An automatic conversion method of activity-dose in nuclear medicine based on generative adversarial networks

Abstract

The present invention discloses a method for automatic conversion of nuclear medicine activity-dose based on generative adversarial networks. Based on the whole-image conversion characteristics of generative adversarial networks, the dose and activity-density overlay maps are connected as reference image pairs, input into a generator G, and a pseudo absorbed dose image is generated, which is then mixed with a standard dose image and identified by a discriminator D. If the pseudo dose image is identified, the information will be fed back, and G will regenerate a new pseudo dose image until the identification rate of D decreases and the true and false dose images cannot be identified, so as to obtain a realistic dose image that can be mistaken for the real thing. The present invention is based on the voxel distribution map of nuclear medicine activity and density, and utilizes the whole-image conversion advantage of the generative adversarial network model from image to image, so as to realize the real-time automatic conversion of the absorbed dose corresponding to the instantaneous activity distribution of nuclear medicine internal irradiation, and takes into account the influence of the material density of the patient's anatomical image.

Description

Technical Field

[0001] The invention relates to a nuclear medicine activity-dose automatic conversion method, and in particular to a nuclear medicine activity-dose automatic conversion method based on a generative adversarial network. Background Art

[0002] In nuclear medicine tumor radiotherapy, after the nuclide is injected into the tumor site of the human body, although the antibody carrying the nuclide has a strong affinity for cancer, some nuclides will still be dispersed in various organs and tissues of the human body, forming an extremely uneven activity distribution. The Medical Internal Radiation Dose (MIRD) method currently used in clinical practice assumes that the activity of all voxels in the organ source is uniform throughout the organ, which is inconsistent with the actual clinical situation. In addition, MIRD obtains the average dose of the organ target unit, rather than the dose distribution of the entire organ.

[0003] The voxel S-value kernel method (VSV) proposed in recent years can obtain the overall dose distribution, but the pre-calculated searchable "voxel V1→voxel V2" dose list assumes that there is a uniform medium (such as water) between voxel V1 and voxel V2, which is inconsistent with the complex human anatomical information and brings deviations to the calculation results. In view of the complexity of the radionuclide activity distribution, the semi-analytical algorithms MIRD and VSV conventionally used in clinical practice are gradually being replaced by more accurate Monte Carlo simulation methods (MC) based on quantitative nuclear medicine images (such as PET, SPECT, etc.).

[0004] The MC method simulates the transport of a large number of radioactive particles based on the ionizing radiation physics model. It has a large amount of calculations and a long simulation time. It often only calculates the dose distribution corresponding to the activity distribution at the initial moment, and then uses the single exponential model of radioactivity decay or the double exponential model of activity absorption-clearance to calculate the cumulative dose distribution. It is difficult to obtain the dose distribution corresponding to the instantaneous activity distribution. It cannot meet the real-time dynamic display of dose changes required by the growing high-end medical services, and assist doctors in timely adjusting treatment plans based on feedback information.

[0005] Deep Learning (DL) is an artificial intelligence technology that has emerged in recent years. It has been widely used in the field of external radiotherapy for tumors, such as automatic delineation of tumor targets, optimization of radiotherapy strategies, and prediction of patient survival. However, the application of DL in the field of internal irradiation dose calculation in nuclear medicine is still very rare. Some abstract works (Lee et al. Journal of Nuclear Medicine, May 2021, 62 (supplement 1) 1174) briefly introduced their dose calculation based on the Convolution Neural Network (CNN) method, but the CNN method can only train the DL model based on image patches, and then classify the activity and dose image patches, and then "assemble" them into 2-D and 3-D dose distributions, not the overall image-to-image conversion. CNN is more suitable for automatic identification and delineation of tumor targets.

[0006] Generative Adversarial Networks (GANs) are known as the smartest deep learning algorithms, similar to the mutual restraint relationship between "police and thieves". This method inputs the dose and activity-density image pairs as reference images into the "thief" (Generator (G) to generate a pseudo absorbed dose image, which is then mixed with the standard dose image (Ground Truth, GT) and recognized by the "police" (Discriminator, (Discriminator, D)). If the pseudo image is recognized, the information will be fed back and G will regenerate the pseudo dose image until the error of D drops to a certain accuracy (determined by the loss function) to obtain a realistic dose image. Summary of the invention

[0007] The present invention aims to avoid the shortcomings of the above-mentioned prior art and provide a method for automatic conversion of nuclear medicine activity-dose based on generative adversarial networks. This method can achieve the dose distribution corresponding to the uniform and non-uniform distribution of nuclear medicine activity, avoid the error caused by the semi-analytical dose calculation method assuming that the activity is uniform inside the volume source, and overcome the weakness of the slow calculation speed of the precise Monte Carlo method.

