Deep learning-based absorbed dose prediction method, apparatus, device, and medium

Abstract

The application provides a kind of based on deep learning's absorbed dose prediction method, device, equipment and medium, it is related to medical image processing technical field, comprising: obtaining the treatment of patient's pre-test multi-modal data;According to clinical examination data and the mask map of the target tumor region and normal key organ region of pre-set drug, space coding is carried out, and the biomarker feature map of the patient to be tested is constructed;According to the pre-treatment medical image and biomarker feature map, a preset first deep learning model is used for prediction, to generate the simulated treatment initial medical image of the patient to be tested;According to the simulated treatment initial medical image, a preset second deep learning model is used for prediction, to generate the post-treatment predicted dosimetry parameter map of the patient to be tested.The application can improve the accuracy of nuclide absorbed dose prediction evaluation.

Description

Technical Field

[0001] This application relates to the field of medical image processing technology, and more specifically, to a method, apparatus, device, and medium for predicting absorbed dose based on deep learning. Background Technology

[0002] Targeted radionuclide therapy (TRT), as an emerging molecular precision therapy, delivers specific radiation doses to lesions by binding radionuclides to tumor-targeting molecules, while minimizing damage to healthy tissues.

[0003] Accurate dosimetric assessment is a crucial prerequisite for achieving personalized and precise TRT treatment. Quantifying the absorbed dose received by tumors and normal organs provides a scientific basis for predicting efficacy and assessing toxicity risks. However, significant differences exist among patients in pharmacokinetics, target molecule expression, and organ function. The fixed-dose approach commonly used in clinical practice often ignores individual patient characteristics, making it difficult to achieve the optimal balance between efficacy and safety.

[0004] Currently, the goal of personalized dosimetry is to predict post-treatment absorbed dose using pre-treatment diagnostic imaging and patient clinical indicators, thereby guiding the individualized adjustment of injection activity. However, existing methods mostly use clinical indicators as single numerical inputs for organ-level dose prediction, lacking a mechanism to encode biomarkers with their corresponding organ / lesion spatial locations. This makes it difficult for the model to learn patterns, thus affecting the accuracy of predicting absorbed doses in key organs and tumor regions. Summary of the Invention

[0005] This application provides a method, apparatus, device, and medium for predicting absorbed dose based on deep learning, which can improve the accuracy of radionuclide absorbed dose prediction and assessment.

[0006] In a first aspect, embodiments of this application provide a deep learning-based method for predicting absorbed dose, the method comprising: Acquire pre-treatment multimodal data of the patient to be tested; wherein, the pre-treatment multimodal data includes: pre-treatment medical imaging and clinical examination data; Spatial coding is performed based on the clinical examination data and a pre-set mask map of the drug tumor target area and the normal key organ area to construct the biomarker feature map of the patient to be tested. Based on the pre-treatment medical images and the biomarker feature maps, a preset first deep learning model is used to predict and generate simulated early-stage medical images of the patient to be tested. Based on the simulated initial medical images, a preset second deep learning model is used to predict and generate a post-treatment predicted dosimetric parameter map for the patient to be tested. The post-treatment predicted dosimetric parameter map includes: a predicted absorbed dose map and / or a predicted time integral activity map.

[0007] Optionally, the step of spatial encoding based on the clinical examination data and a pre-set mask map of the drug-targeted tumor region and normal key organ regions to construct a biomarker feature map of the patient to be tested includes: Based on the mask maps of the drug-targeted tumor regions and the normal critical organ regions, regional examination data indicators corresponding to the drug-targeted tumor regions and the normal critical organ regions are selected from the clinical examination data. The regional inspection data indicators are mapped to the mask map of the corresponding region to generate the coding channels corresponding to each drug-targeted tumor region and the normal key organ region; Multiple encoded channels are spliced ​​together to generate the biomarker feature map.

[0008] Optionally, the method further includes: Acquire sample image pairing data, which includes: pre-treatment sample medical images of treated patients, and actual initial medical images after treatment; The first deep learning model is obtained by training the first model based on the sample image pairing data. Acquire sample treatment cycle data, which includes: actual initial medical images of the treated patients, and corresponding actual post-treatment dosimetric parameter maps; The second deep learning model is obtained by training the second model based on the sample treatment cycle data.

[0009] Optionally, the step of training the first model based on the sample image pairing data to obtain the first deep learning model includes: Based on the pre-treatment sample medical images, a preset first generator is used to generate predictive early-stage medical images of the already treated patients. Based on the predicted initial medical images of the treated patients and the corresponding actual initial medical images, a first discriminator is used to obtain the first prediction accuracy judgment result. Based on the first prediction authenticity judgment result, the first adversarial loss function value is calculated using a preset adversarial loss function; Based on the predicted initial medical images of the treated patients and the corresponding actual initial medical images of the treated patients, the first pixel-level loss function value is calculated using a preset pixel-level loss function. Based on the predicted initial medical images of the treated patients and the corresponding actual initial medical images, a first spatial structure loss function value is calculated using a preset structural similarity loss function. Based on the predicted initial medical images of the treated patients, the corresponding actual initial medical images of the treated patients, and the preset region mask of the treated patients, a preset local loss function is used to calculate the value of the first local region level loss function. Based on the first adversarial loss function value, the first pixel-level loss function value, the first spatial structure loss function value, and the first local region-level loss function value, the preset first generator is tuned to obtain the first deep learning model.

[0010] Optionally, the step of training the second model based on the sample treatment cycle data to obtain the second deep learning model includes: Based on the actual initial medical images of the treated patients, a predicted absorbed dose map of the treated patients is generated using a preset second generator. Based on the predicted absorbed dose map and the corresponding actual absorbed dose map of the treated patients, a second prediction accuracy discrimination result is obtained using a preset second discriminator; Based on the second prediction authenticity judgment result, the second adversarial loss function value is calculated using a preset adversarial loss function; Based on the predicted absorbed dose map and the corresponding actual absorbed dose map of the treated patients, the second pixel-level loss function value is calculated using a preset pixel-level loss function. Based on the predicted absorbed dose map and the corresponding actual absorbed dose map of the treated patients, the value of the second spatial structure loss function is calculated using a preset structural similarity loss function. Based on the predicted absorbed dose map of the treated patient, the corresponding true absorbed dose map, and the preset region mask map of the treated patient, a preset local weighted loss function is used to calculate the value of the second local region level loss function. Based on the first adversarial loss function value, the second pixel-level loss function value, the second spatial structure loss function value, and the second local region-level loss function value, the preset second generator is tuned to obtain the second deep learning model.

[0011] Optionally, the step of generating a post-treatment dosimetric parameter map of the patient under test by predicting using a preset second deep learning model based on the simulated initial medical images includes: Based on the simulated initial medical images and the pre-treatment medical images, the second deep learning model is used to predict and generate a post-treatment dosimetric parameter map of the patient to be tested. or, Based on the simulated initial medical images and the biomarker feature map, the second deep learning model is used to predict and generate the post-treatment dosimetric parameter map of the patient to be tested. or, Based on the simulated initial medical images, the pre-treatment medical images, and the biomarker feature maps, the second deep learning model is used to predict and generate the post-treatment dosimetric parameter map of the patient to be tested.

[0012] Optionally, the method further includes: Dosimetric analysis is performed based on the predicted absorbed dose map to obtain the dosimetric parameters of the treatment site; the dosimetric parameters include at least one of the average dose, maximum dose, and dose-volume distribution histogram. Based on the dosimetric parameters, a personalized treatment plan recommendation is generated.

[0013] Secondly, embodiments of this application also provide a deep learning-based absorbed dose prediction device, the device comprising: The acquisition module is used to acquire pre-treatment multimodal data of the patient to be tested; wherein, the pre-treatment multimodal data includes: pre-treatment medical imaging and clinical examination data; The construction module is used to perform spatial encoding based on the clinical examination data and a preset mask map of the drug-targeted tumor region and the normal key organ region to construct the biomarker feature map of the patient to be tested; The generation module is used to generate simulated initial treatment medical images of the patient under test by using a preset first deep learning model to predict based on the pre-treatment medical images and the biomarker feature maps; and to generate post-treatment predicted dosimetric parameter maps of the patient under test by using a preset second deep learning model to predict based on the simulated initial treatment medical images, wherein the post-treatment predicted dosimetric parameter maps include: predicted absorbed dose maps and / or predicted time integral activity maps.

[0014] Thirdly, embodiments of this application also provide an electronic device, including: a processor, a memory, and a bus, wherein the memory stores program instructions executable by the processor, and when the electronic device is running, the processor communicates with the memory via the bus, and the processor executes the program instructions to perform the steps of the deep learning-based absorbed dose prediction method as described in any of the first aspects.

[0015] Fourthly, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the deep learning-based absorbed dose prediction method as described in any of the first aspects.

