Voxel-level dose distribution prediction method

Abstract

The invention proposes a voxel-level dose distribution prediction method, and the method comprises the steps: obtaining a detection image of a target object when each tracer is independently introduced under the condition that a plurality of tracers are independently introduced into the target object; performing image standardization processing on the detection image corresponding to each tracer agent to obtain a standardized image corresponding to each tracer agent; and taking the standardized images corresponding to the various tracers as input information of a deep learning model, and determining a comprehensive three-dimensional dose prediction result of the target object under the various tracers through the deep learning model. According to the technical scheme, the accuracy of dose prediction in nuclide therapy is improved, full-link automation from image acquisition to dose planning is realized, the generated dose distribution not only meets the clinical requirement of precise coverage of a target region, but also conforms to the radiotherapy principle of normal tissue protection, and reliable technical support is provided for personalized precise radiotherapy.

Description

[Technical Field]

[0001] This application relates to the field of medical imaging technology, and in particular to a method for predicting voxel-level dose distribution. [Background Technology]

[0002] In current radionuclide therapy, traditional dose prediction methods mainly rely on single tracer images for calibration, which cannot accurately resolve tumor heterogeneity and is difficult to integrate complementary biological information from multiple tracers, resulting in insufficient prediction accuracy for complex tumor phenotypes. At the same time, existing deep learning models lack feature fusion mechanisms for multiple tracer inputs, and cannot achieve adaptive weighting of different tracer images in spatial and channel dimensions. Furthermore, due to the lack of differentiated weight optimization for organs at risk and tumor regions, traditional methods cannot balance treatment effectiveness and safety in dose prediction.

[0003] Therefore, how to establish a voxel-level dose distribution prediction method that can integrate multi-tracer information, accurately analyze tumor heterogeneity, and optimize dose prediction in key regions has become an urgent technical problem to be solved. [Summary of the Invention]

[0004] This application provides a voxel-level dose distribution prediction method, which aims to solve the technical problem that the dose distribution prediction level of current radionuclide therapy is difficult to meet the actual treatment needs in related technologies.

[0005] In a first aspect, embodiments of this application provide a voxel-level dose distribution prediction method, including:

[0006] In the case where multiple tracers are introduced into a target object independently, a detection image of the target object is obtained when each of the tracers is introduced independently.

[0007] The detection images corresponding to each of the tracers are subjected to image standardization processing to obtain standardized images corresponding to each of the tracers. The standardized images corresponding to the various tracers are located in the same dose map coordinate system to reflect the distribution of the various tracers in the target object.

[0008] Using standardized images corresponding to each of the various tracers as input information for a deep learning model, the deep learning model determines the comprehensive three-dimensional dose prediction result of the target object under the various tracers. The deep learning model is used to reflect the correlation between the standardized images corresponding to each of the various tracers and the comprehensive three-dimensional dose prediction result of the target object under the various tracers.

[0009] In one embodiment of this application, optionally, the image standardization processing for the detection images corresponding to each of the tracers includes:

[0010] The detection images corresponding to each of the tracers are converted into standard uptake values;

[0011] The standard uptake values ​​corresponding to each of the various tracers are normalized.

[0012] Based on the CT-to-CT deformable registration algorithm, the normalization results corresponding to the various tracers are transformed into the same dose map coordinate system.

[0013] In one embodiment of this application, optionally, the step of transforming the normalization results corresponding to the various tracers to the same dose map coordinate system based on the CT-to-CT deformable registration algorithm includes:

[0014] Affine transformation is performed on the normalization results corresponding to each of the various tracers to obtain the distribution information of each of the various tracers in the same dose map coordinate system.

[0015] Based on the distribution information and reference image information of the various tracers in the same dose map coordinate system, the non-rigid deformation field is determined, wherein...

[0016] The non-rigid deformation field includes the voxel position adjustment displacement of the distribution information corresponding to each of the tracers relative to the reference image information, wherein the reference image information is a preset standard image distribution information, or the reference image information is one of the distribution information corresponding to each of the various tracers;

[0017] The non-rigid deformation field is applied to the distribution information corresponding to each of the various tracers to obtain standardized images corresponding to each of the various tracers.

[0018] In one embodiment of this application, optionally, the step of using standardized images corresponding to each of the multiple tracers as input information for a deep learning model, and determining the comprehensive three-dimensional dose prediction result of the target object under multiple tracers through the deep learning model, includes:

[0019] The standardized images corresponding to each of the various tracers are input into the encoders of the respective tracers, and the encoding results corresponding to the various tracers are subjected to channel attention optimization and spatial attention optimization to obtain the encoding results corresponding to the various tracers.

