Method and system for generating combined functional regions of lung ventilation and perfusion based on planned four-dimensional ct

Abstract

The application provides a lung ventilation and perfusion combined function region generation method and system based on planned four-dimensional CT, and belongs to the technical field of medical image processing.The lung ventilation function image and the lung perfusion function image are simultaneously generated through a deep learning method, so that the dependence of a traditional functional lung imaging method on additional inspection equipment and inspection procedures is reduced; the lung ventilation and lung perfusion related features are simultaneously learned through a joint modeling mode, so that the dynamic information contained in the four-dimensional CT can be more fully utilized; the information sharing between the lung ventilation task and the lung perfusion task is realized, the repeated feature extraction process is reduced, the model parameter size and the calculation complexity are reduced while the generation capabilities of the two types of functional images are maintained; the lung ventilation function image and the lung perfusion function image obtained through fusion prediction are determined to have a spatial overlap region, and a combined high function region capable of simultaneously reflecting the lung ventilation and lung perfusion states is obtained.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to a method and system for generating joint functional regions of lung ventilation and perfusion based on planned 4D CT. More particularly, it relates to using a deep learning model to jointly model lung structure and motion information under different respiratory states in planned 4D CT images acquired during radiotherapy, simultaneously generating spatially corresponding lung ventilation and perfusion function images, and further determining joint functional regions reflecting the common high-function areas of lung ventilation and perfusion. Background Technology

[0002] The lungs are vital gas exchange organs in the human body, and their functional state is related not only to the anatomical structure of lung tissue but also closely to local ventilation capacity and blood perfusion. In patients with thoracic tumors, normal lung tissue may be exposed to varying degrees of radiation during radiotherapy, and lung function exhibits significant spatial heterogeneity. Therefore, protecting vital normal lung tissue while ensuring the effectiveness of tumor treatment is a crucial research direction in precision radiotherapy.

[0003] Currently, clinical radiotherapy planning primarily relies on simulated localization computed tomography (CT) images to delineate target areas and organs at risk, and then evaluates the treatment plan using dose-volume parameters. However, conventional CT mainly reflects tissue density and anatomical structure information, and cannot directly provide functional information such as local lung ventilation and blood perfusion. Therefore, relying solely on anatomical CT to guide radiotherapy planning may not adequately account for the functional differences between different lung tissue regions.

[0004] To obtain spatial distribution information of lung function, existing technologies typically employ functional lung imaging methods, such as single-photon emission computed tomography (SPECT) ventilation / perfusion imaging and positron emission tomography (PET). These methods can provide spatial distribution information of lung ventilation or perfusion function and are applied to lung function evaluation and functional protection radiotherapy research.

[0005] However, functional imaging methods in nuclear medicine typically require additional examination procedures, tracer injections, and specialized equipment, increasing the burden on patients and the complexity of clinical implementation. Furthermore, since functional images and radiotherapy planning CT scans are usually acquired from different devices or at different times, spatial registration is required between them. This registration can be affected by factors such as changes in patient respiratory status, body position, and image resolution, thus limiting their application in routine radiotherapy procedures.

[0006] In recent years, the development of four-dimensional computed tomography (4D-CT) technology has provided new possibilities for evaluating lung function based on radiotherapy planning image data. Compared with traditional three-dimensional CT, 4D-CT can acquire three-dimensional CT images at multiple different time phases within the respiratory cycle. Therefore, it not only contains information on the anatomical structure of the lungs, but also dynamic information such as lung tissue movement, volume changes, and density changes during respiration.

[0007] 4D-CT-based pulmonary function imaging methods have been used to evaluate pulmonary ventilation and perfusion function. For example, some studies utilize deformation information, voxel density changes, or respiratory motion-related parameters between inspiratory and expiratory CT scans to establish methods for evaluating pulmonary ventilation function; others utilize dynamic changes in 4D-CT to estimate pulmonary perfusion-related functional parameters. These studies demonstrate the feasibility of obtaining pulmonary function-related information using 4D-CT data from radiotherapy planning, providing a new technical approach for pulmonary function imaging without the need for additional functional examinations.

[0008] With the development of artificial intelligence and deep learning technologies, methods for generating functional images based on medical imaging data have gradually attracted attention. Deep learning models can learn the complex nonlinear relationship between anatomical and functional images, enabling direct prediction of functional images from CT images. Existing studies have used 3D or 4D CT images to predict lung ventilation or perfusion function images through convolutional neural networks, encoder-decoder networks, and other methods, demonstrating the feasibility of lung function image reconstruction based on CT images.

[0009] However, existing deep learning-based methods for generating lung function images still have the following shortcomings:

[0010] First, most existing studies focus on predicting a single lung function parameter, that is, establishing separate prediction models for lung ventilation function or lung perfusion function. This makes it difficult to obtain the spatial distribution of lung ventilation and lung perfusion function simultaneously, and thus cannot fully reflect the overall functional state of lung tissue.

[0011] Secondly, although 4D-CT contains information on lung structural changes under multiple respiratory states and has been used to predict lung ventilation or lung perfusion function separately, existing deep learning methods that simultaneously generate images of lung ventilation and lung perfusion function are mainly based on single-time-point 3D CT input and have not fully utilized the complementary information contained in different respiratory states in 4D-CT of radiotherapy planning.

[0012] Furthermore, lung ventilation and lung perfusion have different physiological mechanisms and spatial distribution characteristics. Establishing separate prediction models requires repeated feature extraction and model training, increasing the scale of model parameters and computational complexity. Meanwhile, existing methods typically focus on generating individual functional images, and there is still a lack of effective technical solutions for further fusing lung ventilation and lung perfusion information to identify joint functional regions with both high ventilation and high perfusion capabilities. Summary of the Invention

[0013] The purpose of this invention is to provide a method and system for generating combined ventilation-perfusion functional regions of the lungs based on 4D CT for radiotherapy planning. This method can fully utilize the multi-respiratory state information from 4D CT for radiotherapy planning, and simultaneously generate spatially corresponding lung ventilation and lung perfusion functional images within a unified framework. Furthermore, it can determine the combined high-function regions of the lungs to meet the application requirements of functional lung evaluation and functional protection radiotherapy, thereby solving at least one of the technical problems existing in the background art.

[0014] To achieve the above objectives, the present invention adopts the following technical solution:

[0015] In a first aspect, the present invention provides a method for generating combined functional regions of lung ventilation and perfusion based on planned four-dimensional CT, comprising:

[0016] Acquire planned four-dimensional computed tomography image data acquired by the acquisition device;

[0017] The average computed tomography (CT) image, the inspiratory CT image, and the expiratory CT image are obtained from the planned four-dimensional computed tomography image data. Spatial correspondence processing is performed on the average CT image, the inspiratory CT image, and the expiratory CT image to obtain joint data containing information on different respiratory states.

[0018] The combined data is processed using a pre-trained lung ventilation-perfusion combined functional image generation model to generate three-dimensional lung ventilation function images and three-dimensional lung perfusion function images. The lung ventilation-perfusion combined functional image generation model is trained from a training set, which includes multiple sets of planned four-dimensional computed tomography (CT) image data, as well as reference lung ventilation function images and reference lung perfusion function images corresponding to each set of planned four-dimensional CT image data.

[0019] The high-function areas of lung ventilation are determined based on the three-dimensional lung ventilation function images, and the high-function areas of lung perfusion are determined based on the three-dimensional lung perfusion function images.