[0016] This application provides a deep learning-based method, apparatus, device, and medium for predicting absorbed dose. First, it acquires pre-treatment multimodal data of the patient, including pre-treatment medical images and clinical examination data. Then, based on the clinical examination data and a pre-defined mask map of the drug-targeted tumor region and normal key organ regions, spatial encoding is performed to construct a biomarker feature map reflecting the patient's individual information, effectively linking clinical indicators with organ spatial location. Next, the pre-treatment medical images and the biomarker feature map are input into a first deep learning model to generate simulated early-stage medical images to capture the early distribution of the radionuclide in the body. Finally, the simulated images are input into a second deep learning model to generate a post-treatment predicted dosimetric parameter map containing a predicted absorbed dose map and / or a predicted time-integrated activity map, thus providing a scientific basis for individualized dose assessment of radionuclide therapy. By using this method, patients' clinical examination data can be effectively correlated with the spatial location of key areas, enabling deep learning models to perceive individualized biological information of different organ regions during the prediction process. This not only significantly improves the accuracy of dose prediction for key organs such as the kidneys and liver, as well as tumor regions, but more importantly, through the stepwise simulation of the distribution process of radionuclides in the body by a two-stage deep learning model, the generated predicted absorbed dose map can more realistically reflect the dynamic metabolism and cumulative effects of drugs in the patient's body. This significantly improves the accuracy and reliability of individualized dose assessment for radionuclide therapy, reduces the risk of overtreatment or undertreatment caused by fixed-dose methods, and achieves intelligent dose prediction with deep integration of clinical indicators and imaging features, thereby improving the safety and efficacy assurance level of targeted radionuclide therapy. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating a deep learning-based absorbed dose prediction method provided in this application embodiment; Figure 2 A schematic diagram of the process for generating biomarker feature maps in a deep learning-based absorbed dose prediction method provided in this application embodiment; Figure 3 A schematic diagram illustrating the model training process in a deep learning-based absorbed dose prediction method provided in this application embodiment; Figure 4A flowchart illustrating the process of obtaining the first deep learning model in a deep learning-based absorbed dose prediction method provided in this application embodiment; Figure 5 A flowchart illustrating the process of obtaining a second deep learning model in a deep learning-based absorbed dose prediction method provided in this application embodiment; Figure 6 A schematic diagram illustrating the process of generating personalized treatment plan suggestions in a deep learning-based absorbed dose prediction method provided in an embodiment of this application; Figure 7 A schematic diagram comparing a predicted absorbed dose map with an actual absorbed dose map provided in an embodiment of this application; Figure 8 A comparison chart of the mean absolute error of TIA for different models is provided for embodiments of this application; Figure 9 A schematic diagram of an absorption dosing device based on deep learning provided in an embodiment of this application; Figure 10 This is a schematic diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0020] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0021] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0022] Before providing a detailed explanation of this application, let's first introduce its application scenarios.

[0023] When developing targeted radionuclide therapy regimens, clinicians need to determine the radiopharmaceutical injection activity for each patient to ensure sufficient radiation dose to the tumor area while keeping the radiation dose to critical organs such as the kidneys, liver, and bone marrow within safe thresholds. However, significant individual differences exist among patients in terms of pharmacokinetics, target molecule expression levels, and organ function, resulting in vastly different absorbed dose distributions for the same injection activity in different patients. The currently prevalent fixed-dose approach struggles to balance efficacy and safety, while retrospective dose assessments based on multiple post-treatment scans cannot guide regimen optimization before treatment.

[0024] Based on this, this application provides a deep learning-based method, device, equipment, and medium for predicting absorbed dose. First, it acquires pre-treatment multimodal data of the patient, including pre-treatment medical images and clinical examination data. Then, it spatially encodes the clinical examination data and a pre-defined mask map of the drug-targeted tumor region and normal key organ regions to construct a biomarker feature map reflecting the patient's individual information, effectively linking clinical indicators with organ spatial location. Next, it inputs the pre-treatment medical images and the biomarker feature map into a first deep learning model to generate simulated early-stage medical images to capture the early distribution of the radionuclide in the body. Finally, it inputs the simulated images into a second deep learning model to generate a post-treatment predicted dosimetric parameter map containing a predicted absorbed dose map and / or a predicted time-integrated activity map. This provides a scientific basis for individualized dose assessment of radionuclide therapy, significantly improving the accuracy and reliability of individualized dose assessment, reducing the risk of overtreatment or undertreatment due to fixed-dose methods, and achieving intelligent dose prediction through deep integration of clinical indicators and imaging features, thereby improving the safety and efficacy assurance of targeted radionuclide therapy.

[0025] The following explanation, in conjunction with the accompanying drawings, uses several embodiments to illustrate the concepts. Figure 1 A flowchart illustrating a deep learning-based absorbed dose prediction method provided in this application embodiment is shown below. Figure 1 As shown, this deep learning-based absorbed dose prediction method includes: S101, acquire pre-treatment multimodal data of the patient to be tested.

[0026] The pre-treatment multimodal data includes pre-treatment medical imaging and clinical examination data.

[0027] First, multimodal data collected from patients before receiving radionuclide therapy are acquired for subsequent individualized dose prediction. Pre-treatment medical imaging reflects the anatomical structure and functional metabolism of the treatment site, while clinical examination data characterizes the patient's overall condition and the function of key organs.

[0028] Specifically, pre-treatment medical imaging includes positron emission tomography (PET) images and computed tomography (CT) images. PET images can reflect the distribution of tracers in the body, such as... 68 Ga or ¹ 8 F-labeled targeting molecular imaging agents; CT images provide anatomical information for image localization and attenuation correction. Optionally, PET and CT images can be fused in the acquisition device to form a PET / CT image pair, ensuring consistency in image spatial location.

[0029] Clinical examination data includes demographic information and laboratory test data. Demographic information includes, but is not limited to, age, sex, weight, and height; laboratory test data includes hematological parameters (such as platelet count and white blood cell count), biochemical parameters (such as estimated glomerular filtration rate (eGFR), alanine aminotransferase (ALT), serum creatinine, and C-reactive protein (CRP), and tumor markers (such as prostate-specific antigen (PSA)). This data can be extracted from hospital information systems or electronic medical records and is used to assess the patient's critical organ function and tumor burden.

[0030] It should be noted that this application does not limit the specific modalities of pre-treatment medical imaging or the specific types of clinical examination data, and these can be adapted according to the actual treatment scenario and treatment plan. For example, for prostate cancer patients, pre-treatment medical imaging may include... 68 Ga-PSMA PET / CT; for patients with neuroendocrine tumors, this may be an option. 68 Ga-DOTATATEPET / CT.

[0031] S102, based on clinical examination data and a pre-set mask map of the drug-targeted tumor region and normal key organ region, spatial encoding is performed to construct a biomarker feature map of the patient to be tested.

[0032] The tumor-targeting region is the area targeted by radionuclide therapy, such as the tumor area in a prostate cancer patient. Normal critical organ regions include tissue structures requiring special attention, such as critical organs that may suffer radiation damage, like the kidneys, bone marrow, and salivary glands in a prostate cancer patient. The mask image is a spatially labeled binary image of each critical region, with dimensions consistent with the pre-treatment medical image. Each voxel has a value of 0 or 1, indicating whether the voxel belongs to a specific critical region.

[0033] Specifically, after acquiring the clinical examination data of the patient to be tested, it is necessary to associate these non-image-based numerical data with the image space so that the deep learning model can perceive the functional state information corresponding to different key regions. Specifically, the clinical examination data is encoded into feature maps aligned with the image space using a preset region mask, thus constructing biomarker feature maps.

[0034] In one possible implementation, masks of the drug-targeted tumor region and normal critical organ regions are first obtained. These regions can be defined according to the treatment scenario. The normal critical organ regions include vital organs such as the kidneys, liver, salivary glands, and bone marrow, while the drug-targeted tumor region includes one or more tumor lesions. The masks corresponding to each region can be pre-obtained using image segmentation methods, such as using the pre-set open-source segmentation tool TotalSegmentator to automatically segment organ regions, and using ITK-Snap for manual tumor segmentation by clinicians.

[0035] Subsequently, based on the mask map of the drug-targeted tumor region and the normal key organ region, the clinical examination data is spatially encoded. Specifically, information related to the function or pathological state of each region in the clinical examination data is mapped to the spatial location indicated by the mask map of the corresponding key region through a preset encoding strategy, thereby generating a feature representation reflecting the biological state of each region in the image space, i.e., obtaining a biomarker feature map.

[0036] Through the aforementioned spatial encoding method, clinical examination data is converted into image features with spatial location information. The resulting biomarker feature map has the same spatial dimensions as the pre-treatment medical images, and its channel count is correlated with the total number of drug-targeted tumor regions and normal key organ regions. The biomarker feature map and the pre-treatment medical images are jointly input into the subsequent deep learning model, enabling joint modeling of imaging and non-imaging data.

[0037] S103, based on pre-treatment medical images and biomarker feature maps, uses a preset first deep learning model to predict and generate simulated early-stage medical images of the patient to be tested.

[0038] Among these, simulated early-stage medical imaging refers to images of the patient at an early time point (e.g., 2 hours after injection) following the injection of the therapeutic radionuclide, used to reflect the initial distribution of the radiopharmaceutical in the body. In one possible implementation, simulated early-stage medical imaging is¹ 77 Single-photon emission computed tomography (SPECT) images of Lu-labeled drugs.

[0039] Specifically, after acquiring pre-treatment medical images and biomarker feature maps, they are input together into a pre-defined first deep learning model. The first deep learning model is an image-to-image generation model, which can employ a conditional generative adversarial network (cGAN), a transformer-based network structure, or other deep neural network architectures suitable for 3D medical image generation tasks. In a preferred implementation, the first deep learning model uses a cGAN architecture, with its generator being a 3D UNet structure and its discriminator being a PatchGAN structure.

[0040] The input data of the first deep learning model includes multiple channels: PET image channel and CT image channel in pre-treatment medical images, as well as multiple coding channels in the biomarker feature map. If the pre-treatment medical images are PET / CT fusion images, the PET image and CT image can be input as two independent channels; if only PET images are available, only a single-channel PET image can be input, and this application does not impose any limitations on this.

[0041] The first deep learning model processes the input multi-channel data based on the mapping relationships learned during the training phase, generating simulated early-stage medical images with the same spatial dimensions as the input images. These simulated early-stage medical images are designed to be as close as possible to real early-stage medical images, reflecting the early distribution characteristics of radionuclides at the treatment site.

[0042] S104. Based on the simulated initial medical images of the patient, a pre-set second deep learning model is used to predict and generate a post-treatment dosimetric parameter map of the patient to be tested.

[0043] Among them, the post-treatment predicted dosimetric parameters include: the predicted absorbed dose map and / or the predicted time-integrated activity map.

