Image processing device, radiographic system, method for operating image processing device, and program

Abstract

This image processing device for processing a medical image includes an acquisition means for acquiring, using an input image as input data for a trained model obtained by using training data that includes information indicating a femur region in an image of the femur and training data that includes information indicating a neck region-of-interest including the femoral neck, a neck region-of-interest including the femoral neck in the input image.

Description

Image processing device, radiation imaging system, operation method of image processing device, and program

[0001] The disclosure of this specification relates to an image processing device, a radiography system, an operation method of an image processing device, and a program.

[0002] Bone density is measured to diagnose osteoporosis and determine the effectiveness of therapeutic drugs. ２ This is the bone mass per unit area. A known technique for measuring bone density is to use images acquired by DXA (Dual Energy X-ray Absorbtiometry). In DXA, X-ray imaging is performed using two types of tube voltage, and materials such as bone and soft tissue are differentiated based on the difference in absorption rate. Images acquired using DXA in this way are called energy subtraction images (hereinafter referred to as ES images) or material decomposition images.

[0003] For bone density measurement, it is necessary to obtain a target region for bone density measurement from the bone region of the ES image. When measuring bone density of the femur, the target region for bone density measurement is generally the neck region in the proximal part of the femur. Furthermore, the region around the thinnest part of the neck, approximately perpendicular to the cervical axis, is set as a neck ROI (Region of Interest).

[0004] Patent Document 1 discloses a technique for automatically extracting a femoral neck region in an image using machine learning, and obtaining a neck ROI based on this neck region.

[0005] Patent No. 7226199

[0006] In the above-mentioned Patent Document 1, the femoral neck region is extracted using a trained model created by a learning process using training images with labels that enable discrimination between the femoral neck region and the background region. However, if there is an inference error in the output result of the trained model, an error will occur in the angle and position of the neck ROI. For example, if the shape of the bone in the input image is distorted or if the quality of the input image is low due to misalignment between images of two different tube voltages, inference errors in the neck region are likely to occur.

[0007] Therefore, one of the objectives of one embodiment of the disclosure of this specification is to obtain with high accuracy a neck ROI, which is a target region for bone density measurement of the femur.

[0008] An image processing device according to one embodiment disclosed in this specification is an image processing device that processes medical images, and is equipped with an acquisition means that acquires a neck region of interest including the femoral neck in an input image by using an input image as input data for a trained model obtained using training data including information indicating the femur region in an image of the femur and training data including information indicating a neck region of interest including the femoral neck.

[0009] Further features of the present disclosure will become apparent from the following description of exemplary embodiments which proceeds with reference to the accompanying drawings.

[0010] 1 shows an example of a schematic configuration of an X-ray imaging system according to Example 1. 2 shows an example of an input image according to Example 1. 3 shows an example of a cervical ROI according to Example 1. 4 shows an example of a flowchart of image processing according to Example 1. 5 shows an example of an acquisition result by a first acquisition unit according to Example 1. 6 shows an example of an acquisition result by a first acquisition unit according to Example 1. 7 shows an example of an acquisition result by a first acquisition unit according to Example 1. 8 shows an example of an acquisition result by a first acquisition unit according to Example 1. 9 shows an example of an acquisition result by a first acquisition unit according to Example 1. 10 shows an example of a position and angle of a cervical ROI according to Example 1. 11 shows an example of an input image to a second acquisition unit according to Example 1. 12 shows another example of an input image to a second acquisition unit according to Example 1. 13 shows an example of a cervical axis according to Example 1. 14 shows an example of a radiological image according to Example 2. 15 shows an example of a flowchart of image processing according to Example 2.

[0011] Hereinafter, exemplary embodiments for carrying out the present disclosure will be described in detail with reference to the drawings. In the drawings, the same reference numerals are used to indicate identical or functionally similar elements. In addition, in each drawing, some components, parts, and processes that are not important for the description may be omitted.

[0012] In the following examples, an imaging system using X-rays as an example of radiation will be described, but the imaging system according to the present disclosure may use other types of radiation. Here, the term "radiation" includes, for example, electromagnetic radiation such as X-rays and gamma rays, as well as particle radiation such as alpha rays, beta rays, particle beams, proton beams, heavy ion beams, and meson beams.

[0013] In the following, a machine learning model refers to a learning model based on a machine learning algorithm. Specific examples of machine learning algorithms include nearest neighbor algorithms, naive Bayes algorithms, decision trees, and support vector machines. Another example is deep learning, which uses a neural network to generate features and connection weighting coefficients for learning. Algorithms using decision trees include gradient boosting techniques such as LightGBM and XGBoost. Any of the above algorithms can be used as appropriate and applied to the following embodiments and modifications. Furthermore, training data refers to training data and is composed of pairs of input data and output data. Furthermore, output data from training data is also referred to as correct answer data.

[0014] A trained model is a machine learning model that follows any machine learning algorithm, such as deep learning, and has been trained (learned) in advance using appropriate training data (learning data). However, although a trained model is obtained in advance using appropriate training data, it does not mean that it does not undergo further learning, and additional learning can also be performed. Additional learning can also be performed after the device is installed at the site of use.

[0015] (First Embodiment) Hereinafter, a radiation imaging system, an image processing device, and an operation method of the image processing device according to a first embodiment of the present disclosure will be described with reference to Figs. 1 to 9. Fig. 1 shows an example of the configuration of an X-ray imaging system 100, which is an example of a radiation imaging system according to this embodiment. The X-ray imaging system 100 includes an X-ray generator control unit 102, an X-ray tube 103, an FPD (Flat Panel Detector, FPDe) 106, and an image processing device 120. The configuration of the X-ray imaging system 100 is also simply referred to as an X-ray imaging device. The image processing device 120 processes information based on captured X-ray images.

[0016] When the exposure switch is pressed, the X-ray generator control unit 102 applies a high-voltage pulse to the X-ray tube 103 to generate X-rays, and the X-ray tube 103 irradiates the X-rays. When the X-ray tube 103 irradiates X-rays, the X-rays pass through an object (not shown) on a gantry 104 and a grid 105, and reach the FPD 106. The arriving X-rays are converted into visible light by the FPD 106 and output as an X-ray image. The output X-ray image is transferred to the image processing device 120 via the control unit 101.

[0017] The FPD 106 has an X-ray detection unit (not shown) with a pixel array for generating signals corresponding to X-rays. The X-ray detection unit detects X-rays that have passed through a grid as image signals. The X-ray detection unit has pixels arranged in an array (two-dimensional area) that output signals corresponding to incident light. The photoelectric conversion element of each pixel converts X-rays converted into visible light by a phosphor into an electrical signal and outputs the electrical signal as an image signal. In this way, the X-ray detection unit is configured to detect X-rays that have passed through the grid and acquire an image signal (X-ray image). The FPD 106 outputs an image signal (X-ray image) read out in accordance with instructions from the control unit 101 to the control unit 101. The X-ray detection unit may be an indirect conversion type detection unit that first converts X-rays into visible light using a scintillator or the like and then converts the visible light into an electrical signal using a photoelectric conversion element, or it may be a direct conversion type detection unit that directly converts incident X-rays into an electrical signal.