[0008] To achieve the above object, the present invention provides the following technical solutions:

[0009] A method for automatic conversion of nuclear medicine activity-dose based on a generative adversarial network, the method comprising the following steps:

[0010] Step 1: delineate the lesion and organ at risk region in the patient's anatomical image, establish a material and density model of the patient based on the patient's anatomical image, and assign region numbers to all voxels of the delineated lesions and organs at risk in the material and density model in a one-to-one correspondence manner;

[0011] Step 2, according to the image of the activity space distribution of the nuclide at a certain time t provided by the functional imaging of the patient, obtaining the spatial distribution information of the activity of the nuclide at time t in the patient's body;

[0012] Step 3: Weighted superposition of the patient's radionuclide activity spatial distribution image and the corresponding patient's anatomical image density map.

[0013] AD voxel =A voxel *weigt+density voxel

[0014] Among them, A voxel is the activity of a certain voxel obtained in step 2, weigt is the uniform weighting of all voxel activities, density voxel is the material density of the voxel obtained in step 1, AD voxel is the obtained activity-density weighted value;

[0015] Step 4: Use nuclear medicine dose calculation method to obtain the patient's dose spatial distribution;

[0016] Step 5: The dose spatial distribution obtained in step 4 and the weighted activity-density distribution obtained in step 3 are plotted into an image with the same resolution as the original dose distribution using a plotting software, and connected into a left-right arranged image pair;

[0017] Step 6: Set all the dose distributions obtained in step 4 to 0, and use a drawing software to draw the weighted activity-density obtained in step 3 into an image with the same resolution as the original dose distribution, and connect them into equal left-right arranged image pairs;

[0018] Step 7: Send the dose and activity-density image pairs obtained in step 5 as training atlases to the generative adversarial network software, train the deep learning model based on the optimized model parameters, and generate a deep learning model that generates corresponding dose distribution from activity-density distribution;

[0019] Step 8: The 0 dose map and the activity-density overlay map obtained in step 6 are used as a test map set, and sent to the deep learning model generated in step 7 for generating the corresponding dose distribution from the activity-density distribution for testing, so as to generate a dose distribution corresponding to the activity-density distribution;

[0020] Step 9: Using the patient material and the region number in the density model established in step 1, sort out the voxel doses of lesions and organs at risk for calculation of indicators including average dose, gamma pass rate assessment, dose-volume histogram, and off-axis dose distribution.

[0021] Preferably, in step 4, the nuclear medicine dose calculation adopts the Monte Carlo method.

[0022] Preferably, in step 4, the dose distribution of the patient obtained by the nuclear medicine method is determined as follows: particle transport calculations are performed using the particle physics model of the classical Monte Carlo program EGSnrc / DOSXYZnrc to obtain the dose distribution corresponding to the nuclide activity distribution.

[0023] Preferably, in step 4, the nuclear medicine dose is calculated using a voxel S value method.

[0024] Preferably, in step 4, the dose distribution of the patient obtained by the nuclear medicine method is determined as follows: the dose distribution is obtained by looking up the table using the equivalent distance between the source voxel V1 and the target voxel V2.

[0025] Preferably, the equivalent distance between the source voxel V1 and the target voxel V2 in the voxel S value method is determined as follows: for the non-uniform medium distribution between the source voxel V1 and the target voxel V2, the equivalent path method conversion is performed according to the material density information obtained in step 1, that is,

[0026]

[0027] Among them, l i is the path length of any voxel passing through (voxel V1→voxel V2), ρ i is the density of the voxel being traversed, ρ w is the density of water, n is the number of all voxels that pass through (voxel V1 → voxel V2), L homo is the calculated equivalent path.

