X-ray CT scanner and high-resolution image generation device

Abstract

To provide an X-ray CT apparatus that includes a learned model generated by acquiring one or more learning datasets from one pass of imaging without increasing exposure of a subject, and performs machine learning using the acquired learning set.SOLUTION: A learned model is a model after learning in which a low-quality image is input data and a high-quality image is training data. The low-quality image and the high-quality image are obtained on the basis of the same learning measurement data or learning projection data obtained by logarithmically transforming the learning measurement data. The low-quality image is a CT image reconstructed from partial data obtained by dividing the learning measurement data or the learning projection data, and the high-quality image is a CT image obtained by reconstructing the learning projection data.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] The present invention relates to an X-ray CT apparatus equipped with a pre-trained model obtained through machine learning, which is capable of displaying low-noise, high-resolution images. [Background technology]

[0002] An X-ray CT scanner is a device that uses an X-ray source that irradiates a subject with X-rays, and an X-ray detector that detects the amount of X-rays that have passed through the subject as projection data, to reconstruct a tomographic image (hereinafter referred to as "CT image") of the subject using projection data obtained from multiple angles by rotating the X-ray source around the subject, and then displays the reconstructed CT image. The image displayed by the X-ray CT scanner depicts the shape of the organs inside the subject and is used for diagnostic imaging.

[0003] Since X-ray CT scanners use X-rays for imaging, patient exposure is one of the challenges. There is an inverse relationship between the amount of noise and the amount of radiation exposure; reducing noise to improve diagnostic accuracy increases radiation exposure.

[0004] Adaptive filters and iterative reconstruction have been commercialized as technologies to reduce noise without increasing radiation exposure. However, adaptive filters have the problem of blurring of the subject structure due to noise reduction. Iterative reconstruction, due to its iterative nature, generally requires an enormous amount of computation time, and solving this requires expensive computing power. In addition, a decrease in spatial resolution for low-contrast objects and an unnatural appearance due to changes in noise texture can be problematic.

[0005] Therefore, in recent years, attention has been focused on noise reduction and image quality improvement technologies using machine learning, including deep learning. When performing noise reduction using machine learning, at least a high-resolution image (training data) with less noise than the data to be corrected is required. Hereafter, a dataset necessary for learning that includes at least one data to be corrected and at least one training data will be referred to as a "training dataset".

[0006] While noise reduction using machine learning can be highly effective, it faces the challenge of difficulty in preparing training datasets.

[0007] The conventional methods and challenges for obtaining a set of training datasets are described below.

[0008] One typical method involves using images taken at low doses as the data to be corrected and images taken at high doses as the training data. Patent Document 1 also describes a method in which the same area is repeatedly photographed, and the image before superimposing is used as the data to be corrected (low-resolution image), and the image after superimposing is used as the training data (high-resolution image). However, these methods involve photographing the same subject at least twice, which is undesirable from the standpoint of radiation exposure in the case of X-ray CT scanners. Furthermore, when multiple scans are performed in general, the positions of structures such as the body surface, bones, organs, blood vessels, and tumors do not perfectly match due to differences in the patient's posture, respiration, peristalsis, and other body movements, as well as differences in the condition of the equipment, such as differences in the X-ray focal point position, differences in the starting angle of the scan, and the accuracy of the patient's bed position. Images containing positional shifts cannot be considered ideal training data, and machine learning using such training data may create structures that do not actually exist, or erase or blur structures. Thus, training datasets that contain positional shifts due to differences in scanning timing are undesirable.

[0009] Another method involves using images generated by filtered back projection as the data to be corrected and low-noise images generated by iterative reconstruction as the training data. However, iterative reconstruction images tend to have a shift in the noise frequency spectrum to lower frequencies, resulting in reduced spatial resolution for low-contrast objects and altered noise texture. Thus, iterative reconstruction images are not always ideal, and training datasets using different reconstruction methods are undesirable.

[0010] Alternatively, Patent Document 1 describes a method in which a first noise component modeled with Gaussian noise or the like is added to the same image to generate corrected data, and a second noise component is added to generate training data. For example, in the case of optical coherence tomography (OCT) equipment, noise can be estimated based on data acquired without a model eye or the eye under examination, and in the case of OCT angiography (OCTA) equipment, noise appearing in the avascular zone (FAZ) or noise appearing in images of a model eye that schematically reproduces blood flow can be used as a noise model. [Prior art documents] [Patent Documents]

[0011] [Patent Document 1] Japanese Patent Publication No. 2020-166813 [Overview of the project] [Problems that the invention aims to solve]

[0012] Applying the method described in Patent Document 1, which generates corrected data and training data by adding modeled noise to an image, to CT images requires reproducing the effects of electrical circuit noise in the data acquisition device of the X-ray CT scanner, but reproducing the effects of electrical circuit noise is difficult. On the other hand, adding virtually generated noise to an image may lead to an underestimation or overestimation of the actual noise. Therefore, it is difficult to generate corrected data and training data for CT images using the method described in Patent Document 1, which adds modeled noise to an image.

[0013] Thus, in order to perform machine learning, it is necessary to prepare training datasets of several thousand to several hundred thousand pairs, but in the case of X-ray CT scanners, there was a challenge in preparing ideal training datasets.

[0014] An object of the present invention is to obtain one or more sets of learning data sets from a single imaging without increasing the radiation exposure of the subject, perform machine learning using the obtained learning sets to generate a learned model, and mount it on an X-ray CT apparatus.

Means for Solving the Problems

[0015] In order to solve the above problems, the X-ray CT apparatus of the present invention has an X-ray source and an X-ray detector arranged to face each other with a subject interposed therebetween, and rotates the X-ray source and the X-ray detector around the subject to irradiate from the X-ray source, A scan gantry unit that acquires measurement data in which the X-ray detector detects and outputs the X-rays transmitted through the subject, an image reconstruction unit that generates a CT image of the subject using the measurement data, and a high-quality image generation unit. The high-quality image generation unit includes a learned model, and is configured to obtain a high-quality image output by the learned model by receiving the CT image reconstructed by the image reconstruction unit and inputting it into the learned model. The learned model is a model that has been learned in advance using at least one low-quality image generated in advance as input data and at least one high-quality image having higher quality than the low-quality image as teacher data. The low-quality image and the high-quality image are obtained based on the same learning measurement data or learning projection data obtained by logarithmically converting the learning measurement data. The learning measurement data is data obtained by arranging a learning X-ray source and a learning X-ray detector to face each other with a learning subject interposed therebetween, and rotating the learning X-ray source and the learning X-ray detector around the learning subject, and detecting and outputting the X-rays irradiated from the learning X-ray source and transmitted through the learning subject by the learning X-ray detector. The low-quality image is a CT image reconstructed from a part of the learning measurement data or the learning projection data divided, and the high-quality image is a CT image reconstructed from the learning projection data.