[0044] Here, the absorbed dose map reflects the distribution of radiation energy deposited by radionuclides in tissues and organs; the time-integrated activity map reflects the cumulative activity of radionuclides in the body over time.

[0045] Specifically, after generating simulated initial medical images for treatment, these images are used as input to a second deep learning model for further prediction. The second deep learning model is also an image-to-image generation model and can employ a 3D UNet, SwinTransformer-UNet, or other deep neural network architectures suitable for 3D medical image generation tasks. In a preferred implementation, the second deep learning model uses the same cGAN architecture as the first deep learning model to maintain consistency in network design.

[0046] The second deep learning model, based on the mapping relationships learned during the training phase, processes the input simulated initial treatment medical images to generate a predicted absorbed dose map or a predicted time-integrated activity map with the same spatial dimensions as the simulated initial treatment medical images. If the generated model is a predicted time-integrated activity map, it can be further converted into a predicted absorbed dose map using voxel S-value convolution or Monte Carlo simulation methods.

[0047] Through the above two-stage prediction process, pre-treatment multimodal data is gradually converted into post-treatment predictive dosimetric parameter maps with clinical guidance value, realizing the prediction from pre-treatment information to post-treatment dose distribution, and providing a scientific basis for the subsequent individualized treatment plan.

[0048] In this embodiment, the patient's clinical examination data can be effectively correlated with the spatial location of key areas, enabling the deep learning model to perceive individualized biological information of different organ regions during the prediction process. This not only significantly improves the accuracy of dose prediction for key organs such as the kidneys and liver, as well as tumor regions, but more importantly, through the stepwise simulation of the distribution process of radionuclides in the body by the two-stage deep learning model, the generated predicted absorbed dose map can more realistically reflect the dynamic metabolism and cumulative effect of the drug in the patient's body. This significantly improves the accuracy and reliability of individualized dose assessment for radionuclide therapy, reduces the risk of overtreatment or undertreatment caused by the fixed-dose method, and realizes intelligent dose prediction with deep integration of clinical indicators and imaging features, thereby improving the safety and efficacy assurance level of targeted radionuclide therapy.

[0049] In the above Figure 1 Based on the corresponding embodiments, in order to more clearly demonstrate the process of generating biomarker feature maps, this application also provides a possible implementation of generating biomarker feature maps in a deep learning-based absorbed dose prediction method. Figure 2 This is a schematic diagram illustrating the process of generating biomarker feature maps in a deep learning-based absorbed dose prediction method provided in an embodiment of this application. For example... Figure 2 As shown, in S102 above, spatial encoding is performed based on clinical examination data and a pre-set mask map of the drug-targeted tumor region and normal key organ regions to construct a biomarker feature map of the patient to be tested, including: S210, based on the pre-set mask maps of drug-targeted tumor regions and normal critical organ regions, select the corresponding regional examination data indicators from the clinical examination data for each drug-targeted tumor region and normal critical organ region.

[0050] Specifically, based on prior clinical knowledge and pharmacological mechanisms, a correspondence is established between each region and clinical examination data. For example, for normal key organ regions, the estimated glomerular filtration rate (eGFR) and serum creatinine can be selected for the kidney region; alanine aminotransferase (ALT) can be selected for the liver region; platelet count and C-reactive protein (CRP) can be selected for the salivary gland region; white blood cell count (WBC) can be selected for the bone marrow region; and for drug-targeted tumor regions, prostate-specific antigen (PSA) (for prostate cancer) or other corresponding tumor markers can be selected.

[0051] After selecting the regional examination data indicators, these indicators need to be standardized. One possible approach is to use z-score standardization to convert clinical indicators of different dimensions into standard scores with a mean of 0 and a variance of 1, thus eliminating the impact of dimensional differences on model training. Another possible approach is to use min-max normalization to linearly map the indicator values ​​to the [0,1] interval. Standardized regional examination data indicators have a uniform numerical range, facilitating subsequent joint modeling with imaging data.

[0052] It should be noted that the above-mentioned correspondence between the targeted tumor region and the normal key organ region and the regional examination data indicators is only illustrative and can be adaptively adjusted according to the specific therapeutic drug, mechanism of action, and clinical knowledge in practical applications. For example, for radionuclide drugs that are mainly excreted by the liver and gallbladder, liver function-related indicators can be focused on; for bone metastases, bone metabolism indicators such as alkaline phosphatase can be focused on. This application does not limit the specific selection method of regional examination data indicators.

[0053] S220 maps the regional inspection data indicators to the corresponding regional mask map, generating the coding channels corresponding to the tumor regions targeted by each drug and the normal key organ regions.

[0054] Specifically, for each drug-targeted tumor region or normal critical organ region, a mask image corresponding to that region is obtained, denoted as... The preset region mask is a binary image, which is valid for any voxel location. , This indicates that the voxel belongs to the first... Each region This indicates that it does not belong to that area.

[0055] For each region Select corresponding biomarker values ​​from clinical examination data. Or construct a combined function based on multiple relevant indicators. ,in It can be a standardized value of a single marker or a combination function of multiple markers; optionally, the combination markers can be constructed using a weighted summation method, for example, the kidney channel is encoded by a combination of eGFR and creatinine, and the salivary gland channel is encoded by a combination of platelet count and C-reactive protein.

[0056] Subsequently, the standardized regional inspection data indicators or their combined function values ​​are spatially mapped to the corresponding regional mask maps to construct encoding channels. Its mathematical expression is shown in formula (1): Formula (1) in, voxels Binary mask value at the location, For the first Each region corresponds to a biomarker value or a combination function value. Through the above mapping operation, the originally single-valued region inspection data index is expanded into a feature map with spatial distribution information: at each voxel location within each region, the encoded channel value is... This reflects the overall functional state of the region; at voxel locations outside each region, the encoded channel value is 0, indicating that the region is not directly affected by the clinical indicator. This encoding method not only preserves the numerical information of the clinical indicator but also associates it with a specific anatomical spatial location, enabling subsequent deep learning models to perceive differences in the functional state of different regions.

[0057] S230 splices multiple encoding channels to generate a biomarker feature map.

[0058] Generate the encoding channels corresponding to each region Then, these single-channel feature maps need to be integrated into a multi-channel feature map, which will then be used as input for subsequent deep learning models.

[0059] Specifically, according to the preset channel order, the encoded channels of each region are stitched together along the channel dimension to generate a multi-channel three-dimensional image, which is the biomarker feature map. The spatial size of the biomarker feature map is completely consistent with the pre-treatment medical image, and the number of channels is equal to the total number of selected drug-targeted tumor regions and normal key organ regions.

[0060] Through the above splicing operation, the biomarker feature map integrates clinical information from multiple drug-targeted tumor regions and normal key organ regions into the same data tensor, with each channel corresponding to the biological state information of a specific key region.

[0061] In this embodiment, by mapping indicators related to each region in clinical examination data to corresponding spatial mask regions and stitching them together to generate a biomarker feature map, a unified spatial representation of non-imaging clinical data and medical images is achieved. This encoding method enables the deep learning model to perceive individualized biological information corresponding to different regions while maintaining spatial structural consistency. It effectively solves the problem of losing spatial correlation by treating clinical indicators as single numerical inputs in traditional methods, providing richer feature inputs for subsequent two-stage dose prediction, and significantly improving the model's prediction accuracy and interpretability for absorbed doses in key organs and tumor regions.

[0062] In the above Figure 1 Based on the corresponding embodiments, in order to more clearly demonstrate the model training process, this application also provides a possible implementation method for model training in a deep learning-based absorbed dose prediction method. Figure 3 This is a schematic diagram illustrating the model training process in a deep learning-based absorbed dose prediction method provided in an embodiment of this application. Figure 3 As shown, based on the above S101-S104, the method further includes: S310, acquire sample image pairing data.

[0063] The sample image pairing data includes: pre-treatment sample medical images of treated patients, and actual early-stage medical images after treatment.

[0064] To train the first deep learning model, a training dataset containing input images and corresponding labels needs to be constructed. Specifically, patients who have received radionuclide therapy are selected as the sample source, and medical images of these patients taken before treatment and real images taken at early time points after treatment are obtained to form a one-to-one paired data.

[0065] In one possible implementation, the pre-treatment sample medical images are pre-treatment PET / CT images, and the actual initial medical images after treatment are acquired 2 hours after treatment.¹ 77 Lu SPECT images. The sample data volume can be configured according to the actual situation. Preferably, 29 patients who received radionuclide therapy were selected, and their pre-treatment PET images and SPECT images 2 hours after treatment were obtained, for a total of 29 pairs of sample image data. These data will be used for the subsequent training of the first deep learning model, enabling the model to learn the mapping relationship from pre-treatment images to initial treatment images.

[0066] It should be noted that this application does not impose specific limitations on the sample size, imaging modality, and acquisition time point, and these can be adaptively adjusted based on actual clinical data and treatment plans.

[0067] S320: Train the first model based on the sample image pairing data to obtain the first deep learning model.

[0068] After acquiring sample image pairing data, the first deep learning model is trained using supervised learning. The training objective is to minimize the difference between the simulated initial treatment medical images predicted by the model and the actual initial treatment medical images.

[0069] Specifically, the pre-treatment sample medical images in the sample image pairing data are used as the model input, and the corresponding real early-stage medical images are used as supervision labels, so that the model can gradually learn the mapping rules from pre-treatment images to early-stage images.

[0070] After multiple rounds of iterative optimization, when the model's performance metrics on the validation set tend to stabilize, the first deep learning model that has been trained is obtained. This model is then able to generate simulated early-stage medical images that closely resemble the actual early-stage medical images of the patient based on their pre-treatment medical images and biomarker feature maps.

[0071] S330, acquire sample treatment cycle data.

[0072] The sample treatment cycle data includes: real initial medical images of treated patients, and corresponding real post-treatment dosimetric parameter graphs.