[0020] Based on the spatial overlap between the high-functioning pulmonary ventilation region and the high-functioning pulmonary perfusion region in the same spatial coordinate system, the spatial intersection of the two is determined as the combined high-functioning pulmonary ventilation-perfusion region.

[0021] As a further limitation of the first aspect of the present invention, the lung ventilation-perfusion combined functional image generation model includes a shared feature extraction unit, a lung ventilation function reconstruction unit, and a lung perfusion function reconstruction unit; the shared feature extraction unit is used to extract multi-scale spatial features from the combined data that simultaneously contain anatomical structural information and dynamic changes in respiratory motion; the lung ventilation function reconstruction unit is used to generate lung ventilation function images based on the multi-scale spatial features; and the lung perfusion function reconstruction unit is used to generate lung perfusion function images based on the multi-scale spatial features.

[0022] As a further limitation of the first aspect of the present invention, the shared feature extraction unit includes an encoder network and a bottleneck network; a skip connection is set between the encoder network and the decoder network at the corresponding scale to transfer the spatial detail features extracted in the encoding stage to the decoding stage; the bottleneck network is used to further integrate the high-level information extracted by the encoder network at the lowest spatial resolution and the highest feature channel dimension.

[0023] As a further limitation of the first aspect of the invention, the encoder network includes multiple encoding stages, and the lung ventilation prediction task and the lung perfusion prediction task jointly use all of the multiple encoding stages and the bottleneck network.

[0024] As a further limitation of the first aspect of the present invention, both the lung ventilation function reconstruction unit and the lung perfusion function reconstruction unit include task-specific decoder networks; the task specificity means that the lung ventilation function reconstruction unit and the lung perfusion function reconstruction unit have the same network topology, but they have independent sets of trainable parameters, and the parameters are updated according to the lung ventilation supervision target and the lung perfusion supervision target respectively during the training process; wherein, the parameters of the encoder network and the bottleneck network are shared by the lung ventilation function reconstruction unit and the lung perfusion function reconstruction unit, while the task-specific decoder networks are optimized independently.

[0025] As a further limitation of the first aspect of the present invention, in the four consecutive decoding stages, for the first... The input features in the decoding stage are first upsampled by 2×2×2 three-dimensional transposed convolution; then, the upsampled features are concatenated with the skip connection features generated by the encoder at the same spatial scale in the channel dimension; then, the residual features of the corresponding decoding stage are transformed; the last decoding stage outputs 8-channel high-resolution features, which are mapped to an output channel by 1×1×1 three-dimensional convolution.

[0026] As a further limitation of the first aspect of the present invention, during the training process of the lung ventilation-perfusion combined functional imaging generation model, a joint loss function is constructed based on the differences between the predicted lung ventilation function image and the reference lung ventilation function image, and the differences between the predicted lung perfusion function image and the reference lung perfusion function image, and the parameters of the lung ventilation-perfusion combined functional imaging generation model are optimized using the joint loss function.

[0027] Secondly, the present invention provides a system for generating combined functional regions of lung ventilation and perfusion based on planned four-dimensional CT, comprising:

[0028] The acquisition module is used to acquire the planned four-dimensional computed tomography image data collected by the acquisition device;

[0029] The processing module is used to obtain the average computed tomography image, the inspiratory computed tomography image, and the expiratory computed tomography image from the planned four-dimensional computed tomography image data, and to perform spatial correspondence processing on the average computed tomography image, the inspiratory computed tomography image, and the expiratory computed tomography image to obtain joint data containing information on different respiratory states.

[0030] The prediction generation module is used to process the combined data using a pre-trained lung ventilation-perfusion combined functional image generation model to generate three-dimensional lung ventilation function images and three-dimensional lung perfusion function images; wherein, the lung ventilation-perfusion combined functional image generation model is trained by a training set, which includes multiple sets of planned four-dimensional computed tomography (CT) image data, and reference lung ventilation function images and reference lung perfusion function images corresponding to each set of planned four-dimensional CT image data space;

[0031] The determination module is used to determine the high-function region of lung ventilation based on the three-dimensional lung ventilation function image, and to determine the high-function region of lung perfusion based on the three-dimensional lung perfusion function image; based on the spatial overlap relationship between the high-function region of lung ventilation and the high-function region of lung perfusion in the same spatial coordinate system, the spatial intersection of the two is determined as the combined high-function region of lung ventilation and perfusion.

[0032] Thirdly, the present invention provides a non-transitory computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the lung ventilation-perfusion combined functional region generation method based on planned four-dimensional CT as described in the first aspect.

[0033] Fourthly, the present invention provides a computer device including a memory and a processor, wherein the processor and the memory communicate with each other, the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the lung ventilation-perfusion combined functional region generation method based on planned four-dimensional CT as described in the first aspect.

[0034] Fifthly, the present invention provides an electronic device, comprising: a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute instructions for implementing the lung ventilation-perfusion combined functional area generation method based on planned four-dimensional CT as described in the first aspect.

[0035] The beneficial effects of this invention are as follows: Utilizing existing 4D CT image data from radiotherapy planning, lung ventilation and perfusion function images can be generated simultaneously without the need for additional SPECT, PET, or other independent functional images during model application or inference phases, thus reducing reliance on additional functional imaging equipment and procedures. By combining average anatomical information from the planned 4D CT scan with lung tissue structure changes under different respiratory states, lung ventilation and perfusion-related features are learned within a unified model. Information sharing between lung ventilation and perfusion tasks is achieved through a shared feature extraction path, maintaining the ability to generate both types of functional images while reducing repetitive feature extraction processes, lowering model parameter size and computational complexity. Based on the predicted lung ventilation and perfusion function images, the spatial overlap area is identified as a combined high-function region of lung ventilation and perfusion, providing a new technical means for functional lung evaluation, functional protection radiotherapy, and subsequent dose-function relationship studies.

[0036] The advantages of additional aspects of the invention will be set forth more clearly in the following description or will be learned by practice of the invention. Attached Figure Description

[0037] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is a schematic diagram of the overall process of the lung ventilation-perfusion combined functional region generation method based on planned four-dimensional CT as described in an embodiment of the present invention.

[0039] Figure 2This is a schematic diagram of the lung ventilation-perfusion combined functional imaging generation model according to an embodiment of the present invention.

[0040] Figure 3 This is a schematic diagram of an embodiment of the fully shared feature extraction structure described in this invention.

[0041] Figure 4 This is a schematic diagram illustrating an embodiment of the partially shared feature extraction structure described in this invention.

[0042] Figure 5 This is a flowchart of the training method for the lung ventilation-perfusion combined functional imaging generation model according to an embodiment of the present invention.

[0043] Figure 6 This is a flowchart of the lung ventilation-perfusion combined high-function region generation method according to an embodiment of the present invention.

[0044] Figure 7 This is a schematic diagram illustrating example results of the lung ventilation function image, lung perfusion function image, and combined high-function region generated according to an embodiment of the present invention. (A) represents the lung ventilation function image and lung perfusion function image; (B) represents the lung ventilation, lung perfusion, and combined high-function region.

