Device and method for enabling quantitative renal spect without CT by using neural network

Abstract

According to an embodiment, disclosed are a device and method for generating a quantitative renal SPECT without CT by using a neural network. The device for generating a quantitative renal SPECT without CT by using a neural network, according to an embodiment, comprises: one or more processors; and a memory for storing instructions executed by the one or more processors, wherein the one or more processors are configured to generate an attenuation map by inputting a first single photon emission computed tomography (SPECT) image and a second SPECT image capturing a kidney into a model.

Description

Device and method for enabling quantitative renal SPECT without CT using a neural network

[0001] It relates to a technology that enables quantitative renal SPECT without CT using a neural network.

[0002] [Cross-reference to related applications]

[0003] This application claims priority to Republic of Korea Patent Application No. 10-2024-0082093, filed June 24, 2024, the entire contents of which are incorporated herein by reference.

[0004] [Description of Nationally Supported Research and Development]

[0005] This study was supported by the Ministry of Science and ICT [Project ID: 1711197862, Project No.: RS-2023-00241409, Ministry: MSIT, Project Management (Specialized) Institution: National Research Foundation of Korea, Research Project Name: Individual Basic Research (MSIT), Research Project Name: Development of CT-free Quantitative Single Photon Emission Computed Tomography Technology Using Deep Learning, Project Implementing Institution: Bundang Seoul National University Hospital, Research Period: 2023.06.01 ~ 2026.02.28].

[0006] Renal function is assessed by measuring the percentage uptake (%) of a radiopharmaceutical (e.g., Tc-99m DTPA) administered intravenously to the patient. Typically, this uptake is measured using planar imaging from a gamma camera.

[0007] Recently, SPECT / CT-based nuclear medicine imaging methods have emerged as quantitative methods in nuclear medicine. SPECT / CT utilizes attenuation correction to improve accuracy. This attenuation correction is performed using CT, which is preferred due to its superior image quality and short imaging time.

[0008] Moreover, SPECT / CT reports improved precision and improved reproducibility when compared to conventional planar imaging. As shown in Figure 1, SPECT / CT exhibits lower %CV and higher ICC compared to planar imaging.

[0009] However, compared to planar imaging, SPECT / CT inevitably exposes patients to additional radiation. This means that in addition to the approximately 1.813 mSv dose from administering radiopharmaceuticals, CT imaging adds an additional 1.5 to 6.75 mSv. While SPECT / CT improves accuracy and reproducibility, it exposes patients to approximately three to seven times more radiation.

[0010] In other words, there is a need for a method to measure renal function that improves precision and reproducibility while minimizing radiation exposure to patients.

[0011] The disclosed embodiments are intended to enable quantitative renal SPECT without CT using neural networks.

[0012] A device for generating a quantitative renal SPECT without CT using a neural network according to one embodiment comprises one or more processors; and a memory for storing commands executed by the one or more processors, wherein the one or more processors: input a first SPECT (Single Photon Emission Computed Tomography) image and a second SPECT image of a kidney into a model to generate an attenuation map.

[0013] The first SPECT image may be a primary emission SPECT image, and the second SPECT image may be a scattering SPECT image.

[0014] The above model may include an encoder that extracts features by downsampling the first SPECT image and the second SPECT image, and a decoder that upsamples the features and outputs the attenuation map.

[0015] At least one of the encoder and decoder can adjust the data distribution by applying log-maximum normalization to each voxel in the first SPECT and the second SPECT.

[0016] The above decoder can use a loss function defined based on the following [Mathematical Formula 1].

[0017] [Mathematical Formula 1]

[0018]

[0019] (Here, E is the loss function, G(X) is the predicted attenuation map, X is the input first SPECT image and second SPECT image, Y is the correct value of the attenuation map, N is the number of voxels in the attenuation map, M is the product of the number of voxels in the attenuation map and the number of axes of the gradient, is the image gradient operator, a, b can be any real number, n can be any natural number.)

[0020] The above decoder can perform upsampling via maximum adjacent interpolation.

[0021] The above model can be supervised learning based on an attenuation map affected by iodine-contrast agent and an attenuation map not affected by iodine-contrast agent.

