Learning apparatus, method and program, and image processing apparatus, method and program

Abstract

In a learning device, method, and program, and an image processing device, method, and program, artifacts caused by high-absorbent regions in tomographic images are accurately reduced. [Solution] The processor acquires training data including training tomography images containing high absorbers and artifacts caused by high absorbers, and ground truth tomography images that do not contain high absorbers and artifacts caused by high absorbers. By normalizing at least one of the sharpness, contrast, and noise of the training tomography images and ground truth tomography images, the processor derives normalized training tomography images and normalized ground truth tomography images. Using machine learning with the normalized training tomography images and normalized ground truth tomography images, when a target tomography image containing high absorbers and artifacts caused by high absorbers is input, the processor constructs a derivation model that derives a degraded tomography image from which the high absorbers and artifacts contained in the target tomography image have been removed.

Description

Technical Field

[0001] The present disclosure relates to a learning device, method, and program, as well as an image processing device, method, and program.

Background Art

[0002] In a CT (Computed Tomography) device, when an object with a high X-ray absorption rate such as metal is included inside a subject, artifacts caused by the high absorber occur in the reconstructed image. Such artifacts hinder clinical diagnosis. For this reason, various techniques for removing artifacts have been proposed. For example, in Patent Document 1, in addition to data of a pair of a cross-sectional image of a workpiece including artifacts caused by metal in the subject and a photograph of the actual workpiece, a removal model constructed using data not including artifacts is used to remove the artifacts to be predicted.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] In the method described in Patent Document 1, in order to construct a model for removing artifacts with high accuracy, a large number of teacher data with various variations are required. However, when the number of teacher data increases, the accuracy of removing artifacts of the constructed derivation model decreases, the need to generate a large number of teacher data arises, or the learning time becomes extremely long. On the other hand, when the number of teacher data is reduced, the imaging conditions and reconstruction conditions of tomographic images for which artifacts can be removed are limited, so it is impossible to handle tomographic images acquired under various conditions.

[0005] This disclosure is made in view of the above circumstances and aims to accurately reduce artifacts caused by high-absorption regions in tomographic images. [Means for solving the problem]

[0006] The learning device described herein comprises a processor, The processor acquires training data, including training tomography images containing high-absorte and artifacts caused by high-absorte, as well as ground truth tomography images that do not contain high-absorte and artifacts caused by high-absorte. By normalizing at least one of the sharpness, contrast, and noise of the training tomography and ground truth tomography, normalized training tomography and normalized ground truth tomography are derived. Using normalized tomographic images for training and normalized ground truth tomographic images, a derivation model is constructed that, when a target tomographic image containing high-absorting elements and artifacts caused by these elements is input, derives a decomposed tomographic image from which the high-absorting elements and artifacts contained in the target tomographic image have been removed.

[0007] The image processing apparatus according to this disclosure comprises a processor, The processor acquires projection images, including high-attenuation materials and artifacts caused by high-attenuation materials, obtained by scanning a subject containing high-attenuation materials with a CT scanner. A provisional tomographic image is derived by reconstructing the projection image, and a normalized projection image is derived by normalizing at least one of the sharpness, contrast, and noise of the provisional tomographic image, or by normalizing at least one of the sharpness, contrast, and noise of the projection image, and a normalized projection image is derived by reconstructing the normalized projection image. Using the derivation model constructed by the learning device described herein, a de-absorbed tomographic image is derived from a normalized tomographic image from which high-absorbent materials and artifacts have been removed. A removed projection image is derived by denormalizing at least one of the sharpness, contrast, and noise of the removed tomography image and projecting the denormalized removed tomography image forward, or by projecting the removed tomography image forward and denormalizing at least one of the sharpness, contrast, and noise of the projected removed tomography image. A corrected projection image is derived by replacing the areas of high absorbers and artifacts in the projected image with images of the areas corresponding to the high absorbers and artifacts in the desensitized projection image.

[0008] In the image processing apparatus according to this disclosure, the processor may reconstruct a corrected projection image to derive a corrected tomographic image.

[0009] The learning method described herein involves a computer acquiring training data, which includes training tomographic images containing high-absorting materials and artifacts caused by high-absorting materials, as well as ground truth tomographic images that do not contain high-absorting materials and artifacts caused by high-absorting materials. By normalizing at least one of the sharpness, contrast, and noise of the training tomography and ground truth tomography, normalized training tomography and normalized ground truth tomography are derived. Using normalized tomographic images for training and normalized ground truth tomographic images, a derivation model is constructed that, when a target tomographic image containing high-absorting elements and artifacts caused by these elements is input, derives a decomposed tomographic image from which the high-absorting elements and artifacts contained in the target tomographic image have been removed.