Effects of the Invention

[0016] According to the present invention, it is possible to obtain one or more training datasets from a single scan without increasing the radiation exposure of the subject, and to generate a trained model by performing machine learning using the acquired training datasets, thereby providing an X-ray CT apparatus capable of high-precision noise reduction. [Brief explanation of the drawing]

[0017] [Figure 1] This is a diagram showing the overall configuration of an X-ray CT scanner according to an embodiment. [Figure 2] This block diagram shows the configuration of the trained model generation device 130, the image reconstruction device 122, and the high-resolution image generation device 137 according to Embodiment 1. [Figure 3] This flowchart shows the processing flow during model training of the X-ray CT apparatus of Embodiment 1. [Figure 4] This flowchart shows the processing flow during operation (imaging) of the X-ray CT scanner of Embodiment 1. [Figure 5] This block diagram shows the configuration of the trained model generation device 130, the image reconstruction device 122, and the high-resolution image generation device 137 according to Embodiment 2. [Figure 6] This flowchart shows the processing flow according to the first processing method during operation (imaging) of the X-ray CT apparatus of Embodiment 2. [Figure 7] This flowchart shows the processing flow according to the second processing method during operation (imaging) of the X-ray CT apparatus of Embodiment 2. [Figure 8] This flowchart shows the processing flow during operation (imaging) of the X-ray CT scanner of Embodiment 3. [Figure 9] This block diagram shows the configuration of the trained model generation device 130, the image reconstruction device 122, and the high-resolution image generation device 137 according to Embodiment 4. [Figure 10] This flowchart shows the processing flow during model training of the X-ray CT apparatus in Embodiment 4. [Figure 11] This flowchart shows the processing flow during operation (imaging) of the X-ray CT scanner of Embodiment 4.

Embodiment for Carrying Out the Invention

[0018] An X-ray CT apparatus according to an embodiment of the present invention will be described with reference to the drawings. <<Configuration of X-ray CT Apparatus>> First, the X-ray CT apparatus of the present embodiment will be described with reference to the drawings.

[0019] FIG. 1 is an overall configuration diagram of an X-ray CT apparatus 1 to which the present invention is applied. The X-ray CT apparatus 1 includes a scan gantry unit 100 and an operation console 120.

[0020] The scan gantry unit 100 includes an X-ray source 101, a rotating disk 102, a collimator unit 103, a couch 105, an X-ray detector 106, a data acquisition device 107, a gantry control device 108, a couch control device 109, and an X-ray control device 110.

[0021] The rotating disk 102 has an opening 104 into which a subject placed on the couch 105 enters, and mounts the X-ray source 101, the X-ray detector 106, and the data acquisition device 107, and rotates around the imaging region where the subject is disposed.

[0022] The X-ray detector 106 is disposed opposite to the X-ray source 101, and is a device that measures the spatial distribution of transmitted X-rays by detecting the X-rays transmitted through the subject for each rotation angle (view). The X-ray detector 106 is formed by arranging, for example, about 1000 X-ray detection element groups each constituted by a combination of a scintillator and a photodiode in the rotation direction (channel direction) of the rotating disk 102 and about 1 to 320 in the rotation axis direction (column direction).

[0023] The data acquisition device 107 collects the X-ray dose detected by the X-ray detector 106, converts it into digital data, and outputs it sequentially to the image reconstruction device 122. The gantry control device 108 controls the rotation of the rotating disk 102. The bed control device 109 controls the up, down, left, right, forward, and backward movement of the bed 105. The X-ray control device 110 controls the power input to the X-ray source 101.

[0024] The control console 120 includes an input device 121, an image reconstruction device 122, a high-resolution image generation device 137, a display device 125, a storage device 123, and a system control device 124.

[0025] The input device 121 is a device for inputting the subject's name, examination date and time, imaging conditions, etc. Specifically, it is a pointing device such as a keyboard or mouse, various switch buttons, etc. The input device 121 may also be a touch panel type input device that is integrated with the display screen of the display device 125.

[0026] The image reconstruction device 122 is a device that acquires measurement data sent from the data acquisition device 107, performs calculations, and reconstructs CT images.

[0027] The display device 125 is connected to the system control device 124 and displays CT images reconstructed by the image reconstruction device 122, as well as various information handled by the system control device 124. The display device 125 is composed of a display device such as a liquid crystal panel or a CRT monitor, and a logic circuit for performing display processing in cooperation with the display device.

[0028] The storage device 123 is a device that stores measurement data collected by the data acquisition device 107, and / or projection data as described later, and image data of CT images created by the image reconstruction device 122. Specifically, the storage device 123 is a data recording device such as an HDD (Hard Disk Drive) or an SSD (Solid State Drive).

[0029] The system control unit 124 controls the above-mentioned devices in the control console 120, as well as the gantry control unit 108, the patient bed control unit 109, and the X-ray control unit 110. The system control unit 124 is a computer equipped with a CPU (Central Processing Unit), GPU (Graphics Processing Unit), ROM (Read Only Memory), RAM (Random Access Memory), etc.

[0030] The structure of the high-resolution image generation device 137 will be described in Embodiments 1 to 4 below.

[0031] In the X-ray CT apparatus with the configuration described above, the X-ray control device 110 controls the power input to the X-ray source 101 based on the imaging conditions input from the input device 121, particularly the X-ray tube voltage and X-ray tube current. As a result, the X-ray source 101 irradiates the subject with X-rays according to the imaging conditions. The X-ray detector 106 detects the X-rays irradiated from the X-ray source 101 and transmitted through the subject with a number of X-ray detection elements and measures the distribution of transmitted X-rays. The rotating disk 102 is controlled by the gantry control device 108 and rotates at a rotation speed of, for example, about 0.2 to 2 seconds per rotation, based on the imaging conditions input from the input device 121. The patient table 105 is controlled by the patient table control device 109. Known scan types include "axial scan," in which the patient table is not moved during X-ray irradiation, and "helical scan," in which the patient table is moved during X-ray irradiation.

[0032] As the X-ray source 101 irradiates and the X-ray detector 106 measures the transmitted X-ray distribution, the rotation of the rotating disk 102 is repeated, and measurement data from various angles (views) is acquired by the data acquisition device 107. The measurement data is transmitted from the data acquisition device 107 to the image reconstruction device 122. The image reconstruction device 122 performs a logarithmic transformation on the input measurement data and converts the measurement data into "projection data". Furthermore, the image reconstruction device 122 obtains a CT image by performing a back projection on the projection data.

[0033] The high-resolution image generation device 137 receives the CT image and enhances its image quality. The enhanced CT image is then displayed on the display device 125.