[0020] Based on the gating mechanism, the encoding result from the encoder is fused with the upsampled features from the decoder to obtain high-resolution features;

[0021] The high-resolution features are mapped to a comprehensive three-dimensional dose prediction result by performing layer-by-layer upsampling through decoder, skip connections to fuse multi-scale features, and convolution.

[0022] In one embodiment of this application, optionally, before acquiring the detection image of the target object when each of the tracers is independently introduced, the method further includes:

[0023] Obtain training samples;

[0024] Based on the training samples, the deep learning model is pre-trained, wherein the pre-training method specifically includes:

[0025] Acquire detection images of each sample tracer when each sample object in the training sample is independently introduced;

[0026] The detection images corresponding to each of the sample tracers are subjected to image standardization processing to obtain standardized images corresponding to each of the sample tracers;

[0027] The standardized images corresponding to each of the sample objects when multiple sample tracers are introduced are used as the input information of the initial learning model, and the initial prediction results of the sample objects under multiple sample tracers are determined by the initial learning model.

[0028] Based on the initial prediction results and the actual results corresponding to the training samples, determine whether the loss function of this iteration meets the predetermined conditions;

[0029] If the loss function of the current iteration satisfies the predetermined condition, the iteration stops; otherwise, the process returns to the step of obtaining training samples.

[0030] In one embodiment of this application, optionally, before determining whether the loss function of this iteration satisfies a predetermined condition based on the initial prediction result and the true result corresponding to the training sample, the method further includes:

[0031] Based on the CT images in the detected images and a first mask transformation model trained by nnUnet, an organ binary mask for the target object is determined, the organ binary mask including a binary mask for organs at risk; and

[0032] Based on the PET images in the detected images and the second mask conversion model trained by nnUnet, a tumor binary mask is determined;

[0033] Based on the binary mask of the endangered organ and the binary mask of the tumor, the background weight, endangered organ weight, and tumor region weight in the detected image are determined.

[0034] The loss function is determined based on the background weight, the organ at risk weight, and the tumor region weight.

[0035] In one embodiment of this application, optionally, determining the loss function based on the background weight, the organ at risk weight, and the tumor region weight includes:

[0036]

[0037] Where w(v)=α∈{a,b,c}, a, b, c represent the background weight, the organ at risk weight, and the tumor region weight, respectively, D pred (v) represents the predicted value of voxel v of the sample object in the initial prediction result of this iteration, D true (v) represents the true value of voxel v of the sample object in the true result corresponding to the training sample, and λ is the adjustment factor of the structural loss. This represents the loss of structural similarity between the lesions in the initial prediction result and the lesions in the actual result.

[0038] In a second aspect, embodiments of this application provide a computer device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being configured to perform the method described in the first aspect above.

[0039] Thirdly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions for performing the method described in the first aspect above.

[0040] The above technical solution addresses the technical problem that current radionuclide therapy dose distribution prediction levels are insufficient to meet actual treatment needs. It employs multiple targeted tracers to independently acquire multi-dimensional biological information of the target object. Spatial and intensity standardization of multi-source images is achieved through standard uptake value conversion and CT-to-CT deformable registration. Subsequently, a deep learning network with dual attention mechanisms and gated fusion capabilities is constructed to adaptively weight and hierarchically fuse multi-tracer features. Finally, a three-dimensional dose distribution prediction is obtained through optimization using a predetermined loss function. This improves the accuracy and reliability of dose prediction, makes the prediction results more consistent with the biological characteristics of tumors through multi-parameter feature fusion, optimizes the clinical workflow, and achieves end-to-end automation from image acquisition to dose planning. The generated dose distribution meets both the clinical requirements of precise target coverage and the principle of radiotherapy to protect normal tissues, providing reliable technical support for personalized precision radiotherapy. [Attached Image Description]

[0041] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 A flowchart of a voxel-level dose distribution prediction method according to an embodiment of this application is shown;

[0043] Figure 2 A flowchart of a deep learning model training method according to an embodiment of this application is shown;

[0044] Figure 3 A block diagram of a computer device according to one embodiment of this application is shown;

[0045] Figure 4 A block diagram of a computer device according to another embodiment of this application is shown.

Detailed Implementation Methods

[0046] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] Figure 1 A flowchart of a voxel-level dose distribution prediction method according to an embodiment of this application is shown.