[0018] The control unit 101 transfers the X-ray images acquired from the FPD 106 to an image processing device 120 that processes the X-ray images and to a storage unit 109 that stores the results of image processing and various programs. The storage unit 109 is configured, for example, with a ROM (Read Only Memory) or a RAM (Random Access Memory). The storage unit 109 can store images output from the control unit 101, images processed by the image processing device 120, and calculation results calculated by the image processing device 120.

[0019] The image processing device 120 processes the input image and outputs the calculation results, images, etc. The image processing device 120 is provided with a first acquisition unit 121 and a second acquisition unit 122. The first acquisition unit 121 extracts and acquires a proximal femur region (femur region) including the region of the proximal femur from the input image using a first trained model described below. The second acquisition unit 122 acquires a neck ROI, which is a target region for bone density measurement of the femur, based on the proximal femur region acquired by the first acquisition unit 121 using a second trained model described below. Details of the processing of the first acquisition unit 121 and the second acquisition unit 122 will be described later. The image processing device 120 can also perform bone density analysis on the acquired neck ROI and calculate bone density.

[0020] The control unit 101 and the image processing device 120 can be configured by a computer provided with a processor and a memory. The control unit 101 and the image processing device 120 may be configured by a general computer or a computer dedicated to the X-ray imaging system.

[0021] The components of the control unit 101 and each component of the image processing device 120, such as the first acquisition unit 121, are functionally configured using, for example, one or more processors, such as CPUs (Central Processing Units), and programs loaded from the storage unit 109. The processor may be, for example, an MPU (Micro Processing Unit), a GPU (Graphical Processing Unit), or an FPGA (Field-Programmable Gate Array). The components of the control unit 101 and each component of the image processing device 120 may be configured using an integrated circuit that performs a specific function, such as an ASIC. The internal configuration of the control unit 101 and the image processing device 120 may also include a graphics control unit such as a GPU, a communication unit such as a network card, and an input / output control unit such as a keyboard, display, or touch panel.

[0022] The display unit 107 (monitor) displays X-ray images (digital images) received by the control unit 101 from the FPD 106, images processed by the image processing device 120, patient information, user notifications, etc. The operation unit 108 can input instructions to the image processing device 120 and the FPD 106 via a user interface. The operation unit 108 may be configured using input devices such as a mouse and a keyboard. The control unit 101 functions as an example of a display control means that can control the display unit 107 in response to instructions from the operation unit 108, etc.

[0023] In this embodiment, the image processing device 120 and the control unit 101 are described as separate devices, but they can also be configured as the same device, for example, a single computer. The display unit 107 can be configured as a touch panel display, and in this case, the display unit 107 can also be used as the operation unit 108.

[0024] 2 shows an example of an image of the vicinity of the proximal femur acquired by the X-ray imaging system 100. For the purpose of explanation, Fig. 2 shows a head region 21 including the head of the femur, a trochanter region 22 including the trochanter of the femur, a subtrochanteric region 23 including the lower part of the trochanter of the femur (in this specification, the shaft in the image is also treated as this region), and a pelvic region 24 including the pelvis. The proximal femur region is made up of the head region 21, the trochanter region 22, and the subtrochanteric region 23.

[0025] Although the following description will be given for the case where the left leg is imaged as shown in Fig. 2, the same processing may be performed when the opposite leg is imaged. Furthermore, when either the left or right side is used as the reference side and the opposite leg is imaged, the following processing may be performed on an image obtained by flipping the image acquired by the X-ray imaging system 100 horizontally.

[0026] Furthermore, Fig. 3 shows an example of a neck ROI (neck region of interest) according to this embodiment. In Fig. 3, a rectangle 25 corresponding to the neck ROI, which is the target for measuring the bone density of the femur, is shown by a dashed line. The neck ROI is generally set as a rectangle. However, the neck ROI is not limited to a rectangle and may be set in any shape depending on the desired configuration.

[0027] The image processing device 120 in this embodiment acquires a neck ROI as shown in FIG. 3 using an image of the vicinity of the proximal femur as an input image, as shown in FIG. 2 . The input image may be an ES image acquired by DXA. This ES image may be generated from images captured by X-ray imaging at two different tube voltages using a kV switching method. Alternatively, this ES image may be generated from images of each layer obtained by a single exposure using a stacked sensor formed of two types of materials (phosphors, scintillators) with different X-ray absorption rates as the FPD 106. As an example, these two types of materials may contain CsI(Tl) (thallium-activated cesium iodide) and GOS (gadolinium sulfate) as their main components. This stacked sensor may be configured so that the material on the side where X-rays are incident (the subject side) contains CsI(Tl) and the material on the opposite side contains GOS. Alternatively, a plain X-ray image acquired by a single exposure (plain X-ray photography) may be used as the input image. If the FPD 106 is a stacked sensor, an image of one of the layers (for example, the upper layer on the subject side) may be used as the input image.

[0028] Image processing by the image processing device 120 for acquiring a cervical ROI in this embodiment will be described below with reference to Fig. 4. Fig. 4 shows the image processing flow of the image processing device 120 for acquiring a cervical ROI in this embodiment. When the image processing flow according to this embodiment starts, the process proceeds to step S11 (first acquisition step).

[0029] In step S11, the first acquisition unit 121 of the image processing device 120 uses the first trained model to extract (segment) and acquire the femoral head region, trochanter region, subtrochanteric region, and pelvis region from the input image. In this manner, the first acquisition unit 121 is an example of a first acquisition means that acquires the proximal femur region in the input image using the input image as input data for the first trained model. In this embodiment, the first acquisition unit 121 acquires not only the proximal femur region but also the pelvis region, which is a bone region other than the femur region, for use in acquiring the background region in a later step. Hereinafter, the acquisition process of various regions such as the femoral head region, trochanter region, subtrochanteric region, pelvic region, and neck ROI is assumed to include the acquisition of images (information) indicating these regions and the identification of these regions.

[0030] Examples of results extracted (acquired) by the first acquisition unit 121 are shown in Figures 5A to 5D. The white area in Figure 5A represents the extraction result (acquisition result) of the femoral head region, the white area in Figure 5B represents the extraction result of the trochanter region, the white area in Figure 5C represents the extraction result of the subtrochanter region, and the white area in Figure 5D represents the extraction result of the pelvis region. As an example of the brightness values ​​of images representing the extraction results of these regions, for example, the brightness value of the white area, which is the bone region, may be 1, and the brightness value of the black area, which is the background region, may be 0. An annotated image including labels indicating specific regions, such as an image in which the bone region is represented by brightness values, is hereinafter referred to as a labeled image. Note that the labels are not limited to brightness values ​​and may be any numerical values, characters, or the like depending on the desired configuration. Different labels may also be assigned to different regions.