[0010] The image processing method disclosed herein involves a computer acquiring a projection image, including the high-attenuation material and artifacts caused by the high-attenuation material, obtained by scanning a subject containing a high-attenuation material with a CT scanner. A provisional tomographic image is derived by reconstructing the projection image, and a normalized projection image is derived by normalizing at least one of the sharpness, contrast, and noise of the provisional tomographic image, or by normalizing at least one of the sharpness, contrast, and noise of the projection image, and a normalized projection image is derived by reconstructing the normalized projection image. Using the derivation model constructed by the learning device described herein, a de-absorbed tomographic image is derived from a normalized tomographic image from which high-absorbent materials and artifacts have been removed. A removed projection image is derived by denormalizing at least one of the sharpness, contrast, and noise of the removed tomography image and projecting the denormalized removed tomography image forward, or by projecting the removed tomography image forward and denormalizing at least one of the sharpness, contrast, and noise of the projected removed tomography image. A corrected projection image is derived by replacing the areas of high absorbers and artifacts in the projected image with images of the areas corresponding to the high absorbers and artifacts in the desensitized projection image.

[0011] The learning program described herein includes a procedure for obtaining training data, which includes training tomographic images containing high-absorting materials and artifacts caused by high-absorting materials, and ground truth tomographic images that do not contain high-absorting materials and artifacts caused by high-absorting materials. A procedure for deriving normalized training tomography images and normalized ground truth images by normalizing at least one of the sharpness, contrast, and noise of the training tomography images and ground truth images, Using normalized tomographic images for training and normalized ground truth tomographic images, machine learning is used to construct a derivation model that, when a target tomographic image containing high-absorting elements and artifacts caused by these elements is input, derives a derivation tomographic image from which the high-absorting elements and artifacts contained in the target tomographic image have been removed.

[0012] The image processing program disclosed herein includes a procedure for acquiring a projection image containing high-attenuation materials and artifacts caused by high-attenuation materials, obtained by imaging a subject containing high-attenuation materials with a CT scanner, and A procedure for deriving a normalized tomographic image by reconstructing a projection image to derive a provisional tomographic image, normalizing at least one of the sharpness, contrast, and noise of the provisional tomographic image, or by normalizing at least one of the sharpness, contrast, and noise of the projection image to derive a normalized projection image, and reconstructing the normalized projection image, A procedure for deriving a degraded tomographic image from a normalized tomographic image, in which high-absorte and artifacts have been removed, using a derivation model constructed by the learning device described herein, A procedure for deriving a removed projection image by inversely normalizing at least one of the sharpness, contrast, and noise of the removed tomography image and projecting the inversely normalized removed tomography image forward, or by projecting the removed tomography image forward and inversely normalizing at least one of the sharpness, contrast, and noise of the projected removed tomography image, The computer is instructed to perform a procedure to derive a corrected projection image by replacing areas of high absorbers and artifacts in the projected image with images of the corresponding areas of high absorbers and artifacts in the desensitized projection image.

[0013] Furthermore, the technology disclosed herein may be applied to program products. [Effects of the Invention]

[0014] According to this disclosure, artifacts caused by high-absorption regions in tomographic images can be accurately reduced. [Brief explanation of the drawing]

[0015] [Figure 1] Schematic diagram of a medical image acquisition system equipped with a learning device and image processing device according to this embodiment. [Figure 2] This figure shows the hardware configuration of the learning device and image processing device according to this embodiment. [Figure 3] Functional configuration diagram of the learning device according to this embodiment [Figure 4] Diagram showing an example of teacher data [Figure 5] Diagram for explaining the construction of the derived model [Figure 6] Functional configuration diagram of the image processing device according to this embodiment [Figure 7] Diagram showing raw data obtained by photographing the head of a subject containing metal with a CT device [Figure 8] Diagram showing the processing flow performed by the specific part, correction part, and reconstruction part in this embodiment [Figure 9] Diagram for explaining the specification of the metal region [Figure 10] Diagram for explaining the forward projection of the metal region [[ID=z4]] [Figure 11] [[ID=2z]]Diagram for explaining the correction of the metal region [Figure 12] Flowchart showing the learning process performed by the learning device in this embodiment [Figure 13] Flowchart showing the image processing performed by the image processing device in this embodiment [Figure 14] Flowchart showing the processing performed by the correction part

Embodiments for Carrying Out the Invention

[0016] Hereinafter, embodiments of the present disclosure will be described in detail with reference to the drawings. First, an example of the configuration of a medical image imaging system including a learning device and an image processing device according to an embodiment of the present disclosure will be described. FIG. 1 is a schematic configuration diagram of a medical image imaging system including a learning device and an image processing device according to this embodiment.

[0017] As shown in FIG. 1, the medical image imaging system 1 of this embodiment includes a CT device 2 and a console 3. The CT device 2 includes a gantry 4 and a couch 8. In the following description, the horizontal direction in FIG. 1 is the X-axis, the vertical direction is the Y-axis, and the direction perpendicular to the XY plane is the Z-axis. The CT device 2 is an example of a radiation imaging device.

[0018] The gantry 4 has an opening 4A, and the subject H to be photographed is placed inside the opening 4A while on the bed 8. The gantry 4 and the bed 8 are movable relative to each other in the Z-axis direction.

[0019] Inside the gantry 4, a radiation source 5, which has a radiation tube 6 and a bowtie filter 7, and a detector 9 are arranged facing each other with the subject H in between. The bowtie filter 7 optimizes the radiation dose by increasing the dose near the center and decreasing the dose around the periphery in order to reduce the dose in the peripheral area. The radiation emitted from the radiation tube 6 is shaped into a beam shape suitable for the size of the subject H by the bowtie filter 7 and irradiated onto the subject H.