[0022] The one or more processors may: perform attenuation correction on the first SPECT image using the attenuation map, and calculate the radiopharmaceutical uptake in the kidney based on the corrected SPECT image.

[0023] The one or more processors can: perform attenuation correction on the first SPECT image using the attenuation map, and calculate a glomerular filtration rate (GFR) in the kidney based on the corrected SPECT image.

[0024] A method for generating a quantitative renal SPECT without CT using a neural network according to one embodiment is a method performed by an apparatus for generating a quantitative renal SPECT without CT using a neural network, the apparatus comprising: one or more processors; and a memory storing instructions executed by the one or more processors, the method comprising: inputting a first SPECT (Single Photon Emission Computed Tomography) image and a second SPECT image of a kidney into a model to generate an attenuation map.

[0025] The first SPECT image may be a primary emission SPECT image, and the second SPECT image may be a scattering SPECT image.

[0026] The above model may include an encoder that extracts features by downsampling the first SPECT image and the second SPECT image, and a decoder that upsamples the features and outputs the attenuation map.

[0027] At least one of the encoder and decoder can adjust the data distribution by applying log-maximum normalization to each voxel in the first SPECT and the second SPECT.

[0028] The above decoder can use a loss function defined based on the following [Mathematical Formula 1].

[0029] [Mathematical Formula 1]

[0030]

[0031] (Here, E is the loss function, G(X) is the predicted attenuation map, X is the input first SPECT image and second SPECT image, Y is the correct value of the attenuation map, N is the number of voxels in the attenuation map, M is the product of the number of voxels in the attenuation map and the number of axes of the gradient, is the image gradient operator, a, b can be any real number, n can be any natural number.)

[0032] The above decoder can perform upsampling via maximum adjacent interpolation.

[0033] The above model can be supervised learning based on an attenuation map affected by iodine-contrast agent and an attenuation map not affected by iodine-contrast agent.

[0034] The method may include a step of performing attenuation correction on the first SPECT image using the attenuation map, and calculating the radiopharmaceutical uptake in the kidney based on the corrected SPECT image.

[0035] The above method can perform attenuation correction on the first SPECT image using the attenuation map, and calculate the glomerular filtration rate (GFR) in the kidney based on the corrected SPECT image.

[0036] The disclosed embodiments generate an attenuation map using only SPECT images without CT using a neural network-based model, thereby reducing radiation exposure required for functional examination in accordance with the ALARA (As Low As Reasonably Achievable) principle.

[0037] The disclosed embodiments consider the estimated proportion of learning data affected by iodine-contrast agents to overcome biased errors of neural network-based models toward iodine-contrast agents, thereby generating iodine-neutral attenuation maps.

[0038] The disclosed embodiments utilize a model having optimal components that address the problem of checkerboard-like artifacts and weak intensity signals not being reflected, thereby improving the quality of the attenuation map generated by correcting the distorted signal.

[0039] Figure 1 is a graph showing the performance of two existing methods for detecting radiopharmaceuticals using nuclear medicine imaging: planar imaging and SPECT / CT.

[0040] FIG. 2 is a block diagram illustrating a device for generating quantitative renal SPECT without CT using a neural network according to one embodiment.

[0041] Figure 3 is a schematic diagram illustrating the types of models and the utilization of the device as an example.

[0042] Figure 4 is an example diagram illustrating the architecture of an example model.

[0043] Figure 5a is an example diagram showing the distribution of voxels in each image by signal intensity using the first and second SPECT images and maximum normalization.

[0044] Figure 5b is an example diagram showing the distribution of voxels in each image by signal intensity using the first and second SPECT images and log-maximum normalization.

[0045] Figures 6a to 6f are exemplary diagrams showing the results of the attenuation map generated by each model of the disclosed embodiment.

[0046] Figures 7a to 7e are error maps showing the difference in attenuation coefficients between the attenuation maps output by the models of Figures 6a to 6e and the attenuation maps of the correct values, respectively.

[0047] Figure 8a is an error rate map showing the difference between the correct value of the attenuation map affected by the iodine-contrast agent and the attenuation coefficient of the attenuation map output by the model.