[0031] The first trained model can be stored in the storage unit 109, and the first acquisition unit 121 may call up and use the first trained model stored in the storage unit 109. The trained models for extracting the respective regions shown in FIGS. 5A to 5D may be a common model. When an image of the area around the proximal femur is input, such a trained model can output extraction results for the femoral head region, trochanter region, subtrochanteric region, and pelvic region from each channel. Alternatively, trained models for extracting each region may be provided separately. These trained models may have the same structure, with only different weights. In this embodiment, the X-ray imaging system 100 is configured to include the trained model, but the trained model may be provided outside the X-ray imaging system 100.

[0032] The first trained model here can be generated, for example, by training a combination of an image of the area around the proximal femur and an image in which the femoral head region, trochanter region, subtrochanteric region, and pelvic region within that image are labeled as training data.

[0033] In addition, when separate trained models for extracting each region are provided, the labeled images may be trained separately. Specifically, a trained model for extracting the femoral head region may be generated by using, as training data, a combination of an image of the vicinity of the proximal femur and an image within that image labeled with the femoral head region. Similarly, a trained model for extracting the trochanter region may be generated by using, as training data, a combination of an image of the vicinity of the proximal femur and an image within that image labeled with the trochanter region. Furthermore, a trained model for extracting the subtrochanter region may be generated by using, as training data, a combination of an image of the vicinity of the proximal femur and an image within that image labeled with the subtrochanter region. Furthermore, a trained model for extracting the pelvic region may be generated by using, as training data, a combination of an image of the vicinity of the proximal femur and an image within that image labeled with the pelvic region.

[0034] In this embodiment, for example, the control unit 101 can function as an example of a learning unit that learns the learned model. However, the first acquisition unit 121 may use the first learned model that has been learned by the image processing device 120, another learning device, or the like.

[0035] Furthermore, a GPU can perform efficient calculations by processing a larger amount of data in parallel. Therefore, when performing learning multiple times using a machine learning algorithm such as deep learning, it is effective to perform the processing using a GPU. Therefore, in this embodiment, the processing by the control unit 101, which functions as an example of a learning unit, may use a GPU in addition to a CPU. Specifically, when executing a learning program including a learning model, learning can be performed by the CPU and GPU working together to perform calculations. Note that the processing by the learning unit may be performed only by the CPU or the GPU. Furthermore, the extraction processing according to this embodiment may also be realized using a GPU, as with the learning unit. Note that if the trained model is provided in an external device, the control unit 101 does not need to function as a learning unit.

[0036] The learning unit may also include an error detection unit and an update unit (not shown). The error detection unit obtains the error between correct data and output data output from the output layer of the neural network in response to input data input to the input layer. The error detection unit may use a loss function to calculate the error between the output data from the neural network and correct data. The update unit updates the connection weighting coefficients between the nodes of the neural network based on the error obtained by the error detection unit so as to reduce the error. This update unit updates the connection weighting coefficients using, for example, an error backpropagation method. The error backpropagation method is a technique for adjusting the connection weighting coefficients between the nodes of each neural network so as to reduce the error.

[0037] Note that, as the machine learning model according to this embodiment, for example, FCN (Fully Convolutional Network) or SegNet can be used. Furthermore, as the machine learning model for object recognition, for example, RCNN (Region CNN), fastRCNN, or fasterRCNN can be used. Furthermore, as the machine learning model for object recognition in units of regions, YOLO (You Only Look Once) or SSD (Single Shot Detector or Single Shot MultiBox Detector) can also be used.

[0038] Furthermore, the image processing device 120 may perform preprocessing on the input image before the segmentation process. An example of the preprocessing may be a process for improving contrast. For example, the contrast improvement process may be a process for linearly expanding the luminance gradation by setting the maximum luminance value in the image to 1 and the minimum luminance value in the image to 0. Another example of the preprocessing may be a process such as histogram equalization. Another example of the preprocessing may be a process for enhancing edges. For example, the edge enhancement process may be a filter process. Another example of the preprocessing may be a standardization process for setting the average luminance value of the image to 0 and the variance to 1. The image processing device 120 may also perform a combination of the above preprocessing processes. These preprocessing processes may also be performed on the training data used to train the first trained model.

[0039] The input image may be an image obtained by trimming an image acquired by the X-ray imaging system 100 in accordance with an instruction from the operator via the operation unit 108 while viewing the image displayed on the display unit 107. The trimmed image may also be resized to a specified image size. The specified image size may be, for example, the input image size of the trained model. The input image may also be an image automatically trimmed by the image processing device 120 based on the imaging range so as to leave the proximal femur.

[0040] Furthermore, if the input image is an image that has been cropped by an operator, the input image may not include the subtrochanteric region. In this case, the output of the first trained model may not include the subtrochanteric region. Also, the image acquired by the X-ray imaging system 100 may not include the subtrochanteric region.

[0041] In addition, if a particularly small area is also extracted outside the extracted area for each target area, the first acquisition unit 121 may consider this area to be a segmentation error and remove it from the extraction result, leaving only the area with the largest area.

[0042] Furthermore, the control unit 101 may cause the display unit 107 to display the proximal femur region acquired by the first acquisition unit 121, and the operator may use the operation unit 108 to give instructions for correction, such as adding or deleting a bone region. In this case, the control unit 101 sends the operator's instructions to the image processing device 120, and the image processing device 120 can make corrections, such as adding or deleting a bone region, in accordance with the acquired instructions from the operator.

[0043] Here, if the label of the femoral head region and the label of the trochanter region acquired by the first acquisition unit 121 are the same, an instruction to add a bone region can be given by specifying the location where the same label value is to be added to the displayed proximal femoral region.

[0044] Furthermore, if the label of the femoral head region and the label of the trochanter region acquired by the first acquisition unit 121 are different, the operator can individually specify the location of the femoral head region and the trochanter region where the bone region is to be added to the displayed proximal femur region. By performing such an operation, the operator can issue instructions for addition and correction of each of the femoral head region and the trochanter region.

[0045] Even if the label of the femoral head region and the label of the trochanter region acquired by the first acquisition unit 121 are different, the operator may instruct additional correction using a bone region label common to the femoral head region and the trochanter region. In this case, the first acquisition unit 121 may distinguish whether the added bone region is a femoral head region or a trochanter region. The label of the added bone region may be set to, for example, the label of the adjacent pixel of its outer frame.

[0046] Furthermore, the image processing device 120 may use the first acquired model to acquire a labeled image in which the femoral head region and the trochanter region of the input image are designated as regions with a common label. In this case, the first acquired unit 121 may further acquire a labeled image in which the femoral head region and the trochanter region are designated as separate regions from the acquired labeled image. The first acquired unit 121 may use a third acquired model to acquire a labeled image in which the femoral head region and the trochanter region are separately extracted from a labeled image in which the femoral head region and the trochanter region are designated as a common label. Here, the training data of the third acquired model may use a labeled image in which the femoral head region and the trochanter region are designated as a common label as input data, and a labeled image in which the femoral head region and the trochanter region in the labeled image are designated as separate labels as output data. The first acquisition unit 121 may also acquire labeled images in which the femoral head region and the trochanter region are separately extracted from a labeled image in which a common label is assigned to the femoral head region and the trochanter region by rule-based processing. The first acquisition unit 121 may also include a third acquisition unit and a fourth acquisition unit that perform these respective processes.