[0020] The detector 9 detects radiation transmitted through the subject H and generates projection data corresponding to the detected radiation dose. In the detector 9, multiple detection elements 9P are arranged in an arc shape centered on the focal point of the radiation tube 6. The direction in which the multiple detection elements 9P are aligned in the arc is defined as the channel direction.

[0021] In this embodiment, X-rays are used as an example of radiation, but the invention is not limited to this, and gamma rays or other types of radiation can also be used.

[0022] The radiation source 5 and detector 9 are mounted on a rotating plate 4B inside the gantry 4 and are rotated around the subject H by a rotation drive unit (not shown). As radiation irradiation from the radiation source 5 and detection of radiation by the detector 9 are repeated along with the rotation of both, raw data is acquired in multiple view units with different projection angles of radiation onto the subject H, and projection data is generated from the raw data. The generated projection data is output to the console 3. The projection data is derived by arranging the raw data with the horizontal axis representing the channel of the detector 9 and the vertical axis representing the rotation angle of the CT device 2.

[0023] The radiation dose emitted from the radiation tube 6, the rotation speed of the gantry 4, and the relative movement speed between the gantry 4 and the patient bed 8 are all set by the console 3 based on the imaging conditions entered by the operator, such as a technician.

[0024] Console 3 in this embodiment controls the imaging of subject H, generates projection data from raw data acquired by imaging, reconstructs tomographic images from projection data, and sets the storage of image data for projection data and tomographic images. Furthermore, as will be described later, Console 3 in this embodiment also performs the process of constructing a derivation model for deriving a tomographic image from which artifacts have been removed from the tomographic image. Console 3 is an example of the learning device and image processing device of this disclosure.

[0025] Next, the learning device and image processing device according to this embodiment will be described. First, with reference to Figure 2, the hardware configuration of the learning device and image processing device according to this embodiment, which are contained within the console 3, will be described. As shown in Figure 2, the learning device 10A and image processing device 10B (hereinafter sometimes referred to as the image processing device 10) contained within the console 3 are computers such as workstations, server computers, and personal computers, and are equipped with a CPU (Central Processing Unit) 11, non-volatile storage 13, and memory 16 as a temporary storage area.

[0026] The image processing device 10 also includes a display 14, an input device 15, and an I / F (Interface) 17. The CPU 11, storage 13, display 14, input device 15, memory 16, and I / F 17 are connected to a bus 18. The CPU 11 is an example of a processor in this disclosure.

[0027] The storage 13 is implemented using an HDD (Hard Disk Drive), an SSD (Solid State Drive), and flash memory, etc. The learning program 12A and the image processing program 12B installed on the image processing device 10 are stored in the storage 13 as a storage medium. The CPU 11 reads the learning program 12A and the image processing program 12B from the storage 13, expands them into memory 16, and executes the expanded learning program 12A and image processing program 12B.

[0028] The display 14 is a device that displays various types of screens, such as a liquid crystal display or an EL (Electro Luminescence) display.

[0029] The input device 15 is used by the operator to input instructions and various information regarding the shooting conditions, image generation and display, etc., when photographing the subject H. Examples of input devices 15 include various switches, buttons, touch panels, styluses, keyboards, and mice. The display 14 and the input device 15 may be integrated to form a touch panel display.

[0030] I / F17 communicates various types of information with the rotational drive unit (not shown) of the gantry 4, the radiation source 5, and the detector 9 via wired or wireless communication. I / F17 also communicates with an image storage server (not shown) that stores training data used when constructing the derivation model, as will be described later.

[0031] The learning program 12A and the image processing program 12B are stored in a memory device of a server computer connected to the network, or in network storage, in a state that allows external access, and are downloaded and installed on the computers comprising the image processing device 10 upon request. Alternatively, they are recorded on a recording medium such as a DVD (Digital Versatile Disc) or CD-ROM (Compact Disc Read Only Memory) and distributed, and then installed from that recording medium onto the computers comprising the image processing device 10.

[0032] Next, the learning device according to this embodiment will be described. Figure 3 is a diagram showing the functional configuration of the learning device according to this embodiment. As shown in Figure 3, the learning device 10A includes an information acquisition unit 21, a normalization unit 22, and a learning unit 23. The CPU 11 functions as the information acquisition unit 21, the normalization unit 22, and the learning unit 23 by executing the learning program 12A.

[0033] The information acquisition unit 21 acquires training data from the image storage server based on instructions from the input device 15. The acquired training data is stored in the storage 13. If training data is already stored in the storage 13, the information acquisition unit 21 acquires the training data from the storage 13.

[0034] In this embodiment, the derived model constructed by the learning device 10A is constructed to remove metals and artifacts caused by metals contained in the input tomographic image. Figure 4 shows an example of training data. As shown in Figure 4, the training data 30 includes a training tomographic image 31, which is a tomographic image of the head containing metals and artifacts, and a ground truth tomographic image 32, which is a tomographic image of the head that does not contain metals and artifacts. The training tomographic image 31 and the ground truth tomographic image 32 may or may not be of the same subject.