[0045] Figure 8 This is a comparison chart of the overall case-level performance of the three modeling strategies described in the embodiments of the present invention. Figure 8 A represents the Spearman correlation coefficient, SSIM, and Dice coefficient for ventilation synthesis; B represents the corresponding indices for perfusion synthesis; C represents the RMSE for ventilation and perfusion synthesis; and D represents the Dice coefficient and Jaccard index for the V / Q overlap area. Data are expressed as mean ± standard deviation at the case level. The V / Q overlap area for task-specific strategies was obtained from independently predicted ventilation and perfusion maps. Detailed Implementation

[0046] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0047] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0048] It should also be understood that terms such as those defined in general dictionaries should be understood to have meanings consistent with their meanings in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as here.

[0049] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, and / or groups thereof.

[0050] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0051] To facilitate understanding of the present invention, the present invention will be further explained and described below with reference to the accompanying drawings and specific embodiments. However, the specific embodiments do not constitute a limitation on the embodiments of the present invention.

[0052] Those skilled in the art should understand that the accompanying drawings are merely schematic diagrams of embodiments, and the components in the drawings are not necessarily essential for implementing the present invention.

[0053] Example 1

[0054] Combination Figure 1 , Figure 5 As shown in this embodiment 1, a training method for a lung ventilation-perfusion combined functional imaging generation model is first provided. The lung ventilation-perfusion combined functional imaging generation model is obtained through training by this method. Using this model, lung ventilation functional images and lung perfusion functional images can be synthesized.

[0055] Specifically, the lung ventilation-perfusion combined functional imaging generation model adopts a three-dimensional encoder-decoder network structure.

[0056] like Figure 2 As shown, the model includes: an input layer, a shared feature extraction unit, a lung ventilation function reconstruction unit, and a lung perfusion function reconstruction unit.

[0057] The input layer is used to input the obtained three-channel CT data; wherein, of the three input channels, the first channel corresponds to the average CT image, the second channel corresponds to the peak inspiratory CT image, and the third channel corresponds to the peak expiratory CT image.

[0058] By combining CT information from different respiratory states, the model can simultaneously obtain information on the static anatomical structure of the lungs, changes in respiratory state, and lung tissue movement.

[0059] The shared feature extraction unit is used to extract multi-scale spatial features from the input data. For example... Figure 3 As shown, in this embodiment, the shared feature extraction unit adopts a three-dimensional residual coding structure, which includes multiple consecutive coding stages. Each coding stage includes a three-dimensional convolutional layer, a normalization layer, a nonlinear activation function, and a residual connection structure. Through layer-by-layer feature extraction, a multi-scale representation from low-level spatial structure information to high-level functionally related features is obtained.

[0060] The lung ventilation function reconstruction unit receives multi-scale features output by the shared feature extraction unit, restores spatial resolution through a decoding process, and finally outputs a three-dimensional lung ventilation function prediction image. The lung perfusion function reconstruction unit adopts a similar structure to the lung ventilation function reconstruction module. It generates a three-dimensional lung perfusion prediction function image based on shared features. .

[0061] In this embodiment, the lung ventilation and lung perfusion tasks share the encoding phase and bottleneck layer parameters; the two tasks each employ independent decoding paths. This structure enables the two tasks to share common features related to lung structure, while preserving the differential feature expressions between the ventilation and perfusion tasks.

[0062] In this embodiment, the steps for obtaining and processing the training data of the above model are as follows:

[0063] Specifically, four-dimensional CT image data of radiotherapy planning for patients with thoracic tumors and their paired pulmonary function reference images were acquired. In this embodiment, data from 65 patients with thoracic malignant tumors were acquired, including patients with lung cancer and esophageal cancer. All patients underwent four-dimensional CT scans for radiotherapy planning and completed pulmonary ventilation / perfusion SPECT examinations within a similar timeframe. The four-dimensional CT images for planning were acquired using a radiotherapy simulation positioning CT device, and the pulmonary ventilation / perfusion images for SPECT were acquired using nuclear medicine equipment. The acquired pulmonary ventilation and pulmonary perfusion function images were used as supervised targets during the training of the deep learning model. For each patient's images, average CT images, peak inspiratory CT images, and peak expiratory CT images were selected from the four-dimensional CT planning data. The peak inspiratory CT images and peak expiratory CT images reflect the expansion and contraction states of lung tissue during the patient's respiratory cycle, respectively.

[0064] Specifically, this embodiment includes 65 patients with histopathologically confirmed thoracic malignancies who underwent both planned 4D-CT and paired SPECT V / Q examinations between 2017 and 2021. These included 27 cases of lung cancer and 38 cases of esophageal cancer. All data used were legally obtained. The median age of the patients was 67 years (range: 40–86 years), with 53 males and 12 females. The median time interval between the planned 4D-CT and paired SPECT V / Q examinations was 2.1 days (range: 0.1–13.2 days). Inclusion criteria were the availability of both planned 4D-CT and paired SPECT V / Q data, with image quality sufficient for preprocessing, registration, and quantitative analysis. Cases with severe image artifacts, incomplete lung coverage, or failed image registration were excluded. In this embodiment, AVG-CT, T00, and T50 come from the same 4D-CT and serve as three-channel inputs in a common space. The actual rigid registration is performed between the concurrent low-dose CT and AVG-CT scans of SPECT / CT. Spatial correspondence processing mainly includes maintaining / unifying common coordinates, resampling, and cropping, etc., and does not necessarily require separate rigid or deformation registration between the three-phase CT scans. Basic patient characteristics are shown in Table 1.

[0065] Table 1

[0066]

[0067] All patients underwent simulated positioning in the radiotherapy position using a dedicated CT simulator (Philips BigBore Brilliance, Philips Healthcare, Andover, Massachusetts, USA). Respiratory monitoring was performed using a real-time position management system (RPM, Varian Medical Systems, Palo Alto, California, USA), and 4D-CT images were acquired during free breathing. Images were reconstructed using a standard filtered backprojection algorithm and segmented into 10 time phases (T00–T90) based on respiratory timing, with T00 and T50 corresponding to peak inspiration and peak expiration, respectively. An average CT with the same spatial resolution (AVG-CT) was further reconstructed from the 10 time-phase data. All images had an intra-slice matrix of 512×512 and a slice thickness of 3 mm.

[0068] Paired SPECT / CT V / Q imaging was performed using an Infinia Hawkeye4 SPECT / CT system (GE Healthcare, Milwaukee, Wisconsin, USA) equipped with a low-energy, high-resolution parallel-aperture collimator. Ventilation SPECT / CT was acquired after the patient inhaled 99mTc-Technegas prepared from 555 MBq of 99mTc-pertechnetate, with a 64×64 acquisition matrix, 3° angular step, and 20 s acquisition time per projection. Subsequently, while the patient remained supine, approximately 222 MBq of 99mTc-large particle polymerized albumin (99mTc-MAA, radiochemical purity >95%) was administered intravenously, followed immediately by perfusion SPECT / CT acquisition; all parameters remained the same except for a 15 s acquisition time per projection. Low-dose CT was acquired simultaneously for attenuation correction and anatomical localization. Ventilation and perfusion SPECT images were reconstructed using an iterative algorithm incorporating Butterworth filtering (cutoff frequency 0.5) and fused with concurrent CT images on a GE Xeleris 3.0 workstation.

[0069] Further preprocessing of the acquired CT images and reference functional images includes:

[0070] (1) Spatial registration processing. Using the average CT image as a unified reference space, the lung ventilation function images and lung perfusion function images obtained by SPECT / CT are converted to the average CT space through image registration. The registration method can be rigid registration.

[0071] (2) Image resampling processing. Images from different sources and at different resolutions are resampled to a uniform spatial resolution to ensure the spatial correspondence between CT images and functional images.