[0047] When the neck region is extracted from the input image using a trained model, it is difficult to identify the cause of any abnormalities in the results or to correct them midway. In contrast, the image processing device 120 can more reliably obtain the final neck ROI by displaying and correcting the proximal femur region obtained midway as described above.

[0048] Next, in step S12 (second acquisition process), the second acquisition unit 122 of the image processing device 120 acquires a cervical ROI from the acquisition results of step S11 (Figures 5A to 5D) using the second learned model.

[0049] As described above, when an image processing device is used to extract the femoral neck region and neck ROI from an input image using a single trained model, inference errors in the output results of the trained model result in errors in the angle and position of the neck ROI. For example, if the bone shape of the input image is distorted or the quality of the input image is low due to misalignment between images of two different tube voltages, inference errors in the neck region are likely to occur. Furthermore, the size of the set neck ROI may vary depending on the policy of the facility or the technician. When extracting the neck region using a trained model, the size of the neck region and neck ROI depend on the training data. Changing the output of the trained model requires re-training, making it difficult to change the neck ROI settings in accordance with the facility's policy. Alternatively, the operator may have to go through the trouble of correcting the output of the trained model.

[0050] In contrast, the second acquisition unit 122 according to this embodiment acquires the position and angle of the neck ROI in the input image using the second trained model from the proximal femur region that is the acquisition result of step S11. In this way, the second acquisition unit 122 is an example of a second acquisition means that acquires a neck region of interest, which is a target region for bone density measurement of the femur, using the proximal femur region acquired by the first acquisition means as input data for the second trained model.

[0051] FIG. 6 shows an example of the position and angle of the neck ROI (rectangle 25) according to this embodiment. The position of the neck ROI here can be set as the center position RP of the neck ROI (rectangle 25) as shown in FIG. 6 . The angle of the neck ROI can be set as the angle of a line parallel to the long side of the rectangle (crossing the neck), as shown by line RA in FIG. 6 , relative to a line in the left-right direction of the image. However, the position and angle of the neck ROI are not limited to these, and any representative position and angle can be set.

[0052] The second acquisition unit 122 can also set a fixed value in advance as the size of the neck ROI (the length and width of the rectangle). In this way, the second acquisition unit 122 can acquire a neck ROI positioned correctly by setting the size of the neck ROI in accordance with the facility's policy, etc., and then specifying the output of the trained model as the position and angle of the neck ROI. The size of the neck ROI may also be changed depending on the subject's physique, the size and positional relationship of the femoral head region and trochanter region acquired in step S11, etc. In this case, the second acquisition unit 122 may determine the size of the neck ROI by, for example, referring to a table storing the correspondence between the subject's physique, the size and positional relationship of the femoral head region and trochanter region, and the size of the neck ROI, or by using a function related to the correspondence.

[0053] The image input by the second acquisition unit 122 to the second trained model may include a labeled image of the femoral head region as shown in FIG. 5A and a labeled image of the trochanter region as shown in FIG. 5B. Here, the second trained model may be configured to receive a tensor having these two images in separate channels. Furthermore, the input data of the second trained model may be a single labeled image in which the femoral head region and the trochanter region are labeled, as shown in FIG. 7. In this case, the labels of the femoral head region and the trochanter region may be a common label. Furthermore, the input data of the second trained model may be a labeled image in which the brightness values ​​of the femoral head region (the shaded area in FIG. 8) and the trochanter region (the white area in FIG. 8) are individually set, as shown in FIG. 8.

[0054] The second trained model used in this embodiment may be a regression model that uses the above-described labeled image as input data and the numerical values ​​of the position and angle of the cervical ROI in the input image or the labeled image as output data. In this case, the second trained model can be generated by, for example, training a combination of the above-described labeled image and the input image corresponding to the labeled image or the position and angle of the cervical ROI in the labeled image as training data.

[0055] The task of specifying the position and angle of the cervical ROI in the training data may involve specifying numerical values. Alternatively, the task of specifying the position and angle of the cervical ROI in the training data may involve specifying a center position RP of the cervical region within the image and a line RA that passes through the center position RP and is parallel to the long side of a rectangle, as shown in FIG. 6 . In this case, the angle of the line RA can be obtained by post-processing, and the angle of the line RA can be specified as the angle of the cervical ROI. Alternatively, the task of specifying the position and angle of the cervical ROI in the training data may involve specifying a rectangle 25 as shown in FIG. 6 , and the center position and angle of the rectangle can be obtained by post-processing. The rectangle specified at this time may be a different size from the cervical ROI.

[0056] The second trained model may also be a model that extracts a region corresponding to the line RA shown in FIG. 6 , which is specified as training data, from an input labeled image using processing such as semantic segmentation. In this case, the second trained model may be obtained by training using training data in which the labeled image described above is used as input data and the region corresponding to the line RA in the input image or the labeled image is used as output data. The second acquisition unit 122 may acquire an approximation line of the line RA by performing post-processing such as linear approximation on the extracted region, and acquire the center position and angle of this approximation line as the position and angle of the cervical ROI. In this case, the center position of the cervical ROI may be the center of gravity of the region where the acquired approximation line overlaps with the bone region extracted by the first acquisition unit 121.

[0057] The second trained model may also be a model that extracts, by segmentation processing, a rectangular region 25 corresponding to the cervical ROI, as shown in FIG. 6 , specified as training data, from an input labeled image. In this case, the second trained model may be obtained by training data in which the above-described labeled image is used as input data and the input image or the rectangular region 25 within the labeled image is used as output data. The second acquisition unit 122 may approximate the extracted region to a rectangle, and then acquire the center position and the angle of the long side of the approximated rectangular region as the position and angle of the cervical ROI. For example, by providing training data as a rectangular region for the input labeled image, the appropriate size of the cervical ROI can also be learned.

[0058] The second trained model may be a single model that outputs the position and angle of the cervical ROI on separate channels. Alternatively, the second trained model may include a model that outputs the position of the cervical ROI and a model that outputs the angle of the cervical ROI. In this case, the model that outputs the position of the cervical ROI may be obtained by training training data using the above-described labeled image as input data and the position of the cervical ROI in the input image or the labeled image as output data. Alternatively, the model that outputs the angle of the cervical ROI may be obtained by training training data using the above-described labeled image as input data and the angle of the cervical ROI in the input image or the labeled image as output data.

[0059] The second trained model may output the position of the cervical ROI and the angle of the cervical axis 26 shown in Fig. 9. The cervical axis 26 may be set, for example, as a line passing through the center position of the cervical ROI and perpendicular to the long side of the cervical ROI. In this case, the second trained model may be obtained by training training data using the above-described labeled image as input data and outputting the position of the cervical ROI and the angle of the cervical axis 26 in the input image or the labeled image.

[0060] In addition, the control unit 101 may calculate the cervical axis 26 based on the position of the cervical ROI and the angle of the long side output by the second learned model, and display the cervical axis 26 on the display unit 107.

[0061] Furthermore, the second acquisition unit 122 may acquire an offset value for the position or angle of the cervical ROI output by the second trained model. This offset value can be determined based on the policy of the facility or the technician.