[0034] A CT image is a collection of one or more cross-sectional images, and the position of these cross-sectional images will be referred to as the "cross-sectional position" in the following explanation. <<Embodiment 1>> <Configuration of the X-ray CT apparatus in Embodiment 1> The configuration of the X-ray CT apparatus according to Embodiment 1 of the present invention will be further explained with reference to Figure 2.

[0035] As described above, the control console 120 is equipped with an image reconstruction device 122 and a high-resolution image generation device 137. A storage unit 135 is located inside the high-resolution image generation device 137, and a trained model is pre-stored in the storage unit 135.

[0036] The image reconstruction device 122 receives measurement data from the data acquisition device 107, performs logarithmic transformation to obtain projection data, and generates a CT image by back-projecting the projection data.

[0037] The high-resolution image generation device 137 receives the CT image reconstructed by the image reconstruction device 122, reads the trained model 136 stored in the memory unit 135, and inputs it into the trained model to obtain an image output by the trained model 136. The image output by the trained model 136 has reduced noise and improved image quality compared to the input CT image. This image is displayed on the display device 125.

[0038] Any machine learning model can be used; for example, a well-known convolutional neural network (CNN) can be used. A CNN consists of an input layer that takes at least one image as input, an intermediate layer, and an output layer that outputs at least one image.

[0039] The trained model 136 is generated by the trained model generation device 130. The trained model generation device 130 is generated using a projection data splitting unit 131, a low-resolution image generation unit 132, a training image generation unit 133, and a trained model generation unit 134.

[0040] The pre-trained model 136 is a pre-trained model generated using a training dataset that consists of at least one low-resolution image as input data and at least one high-resolution image (higher resolution than the low-resolution image) as training data. During model generation (machine learning), several thousand to several hundred thousand sets of training datasets are used.

[0041] Here, the low-resolution and high-resolution images are obtained based on the same training measurement data, or training projection data obtained by logarithmically transforming the training measurement data. The training measurement data is obtained by rotating the training X-ray source and training X-ray detector around the training subject, with the training X-ray source and training X-ray detector facing each other with the training subject in between, and detecting the X-rays irradiated from the training X-ray source and transmitted through the training subject by the training X-ray detector.

[0042] Low-resolution images are CT images reconstructed from a portion of the training measurement data or training projection data. High-resolution images are CT images created by combining multiple low-resolution images, or CT images directly reconstructed from the training projection data. The low-resolution and high-resolution images are images from the same cross-sectional location.

[0043] Thus, in Embodiment 1, low-resolution and high-resolution images can be generated from the same training measurement data or training projection data. Therefore, without increasing the radiation exposure of the training subjects, it is possible to easily obtain training datasets (low-resolution and high-resolution images) with no temporal lag and different noise levels and artifacts from a single shooting session.

[0044] Furthermore, since the low-resolution images are reconstructed from a portion of the training measurement data or training projection data, they are essentially equivalent to images reconstructed from logarithmically transformed projection data obtained with a small number of samples. Therefore, the low-resolution images are noisy. On the other hand, the high-resolution images are high-resolution images reconstructed from training projection data with a large number of samples, and therefore have less noise. Thus, the trained model 136 generated by machine learning the model using these low-resolution images as input data and the high-resolution images as training data has a high noise reduction effect and an effect of reducing artifacts caused by insufficient sampling.

[0045] The training projection data may be projection data acquired by the scan gantry unit 100 of the X-ray CT device in which the trained model 136 is stored, or projection data acquired by another X-ray CT device may be used.

[0046] Below, we will explain in detail, using Figures 2 and 3, an example of obtaining a low-resolution image by reconstructing data obtained by dividing training projection data (hereinafter referred to as training partial projection data). Note that when dividing training measurement data, training partial projection data can be obtained by logarithmically transforming the divided training measurement data.

[0047] The trained model generation device 130 receives projection data acquired by the scan gantry unit 100 of the X-ray CT device in which the trained model 136 is stored, or projection data acquired by another X-ray CT device.

[0048] The projection data splitting unit 131 divides the training projection data into multiple partial projection data. For example, it extracts some data (training partial projection data) from the training projection data in one of the following directions: the view direction, column direction, or channel direction of the training X-ray detector.

[0049] The low-resolution image generation unit 132 generates low-resolution CT images (low-resolution images) by reconstructing training partial projection data.

[0050] Here, the low-resolution image generation unit 132 generates a low-resolution image for each of the divided partial projection data. This generates multiple low-resolution images of the same cross-sectional position. The training image generation unit 133 generates a high-resolution image (training image) by combining the multiple low-resolution images. Alternatively, the training projection data could be directly reconstructed to generate a high-resolution image (training image). However, generating a low-resolution image for each of the divided partial projection data, and then combining the multiple low-resolution images to generate a high-resolution image (training image), allows for the generation of a set of high-resolution and low-resolution images in a shorter time than when the undivided training projection data is directly reconstructed to generate a high-resolution image (training image). This is because the image reconstruction time increases with the amount of projection data.

[0051] The trained model generation unit 134 uses one of the low-resolution images as the data to be corrected (input data) and a high-resolution image (training image) as the training data as a training dataset, and generates at least one trained model 136 by performing supervised machine learning.

[0052] Furthermore, all or part of the processing of each part 131 to 134 within the trained model generation device 130 may be performed within the X-ray CT device 1. Alternatively, the processing of each part 131 to 134 within the trained model generation device 130 may be performed using a personal computer equipped with a processor such as a CPU or GPU, or a cloud environment. However, since the machine learning of the trained model generation unit 134 is computationally intensive and requires expensive computing resources such as CPUs and GPUs, it is preferable to perform machine learning in a high-performance computing environment beforehand and then install the results into the X-ray CT device 1.

[0053] In the following, the process by which each part 131 to 134 within the trained model generation device 130 generates a trained model will be referred to as "training," and the process by which the high-resolution image generation device 137 generates high-resolution images using the trained model 136 will be referred to as "operation" of the X-ray CT device.

[0054] <Operation of the trained model generation device 130 in Embodiment 1 during training> The operation of the trained model generation device 130 in Embodiment 1 during training will be explained using the flowchart in Figure 3.

[0055] In Embodiment 1, the trained model generation device 130 is configured using software. Specifically, the trained model generation device 130 includes a processor such as a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit) and memory. The CPU reads and executes a program stored in memory to realize the functions of each part of the trained model generation device 130. Alternatively, some or all of the trained model generation device 130 can be configured using hardware. For example, a custom IC such as an ASIC (Application Specific Integrated Circuit) or a programmable IC such as an FPGA (Field-Programmable Gate Array) can be used to design a circuit that realizes the functions of each part of the trained model generation device 130.