[0072] (3) Lung region extraction. Initial lung region segmentation was performed based on the CT value range (-999 to -250 HU), and the final lung region mask was obtained by combining manual correction.

[0073] (4) Image normalization processing. The CT values ​​and functional image intensities in the lung region are normalized to reduce the differences in image intensity among different patients.

[0074] All images were preprocessed and analyzed after anonymization. For each patient, planned 4D-CT and paired SPECT / CT V / Q data were unified to a reference space defined by AVG-CT. Low-dose CT acquired concurrently with SPECT / CT was rigidly registered to AVG-CT using MIMvista 6.8 (MIM Software Inc., Cleveland, Ohio, USA), and the resulting transformation was applied to the corresponding ventilation and perfusion SPECT images. The registration results were reviewed by an experienced radiation oncologist.

[0075] Subsequently, all images were resampled to isotropic voxels of 1.5 mm × 1.5 mm × 1.5 mm and cropped to a fixed size of 192 × 192 × 192 mm according to the lung bounding box, with 16 voxels retained at the edges. A lung mask was generated on the reference CT using a threshold range of -999 to -250 HU, and the trachea and main bronchus were manually excluded. The CT values ​​were truncated to -1000 to 400 HU and then linearly scaled by dividing by -2000; ventilation and perfusion SPECT images were normalized to minimum and maximum within the lung mask. The registered ventilation and perfusion SPECT images were used as reference images for ventilation and perfusion, respectively.

[0076] After the above processing, the final combined input data is obtained: ;in, Represents the average CT image; This represents a CT image during inspiration. This is a CT image showing the expiratory phase. Supervision label: This refers to imaging of lung ventilation function. This refers to the lung perfusion function imaging.

[0077] In this embodiment, after acquiring the training data, the training input data for the lung ventilation-perfusion combined functional image generation model is the planned four-dimensional CT three-channel data (i.e., the final combined input data), and the training objective and model output data are the predicted lung ventilation function images. Images of lung perfusion function .

[0078] Furthermore, this embodiment provides a detailed description of the construction and training process of the lung ventilation-perfusion combined functional imaging generation model. The lung ventilation-perfusion combined functional imaging generation model is trained using paired planning 4D CT images, reference lung ventilation function images, and reference lung perfusion function images. After training, the planning 4D CT image of the object to be analyzed is input into the lung ventilation-perfusion combined functional imaging generation model, which simultaneously yields predicted lung ventilation function images and predicted lung perfusion function images.

[0079] In this embodiment, the lung ventilation-perfusion combined functional imaging generation model adopts an encoder-decoder network structure based on 3D residual U-Net, including an input layer, a shared feature extraction unit, a lung ventilation function reconstruction unit, and a lung perfusion function reconstruction unit. The shared feature extraction unit includes four sequentially arranged encoding stages and a bottleneck network; the lung ventilation function reconstruction unit and the lung perfusion function reconstruction unit each include independent task-specific decoding networks. Skip connections are set between the encoder and the decoder at the corresponding scale to transfer the spatial detail features extracted in the encoding stage to the decoding stage.

[0080] The data processing procedure of the model can be represented as follows: ;in, This represents the three-channel planned four-dimensional CT data input to the model; The parameter is A model for generating combined ventilation-perfusion functional images of the lungs; This represents the predicted lung ventilation function images generated by the model. This represents the predicted lung perfusion function images generated by the model.

[0081] The input layer is used to input the spatially mapped and preprocessed three-channel CT data. Specifically, the average CT image, the peak inspiratory CT image, and the peak expiratory CT image are combined along the channel dimension to obtain: ;in, Represents the average CT image. This represents a CT image showing peak inspiration. This represents a CT image showing the peak expiratory state. This indicates a stitching operation along the channel dimension. In this embodiment, the peak inspiratory CT image corresponds to the T00 phase in the planned 4D-CT, and the peak expiratory CT image corresponds to the T50 phase. After preprocessing, the three images have the same spatial dimensions and voxel correspondences, therefore they are spatially similar. A three-dimensional channel vector is formed at this location:

[0082] ;

[0083] This allows the model to simultaneously obtain average anatomical structure information and spatial changes in lung tissue under different respiratory states from the same input.

[0084] The shared feature extraction unit is used to extract multi-scale three-dimensional spatial features layer by layer from the three-channel input data. In the fully shared joint model of this embodiment, the lung ventilation prediction task and the lung perfusion prediction task jointly use all four encoding stages and the bottleneck network. The number of feature channels in the four encoding stages are 8, 16, 32 and 64 respectively, and the number of feature channels in the bottleneck network is 128.

[0085] Let the first The input features of each encoding stage are Features are obtained after this encoding stage. The encoding stage can be summarized as follows: ;in, Indicates the first A three-dimensional residual feature extraction module, .

[0086] Each residual feature extraction module includes two 3×3×3 three-dimensional convolutional layers, batch normalization processing corresponding to each convolutional operation, LeakyReLU nonlinear activation, and residual shortcut connections.

[0087] For a 3D convolution operation, let the input feature tensor be... The output feature tensor is Then the first Position in each output channel The convolution result at point can be represented as:

[0088] ;

[0089] in, Indicates the number of input feature channels; Indicates the input channel index; Indicates the output channel index; Indicates input channel Mapped to output channel 3D convolution kernel weights; This indicates the position index of the convolution kernel in three spatial directions; in this embodiment, the convolution kernel size is 3×3×3. Indicates the first The bias parameters of each output channel.

[0090] The batch normalization layer is used to normalize the feature channels of the convolution output. For the input value of a specific feature channel within a batch... Its normalized output can be expressed as:

[0091] ;

[0092] in, This represents the mean value of the feature channel corresponding to the current batch; This represents the variance of the feature channel corresponding to the current batch; It is the stability constant; and These are the scale parameters and translation parameters learned during model training, respectively.

[0093] The nonlinear activation function is LeakyReLU, which can be expressed as:

[0094] ;

[0095] in, Indicates the input to the activation function; This indicates the output after activation; The slope coefficient of the negative half-axis is preset, and satisfies the following conditions: .

[0096] Each residual module contains two consecutive 3D convolutional transformations. Let the nonlinear mapping consisting of two convolutional layers, a batch normalization layer, and an activation function be denoted as... Then the output of the residual module can be expressed as:

[0097] ;

[0098] in, This represents the input characteristics of the residual module; This indicates the output characteristics of the residual module; Represents the residual mapping of the main branch; This represents the trainable parameters in the main branch; This indicates a shortcut branch.

[0099] When the spatial dimensions and number of channels of the input and output features are the same: ;

[0100] When the number of channels in the input features and the output features are inconsistent, the shortcut branch uses 1×1×1 3D convolution combined with batch normalization for projection transformation: ;in, This represents a 1×1×1 projective convolution kernel. Indicates the corresponding bias parameter, symbol " "" indicates a three-dimensional convolution operation.

[0101] During the encoding process, 2×2×2 max pooling is used for stepwise downsampling of the input features. The result of max pooling can be expressed as: ;in, This represents the max pooling operation; Represents a 2×2×2 local neighborhood; Indicates the feature channel.

[0102] In this embodiment, the preprocessed input image size is 192×192×192. When using 2×2×2 max pooling for stepwise downsampling, the four encoding stages and the bottleneck layer form multi-scale spatial features, the typical scale of which is expressed as follows:

[0103]

[0104]

[0105]

[0106]

[0107]

[0108] in, , , and These represent the multi-scale features obtained in the first to fourth encoding stages, respectively. This represents the high-level characteristics of the bottleneck network output.