[0073] To train the second deep learning model, another training dataset needs to be constructed. Specifically, treated patients who have completed multi-timepoint imaging are selected as the sample source. Real medical images of these patients collected in the early stages of treatment, as well as real post-treatment dosimetric parameter maps calculated through subsequent multi-timepoint imaging, are obtained to form a one-to-one paired data set.

[0074] In one possible implementation, the actual initial medical images are SPECT images taken 2 hours after treatment, and the actual post-treatment dosimetric maps are time-integrated activity maps or absorbed dose maps calculated through multi-timepoint SPECT imaging. Considering that some patients may receive multiple treatment cycles, a pair of paired data can be extracted from each treatment cycle. Preferably, data from multiple treatment cycles of 38 patients are selected, resulting in 72 pairs of 2-hour SPECT images and their corresponding time-integrated activity maps or absorbed dose maps, which serve as sample treatment cycle data. This data will be used to train a subsequent second deep learning model, enabling the model to learn the mapping relationship from initial treatment images to the final absorbed dose distribution.

[0075] It should be noted that the actual post-treatment dosimetric parameters can be obtained using a voxel-level multi-exponential curve fitting method: based on data obtained at least three time points¹ 77Lu SPECT image data are used to fit the time-activity curves of each voxel to calculate the time-integrated activity map. Furthermore, the time-integrated activity map can be converted into an absorbed dose map by voxel S-value convolution or Monte Carlo simulation.

[0076] S340, the second model is trained based on the sample treatment cycle data to obtain the second deep learning model.

[0077] After acquiring sample treatment cycle data, a second deep learning model was trained using supervised learning. The training objective was to minimize the voxel-level difference between the model-predicted absorbed dose map or predicted time-integrated activity map and the actual absorbed dose map or time-integrated activity map.

[0078] Specifically, the model input is the actual medical images from the initial treatment phase of the sample treatment cycle data, and the corresponding actual post-treatment dosimetric parameter map is used as the supervision label, so that the model can gradually learn the mapping law from the initial treatment images to the final dose distribution.

[0079] After multiple rounds of iterative optimization, when the model's performance metrics on the validation set tend to stabilize, a second deep learning model that has been trained is obtained. This model can generate a predicted absorbed dose map or a predicted time integral activity map that closely approximates the actual dose distribution based on simulated early-stage medical images of the patient being tested.

[0080] In this embodiment, by constructing a two-stage independent training process, a mapping model from pre-treatment images to early-stage treatment images and a mapping model from early-stage treatment images to dose distribution are obtained respectively. This allows the model in each stage to focus on learning the mapping rules of a specific stage, effectively reducing the learning difficulty of single-stage end-to-end mapping. The two-stage model can fully learn the early distribution characteristics and cumulative metabolic rules of radionuclides in vivo, providing a reliable model foundation for subsequent individualized dose prediction and significantly improving the accuracy and stability of the prediction results.

[0081] In the above Figure 1 Based on the corresponding embodiments, in order to more clearly demonstrate the process of obtaining the first deep learning model, this application also provides a possible implementation of obtaining the first deep learning model in a deep learning-based absorbed dose prediction method. Figure 4 This is a schematic flowchart illustrating the process of obtaining the first deep learning model in a deep learning-based absorbed dose prediction method provided in an embodiment of this application. Figure 4 As shown, in step S320 above, the first deep learning model is trained based on the sample image pairing data, and includes: S410, based on pre-treatment sample medical images, uses a preset first generator to generate predictive early-stage medical images of already treated patients.

[0082] The first generator is a generative network in the first deep learning model, used to learn the mapping relationship from pre-treatment images to initial-treatment images. Optionally, the first generator adopts a 3D UNet architecture, with pre-treatment sample medical images as input and predicted initial-treatment medical images with the same spatial dimensions as the input images as output.

[0083] Specifically, the pre-treatment sample medical images from the sample image pairing data are input into a preset generator. After layer-by-layer feature extraction and upsampling reconstruction by the encoder and decoder structure, corresponding predicted early-stage medical images are generated. The predicted early-stage medical images are consistent with the real early-stage medical images in terms of spatial size and voxel spacing, and are used for subsequent loss calculation and model optimization.

[0084] S420: Based on the predicted initial medical images of treated patients and the corresponding actual initial medical images, a preset first discriminator is used to obtain the prediction authenticity judgment result.

[0085] The first discriminator is a discriminative network in the first deep learning model, used to distinguish between the generated predicted images and the actual initial treatment images. Optionally, the first discriminator adopts a PatchGAN structure, which performs authenticity judgment on local regions of the input image and outputs the probability distribution of each region belonging to the actual image.

[0086] Specifically, the predicted initial medical images and the actual initial medical images are input separately or in pairs into a preset discriminator. The discriminator extracts and classifies the input images and outputs the prediction authenticity judgment result, which reflects the degree to which the generated images are statistically close to the actual images, providing a basis for the subsequent calculation of adversarial loss.

[0087] S430, based on the prediction accuracy judgment result, the adversarial loss function is used to calculate the value of the first adversarial loss function.

[0088] Among them, the adversarial loss function is used to guide the generator to generate prediction results that can deceive the discriminator, so that the distribution of the generated image is as close as possible to the distribution of the real image.

[0089] Specifically, based on the prediction realism judgment result output by the preset first discriminator, the first adversarial loss function value is calculated using the standard adversarial loss function of generative adversarial networks. The first adversarial loss function value measures the degree to which the generated image is judged as a real image by the discriminator; the smaller the value, the closer the distribution of the generated image is to the distribution of the real image. By minimizing the adversarial loss, the preset first generator gradually learns to generate more realistic predictive medical images for the early stages of treatment.

[0090] S440: Based on the predicted initial medical images of the treated patients and the corresponding actual initial medical images, a preset pixel-level loss function is used to calculate the first pixel-level loss function value.

[0091] The preset pixel-level loss function is used to constrain the difference between the generated image and the real image at the voxel intensity level, thereby improving the quantitative consistency of the prediction results. Optionally, the preset pixel-level loss function is the L1 loss function, which calculates the mean of the absolute error between corresponding voxels of the predicted image and the real image.

[0092] Specifically, the system iterates through all voxel positions in the predicted initial medical images and the actual initial medical images of treated patients, calculates the absolute value of the difference between the predicted value and the actual value at each voxel position, sums the absolute errors of all voxels and takes the average value to obtain the first pixel-level loss function value. The smaller the first pixel-level loss function value, the closer the generated image is to the actual image at the voxel intensity level.

[0093] S450, based on the predicted initial medical images and the actual initial medical images of the treated patients, calculate the value of the first spatial structure loss function using a preset structural similarity loss function.

[0094] The preset structural similarity loss function is used to enhance the consistency between the generated image and the real image at the structural level, preserving the spatial structural features of organs and lesions. Optionally, the preset structural similarity loss function adopts the negative form of the Structural Similarity Index Measure (SSIM).

[0095] Specifically, local image patches are extracted from both predicted and actual early-stage medical images of treated patients. Brightness contrast, contrast contrast, and structural contrast are calculated between these image patches to obtain a structural similarity index. The structural similarity index is then subtracted from 1 to obtain the first structural similarity loss value. A smaller first structural similarity loss value indicates a greater spatial structural similarity between the generated and actual images. By introducing structural similarity loss, the model can better preserve the boundary clarity and texture features of key organs and tumor regions.

[0096] S460: Based on the predicted initial medical images of the treated patients, the actual initial medical images of the treated patients, and the preset region mask of the treated patients, the first local region level loss function value is calculated using a preset local loss function.

[0097] The preset local loss function is a local weighted loss based on organ and tumor masks, which is used to enhance the model's ability to perceive key anatomical structures and treatment-related areas, so that the prediction errors of tumor and key organ regions have a higher weight in the overall loss.

[0098] Specifically, a pre-defined region mask map of the treated patient is first obtained, including masks of key organs (such as kidneys, liver, salivary glands, bone marrow, etc.) and tumor lesions. The pre-defined region mask map has the same spatial dimensions as the predicted image and the real image, and each voxel location identifies its corresponding key region category.

[0099] Subsequently, voxel values ​​of the predicted and actual medical images in the early stages of treatment are extracted within the masks of each key region, and the prediction error in each region is calculated. For the tumor region, each tumor lesion mask is traversed, and the absolute value of the difference between the predicted and actual values ​​in that region is calculated and summed; for the key organ region, each organ mask is traversed, and the absolute value of the difference between the predicted and actual values ​​in that region is calculated and summed. The total error of the tumor region is added to the total error of the organ region to obtain the first local region-level loss function value, the mathematical expression of which is given in formula (2): Formula (2) in, and These represent the total number of tumors and organs, respectively. and Do not index individual tumors and organs; For the first Prediction error within a tumor region, This is the predicted value within the mask of the tumor region. For the corresponding true value; For the first Prediction error within individual organ regions, This is the predicted value within the mask of the organ region. This corresponds to the actual value.

[0100] By using the aforementioned local weighted loss, the model will pay more attention to the prediction accuracy of tumor and key organ regions during the optimization process, thereby improving the dose prediction accuracy of these clinically important regions.

[0101] S470, based on the first adversarial loss function value, the first pixel-level loss function value, the first spatial structure loss function value, and the first local region-level loss function value, the preset first generator is tuned to obtain the first deep learning model.

[0102] After calculating the values ​​of each loss function, they are weighted and combined to form the overall optimization objective of the generator. The expression for the overall loss function is formula (3): Formula (3) in, This is the value of the first adversarial loss function. This is the value of the first pixel-level loss function. The value of the first spatial structure loss function. This represents the value of the loss function at the first local region level. , , , These are the weighting coefficients for the corresponding loss terms, used to adjust the relative contributions of different losses in the model parameter optimization process.

