A method and system for personalized sub-target region radiotherapy planning of brain glioma

Abstract

This invention discloses a method and system for personalized radiotherapy planning of subtarget areas in gliomas. The method includes: inputting the target T1C image and its corresponding AAL image into a segmentation model to obtain a brain region segmentation mask file; inputting the target T1C image into a mapping model to obtain a pseudo-MRS image; inputting the target T1C image and the registered glioma annotation file into a habitat model to obtain a habitat atlas; based on the brain region segmentation mask file, identifying whether the habitat subregion is located in a functionally critical brain region to determine whether its safe dose threshold is set to the specific tolerable dose corresponding to the functionally critical brain region; and determining the radiation dose of the glioma subregion based on the target T1C image, ASL, pseudo-MRS image, and safe dose threshold. This effectively improves the stability, repeatability, and efficiency of target delineation and achieves biological conformity of radiotherapy dose, thus providing an objective and accurate quantitative reference for developing radiotherapy plans.

Description

Technical Field

[0001] This invention relates to the field of radiotherapy technology, and in particular to a method and system for personalized radiotherapy planning for subtargeted areas of glioma. Background Technology

[0002] Gliomas are the most common refractory malignant tumors in the brain. Radiotherapy, with its non-invasive nature and ability to precisely target and kill tumor cells, has become one of the important methods recommended internationally for the treatment of gliomas and occupies a key position in the comprehensive treatment system for gliomas.

[0003] Tumor target delineation is a crucial step in the effective implementation of radiotherapy, and its accuracy directly affects the radiotherapy outcome. However, current radiotherapy target delineation mainly relies on manual operation by professional physicians, which has significant shortcomings. On the one hand, manual delineation is labor-intensive and extremely time-consuming. Physicians need to carefully study the patient's multimodal imaging data, such as MRI (Magnetic Resonance Imaging) and CT (Computed Tomography) scans, to accurately determine the tumor boundaries. A single delineation often takes 1-2 hours, and for patients with complex conditions, it may take even longer. This not only increases the workload of physicians but also limits the number of patients that can be treated per unit of time, affecting the rational use of medical resources.

[0004] On the other hand, manual delineation has poor reproducibility among different doctors. Different doctors, due to differences in experience, professional background, and interpretation of imaging data, exhibit subjectivity in delineating target areas. This difference leads to a lack of stability in radiotherapy planning and implementation, resulting in inconsistent treatment outcomes.

[0005] In addition, the radiation dose inside the tumor in current conventional radiotherapy plans is relatively fixed, and it does not fully take into account the intermingling of glioma and normal brain tissue, the different cell proliferation activities in different regions of the tumor, and the strict dose tolerance limits of key neural structures, making it difficult to achieve biological conformity of radiotherapy dose.

[0006] Therefore, improving the stability of target delineation and achieving biological conformity of radiotherapy doses have become key issues that urgently need to be addressed to improve the efficacy of radiotherapy for gliomas. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for personalized radiotherapy planning of subtarget areas in gliomas, which can effectively improve the stability, repeatability and efficiency of target delineation, and achieve biological conformity of radiotherapy dose, thereby providing objective and accurate quantitative reference for the formulation of radiotherapy plans.

[0008] The objective of this invention is achieved through the following technical solution:

[0009] A personalized radiotherapy planning method for subtargeted areas of glioma includes:

[0010] The brain region's T1C image and corresponding AAL image are input into a segmentation model trained by a convolutional neural network to obtain a brain region segmentation mask file.

[0011] The T1C image to be tested is input into a mapping model trained by a mapping deep learning algorithm to obtain a pseudo MRS image.

[0012] The T1C image to be tested and the registered glioma annotation file are input into the habitat model trained by the habitat algorithm to obtain the habitat map;

[0013] Using the brain region segmentation mask file, it is identified whether the habitat subregion of the glioma is located in a functional critical brain region; if so, the safe dose threshold for the habitat subregion is configured to the specific tolerable dose corresponding to the functional critical brain region; if not, the safe dose threshold for the habitat subregion is configured to the conventional tolerable dose.

[0014] Obtain the ASL corresponding to the T1C image to be tested; determine the radiation dose of the glioma subregion based on the T1C image to be tested, the ASL, the pseudo-MRS image and the safe dose threshold.