[0062] Note that the input data to the second trained model may be an image in which an input image (such as an ES image) is superimposed on the label image acquired by the first acquisition unit 121. In this case, the input data of the training data may also be a similar image.

[0063] Furthermore, the input data to the second trained model may be two images: a labeled image acquired by the first acquisition unit 121 and an input image (such as an ES image). These two images may be input to separate channels, or may be combined and input to a single channel. In this case, the input data of the training data may also be similar images.

[0064] In this embodiment, the second acquisition unit 122 acquires the position and angle of the neck ROI based on the label image (information) indicating the femur region acquired by the first acquisition unit 121. In this way, by approximately acquiring the overall structure of the femur region once and then searching for the neck ROI relative to it, it is possible to acquire the neck ROI with higher accuracy even when the shape of the bone is distorted or the quality of the input image is low. Furthermore, in this embodiment, the position and angle of the neck ROI are acquired using a trained model, and rules for the size are set arbitrarily, allowing examination settings to be made in accordance with the policies of the facility or the medical technician.

[0065] In this way, the second acquisition unit 122 can extract and acquire the femoral neck ROI with high accuracy by referring to the acquisition results of the proximal femoral region including the femoral head region and the trochanter region. After the neck ROI is acquired in step S12, the process proceeds to step S13.

[0066] Next, in step S13, the second acquisition unit 122 sets a background region for calculating the transmittance of bones. An example of a method for setting the background region will be described below.

[0067] An example of the background region may be a region other than the bone region in the input image as shown in FIG. 2 , which is automatically selected. The bone region here refers to the region including the femoral head region, trochanter region, subtrochanteric region, and pelvis region acquired in step S11. Therefore, the second acquisition unit 122 can, for example, set the region of the input image excluding the bone region acquired in step S11 as the background region. Furthermore, if the input image is an image cropped by the operator, the background region may be a region other than the bone region in the cropped image. In this case, the second acquisition unit 122 can, for example, set the region of the cropped image excluding the bone region acquired in step S11 as the background region. However, the method for setting the background region is not limited to this, and any method may be used. For example, the operator may specify the background region using the operation unit 108.

[0068] The control unit 101 may display the cervical ROI set as described above on the display unit 107. The control unit 101 may also display the cervical axis. Furthermore, based on the cervical ROI displayed on the display unit 107, the operator may use the operation unit 108 to modify and determine the position and angle of the cervical ROI. The operator may also modify the size of the cervical ROI by specifying the positions of the four corners of the cervical ROI using the operation unit 108. Note that the control unit 101 does not need to display the position of the background region on the display unit 107.

[0069] Furthermore, the control unit 101 can cause the display unit 107 to display a warning message when it is determined that at least one of the acquired results of the position and angle of the neck ROI is abnormal.

[0070] An example of a condition for determining an abnormality is when the position of the neck ROI is closer to the pelvis than the center of the femoral head region. The position of the neck ROI here may refer to the left-right position (X coordinate) of the center of the neck ROI (center position RP in FIG. 6 ) acquired in step S12. The center position of the femoral head region here may be determined by, for example, calculating the center of gravity of the femoral head region ( FIG. 5A ) acquired in step S11 and then referencing the left-right position (X coordinate) of the calculated center of gravity. An abnormality may be determined when the X coordinate of the neck ROI is closer to the pelvis than the X coordinate of the center of the femoral head region. Here, the "pelvis side" refers to the left side of the image when the input image is an image of the left foot, as shown in FIG. 6 , and conversely, refers to the right side of the image when the input image is an image of the right foot.

[0071] Another example of a condition for determining an abnormality is when the angle of the neck ROI is different from the reference angle by more than a threshold value. The reference angle and threshold value here can be arbitrarily specified. For example, as shown in FIG. 6 , if the input image is an image of the left leg, the reference angle of the neck ROI (the angle of line RA in FIG. 6 relative to the line in the horizontal direction of the image) may be set to 45 degrees, and the threshold angle may also be set to 45 degrees. That is, in this example, an abnormality may be determined when the angle of the neck ROI is 90 degrees or more. As in these examples, the determination of an abnormality may be performed using rule-based processing.

[0072] The image processing device 120 calculates bone density in the cervical ROI by referencing the cervical ROI and background region acquired as described above. Bone density can be calculated using a known method, such as information on the thickness of bones and soft tissues in an image acquired by DXA. Alternatively, bone density may be calculated using information on atomic number and surface density acquired by DXA. The control unit 101 may display the calculated bone density on the display unit 107. Furthermore, the bone region referenced for bone density calculation is the cervical region within the cervical ROI, and bone regions other than the cervical region included in the cervical ROI (such as the pelvic region) may not be referenced.

[0073] The control unit 101 may cause the display unit 107 to display any value useful for diagnosis based on the calculated bone density. For example, the control unit 101 may cause the display unit 107 to display a value compared with the average bone density of people of the same age or other age groups. The control unit 101 may also cause the display unit 107 to display time-series data compared with, for example, past measurement results. Furthermore, in addition to the bone density value calculated within the cervical ROI, the display unit 107 may also display a bone density value calculated using the entire bone region or a portion of the bone region in the input image and the background region.

[0074] As described above, the X-ray imaging system 100 according to this embodiment functions as an example of a radiographic system and includes the image processing device 120, the X-ray tube 103, and the FPD 106. The X-ray tube 103 functions as an example of a radiation generating device that generates radiation. The FPD 106 functions as an example of a radiation detecting device that detects radiation and generates a radiographic image. The image processing device 120 functions as an example of an image processing device that processes medical images and includes the first acquisition unit 121 and the second acquisition unit 122. The first acquisition unit 121 and the second acquisition unit 122 can function as an example of an acquisition means that acquires a neck region of interest including the femoral neck in an input image by using an input image as input data for a trained model obtained using training data including information indicating a proximal femur region (femur region) in an image of the femur and training data including information indicating a neck region of interest including the femoral neck.

[0075] More specifically, the trained model may include a first trained model obtained using training data including information indicating a femur region in an image of a femur, and a second trained model obtained using training data including information indicating a neck region of interest including the neck of the femur. The first acquisition unit 121 may function as an example of a first acquisition means that acquires information indicating the femur region in the input image by using the input image as input data for the first trained model. The second acquisition unit 122 may function as an example of a second acquisition means that acquires a neck ROI (neck region of interest) in the input image by using information indicating the femur region acquired as input data for the second trained model. The neck region of interest in the input image may be a target region for bone density measurement. The femur region may include a region of the proximal femur.

[0076] With this configuration, the image processing device 120 according to this embodiment first obtains an approximate global structure of the femur region and then searches for the neck ROI based on that. Therefore, even when the bone shape is distorted or the quality of the input image is low, the neck ROI can be obtained with higher accuracy.