[0048] Figure 8b is an error rate map showing the difference between the correct value of the attenuation map that is not affected by the iodine-contrast agent and the attenuation coefficient of the attenuation map output by the model.

[0049] FIG. 9 is a flowchart illustrating a method for generating quantitative renal SPECT without CT using a neural network according to one embodiment.

[0050] The terms used in this specification have been selected from widely used and commonly accepted terms, taking functionality into consideration. However, these terms may vary depending on the intentions or practices of those skilled in the art, the emergence of new technologies, etc. Furthermore, in certain cases, the applicant may arbitrarily select terms, in which case their meanings will be described in the relevant explanatory section of the specification. Therefore, it should be noted that the terms used in this specification should be interpreted based on their substantive meaning and the overall content of the specification, rather than simply their names.

[0051] While the terms "first" and "second" used in this application may be used to describe various components, these components should not be limited by these terms. These terms are intended solely to distinguish one component from another. For example, without departing from the scope of the invention, a first component may be referred to as a "second component," and similarly, a second component may also be referred to as a "first component."

[0052] The singular expressions used in this application include plural expressions unless the context clearly dictates otherwise. In this application, terms such as "include," "have," and "have" are intended to indicate the presence of components or combinations thereof described in the specification, but do not preclude the possibility of other components or features being present or added.

[0053] Additionally, the embodiments described herein may have aspects that are entirely hardware, partially hardware and partially software, or entirely software. As used herein, "unit," "module," "device," "server," or "system" refer to hardware, a combination of hardware and software, or a computer-related entity such as software. For example, a unit, module, device, server, or system may refer to hardware that constitutes part or all of a platform and / or software such as an application for operating the hardware.

[0054] The embodiments are described in detail with reference to the attached drawings and the contents described in the attached drawings, but the scope of the claims is not limited or restricted by the embodiments.

[0055] FIG. 2 is a block diagram illustrating a device (100) for generating quantitative renal SPECT without CT using a neural network according to one embodiment.

[0056] Referring to FIG. 2, a device (100) for generating quantitative renal SPECT without CT using a neural network includes a processor (110) and a memory (120).

[0057] The processor (110) inputs the first SPECT (Single Photon Emission Computed Tomography) image and the second SPECT image (2) of the kidney into the model (10) to generate an attenuation map.

[0058] Here, single-photon emission computed tomography (SPET) is a nuclear medicine imaging technique. It involves injecting a radiopharmaceutical that emits single photons (gamma rays) into a tumor, measuring the gamma rays transmitted through the tumor, and reconstructing the body's distribution of the SPET into an image. In particular, SPET can provide biofunctional images, which are biochemical phenomena within organs.

[0059] The first SPECT image (1) may be a primary emission SPECT image generated using single photon emission computed tomography. The primary emission SPECT image may be a SPECT image for the primary gamma ray in single photon emission computed tomography. Preferably, the energy level for the primary gamma ray is about 140 keV, and may be 126-154 KeV within a 10% error range.

[0060] The second SPECT image (2) may be a scattering SPECT image generated using single photon emission computed tomography. The scattering SPECT image may be a SPECT image of gamma rays scattered during single photon emission computed tomography. Preferably, the level of the scattering energy is about 120 keV, and may be 115-125 keV within a 5% error range.

[0061] At this time, it may be desirable to acquire the first SPECT image (1) and the second SPECT 2-3 minutes after intravenous injection of a radiopharmaceutical (e.g., Tc99m-DTPA) to a patient suspected of having kidney disease.

[0062] At least one of the first SPECT image (1) and the second SPECT image (2) may be an image corrected for resolution recovery using a Butterworth low-pass filter (e.g., order 10, cutoff frequency: 0.48 cycles / cm) for statistical noise reduction.

[0063] At least one of the first SPECT image (1) and the second SPECT image (2) has an original size of 128x128x128 and a voxel size of 3.45×3.45×3.45 mm. 3 It may be. Preferably, at least one of the first SPECT image (1) and the second SPECT image (2) may be an image cropped to 64×128×128 to reduce the span in the z-axis to ensure consistency of the kidney volume.

