatomical features or single functional indicators, providing a more comprehensive biological basis for tumor target localization. Secondly, this application innovatively proposes a boron-sensitive attention module. By mining the correlation between PET delayed-phase data and tumor boundaries, it guides the network to dynamically focus on high-boron-concentration areas, significantly improving the ability to distinguish tumors from surrounding normal tissues and effectively reducing segmentation errors caused by uneven boron drug distribution. It performs particularly well in areas prone to errors in traditional methods, such as tumor edges and invasive growth areas. Thirdly, this application innovatively proposes a joint optimization of basic training loss and additional training loss. The model ensures both the statistical accuracy of segmentation and enhances the anatomical consistency between segmentation results and boron drug distribution, providing a more reliable imaging basis for tumor target localization and boron drug dosage planning in BNCT treatment. Finally, this application innovatively combines the BNCT boron drug metabolism kinetics characteristics in the post-processing stage, and avoids the "oversegmentation" or "undersegmentation" problems of simple probability segmentation by optimizing connected components and dynamically adjusting the segmentation threshold. The generated segmentation results are closer to the actual distribution boundary of boron drugs, providing a more accurate basis for dose calculation for the treatment planning system.

[0122] The scope of protection for the tumor dynamic segmentation method and model training method applicable to BNCT treatment in this application is not limited to the execution order of the steps listed in this embodiment. Any solution implemented by adding, subtracting, or replacing steps in the prior art based on the principles of this application is included within the scope of protection of this application.

[0123] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0124] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. A dynamic tumor segmentation method suitable for BNCT treatment, characterized in that, The method includes: Acquire PET-CT multimodal data; The PET-CT multimodal data is input into the trained tumor segmentation model to obtain a tumor segmentation probability map based on the spatial location dimension and boron concentration distribution dimension of the tumor segmentation model. Dynamic threshold segmentation is performed on the tumor segmentation probability map based on the boron drug metabolism kinetics characteristics to obtain the tumor segmentation results.

2. The dynamic tumor segmentation method for BNCT treatment according to claim 1, characterized in that, Acquiring PET-CT multimodal data includes: Acquire CT images and extract their spatial and density information as first CT modality data; The CT image is segmented into bones and soft tissues to generate a binary mask image, and the binary mask image is used as the second CT modality data; First PET modal data and second PET modal data are acquired, and the first PET modal data and second PET modal data are weighted and fused to obtain PET fused data; the first PET modal data reflects the initial distribution characteristics of boron drug, and the second PET modal data reflects the retention distribution characteristics of boron drug; The first CT modal data, the second CT modal data, and the PET fusion data are registered together to form the PET-CT multimodal data.

3. The dynamic tumor segmentation method for BNCT treatment according to claim 1, characterized in that, The tumor segmentation probability map obtained based on the tumor segmentation model, using both spatial location and boron concentration distribution dimensions, includes: The PET-CT multimodal data is encoded using the encoder module of the tumor segmentation model to obtain spatial feature maps and concentration feature maps accordingly. The concentration feature map is spatially attention-weighted according to the boron-sensitive attention module of the tumor segmentation model to obtain boron concentration distribution features. The spatial feature map and the boron concentration distribution feature are fused into the decoder module of the tumor segmentation model according to the skip connection path of the tumor segmentation model, so as to obtain the tumor segmentation probability map according to the decoder module.

4. The dynamic tumor segmentation method for BNCT treatment according to claim 3, characterized in that, Obtaining boron concentration distribution characteristics includes: Obtain the distribution gradient map of the PET fusion data; The distribution gradient map is normalized to obtain the spatial attention weight map; The concentration feature map is subjected to spatial attention weighting based on the spatial attention weight map to obtain the boron concentration distribution characteristics.

5. The dynamic tumor segmentation method for BNCT treatment according to claim 2, characterized in that, Dynamic threshold segmentation of the tumor segmentation probability map based on boron drug metabolism kinetics characteristics is performed to obtain tumor segmentation results, including: The tumor segmentation probability map is optimized for connected components, and the segmentation threshold is dynamically adjusted based on the standard uptake value of the second PET modality data. Based on the segmentation threshold, the tumor segmentation probability map is used to label the tumor target region to obtain the tumor segmentation result; wherein... When the standard uptake value of the second PET modality data is greater than a predetermined threshold, the segmentation threshold is dynamically reduced to expand the tumor target area; When the standard uptake value of the second PET modality data is less than a predetermined threshold, the segmentation threshold is dynamically increased to reduce the tumor target area.

6. The dynamic tumor segmentation method for BNCT treatment according to claim 1, characterized in that, The training method for the tumor segmentation model includes: Construct the first multimodal dataset for PET-CT; A first training set is read from the PET-CT multimodal first dataset to train the original segmentation model, and a first basic training loss and a first additional training loss are obtained. The parameters of the original segmentation model are jointly optimized based on the first basic training loss and the first additional training loss to obtain the trained tumor segmentation model.

7. The dynamic tumor segmentation method for BNCT treatment according to claim 6, characterized in that, The construction of the first PET-CT multimodal dataset includes: Multiple CT training images are acquired and their spatial and density information is extracted to serve as the first CT modality training data. Multiple CT training images are segmented into bones and soft tissues to generate multiple binarized mask training images, and these binarized mask training images are used as second CT modality training data. Acquire multiple first PET modality training data and multiple second PET modality training data, and perform standard uptake value correction on all first PET modality training data and second PET modality training data; Weighted fusion is performed on every two corresponding first PET modality training data and second PET modality training data to obtain multiple PET fused first training data; The first CT modality training data and the PET fusion first training data are jointly annotated to obtain a first structure file; After registering the first CT modality training data, the second CT modality training data, and the PET fusion first training data, they are used together with the first structure file as the first PET-CT multimodal dataset.

8. The dynamic tumor segmentation method for BNCT treatment according to claim 6, characterized in that, Obtaining the first base training loss and the first additional training loss includes: Obtain the first predicted tumor segmentation probability map output by the original segmentation model; A first boundary gradient map is obtained based on the first predicted tumor segmentation probability map; The first basic training loss is obtained based on the first predicted tumor segmentation probability map and the first structure file; The first additional training loss is obtained based on the first boundary gradient map and the distribution gradient map of the PET fused first training data.

9. A model training method, characterized in that, The method includes: Construct a second multimodal dataset for PET-CT; The second training set is read from the PET-CT multimodal second dataset to train the original segmentation model, and the second basic training loss and the second additional training loss are obtained. The parameters of the original segmentation model are jointly optimized based on the second basic training loss and the second additional training loss to obtain a trained tumor segmentation model; the tumor segmentation model obtains a tumor segmentation probability map based on the spatial location dimension and the boron concentration distribution dimension.

10. The model training method according to claim 9, characterized in that, Obtaining the second base training loss and the second additional training loss includes: Obtain the second predicted tumor segmentation probability map output by the original segmentation model; A second boundary gradient map is obtained based on the second predicted tumor segmentation probability map; The second basic training loss is obtained based on the second predicted tumor segmentation probability map and the second structure file; The second additional training loss is obtained based on the second boundary gradient map and the distribution gradient map of the second training data fused with PET.

Images