[0109] The bottleneck network, located between the encoder and decoder networks, is used to further integrate the high-level information extracted by the encoder at the lowest spatial resolution and the highest feature channel dimension. If the features after downsampling in the fourth encoding stage are represented as... The bottleneck characteristics can then be expressed as: ;in, This represents the residual feature transformation in the bottleneck layer.

[0110] The "task specificity" mentioned in this embodiment refers to the fact that although the lung ventilation function reconstruction unit and the lung perfusion function reconstruction unit have the same or corresponding network topology, they have independent sets of trainable parameters. During the model training process, the parameters are updated according to the lung ventilation supervision target and the lung perfusion supervision target, respectively.

[0111] The lung ventilation task decoding network is denoted as... The lung perfusion task decoding network is denoted as ,but:

[0112]

[0113]

[0114] The lung ventilation function reconstruction unit comprises four consecutive decoding phases. Let the first... The input features of the decoding stage are First, perform 2×2×2 three-dimensional transposed convolution upsampling: ;

[0115] in, Indicates the lung ventilation task Level 3 transposed convolution; This indicates the lung ventilation task characteristics after upsampling.

[0116] Subsequently, the upsampled features are concatenated with the skip connection features generated by the encoder at the same spatial scale using channel dimensions: ;

[0117] Then, through the residual feature transformation of the corresponding decoding stage, the following can be obtained: ;in, Indicates the lung ventilation task Each task-specific decoding residual module.

[0118] The final decoding stage outputs 8-channel high-resolution features. Mapped to a single output channel via a 1×1×1 3D convolution: ;in, A 1×1×1 convolutional kernel representing the output layer of the lung ventilation task; Indicates the corresponding bias; Images that predict lung ventilation function .

[0119] The lung perfusion function reconstruction unit adopts a four-level decoding structure corresponding to the lung ventilation function reconstruction unit, but its model parameters are independent. Let the lung perfusion task be... Level decoding features are ,but:

[0120]

[0121]

[0122]

[0123] High-resolution features are obtained after four levels of decoding. Single-channel lung perfusion functional volume data were generated through 1×1×1 three-dimensional convolution: ;in, and These represent the convolution kernel parameters and bias parameters of the output layer for the lung perfusion task, respectively; the obtained... Images that predict lung perfusion function .

[0124] The data flow of a preferred fully shared architecture can be summarized as follows:

[0125]

[0126]

[0127]

[0128] in, This indicates a shared encoder and a bottleneck network. Indicates shared parameters; and These are independent parameters for the lung ventilation and lung perfusion task decoding networks, respectively.

[0129] During model training, reference lung ventilation SPECT images and reference lung perfusion SPECT images spatially registered with the planned 4D-CT were used as supervised targets. The lung region mask was set as follows: The number of voxels in the lung region is:

[0130]

[0131] Lung ventilation task using masked L1 loss:

[0132]

[0133] Lung perfusion mission using masked L1 loss:

[0134]

[0135] The total loss of the joint model is:

[0136]

[0137] in, and They represent the first Reference lung ventilation function value and reference lung perfusion function value for each lung endogenous body element; and These represent the lung ventilation function value and lung perfusion function value predicted by the model, respectively.

[0138] This embodiment uses the Adam optimizer for parameter updates, with an initial learning rate of... The batch size is 1, and the maximum number of training rounds is 300. Training stops when the validation set loss fails to improve for 30 consecutive rounds. During training, random spatial flipping and translation are used for data augmentation, and the same spatial transformation is applied simultaneously to the three input CT channels, the reference lung ventilation function image, the reference lung perfusion function image, and the lung region mask to maintain the voxel-by-voxel spatial correspondence.

[0139] The training process described above yields a trained lung ventilation-perfusion combined functional image generation model. For a new object to be analyzed, the average CT image, peak inspiratory CT image, and peak expiratory CT image obtained through the same preprocessing are used as a three-channel input. This allows for the simultaneous generation of spatially corresponding three-dimensional predicted lung ventilation functional images and three-dimensional predicted lung perfusion functional images through a single forward inference. Figure 7 As shown.

[0140] Example 2

[0141] In this second embodiment, a performance comparison experiment of the model trained in the first embodiment is provided.

[0142] In this embodiment, four models based on planned 4D-CT synthetic lung ventilation and perfusion images are constructed, including a mission-specific ventilation model, a mission-specific perfusion model, a fully shared joint model, and a partially shared joint model. All models use three registered CT volumes (AVG-CT, T00, and T50) as input, which are then stitched together into three channels. The networks all use a three-dimensional residual U-Net (ResUNet3D) encoder-decoder structure with multi-layer skip connections as their backbone.

[0143] All networks have a base channel count of 8. The task-specific ventilation and perfusion models use the same ResUNet3D architecture but are trained independently for different target images. The network consists of four encoder stages, one bottleneck layer, and four decoder stages, with skip connections between the encoder and decoder layers. The encoder channels are 8, 16, 32, and 64 respectively, and the bottleneck layer has 128 channels. The decoder symmetrically reduces the number of feature channels from 128 to 64, 32, 16, and 8, and finally generates a single-channel synthetic functional image through a 1×1×1 convolution. The ventilation task-specific model uses SPECT ventilation images as the prediction target, while the perfusion task-specific model uses SPECT perfusion images as the prediction target.

[0144] The joint learning model employs two dual-output ResUNet3D architectures, simultaneously synthesizing ventilation and perfusion images from the same set of three-channel CT inputs. In the fully shared joint model, the two prediction tasks share all encoder stages and bottleneck layers. The shared encoder extracts multi-scale features and passes them to the two task-specific decoders via skip connections. The ventilation and perfusion decoders are structurally identical but parameter-independent, each outputting a single-channel synthetic functional image. This model shares all encoding parameters between the two tasks while preserving task-specific reconstruction pathways.

[0145] like Figure 4As shown, in the partially shared joint model, ventilation and perfusion predictions share only the low-level encoder. Specifically, the first three encoder stages are shared, with output channels of 8, 16, and 32 respectively. The network then splits into ventilation-specific and perfusion-specific high-level coding branches, each containing one additional encoder stage and one bottleneck layer, corresponding to 64 and 128 channels respectively, and connected to independent decoders. Shared low-level features are simultaneously input to both decoders via skip connections, while high-level coding features are only connected to the corresponding task-specific decoder. This structure aims to share low-level anatomical representations while preserving task-specific high-level feature extraction and image reconstruction capabilities.

[0146] All residual blocks in the network contain two 3×3×3 convolutional layers, a batch normalized layer, and a LeakyReLU activation function, with identity or projective shortcut connections configured. When the number of input and output channels is inconsistent, the shortcut branch uses a 1×1×1 convolution with batch normalization. Downsampling uses 2×2×2 max pooling, and upsampling uses 2×2×2 transposed convolutions, followed by concatenation with the skip connection features of the corresponding layer.

[0147] All models employed patient-level stratified five-fold cross-validation to obtain robust out-of-fold performance estimates. Stratification was performed based on diagnostic type to ensure a roughly consistent proportion of lung and esophageal cancer cases across all folds. For each fold, one subset was used as an independent test set, with the remaining patients used for training and validation. Further stratified randomization was performed from the training subset, with 20% of patients used as the validation set for model selection. All learning strategies used identical training, validation, and test set partitions to ensure fairness in comparisons.