[0064] An attenuation map is an image that represents the degree of reduction in energy, such as radiation or ultrasound, when it passes through a medium (especially an organ), and can express the degree of absorption and scattering of radiation through the attenuation coefficient.

[0065] The processor (110) can perform attenuation correction on the first SPECT image (1) using the generated attenuation map, and calculate the amount of radiopharmaceutical uptake in the kidney based on the corrected SPECT image.

[0066] The processor (110) can perform attenuation correction on the first SPECT image (1) using the generated attenuation map, and calculate the glomerular filtration rate (GFR) in the kidney based on the corrected SPECT image.

[0067] The memory (120) stores one or more instructions to be executed by the processor (110).

[0068] The memory (120) can store various data used by the processor (110). For example, the memory (120) can include software (e.g., input data or output data for a program executed by the processor (110) and / or instructions related to the program).

[0069] The memory (120) may include volatile memory or non-volatile memory.

[0070] The memory (120) can store a model (10) for which learning has been completed.

[0071] Figure 3 is a schematic diagram explaining the type of model (10) and the use of the device as an example.

[0072] The model (10) receives the first SPECT image (1) and the second SPECT image (2) as input and outputs an attenuation map. When the first SPECT image (1) and the second SPECT image (2) are combined and used as input data, the model (10) has a performance of R according to the performance of Table 1 below. 2 As this increases, MSE and NMAE decrease. That is, for improved performance, it is desirable for the model (10) to use the first SPECT image (1) and the second SPECT image (2) as input data.

[0073] Input regularization loss function upsampling R 2 MSE (*10 -4 )%NMAEPMaxL1TC0.9802±0.0103001.1081±0.57671.7975±0.4452P+SMaxL1TC0.9814±0.0093811.0417±0.53671.7762±0.4253

[0074] Here, P may represent the first SPECT image (1), S may represent the second SPECT image (2), and P+S may represent the combination of the first SPECT image (1) and the second SPECT image (2). The model (10) may output an attenuation map. That is, the model (10) may output an attenuation map provided by a CT without a CT. At this time, the model (10) may use the first loss function.

[0075] The first loss function may be a function that calculates the sum of the absolute value differences between the correct value of the attenuation map and the generated attenuation map.

[0076] Specifically, the first loss function can be defined as in [Mathematical Formula 1] below.

[0077] [Mathematical Formula 1]

[0078]

[0079] Here, X may represent the first and / or second SPECT images (2) being input, G(X) may represent the attenuation map predicted by the model (10), Y may represent the correct value of the attenuation map, and N may represent the number of voxels in the attenuation map.

[0080] As another example, the model (10) may use a second loss function based on GDL (Gradient Difference Loss). Specifically, the model (10) may use a second loss function defined based on the following [Mathematical Formula 2].

[0081] [Equation 2]

[0082]

[0083] Here, n can be a natural number greater than or equal to 1, and M can mean the product of the axis of the gradient and the number of voxels in the attenuation map.

[0084] As another example, model (10) may use a third loss function composed of a first loss function and a second loss function. Specifically, model (10) may use a second loss function defined based on [Mathematical Formula 3], for example.

[0085] [Equation 3]

[0086]

[0087] Here, n can mean a natural number greater than or equal to 1.

[0088] At this time, when the model (10) uses the first SPECT image (1) and the second SPECT image (2) as input, considering the performance of Table 2 below, L1+3 L GDL 1 It may be desirable to use as a loss function.

[0089] Input regularization loss function upsampling R 2 MSE (*10 -4)%NMAEPSLog-maxL1+1*L GDL 1 TC0.9817±0.0100761.0252±0.5954551.7019±0.459276PSLog-maxL1+1*L GDL 2 TC0.9812±0.0096781.0523±0.5578601.7409±0.425118PSLog-maxL1+3*L GDL 1 TC0.9822±0.0095780.9998±0.5672571.6790±0.431515PSLog-maxL1+3*L GDL 2 TC0.9810±0.0101221.0676±0.5872521.7369±0.446755PSLog-maxL1+5*L GDL 1 TC0.9819±0.0098451.0153±0.5693241.7131±0.428341PSLog-maxL1+5*L GDL 2 TC0.9817±0.0093321.0271±0.5489561.6995±0.430501

[0090] Model (10) can use pre-convolution or nearest neighbor interpolation as an upsampling method. In this case, considering the performance of [Table 3] below, it may be preferable for model (10) to use nearest neighbor interpolation as an upsampling method.