[0035] Here, if the subject H contains an object with a high radiation absorption rate, such as metal, the tomographic image obtained by reconstructing the projection image represented by the projection data acquired by imaging will contain artifacts caused by the metal. The learning tomographic image 31 is obtained by imaging subject H, which contains metal in its head, with a CT scanner 2. Metal is an example of a high absorber in this disclosure.

[0036] The ground truth tomographic image 32 is obtained by imaging subject H, whose head does not contain any metal, using the CT scanner 2. The ground truth tomographic image 32 does not contain artifacts caused by metal.

[0037] The normalization unit 22 normalizes the training data 30. In this embodiment, normalization refers to normalizing at least one of the sharpness, contrast, and noise of the training tomography image 31 and the ground truth tomography image 32 included in the training data 30. Normalization of sharpness is performed by applying frequency processing to the training tomography image 31 and the ground truth tomography image 32 to emphasize or suppress predetermined high-frequency components. For example, the modulation transfer function (MTF) of the training tomography image 31 and the ground truth tomography image 32 can be adjusted for each frequency band. Specifically, this can be done by approximately adjusting the 50% MTF to 0.5 (cycles / mm).

[0038] Contrast normalization is performed by transforming the pixel values ​​(CT values) of the training tomography image 31 and the ground truth tomography image 32 so that the pixel values ​​of the training tomography image 31 and the ground truth tomography image 32 fall within predetermined lower and upper limits. In this case, if the CT value exceeds the upper limit, it will be fixed to the upper limit. Alternatively, contrast normalization may be performed by linearly transforming the difference between the CT values ​​of soft tissues and bones and the CT values ​​of water, using the CT value of water as a reference (i.e., dynamic range compression / expansion processing).

[0039] Regarding noise normalization, since the pixel values ​​(CT values) of the training tomography image 31 and the ground truth tomography image 32 are normalized, the noise in the training tomography image 31 and the ground truth tomography image 32 is represented by the standard deviation (SD) of the CT values ​​within the region of interest. In this embodiment, noise is normalized by filtering with a noise reduction filter or by adding noise so that the SD falls within a predetermined range (for example, SD = 5 to 30 HU).

[0040] The learning unit 23 uses normalized training data 30S to machine-learn a neural network, thereby constructing a derivation model that, when a normalized tomographic image to be processed, including metals and metal-induced artifacts, is input, derives a normalized de-processed tomographic image from which the metals and artifacts contained in the tomographic image to be processed have been removed. Figure 5 is a diagram illustrating the construction of the derivation model.

[0041] Examples of machine learning models used to construct derivation models include neural network models. Examples of neural network models include simple perceptrons, multilayer perceptrons, deep neural networks, convolutional neural networks, deep belief networks, recurrent neural networks, and stochastic neural networks.

[0042] The learning unit 23 inputs the normalized training tomography 31S contained in the normalized training data 30S into the machine learning model 35 and outputs a de-metallic training tomography 36S from which metals and artifacts caused by metals in the normalized training tomography 31S have been removed. The learning unit 23 derives the difference between the normalized ground truth tomography 32S contained in the training data 30S and the de-metallic training tomography 36S as the loss L. The learning unit 23 trains the machine learning model 35 based on the loss L. For example, if the machine learning model 35 is a convolutional neural network, it derives the kernel coefficients and connection weights of the neural network in the convolutional neural network in order to minimize the loss L1.

[0043] The learning unit 23 repeatedly trains the machine learning model 35 using multiple normalized training data 30S until the loss L falls below a predetermined threshold. Alternatively, the learning unit 23 repeats the training a predetermined number of times. As a result, when a normalized tomographic image to be processed, including metal and metal-induced artifacts, is input, a derivation model 38 is constructed that derives a normalized decomposed tomographic image from which the metal regions and artifacts contained in the tomographic image to be processed have been removed. The constructed derivation model 38 is stored in the storage 13.

[0044] Next, an image processing apparatus according to this embodiment will be described. Figure 6 is a diagram showing the functional configuration of the image processing apparatus according to this embodiment. As shown in Figure 6, the image processing apparatus 10B includes an image capture control unit 41, an information acquisition unit 42, a specific unit 43, a correction unit 44, and a reconstruction unit 45. The CPU 11 functions as the image capture control unit 41, the information acquisition unit 42, the specific unit 43, the correction unit 44, and the reconstruction unit 45 by executing the image processing program 12B.

[0045] The imaging control unit 41 controls each part of the CT scanner 2 to perform imaging of the subject H based on instructions from the input device 15. In this embodiment, the head of the subject H is to be imaged. For illustrative purposes, the head is assumed to contain metal. The metal is an example of a highly absorbent material in this disclosure.

[0046] The information acquisition unit 42 acquires projection data obtained by imaging the subject H from the CT device 2. The image represented by the projection data is the projection image.

[0047] Figure 7 shows the raw data obtained by imaging the head of subject H containing metal using the CT scanner 2. In Figure 7, the raw data 50 is shown when radiation is irradiated to the head 51 in the direction of arrow A. In the raw data 50 shown in Figure 7, the horizontal axis represents the channel direction of the detector 9, and the vertical axis represents the data value. In the raw data 50, the data value is small in the channel that detected radiation that did not pass through the head 51 (i.e., the detection element 9P), the data value is large in the channel that detected radiation that passed through the head 51, and there is a peak in the data value in the channel that detected radiation that passed through the metal 52 inside the head 51. When such raw data is arranged with the horizontal axis representing the channel direction of the detector 9 and the vertical axis representing the rotation angle, it becomes projection data.