[0056] (Step S001) First, in step S001, the projection data splitting unit 131 splits the learning projection data acquired by X-ray CT apparatus 1 or another X-ray CT apparatus into multiple learning partial projection data sets. Specifically, for example, the projection data splitting unit 131 assigns learning projection data sets with relatively close numbers in any of the view number (rotation angle), column number, and channel number to separate learning partial projection data sets. In this way, the projection data splitting unit 131 generates multiple learning partial projection data sets with the same acquisition timing and the same cross-sectional position.

[0057] For example, consider the case where we split by view number. We have training projection data with view count I and view numbers i=0, 1, 2, ..., I-1. We want to split this into two sets: a first training subprojection with view count a, and a second training subprojection with view count b. As one method of splitting, let's take the case where I is even. i = 0, 2, 4, ..., I-2 are even view numbers (view count a = I / 2), Odd view numbers i=1, 3, 5, ..., I-1 (view count b=I / 2), This method of division is used. If I is odd, there will be one more view with an even number than with an odd number, but the division can be done similarly. This method of dividing into even and odd views will be called "even-odd division" below.

[0058] As another method of division, let's take the case where the number of views I is a multiple of 3 as an example. i=0, 3, 6, ..., I-3 view numbers (view count a = I × 1 / 3), i = 1, 2, 4, 5, 7, 8, ..., I-2, I-1 are view numbers (view count b = I × 2 / 3), It may also be divided into two sets. This allows the training projection data to be divided into two sets of training segmented projection data with an uneven ratio of 1:2 in terms of the number of views. If the number of views I is not a multiple of 3, one of the sets will have one more view, but it can be divided in the same way. The ratio of the number of views between the first and second training segmented projection data can be arbitrary, such as 2:1, 1:3, or 2:3. This method of unevenly dividing the training projection data will be referred to as "uneven partitioning" below.

[0059] Furthermore, in the above example, the projection data splitting unit 131 splits the training projection data into two parts in the view direction, but it may also split it into three or more parts. Additionally, the projection data splitting unit 131 may split the training projection data in the column direction and the channel direction.

[0060] (Step S002) Next, in step S002, the low-quality image generation unit 132 reconstructs the first learning partial projection data and the second learning partial projection data into images respectively, to generate the first low-quality image and the second low-quality image. As a result, two low-quality images at the same cross-sectional position are obtained. The low-quality image generation unit 132 uses, for example, a known filter-corrected backprojection method or a convolution-corrected backprojection method for image reconstruction.

[0061] (Step S003) Next, in step S003, the first low-quality image and the second low-quality image at the same cross-sectional position are combined to generate a teacher image. As the combination method, a weighted average method can be used to weight the pixel values of the corresponding positions of the first low-quality image and the second low-quality image, for example, by variance value or data amount.

[0062] The weighted average method will be described using specific mathematical formulas. In image reconstruction by linear processing such as the filter-corrected backprojection method, the data amount and the variance value of the image (the square value of image noise) are in an inverse proportional relationship. Therefore, from the ratio of the projection data amounts, the variance values σ1 2 , σ2 2 of the first low-quality image and the second low-quality image at a certain cross-sectional position are represented by Formula 1 and Formula 2 respectively. σ1 2 =(a + b) / a × σ 2 (Formula 1) σ2 2 =(a + b) / b × σ 2 (Formula 2) However, a is the number of views of the first learning partial projection data, b is the number of views of the second learning partial projection data. σ 2 is the variance value of the image when reconstructed using the total number of views I (= a + b).

[0063] Using these variance values σ1 2 , σ2 2 , the pixel values of the first low-quality image and the second low-quality image are weighted and averaged according to Formulas 3 - 5. (The first low-quality image) × k1 + (The second low-quality image) × k2 = (The teacher image) (Formula 3) however, k1 = σ² 2 / (σ1 2 +σ2 2 ) = a / (a+b) (Equation 4) k2=σ1 2 / (σ1 2 +σ2 2 )=b / (a+b) (Equation 5)

[0064] As shown in equations 3-5, the weight of the pixel values ​​of the lower-resolution images with less data and higher variance is reduced. When a=b, (equation 3) is equivalent to the arithmetic average.

[0065] At this time, the variance of the training image is σ3 2 This is expressed by the following equation 6. σ3 2 =σ1 2 ×k1 2 +σ2 2 ×k2 2 =σ 2 (Formula 6) This is the result.

[0066] Comparing (Equation 1), (Equation 2), and (Equation 6), the high-resolution image (training image) obtained by image stitching has less image noise than the first low-resolution image and the second low-resolution image.

[0067] While this example demonstrates combining two low-resolution images, the same method can be used to combine three or more low-resolution images.

[0068] In this way, a set of high-resolution images and lower-resolution images with higher noise can be easily obtained. In this embodiment, the first and second training partial projection data, which have a small amount of data, are reconstructed to obtain the first low-resolution image and the second low-resolution image, and then these are combined to obtain the high-resolution image. Therefore, the reconstruction time required to obtain the high-resolution and low-resolution images is proportional to a+b. If the high-resolution image is reconstructed from undivided training projection data, the reconstruction time required to obtain the high-resolution and low-resolution images will be proportional to I+a or I+b, which is longer than in this embodiment. Therefore, in this embodiment, a set of training images and low-resolution images can be obtained in a shorter time.

[0069] Furthermore, we will explain the features of this training image. In X-ray CT scanners, it is known that streak artifacts, partial volume effects, and alias artifacts occur when sampling is insufficient in the view direction, column direction, and channel direction, respectively. The first and second training partial projection data sets have undergone data downsampling, resulting in insufficient sampling. The first and second low-resolution images generated based on these data sets contain artifacts caused by insufficient sampling. On the other hand, the training image has a sufficient amount of data, and the artifacts caused by insufficient sampling have been improved. Therefore, the training image is a high-resolution image with low image noise and reduced artifacts caused by insufficient sampling.

[0070] The training images may also be obtained by reconstructing undivided training projection data.

[0071] (Step S004) Next, in step S004, the trained model generation unit 134 obtains a training dataset using one of the low-resolution images of the same cross-sectional position as input data and the training image as training data.

[0072] In step S001, if the projection data division unit 131 performs an even-odd division, the first low-resolution image and the second low-resolution image are equivalent, so the trained model generation unit 134 only needs to use one of them as input data. On the other hand, if the projection data division unit 131 performs an uneven division in step S001, it is expected that a trained model 136 with a higher noise reduction effect can be obtained by using the low-resolution image with greater image noise from the first low-resolution image and the second low-resolution image as input data. Alternatively, the first trained model 136 can be generated using the first low-resolution image as input data, and the second trained model 136 can be generated using the second low-resolution image as input data, thereby obtaining multiple trained models 136 to use interchangeably during operation.