[0091] Input regularization loss function upsampling R 2 MSE (*10 -4 )%NMAEPSLog-maxL1+3*L GDL 1 TC0.9821±0.0109990.9901±0.6634091.6586±0.492409PSLog-maxL1+3*L GDL 1 Interpolation0.9820±0.0114430.9938±0.6655761.6600±0.495213

[0092] The model (10) can be trained to learn an attenuation map by inputting a SPECT image using learning data having a 1:1 ratio of attenuation maps affected by the iodine-contrast agent and attenuation maps not affected by the iodine-contrast agent in order to generate an attenuation map that is neutral to the iodine-contrast agent. In this case, the contrast agent is described as an iodine-contrast agent used in the kidney, but any contrast agent that can be used in the kidney belongs to this category and is not necessarily limited to the described example. The processor (110) can perform attenuation correction on the first SPECT image (1) and the second SPECT image (2) using the attenuation map output by the model (10). The processor (110) can also analyze the glomerular filtration rate of the kidney through the first and second SPECT images (2) on which attenuation correction has been completed.

[0093] The model (10) can be trained to learn an attenuation map by inputting a SPECT image using training data having a 1:1 ratio of attenuation maps affected by the iodine-contrast agent and attenuation maps not affected by the iodine-contrast agent, in order to generate an attenuation map that is neutral to the iodine-contrast agent. In this case, the contrast agent is described as an iodine-contrast agent used in the kidney, but any contrast agent that can be used in the kidney falls under this category and is not necessarily limited to the described example.

[0094] The processor (110) can perform attenuation correction on the first SPECT image (1) and the second SPECT image (2) using the attenuation map output by the model (10). The processor (110) can also analyze the glomerular filtration rate of the kidney through the first and second SPECT images (2) on which attenuation correction has been completed.

[0095] Figure 4 is an example diagram explaining the architecture of an example model (10).

[0096] Referring to Fig. 4, an example model (10) is a U-net structure and includes an encoder and a decoder.

[0097] The encoder can perform downsampling using a convolution block to extract features of the first SPECT image (1) and the second SPECT image (2). At this time, the encoder can gradually downsample the first SPECT image (1) and the second SPECT image (2) to extract high-level abstract features.

[0098] In the shrinkage path, the convolution block can be designed by successively repeating a 3X3X3 convolution layer, an instance normalization layer (batch normalization), and a ReLU function, for example, twice. Afterwards, a max pooling layer (e.g., 2X2X2) can be connected to the convolution block and used for downsampling.

[0099] In particular, the encoder is preferably designed to adopt an instance normalization layer to reflect low-intensity signals from the first SPECT signal and the second SPECT signal in feature extraction. The following description will be given with reference to FIGS. 5a and 5b.

[0100] The decoder can perform upsampling using a convolution block to reconstruct the features extracted from the encoder to the same size as the input. This allows the decoder to gradually restore the spatial resolution of the extracted features and reconstruct details.

[0101] In the expansion path, the convolution block can be connected to a skip connection. The skip connection is located between the encoder and decoder, connecting the contraction path and the expansion path. That is, features extracted from the encoder can be passed to the decoder. The convolution block can perform a convolution operation on the features received from the skip connection.

[0102] The convolution block in the expansion path can be designed, similar to the convolution block in the contraction path, by repeating, for example, a 3X3X3 convolution layer, an instance normalization layer (batch normalization), and a ReLU function four times in succession. Afterwards, an upsampling layer (e.g., 2X2X2) can be connected to the convolution block and used for upsampling.

[0103] Meanwhile, model (10) is described as a U-net structure frequently used in the medical field, but this is an example and can be designed based on various known structures such as Seg-NET and MFU-net, and is not necessarily limited to the U-net structure.