[0077] The second acquisition unit 122 acquires a neck region of interest in the input image using the position and angle of the neck region of interest based on the output from the second trained model. The size of the neck region of interest in the input image may be set in advance or may be set by the second acquisition unit 122. With this configuration, the size of the neck ROI can be set in accordance with the facility's policy, and then the output of the trained model can be specified as the position and angle of the neck ROI, thereby acquiring a neck ROI positioned in the correct position. The size of the neck region of interest in the input image can be set according to the subject's physique and the size and positional relationship of the acquired femoral head and trochanter regions. This allows for examination settings to be configured in accordance with the facility's or technician's policy.

[0078] The first acquisition unit 121 can also acquire a femoral head region including a femoral head region in the input image. Furthermore, the first acquisition unit 121 can acquire at least one of a trochanter region including a trochanter region of the femur, a subtrochanter region including a subtrochanter region of the femur, and a pelvic region including a pelvis region in the input image. The first acquisition unit 121 may separately acquire information indicating the femoral head region and information indicating the trochanter region using the first trained model, or may acquire information indicating a combined region of the femoral head region and the trochanter region.

[0079] The input image may be an energy subtraction image, at least one of two images acquired using radiation at two different tube voltages, at least one of two images acquired using a sensor having two types of materials with different radiation absorption rates (a stacked sensor), or a simple radiographic image, where one of the two types of materials may contain CsI and the other of the two types of materials may contain GOS.

[0080] The second acquisition unit 122 may further acquire a cervical axis. Furthermore, the second acquisition unit 122 may acquire a cervical region of interest in the input image by using an image in which the information indicating the femur region acquired by the first acquisition unit 121 is superimposed on the input image. In this case, the second trained model can extract more features, which is expected to enable acquisition of a cervical ROI with higher accuracy.

[0081] Furthermore, the second acquisition unit 122 can acquire a background region within the input image based on the femur region acquired by the first acquisition unit 121. Therefore, the image processing device 120 can measure bone density of the femur using the pixel values, etc., of the neck ROI set according to the position and angle of the acquired neck ROI and the acquired background region. Note that the first acquisition unit 121 may further extract a bone region other than the femur region within the input image, and the second acquisition unit 122 may use the bone region other than the femur region to extract the background region.

[0082] The image processing device 120 further includes a control unit 101. The control unit 101 can function as an example of a display control means that displays the intermediate results of the femur region acquired by the first acquisition unit 121 and the results of the neck ROI acquired by the second acquisition unit 122 on the display unit 107. In this case, the operator can check the results of the acquisition by the first acquisition unit 121 or the second acquisition unit 122 displayed on the display unit 107 and confirm whether they correspond to the desired position and angle of the femur region or neck ROI.

[0083] Furthermore, the image processing device 120 may further include an operation unit 108 that allows the operator to operate the correction of the results acquired by the first acquisition unit 121 and the second acquisition unit 122 while viewing the display on the display unit 107. In this case, the operator can confirm the display on the display unit 107 and correct and determine the position and angle of the femur region and the neck ROI. Note that when the results acquired by the first acquisition unit 121 are corrected, the second acquisition unit 122 can acquire a neck region of interest in the input image by using the femur region corrected using the operation unit 108 as input data for the second trained model.

[0084] Furthermore, when the cervical ROI (acquisition result) or the acquisition process acquired by the second acquisition unit 122 is determined to be abnormal, the control unit 101 can display a warning such as a message indicating the abnormality on the display unit 107. In this case, the operator can understand the abnormality in the acquisition result or the acquisition process, and change the settings or stop image processing according to a desired policy, etc.

[0085] Furthermore, the first acquisition unit 121 may acquire information indicating a single region in which the femoral head region and the trochanter region of the input image are combined into a single region by using the input image as input data for the first trained model. In this case, the first acquisition unit 121 may acquire information indicating the femur region in the input image by separately extracting information indicating the femoral head region and information indicating the trochanter region from the information indicating the single region. In this case, the operation unit 108 may be used to operate the correction of the single region.

[0086] The first acquisition unit 121 may separately acquire information indicating the femoral head region and information indicating the trochanter region in the acquired single region by using information indicating the acquired single region as input data for the trained model. The trained model can be obtained using training data including information indicating femoral regions obtained by separately extracting the femoral head region and the trochanter region in a single region including the femoral head region and the trochanter region. In contrast, the first acquisition unit 121 may separately acquire information indicating the femoral head region and information indicating the trochanter region in the single region by rule-based processing.

[0087] In the present embodiment, the image processing device 120 has a first acquisition unit 121 that acquires a proximal femur region using a first trained model, and a second acquisition unit 122 that acquires a neck ROI using a second trained model. Here, the first acquisition unit 121 and the second acquisition unit 122 may configure a single acquisition unit that acquires the neck ROI from an input image using a trained model obtained using training data including information indicating the femur region in an image of the femur and training data including information indicating a neck region of interest including the neck of the femur.

[0088] Furthermore, such an acquisition unit may be configured to acquire a neck ROI using an input image as input data for a trained model in which the first trained model and the second trained model are integrated. The integration of the first trained model and the second trained model can be performed, for example, by connecting the output of the first trained model to the input of the second trained model. Here, as described above, the first trained model may be obtained using training data including information indicating a femur region in an image of a femur. Furthermore, the second trained model may be obtained using training data including information indicating a neck region of interest including the neck of the femur.

[0089] The trained model, which is an integration of the first trained model and the second trained model, has parameters corresponding to the first trained model and extracts a proximal femur region, including a femoral head region and a trochanter region, from an input image used as input data. The trained model also has parameters corresponding to the second trained model and outputs the position and angle of a neck ROI based on the extracted proximal femur region. The acquisition unit using the trained model, like the second acquisition unit 122, can acquire a neck ROI in an input image using the position and angle of the neck ROI based on the output from the trained model. Even with this configuration, the overall structure of the femur region is approximately acquired, and the neck ROI is searched for in the acquired femur region. This allows for more accurate acquisition of the neck ROI even when the bone shape is deformed or the quality of the input image is low.

[0090] Such an acquisition unit can function as an example of an acquisition means for acquiring a neck region of interest including the femoral neck in an input image by using the input image as input data for a trained model obtained using training data including information indicating the femoral region in an image of the femur and training data including information indicating the neck region of interest including the femoral neck.

[0091] Second Embodiment Next, a second embodiment of the present disclosure will be described. A configuration example of an X-ray imaging system according to this embodiment is similar to the configuration example of the X-ray imaging system 100 according to the first embodiment, and is shown in Fig. 1. Note that the configuration example of the X-ray imaging system according to this embodiment will be assigned the same reference numerals as the configuration example of the X-ray imaging system 100 according to the first embodiment, and description thereof will be omitted.

[0092] The image processing device 120 in this embodiment extracts the femoral neck ROI using a radiographic image such as that shown in Fig. 10. The radiographic image shown in Fig. 10 may be, for example, a plain X-ray image, which is an X-ray image acquired over a wide area rather than only around the neck ROI. The input image may also be an image acquired by DXA. This embodiment will be described below, focusing on the differences from the first embodiment.

[0093] 11 shows the image processing flow of the image processing device 120 for extracting the cervical ROI in this embodiment. When the image processing according to this embodiment is started, the process proceeds to step S21.