[0073] The trained model generation unit 134 uses, for example, a well-known convolutional neural network (CNN) as the machine learning model. A CNN is composed of an input layer that takes at least one image as input, an intermediate layer, and an output layer that outputs at least one image. The trained model generation unit 134 takes input data (low-resolution images) as input data (high-resolution images) as input data as input data as the input layer of the machine learning model, and training data (high-resolution images) as input as the output layer. The intermediate layer constructs the network (CNN) so that the output image from the output layer is close to the training data, that is, so that the difference between the output image and the training data is small. Through this machine learning, a trained model 136 is obtained that improves the image quality of the input data.

[0074] Furthermore, the number of input data (low-resolution images) and training data (high-resolution images) is not limited to one image each. For example, to reflect features in the axial direction, three input data (low-resolution images) may be used, and one or more images may be output. Alternatively, residual learning based on a CNN may be used.

[0075] The trained model generation unit 134 stores the obtained trained model 136 in the storage device 123.

[0076] Typically, one to several thousand cross-sectional images can be obtained from a single X-ray CT scan of a training subject (one case). Therefore, according to Embodiment 1, one to several thousand training datasets can be obtained from a single X-ray CT scan. Thus, it is not necessary to scan the same training subject multiple times to obtain training datasets. Furthermore, by increasing the number of cases, it is possible to obtain training datasets with the amount of data necessary for machine learning (e.g., several thousand to several hundred thousand sets). By preparing a large number of training datasets, it is possible to create a trained model that can reduce noise and output high-resolution images for various cases.

[0077] Since machine learning typically involves validation and testing after training, it is acceptable to use a portion of the training dataset for training and another portion for validation and testing.

[0078] <During operation of the X-ray CT scanner of Embodiment 1> The operation of the X-ray CT scanner 1 of Embodiment 1 during operation (when imaging a subject) will be explained using the flowchart in Figure 4.

[0079] (Step S012) First, in step S012, the image reconstruction device 122 receives projection data of the subject obtained by the scan gantry unit 100 irradiating the subject with X-rays, reconstructs the image, and generates a CT image of the subject. The image reconstruction method may be, for example, a known filter-corrected back projection method or a known iterative reconstruction method.

[0080] (Step S014) Next, in step S014, the high-resolution image generation device 137 reads the trained model 136 that has been previously stored in the storage device 123 and inputs the CT image of the subject generated in step S012 into the input layer. As a result, the trained model 136 outputs a high-resolution image of the subject's CT image from the output layer. The high-resolution image generation device 137 displays the high-resolution image on the display device 125 and stores it in the storage device 123.

[0081] As described above, in Embodiment 1, multiple training datasets (low-resolution images and high-resolution images) with different noise levels and artifacts, without temporal lag, can be easily obtained from a single scan without increasing the radiation exposure of the training subjects. Furthermore, the low-resolution images input to the trained model 136 are essentially noisy images acquired with a small number of samples, while the high-resolution images input as training data are reconstructed from training projection data with a large number of samples. Therefore, the trained model 136 using these training datasets has a high noise reduction effect and an artifact reduction effect due to insufficient sampling.

[0082] <<Embodiment 2>> An X-ray CT apparatus according to Embodiment 2 of the present invention will be described.

[0083] In Embodiment 1, the projection data was split during training of the pre-trained model 136, and the low-resolution images were used as input data. However, during operation, the projection data of the subject was not split, and the entire projection data of the subject was reconstructed to obtain a CT image, which was then used as input data for the pre-trained model 136. As a result, the CT image input to the pre-trained model 136 during operation had twice the data volume (in the case of even-odd splitting) compared to the low-resolution image input during training, resulting in a significant difference in the amount of input data between training and operation.

[0084] If there is a difference in the amount of data between the images used as input data during training and during operation, there are concerns that the noise reduction and artifact reduction performance of the trained model 136 may not be fully utilized during operation, and that unexpected image quality degradation may occur.

[0085] In particular, streak artifacts are affected by the number of views in the projection data. For example, if CT images reconstructed from projection data with a higher number of views than during training are used as input data for the trained model 136 during operation, there is a concern that the streak artifact reduction effect will be too strong, resulting in image quality degradation due to overcorrection in the output data of the trained model 136. Therefore, in Embodiment 2, with the aim of making the amount of input data as similar as possible between training and operation, the amount of CT image data input to the trained model 136 during operation is reduced to be equivalent to the input data during training.

[0086] Specifically, as shown in Figure 5, the X-ray CT apparatus of Embodiment 2 includes an image reconstruction apparatus 122 comprising a projection data division unit 141 that divides the measurement data of the subject or projection data obtained by logarithmically transforming the measurement data, and a low-resolution image generation unit 142 that reconstructs a CT image (low-resolution image) from a portion of the data after the projection data division unit 141 has divided it. For example, the projection data division unit 141 selects a portion of the measurement data or projection data in either the view direction, column direction, or channel direction. The low-resolution image generation unit 142 generates a reconstructed CT image using the selected portion of the data.

[0087] In this case, it is preferable that the projection data splitting unit 141 splits the data such that the number of views of the partial projection data used for training when generating the trained model is approximately equal to the number of views of the partial data after the projection data splitting unit 141 has split it during operation. Specifically, the projection data splitting unit 141 splits the subject's measurement data or projection data under the same conditions as when the training measurement data or training projection data was split when generating low-resolution images for training the trained model. Alternatively, the projection data splitting unit 141 splits the subject's measurement data or projection data such that the amount of data contained in the low-resolution images used for training the trained model 136 matches the amount of data contained in the CT images (low-resolution images) reconstructed by the low-resolution image generation unit 142. These splitting processes will be explained in detail later.

[0088] The high-resolution image generation device 137 inputs a CT image (low-resolution image) reconstructed from a portion of the data after division by the projection data division unit 141 into the trained model 136, thereby obtaining a high-resolution image from the trained model 136.

[0089] Furthermore, if the amount of input data is the same during training and operation, even without splitting the measurement data or projection data during operation (for example, if 2000 views of projection data are split into even-odd 1000-view segments for training, and 1000 views of projection data are used during operation), then the procedure in Embodiment 1 can be followed, and this will be omitted. In this embodiment, the following explanation assumes that the projection data splitting unit 141 splits the projection data during operation. The projection data after splitting by the projection data splitting unit 141 is referred to as partial projection data.

[0090] (Configuration of the X-ray CT scanner in Embodiment 2) Figure 5 shows the configuration of Embodiment 2. The difference from Embodiment 1 is that the image reconstruction device 122 includes a projection data splitting unit 141 and a low-resolution image generation unit 142. The high-resolution image generation device 137 includes an image merging unit 138 in addition to a storage unit 135. The projection data splitting unit 141 splits the projection data of the subject into multiple partial projection data. The low-resolution image generation unit 142 reconstructs each of the partial projection data to generate multiple low-resolution images. The image merging unit 138 generates a single image by merging multiple high-resolution images.

[0091] (During operation of Embodiment 2) Regarding the operation of Embodiment 2, two processing methods are possible.