[0104] Figure 5a is an example diagram showing the distribution of voxels in each image by signal intensity using the first and second SPECT images (2) and maximum normalization.

[0105] Referring to Fig. 5a, a graph is shown that normalizes the distribution of signal intensities per voxel corresponding to the first and second SPECT images (2).

[0106] In Fig. 5a, the first and second SPECT images (2) are depicted as having low-intensity signals in areas including the renal parenchyma. In particular, the second SPECT image (2) has a large distribution of low-intensity voxels, making it difficult to identify the captured kidney.

[0107] In Fig. 5a, the signal intensity distribution table of voxels in the first and second SPECT images (2) shows that the signal intensity of the voxels is distributed asymmetrically. That is, when applying maximum normalization to the first and second SPECT images (2), distortion may occur that underrepresents areas of low signal when generating an attenuation map.

[0108] Figure 5b is an example diagram showing the distribution of voxels in each image by signal intensity using the first and second SPECT images (2) and log-maximum normalization.

[0109] Here, log-maximum normalization can be a normalization method that takes the log of each data and then scales the log value to the range of the maximum and minimum values.

[0110] In Fig. 5b, it can be qualitatively confirmed that the first and second SPECT images (2) have improved contrast and clarity compared to Fig. 5a. In particular, the second SPECT image (2) has been improved so that the captured kidney is visually identifiable and details are better revealed.

[0111] In Fig. 5b, the distribution of voxels within the first and second SPECT images (2) confirms that the signal intensity has increased and the variance has decreased. That is, when log-maximum normalization is applied to the first and second SPECT images (2), an attenuation map can be generated by increasing the proportion of low-intensity signals.

[0112] That is, it may be desirable for the model to use log-maximum normalization.

[0113] Figures 6a to 6f are exemplary diagrams showing the results of the attenuation map generated by each model (10) of the disclosed embodiment.

[0114] Fig. 6a is an attenuation map predicted by the model (10) when only the first SPECT image (1) was used as input to the model (10). In Fig. 6a, the model (10) exhibits high focal activity in both kidneys, which mostly causes checkerboard-shaped artifacts in both kidneys within the attenuation map.

[0115] Fig. 6b is an attenuation map predicted by the model (10) when the first SPECT image (1) and the second SPECT image (2) are used as inputs to the model (10). In Fig. 6b, the model (10) performs a focus activity with a bias toward one kidney, which mostly causes a checkerboard-shaped artifact in one kidney within the attenuation map.

[0116] Figure 6c is an attenuation map predicted by model (10) when a log-maximum normalization layer is added to the convolution block of model (10). It can be confirmed that the attenuation map output by model (10) in Figure 6c has reduced artifacts.

[0117] Figure 6d is the loss function used by model (10), L1+3 L GDL 1 When using, this is the attenuation map predicted by the model (10). In Fig. 6d, the attenuation map output by the model (10) confirms that the artifact location moves from the renal parenchyma to the center of the renal pelvis.

[0118] In Fig. 6e, the model (10) predicts an attenuation map when the pre-convolution is replaced with the nearest neighbor interpolation under the conditions of Figs. 6c and 6d. It can be confirmed that the attenuation map output by the model (10) in Fig. 6e has completely removed artifacts.

[0119] Fig. 6f is an attenuation map of the correct values ​​corresponding to the first SPECT image (1) and the second SPECT image (2). It can be confirmed that Fig. 6f qualitatively shows the greatest similarity to Fig. 6e.

[0120] Figures 7a to 7e are error maps showing the difference in attenuation coefficients between the attenuation map output by the model (10) of Figures 6a to 6e and the attenuation map of the correct value, respectively. As shown, the errors can be distinguished by color.

[0121] As shown in Fig. 7e, a model (10) trained to receive the first SPECT image (1) and the second SPECT image (2) as input and output an attenuation map, with log-maximum normalization and loss function: L1+3 L GDL 1 And it was confirmed that the model (10) using the nearest neighbor interpolation had the smallest quantitative error when compared to the correct value.

[0122] Figure 8a is an error rate map showing the difference between the correct value of the attenuation map affected by the iodine contrast agent and the attenuation coefficient of the attenuation map output by the model. As shown, the error rates can be distinguished by color.