[0148] Both the task-specific ventilation and perfusion models were trained using the masked L1 loss between the predicted lung image and the corresponding SPECT reference image. The total loss of the joint model was defined as the weighted sum of the masked L1 losses for ventilation and perfusion prediction. All networks used the Adam optimizer with an initial learning rate of [missing information]. The batch size is 1, and the maximum number of training rounds is 300. An early stopping strategy is adopted based on the validation loss, and the patience value is set to 30 rounds.

[0149] During training, random spatial flipping and translation were used for data augmentation, and the same spatial transformation was applied simultaneously to the input CT, reference SPECT image, and lung mask. For the joint model, the ventilation and perfusion target images were also transformed synchronously to maintain voxel-level correspondence. For each fold, the model weights with the lowest validation loss were selected to predict independent test cases. Finally, all out-of-fold prediction results from the 5-fold cross-validation were summarized at the case level and quantitatively compared.

[0150] All models were implemented in PyTorch and trained on a workstation equipped with an NVIDIA GeForce RTX 2080 Ti GPU (11 GB VRAM). To evaluate computational efficiency, the number of trainable parameters for each model strategy was recorded. The forward inference time for each independent test case was measured by performing CUDA synchronization before and after model execution. The inference time for task-specific strategies was defined as the sum of the inference times for the independent ventilation and perfusion models.

[0151] Model performance was evaluated at three levels: voxel-level image similarity, spatial consistency of high-functional areas, and spatial consistency of V / Q overlap regions. All metrics were calculated within the lung mask.

[0152] Specifically, for voxel-level image similarity, Spearman correlation coefficient, structural similarity index (SSIM), mean absolute error (MAE), and root mean square error (RMSE) are used to quantitatively evaluate the voxel-level consistency between the synthesized ventilation and perfusion maps and the corresponding SPECT reference images. Ventilation and perfusion indices are calculated separately; the two types of indices of the joint model are obtained based on the corresponding task output branches.

[0153] For high-function areas consistency, the highest voxel intensities were selected within the lung mask of both the synthetic and reference images to define high-function ventilation and high-function perfusion areas. The Dice similarity coefficient and Jaccard index were used to evaluate the spatial consistency between the high-function areas identified in the synthetic and reference images.

[0154] For ventilation / perfusion overlap consistency, the V / Q overlap is defined as the intersection of the high-function ventilation area and the high-function perfusion area. The reference overlap is obtained from paired SPECT ventilation and perfusion maps; the overlap of the joint model is obtained from synchronously predicted ventilation and perfusion outputs; and the overlap of the task-specific strategy is obtained from the predictions of two independent models. The Dice similarity coefficient and Jaccard index are used to evaluate the spatial consistency between the predicted overlap and the reference overlap.

[0155] Based on the data distribution, quantitative data are expressed as mean ± standard deviation or median (interquartile range). Final statistical comparisons were performed at the case level, including independent test results from a 5-fold cross-validation summary; each patient had only one out-of-fold prediction result for each model strategy. Since all three model strategies were evaluated within the same group of cases, paired statistical methods were used for comparison. First, the Friedman test was used to evaluate the overall differences between the task-specific, fully shared, and partially shared strategies; when the overall differences were statistically significant, the Wilcoxon signed-rank test was used for pairwise comparisons, and the Bonferroni method was used to correct for multiple comparisons. Ventilation, perfusion, and V / Q overlap indices were statistically analyzed separately. The V / Q overlap indices for the task-specific strategy were calculated from independently predicted ventilation and perfusion maps and included in the same comparison frame as the two combined models. An exploratory subgroup analysis was also performed for lung cancer and esophageal cancer patients by diagnostic type. Given the limited sample size of the subgroups, the results are primarily used to describe the consistency of model performance across different diagnostic groups and are not used to establish the superiority of any particular model in a specific disease. A two-sided p-value < 0.05 was considered statistically significant. Statistical analysis was performed using Python 3.8.3 and SciPy 1.10.1.

[0156] Overall case-level performance Figure 8 All three learning strategies were able to recover the major spatial distributions of ventilation and perfusion from planned 4D-CT scans. Regardless of the strategy used, perfusion prediction showed higher voxel-level correlation, SSIM, and high-function area consistency than ventilation prediction. Joint learning and task-specific learning had similar overall performance but did not show a consistent advantage. The V / Q overlap of the task-specific strategy was obtained from independently predicted ventilation and perfusion maps; the high-function area was defined as the 30% with the highest voxel intensity within the lung mask.

[0157] The fully shared model showed the highest numerical values ​​for the ventilation Spearman correlation coefficient, the ventilation high-function area Dice coefficient, and the consistency of the V / Q overlap region; the partially shared model showed the highest numerical values ​​for the perfusion Spearman correlation coefficient and the consistency of the perfusion high-function area. However, the absolute differences between the strategies were small, and neither of the two joint learning strategies showed a consistent advantage across all evaluation metrics.

[0158] The fully shared joint model has 1.98 million trainable parameters, significantly fewer than the 2.87 million parameters of the task-specific joint strategy; the partially shared joint model has 2.82 million parameters. The single-instance inference time for the task-specific strategy is (0.4164±0.0125) s, while that for the fully shared and partially shared joint models are (0.3098±0.0121) s and (0.3114±0.0113) s, respectively. Therefore, both joint models reduce inference time by approximately 25% while simultaneously generating ventilation and perfusion images.

[0159] Compared to perfusion, ventilation prediction showed lower overall consistency with SPECT-referenced images. The Spearman correlation coefficient, SSIM, MAE, and RMSE for the task-specific model were 0.722±0.113, 0.651±0.127, 0.135±0.033, and 0.173±0.037, respectively; for the fully shared model, they were 0.727±0.106, 0.651±0.126, 0.137±0.036, and 0.176±0.041, respectively; and for the partially shared model, they were... The values ​​were 0.720±0.118, 0.637±0.130, 0.136±0.035, and 0.176±0.039, respectively. The Dice coefficient / Jaccard index of the high-functional region of the three models were 0.650±0.082 / 0.487±0.086, 0.653±0.080 / 0.490±0.086, and 0.647±0.087 / 0.483±0.090, respectively.

[0160] Friedman's test showed that only the overall difference in ventilation-specific SSIM was statistically significant (P=0.022). Pairwise comparisons showed that the partially shared model was inferior to the task-specific model (P=0.045 after Bonferroni correction) and the fully shared model (P=0.024 after Bonferroni correction). There were no statistically significant differences in ventilation Spearman correlation coefficient, MAE, RMSE, Dice coefficient for high-functioning areas, and Jaccard index.

[0161] Perfusion prediction showed better agreement with SPECT images than ventilation prediction, and the results were more stable. The Spearman correlation coefficient, SSIM, MAE, and RMSE of the task-specific model were 0.813±0.099, 0.783±0.110, 0.126±0.037, and 0.160±0.047, respectively; those of the fully shared model were 0.817±0.087, 0.787±0.103, 0.124±0.037, and 0.159±0.045, respectively; and those of the partially shared model were... The values ​​were 0.819±0.090, 0.782±0.112, 0.124±0.039, and 0.159±0.047, respectively. The Dice coefficient / Jaccard index of the high-functional region of the three models were 0.733±0.057 / 0.581±0.068, 0.733±0.061 / 0.582±0.072, and 0.737±0.056 / 0.587±0.068, respectively.