[0048] Figure 8 is a diagram showing the processing flow performed by the identification unit 43, the correction unit 44, and the reconstruction unit 45 in this embodiment. As shown in Figure 8, first, the metallic region A0 is extracted from the projection image P0 represented by projection data. Next, the metallic region A0 in the projection image P0 is corrected to derive a corrected projection image P1. Furthermore, the corrected projection image P1 is reconstructed to derive a corrected tomographic image D1. The individual processes performed by the identification unit 43, the correction unit 44, and the reconstruction unit 45 will be described below.

[0049] The identification unit 43 identifies metallic regions in the projected image. Figure 9 is a diagram illustrating the identification of metallic regions. The identification unit 43 first derives a provisional tomographic image D0 by reconstructing the projected image P0 using the reconstruction unit 45. Here, the provisional tomographic image D0 includes metallic regions and artifacts due to the influence of metal. The identification unit 43 removes artifacts from the provisional tomographic image D0. For example, the identification unit 43 removes artifacts from the provisional tomographic image D0 using a removal model constructed to remove artifacts from a tomographic image. Note that the removal model used by the identification unit 43 differs from the derivation model 38 constructed by the learning device 10A according to this embodiment, and is constructed to remove only artifacts from the provisional tomographic image D0. Also, the input provisional tomographic image D0 does not need to be normalized.

[0050] The identification unit 43 identifies the metallic region A1 in the provisional tomographic image from which artifacts have been removed. Since the metallic region A1 is a high-brightness region in the provisional tomographic image from which artifacts have been removed, the identification unit 43 extracts the metallic region A1 using an extraction model constructed to extract such high-brightness regions. Then, as shown in Figure 10, the identification unit 43 identifies the metallic region A0 in the projected image P0 by projecting the metallic region A1 in the direction of arrow B.

[0051] The correction unit 44 derives a corrected projected image P1 by performing corrections on the projected image P0. Figure 11 is a diagram illustrating the correction of metallic regions. As shown in Figure 11, the correction unit 44 has a normalization unit 61 and an inverse normalization unit 62. The correction unit 44 derives a provisional tomographic image D0 by reconstructing the projected image P0 with the reconstruction unit 45. The provisional tomographic image D0 may be one derived by the identification unit 43. The provisional tomographic image D0 includes metal and artifacts due to the influence of metal. The normalization unit 61 of the correction unit 44 normalizes the provisional tomographic image D0. At this time, the normalization unit 61 normalizes at least one of the sharpness, contrast, and noise of the provisional tomographic image D0, similar to the normalization performed by the normalization unit 22 of the learning device 10A described above. This derives a normalized tomographic image Ds0.

[0052] Furthermore, if the tube voltage when subject H is imaged differs from the tube voltage when the training tomography image 31 is acquired, the contrast will differ between the provisional tomography image D0 and the training tomography image 31, even for the same part of the same subject. Therefore, in the case of dynamic range compression / expansion, when the normalization process performed by the normalization unit 61 involves compressing / expanding the contrast, if the tube voltage when subject H is imaged differs from the tube voltage when the training tomography image 31 is acquired, the normalization is performed after the pixel values ​​of the provisional tomography image D0 are multiplied by a coefficient according to the difference in tube voltage to match the pixel values ​​corresponding to the tube voltage during training.

[0053] The normalization unit 61 may also normalize the projection image P0 instead of the temporary tomographic image D0. In this case, the correction unit 44 derives the normalized tomographic image Ds0 by reconstructing the normalized projection image P0 using the reconstruction unit 45.

[0054] The correction unit 44 inputs the normalized tomographic image Ds0 into the derivation model 38 to derive a decompressed tomographic image Ds2 from which metals and metal-related artifacts contained in the normalized tomographic image Ds0 have been removed.

[0055] In this embodiment, since the removed tomography image Ds2 is derived from the normalized tomography image Ds0, at least one of the sharpness, contrast, and noise is normalized. Therefore, in this embodiment, the inverse normalization unit 62 of the correction unit 44 derives the inverse normalized removed tomography image D2 by inverse normalizing the removed tomography image Ds2. Inverse normalization is a process that matches at least one of the sharpness, contrast, and noise of the removed tomography image Ds2 with at least one of the sharpness, contrast, and noise of the projection image P0 or the temporary tomography image D0.

[0056] For example, inverse normalization of sharpness can be performed by applying frequency processing to the removed tomography image Ds2 to suppress or enhance predetermined high-frequency components, thereby matching the spatial frequency of the removed tomography image Ds2 with the spatial frequency of the projected image P0 or the temporary tomography image D0. Inverse normalization of contrast can be performed by transforming the pixel values ​​(CT values) of the removed tomography image Ds2 so that they fall within the lower and upper limits of the projected image P0 or the temporary tomography image D0. Alternatively, the inverse normalization process can be performed by applying dynamic range compression / expansion processing to the removed tomography image Ds2. Inverse normalization of noise can be performed by filtering with a noise reduction filter or by adding noise so that the standard deviation (SD) of the removed tomography image Ds2 is within the same range as the SD of the projected image P0 or the temporary tomography image D0. Through this inverse normalization process, the inverse normalization unit 62 derives an inversely normalized tomographic image D2 from the deremoved tomographic image Ds2.