[0048] like Figure 1 As shown, a voxel-level dose distribution prediction method according to an embodiment of this application includes:

[0049] Step 102: When multiple tracers are introduced into the target object independently, acquire the detection image of the target object when each of the tracers is introduced independently.

[0050] A single tracer is a radiolabeled molecular probe used in nuclear medicine imaging (such as PET or SPECT) to obtain functional or metabolic information in a living organism.

[0051] In the fields of nuclear medicine imaging and radionuclide therapy, commonly used tracers mainly include the following categories: radioligands targeting tumor-specific receptors, such as PSMA tracers, Ga-PSMA and Lu-PSMA used for prostate cancer diagnosis and treatment; DOTATATE and DOTATOC targeting somatostatin receptors, used for neuroendocrine tumor imaging and treatment; metabolic tracers, such as FDG reflecting glucose metabolism, widely used in the diagnosis of various tumors and inflammation detection; hypoxia imaging agents, such as FMISO, used to assess tumor hypoxia status; and bone imaging agents, such as NaF, used to detect bone metastases. In addition, there are novel tracers such as Pentixafor targeting the CXCR4 receptor for hematologic malignancy imaging and FAPI targeting fibrin for tumor microenvironment imaging. These tracers are distributed in vivo through different biological mechanisms, thus providing complementary diagnostic and therapeutic information. Of course, the various tracers described in this application may include at least two of the above-listed tracers, or any other tracers.

[0052] The target object includes, but is not limited to, any living organism such as a human or animal. The detected images, as raw data, reflect the initial distribution of the tracer in the target object, including, but not limited to, images obtained by any means such as CT (Computed Tomography), PET / CT (Positron Emission Tomography / Computed Tomography), SPECT / CT (Single-Photon Emission Computed Tomography / Computed Tomography), and PET / MR (Positron Emission Tomography / Magnetic Resonance Imaging).

[0053] By introducing multiple tracers independently into the target object and acquiring corresponding detection images, the distribution of the target object in the examination results of each tracer can be captured separately. This step provides a multimodal imaging basis for subsequent analysis, realizes independent observation of the biological characteristics of different tracers, provides a comprehensive data source for integrated dose prediction, and avoids mutual interference between tracers, thereby improving the accuracy and reliability of subsequent processing.

[0054] Step 104: Perform image standardization processing on the detection image corresponding to each of the tracers to obtain a standardized image corresponding to each of the tracers. The standardized images corresponding to the various tracers are located in the same dose map coordinate system to reflect the distribution of the various tracers in the target object.

[0055] Image standardization processing was performed on the detection images corresponding to each tracer, including but not limited to: conversion to standard uptake values, normalization, and spatial alignment based on a CT-to-CT deformable registration algorithm. The standardized images served as the result of standardization, ensuring the consistency of multimodal data in the dose map coordinate system. By eliminating the influence of individual differences and scanning conditions, images of different tracers could be compared and fused within the same spatial framework, enabling quantitative analysis and accurate registration of multi-source image data. This provided high-quality input for deep learning models, thereby improving the spatial consistency and biological relevance of dose prediction.

[0056] In one possible design, the detection image corresponding to each of the tracers is first converted into a Standardized Uptake Value (SUV), and then the Standardized Uptake Values ​​corresponding to each of the tracers are normalized.

[0057] The primary function of standardized uptake values ​​is to eliminate the influence of individual differences and scanning conditions, enabling quantitative analysis and thus more accurately assessing the metabolic activity or receptor expression levels of biological tissues. Converting the detection images corresponding to each tracer into standardized uptake values ​​eliminates data biases caused by different scanning conditions and equipment parameters. Standardized uptake values, as quantitative indicators, ensure the comparability of image data from different patients. By normalizing the standardized uptake values ​​corresponding to various tracers, individual metabolic differences and tracer injection dosage variations are further eliminated. This achieves standardized preprocessing of multi-source image data, laying a quantitative foundation for subsequent accurate registration and feature fusion, and significantly improving the consistency and comparability of multimodal data.

[0058] Based on this, further, using the CT-to-CT deformable registration algorithm, the normalization results corresponding to the various tracers are transformed into the same dose map coordinate system.

[0059] This ensures the precise alignment of different modal images in anatomical space. The CT-to-CT deformable registration algorithm, by combining rigid transformation and nonlinear deformation, can effectively compensate for differences in patient position and organ deformation during different scans, thereby establishing a unified spatial reference frame. This allows biological information from different tracers to be accurately superimposed at the voxel level, providing accurate spatial correspondence for subsequent multimodal feature fusion and dose prediction.