[0162] Although the Spearman correlation coefficient and high-functional-area consistency were numerically highest in the partially shared model, and the SSIM was numerically highest in the fully shared model, the Friedman test showed no statistically significant differences in any of the perfusion indices among the three strategies. The perfusion Spearman correlation coefficient only showed a trend of difference between models (P=0.060).

[0163] Example 3

[0164] In this embodiment 3, a lung ventilation and perfusion combined functional region generation system based on planned 4D CT is first provided. This system includes: an acquisition module for acquiring planned 4D computed tomography (CT) image data acquired by an acquisition device; a processing module for acquiring average CT images, inspiratory CT images, and expiratory CT images from the planned 4D CT image data, and performing spatial correspondence processing on the average CT images, inspiratory CT images, and expiratory CT images to obtain combined data containing information on different respiratory states; and a prediction generation module for processing the combined data using a pre-trained lung ventilation-perfusion combined functional image generation model to generate three-dimensional lung ventilation functional images and three-dimensional lung perfusion functional images. The lung ventilation-perfusion combined functional image generation model is trained using a training set, which includes multiple sets of planned 4D CT image data, and reference lung ventilation functional images and reference lung perfusion functional images spatially corresponding to each set of planned 4D CT image data. The determination module is used to determine the high-function region of lung ventilation based on the three-dimensional lung ventilation function image, and to determine the high-function region of lung perfusion based on the three-dimensional lung perfusion function image; based on the spatial overlap relationship between the high-function region of lung ventilation and the high-function region of lung perfusion in the same spatial coordinate system, the spatial intersection of the two is determined as the combined high-function region of lung ventilation and perfusion.

[0165] like Figure 6As shown, in this embodiment, the above-described system is used to implement a method for generating combined ventilation and perfusion functional regions based on planned four-dimensional CT, including: acquiring planned four-dimensional computed tomography (CT) image data collected by an acquisition device, such as a radiotherapy simulation positioning CT device, using an acquisition module; acquiring average CT images, inspiratory CT images, and expiratory CT images from the planned four-dimensional CT image data, and performing spatial correspondence processing on the average CT images, inspiratory CT images, and expiratory CT images to obtain combined data containing information on different respiratory states; using a pre-trained lung ventilation-perfusion combined functional image generation model, processing the combined data to generate three-dimensional lung ventilation functional images and three-dimensional lung perfusion functional images; wherein, the lung ventilation-perfusion combined functional image generation model... The combined functional image generation model is trained using a training set, which includes multiple sets of planned four-dimensional computed tomography (CT) image data, as well as reference lung ventilation function images and reference lung perfusion function images spatially corresponding to each set of planned four-dimensional CT image data. High-function lung ventilation regions are determined based on the three-dimensional lung ventilation function images, and high-function lung perfusion regions are determined based on the three-dimensional lung perfusion function images. The spatial intersection of the high-function lung ventilation regions and the high-function lung perfusion regions in the same spatial coordinate system is defined as the combined high-function lung ventilation-perfusion region.

[0166] In this embodiment, the model trained in Example 1 is used to obtain the predicted lung ventilation function map. and predictive lung perfusion function map First, the analysis area is defined based on the lung region mask; then, the functional values ​​of all voxels within the lung region are sorted.

[0167] Based on the ranking of pulmonary ventilation function values: select a certain proportion of voxels with the highest function values:

[0168]

[0169] in: Indicates the selection ratio.

[0170] Identification of high-functioning areas of lung perfusion

[0171] Similarly:

[0172] Calculation for determining the high-functioning area of ​​the combined ventilation-perfusion system: ;

[0173] The resulting area is the high-functioning area of ​​combined ventilation and perfusion.

[0174] In this embodiment: Select ratio: That is, select the region with the highest 30% pulmonary ventilation function value and the region with the highest 30% pulmonary perfusion function value respectively; the spatial intersection of the two is the joint high-function region.

[0175] Example 4

[0176] This embodiment 4 provides a non-transitory computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, they implement the lung ventilation-perfusion combined functional region generation method based on planned four-dimensional CT as described above. The method includes:

[0177] Acquire planned four-dimensional computed tomography image data acquired by the acquisition device;

[0178] The average computed tomography (CT) image, the inspiratory CT image, and the expiratory CT image are obtained from the planned four-dimensional computed tomography image data. Spatial correspondence processing is performed on the average CT image, the inspiratory CT image, and the expiratory CT image to obtain joint data containing information on different respiratory states.

[0179] The combined data is processed using a pre-trained lung ventilation-perfusion combined functional image generation model to generate three-dimensional lung ventilation function images and three-dimensional lung perfusion function images. The lung ventilation-perfusion combined functional image generation model is trained from a training set, which includes multiple sets of planned four-dimensional computed tomography (CT) image data, as well as reference lung ventilation function images and reference lung perfusion function images corresponding to each set of planned four-dimensional CT image data.

[0180] The high-function areas of lung ventilation are determined based on the three-dimensional lung ventilation function images, and the high-function areas of lung perfusion are determined based on the three-dimensional lung perfusion function images.

[0181] Based on the spatial overlap between the high-functioning pulmonary ventilation region and the high-functioning pulmonary perfusion region in the same spatial coordinate system, the spatial intersection of the two is determined as the combined high-functioning pulmonary ventilation-perfusion region.

[0182] Example 5

[0183] This embodiment 5 provides a computer device, including a memory and a processor, wherein the processor and the memory communicate with each other, and the memory stores program instructions that can be executed by the processor. The processor calls the program instructions to execute the lung ventilation-perfusion combined functional region generation method based on planned four-dimensional CT as described above, the method including:

[0184] Acquire planned four-dimensional computed tomography image data acquired by the acquisition device;

[0185] The average computed tomography (CT) image, the inspiratory CT image, and the expiratory CT image are obtained from the planned four-dimensional computed tomography image data. Spatial correspondence processing is performed on the average CT image, the inspiratory CT image, and the expiratory CT image to obtain joint data containing information on different respiratory states.

[0186] The combined data is processed using a pre-trained lung ventilation-perfusion combined functional image generation model to generate three-dimensional lung ventilation function images and three-dimensional lung perfusion function images. The lung ventilation-perfusion combined functional image generation model is trained from a training set, which includes multiple sets of planned four-dimensional computed tomography (CT) image data, as well as reference lung ventilation function images and reference lung perfusion function images corresponding to each set of planned four-dimensional CT image data.

[0187] The high-function areas of lung ventilation are determined based on the three-dimensional lung ventilation function images, and the high-function areas of lung perfusion are determined based on the three-dimensional lung perfusion function images.

[0188] Based on the spatial overlap between the high-functioning pulmonary ventilation region and the high-functioning pulmonary perfusion region in the same spatial coordinate system, the spatial intersection of the two is determined as the combined high-functioning pulmonary ventilation-perfusion region.

[0189] Example 6

[0190] This embodiment 6 provides an electronic device, including: a processor, a memory, and a computer program; wherein, the processor is connected to the memory, and the computer program is stored in the memory. When the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute instructions for implementing the lung ventilation-perfusion combined functional region generation method based on planned four-dimensional CT as described above, the method including:

[0191] Acquire planned four-dimensional computed tomography image data acquired by the acquisition device;

[0192] The average computed tomography (CT) image, the inspiratory CT image, and the expiratory CT image are obtained from the planned four-dimensional computed tomography image data. Spatial correspondence processing is performed on the average CT image, the inspiratory CT image, and the expiratory CT image to obtain joint data containing information on different respiratory states.