[0123] Referring to Fig. 8a, from the left, in order, the correct value of the attenuation map affected by the iodine contrast agent, the attenuation map predicted by the model (10), and the result quantitatively expressing the difference in radioactivity between the correct value and the predicted attenuation map.

[0124] In Figure 8a, the correct value attenuation map affected by the iodine contrast agent shows iodine contrast agent with a high concentration in the renal pelvis. In contrast, the attenuation map predicted by the model (10) does not reflect iodine contrast agent in the renal pelvis to the extent of the correct value. Therefore, the difference between the correct and predicted attenuation maps is a positive value in the renal pelvis region.

[0125] Figure 8b is an error rate map showing the difference between the correct value of the attenuation map that is not affected by the iodine-contrast agent and the attenuation coefficient of the attenuation map output by the model.

[0126] Referring to Fig. 8b, from the left, in order, the correct attenuation map value not affected by the iodine contrast agent, the attenuation map predicted by the model (10), and the result quantitatively expressing the difference in radioactivity between the correct and predicted attenuation maps.

[0127] In Fig. 8b, the attenuation map predicted by the model (10) does not show the influence of the iodo-contrast agent identified in the renal pelvis. However, the difference between the correct and predicted attenuation maps shows a negative value in the renal pelvis region.

[0128] That is, it can be confirmed that model (10) generates an attenuation map that is neutral to iodine-contrast agents, assuming both cases where it is affected by iodine-contrast agents and cases where it is not affected by iodine-contrast agents.

[0129] FIG. 9 is a flowchart illustrating a method for generating quantitative renal SPECT without CT using a neural network according to one embodiment.

[0130] Referring to FIG. 9, a method for generating quantitative renal SPECT without CT using a neural network according to one embodiment can be performed by the device of FIG. 2.

[0131] First, a device that generates a quantitative kidney SPECT without CT using a neural network inputs a first SPECT (Single Photon Emission Computed Tomography) image and a second SPECT image (2) of the kidney into a model (10) to generate an attenuation map (910).

[0132] The method illustrated in FIG. 9 has been described with reference to the flowchart presented in the drawings. While the method has been depicted and described as a series of blocks for purposes of illustration, the present invention is not limited to the order of the blocks, and some blocks may occur in a different order or concurrently with other blocks than depicted and described herein, and various other branches, flow paths, and block orders that achieve the same or similar results may be implemented. Furthermore, not all depicted blocks may be required to implement the method described herein.

[0133] Furthermore, the method according to one embodiment of the present invention may be implemented in the form of a computer program for performing a series of processes, and the computer program may be recorded on a computer-readable recording medium. Examples of the computer-readable recording medium include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical recording media such as CD-ROMs and DVDs, magneto-optical media such as floptical disks, and hardware devices specifically configured to store and execute program instructions such as ROMs, RAMs, and flash memories.

[0134] Although the present invention has been described above with reference to embodiments, it will be understood by those skilled in the art that various modifications and changes can be made to the present invention without departing from the spirit and scope of the present invention as set forth in the claims below.

[0135] A device for enabling quantitative renal SPECT without CT using a neural network according to one embodiment generates an attenuation map using only SPECT images without CT using a neural network-based model to calculate radiopharmaceutical uptake in the kidney, and can be used in the medical device and digital medical industries.

Claims

1. One or more processors; and A device for generating quantitative renal SPECT without CT using a neural network, the device having a memory for storing instructions executed by one or more processors, One or more of the above processors: A device for generating quantitative kidney SPECT without CT using a neural network by inputting a first SPECT (Single Photon Emission Computed Tomography) image and a second SPECT image of the kidney into a model to generate an attenuation map.

2. In paragraph 1, A device for generating quantitative renal SPECT without CT using a neural network, wherein the first SPECT image is a primary emission SPECT image and the second SPECT image is a scattering SPECT image.

3. In paragraph 1, A device for generating quantitative renal SPECT without CT using a neural network, the model including an encoder for extracting features by downsampling the first SPECT image and the second SPECT image and a decoder for upsampling the features and outputting the attenuation map.