[0092] <First processing method> The first processing method of Embodiment 2 will be explained using the flow chart in Figure 6. In the processing method of the flow chart in Figure 6, the projection data of the subject is divided using the same projection data division method as during training.

[0093] (Step S011) In step S011, the projection data splitting unit 141 splits the subject's projection data using the same projection data splitting method as used during training of the trained model 136. For example, if even-odd splitting was performed during training, even-odd splitting is also performed during operation, splitting the subject's projection data into a first partial projection data and a second projection data. Generally, the number of views I of the training projection data during training and the number of views J of the subject's projection data during operation do not necessarily match. Therefore, if we let J be the number of views of the subject's projection data and j be the view number, Even view numbers j=0, 2, 4, ..., J-2, Odd view numbers j=1, 3, 5, ..., J-1, The data is divided into even-odd and odd parts. Here, J is shown as an even number, but the same applies if it is an odd number.

[0094] If the number of views I of the training projection data during training and the number of views J of the subject projection data during operation can be considered to be approximately equal to I ≈ J, then in Embodiment 2, the amount of input data during training and operation can be made closer compared to Embodiment 1, which does not involve division.

[0095] (Step S012) In step S012, the low-resolution image generation unit 142 performs image reconstruction on the first partial projection data and the second projection data that were divided in step S011 to obtain a first low-resolution image and a second low-resolution image.

[0096] (Step S014) In step S014, the first low-resolution image and the second low-resolution image are input to the trained model 136 as input data, and the trained model 136 outputs a noise-reduced first high-resolution image and a second high-resolution image.

[0097] Since the amount of low-resolution images input to the trained model 136 during training and operation is the same, artifact reduction is achieved for the first high-resolution image and the second high-resolution image, just as during training.

[0098] (Step S015) In step S015, the first high-resolution image and the second high-resolution image are combined to generate a third high-resolution image. For the combining method, a weighted average similar to (Equation 3) to (Equation 5) is used.

[0099] <Second processing method> Next, the second processing method of Embodiment 2 will be described with reference to Figure 7. In the second processing method, the projection data splitting unit 141 splits the subject's measurement data or projection data so that the amount of data contained in the low-resolution images used for training the trained model 136 matches the amount of data contained in the CT images (low-resolution images) reconstructed by the low-resolution image generation unit 142.

[0100] (Step S010) The storage device 123 of the X-ray CT apparatus 1 pre-stores a typical amount of input data (low-resolution images) used to generate the trained model 136, for example, the number of views a. The projection data division unit 141 obtains the amount of projection data of the subject acquired by the scan gantry unit 100, for example, the number of views J.

[0101] In step S010, the projection data division unit 141 calculates the number of divisions N using the following formula so that the amount of data in the input data during training and during operation are the same.

[0102] N=J / a Note that N will be rounded or otherwise modified so that it becomes an integer.

[0103] (Step S011) In step S011, the projection data division unit 141 divides the subject's projection data into N parts. For example, the partial projection data n=0, 1, ..., N-1 is j n Divide the area evenly so that the view numbers are =n, N+n, 2N+n, ... If N is 1 or less, do not divide the area and follow Embodiment 1. Figure 7 shows an example when N=2.

[0104] (Steps S012~S014) Steps S012 to S014 and beyond are the same as the first implementation method of Embodiment 2, so their explanation will be omitted.

[0105] In the X-ray CT apparatus of Embodiment 2, the amount of input data to the trained model 136 during training and operation is the same, making it possible to reduce noise and artifacts with high accuracy.

[0106] <<Embodiment 3>> The X-ray CT apparatus of Embodiment 3 will now be described.

[0107] In Embodiment 3, the projection data splitting unit 141 splits the subject's measurement data or projection data such that the amount of noise contained in the low-resolution images used to train the trained model 136 matches the amount of noise contained in the CT images (low-resolution images) reconstructed by the low-resolution image generation unit 142.

[0108] Specifically, the projection data division unit 141 includes a prediction unit (not shown) that predicts the amount of noise obtained when the projection data (undivided) of the subject is reconstructed as an image. Based on the amount of noise predicted by the prediction unit, the projection data division unit 141 determines the number of divisions N of the subject's measurement data or projection data so that the predicted amount of noise matches the amount of noise in the low-resolution image used to train the trained model 136.

[0109] Generally, image noise can take on various values ​​depending on differences in imaging and reconstruction conditions such as the subject's physique, the area being scanned, the internal structure of the subject, the dose, the slice thickness, and the reconstruction filter. If sufficient computing resources are available during training, it would be possible to create dozens or more pre-trained models 136 for each slice thickness or reconstruction filter, anticipating all operational conditions. However, in reality, due to the limitations of computing resources, machine learning takes a considerable amount of processing time, so it is more practical to create pre-trained models 136 for one to at most a few representative conditions. Therefore, the image noise of the input data used to create the pre-trained models 136 may differ significantly from the image noise of the input data used in actual operation.

[0110] If the noise level of the input data for the trained model 136 differs between the training and operation phases, there is a concern that noise reduction and artifact reduction performance may decrease during operation, or that unexpected image quality degradation may occur. Therefore, in Embodiment 3, the noise level of the input data is made as similar as possible between the training and operation phases.

[0111] (Configuration of Embodiment 3) The configuration of the X-ray CT apparatus in Embodiment 3 is the same as in Figure 5. However, the projection data splitting unit 141 is equipped with a prediction unit (not shown) that predicts the amount of noise obtained when the projection data of the subject (unsplit) is reconstructed as an image.

[0112] (During operation of Embodiment 3) The operation of the X-ray CT apparatus in Embodiment 3 will be explained using the flowchart in Figure 8.

[0113] (Step S020) The typical image noise level or variance of the input data (low-resolution images) used to generate the trained model 136, for example, the image noise level σ1 or variance σ1 of the first low-resolution image. 2 However, this information is pre-stored in the storage device 123 of the X-ray CT scanner 1.

[0114] The prediction unit calculates the image noise σ obtained when the projection data (undivided) of the subject is reconstructed during operation. X or variance value σ X 2 This predicts the image noise or variance. Here, the prediction method for predicting the image noise or variance can be any known CT-AEC (Auto Exposure Control) method, for example, the method described in Japanese Patent No. 4731151.

[0115] In step S020, the projection data division unit 141 calculates the number of divisions N such that the amount of noise obtained when reconstructing the image using the input data (low-resolution image) used during training of the trained model 136 and the projection data of the subject during operation (undivided) is the same. Specifically, σ1 2 , σX 2 The number of divisions N is calculated using the following formula. N=σ² 2 / σ X 2 Note that N will be rounded or otherwise modified so that it becomes an integer.

[0116] (Steps S011~S014) Steps S011 to S014 are the same as in Embodiment 2, so their explanation is omitted.