[0028] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0029] The present invention aims at the errors caused by the assumption that the activity is uniform inside the volume source by the MIRD currently used in clinical practice, and the assumption that the medium between the source voxel and the target voxel is uniform by the VSV, as well as the slow calculation speed of the precise Monte Carlo method. Based on the smartest deep learning algorithm, the generative adversarial network GAN is used to overlap the activity distribution with the material density distribution. Based on the dose and activity-material density image pairs, the GAN generator is used to generate a dose distribution map that is indistinguishable from the real thing, thereby realizing the rapid and automatic generation of the corresponding dose distribution. This nuclear medicine activity-dose automatic conversion method based on the generative adversarial network can overcome the subjective assumption disadvantage of the semi-analytical dose calculation method assuming uniform activity or material density, while avoiding the objective disadvantage of the slow calculation speed of the precise Monte Carlo method, and improving the speed and accuracy of the dose calculation of non-uniform activity distribution in nuclear medicine internal irradiation. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a flow chart of a method for automatic conversion of nuclear medicine activity-dose based on a generative adversarial network in an embodiment of the present invention;

[0031] Figure 2 This is an image pair of a dose map and an activity-density overlay map required for training the GAN deep learning model in an embodiment of the present invention, wherein the dose map is on the left and the activity-density overlay map is on the right.

[0032] Figure 3 The GAN deep learning model trained with a simple single-region source model image in the embodiment of the present invention predicts the absorption dose distribution map of the three-source organ model and the comparison with the standard GT absorption dose distribution map. Left column: input activity-density overlay map; middle: predicted dose distribution map; right column: standard GT dose distribution map. Figure 3 (a) is the case of uniform distribution Gauss 0.0 (i.e. σ = 0); Figure 3 (b) and (c) are the cases of non-uniform distribution Gauss0.3 (i.e., σ=0.3) and Gauss0.6 (i.e., σ=0.6). DETAILED DESCRIPTION

[0033] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0034] A method for automatic conversion of nuclear medicine activity-dose based on generative adversarial networks, such as Figure 1 As shown, the method is carried out in the following steps:

[0035] Step 1. Based on the patient's anatomical image (such as CT, MRI, etc.), establish the patient's material and density model. This part can be generated by the DOSXYZnrc front-end program ctcreate, which automatically converts CT values, density and materials, and generates phantoms of four materials (air, soft tissue, bone and lung tissue) (the user can also increase the division of material density). Outline the lesions and organs at risk in the anatomical image, and use different area numbers to arrange the voxels belonging to the lesions and organs at risk one by one, so as to facilitate the data collection and processing of the calculation results in the later stage;

[0036] Step 2, according to the spatial distribution image of the activity of the nuclide at a certain time t provided by the patient's functional imaging (such as SPECT, PET, etc.), obtain the spatial distribution information of the activity of the nuclide at time t in the patient's body;

[0037] Step 3: Weighted superposition of the patient's radionuclide activity spatial distribution image and the corresponding patient's anatomical image density map.

[0038] AD voxel =A voxel *weigt+density voxel

[0039] Among them, A voxel is the activity of a voxel obtained in step 2, weigt is the uniform weighting of all voxel activities (e.g. 10 times to highlight the distribution of activity), density voxel is the material density of the voxel obtained in step 1, AD voxel is the obtained activity-density weighted value;

[0040] Step 4: Using the Monte Carlo method or other nuclear medicine dose calculation methods to obtain the patient's dose distribution;

[0041] For example, the particle transport physics model of the classical Monte Carlo program EGSnrc / DOSXYZnrc is used to simulate particle transport in the patient anatomical model, and the simulation parameters are optimized (for example, a low particle transport cutoff energy is selected, such as 1 keV for both electron and photon cutoff energies, PRESTA-I boundary crossing rules and PRESTA-II electron step rules, etc.). The appropriate number of transported particles is set (for example, for a single organ source, the number of particles can be set to 2×10 8 ; For distributed source, the number of transported particles can reach 2×10 9 The above is to ensure that the data uncertainty of the voxels in the tumor area is generally less than 2%). The voxel dose distribution corresponding to the nuclide activity distribution is obtained.