[0015] Optionally, the step of determining the radiation dose of the habitat subregion based on the T1C image to be measured, the ASL, and the pseudo-MRS image includes:

[0016] The T1C image to be tested, the ASL and the pseudo MRS image are registered to obtain the MRS metabolite function reference value and the ASL metabolite function reference value for each habitat subregion.

[0017] The functional reference values ​​of the MRS metabolites and the functional reference values ​​of the ASL metabolites are weighted and fused to obtain the corresponding malignancy index;

[0018] The radiation dose of the habitat sub-region is configured to be positively correlated with the malignancy index, with the safe dose threshold of the habitat sub-region as the upper limit.

[0019] Optionally, before the step of inputting the T1C image of the brain region to be tested and the corresponding AAL image into a segmentation model trained by a convolutional neural network to obtain a brain region segmentation mask file, the method further includes:

[0020] The T1C image registered with the AAL image is resampled to obtain a high-resolution image and a low-resolution image;

[0021] The high-resolution image and the low-resolution image are normalized respectively, and corresponding high-resolution image blocks and low-resolution image blocks are cropped out.

[0022] The high-resolution image patch is input into a first convolutional neural network for training. After iterating through a loss function, a fine segmentation network model file of the segmentation model is obtained. The low-resolution image patch is input into a second convolutional neural network for training. After iterating through a loss function, a coarse segmentation network model file of the segmentation model is obtained. The segmentation model is composed of the cascaded fine segmentation network model file and the coarse segmentation network model file.

[0023] Alternatively, the loss function may be the Dice loss function.

[0024] Optionally, the mapping deep learning algorithm may employ any one of the following: GAN network, diffusion model, CNN network, or Transformer network.

[0025] Optionally, the mapping deep learning algorithm employs a GAN network;

[0026] Before the step of inputting the T1C image to be tested into a mapping model trained by a mapping deep learning algorithm to obtain a pseudo MRS image, the method further includes:

[0027] The registered T1C image and MRS image pair dataset is input into a generative adversarial network. After iterating through a total loss function consisting of joint adversarial loss, pixel-level loss, and perceptual loss, the mapping model that can output the pseudo MRS image is generated.

[0028] Optionally, the habitat algorithm employs at least one of K-means, GMM, Otsu, SLIC, MSI matrix, and ITHscore.

[0029] Optionally, the habitat algorithm is a superpixel segmentation algorithm;

[0030] Before the step of inputting the T1C image to be tested and the corresponding AAL image into the habitat model trained by the habitat algorithm to obtain the habitat map, the method further includes:

[0031] Input the registered T1C image and the registered glioma annotation file, use the superpixel segmentation algorithm to divide the brain tumor region into multiple superpixel units, and obtain the image features of the superpixel units;

[0032] Based on the image features, superpixel units are grouped to determine habitat subregions and generate subregion label masks.

[0033] By learning the mapping relationship from the image features to the sub-region labels, the habitat analysis model file of the habitat model is obtained.

[0034] Optionally, the convolutional neural network is any one of VB-Net network, U-Net network, variants of U-Net network, nnUNet network, or a hybrid neural network based on CNN and Transformer architecture.

[0035] A personalized radiotherapy planning system for subtargeted areas of glioma, comprising:

[0036] The segmentation mask generation module is used to input the T1C image of the brain region to be tested and the corresponding AAL image into the segmentation model trained by the convolutional neural network to obtain the brain region segmentation mask file.

[0037] The pseudo-MRS image generation module is used to input the T1C image to be tested into a mapping model trained by a mapping deep learning algorithm to obtain a pseudo-MRS image.

[0038] The habitat map generation module is used to input the T1C image to be tested and the registered glioma annotation file into the habitat model trained by the habitat algorithm to obtain the habitat map;

[0039] The first dose carving prescription module identifies whether the habitat subregion of the glioma is located in a functional critical brain region through the brain region segmentation mask file; if so, the safe dose threshold of the habitat subregion is configured to the specific tolerable dose corresponding to the functional critical brain region; if not, the safe dose threshold of the habitat subregion is configured to the conventional tolerable dose.

[0040] The second dose carving prescription module is used to acquire the ASL corresponding to the T1C image to be tested; and to determine the radiation dose of the glioma subregion based on the T1C image to be tested, the ASL, the pseudo-MRS image and the safe dose threshold.