[0103] In one possible implementation, to evaluate the impact of different loss functions on prediction accuracy, comparative experiments can be conducted on combinations of loss functions. For example, a combination of pixel-level loss and structural similarity loss can be used, while simultaneously constraining voxel intensity differences and structural similarity; alternatively, only pixel-level loss can be used as a baseline control. By comparing the model performance under different combinations, the optimal loss function configuration can be selected.

[0104] Optionally, other types of loss functions can be combined during model training to further improve prediction performance. For example, in a comparative experimental setting, a combination of pixel-level loss and mutual information loss can be used to enhance the matching ability between generated images and real images in terms of statistical correlation and multimodal information consistency; other loss functions suitable for image generation tasks, such as perceptual loss and gradient consistency loss, can also be combined. This application does not limit the specific combination of loss functions and can be adaptively adjusted according to actual application scenarios and performance requirements.

[0105] Subsequently, based on the overall loss function value, the network parameters of the preset first generator are iteratively optimized. After multiple rounds of training, when the model's performance metrics on the validation set tend to stabilize, training is stopped, resulting in the first trained deep learning model, which can generate high-precision simulated early-stage medical images based on the pre-treatment medical images and biomarker feature maps of the patients to be tested.

[0106] In this embodiment, the first generator is optimized using a multi-objective approach by combining adversarial loss, pixel-level loss, structural similarity loss, and locally weighted loss based on organ-tumor masks. This enables the first deep learning model to not only generate realistic early-stage treatment images in terms of overall distribution, but also to achieve high-precision quantitative predictions of key organs and tumor regions. Adversarial loss ensures that the statistical distribution of the generated images approximates the real distribution; pixel-level loss guarantees the quantitative accuracy of voxel intensity; structural similarity loss maintains the spatial structural integrity of organ boundaries; and locally weighted loss guides the model to focus on key clinical areas. The synergistic effect of multiple loss functions significantly improves the prediction accuracy of simulated early-stage treatment medical images.

[0107] In the above Figure 1Based on the corresponding embodiments, in order to more clearly demonstrate the process of obtaining the second deep learning model, this application also provides a possible implementation of obtaining the second deep learning model in a deep learning-based absorbed dose prediction method. Figure 5 This is a flowchart illustrating the process of obtaining the second deep learning model in a deep learning-based absorbed dose prediction method provided in an embodiment of this application. Figure 5 As shown, in S340 above, the second deep learning model is trained based on the sample treatment cycle data, and includes: S510, based on the actual initial medical images of treated patients, uses a preset second generator to generate a predicted absorbed dose map of the treated patients.

[0108] The preset second generator is the generative network in the second deep learning model, used to learn the mapping relationship from initial treatment images to absorbed dose maps. Optionally, the preset second generator may adopt the same network architecture as the preset first generator, such as a three-dimensional UNet structure, to maintain consistency in network design.

[0109] Specifically, the actual initial medical images (such as SPECT images 2 hours after treatment) from the sample treatment cycle data are input into a preset second generator. After feature extraction and reconstruction by the encoder and decoder structure, a corresponding predicted absorbed dose map is generated. The predicted absorbed dose map is consistent with the actual absorbed dose map in terms of spatial size and voxel spacing, and is used for subsequent loss calculation and model optimization. If the actual post-treatment dosimetric parameter map in the sample treatment cycle data is a time-integrated activity map, the preset second generator can also generate a predicted time-integrated activity map; this application does not limit this.

[0110] S520: Based on the predicted absorbed dose map and the corresponding actual absorbed dose map of the treated patients, a preset second discriminator is used to obtain the second prediction authenticity discrimination result.

[0111] The preset second discriminator is a discriminant network in the second deep learning model, used to distinguish between the generated predicted absorbance-dose map and the actual absorbance-dose map. Optionally, the preset second discriminator may adopt the same PatchGAN structure as the preset first discriminator.

[0112] Specifically, the predicted absorbed dose map and the actual absorbed dose map of the treated patients are input into a preset second discriminator. The discriminator performs feature extraction and classification on the input images and outputs the second prediction authenticity discrimination result, which reflects the degree of similarity between the generated image and the real image in terms of statistical distribution, and provides a basis for the subsequent calculation of adversarial loss.

[0113] S530, based on the second prediction authenticity judgment result, the second adversarial loss function value is calculated using a preset adversarial loss function.

[0114] The preset adversarial loss function is the same as the adversarial loss function used in S430, and is used to guide the preset second generator to generate prediction results that can deceive the discriminator.

[0115] Specifically, based on the second prediction realism discrimination result output by the preset second discriminator, the second adversarial loss function value is calculated using the standard adversarial loss function of generative adversarial networks. The smaller the second adversarial loss function value, the closer the distribution of the generated image is to the distribution of the real image. By minimizing the adversarial loss, the preset second generator gradually learns to generate more realistic predicted absorbance maps.

[0116] S540, based on the predicted absorbed dose map and the corresponding actual absorbed dose map of the treated patients, calculate the second pixel-level loss function value using a preset pixel-level loss function.

[0117] The preset pixel-level loss function is the same as the pixel-level loss function used in S440, and is used to constrain the difference between the generated image and the real image at the voxel intensity level.

[0118] Specifically, the system iterates through all voxel positions in the predicted and actual absorbed dose maps of the treated patients, calculates the absolute value of the difference between the predicted and actual values ​​at each voxel position, sums the absolute errors of all voxels, and takes the average value to obtain the second pixel-level loss function value. The smaller the second pixel-level loss function value, the closer the generated image is to the real image in terms of voxel intensity.

[0119] S550, based on the predicted absorbed dose map and the corresponding actual absorbed dose map of the treated patients, the second spatial structure loss function value is calculated using a preset structural similarity loss function.

[0120] The preset structural similarity loss function is the same as the structural similarity loss function used in S450, which is used to enhance the consistency between the generated image and the real image at the structural level.

[0121] Specifically, local image patches are extracted from the predicted and actual absorbed dose maps of treated patients. The brightness contrast, contrast contrast, and structural contrast between each image patch are calculated, and a structural similarity index is obtained by combining these factors. The second spatial structural loss function value is then calculated. The specific calculation method can be found in the relevant description in S450, and will not be repeated here.

[0122] S560, based on the predicted absorbed dose map of the treated patient, the corresponding true absorbed dose map, and the preset region mask map of the treated patient, the second local region level loss function value is calculated using a preset local weighted loss function.

[0123] The preset local weighted loss function is the same as the local loss function used in S460. Both are local weighted losses based on organ and tumor masks, used to enhance the model's ability to perceive key anatomical structures and treatment-related areas.

[0124] Specifically, firstly, a preset region mask map of the treated patient is obtained, including masks for each key organ and a tumor lesion mask. Then, the voxel values ​​of the predicted and actual absorbed dose maps within each key region mask are extracted, and the prediction error within each region is calculated. For the tumor region, each tumor lesion mask is traversed, and the absolute value of the difference between the predicted and actual values ​​within that region is calculated and summed. For the key organ region, each organ mask is traversed, and the absolute value of the difference between the predicted and actual values ​​within that region is calculated and summed. The total error of the tumor region is added to the total error of the organ region to obtain the second local region-level loss function value. Its mathematical expression can be found in the relevant description of formula (2), and will not be repeated here.

[0125] By using the aforementioned local weighted loss, the model will pay more attention to the prediction accuracy of tumor and key organ regions during the optimization process, thereby further improving the prediction accuracy of these clinically important regions.

[0126] S570, based on the second adversarial loss function value, the second pixel-level loss function value, the second spatial structure loss function value, and the second local region-level loss function value, the preset second generator is tuned to obtain the second deep learning model.

[0127] After calculating the values ​​of each loss function, they are weighted and combined to form the overall optimization objective of the preset second generator. The expression of the overall loss function can be found in formula (3). The losses of each item in the first stage are replaced with the corresponding losses of the second stage and the weight coefficients of the corresponding loss items. The specific calculation method can be found in the relevant description of S470, and will not be repeated here.

[0128] In one possible implementation, comparative experiments can be conducted on combinations of loss functions to evaluate the impact of different combinations on the accuracy of absorbed dose prediction. For example, a combination of pixel-level loss and structural similarity loss can be used, or only pixel-level loss can be used as the baseline control scheme. Optionally, other types of loss functions, such as mutual information loss, can be combined to further improve prediction performance; details can be found in the description of the first stage, and will not be elaborated here.

[0129] Subsequently, based on the overall loss function value, the network parameters of the preset second generator are iteratively optimized. After multiple rounds of training, when the model's performance metrics on the validation set tend to stabilize, training is stopped, resulting in a trained second deep learning model that can generate high-precision predicted absorbed dose maps or predicted time-integrated activity maps based on simulated early-stage medical images of the patient.

[0130] In this embodiment, the second generator is optimized using a multi-objective approach, combining adversarial loss, pixel-level loss, structural similarity loss, and locally weighted loss based on organ-tumor masks. This ensures that the second deep learning model maintains a high level of focus on key organs and tumor regions during the mapping process from initial treatment images to dose distribution. The synergistic effect of multiple loss functions ensures that the generated predicted absorbed dose map not only closely approximates the true dose distribution in its overall distribution but also achieves high-precision quantitative prediction in clinically critical areas, providing a reliable dosimetric basis for the subsequent development of individualized treatment plans.

[0131] In the above Figure 1 Based on the corresponding embodiments, to more clearly demonstrate the process of generating a post-treatment predicted dosimetry parameter map for the patient under test, this application also provides a possible implementation method for generating a post-treatment predicted dosimetry parameter map for the patient under test in a deep learning-based absorbed dose prediction method. Optionally, in S104 above, based on simulated initial medical images of treatment, a preset second deep learning model is used to predict and generate a post-treatment predicted dosimetry parameter map for the patient under test, including: S610 uses a second deep learning model to predict and generate a post-treatment dosimetric parameter map for the patient based on simulated medical images from the initial stage of treatment and pre-treatment medical images.