[0060] Specifically, an affine transformation is performed on the normalization results corresponding to each of the various tracers to obtain the distribution information of each of the various tracers in the same dose map coordinate system; based on the distribution information of each of the various tracers in the same dose map coordinate system and reference image information, a non-rigid deformation field is determined, wherein the non-rigid deformation field includes the voxel position adjustment displacement of the distribution information corresponding to each of the tracers relative to the reference image information, and the reference image information is a preset standard image distribution information, or the reference image information is one of the distribution information corresponding to each of the various tracers; the non-rigid deformation field is applied to the distribution information corresponding to each of the various tracers to obtain the standardized image corresponding to each of the various tracers.

[0061] Affine transformations were applied to the normalized results of various tracers to achieve global alignment between different images. The affine transformation, by adjusting translation, rotation, and scaling parameters, initially matched the overall anatomical structures of the different images. Next, a further determined non-rigid deformation field further optimized the alignment accuracy of local anatomical structures. This non-rigid deformation field, through voxel-level displacement adjustment, accurately compensated for organ deformation and local tissue displacement. Applying the non-rigid deformation field to the distribution information of each tracer resulted in a standardized image with voxel-level spatial alignment. In summary, a multi-level registration strategy achieved accurate spatial normalization from global to local levels, ensuring high consistency in anatomical structures across multimodal image data and providing a reliable spatial foundation for subsequent deep learning models to extract accurate cross-modal features.

[0062] Step 106: Using the standardized images corresponding to each of the various tracers as input information for the deep learning model, the deep learning model determines the comprehensive three-dimensional dose prediction result of the target object under the various tracers. The deep learning model is used to reflect the correlation between the standardized images corresponding to each of the various tracers and the comprehensive three-dimensional dose prediction result of the target object under the various tracers.

[0063] Using standardized images as input to a deep learning model, the model employs a multi-branch encoder, attention mechanism, and feature fusion module to ultimately output a comprehensive three-dimensional dose prediction result. This prediction reflects the dose distribution under the combined effects of multiple tracers. In other words, by learning the complex mapping relationship between multimodal images and dose distribution, the deep learning model achieves accurate modeling of tumor heterogeneity and key regions, automatically integrates complementary information from multiple tracers, and generates prediction results highly consistent with the actual dose, thus providing a reliable basis for dose planning in personalized radionuclide therapy.

[0064] In one possible design, the standardized images corresponding to each of the various tracers can be input into the encoders of the respective tracers, and the encoding results corresponding to the various tracers can be subjected to channel attention optimization and spatial attention optimization to obtain the encoding results corresponding to the various tracers.

[0065] The encoder, as the feature extraction module of the deep learning network, can extract specific features based on the biological characteristics of different tracers. The channel attention module automatically learns the importance weights of different feature channels to highlight key biological information, while the spatial attention module focuses on specific anatomical regions in the image. This dual attention mechanism enables the model to adaptively weight and fuse feature information from multiple tracers. Thus, the final encoded result retains the unique biological characteristics of each tracer while enhancing the feature representation of key regions, achieving feature-level deep fusion of multimodal medical images and providing a more discriminative feature representation for subsequent dose prediction.

[0066] Next, based on the gating mechanism, the encoding result from the encoder is fused with the upsampled features from the decoder to obtain high-resolution features.

[0067] The gating mechanism dynamically adjusts the contribution of different feature sources through attention weights. The upsampled features carry high-level semantic information recovered by the decoder. This fusion method preserves the detailed features extracted by the encoder and integrates the contextual information reconstructed by the decoder. The resulting high-resolution features achieve an optimal balance between spatial detail and semantic understanding. Thus, information complementarity between the encoding and decoding paths is realized. By selectively strengthening the feature combinations most relevant to dose prediction through gating, the quality of key feature representation is significantly improved.

[0068] Furthermore, the high-resolution features are mapped to a comprehensive three-dimensional dose prediction result by performing layer-by-layer upsampling through the decoder, skip connections to fuse multi-scale features, and convolution.

[0069] By performing layer-by-layer upsampling through the decoder, spatial resolution can be gradually restored. Skip connections are used to fuse multi-scale features, integrating feature information from different levels. Convolutional operations further optimize feature representation, ultimately mapping high-resolution features into a comprehensive 3D dose prediction result. The layer-by-layer upsampling process progressively refines the spatial accuracy of the prediction result, multi-scale feature fusion ensures a complete information flow from local details to global semantics, and convolutional layers achieve a non-linear mapping from features to dose values. In summary, through progressive feature reconstruction and optimization, voxel-level dose distribution predictions aligned with the input image space can be generated, meeting clinical requirements for dose prediction accuracy while maintaining computational efficiency.