[0041] Compared with existing technologies, the beneficial effects of this invention include at least the following: It can efficiently and accurately predict the distribution of metabolites reflecting tumor heterogeneity using only conventional T1C images, enabling the development of dose-sculpting prescriptions that integrate anatomy, function, and metabolism. It can effectively improve the stability, repeatability, and efficiency of target delineation, and achieve biological conformity of radiotherapy doses, thereby providing objective and precise quantitative reference for developing radiotherapy plans. Attached Figure Description

[0042] Figure 1 This is a flowchart of a personalized radiotherapy planning method for subtargeted areas of glioma according to an embodiment of the present invention;

[0043] Figure 2 This is a schematic diagram of the training and testing of the segmentation model according to an embodiment of the present invention;

[0044] Figure 3 This is a schematic diagram of the training and testing of the mapping model according to an embodiment of the present invention;

[0045] Figure 4 This is a schematic diagram of the training and testing of the habitat model according to an embodiment of the present invention;

[0046] Figure 5 This is a schematic diagram illustrating the formulation of a dosage-carving prescription according to an embodiment of the present invention;

[0047] Figure 6 This is a structural block diagram of a personalized radiotherapy planning system for subtargeted areas of glioma, according to an embodiment of the present invention. Detailed Implementation

[0048] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the embodiments set forth herein; rather, they are provided to make the invention more comprehensive and complete, and to fully convey the concept of the exemplary embodiments to those skilled in the art. The same reference numerals in the drawings denote the same or similar structures, and therefore repeated descriptions of them will be omitted.

[0049] The terms used to express position and direction in this invention are illustrated with reference to the accompanying drawings, but changes can be made as needed, and all such changes are included within the scope of protection of this invention.

[0050] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0051] First, it should be noted that current radiotherapy target delineation mainly relies on manual operation by professional physicians, which has significant shortcomings. Not only is manual delineation labor-intensive and extremely time-consuming, but it also suffers from poor repeatability among different physicians. Furthermore, it is difficult to achieve biological conformity of radiotherapy doses.

[0052] Among related technologies, MRS (Magnetic Resonance Spectroscopy) can detect the concentration of metabolites in brain tissue, among which a significant increase in Cho (choline) concentration is a key indicator for assessing the proliferative activity of tumor cells. However, MRS sequences have inherent limitations such as long scan time and complex operation and analysis, which makes it difficult to use as a routine examination in clinical radiotherapy planning and limits the accurate delineation of target areas based on metabolic information.

[0053] T1C (T1-weighted contrast-enhanced scan) is an indispensable standardized sequence in the diagnosis of gliomas and the delineation of radiotherapy target areas.

[0054] 3D-ASL (3D-Arterial Spin Labeling), a mature magnetic resonance perfusion technique, is now widely used in clinical practice. This technique allows for the quantitative measurement of cerebral blood flow (CBF) without the need for contrast agent injection, and the value directly reflects the metabolic activity of local tissues. In high-grade gliomas, abnormally elevated CBF "hot zones" typically correspond to the most aggressive subregions of the tumor, providing crucial biological evidence for precise dose sculpting. However, research on automatically and systematically integrating CBF information into radiotherapy planning optimization processes remains relatively limited, and its clinical translation and application potential requires further exploration and validation.

[0055] Based on this, the present invention provides a method and system for personalized radiotherapy planning of subtarget areas in gliomas. Even with only conventional T1C images, it can efficiently and accurately predict the distribution of metabolites reflecting tumor heterogeneity. This effectively improves the stability, repeatability, and efficiency of target delineation and achieves biological conformity of radiotherapy dose, thereby providing objective and precise quantitative reference for developing radiotherapy plans.

[0056] Figure 1 This is a flowchart of a personalized radiotherapy planning method for subtargeted areas of glioma according to an embodiment of this application. Figures 2-5 This is a schematic diagram illustrating the specific process of this method. The following will refer to... Figures 1-5 This application provides a detailed description of the personalized radiotherapy planning method for subtargeted areas of glioma.

[0057] In step 101, the T1C image of the brain region to be tested and the corresponding AAL image are input into the segmentation model generated by the convolutional neural network to obtain the brain region segmentation mask file.

[0058] Specifically, a convolutional neural network can be any of the following: U-Net (U-shaped network), U-Net variant, VB-Net, nnUNet (no new network), or a hybrid neural network based on CNN (Convolutional Neural Network) and Transformer architecture.