[0132] In one possible implementation, simulated early-stage medical images and pre-treatment medical images are combined as input to a second deep learning model. The simulated early-stage medical images reflect the early distribution information of radionuclides in the body, while the pre-treatment medical images provide baseline information on anatomical structure and functional metabolism before treatment.

[0133] Specifically, simulated initial treatment medical images are stitched together with pre-treatment medical images along the channel dimension to form multi-channel input data, which is then fed into a pre-trained second deep learning model. Based on the mapping relationships learned during training, the second deep learning model processes the input data to generate a predicted absorbed dose map or a predicted time-integrated activity map with the same spatial dimensions as the input images. By simultaneously referencing pre-treatment baseline information and predicted early distribution information, the model can more comprehensively understand the anatomical features and metabolic patterns of individual patients, thereby improving the accuracy of dose prediction.

[0134] S620 uses a second deep learning model to predict and generate a post-treatment dosimetric parameter map for the patient based on simulated medical images and biomarker feature maps from the initial stage of treatment.

[0135] In another possible implementation, simulated early-stage medical images combined with biomarker feature maps are used as input to a second deep learning model. The biomarker feature maps contain clinical examination data corresponding to key regions, reflecting the functional status of key organs and tumor burden.

[0136] Specifically, simulated initial treatment medical images and biomarker feature maps are stitched together along the channel dimension to form multi-channel input data, which is then fed into a pre-trained second deep learning model. Based on the mapping relationships learned during training, the second deep learning model processes the input data to generate a predicted absorbed dose map or a predicted time-integrated activity map. By introducing biomarker feature maps, the model can continuously perceive the biological state information of key regions during dose prediction, enabling the prediction results to reflect individualized organ function and tumor characteristics, further improving the dose prediction accuracy of key regions.

[0137] S630 uses a second deep learning model to predict and generate a post-treatment dosimetric parameter map for the patient based on simulated initial medical images, pre-treatment medical images, and biomarker feature maps.

[0138] In another possible implementation, simulated early-stage medical images, pre-treatment medical images, and biomarker characteristics will be used. Figure Three The original information and intermediate prediction information are combined as input to the second deep learning model to preserve the original information and intermediate prediction information to the greatest extent.

[0139] Specifically, simulated initial treatment medical images, pre-treatment medical images, and biomarker feature maps are stitched together along the channel dimension to form joint input data containing multi-source information, which is then input into a pre-trained second deep learning model. Based on the mapping relationships learned during training, the second deep learning model processes the input data to generate a predicted absorbed dose map or a predicted time-integrated activity map. By simultaneously fusing pre-treatment baseline information, predicted early distribution information, and clinical biological information from key regions, the model can comprehensively understand the factors influencing radionuclide metabolism and dose deposition from multiple dimensions, providing the richest feature support for dose prediction.

[0140] It should be noted that S610, S620, and S630 are three parallel optional implementation methods, which can be flexibly selected according to data accessibility, computing resources, and prediction accuracy requirements in practical applications. For example, when the clinical data required for constructing biomarker feature maps is incomplete, the S610 implementation method can be used; when the quality of pre-treatment medical images is poor or there are acquisition differences, the S620 implementation method can be used; when pursuing the highest prediction accuracy, the S630 implementation method can be used, simultaneously fusing the three types of input information. This application does not limit the specific input configuration of the second deep learning model.

[0141] In this embodiment, by providing prediction methods with multiple input configurations, the second deep learning model can flexibly adjust the combination of input data according to the actual application scenario. Regardless of the input configuration used, the model can generate high-quality predicted dosimetric parameter maps based on the mapping relationships learned during the training phase. The parallel existence of multiple input configurations ensures the universality and flexibility of the method, and also provides diversified technical paths for individualized dose prediction in different clinical scenarios, effectively improving the adaptability and operability of this method in actual clinical applications.

[0142] In the above Figure 1 Based on the corresponding embodiments, in order to more clearly demonstrate the process of generating personalized treatment plan recommendations, this application also provides a possible implementation of generating personalized treatment plan recommendations in a deep learning-based absorbed dose prediction method. Figure 6 This is a schematic diagram illustrating the process of generating personalized treatment plan suggestions in a deep learning-based absorbed dose prediction method provided in an embodiment of this application. Figure 6 As shown, based on the above S101-S104, the method further includes: S710, based on the predicted absorbed dose map, performs dosimetric analysis to obtain the dosimetric parameters of the treatment site.

[0143] The dosimetric parameters include at least one of the following: average dose, maximum dose, and dose-volume distribution histogram.

[0144] Specifically, firstly, based on a preset region mask map, dose distribution data corresponding to each key region (such as key organs like the kidneys, liver, salivary glands, and bone marrow, as well as tumor lesions) is extracted from the predicted absorbed dose map. Subsequently, for each key region, the corresponding dosimetric parameters are calculated.

[0145] In one possible implementation, for each critical area, its average dose—the arithmetic mean of the dose values ​​of all voxels within that area—is calculated to assess the overall radiation exposure level of that area. Simultaneously, its maximum dose—the maximum value of the dose values ​​of all voxels within that area—is calculated to identify potential dose hotspots and assess the risk of localized overdose.

[0146] In another possible implementation, a dose-volume distribution histogram is generated. Specifically, for each key region, the cumulative percentage of volume covered by different dose levels is calculated, and a curve is plotted with the dose value on the x-axis and the volume percentage on the y-axis. The dose-volume distribution histogram can visually demonstrate the uniformity of dose distribution within a region, as well as the proportion of irradiated volume corresponding to a specific dose threshold, providing important evidence for assessing organ damage risk and tumor treatment efficacy.

[0147] It should be noted that this application does not limit the specific types and calculation methods of dosimetric parameters, and appropriate selections can be made according to clinical needs and treatment guidelines. For example, parameters such as median dose, standard deviation of dose, and volume percentage reaching a specific dose level can also be calculated.

[0148] S720 generates personalized treatment plan recommendations based on dosimetric parameters.

[0149] After obtaining the dosimetric parameters of the treatment site, these parameters are compared and analyzed with preset dose constraints. Combined with clinical treatment goals and safety limits, an individualized treatment plan recommendation is generated for the patient.

[0150] Specifically, the first step is to establish dose constraints, including safe dose limits for key organs and treatment target values ​​for tumor lesions.

[0151] Subsequently, the dosimetric parameters of each key region calculated by S710 were compared and analyzed with the above constraints. If the dosimetric parameters of all key organs are within the safety limits and the dosimetric parameters of the tumor region meet the treatment target, it can be recommended to maintain the current treatment plan; if the dosimetric parameters of some key organs exceed the safety limits, it can be recommended to reduce the injection activity or adjust the treatment plan; if the dosimetric parameters of the tumor region do not meet the treatment target, it can be recommended to increase the injection activity or consider combination therapy.

[0152] In one possible implementation, a structured treatment plan recommendation report is generated based on the comparison results, including dosimetric parameter values ​​for each key region, comparison results with constraints, and specific adjustment recommendations. For example, "The average renal dose is A, exceeding the safety limit B, it is recommended to reduce the injection activity by 10%" or "The minimum tumor coverage dose is C, below the treatment target D, it is recommended to increase the injection activity or consider fractionated treatment."

[0153] In another possible implementation, machine learning or decision tree models are combined to automatically recommend the optimal injection activity adjustment plan based on dosimetric parameters, thereby achieving intelligent assisted decision-making for treatment planning.

[0154] In this embodiment, by quantitatively analyzing the predicted absorbed dose map, dosimetric parameters of key regions are extracted and compared with preset dose constraints to generate personalized treatment plan suggestions, extending from dose prediction to clinical decision support. The quantitative analysis of dosimetric parameters transforms the dose distribution map into specific clinical indicators, facilitating physicians to quickly grasp key patient information; the personalized treatment plan suggestions directly translate the prediction results into actionable treatment adjustment plans, providing physicians with a scientific basis and significantly improving the accuracy and safety of targeted radionuclide therapy.

[0155] Based on the above embodiments, in order to verify the effectiveness of the deep learning-based absorbed dose prediction method provided in this application, the following specific embodiment will be used to explain and illustrate this application.

[0156] During the model testing phase, data from an additional 15 patients who simultaneously underwent pre-treatment PET scans and multi-timepoint SPECT scans were selected as the test set. The testing procedure was as follows: First, the pre-treatment PET / CT images of the patients to be tested, along with biomarker feature maps constructed based on clinical examination data, were input into the trained first deep learning model to generate simulated initial-treatment SPECT images. Subsequently, the predicted simulated initial-treatment SPECT images were used as input to the second deep learning model, and pre-treatment CT images and / or biomarker feature maps could be selectively combined and input into the second deep learning model to generate predicted time-integrated activity maps or absorbed dose maps. Finally, the predicted absorbed dose map was compared with the actual absorbed dose map calculated based on multi-timepoint SPECT imaging, and evaluation metrics such as root mean square error (RMSE), peak signal-to-noise ratio (PSNR), and SSIM were calculated to quantify the prediction accuracy.

[0157] To verify the technical advantages of this application, a set of comparative experiments were set up. First, a one-step end-to-end control network was constructed. This network adopts an adversarial generative model with a three-dimensional UNet architecture. The input is a pre-treatment PET standardized uptake value (SUV) image, and the output is a voxel-level time-integrated activity map or an absorbed dose map, skipping the two-stage prediction process proposed in this application. Further ablation experiments were set up to remove biomarker feature maps or locally weighted loss terms to evaluate independent contributions.

[0158] In the first-stage network loss function comparison experiment, the performance of five different loss function combinations was compared under the same dataset and training parameters. Experimental results show that the combination of L1 loss and structural similarity loss (SSIMloss) performs best in all three evaluation metrics: RMSE, PSNR, and SSIM. Specific values ​​are shown in Table 1, which compares the performance of different loss functions. This result verifies the importance of jointly optimizing voxel intensity differences and spatial structure consistency for improving the accuracy of image prediction in the early stages of treatment.