[0070] The above technical solution employs multiple targeted tracers to independently acquire multi-dimensional biological information of the target object. Spatial and intensity standardization of multi-source images is achieved through standard uptake value conversion and CT-to-CT deformable registration. Subsequently, a deep learning network with dual attention mechanisms and gated fusion capabilities is constructed to adaptively weight and hierarchically fuse multi-tracer features. Finally, a three-dimensional dose distribution prediction is obtained through optimization using a predetermined loss function. This establishes a fully automated mapping relationship from multi-parameter biological images to precise dose distribution, overcoming the technical bottlenecks of traditional methods in tumor heterogeneity analysis and multi-modal information fusion. Furthermore, through tracer-specific encoders and attention mechanisms, it can fully exploit the complementary biological information of different tracers.

[0071] In summary, this technical solution significantly improves the accuracy and reliability of dose prediction. By fusing multiple parameters, the prediction results are more consistent with the biological characteristics of tumors. Furthermore, it optimizes the clinical workflow, achieving full-chain automation from image acquisition to dose planning. The generated dose distribution not only meets the clinical requirements for precise target coverage but also conforms to the radiotherapy principle of protecting normal tissues, providing reliable technical support for personalized precision radiotherapy.

[0072] Figure 2 A flowchart of a deep learning model training method according to an embodiment of this application is shown.

[0073] like Figure 2 As shown, a deep learning model training method according to an embodiment of this application includes:

[0074] Step 202: Before acquiring the detection images of the target object when each tracer is introduced independently, acquire training samples so as to pre-train the deep learning model described above based on the training samples.

[0075] By acquiring training samples in advance, a data foundation can be laid for the pre-training of deep learning models. The training samples, as the knowledge source for model learning, contain diverse clinical cases and dose distribution patterns, ensuring that the model can learn the complex mapping relationship between multi-tracer images and dose distribution from a large amount of labeled data before formal application. This is the basis for establishing a transformation channel from data to knowledge, providing reliable learning materials and optimization directions for subsequent accurate predictions.

[0076] Step 204: Obtain the detection image of each sample tracer when the sample object in the training sample is independently introduced.

[0077] The detection images of the sample objects under each sample tracer were used to construct a multimodal training dataset. The detection images, as the original input, reflect the initial distribution characteristics of different tracers in the organism. Thus, the tracer-specific biological characteristics can be captured, providing a comprehensive information source for the model to learn multi-parameter joint analysis, enabling the model to understand tumor heterogeneity and normal tissue response from different perspectives.

[0078] Step 206: Perform image standardization processing on the detection image corresponding to each of the sample tracers to obtain a standardized image corresponding to each of the sample tracers.

[0079] Standardization, by unifying data formats and spatial coordinate systems, can eliminate noise interference caused by equipment differences and individual variations. Standardized images, as normalized input training samples, significantly improve the model's generalization ability and robustness. In summary, data preprocessing can transform raw images into the effective input required by the model, creating the necessary conditions for stable and efficient model training.

[0080] Step 208: Using the standardized images of the sample object corresponding to each of the multiple sample tracers as input information for the initial learning model, the initial prediction results of the sample object under multiple sample tracers are determined through the initial learning model.

[0081] By inputting standardized multi-tracer images into the initial learning model, initial prediction results can be obtained. The initial learning model serves as the network architecture with parameters to be optimized, responsible for establishing a preliminary mapping relationship from the input images to the dose distribution. Thus, the encoding and decoding of multimodal features are realized through the forward propagation process, constructing an end-to-end dose prediction framework and providing benchmark performance for subsequent iterative optimization.

[0082] Step 210: Based on the initial prediction result and the actual result corresponding to the training sample, determine whether the loss function of this iteration meets the predetermined conditions.

[0083] By comparing the initial prediction results with the actual results, we can assess whether the loss function meets the standard. As a quantitative indicator of model performance, the loss function actually reflects the degree of deviation between the predicted dose and the actual dose. This step uses the backpropagation algorithm to identify the optimization direction of the model parameters and establishes a feedback mechanism for model self-improvement, ensuring that the prediction accuracy can be continuously improved through continuous iteration.