[0059] In this embodiment of the application, before step 101, the method further includes: resampling the T1C image registered with the AAL (Automated Anatomical Labeling) brain template annotation image to obtain a high-resolution image and a low-resolution image. The high-resolution image and the low-resolution image are then normalized, and corresponding high-resolution image patches and low-resolution image patches are cropped. The high-resolution image patches are input into a first convolutional neural network for training, and after iterative processing using a loss function, a fine-segmentation network model file is obtained; the low-resolution image patches are input into a second convolutional neural network for training, and after iterative processing using a loss function, a coarse-segmentation network model file is obtained.

[0060] In other words, the T1C brain region segmentation model is trained first. Specifically, the input raw body data images can be resampled into two different resolution images, such as low resolution [3mm, 3mm, 3mm] and high resolution [0.3mm, 0.3mm, 0.3mm], and the high resolution and low resolution images are normalized respectively (e.g., the pixel values ​​are normalized to 0-1024).

[0061] Then, the high-resolution and low-resolution images are randomly cropped into fixed-size image patches. The low-resolution image patches are then fed into a first convolutional neural network (coarse segmentation network, image resolution [3mm, 3mm, 3mm]) for training, while the high-resolution image patches are fed into a second convolutional neural network (fine segmentation network, image resolution [0.3mm, 0.3mm, 0.3mm]) for training. Both the coarse and fine segmentation networks use the VB-Net architecture and are iteratively optimized using the Dice loss function until the loss value falls below a preset threshold, generating coarse and fine segmentation model files respectively.

[0062] In this embodiment, a multi-scale cascaded VB-Net network model can be used. The VB-Net network improves segmentation accuracy by fusing multi-scale features through skip connections; multi-scale cascading refers to cascading a coarse segmentation network model and a fine segmentation network model. The coarse segmentation network model is used for general organ localization, while the fine segmentation network model is used for precise segmentation of organ edges. It is understood that a single network or multiple networks can also be used for organ segmentation, and this invention does not impose specific limitations on this.

[0063] It is also understood that this embodiment uses a three-dimensional convolutional neural network, but the present invention is not limited to this and is applicable to both two-dimensional and three-dimensional spaces.

[0064] The reasoning process of this segmentation model is as follows: The T1C image to be tested is input into the deep learning segmentation algorithm. The segmentation algorithm first resamples the image to a specified resolution image. After normalization, the image is input into the multi-scale cascaded segmentation model, which automatically segments the brain region segmentation mask file of the AAL brain template and marks each of them with labels from 1 to 112, thus obtaining the brain region segmentation mask file.

[0065] The same method was used for training and testing the segmentation model for glioma lesions. After training, a network model file for the glioma was obtained. During inference, the T1C image was input into the glioma network model file, and the segmentation mask file for the glioma lesions was automatically generated.

[0066] In specific examples, during independent testing, the segmentation results predicted by the model are compared with the AAL annotations. Metrics such as the Dice coefficient and Hausdorff distance can be calculated to quantify segmentation performance. For instance, a Dice coefficient of 0.9 for the hippocampus indicates that the model achieves 90% segmentation accuracy for that region.

[0067] In step 102, the T1C image to be tested is input into the mapping model trained by the mapping deep learning algorithm to obtain a pseudo MRS image.

[0068] Specifically, this study innovatively proposes an image mapping model designed to learn the end-to-end conversion relationship from T1C images to MRS images, generating high-fidelity "pseudo-MRS" images. This transforms the scarce resource of MRS metabolic information—characterized by long scan times and complex analysis—into a standardized output based on conventional T1C images, enabling the widespread adoption of precision radiotherapy. In this "pseudo-MRS" image, grayscale values ​​represent the concentration of metabolites in brain tissue. Cho concentration is calculated using a formula, and a significant increase in Cho concentration is a key indicator for assessing tumor cell proliferation activity, laying the foundation for subsequent dose sculpting.

[0069] It should be noted that the mapping deep learning algorithm can be any of the following: GAN (Generative Adversarial Network), diffusion model, CNN (Convolutional Neural Network), or Transformer network.

[0070] In this embodiment of the application, the mapping deep learning algorithm adopts a GAN network; before step 102, the method further includes: inputting the registered T1C image and MRS image pair dataset into the GAN network, and generating a mapping model that can output pseudo MRS images through the total loss function of joint adversarial loss, pixel-level loss (L1 loss) and perceptual loss.