[0159] Table 1: Performance Comparison of Different Loss Functions

[0160] In comparative experiments between the two-stage method and the one-step end-to-end method, the absolute percentage error of absorbed dose in each key organ and tumor region was calculated. Experimental results show that the dose prediction error of the proposed two-stage method in the kidney, liver, salivary glands, bone marrow, and tumor region is significantly smaller than that of the one-step end-to-end method. Furthermore, by introducing biomarker feature maps or local weighted losses into the two-step method, the prediction error in each key region is further reduced; when both biomarker feature maps and local weighted losses are introduced simultaneously, the model achieves the lowest prediction error in each key region, as shown in Table 2.

[0161] Table 2: Comparison of Absolute Percentage Error of Absorbed Dose for Different Organs and Tumors

[0162] In the comparison of global error metrics, the two-stage method outperformed the one-step end-to-end method in terms of RMSE, SSIM, and PSNR. Specifically, the two-stage method achieved the lowest RMSE value after introducing local weighted loss, and the highest SSIM value after simultaneously introducing biomarker feature maps and local weighted loss (see Table 3 for details). These results demonstrate that the proposed two-stage prediction process, along with the introduction of biomarker feature maps and local weighted loss, effectively improves the overall accuracy and structural fidelity of absorbed dose prediction.

[0163] Table 3: Comparison of global error indices for different methods

[0164] To further visually demonstrate the differences in the performance of different methods in absorbed dose prediction, this embodiment simultaneously visualizes and compares the time-activity integral plots generated by each method. The comparison results are as follows: Figure 7 As shown. Figure 7 This is a schematic diagram comparing a predicted absorbed dose map with an actual absorbed dose map provided in an embodiment of this application. The comparison diagram uses the pre-treatment PET input image as a reference and sequentially shows the time activity integral prediction results generated by the one-step end-to-end method, the basic two-stage two-step method, the two-step method incorporating biomarker feature maps, the two-step method incorporating local weighted loss, and the two-step method incorporating both biomarker feature maps and local weighted loss. The gold standard of time activity integral calculated based on multi-time point SPECT is used as the control benchmark, which can intuitively reflect the degree of fit between the prediction results of different methods and the actual distribution.

[0165] The visualization results clearly show that the predicted images generated by the one-step end-to-end method have high overall ambiguity, with significant deviations from the gold standard in the distribution of hotspot regions of lesions and key organs, unclear boundary contours, and a large amount of background noise, failing to accurately reflect the true spatial characteristics of nuclide distribution. In contrast, the basic two-stage, two-step method proposed in this application generates predicted images that match the gold standard in terms of hotspot distribution and spatial structure. The contours of lesion regions and nuclides are presented more clearly, and background noise is significantly reduced, intuitively demonstrating the improvement in prediction accuracy brought about by the two-stage prediction process. By introducing biomarker feature maps into the basic two-step method, the details of key organs and tumor lesion regions in the predicted images are further optimized. The hotspot intensity and spatial distribution of nuclides are closer to the gold standard, the boundary contours of key organs such as the kidneys and liver are clearer, and the continuous distribution characteristics of tumor lesions are more consistent with the actual situation. This intuitively reflects that biomarker feature maps can effectively guide the model to learn the correlation between clinical indicators and corresponding organs and lesions, thereby improving the prediction fidelity of key regions. By introducing local weighted loss into the basic two-step method, the focusing effect of the predicted image on the tumor target area and key organs is significantly improved. The localization accuracy of tumor hotspot areas and the boundary discrimination of key organs are significantly optimized, and the spatial matching with the gold standard is further improved. This intuitively demonstrates that the local weighted loss function based on key region mask can effectively enhance the model's learning ability on treatment-related core areas and reduce prediction bias in key areas. When biomarker feature maps and local weighted loss are introduced simultaneously into the two-step method, the generated predicted image achieves a high degree of visual consistency with the time-activity integral gold standard in terms of overall spatial distribution, hotspot intensity, and organ and lesion boundary contours. There is no obvious background noise, and the continuous distribution characteristics of tumor lesions and the radionuclide aggregation characteristics of key organs are accurately restored. Both the global structure and local details are highly consistent with the real distribution. The visualization results are consistent with the conclusions of the aforementioned quantitative experiments that the scheme achieved the lowest prediction error and the highest SSIM value in various key organs and tumor regions. This fully verifies that the two-stage prediction architecture, biomarker spatial encoding mechanism and local weighted loss optimization strategy proposed in this application can synergistically improve the accuracy and reliability of absorbed dose prediction, and provide accurate visualization and quantitative support for individualized dose assessment of radionuclide therapy.

[0166] To further quantify and verify the improvement effect of the proposed method on the prediction accuracy of different key organs and tumor regions, this embodiment statistically compares the mean absolute error of time integral activity of different models in five core regions: kidney, liver, parotid gland, bone marrow, and tumor. The results are as follows: Figure 8 As shown. Figure 8 A comparison chart of the mean absolute error of TIA for different models provided in this application embodiment.

[0167] likeFigure 8 As shown in the figure, the horizontal axis represents the organs at risk and tumor target area of ​​radionuclide therapy, and the vertical axis represents the mean absolute error of TIA (the lower the value, the higher the prediction accuracy). A lower value indicates a smaller deviation between the model's prediction results and the gold standard, and higher prediction accuracy. Different colored bars correspond to five control models, in order: single-stage dose prediction network, two-stage dose prediction network, biomarker-enhanced two-stage dose prediction network, two-stage dose prediction network with local weighted loss, and biomarker-enhanced two-stage dose prediction network with local weighted loss. The error bars above the bars represent the dispersion of the corresponding data, which can intuitively reflect the stability of the model's prediction results. The shorter the error bars, the better the model's generalization and the smaller the fluctuation in prediction results between individuals. It can be seen that the single-stage dose prediction network has the highest error in all evaluation regions, with the bone marrow region having an error exceeding 100% and the largest error fluctuation, verifying the defects in accuracy and stability of single-stage end-to-end prediction. The basic two-stage model proposed in this application achieves a significant reduction in error across all regions compared to the single-stage model, especially in the bone marrow region, with a simultaneous improvement in prediction stability. Based on the two-stage architecture, the prediction error of each region was further reduced by introducing biomarker feature maps and local weighted loss respectively; when the two were introduced together, the error of all regions was reduced to the lowest level in each group, while the error fluctuation was minimized, achieving the best prediction accuracy and generalization stability.

[0168] The above comparison clearly shows that the method of this application has higher prediction accuracy, smaller difference between the predicted dose distribution and the actual dose distribution, and improves the accuracy of radionuclide absorbed dose prediction.

[0169] The following describes the deep learning-based absorbed dose prediction device and electronic device provided in this application, which are used to implement the application. The specific implementation process and technical effects are described above and will not be repeated below.

[0170] Figure 9 A schematic diagram of a deep learning-based absorbed dose prediction device provided in an embodiment of this application is shown below. Figure 9 As shown, the deep learning-based absorbed dose prediction device includes: The acquisition module 1000 is used to acquire pre-treatment multimodal data of the patient to be tested. This pre-treatment multimodal data includes pre-treatment medical imaging and clinical examination data.

[0171] Module 2000 is used to construct a biomarker feature map of the patient under test by spatial encoding based on clinical examination data and a preset mask map of the drug tumor target area and normal key organ area.

[0172] The generation module 3000 is used to generate simulated initial treatment medical images of the patient based on pre-treatment medical images and biomarker feature maps using a preset first deep learning model; and to generate post-treatment dosimetric parameter maps of the patient based on simulated initial treatment medical images using a preset second deep learning model. The post-treatment dosimetric parameter maps include: a predicted absorbed dose map and / or a predicted time-integrated activity map.

[0173] Optionally, module 2000 is specifically used to select regional examination data indicators corresponding to the tumor regions targeted by each drug and the normal key organ regions from clinical examination data based on the mask maps of the tumor regions targeted by each drug and the normal key organ regions; map the regional examination data indicators to the mask maps of the corresponding regions to generate the coding channels corresponding to the tumor regions targeted by each drug and the normal key organ regions; and splice multiple coding channels to generate a biomarker feature map.

[0174] Optionally, the acquisition module 1000 is also used to acquire sample image pairing data.

[0175] Optionally, the deep learning-based absorbed dose prediction device further includes a training module 4000 for training a first model based on sample image pairing data to obtain a first deep learning model.

[0176] Optionally, the acquisition module 1000 is also used to acquire sample treatment cycle data. The sample treatment cycle data includes: actual initial medical images of treated patients, and corresponding actual post-treatment dosimetric parameter maps.

[0177] Optionally, the training module 4000 is also used to train a second model based on sample treatment cycle data to obtain a second deep learning model.

[0178] Optionally, the training module 4000 is specifically used to generate, based on pre-treatment sample medical images, a preset first generator to generate predicted initial treatment medical images of treated patients; to obtain a prediction authenticity judgment result using a preset first discriminator based on the predicted initial treatment medical images of treated patients and the corresponding real initial treatment medical images; to calculate a first adversarial loss function value using an adversarial loss function based on the prediction authenticity judgment result; to calculate a first pixel-level loss function value using a preset pixel-level loss function based on the predicted initial treatment medical images of treated patients and the corresponding real initial treatment medical images; to calculate a first spatial structure loss function value using a preset structural similarity loss function based on the predicted initial treatment medical images of treated patients and the real initial treatment medical images; to calculate a first local region-level loss function value using a preset local loss function based on the predicted initial treatment medical images of treated patients, the real initial treatment medical images, and a preset region mask of treated patients; and to tune the preset first generator based on the first adversarial loss function value, the first pixel-level loss function value, the first spatial structure loss function value, and the first local region-level loss function value to obtain a first deep learning model.