[0117] Note that the flowchart in Figure 8 is an example for the case where the number of divisions N=2.

[0118] In the X-ray CT apparatus of Embodiment 3, the amount of noise in the input data during training and operation of the trained model 136 is the same, so an X-ray CT apparatus capable of high-precision noise reduction and artifact reduction can be provided.

[0119] <<Embodiment 4>> The X-ray CT apparatus of Embodiment 4 will now be described.

[0120] The X-ray CT apparatus of Embodiment 4 has the same configuration as the X-ray CT apparatus of Embodiment 2, but differs from Embodiment 2 in that it uses difference images of low-resolution images as additional data during training and operation of the trained model 136.

[0121] As disclosed in U.S. Patent No. 7,706,497 and Japanese Patent No. 6,713,860, a noise image with the signal removed can be obtained by dividing the projection data into even and odd divisions, reconstructing each image, and taking the difference between the two images.

[0122] In Embodiment 4, the noise reduction accuracy of the trained model 136 is improved by inputting this difference image (noise image) into the model as additional data for machine learning.

[0123] (Configuration of Embodiment 4) Figure 9 shows the configuration of the X-ray CT apparatus of Embodiment 4. The X-ray CT apparatus of Embodiment 4 has a similar configuration to the apparatus of Embodiment 2, but differs from the apparatus of Embodiment 4 in that the trained model generation device 130 is equipped with a difference image generation unit 139 and the image reconstruction device 122 is equipped with a difference image generation unit 143.

[0124] (During the learning phase of the X-ray CT scanner in Embodiment 4) The operation of the model in Embodiment 4 during training will be explained using Figure 10.

[0125] (Step S001) In step S001, the projection data division unit 131 divides the training projection data into even and odd divisions to obtain the first partial projection data and the second partial projection data.

[0126] (Steps S002, S003) Steps S002 and S003 are the same as in Embodiment 1, in which the low-resolution image generation unit 132 generates a first low-resolution image and a second low-resolution image from the first partial projection data and the second partial projection data. The training image generation unit 133 combines the first low-resolution image and the second low-resolution image to obtain a high-resolution image (training image).

[0127] (Step S005) On the other hand, in step S005, the difference image generation unit 139 takes the difference between the first low-resolution image and the second low-resolution image and generates a difference image. Since the images of the training subject included in the first low-resolution image and the second low-resolution image are at almost the same cross-sectional position, the difference image is an image consisting only of noise.

[0128] (Step S006) In step S006, supervised machine learning is performed using either the first or second low-resolution image as input data, the training image as training data, and the difference image as additional data. In machine learning, as described in Embodiment 1, for example, a known convolutional neural network (CNN) is used. The CNN is composed of an input layer that takes at least one image as input, an intermediate layer, and an output layer that outputs at least one image. When the input data, training data, and difference image are input, the intermediate layer constructs a network (trained model 136) so that the output image and the training data are close together, that is, so that the difference between the output image and the training data is small. Through machine learning, a trained model 136 that reduces image noise and artifacts can be generated. Since the difference image contains information on the noise distribution for each pixel, by adding the difference image, a trained model 136 can be obtained that can remove noise components more accurately while retaining signal components.

[0129] (During operation of the X-ray CT scanner of Embodiment 4) The operation of Embodiment 4 during use will be explained with reference to Figure 11.

[0130] (Step S006) In step S011, the projection data division unit 141 divides the subject's projection data into even and odd divisions to obtain the first partial projection data and the second partial projection data.

[0131] (Step S012) In step S012, the low-resolution image generation unit 142 performs image reconstruction on the first partial projection data and the second partial projection data, respectively, to obtain the first low-resolution image and the second low-resolution image.

[0132] (Step S023) In step S023, the difference image generation unit 143 calculates the difference between the first low-resolution image and the second low-resolution image and generates a difference image.

[0133] (Step S024) In step S024, the first low-resolution image and the difference image are input to the trained model 136 as input data and additional data, respectively, to obtain a third high-resolution image with reduced noise. Similarly, the second low-resolution image and the difference image are input to the trained model 136 as input data and additional data, respectively, to generate a fourth high-resolution image with reduced noise.

[0134] (Step S025) In step S025, the third high-resolution image and the fourth high-resolution image are combined to generate a fifth high-resolution image. For the combining method, a weighted average similar to (Equation 3) to (Equation 5) is used.

[0135] Embodiment 4 provides an X-ray CT apparatus that can accurately reduce only the noise by adding a difference image that reflects the noise component as additional data for machine learning.

[0136] Although preferred embodiments of the X-ray CT apparatus according to the present invention have been described above, the present invention is not limited to the embodiments described above. Furthermore, it is clear to those skilled in the art that various modifications or alterations can be conceived within the scope of the technical idea disclosed herein, and these will naturally also fall within the technical scope of the present invention. [Explanation of Symbols]

[0137] 1 X-ray CT device 100 Scan Gunner Unit 101 X-ray source 102 Rotating Discs 103 Collimator Unit 104 Opening 105 berths 106 X-ray detector 107 Data acquisition device 108 Gantry Control Device 109 Bed control device 110 X-ray control device 120 Control console 121 Input device 122 Image reconstruction device 123 Storage device 124 System Control Unit 125 Display device 130 Model Generator 131 Projection data division section 132 Low-resolution image generation unit 133 Training Image Generation Unit 134 Pre-trained model generation unit 135 Storage section 136 pre-trained models 137 High-resolution image generation device 138 Image merging section 139 Difference Image Generation Unit 141 Projection data division section 142 Low-resolution image generation unit 143 Difference Image Generation Unit

Claims

1. A scanning gantry unit that acquires measurement data by rotating the X-ray source and the X-ray detector around the subject with the subject in the position of facing each other with the subject in between, thereby detecting and outputting X-rays irradiated from the X-ray source and transmitted through the subject by the X-ray detector; an image reconstruction unit that generates a CT image of the subject using the measurement data; and a high-resolution image generation unit. The high-resolution image generation unit includes a trained model, and receives the CT image reconstructed by the image reconstruction unit and inputs it to the trained model to obtain a high-resolution image output by the trained model. The aforementioned trained model is a model that has been trained using at least one pre-generated low-resolution image as input data and at least one high-resolution image, which is of higher resolution than the low-resolution image, as training data. The low-resolution image and the high-resolution image are obtained based on the same training measurement data, or training projection data obtained by logarithmically transforming the training measurement data. The training measurement data is data detected and output by the training X-ray detector when the training X-ray source and the training X-ray detector are rotated around the training subject with the training subject in between, while the training X-ray source and the training X-ray detector are positioned opposite each other with the training subject in between. The aforementioned low-resolution image is a CT image reconstructed from a portion of the learning measurement data or learning projection data obtained by dividing it. The aforementioned high-resolution image is a CT image reconstructed from the aforementioned learning projection data. The image reconstruction unit has a division unit that divides the measurement data or projection data obtained by logarithmically transforming the measurement data, and reconstructs a CT image from a portion of the divided data. The high-resolution image generation unit inputs the CT image reconstructed from the portion of the data after segmentation into the trained model, thereby obtaining a high-resolution image output by the trained model. An X-ray CT apparatus characterized by the following features.