[0071] In other words, the mapping model for MRS images is trained before application. During training, pre-registered T1C-MRS image pairs are used as training samples. The network architecture is based on a Generative Adversarial Network (GAN) to build the mapping model, including a generator and a discriminator. In specific examples, the generator can use a U-Net or ResNet structure to learn the complex nonlinear pixel mapping relationship from T1C to MRS; the discriminator can use a PatchGAN structure to distinguish between generated pseudo-MRS images and real MRS images.

[0072] In this embodiment, a combined loss function comprising adversarial loss, pixel-level loss, and perceptual loss is used for joint optimization. The adversarial loss improves the global realism of the pseudo-MRS image through a game between the generator and the discriminator; the pixel-level loss calculates the pixel-wise absolute error between the generated and real images, preserving low-frequency structural information; and the perceptual loss constrains the high-frequency texture details of the generated image based on the feature layer differences of the pre-trained VGG network. The weight coefficients of the three losses balance the contributions of each loss term.

[0073] In step 103, the T1C image to be tested and the registered glioma annotation file are input into the habitat model generated by the habitat algorithm to obtain the habitat map.

[0074] The glioma annotation file is used to accurately identify key information such as the location and extent of gliomas in an image. In specific examples, it may include at least one of the following: tumor region outline, tumor category information, and tumor subregion information.

[0075] The habitat algorithm can be one of the following: K-means (K-means Clustering), GMM (Gaussian Mixture Model), Otsu (Otsu's Method), SLIC (Simple Linear Iterative Clustering), MSI (Multi-Spectral Imaging Matrix), or ITHscore (Intratumoral Heterogeneity Score), or a combination of at least two of them.

[0076] In this embodiment, the habitat algorithm is a superpixel segmentation algorithm. Before step 103, the method further includes: inputting a registered T1C image and a registered glioma annotation file; using a superpixel segmentation algorithm to divide the brain tumor region into multiple superpixel units and obtaining the image features of the superpixel units; grouping the superpixel units based on the image features to determine habitat subregions and generating subregion label masks; learning the mapping relationship from image features to subregion labels to obtain the habitat analysis model file of the habitat model.

[0077] In other words, a superpixel segmentation algorithm is used to analyze the tumor region of each sample, dividing it into multiple units with uniform internal features. The training process will produce a habitat model that learns and solidifies the mapping relationship from T1C image features to the division of tumor internal subregions.

[0078] The inference process of this habitat model is as follows: For a trained habitat analysis model, paired T1C images and registered glioma annotation files are input, and a habitat atlas is automatically output. This habitat atlas is presented in the form of a segmentation mask, where each voxel is assigned a label identifying its corresponding habitat subregion with specific imaging characteristics. This allows for a quantitative representation of tumor internal heterogeneity, providing a direct basis for subsequent dose sculpting.

[0079] In step 104, the habitat subregion of the glioma is identified by using a brain region segmentation mask file to determine whether it is located in a functional critical brain region. If so, the safe dose threshold for the habitat subregion is configured to the specific tolerable dose corresponding to the functional critical brain region. If not, the safe dose threshold for the habitat subregion is configured to the conventional tolerable dose.

[0080] It should be noted that the functional critical brain region refers to the area where the hippocampus, optic nerve, pituitary gland, or other key parts are located, or their immediate surrounding areas; being located in the functional critical brain region means that the habitat subregion of the glioma at least partially overlaps with the functional critical brain region.

[0081] Specifically, if a habitat subregion is located in or near a critical organ, the tolerance dose of that organ is set as the upper limit of the dose for that habitat subregion, serving as a safety red line for the dose. In other words, a hard constraint is applied to ensure that the radiation dose for that habitat subregion in subsequent optimizations is lower than the tolerance dose of the organ.

[0082] In step 105, the ASL (Arterial spin labeling) corresponding to the T1C image to be tested is obtained; and the radiation dose of the glioma subregion is determined based on the T1C image to be tested, the ASL, the pseudo-MRS image and the safe dose threshold.

[0083] In this embodiment of the application, step 105 specifically includes the following sub-steps:

[0084] Step 1051: Register the T1C image, ASL and pseudo-MRS image to be tested, and obtain the MRS metabolite function reference value and ASL metabolite function reference value for each habitat subregion.

[0085] The functional reference values ​​for MRS metabolites include, but are not limited to, one or a combination of Cho, NAA (acetic acid), Cr (creatine), Lip (lipids), and Lac (lactic acid); the functional reference values ​​for ASL metabolites include, but are not limited to, one or a combination of CBF (cerebral blood flow), ATT (arterial transit time), CBV (cerebral blood volume), and VEGF (vascular endothelial growth factor).