4. In paragraph 3, A device for generating quantitative renal SPECT without CT using a neural network, wherein at least one of the encoder and decoder adjusts the data distribution by applying log-maximum normalization to each voxel in the first SPECT and the second SPECT.

5. In paragraph 3, The above decoder is a device that generates quantitative renal SPECT without CT using a neural network using a loss function defined based on the following [Mathematical Formula 1]. [Mathematical Formula 1] (Here, E is the loss function, G(X) is the predicted attenuation map, X is the input first SPECT image and second SPECT image, Y is the correct value of the attenuation map, N is the number of voxels in the attenuation map, M is the product of the number of voxels in the attenuation map and the number of axes of the gradient, is the image gradient operator, a, b can be any real number, n can be any natural number.) 6. In paragraph 3, The above decoder is a device for generating quantitative renal SPECT without CT using a neural network that performs upsampling through maximum neighbor interpolation.

7. In paragraph 1, The above model is a device for generating quantitative renal SPECT without CT using a neural network that is supervised by learning based on an attenuation map affected by iodine-contrast agent and an attenuation map not affected by iodine-contrast agent.

8. In paragraph 1, One or more of the above processors: A device for generating quantitative kidney SPECT without CT using a neural network, which performs attenuation correction on the first SPECT image using the attenuation map and calculates radiopharmaceutical uptake in the kidney based on the corrected SPECT image.

9. In paragraph 1, One or more of the above processors: A device for generating a quantitative kidney SPECT without CT using a neural network, which performs attenuation correction on the first SPECT image using the attenuation map and calculates the glomerular filtration rate (GFR) in the kidney based on the corrected SPECT image.

10. One or more processors; and A method performed by a device for generating quantitative renal SPECT without CT using a neural network, the device having a memory for storing instructions executed by the one or more processors, The above method is: A method for generating a quantitative kidney SPECT without CT using a neural network, comprising the step of inputting a first SPECT (Single Photon Emission Computed Tomography) image and a second SPECT image of the kidney into a model to generate an attenuation map.

11. In paragraph 10, A method for generating quantitative renal SPECT without CT using a neural network, wherein the first SPECT image is a primary emission SPECT image and the second SPECT image is a scattering SPECT image.

12. In paragraph 10, A method for generating quantitative renal SPECT without CT using a neural network, wherein the model comprises an encoder for extracting features by downsampling the first SPECT image and the second SPECT image and a decoder for upsampling the features and outputting the attenuation map.

13. In paragraph 12, A method for generating quantitative renal SPECT without CT using a neural network, wherein at least one of the encoder and decoder adjusts the data distribution by applying log-maximum normalization to each voxel in the first SPECT and the second SPECT.

14. In paragraph 12, The above decoder is a method for generating quantitative renal SPECT without CT using a neural network, using a loss function defined based on the following [Mathematical Formula 1]. [Mathematical Formula 1] (Here, E is the loss function, G(X) is the predicted attenuation map, X is the input first SPECT image and second SPECT image, Y is the correct value of the attenuation map, N is the number of voxels in the attenuation map, M is the product of the number of voxels in the attenuation map and the number of axes of the gradient, is the image gradient operator, a, b can be any real number, n can be any natural number.) 15. In paragraph 12, The above decoder is a method for generating quantitative renal SPECT without CT using a neural network that performs upsampling through maximum neighbor interpolation.

16. In paragraph 10, The above model is a method for generating quantitative renal SPECT without CT using a neural network that is supervised by learning based on an attenuation map affected by iodine-contrast agent and an attenuation map not affected by iodine-contrast agent.

17. In paragraph 10, The above method A method for generating a quantitative kidney SPECT without CT using a neural network, comprising the step of performing attenuation correction on the first SPECT image using the attenuation map and calculating the radiopharmaceutical uptake in the kidney based on the corrected SPECT image.

18. In paragraph 10, The above method A method for generating a quantitative kidney SPECT without CT using a neural network, comprising the step of performing attenuation correction on the first SPECT image using the attenuation map and calculating the glomerular filtration rate (GFR) in the kidney based on the corrected SPECT image.

Images