[0084] Previously, the loss function needed to be determined.

[0085] In each iteration, the process of determining the loss function includes: determining the organ binary mask of the target object based on the CT image in the detected image and a first mask transformation model trained by nnUnet, wherein the organ binary mask includes a binary mask of organs at risk; determining the tumor binary mask based on the PET image in the detected image and a second mask transformation model trained by nnUnet; determining the background weight, organ at risk weight, and tumor region weight in the detected image based on the binary mask of the organs at risk and the binary mask of the tumor; and determining the loss function based on the background weight, the organ at risk weight, and the tumor region weight.

[0086] In determining the loss function, the first step is to use CT images from the detected images and the first mask transformation model trained by nnUnet to generate organ binary masks. These masks can accurately identify the spatial distribution of organs at risk, such as outlining the contours of key anatomical structures like the liver and kidneys. This provides accurate anatomical references for subsequent dose optimization, ensuring that the model can identify normal tissue areas that require special protection.

[0087] Simultaneously, a tumor binary mask can be obtained by detecting PET images in the imaging and another second mask transformation model trained by nnUnet. This mask can clearly define the shape and extent of the tumor. For example, it can specifically mark the three-dimensional boundary of prostate cancer lesions. This processing enables the model to focus on the treatment target area that requires high-dose irradiation.

[0088] Next, based on the two types of mask information mentioned above, the background weight, the weight of organs at risk, and the weight of the tumor region are further determined. These weight parameters reflect the relative importance of different regions in dose optimization. For example, the weight of the tumor region can be set to be significantly higher than that of the background region. This differentiated weight allocation mechanism enables the model to optimize the dose distribution of key regions in a targeted manner during training.

[0089] Finally, the final loss function is determined based on these regional weights. This function balances the numerical accuracy and spatial rationality of dose prediction, ensuring sufficient dose to the tumor region while strictly controlling the radiation dose to organs at risk. This allows the model to selectively optimize the dose prediction accuracy of key regions during training. Thus, a loss function highly aligned with clinical needs is established, enhancing the model's ability to balance tumor control and normal tissue protection.

[0090] Specifically, the loss function is as follows:

[0091]

[0092] Where w(v)=α∈{a,b,c}, a, b, c represent the background weight, the organ at risk weight, and the tumor region weight, respectively, D pred (v) represents the predicted value of voxel v of the sample object in the initial prediction result of this iteration, D true (v) represents the true value of voxel v of the sample object in the true result corresponding to the training sample, and λ is the adjustment factor of the structural loss. This represents the loss of structural similarity between the lesions in the initial prediction result and the lesions in the actual result.

[0093] The weighted mean squared error term, by assigning differentiated weights to different anatomical regions, enables focused attention on tumor areas and organs at risk. The differentiated settings for background weight, organ at risk weight, and tumor region weight reflect varying clinical requirements for dose distribution. In one possible design, the background weight, organ at risk weight, and tumor region weight are 1.0, 1.5, and 2.0, respectively.

[0094] The structural similarity loss function is specifically optimized for the morphological features of the lesion region, ensuring that the predicted dose distribution is consistent with the actual distribution in terms of spatial structure through the SSIM index. The introduction of the adjustment factor λ balances the two optimization objectives of dose numerical accuracy and spatial structural fidelity. This composite loss function achieves synergistic optimization of dose prediction in both numerical accuracy and spatial rationality, ensuring both voxel-level dose calculation accuracy and maintaining the overall morphological characteristics of the dose distribution. The final dose prediction result meets both clinical accuracy requirements and conforms to the spatial distribution patterns of radiotherapy.

[0095] Step 212: If the loss function of this iteration satisfies the predetermined condition, stop the iteration; otherwise, return to the step of obtaining training samples.

[0096] In other words, the decision to terminate the training process is made based on the evaluation of the loss function. Predefined conditions, serving as the criteria for stopping training, balance model performance with computational cost. Generally, these predetermined conditions are a specified range for the loss function, such as being less than or equal to a specified threshold. Through this intelligent training control mechanism, underfitting or overfitting problems can be avoided, resulting in an optimized model that achieves the best balance between prediction accuracy and generalization ability, providing reliable support for clinical dosing planning.