[0094] In step S21, the first acquisition unit 121 uses the trained model to extract the femoral head region and trochanter region of the leg to be measured from the acquired X-ray image of either the left or right leg. Here, a case will be described in which the left leg (the side showing the reference region 31) shown in FIG. 10 is the measurement target.

[0095] The trained model here can be generated by training a combination of, for example, an X-ray image of a human body or phantom including the femur and an image in which the femoral head region and the trochanter region within the X-ray image are labeled. The femoral head region and the trochanter region may be extracted using a process similar to that of step S11 in the first embodiment, and the first acquisition unit 121 may preprocess the X-ray image before inputting it into the trained model. The trained model may extract the femoral head region and the trochanter region separately, or may extract them together as a single region.

[0096] The trained model used in step S21 may extract the femoral head region and the trochanter region of both legs. The first acquisition unit 121 may extract the femoral head region and the trochanter region of the leg to be measured by referring to whether the center coordinates of the femoral head region are on the right or left side of the image.

[0097] The trained model used in step S21 may also extract the femoral head region, trochanter region, subtrochanteric region, and pelvic region from the X-ray image, allowing the trained model used in step S21 to share the same model structure as the trained model used in the subsequent step S23.

[0098] Next, in step S22, the first acquisition unit 121 extracts a reference region 31 in the X-ray image based on the femoral head region and trochanter region extracted in step S21. The reference region 31 is a region around the proximal femur, such as the range region of the input image (FIG. 2) in Example 1, as shown in FIG.

[0099] When the reference region 31 is rectangular, the first acquisition unit 121 can draw each side of the rectangle at positions with margins extending outward from the upper, lower, left, and right ends of the femoral head region and the trochanter region. The margins here can be set to any value, for example, 1 cm. The method for setting the reference region is not limited to this, and the reference region may be a rectangular region with margins on the outside, based on at least one of the femoral head region and the trochanter region. The reference region is not limited to a rectangle, and may be any shape, such as a circle.

[0100] In this way, the first acquisition unit 121 can extract the reference region from the acquired X-ray image.

[0101] Next, in step S23, the first acquisition unit 121 extracts the femoral head region, trochanter region, subtrochanteric region, and pelvis region from the reference region. In this process, the first acquisition unit 121 can perform the same process as step S11 in Example 1 by using the image of the reference region as the input image. In this example, by extracting the reference region from the X-ray image and then extracting the femoral head region, trochanter region, subtrochanteric region, and pelvis region, it is possible to prevent a decrease in model accuracy due to other objects being captured during learning. In this way, the first acquisition unit 121 according to this example is also an example of a first acquisition means that acquires the proximal femur region from the input image using a trained model.

[0102] In step S24, the second acquisition unit 122 extracts a neck ROI based on the femoral head region, trochanter region, and subtrochanteric region within the extracted reference region. That is, the second acquisition unit 122 is an example of a second acquisition means that acquires the neck ROI, which is the target region for bone density measurement of the femur, based on the acquisition result of the first acquisition unit 121. The processing of this step can be performed using the same processing as step S12 of the first embodiment.

[0103] In step S25, the second acquisition unit 122 extracts a background region from the reference region based on the pelvic region from the extracted reference region. The method for extracting the background region is the same as in step S13 of the first embodiment.

[0104] By referring to the extracted cervical ROI and background region as described above, the image processing device 120 according to this embodiment can calculate the bone density in the cervical ROI in the same manner as in embodiment 1. Furthermore, the control unit 101 according to this embodiment may cause the display unit 107 to display the calculated bone density, the set cervical axis and cervical ROI, the acquisition results or the acquisition process of the cervical ROI, etc., in the same manner as in embodiment 1.

[0105] As described above, the first acquisition unit 121 according to this embodiment can extract a reference region including a measurement target from a medical image and use an image showing the reference region as an input image. Here, the reference region can be extracted based on at least one of a femoral head region including the femoral head region and a trochanter region including the femoral trochanter region in the radiographic image. With this configuration, even if an image obtained by capturing a wide area is acquired as the radiographic image, a reference region limited to the area to be measured for bone density can be extracted. This makes it possible to suppress a decrease in the accuracy of extraction processing using a trained model due to unnecessary features appearing in the image.

[0106] In this embodiment, an acquisition unit may be provided that acquires a cervical ROI using a reference region as input data for a trained model that integrates the first trained model and the second trained model. In this case, the trained model may be configured, for example, by connecting the output of the first trained model described above to the input of the second trained model. Furthermore, the acquisition unit may extract a reference image from the input image, similar to the first acquisition unit 121.

[0107] A modified example of at least one of the first and second embodiments of the present disclosure will be described below.

[0108] (Modification) In the above-described Example 2, the neck ROI and background region are extracted after extracting the reference region of the input image, but the neck ROI and background region may be extracted directly from an input image captured over a wide range using the same processing flow as in Example 1. In this case, for example, the first acquisition unit 121 may acquire a femur approximation region from an input image captured over a wide range, and the second acquisition unit 122 may acquire the neck ROI and background region using the femur approximation region.

[0109] The trained models according to the first and second embodiments and the modified examples may extract the femoral head region and other bone regions in the input image, which are categorized as a single label for the trochanteric region, subtrochanteric region, and pelvic region. Then, the neck ROI and background region may be acquired, similar to the first and second embodiments.

[0110] The image processing device 120 according to the first and second embodiments and the modified examples may extract a neck ROI using a total of three images: two images acquired by DXA for different radiation energies and an ES image generated from these images. For example, the first acquisition unit 121 may extract the proximal femur region from the three images when extracting the proximal femur region (or each of the femoral head region, trochanter region, subtrochanteric region, and pelvic region) using a trained model in step S11 of the first embodiment. In this case, the second acquisition unit 122 may use only the region where the extraction results overlap as the final proximal femur region. Alternatively, the second acquisition unit 122 may use only the region extracted as the proximal femur region in two or more of the three images as the final proximal femur region.

[0111] In this case, the X-ray imaging system 100 may have individual trained models for the three images. In this case, each trained model can be generated by training the three images (first and second images related to different radiation energies, and ES images generated from these images) as input data for the respective training data. Note that, as output data for the training data, images in which the proximal femur region is labeled for each type of image may be used. Furthermore, each input channel of one trained model may correspond to these three images. Similarly, the image processing device 120 may extract the proximal femur region using two of these three images.

[0112] As described above, the first acquisition unit 121 according to this modification uses at least two of three images, namely, two images acquired using the DXA method and an energy subtraction image generated from the two images, as input images to acquire femur regions for the at least two images. The second acquisition unit 122 acquires a neck region of interest based on the femur region acquisition results for the at least two images. By performing the processing according to this modification, the influence of segmentation errors can be reduced.

[0113] The training data of the trained models according to the first and second embodiments and the modified examples is not limited to data obtained using the X-ray imaging system itself that actually performs imaging. The image data may be data obtained using the same type of X-ray imaging system or data obtained using the same kind of X-ray imaging system, depending on the desired configuration.