[0042] Alternatively, the voxel S value method VSV can be used to calculate the overall dose distribution by looking up the table. For the non-uniform medium distribution between the source voxel V1 and the target voxel V2 (voxel V1 → voxel V2), the equivalent path method should be used to convert the material density information obtained in step 1, that is,

[0043]

[0044] Among them, l i is the path length of any voxel passing through (voxel V1→voxel V2), ρ i is the density of the voxel being traversed, ρ w is the density of water, n is the number of all voxels that pass through (voxel V1 → voxel V2), L homo is the equivalent path calculated, i.e. the path length based on which the VSV table is searched;

[0045] Step 5: Draw the dose distribution obtained in step 4 and the weighted activity-density distribution obtained in step 3 into an image with the same resolution as the original dose distribution using a drawing software (such as Matlab with powerful graphics functions), and connect them into a left-right arranged image pair, such as Figure 2 As shown, Figure 2 The image pairs of dose map and activity-density overlay map required for GAN deep learning model training are given, where the dose map is on the left and the activity-density overlay map is on the right; the input image pair used to test the trained GAN deep learning model can be a blank image on the left (the original dose map can also be retained, which has no effect on the result), and the image on the right is an activity-density overlay map of the dose to be determined;

[0046] Step 6: Set the voxel dose distribution obtained in step 4 to 0, and use a drawing software (such as Matlab with powerful graphics functions) to draw the weighted activity-density distribution obtained in step 3 into an image with the same resolution as the original dose distribution, and connect them into equal left-right arranged image pairs (in the form of Figure 2 shown, but the left image is blank);

[0047] Step 7: Send the dose and activity-density image pairs obtained in step 5 as training atlases to the core module of this method, namely, the generative adversarial network software (such as conditional generative adversarial network (conditional GANs) pix2pix (https: / / github.com / affinelayer / pix2pix-tensorflow), to train the deep learning model based on the optimized model parameters (learning rate, iteration rounds, etc.), and generate a deep learning model capable of generating corresponding dose distribution from activity-density distribution;

[0048] Step 8, sending the blank and activity-density test image pairs obtained in step 6 to the deep learning model generated in step 7 for generating corresponding dose distribution from activity-density distribution for testing, and generating a dose distribution image corresponding to the activity-density distribution;

[0049] Step 9. Using the regional numbers in the patient material and density model established in step 1, sort out the doses corresponding to all voxels of lesions and organs at risk from the predicted dose image and GT image. These data can be used to calculate indicators including the mean dose (MD: Mean Dose) of the lesion area and organs at risk, gamma pass rate (GPR: Gamma-pass-rate), dose-volume histogram (DVH: Dose-Volume Histogram) and off-axis dose distribution (OAR: Off-Axis Ratio), and compare them with the corresponding GT results.

[0050] To verify the feasibility of the method, we performed Monte Carlo dose calculation and activity distribution conversion dose distribution method tests on the Zubal phantom (http: / / noodle.med.yale.edu / zubal / ), a real human body model with well-segmented organs and commonly used in nuclear medicine internal irradiation dose calculation.

[0051] The Zubal phantom is based on the conversion of CT anatomical images. It is a 128×128×243 data matrix with a voxel size of 4mm×4mm×4mm and a total of 68 segmented organs or tissues.

[0052] First, the Monte Carlo procedure EGSnrc / DOSXYZnrc was used to calculate a clinical 131 In the case of abdominal tumors treated with I, the typical activity distribution information displayed by functional imaging is: if the activity absorbed by the tumor is set to 1, then kidneys: 0.80; liver: 0.28; spleen: 0.52; lungs: 0.28; heart: 0.48; blood pool: 0.48; rest of the body: 0.04. Assume that the tumor is spherical, located in the abdomen of a digital human body (Zubal phantom), with a volume of 135 mL (i.e., the radius of the sphere is 3.18 cm).

[0053] In order to simulate the non-uniform activity distribution formed by the metabolism, absorption and transport of radionuclides after entering human cells, different variance σ changes of Gaussian function are used to describe it. In the first case, the activity in the tumor and each organ is uniformly distributed, that is, sigma=0.0 in the Gaussian distribution (hereinafter referred to as Gauss0.0); in the second case, the activity in the tumor and each organ is distributed according to Gaussian sigma=0.3 (hereinafter referred to as Gauss0.3); in the third case, the activity in the tumor and each organ is distributed according to Gaussian sigma=0.6 (hereinafter referred to as Gauss0.6). The latter two cases are used to verify the dose conversion capability of a nuclear medicine activity-dose automatic conversion method based on a generative adversarial network in this embodiment for completely non-uniform activity distribution.