[0057] The correction unit 44 derives a de-removed projection image P2 by projecting the inverse normalized de-removed tomographic image D2 forward. Furthermore, the correction unit 44 corrects the metal region A0 in the projection image P0 by replacing it with an image of the region corresponding to the metal region A0 in the de-removed projection image P2, thereby deriving a corrected projection image P1.

[0058] The reconstruction unit 45 derives a corrected tomographic image D1 by reconstructing the corrected projection image P1 at multiple projection angles (see Figure 8).

[0059] Alternatively, instead of inversely normalizing the removed tomographic image Ds2, the removed projection image P2 may be derived by forward-projecting the removed tomographic image Ds2 and then inversely normalizing the forward-projected removed tomographic image Ds2.

[0060] Next, the processing performed in this embodiment will be described. Figure 12 is a flowchart showing the learning process performed by the learning device 10A in this embodiment. It is assumed that the training data 30 is obtained from the image storage server and stored in the storage 13. First, the information acquisition unit 21 acquires the training data 30 stored in the storage 13 (step ST1). Next, the normalization unit 22 normalizes the training tomography image 31 and the ground truth tomography image 32 included in the training data 30 to derive the normalized training tomography image 31S and the normalized ground truth tomography image 32S (step ST2).

[0061] The learning unit 23 inputs the normalized training tomographic image 31S to the machine learning model 35 and outputs a degraded training tomographic image 36S from which metallic regions and artifacts caused by metal in the normalized training tomographic image 31S have been removed, and derives the loss L with respect to the normalized ground truth tomographic image 32S (step ST3). Then, the learning unit 23 trains the machine learning model 35 so that the loss L is less than or equal to a predetermined threshold (step ST4).

[0062] The learning unit 23 returns to the process of step ST1, retrieves the next training data 30S from storage 13, and repeats the process of steps ST1 to ST4. This constructs the derived model 38.

[0063] Next, the image processing performed in this embodiment will be described. Figure 13 is a flowchart showing the image processing performed by the image processing device in this embodiment. First, the imaging control unit 41 takes an image of the subject H in the CT device 2 according to the operator's instructions (step ST11), and the information acquisition unit 42 acquires projection data (step ST12). The identification unit 43 identifies the metal region A0 in the projection image P0 represented by the projection data (step ST13). The correction unit 44 corrects the projection image P0 to derive a corrected projection image P1 (step ST14).

[0064] Figure 14 is a flowchart showing the processing performed by the correction unit 44. The correction unit 44 derives a provisional tomographic image D0 by reconstructing the projected image P0 with the reconstruction unit 45 (step ST21). The normalization unit 61 of the correction unit 44 derives a normalized tomographic image Ds0 by normalizing the provisional tomographic image D0 (step ST22). The correction unit 44 derives a decompressed tomographic image Ds2 from which metallic regions and artifacts caused by metallic regions included in the normalized tomographic image Ds0 have been removed by inputting the normalized tomographic image Ds0 into the decomposition model 38 (step ST23).

[0065] The inverse normalization unit 62 of the correction unit 44 derives an inverse normalized removed tomographic image D2 by inverse normalizing the removed tomographic image Ds2 (step ST24). The correction unit 44 derives a removed projected image P2 by forward projecting the inverse normalized removed tomographic image D2 (step ST25). Furthermore, the correction unit 44 corrects the metal region A0 in the projected image P0 by replacing it with an image of the region corresponding to the metal region A0 in the removed projected image P2, thereby deriving a corrected projected image P1 (replacement; step ST26).

[0066] Returning to Figure 13, the reconstruction unit 45 reconstructs the corrected projection image P1 to derive the corrected tomographic image D1 (step ST15), and the process ends.

[0067] As described above, in the learning device 10A according to this embodiment, the training tomographic image 31 and the ground truth tomographic image 32 are normalized, and a derivation model 38 is constructed by machine learning using the normalized training tomographic image 31S and the normalized ground truth tomographic image 32S. When a target tomographic image containing metal and artifacts caused by metal is input, the derivation model 38 is constructed that derives a de-metallic tomographic image from which the metal and artifacts contained in the target tomographic image have been removed. As a result, the variation in training data can be reduced, which prevents a decrease in the accuracy of artifact removal in the derivation model and shortens the learning time.

[0068] Furthermore, in the image processing apparatus 10B according to this embodiment, a provisional tomographic image derived by reconstructing the projected image is normalized, a de-masked tomographic image is derived from the normalized provisional tomographic image using the derivation model 38, the de-masked tomographic image is de-normalized, and the de-normalized de-masked tomographic image is forward-projected to derive a de-projected image. In addition, a corrected projected image is derived by replacing the metallic regions in the projected image with images of the corresponding regions in the de-masked projected image, and a corrected tomographic image is derived by reconstructing the corrected projected image. As a result, a corrected tomographic image from which metallic and metallic artifacts have been removed can be derived. Also, since the de-masked tomographic image is de-normalized and the de-normalized de-masked tomographic image is forward-projected to derive a de-projected image, at least one of the sharpness, contrast, and noise can be matched between the de-projected image and the projected image. Therefore, when metallic regions in the projected image are replaced with images of the corresponding regions in the de-masked projected image, there will be no difference in at least one of the sharpness, contrast, and noise between the replaced region and the region other than the replaced region in the projected image. Therefore, high-resolution corrected tomographic images can be obtained.