[0097] The above technical solution constructs a complete deep learning system for voxel-level dose prediction based on multi-tracer medical images. First, a multimodal training dataset is established, employing standardization and spatial registration to ensure data quality. Then, multi-tracer features are extracted through an encoder-decoder network with a dual attention mechanism. Finally, a composite loss function guided by anatomical structure is used to optimize model parameters. This achieves high-precision end-to-end prediction of three-dimensional dose distribution from multi-source medical images, while also achieving the dual goals of dose numerical accuracy optimization and spatial structure preservation. A differentiated weight allocation mechanism automatically balances the needs of tumor control and normal tissue protection, while the introduction of structural similarity loss ensures the clinical rationality of the dose distribution. In summary, the above model training method significantly improves the accuracy and reliability of dose prediction, enhances the level of multi-tracer feature fusion, and facilitates the model to more accurately capture tumor heterogeneity. Simultaneously, this technical solution optimizes the clinical workflow, enabling fully automated processing from image to dose planning. Furthermore, the final generated dose distribution prediction meets both target coverage requirements and the principle of organ-at-risk protection, providing a scientific basis and technical support for personalized precision radiotherapy.

[0098] In another embodiment, this application provides a computer device, which may be a server, and its internal structure diagram may be as follows. Figure 3 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When the computer program is executed by the processor, it can implement the methods described in any of the above embodiments.

[0099] In one embodiment, this application also provides a computer device, which can be a client, and its internal structure diagram can be as follows: Figure 4 As shown, the computer device includes a processor, memory, network interface, display screen, and input device connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it can implement the methods described in any of the above embodiments.

[0100] Any of the computer devices described in the embodiments of this application exist in various forms, including but not limited to:

[0101] (1) Mobile communication devices: These devices are characterized by their mobile communication capabilities and primarily aim to provide voice and data communication. These terminals include smartphones, multimedia phones, feature phones, and low-end phones.

[0102] (2) Ultra-mobile personal computer devices: These devices fall under the category of personal computers, possessing computing and processing capabilities, and generally also have mobile internet access features. These terminals include PDAs, MIDs, and UMPCs, etc.

[0103] (3) Portable entertainment devices: These devices can display and play multimedia content. This category includes: audio and video players, handheld game consoles, e-books, as well as smart toys, wearable devices, and portable car navigation devices.

[0104] (4) Server: A device that provides computing services. The components of a server include a processor, hard disk, memory, system bus, etc. Servers are similar to general computer architectures, but because they need to provide highly reliable services, they have higher requirements in terms of processing power, stability, reliability, security, scalability, and manageability.

[0105] (5) Other electronic devices with data interaction functions.

[0106] Additionally, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which are used to perform the following steps:

[0107] In the case where multiple tracers are introduced into a target object independently, a detection image of the target object is obtained when each of the tracers is introduced independently.

[0108] The detection images corresponding to each of the tracers are subjected to image standardization processing to obtain standardized images corresponding to each of the tracers. The standardized images corresponding to the various tracers are located in the same dose map coordinate system to reflect the distribution of the various tracers in the target object.

[0109] Using standardized images corresponding to each of the various tracers as input information for a deep learning model, the deep learning model determines the comprehensive three-dimensional dose prediction result of the target object under the various tracers. The deep learning model is used to reflect the correlation between the standardized images corresponding to each of the various tracers and the comprehensive three-dimensional dose prediction result of the target object under the various tracers.

[0110] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0111] The technical solution of this application has been described in detail above with reference to the accompanying drawings. The technical solution of this application improves the accuracy and reliability of dose prediction, makes the prediction results more consistent with the biological characteristics of tumors through multi-parameter feature fusion, optimizes the clinical workflow, and realizes full-link automation from image acquisition to dose planning. The generated dose distribution not only meets the clinical requirements of precise target coverage, but also conforms to the radiotherapy principle of protecting normal tissues, providing reliable technical support for personalized precision radiotherapy.

[0112] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0113] The terminology used in the embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. The singular forms “a,” “the,” and “the” used in the embodiments of this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0114] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0115] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in a combination of hardware and software functional units.

[0116] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0117] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for predicting voxel-level dose distribution, characterized in that, include: In the case where multiple tracers are introduced into a target object independently, a detection image of the target object is obtained when each of the tracers is introduced independently. The detection images corresponding to each of the tracers are subjected to image standardization processing to obtain standardized images corresponding to each of the tracers. The standardized images corresponding to the various tracers are located in the same dose map coordinate system to reflect the distribution of the various tracers in the target object. Using standardized images corresponding to each of the various tracers as input information for a deep learning model, the deep learning model determines the comprehensive three-dimensional dose prediction result of the target object under the various tracers. The deep learning model is used to reflect the correlation between the standardized images corresponding to each of the various tracers and the comprehensive three-dimensional dose prediction result of the target object under the various tracers.