[0086] In a specific example, the CBF value is obtained as the functional reference value for the above MRS metabolites, and the Cho value is obtained as the functional reference value for the above ASL metabolites.

[0087] Step 1052: Weighted fusion of MRS metabolite functional reference values ​​and ASL metabolite functional reference values ​​to obtain the corresponding malignancy index.

[0088] Step 1053: Using the safe dose threshold of the habitat sub-region as the upper limit, the radiation dose of the habitat sub-region is configured in a way that is positively correlated with the malignancy index.

[0089] In other words, if the habitat subregion is located in a functionally critical brain region, the upper limit is the specific tolerable dose corresponding to that functionally critical brain region; if the habitat subregion is not located in a functionally critical brain region, the upper limit is the conventional tolerable dose. Furthermore, a higher malignancy index indicates more active cell proliferation in the habitat subregion, resulting in higher malignancy, and the system will allocate a higher radiation dose within safe limits. Conversely, for subregions with a lower malignancy index, a lower baseline dose is automatically configured. This establishes a dose-sculpting prescription that integrates anatomy, function, and metabolism. In a specific example, this positive correlation configuration can be a proportional allocation.

[0090] like Figure 5 As shown, in a specific example, for different brain regions (BR1-BR14, ...), there are one-to-one matching of cerebral blood flow (CBF1-CBF14, ...) and choline complex (Cho1-Cho14, ...), and the corresponding radiation doses (TD1-TD14, ...) for priority decision under the constraints are obtained.

[0091] In this specific example, the model training and testing are performed using separate training and testing datasets. During the training phase, a neural network model file is generated, containing numerous parameters obtained through deep learning algorithms. During the testing phase, the model obtained during training is first used to test images. The file generated by the model is then used to generate tumor habitat subregions and make anatomical structure decisions.

[0092] In summary, the personalized radiotherapy planning method for glioma subtarget areas according to embodiments of the present invention involves inputting the target T1C image and the corresponding AAL image of the brain region into a segmentation model to obtain a brain region segmentation mask file; inputting the target T1C image into a mapping model to obtain a pseudo-MRS image; and inputting the target T1C image and the registered glioma annotation file into a habitat model to obtain a habitat atlas. The brain region segmentation mask file is used to identify whether the habitat subregion of the glioma is located in a functionally critical brain region. If so, the safe dose threshold for that habitat subregion is configured as a specific tolerable dose corresponding to that functionally critical brain region; otherwise, it is configured as a conventional tolerable dose. Furthermore, by acquiring the ASL corresponding to the target T1C image and based on the target T1C image, ASL, pseudo-MRS image, and safe dose threshold, the radiation dose of the glioma subregion is determined. Therefore, even with only conventional T1C images, a metabolite distribution map reflecting tumor heterogeneity can be predicted efficiently and accurately, enabling the development of a dose-sculpting prescription that integrates anatomy, function, and metabolism. It can effectively improve the stability, repeatability and efficiency of target delineation, and can achieve biological conformity of radiotherapy dose, thus providing objective and accurate quantitative reference for the formulation of radiotherapy plans.

[0093] Figure 6 This is a structural block diagram of a personalized radiotherapy planning system for subtargeted areas of glioma according to an embodiment of this application. (Reference) Figure 6 As shown, the personalized radiotherapy planning system 200 for subtargeted gliomas includes: a segmentation mask generation module 201, a pseudo-MRS image generation module 202, a habitat map generation module 203, a first dose sculpting prescription module 204, and a second dose sculpting prescription module 205.

[0094] The brain region segmentation mask generation module 201 inputs the T1C image to be tested and the corresponding AAL image of the brain region into a segmentation model trained by a convolutional neural network to obtain a brain region segmentation mask file. The pseudo-MRS image generation module 202 inputs the T1C image to be tested into a mapping model trained by a mapping deep learning algorithm to obtain a pseudo-MRS image. The habitat map generation module 203 inputs the T1C image to be tested and the registered glioma annotation file into a habitat model trained by a habitat algorithm to obtain a habitat map. The first dose sculpting prescription module 204, based on the brain region segmentation mask file, identifies whether the habitat subregion of the glioma is located in a functionally critical brain region; if so, the safe dose threshold for that habitat subregion is configured as the specific tolerable dose corresponding to that functionally critical brain region; if not, the safe dose threshold for that habitat subregion is configured as the conventional tolerable dose. The second dose carving prescription module 205 is used to acquire the ASL corresponding to the T1C image to be tested, and to determine the radiation dose of the glioma subregion based on the T1C image to be tested, the ASL, the pseudo-MRS image and the safe dose threshold.