[0069] In the above embodiment, the learning device 10A is applied to the console 3 to construct the derived model 38, but this is not the only way. The console 3 may also construct the derived model 38 by applying the learning device 10A to another computer or the like. In this case, the constructed derived model 38 is transmitted to the console 3 for storage and used for processing metal and artifact correction.

[0070] In this embodiment, each process is executed on any computer. Furthermore, any computer may execute these processes using a processor as hardware, a program as software, or a combination thereof. In that case, the processor is configured to work in cooperation with the program to execute the various processes in this embodiment, and can function as a unit or means in this embodiment. Also, the execution order of the processes by the processor is not limited to the order described and may be changed as appropriate. Any computer may be a general-purpose computer, a computer designed for a specific purpose, a workstation, or any other system capable of executing each process.

[0071] A processor may consist of one or more hardware components, and the type of hardware is not limited. For example, a processor may consist of a CPU (Central Processing Unit), an MPU (Micro Processing Unit), a programmable logic device such as an FPGA (Field Programmable Gate Array), a dedicated circuit for executing a specific process such as an ASIC (Application Specific Integrated Circuit), a GPU (Graphic Processing Unit), or an NPU (Neural Processing Unit). Furthermore, the type of hardware may be a combination of different types of hardware. When multiple hardware components are configured to execute one or more processes of a given processor, these components may reside in physically separate devices or in the same device. Also, in any embodiment, the order of each process performed by the processor is not limited to the order described above and may be changed as appropriate. Hardware is composed of electrical circuits (circuitry) that combine circuit elements such as semiconductor elements.

[0072] Furthermore, the program may be firmware or software such as microcode. Alternatively, the program may be, for example, a set of program modules, each function of which may be implemented by a processor configured to perform its respective function. The program may be program code or multiple code segments stored on one or more non-temporary computer-readable media (e.g., storage media or other storage). The program may be divided and stored on multiple non-temporary computer-readable media located in physically separate devices. Program code or code segments may represent any combination of procedures, functions, subprograms, routines, subroutines, modules, software packages, classes, or instructions, data structures, or program statements. Program code or code segments may be connected to other code segments or hardware circuits by sending and receiving information, data, arguments, parameters, or memory contents.

[0073] Furthermore, although the above embodiment describes a configuration in which the learning program 12A and the image processing program 12B are pre-stored (installed) in the storage 13, the invention is not limited to this configuration. The learning program 12A and the image processing program 12B may be provided in the form of recording media such as CD-ROM (Compact Disc Read Only Memory), DVD-ROM (Digital Versatile Disc Read Only Memory), and USB (Universal Serial Bus) memory. Alternatively, the learning program 12A and the image processing program 12B may be provided in the form of download from an external device via a network.

[0074] The technology disclosed herein extends to all program products. Program products include all forms of products for providing programs. For example, program products include programs provided via networks such as the Internet, and non-temporary computer-readable recording media such as CD-ROMs, DVDs, and USB memory sticks on which programs are stored. [Explanation of Symbols]

[0075] 1. Medical imaging system 2 CT device 3 Console 4 Gantry 4A opening 4B Rotating Plate 5 Radiation source 6 Radiation tubes 7 Bowtie Filter 8 berths 9 Detectors 9P detection element 10A Learning Device 10B Image Processing Device 11 CPU 12A Learning Program 12B Image Processing Program 13 Storage 14 displays 15 Input Devices 16 memory 17 I / F 18 bus 21 Information Acquisition Department 22 Normalization section 23 Learning Department 30 Training data 31. Tomographic images for learning 31S Normalized tomographic images for training 32 Ground truth tomographic images 32S normalized true tomographic images 35 Machine Learning Models 36S removed tomographic image for learning 38 Derivation Model 41. Image capture control unit 42 Information Acquisition Department 43 Specific part 44 Correction section 45 Reconstruction part 50 raw data 51 Head 52 metal 61 Normalization section 62 Denormalization part A0,A1 Metal area D0 Provisional fault image D1 Corrected Tomographic Image Ds0 normalized hypothetical tomographic image D2 Inverse Normalized Depleted Tomographic Image Ds2 Removal Tomographic Image P0 Projection Image P1 Corrected projection image P2 removed projection image

Claims

1. Equipped with a processor, The aforementioned processor, Training data is obtained that includes training tomographic images containing high-absorting elements and artifacts caused by the high-absorting elements, and ground truth tomographic images that do not contain the high-absorting elements and artifacts caused by the high-absorting elements. By normalizing at least one of the sharpness, contrast, and noise of the training tomography and the ground truth tomography, a normalized training tomography and a normalized ground truth tomography are derived. A learning device that, through machine learning using the normalized tomographic images for learning and the normalized ground truth tomographic images, constructs a derivation model that derives a degraded tomographic image from which the high-absorbent material and artifacts caused by the high-absorbent material have been removed, when a target tomographic image containing the high-absorbent material and artifacts caused by the high-absorbent material is input.