[0179] Optionally, the training module 4000 is specifically used to generate a predicted absorbed dose map of the treated patient based on the actual initial medical images of the treated patient using a preset second generator; to obtain a second prediction authenticity discrimination result using a preset second discriminator based on the predicted absorbed dose map and the corresponding actual absorbed dose map of the treated patient; to calculate a second adversarial loss function value using a preset adversarial loss function based on the second prediction authenticity discrimination result; to calculate a second pixel-level loss function value using a preset pixel-level loss function based on the predicted absorbed dose map and the corresponding actual absorbed dose map of the treated patient; to calculate a second spatial structure loss function value using a preset structural similarity loss function based on the predicted absorbed dose map and the corresponding actual absorbed dose map of the treated patient; to calculate a second local region-level loss function value using a preset local weighted loss function based on the predicted absorbed dose map, the corresponding actual absorbed dose map, and the preset region mask map of the treated patient; and to tune the preset second generator based on the second adversarial loss function value, the second pixel-level loss function value, the second spatial structure loss function value, and the second local region-level loss function value to obtain a second deep learning model.

[0180] Optionally, the generation module 3000 is specifically used to generate a post-treatment predicted dosimetric parameter map of the patient under test by using a second deep learning model to predict based on simulated initial treatment medical images and pre-treatment medical images; or, by using simulated initial treatment medical images and biomarker feature maps to predict based on a second deep learning model to generate a post-treatment predicted dosimetric parameter map of the patient under test; or, by using simulated initial treatment medical images, pre-treatment medical images, and biomarker feature maps to predict based on a second deep learning model to generate a post-treatment predicted dosimetric parameter map of the patient under test.

[0181] Optionally, the deep learning-based absorbed dose prediction device further includes: an analysis module 5000, used to perform dosimetric analysis based on the predicted absorbed dose map to obtain dosimetric parameters of the treatment site; wherein the dosimetric parameters include at least one of the average dose, maximum dose, and dose-volume distribution histogram.

[0182] Optionally, the generation module 3000 is also used to generate individualized treatment plan recommendations based on dosimetric parameters.

[0183] These modules can be one or more integrated circuits configured to implement the above methods, such as one or more Application Specific Integrated Circuits (ASICs), one or more digital signal processors (DSPs), or one or more Field Programmable Gate Arrays (FPGAs). Alternatively, when a module is implemented using processing element scheduler code, the processing element can be a general-purpose processor, such as a Central Processing Unit (CPU) or other processor capable of calling program code. Furthermore, these modules can be integrated together as a system-on-a-chip (SOC).

[0184] Figure 10 This is a schematic diagram of an electronic device provided in an embodiment of this application. The device may be a computing device or a server with computing processing capabilities.

[0185] The electronic device 10 includes a processor 11, a storage medium 12, and a bus 13. The storage medium 12 stores program instructions executable by the processor 11. When the electronic device 10 is executed, the processor 11 communicates with the storage medium 12 via the bus 13, and the processor 11 executes the program instructions to perform the above-described method embodiment. The specific implementation and technical effects are similar and will not be described in detail here.

[0186] Optionally, this application also provides a program product, such as a computer-readable storage medium, including a program that, when executed by a processor, performs the above-described method embodiments.

[0187] In the several embodiments provided in this application, it should be understood that the disclosed apparatus 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.

[0188] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0189] 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.

[0190] The integrated units implemented as software functional units described above can be stored in a computer-readable storage medium. These software functional units, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0191] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A deep learning-based method for predicting absorbed dose, characterized in that, The method includes: Acquire pre-treatment multimodal data of the patient to be tested; wherein, the pre-treatment multimodal data includes: pre-treatment medical imaging and clinical examination data; Spatial coding is performed based on the clinical examination data and a pre-set mask map of the drug-targeted tumor area and the normal key organ area to construct the biomarker feature map of the patient to be tested; Based on the pre-treatment medical images and the biomarker feature maps, a preset first deep learning model is used to predict and generate simulated early-stage medical images of the patient to be tested. Based on the simulated initial medical images, a preset second deep learning model is used to predict and generate a post-treatment predicted dosimetric parameter map for the patient to be tested. The post-treatment predicted dosimetric parameter map includes: a predicted absorbed dose map and / or a predicted time integral activity map.

2. The method according to claim 1, characterized in that, The process involves spatial encoding based on the clinical examination data and a pre-defined mask map of the drug-targeted tumor region and normal key organ regions to construct a biomarker feature map of the patient under test, including: Based on the mask maps of the drug-targeted tumor regions and the normal critical organ regions, regional examination data indicators corresponding to the drug-targeted tumor regions and the normal critical organ regions are selected from the clinical examination data. The regional inspection data indicators are mapped to the mask map of the corresponding region to generate the coding channels corresponding to each drug-targeted tumor region and the normal key organ region; Multiple encoded channels are spliced ​​together to generate the biomarker feature map.

3. The method according to claim 1, characterized in that, The method further includes: Acquire sample image pairing data, which includes: pre-treatment sample medical images of treated patients, and actual initial medical images after treatment; The first deep learning model is obtained by training the first model based on the sample image pairing data. Acquire sample treatment cycle data, which includes: actual initial medical images of the treated patients, and corresponding actual post-treatment dosimetric parameter maps; The second deep learning model is obtained by training the second model based on the sample treatment cycle data.

4. The method according to claim 3, characterized in that, The step of training a first model based on the sample image pairing data to obtain the first deep learning model includes: Based on the pre-treatment sample medical images, a preset first generator is used to generate predictive early-stage medical images of the already treated patients. Based on the predicted initial medical images of the treated patients and the corresponding actual initial medical images, a first discriminator is used to obtain the first prediction accuracy judgment result. Based on the first prediction authenticity judgment result, the first adversarial loss function value is calculated using a preset adversarial loss function; Based on the predicted initial medical images of the treated patients and the corresponding actual initial medical images of the treated patients, the first pixel-level loss function value is calculated using a preset pixel-level loss function. Based on the predicted initial medical images of the treated patients and the corresponding actual initial medical images, a first spatial structure loss function value is calculated using a preset structural similarity loss function. Based on the predicted initial medical images of the treated patients, the corresponding actual initial medical images of the treated patients, and the preset region mask of the treated patients, a preset local loss function is used to calculate the value of the first local region level loss function. Based on the first adversarial loss function value, the first pixel-level loss function value, the first spatial structure loss function value, and the first local region-level loss function value, the preset first generator is tuned to obtain the first deep learning model.

5. The method according to claim 3, characterized in that, The step of training the second model based on the sample treatment cycle data to obtain the second deep learning model includes: Based on the actual initial medical images of the treated patients, a predicted absorbed dose map of the treated patients is generated using a preset second generator. Based on the predicted absorbed dose map and the corresponding actual absorbed dose map of the treated patients, a second prediction accuracy discrimination result is obtained using a preset second discriminator; Based on the second prediction authenticity judgment result, the second adversarial loss function value is calculated using a preset adversarial loss function; Based on the predicted absorbed dose map and the corresponding actual absorbed dose map of the treated patients, the second pixel-level loss function value is calculated using a preset pixel-level loss function. Based on the predicted absorbed dose map and the corresponding actual absorbed dose map of the treated patients, the value of the second spatial structure loss function is calculated using a preset structural similarity loss function. Based on the predicted absorbed dose map of the treated patient, the corresponding true absorbed dose map, and the preset region mask map of the treated patient, a preset local weighted loss function is used to calculate the value of the second local region level loss function. Based on the first adversarial loss function value, the second pixel-level loss function value, the second spatial structure loss function value, and the second local region-level loss function value, the preset second generator is tuned to obtain the second deep learning model.

6. The method according to claim 1, characterized in that, The step of generating a post-treatment predicted dosimetric parameter map for the patient under test by using a preset second deep learning model to predict the parameters based on the simulated initial medical images includes: Based on the simulated initial medical images and the pre-treatment medical images, the second deep learning model is used to predict and generate a post-treatment predicted dosimetric parameter map for the patient to be tested. or, Based on the simulated initial medical images and the biomarker feature map, the second deep learning model is used to predict and generate the post-treatment predicted dosimetric parameter map of the patient to be tested. or, Based on the simulated initial medical images, the pre-treatment medical images, and the biomarker feature maps, the second deep learning model is used to predict and generate a post-treatment predicted dosimetric parameter map for the patient to be tested.

7. The method according to claim 1, characterized in that, The method further includes: Dosimetric analysis is performed based on the predicted absorbed dose map to obtain the dosimetric parameters of the treatment site; the dosimetric parameters include at least one of the average dose, maximum dose, and dose-volume distribution histogram. Based on the dosimetric parameters, a personalized treatment plan recommendation is generated.

8. A deep learning-based therapeutic absorption dose prediction device, characterized in that, The device includes: The acquisition module is used to acquire pre-treatment multimodal data of the patient to be tested; wherein, the pre-treatment multimodal data includes: pre-treatment medical imaging and clinical examination data; The construction module is used to perform spatial encoding based on the clinical examination data and a preset mask map of the drug-targeted tumor region and the normal key organ region to construct the biomarker feature map of the patient to be tested; The generation module is used to generate simulated initial treatment medical images of the patient under test by using a preset first deep learning model to predict based on the pre-treatment medical images and the biomarker feature maps; and to generate post-treatment predicted dosimetric parameter maps of the patient under test by using a preset second deep learning model to predict based on the simulated initial treatment medical images, wherein the post-treatment predicted dosimetric parameter maps include: predicted absorbed dose maps and / or predicted time integral activity maps.

9. An electronic device, characterized in that, include: The device includes a processor, a memory, and a bus. The memory stores program instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus, and the processor executes the program instructions to perform the steps of the deep learning-based absorbed dose prediction method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, performs the deep learning-based absorbed dose prediction method as described in any one of claims 1-7.

Images