[0095] In this embodiment of the application, the second dose carving prescription module 205 is specifically used to: register the T1C image to be tested, the ASL and the pseudo MRS image to obtain the MRS metabolite function reference value and the ASL metabolite function reference value for each habitat subregion; perform weighted fusion of the MRS metabolite function reference value and the ASL metabolite function reference value to obtain the corresponding malignancy index; and configure the radiation dose of the habitat subregion positively correlated with the malignancy index, using the safe dose threshold of the habitat subregion as the upper limit.

[0096] In this embodiment, the segmentation mask generation module 201 is further configured to: resample the T1C image registered with the AAL image to obtain a high-resolution image and a low-resolution image; normalize the high-resolution image and the low-resolution image respectively, and crop out corresponding high-resolution image blocks and low-resolution image blocks; input the high-resolution image blocks into a first convolutional neural network for training, and obtain a fine segmentation network model file of the network model through loss function iteration; input the low-resolution image blocks into a second convolutional neural network for training, and obtain a coarse segmentation network model file of the network model through loss function iteration.

[0097] In this embodiment of the application, the segmentation mask generation module 201 is specifically used to: configure the loss function as the Dice loss function.

[0098] In this embodiment of the application, the pseudo MRS image generation module 202 is further configured to: employ any one of the following deep learning algorithms: GAN network, diffusion model, CNN network, and Transformer network.

[0099] In this embodiment of the application, the pseudo MRS image generation module 202 is further configured to: input the registered T1C image and MRS image pair dataset into the GAN network, and generate the mapping model that can output the pseudo MRS image through the total loss function of joint adversarial loss, pixel-level loss and perceptual loss.

[0100] In this embodiment of the application, the habitat map generation module 203 is further configured to: configure the habitat algorithm with at least one of K-means, GMM, Otsu, SLIC, MSI matrix, and ITHscore.

[0101] In this embodiment, the habitat map generation module 203 is further configured to: input a registered T1C image and a registered glioma annotation file; divide the brain tumor region into multiple superpixel units using a superpixel segmentation algorithm and obtain the image features of the superpixel units; group the superpixel units based on the image features, determine habitat subregions, and generate subregion label masks; learn the mapping relationship from the image features to the subregion labels to obtain the habitat analysis model file of the habitat model.

[0102] In this embodiment of the application, the segmentation mask generation module 201 is further configured to: configure the convolutional neural network as any one of VB-Net network, U-Net network, a variant of U-Net network, nnUNet network, or a hybrid neural network based on CNN and Transformer architecture.

[0103] It should be noted that the explanation of the personalized radiotherapy planning method for subtargeted areas of glioma in the above embodiments also applies to the personalized radiotherapy planning system for subtargeted areas of glioma in the above embodiments, and will not be repeated here.

[0104] It is understood that the methods in this invention patent are applicable to scenarios including but not limited to MR, CT, PET (positron emission tomography), MR T1, T2, ADC (apparent diffusion coefficient), DWI (diffusion-weighted imaging), as well as enhanced images, H-MRS (proton magnetic resonance imaging), and 3D-ASL.

[0105] It should be noted that the specific values ​​mentioned above are only for illustrating the implementation of this application in detail, and should not be construed as limitations on this application.

[0106] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the invention without departing from the principles and spirit of the invention, and all such changes should fall within the protection scope of the claims of the present invention.

Claims

1. A method for personalized radiotherapy planning for subtargeted areas of glioma, characterized in that, The method includes: The brain region's T1C image and corresponding AAL image are input into a segmentation model trained by a convolutional neural network to obtain a brain region segmentation mask file. The T1C image to be tested is input into a mapping model trained by a mapping deep learning algorithm to obtain a pseudo MRS image. The T1C image to be tested and the registered glioma annotation file are input into the habitat model trained by the habitat algorithm to obtain the habitat map; Using the brain region segmentation mask file, it is identified whether the habitat subregion of the glioma is located in a functional critical brain region; if so, the safe dose threshold for the habitat subregion is configured to the specific tolerable dose corresponding to the functional critical brain region; if not, the safe dose threshold for the habitat subregion is configured to the conventional tolerable dose. Obtain the ASL corresponding to the T1C image to be tested; determine the radiation dose of the glioma subregion based on the T1C image to be tested, the ASL, the pseudo-MRS image and the safe dose threshold.