[0054] As a feasibility study, we used the dose and activity-density distribution image pairs obtained from Monte Carlo calculations of different organ sources and different activity distributions in the Zubal phantom as the dataset for deep learning model training and testing. The image pairs used for model training came from the following four types of sources:

[0055] (1) 1-region source (S1): There is one independent activity region (i.e., tumor);

[0056] (2) 2-region source (S2): There are two independent activity regions (i.e., tumor and kidney);

[0057] (3) 4-region source (S4): There are 4 independent activity regions (i.e., tumor, kidney, liver, and lung);

[0058] (4) 7-region source (S7): There are 7 independent activity regions (i.e. tumor, kidney, liver, heart, lung, blood and spleen);

[0059] The following blank dose and activity-density image pairs of one type of source are used for model testing:

[0060] (5) 3-region source (S3): There are 3 independent activity regions (i.e., tumor, kidney, and liver);

[0061] The activity ratio of various organs is as mentioned above. 131 Typical activity distribution information displayed by functional imaging in the case of abdominal tumor treatment. The activity in various regional sources is divided into three situations: Gauss0.0, Gauss0.3 and Gauss0.6.

[0062] The learning rate of cGANs pix2pix is ​​set to 0.0002 and the number of iterations is 200.

[0063] Through experimental training, it was found that the GAN deep learning model trained with the simplest S1 regional source model image (i.e., spherical tumor source) had the best prediction effect and the highest training efficiency. A total of 21*3 images of the S1 single source were used for training, i.e., 21 images for Gauss0.0, 21 images for Gauss0.3, and 21 images for Gauss0.6.

[0064] Figure 3 The comparison between the absorption dose distribution map of the three-source organ model predicted by the GAN deep learning model trained with a simple S1 single-source model image and the standard GT absorption dose distribution map is given. Left column: input activity-density overlay map; middle: predicted dose distribution map; right column: standard GT dose distribution map. Figure 3 (a) is the case of uniform distribution Gauss 0.0 (i.e. σ = 0); Figure 3 (b) and (c) are the cases of non-uniform distribution Gauss0.3 (i.e., σ=0.3) and Gauss0.6 (i.e., σ=0.6).

[0065] Table 1 shows the difference between the predicted image results and the GT image results of the two methods. The evaluation based on the whole phantom, that is, the predicted image and the GT image are compared, and the percentage is normalized to the highest dose point value of the entire Zubal phantom, that is, the highest dose point value of both is 100. Then the average value of the dose of the key organs, the gamma pass rate evaluation (3% / 3mm), the dose-volume histogram and the off-axis dose distribution of the two are calculated and compared;

[0066] Before comparing the predicted image with the GT image, the evaluation based on a single organ is normalized to the highest dose point value of a specific organ (such as the target organ and the source organ) in the Zubal phantom, that is, the highest dose point values ​​of both are 100. Then, the average value of the dose of the key organs, the gamma pass rate evaluation, the dose-volume histogram and the off-axis dose distribution of the two are calculated and compared.

[0067] It can be seen that the results obtained by the evaluation method based on a single organ are better than those obtained by the evaluation method based on the entire phantom. Although the average dose difference between the two is still large, both have a high gamma evaluation pass rate (>93%), which shows that the predicted image and the GT image are very similar (from Figure 3 This can be seen). However, more model calculation data are needed for dose quantitative calibration to obtain accurate empirical parameters for data calibration.

[0068] In addition, the standard deviation of the organ-based evaluation results was also calculated. The standard deviation calculation formula is:

[0069]

[0070] Among them, N is the total number of voxels occupied by a certain organ; and are the percentage dose values ​​of the i-th voxel (normalized to the peak dose of the key source organ) corresponding to the predicted and standard GT (i.e., calculated by Monte Carlo procedure) respectively; in order to correct the influence of the deviation between the predicted mean dose and the standard GT mean dose (see Table 1), the percentage dose values ​​of all voxels in the organ are Multiply After the factor, For comparison, it can be seen that from the perspective of the large number of voxels in the organ, the difference between the predicted distribution and the standard GT distribution is about 10%σ in voxel dose.

[0071] Table 1 The GAN deep learning model trained on a simple single-source model image (i.e., spherical tumor source) predicts the percentage difference between the absorbed dose of the three-source organ model and the standard GT absorbed dose (the first data in the square brackets), as well as the Gamma evaluation percentage pass rate of the dose distribution similarity between the two (the second data in the square brackets). The standard deviation Std. is given in the evaluation based on a single organ.