2. The method according to claim 1, characterized in that, The image standardization process for the detection images corresponding to each of the tracers includes: The detection images corresponding to each of the tracers are converted into standard uptake values; The standard uptake values ​​corresponding to each of the various tracers are normalized. Based on the CT-to-CT deformable registration algorithm, the normalization results corresponding to the various tracers are transformed into the same dose map coordinate system.

3. The method according to claim 2, characterized in that, The CT-to-CT deformable registration algorithm transforms the normalized processing results corresponding to various tracers to the same dose map coordinate system, including: Affine transformation is performed on the normalization results corresponding to each of the various tracers to obtain the distribution information of each of the various tracers in the same dose map coordinate system. Based on the distribution information and reference image information of the various tracers in the same dose map coordinate system, the non-rigid deformation field is determined, wherein... The non-rigid deformation field includes the voxel position adjustment displacement of the distribution information corresponding to each of the tracers relative to the reference image information, wherein the reference image information is a preset standard image distribution information, or the reference image information is one of the distribution information corresponding to each of the various tracers; The non-rigid deformation field is applied to the distribution information corresponding to each of the various tracers to obtain standardized images corresponding to each of the various tracers.

4. The method according to any one of claims 1 to 3, characterized in that, The step of using standardized images corresponding to each of the various tracers as input information for a deep learning model, and determining the comprehensive three-dimensional dose prediction result of the target object under various tracers through the deep learning model, includes: The standardized images corresponding to each of the various tracers are input into the encoders of the respective tracers, and the encoding results corresponding to the various tracers are subjected to channel attention optimization and spatial attention optimization to obtain the encoding results corresponding to the various tracers. Based on the gating mechanism, the encoding result from the encoder is fused with the upsampled features from the decoder to obtain high-resolution features; The high-resolution features are mapped to a comprehensive three-dimensional dose prediction result by performing layer-by-layer upsampling through decoder, skip connections to fuse multi-scale features, and convolution.

5. The method according to claim 4, characterized in that, Before acquiring the detection image of the target object when each of the tracers is introduced independently, the method further includes: Obtain training samples; Based on the training samples, the deep learning model is pre-trained, wherein the pre-training method specifically includes: Acquire detection images of each sample tracer when each sample object in the training sample is independently introduced; The detection images corresponding to each of the sample tracers are subjected to image standardization processing to obtain standardized images corresponding to each of the sample tracers; The standardized images corresponding to each of the sample objects when multiple sample tracers are introduced are used as the input information of the initial learning model, and the initial prediction results of the sample objects under multiple sample tracers are determined by the initial learning model. Based on the initial prediction results and the actual results corresponding to the training samples, determine whether the loss function of this iteration meets the predetermined conditions; If the loss function of the current iteration satisfies the predetermined condition, the iteration stops; otherwise, the process returns to the step of obtaining training samples.

6. The method according to claim 5, characterized in that, Before determining whether the loss function of this iteration satisfies the predetermined conditions based on the initial prediction result and the true results corresponding to the training samples, the method further includes: Based on the CT images in the detected images and a first mask transformation model trained by nnUnet, an organ binary mask for the target object is determined, the organ binary mask including a binary mask for organs at risk; and Based on the PET images in the detected images and the second mask conversion model trained by nnUnet, a tumor binary mask is determined; Based on the binary mask of the endangered organ and the binary mask of the tumor, the background weight, endangered organ weight, and tumor region weight in the detected image are determined. The loss function is determined based on the background weight, the organ at risk weight, and the tumor region weight.

7. The method according to claim 6, characterized in that, The step of determining the loss function based on the background weight, the organ at risk weight, and the tumor region weight includes: Where w(v)=α∈{a,b,c}, a, b, c represent the background weight, the organ at risk weight, and the tumor region weight, respectively, D pred (v) represents the predicted value of voxel v of the sample object in the initial prediction result of this iteration, D true (v) represents the true value of voxel v of the sample object in the true result corresponding to the training sample, and λ is the adjustment factor of the structural loss. This represents the loss of structural similarity between the lesions in the initial prediction result and the lesions in the actual result.

8. A computer device, characterized in that, include: At least one processor; And, a memory communicatively connected to the at least one processor; The memory stores instructions executable by the at least one processor, the instructions being configured to cause the processor to perform the method according to any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, The device stores computer-executable instructions configured to perform the method as described in any one of claims 1 to 7.

Images