2. The method of claim 1, wherein, The step of determining the radiation dose of the habitat subregion based on the T1C image to be measured, the ASL, and the pseudo-MRS image includes: The T1C image to be tested, the ASL and the pseudo MRS image are registered to obtain the MRS metabolite function reference value and the ASL metabolite function reference value for each habitat subregion. The functional reference values ​​of the MRS metabolites and the functional reference values ​​of the ASL metabolites are weighted and fused to obtain the corresponding malignancy index; The radiation dose of the habitat sub-region is configured to be positively correlated with the malignancy index, with the safe dose threshold of the habitat sub-region as the upper limit.

3. The method of claim 1, wherein, Before the step of inputting the T1C image of the brain region to be tested and the corresponding AAL image into a segmentation model trained by a convolutional neural network to obtain a brain region segmentation mask file, the method further includes: The T1C image registered with the AAL image is resampled to obtain a high-resolution image and a low-resolution image; The high-resolution image and the low-resolution image are normalized respectively, and corresponding high-resolution image blocks and low-resolution image blocks are cropped out. The high-resolution image patch is input into a first convolutional neural network for training. After iterating through a loss function, a fine segmentation network model file of the segmentation model is obtained. The low-resolution image patch is input into a second convolutional neural network for training. After iterating through a loss function, a coarse segmentation network model file of the segmentation model is obtained. The segmentation model is composed of the cascaded fine segmentation network model file and the coarse segmentation network model file.

4. The method of claim 3, wherein, The loss function is the Dice loss function.

5. The method of claim 1, wherein, The mapping deep learning algorithm employs any one of the following: GAN network, diffusion model, CNN network, or Transformer network.

6. The method of claim 1, wherein, The mapping deep learning algorithm uses a GAN network; Before the step of inputting the T1C image to be tested into a mapping model trained by a mapping deep learning algorithm to obtain a pseudo MRS image, the method further includes: The registered T1C image and MRS image pair dataset is input into the GAN network. After iterating through the total loss function of joint adversarial loss, pixel-level loss and perceptual loss, the mapping model that can output the pseudo MRS image is generated.

7. The method of claim 1, wherein, The habitat algorithm employs at least one of K-means, GMM, Otsu, SLIC, MSI matrix, and ITHscore.

8. The method of claim 1, wherein, The habitat algorithm is a superpixel segmentation algorithm; Before the step of inputting the T1C image to be tested and the registered glioma annotation file into the habitat model trained by the habitat algorithm to obtain the habitat atlas, the method further includes: Input the registered T1C image and the glioma annotation file, use the superpixel segmentation algorithm to divide the brain tumor region into multiple superpixel units, and obtain the image features of the superpixel units; Based on the image features, superpixel units are grouped to determine habitat subregions and generate subregion label masks. By learning the mapping relationship from the image features to the sub-region labels, the habitat analysis model file of the habitat model is obtained.

9. The method according to any one of claims 1 to 8, characterized in that, The convolutional neural network is any one of VB-Net, U-Net, a variant of U-Net, nnUNet, or a hybrid neural network based on CNN and Transformer architecture.

10. A brain glioma sub-target personalized radiotherapy planning system, characterized in that, The system includes: The segmentation mask generation module is used to input the T1C image of the brain region to be tested and the corresponding AAL image into the segmentation model trained by the convolutional neural network to obtain the brain region segmentation mask file. The pseudo-MRS image generation module is used to input the T1C image to be tested into a mapping model trained by a mapping deep learning algorithm to obtain a pseudo-MRS image. The habitat map generation module is used to input the T1C image to be tested and the registered glioma annotation file into the habitat model trained by the habitat algorithm to obtain the habitat map; The first dose carving prescription module identifies whether the habitat subregion of the glioma is located in a functional critical brain region through the brain region segmentation mask file; if so, the safe dose threshold of the habitat subregion is configured to the specific tolerable dose corresponding to the functional critical brain region; if not, the safe dose threshold of the habitat subregion is configured to the conventional tolerable dose. The second dose carving prescription module is used to acquire the ASL corresponding to the T1C image to be tested; and to determine the radiation dose of the glioma subregion based on the T1C image to be tested, the ASL, the pseudo-MRS image and the safe dose threshold.

Images