2. An X-ray CT apparatus according to claim 1, characterized in that the low-resolution image is a CT image reconstructed from data obtained by extracting a portion of the learning measurement data or learning projection data in any of the view direction, column direction, and channel direction of the learning X-ray detector.

3. An X-ray CT apparatus according to claim 2, characterized in that the low-resolution image is a CT image reconstructed using even-numbered or odd-numbered data from the learning measurement data or learning projection data, selected in any of the view direction, column direction, or channel direction.

4. An X-ray CT apparatus according to claim 1, characterized in that the high-resolution image is an image obtained by synthesizing a plurality of CT images obtained by reconstructing the divided data obtained by dividing the learning measurement data or learning projection data into a plurality of parts.

5. An X-ray CT apparatus according to claim 1, characterized in that the measurement data acquired by the scan gantry unit is used as the learning measurement data.

6. An X-ray CT apparatus according to claim 1, wherein the image reconstruction unit selects some of the measurement data or projection data in the view direction, column direction, or channel direction, and generates a CT image reconstructed with respect to the selected portion of the data.

7. An X-ray CT apparatus according to claim 1, characterized in that the measurement data or projection data of the subject is divided under the same conditions as when the training measurement data or training projection data is divided when generating the low-resolution images for training the trained model.

8. An X-ray CT apparatus according to claim 1, wherein the division unit divides the measurement data or projection data of the subject so that the amount of noise contained in the low-resolution image used for training the trained model matches the amount of noise contained in the CT image reconstructed by the image reconstruction unit.

9. An X-ray CT apparatus according to claim 8, wherein the division unit includes a prediction unit that predicts the amount of noise included in the CT image reconstructed by the image reconstruction unit, and the X-ray CT apparatus is characterized in that the number of divisions of the measurement data or projection data of the subject is determined based on the amount of noise predicted by the prediction unit.

10. A scan gantry unit that, with an X-ray source and an X-ray detector positioned opposite each other with the subject in between, rotates the X-ray source and the X-ray detector around the subject to acquire measurement data that is detected and output by the X-ray detector as X-rays irradiated from the X-ray source and transmitted through the subject; an image reconstruction unit that generates a CT image of the subject using the measurement data; and a high-resolution image generation unit. The high-resolution image generation unit includes a trained model, and receives the CT image reconstructed by the image reconstruction unit and inputs it to the trained model to obtain a high-resolution image output by the trained model. The aforementioned trained model is a model that has been trained using at least one pre-generated low-resolution image as input data and at least one high-resolution image, which is of higher resolution than the low-resolution image, as training data. The low-resolution image and the high-resolution image are obtained based on the same training measurement data, or training projection data obtained by logarithmically transforming the training measurement data. The training measurement data is data detected and output by the training X-ray detector when the training X-ray source and the training X-ray detector are rotated around the training subject with the training subject in between, while the training X-ray source and the training X-ray detector are positioned opposite each other with the training subject in between. The aforementioned low-resolution image is a CT image reconstructed from a portion of the learning measurement data or learning projection data obtained by dividing it. The aforementioned high-resolution image is a CT image reconstructed from the aforementioned learning projection data. The aforementioned trained model has been trained using the aforementioned low-resolution images as input data, in addition to the training noise images. The aforementioned training noise image is a difference image of two training CT images obtained by reconstructing the training measurement data or the training projection data divided into two parts. The image reconstruction unit includes a noise image generation unit that generates a noise image from the measurement data or projection data. The noise image generation unit reconstructs the measurement data, or the projection data obtained by logarithmically transforming the measurement data, and divides it into two parts. It then generates the difference image of the two resulting CT images as the noise image. The X-ray CT apparatus is characterized in that the high-resolution image generation unit receives the CT image reconstructed by the image reconstruction unit and the noise image, and inputs these to the trained model to obtain a high-resolution image output by the trained model.

11. The device comprises an image reconstruction unit that receives measurement data output by the X-ray detector of an X-ray CT apparatus and generates a CT image of the subject, and a high-resolution image generation unit, The high-resolution image generation unit includes a trained model, and receives the CT image reconstructed by the image reconstruction unit and inputs it to the trained model to obtain a high-resolution image output by the trained model. The aforementioned trained model is a model that has been trained using at least one pre-generated low-resolution image as input data and at least one high-resolution image, which is of higher resolution than the low-resolution image, as training data. The low-resolution image and the high-resolution image are obtained based on the same training measurement data, or training projection data obtained by logarithmically transforming the training measurement data. The training measurement data is data detected and output by the training X-ray detector when the training X-ray source and the training X-ray detector are rotated around the training subject with the training subject in between, while the training X-ray source and the training X-ray detector are positioned opposite each other with the training subject in between. The aforementioned low-resolution image is a CT image reconstructed from a portion of the learning measurement data or learning projection data obtained by dividing it. The aforementioned high-resolution image is a CT image reconstructed from the aforementioned learning projection data. The image reconstruction unit has a division unit that divides the measurement data received from the X-ray CT apparatus, or the projection data obtained by logarithmically transforming the measurement data, and reconstructs a CT image from a portion of the divided data. The high-resolution image generation unit inputs the CT image reconstructed from the portion of the data after segmentation into the trained model, thereby obtaining a high-resolution image output by the trained model. An image processing apparatus characterized by the following:

12. A step of generating a trained model by training the model using at least one pre-generated low-resolution image as input data and at least one high-resolution image which is of higher resolution than the low-resolution image as training data, Here, the low-resolution image and the high-resolution image are obtained based on the same learning measurement data, or learning projection data obtained by logarithmically transforming the learning measurement data. The learning measurement data is data detected and output by the learning X-ray detector when the learning X-ray source and the learning X-ray detector are rotated around the learning subject with the learning subject in the opposite position, with the learning X-ray source and the learning X-ray detector positioned opposite each other with the learning subject in between. The low-resolution image is a CT image reconstructed from a portion of the learning measurement data or learning projection data, and the high-resolution image is a CT image reconstructed from the learning projection data. The process involves receiving measurement data output from the X-ray detector of an X-ray CT scanner, and dividing the measurement data, or projection data obtained by logarithmically transforming the measurement data, The steps include: reconstructing a CT image from a portion of the data after segmentation, An image processing method comprising the step of inputting a CT image reconstructed from a portion of the data after segmentation into the trained model to obtain a high-resolution image output by the trained model.

Images