[0072]

[0073]

[0074] The above contents are merely examples and explanations of the structure of the present invention. The technicians in this technical field may make various modifications or additions to the specific embodiments described or replace them in a similar manner. As long as they do not deviate from the structure of the present invention or exceed the scope defined by the claims, they should all fall within the protection scope of the present invention.

Claims

1. A method for automatic conversion of nuclear medicine activity-dose based on generative adversarial networks, characterized in that: The method comprises the following steps: Step 1: delineate the lesion and organ at risk region in the patient's anatomical image, establish a material and density model of the patient based on the patient's anatomical image, and assign region numbers to all voxels of the delineated lesions and organs at risk in the material and density model in a one-to-one correspondence manner; Step 2, according to the image of the activity space distribution of the nuclide at a certain time t provided by the functional imaging of the patient, obtaining the spatial distribution information of the activity of the nuclide at time t in the patient's body; Step 3: Weighted superposition of the patient's radionuclide activity spatial distribution image and the corresponding patient's anatomical image density map. HR voxel =A voxel *weight+density voxel Among them, A voxel is the activity of a certain voxel obtained in step 2, weigt is the uniform weighting of all voxel activities, density voxel is the material density of the voxel obtained in step 1, AD voxel is the obtained activity-density weighted value; Step 4: Use nuclear medicine dose calculation method to obtain the patient's dose spatial distribution; Step 5: The dose spatial distribution obtained in step 4 and the weighted activity-density distribution obtained in step 3 are plotted into an image with the same resolution as the original dose distribution using a plotting software, and connected into a left-right arranged image pair; Step 6, setting all the dose distributions obtained in step 4 to 0, and using a drawing software to draw the weighted activity-density obtained in step 3 into an image with the same resolution as the original dose distribution, and connecting them into equal left-right arranged image pairs; Step 7: Send the dose and activity-density image pairs obtained in step 5 as training atlases to the generative adversarial network software, train the deep learning model based on the optimized model parameters, and generate a deep learning model that generates corresponding dose distribution from activity-density distribution; Step 8: The 0 dose map and the activity-density overlay map obtained in step 6 are used as a test map set, and sent to the deep learning model generated in step 7 for generating the corresponding dose distribution from the activity-density distribution for testing, so as to generate a dose distribution corresponding to the activity-density distribution; Step 9: Using the patient material and the region number in the density model established in step 1, sort out the voxel doses of lesions and organs at risk for calculation of indicators including average dose, gamma pass rate assessment, dose-volume histogram, and off-axis dose distribution.

2. The method for automatic conversion of nuclear medicine activity-dose based on generative adversarial networks according to claim 1, characterized in that: In step 4, the nuclear medicine dose calculation adopts the Monte Carlo method.

3. The method for automatic conversion of nuclear medicine activity-dose based on generative adversarial networks according to claim 1 or 2, characterized in that: In step 4, the dose distribution of the patient obtained by the nuclear medicine method is determined as follows: the particle physics model of the classical Monte Carlo program EGSnrc / DOSXYZnrc is used to perform particle transport calculations to obtain the dose distribution corresponding to the nuclide activity distribution.

4. The method for automatic conversion of nuclear medicine activity-dose based on generative adversarial networks according to claim 1, characterized in that: In step 4, the nuclear medicine dose is calculated using the voxel S value method.

5. The method for automatic conversion of nuclear medicine activity-dose based on generative adversarial network according to claim 1 or 4, characterized in that: In step 4, the dose distribution of the patient obtained by nuclear medicine method is determined as follows: the dose distribution is obtained by looking up the table using the equivalent distance between the source voxel V1 and the target voxel V2.

6. A method for automatic conversion of nuclear medicine activity-dose based on generative adversarial networks according to claim 4 or 5, characterized in that: The voxel S value method uses the equivalent distance between the source voxel V1 and the target voxel V2 to determine the non-uniform medium distribution between the source voxel V1 and the target voxel V2, and the equivalent path method conversion is performed according to the material density information obtained in step 1, that is, Among them, l i is the path length of any voxel passing through (voxel V1→voxel V2), ρ i is the density of the voxel being traversed, ρ w is the density of water, n is the number of all voxels that pass through (voxel V1 → voxel V2), L homo is the calculated equivalent path.

Images