2. Equipped with a processor, The aforementioned processor, A projection image including the high-attenuation material and artifacts caused by the high-attenuation material is obtained by imaging a subject containing a high-attenuation material using a CT scanner. A provisional tomographic image is derived by reconstructing the aforementioned projection image, and a normalized projection image is derived by normalizing at least one of the sharpness, contrast, and noise of the provisional tomographic image, or by normalizing at least one of the sharpness, contrast, and noise of the aforementioned projection image, and a normalized projection image is derived by reconstructing the aforementioned normalized projection image. Using the derivation model constructed by the learning device described in claim 1, a de-absorbed tomographic image is derived from the normalized tomographic image, from which the high-absorbent material and the artifacts have been removed. A projected removal image is derived by inversely normalizing at least one of the sharpness, contrast, and noise of the removed tomography image and projecting the inversely normalized removed tomography image forward, or by projecting the removed tomography image forward and inversely normalizing at least one of the sharpness, contrast, and noise of the projected removed tomography image. An image processing apparatus for deriving a corrected projection image by replacing the regions of the high absorber and the artifact in the projection image with images of the regions corresponding to the high absorber and the artifact in the degraded projection image.

3. The image processing apparatus according to claim 2, wherein the processor reconstructs the corrected projection image to derive a corrected tomographic image.

4. The computer acquires training data, which includes training tomographic images containing high-absorting elements and artifacts caused by the high-absorting elements, and ground truth tomographic images that do not contain the high-absorting elements and artifacts caused by the high-absorting elements. By normalizing at least one of the sharpness, contrast, and noise of the training tomography and the ground truth tomography, a normalized training tomography and a normalized ground truth tomography are derived. A learning method that constructs a derivation model by machine learning using the normalized tomographic image for learning and the normalized ground truth tomographic image, which, when a target tomographic image containing the high-absorbent material and artifacts caused by the high-absorbent material is input, derives a decomposed tomographic image from which the high-absorbent material and artifacts contained in the target tomographic image have been removed.

5. A computer acquires a projection image, including the high-attenuation material and artifacts caused by the high-attenuation material, obtained by scanning a subject containing the high-attenuation material with a CT scanner. A provisional tomographic image is derived by reconstructing the aforementioned projection image, and a normalized projection image is derived by normalizing at least one of the sharpness, contrast, and noise of the provisional tomographic image, or by normalizing at least one of the sharpness, contrast, and noise of the aforementioned projection image, and a normalized projection image is derived by reconstructing the aforementioned normalized projection image. Using the derivation model constructed by the learning device described in claim 1, a de-absorbed tomographic image is derived from the normalized tomographic image, from which the high-absorbent material and the artifacts have been removed. A projected removal image is derived by inversely normalizing at least one of the sharpness, contrast, and noise of the removed tomography image and projecting the inversely normalized removed tomography image forward, or by projecting the removed tomography image forward and inversely normalizing at least one of the sharpness, contrast, and noise of the projected removed tomography image. An image processing method for deriving a corrected projection image by replacing the regions of the high absorber and the artifact in the projection image with images of the regions corresponding to the high absorber and the artifact in the degraded projection image.

6. A procedure for obtaining training data including training tomographic images containing high-absorting elements and artifacts caused by the high-absorting elements, and ground truth tomographic images that do not contain the high-absorting elements and artifacts caused by the high-absorting elements, A procedure for deriving a normalized training tomography image and a normalized ground truth tomography image by normalizing at least one of the sharpness, contrast, and noise of the training tomography image and the ground truth tomography image, A learning program that, using machine learning with the normalized tomographic images and the normalized ground truth tomographic images, causes a computer to execute a procedure to construct a derivation model that derives a degraded tomographic image from which the high-absorbent material and artifacts caused by the high-absorbent material have been removed, when a target tomographic image containing the high-absorbent material and artifacts caused by the high-absorbent material is input.

7. A procedure for obtaining a projection image including the high-attenuation material and artifacts caused by the high-attenuation material, obtained by imaging a subject containing a high-attenuation material using a CT scanner, A procedure for deriving a normalized tomographic image by reconstructing the aforementioned projection image, normalizing at least one of the sharpness, contrast, and noise of the aforementioned projection image, or by normalizing at least one of the sharpness, contrast, and noise of the aforementioned projection image to derive a normalized projection image, and reconstructing the normalized projection image, A procedure for deriving a degraded tomographic image from which the high-absorption elements and artifacts have been removed from the normalized tomographic image using a derivation model constructed by the learning device described in claim 1, A procedure for deriving a projected removal image by inversely normalizing at least one of the sharpness, contrast, and noise of the removed tomography image and projecting the inversely normalized removed tomography image forward, or by projecting the removed tomography image forward and inversely normalizing at least one of the sharpness, contrast, and noise of the projected removed tomography image, An image processing program that causes a computer to perform the following steps: replace the regions of the high-absorbent material and the artifact in the projected image with images of the regions corresponding to the high-absorbent material and the artifact in the degraded projected image to derive a corrected projected image.

Images