[0193] The combined data is processed using a pre-trained lung ventilation-perfusion combined functional image generation model to generate three-dimensional lung ventilation function images and three-dimensional lung perfusion function images. The lung ventilation-perfusion combined functional image generation model is trained from a training set, which includes multiple sets of planned four-dimensional computed tomography (CT) image data, as well as reference lung ventilation function images and reference lung perfusion function images corresponding to each set of planned four-dimensional CT image data.

[0194] The high-function areas of lung ventilation are determined based on the three-dimensional lung ventilation function images, and the high-function areas of lung perfusion are determined based on the three-dimensional lung perfusion function images.

[0195] Based on the spatial overlap between the high-functioning pulmonary ventilation region and the high-functioning pulmonary perfusion region in the same spatial coordinate system, the spatial intersection of the two is determined as the combined high-functioning pulmonary ventilation-perfusion region.

[0196] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0197] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0198] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0199] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment, whereby a series of operational steps are performed to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0200] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that, based on the technical solutions disclosed in the present invention, various modifications or variations that can be made by those skilled in the art without creative effort should be included within the scope of protection of the present invention.

Claims

1. A method for generating combined pulmonary ventilation and perfusion functional regions based on planned four-dimensional CT, characterized in that, include: Acquire planned four-dimensional computed tomography image data acquired by the acquisition device; The average computed tomography (CT) image, the inspiratory CT image, and the expiratory CT image are obtained from the planned four-dimensional computed tomography image data. Spatial correspondence processing is performed on the average CT image, the inspiratory CT image, and the expiratory CT image to obtain joint data containing information on different respiratory states. The combined data is processed using a pre-trained lung ventilation-perfusion combined functional image generation model to generate three-dimensional lung ventilation function images and three-dimensional lung perfusion function images. The lung ventilation-perfusion combined functional image generation model is trained from a training set, which includes multiple sets of planned four-dimensional computed tomography (CT) image data, as well as reference lung ventilation function images and reference lung perfusion function images corresponding to each set of planned four-dimensional CT image data. The high-function areas of lung ventilation are determined based on the three-dimensional lung ventilation function images, and the high-function areas of lung perfusion are determined based on the three-dimensional lung perfusion function images. Based on the spatial overlap between the high-functioning pulmonary ventilation region and the high-functioning pulmonary perfusion region in the same spatial coordinate system, the spatial intersection of the two is determined as the combined high-functioning pulmonary ventilation-perfusion region.

2. The method for generating combined functional regions of lung ventilation and perfusion based on planned four-dimensional CT according to claim 1, characterized in that, The lung ventilation-perfusion combined functional imaging generation model includes a shared feature extraction unit, a lung ventilation function reconstruction unit, and a lung perfusion function reconstruction unit. The shared feature extraction unit is used to extract multi-scale spatial features from the combined data that simultaneously contain anatomical structural information and dynamic changes in respiratory motion. The lung ventilation function reconstruction unit is used to generate lung ventilation function images based on the multi-scale spatial features. The lung perfusion function reconstruction unit is used to generate lung perfusion function images based on the multi-scale spatial features.

3. The method for generating combined pulmonary ventilation and perfusion functional regions based on planned four-dimensional CT according to claim 2, characterized in that, The shared feature extraction unit includes an encoder network and a bottleneck network; the encoder network is connected to the decoder network at the corresponding scale to transfer the spatial detail features extracted in the encoding stage to the decoding stage; the bottleneck network is used to further integrate the high-level information extracted by the encoder network at the lowest spatial resolution and the highest feature channel dimension.

4. The method for generating combined pulmonary ventilation and perfusion functional regions based on planned four-dimensional CT according to claim 3, characterized in that, The encoder network includes multiple encoding stages, and the lung ventilation prediction task and the lung perfusion prediction task share all of the multiple encoding stages as well as the bottleneck network.

5. The method for generating combined functional regions of lung ventilation and perfusion based on planned four-dimensional CT according to claim 4, characterized in that, Both the lung ventilation function reconstruction unit and the lung perfusion function reconstruction unit include task-specific decoder networks. The task specificity means that the lung ventilation function reconstruction unit and the lung perfusion function reconstruction unit have the same network topology, but they have independent sets of trainable parameters. During training, the parameters are updated according to the lung ventilation supervision target and the lung perfusion supervision target, respectively. The parameters of the encoder network and the bottleneck network are shared by the lung ventilation function reconstruction unit and the lung perfusion function reconstruction unit, while the task-specific decoder networks are optimized independently.

6. The method for generating combined functional regions of lung ventilation and perfusion based on planned four-dimensional CT according to claim 4, characterized in that, In the four consecutive decoding stages, for the first The input features in the decoding stage are first upsampled by 2×2×2 three-dimensional transposed convolution; then, the upsampled features are concatenated with the skip connection features generated by the encoder at the same spatial scale in the channel dimension; and then transformed by the residual features corresponding to the decoding stage. The final decoding stage outputs 8-channel high-resolution features, which are mapped to one output channel through a 1×1×1 3D convolution.

7. The method for generating combined pulmonary ventilation and perfusion functional regions based on planned four-dimensional CT according to claim 1, characterized in that, During the training of the lung ventilation-perfusion combined functional imaging generation model, a joint loss function is constructed based on the differences between the predicted lung ventilation function image and the reference lung ventilation function image, as well as the differences between the predicted lung perfusion function image and the reference lung perfusion function image. The parameters of the lung ventilation-perfusion combined functional imaging generation model are then optimized using the joint loss function.

8. A system for generating combined functional regions of lung ventilation and perfusion based on planned four-dimensional CT, characterized in that, include: The acquisition module is used to acquire the planned four-dimensional computed tomography image data collected by the acquisition device; The processing module is used to obtain the average computed tomography image, the inspiratory computed tomography image, and the expiratory computed tomography image from the planned four-dimensional computed tomography image data, and to perform spatial correspondence processing on the average computed tomography image, the inspiratory computed tomography image, and the expiratory computed tomography image to obtain joint data containing information on different respiratory states. The prediction generation module is used to process the combined data using a pre-trained lung ventilation-perfusion combined functional image generation model to generate three-dimensional lung ventilation function images and three-dimensional lung perfusion function images; wherein, the lung ventilation-perfusion combined functional image generation model is trained by a training set, which includes multiple sets of planned four-dimensional computed tomography (CT) image data, and reference lung ventilation function images and reference lung perfusion function images corresponding to each set of planned four-dimensional CT image data space; The determination module is used to determine the high-function region of lung ventilation based on the three-dimensional lung ventilation function image, and to determine the high-function region of lung perfusion based on the three-dimensional lung perfusion function image; based on the spatial overlap relationship between the high-function region of lung ventilation and the high-function region of lung perfusion in the same spatial coordinate system, the spatial intersection of the two is determined as the combined high-function region of lung ventilation and perfusion.

9. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium is used to store computer instructions, which, when executed by a processor, implement the method for generating combined pulmonary ventilation and perfusion functional regions based on planned four-dimensional CT as described in any one of claims 1-6.

10. An electronic device, characterized in that, include: The device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute instructions for implementing the method for generating combined functional regions of lung ventilation and perfusion based on planned four-dimensional CT as described in any one of claims 1-6.

Images