[0114] Furthermore, when multiple images are used as input to a trained model, the multiple input images can be input to multiple input channels of the trained model, respectively. Alternatively, the multiple input images can be combined into a single image and input to one channel of the trained model. In this case, the single image obtained by combining the input images can be used as input data for the training data of the trained model as well.

[0115] Furthermore, in the above-described first and second embodiments and the modified examples, the control unit 101 may superimpose the segmentation results of the proximal femur on the display image and display them on the display unit 107 in response to an operator's instruction, such as the operation of a button (not shown). Note that, when there are multiple display images and corresponding segmentation results, the control unit 101 may superimpose the segmentation results on each display image collectively and display them on the display unit 107 in response to an operator's instruction. Similarly, the control unit 101 may superimpose the cervical ROI, cervical axis, and the acquisition results of the femur region by the first acquisition unit 121 on each display image collectively and display them on the display unit 107 in response to an operator's instruction. The control unit 101 may also display the acquisition results of the cervical ROI and the background region in different colors. In this case, multiple images may be displayed collectively in different colors in response to an operator's instruction.

[0116] In addition, in the trained models according to the above-mentioned Examples 1 and 2 and the modified examples, it is considered that the magnitude of the brightness values ​​of the input data image, the order and gradient of the bright and dark areas, position, distribution, continuity, etc. are extracted as part of the features and used in the inference process.

[0117] Furthermore, the various trained models according to the first and second embodiments and the modified examples described above can be provided in the control unit 101 or the image processing device 120. The trained model may be configured, for example, as a software module executed by a processor such as a CPU, MPU, GPU, or FPGA, or as a circuit performing a specific function such as an ASIC. The trained model may also be provided in a separate server device connected to the control unit 101 or the image processing device 120. In this case, the image processing device 120 can use the trained model by connecting to a server or the like equipped with the trained model via any network such as the Internet. Here, the server equipped with the trained model may be, for example, a cloud server, a fog server, or an edge server. Note that when a network within a facility, a site including a facility, or a region including multiple facilities is configured to enable wireless communication, the reliability of the network may be improved by, for example, configuring it to use radio waves in a dedicated wavelength band allocated exclusively to the facility, site, region, etc. The network may also be configured using wireless communication that is capable of high speed, large capacity, low latency, and multiple simultaneous connections.

[0118] Furthermore, the bone region may be extracted by a known threshold process instead of using a trained model. For example, the first acquisition unit 121 can extract and acquire, as the bone region, pixels in the input image whose brightness is equal to or greater than a threshold.

[0119] In the first and second embodiments and the modified examples, images captured using radiation are used as examples of medical images processed by the image processing device 120. The medical images to be processed may include, for example, images captured by digital radiography (DR) using an FPD or computed tomography (CT). Furthermore, the medical images may include images captured by photon-counting CT, magnetic resonance imaging (MRI), ultrasound imaging, etc. Furthermore, the medical images may include images captured by positron emission tomography (PET), single photon emission computed tomography (SPECT), etc. Furthermore, the medical image may be a still image or a moving image.

[0120] (Other Embodiments) The present disclosure can also be realized by providing software (programs) that realize one or more functions of the various embodiments and modifications described above to a system or device via a network or a storage medium, and having a computer in the system or device read and execute the programs. It can also be realized by a circuit (e.g., an ASIC) that realizes one or more functions. The computer may have one or more processors or circuits, and may include multiple separate computers or a network of multiple separate processors or circuits to read and execute computer-executable instructions.

[0121] In this case, the processor or circuitry may include a central processing unit (CPU), a microprocessing unit (MPU), a graphics processing unit (GPU), an application specific integrated circuit (ASIC), or a field programmable gate array (FPGA), and may also include a digital signal processor (DSP), a data flow processor (DFP), or a neural processing unit (NPU).

[0122] According to at least one embodiment of the present disclosure, the neck ROI, which is the target region for bone density measurement of the femur, can be obtained with high accuracy.

[0123] The present invention is not limited to the above-described embodiments, and various modifications and variations can be made without departing from the spirit and scope of the present invention. Therefore, the following claims are appended to apprise the public of the scope of the present invention.

[0124] This application claims priority based on Japanese Patent Application No. 2023-096070, filed on June 12, 2023, the entire contents of which are incorporated herein by reference.

Claims

1. An image processing device for processing medical images, comprising an acquisition means for acquiring a neck region of interest including the femoral neck in an input image by using an input image as input data for a trained model obtained using training data including information indicating a femur region in an image of the femur and training data including information indicating a neck region of interest including the femoral neck.

2. The image processing device of claim 1, wherein the trained model includes: a first trained model obtained using training data including information indicating a femur region in an image of the femur; and a second trained model obtained using training data including information indicating a neck region of interest including the neck of the femur; and the acquisition means includes: a first acquisition means for acquiring information indicating the femur region in the input image by using the input image as input data for the first trained model; and a second acquisition means for acquiring a neck region of interest in the input image by using the acquired information indicating the femur region as input data for the second trained model.

3. The image processing device according to claim 2, wherein the second acquisition means acquires a cervical region of interest in the input image using a position and angle of the cervical region of interest based on an output from the second trained model.

4. An image processing device as described in claim 2 or 3, wherein the femoral region includes a femoral head and a trochanter, and the first acquisition means separately acquires information indicating the femoral head region and information indicating the trochanter region using the first trained model.

5. An image processing device as described in claim 2 or 3, wherein the femoral region includes a femoral head and a trochanter, and the first acquisition means acquires information indicating a region combining the femoral head region and the trochanter region using the first trained model.

6. An image processing device according to any one of claims 2 to 5, wherein the second acquisition means further acquires a neck axis.

7. An image processing device according to any one of claims 1 to 6, wherein the size of the cervical region of interest in the input image is preset.

8. An image processing device according to any one of claims 2 to 6, wherein the size of the cervical region of interest in the input image is set by the second acquisition means.

9. An image processing device according to any one of claims 2 to 6 and 8, wherein the first acquisition means further extracts bone regions other than the femur region in the input image, and the bone regions other than the femur region are used to extract a background region.

10. An image processing device as described in any one of claims 2 to 6, 8 and 9, further comprising an operation unit for operating the modification of the acquired femur region, wherein the second acquisition means acquires a neck region of interest in the input image by using the femur region modified using the operation unit as input data for the second trained model.

11. An image processing device as claimed in any one of claims 2 to 6 and 8 to 10, wherein the second acquisition means acquires a neck region of interest in the input image using an image in which information indicating the acquired femur region and the input image are superimposed.

12. An image processing device as claimed in any one of claims 2 to 6 and 8 to 11, wherein the first acquisition means acquires information indicating a single region of the femoral head region and the trochanter region of the input image by using an input image as input data for the first learned model, and acquires information indicating the femoral region in the input image by separately extracting information indicating the femoral head region and information indicating the trochanter region from the information indicating the single region.

13. The image processing device according to claim 12, further comprising an operation unit for operating the modification of the single region.

14. An image processing device as described in claim 12 or 13, wherein the first acquisition means acquires information indicating the femoral head region and information indicating the trochanter region in the acquired single region separately by using information indicating the acquired single region as input data for a trained model obtained using training data including information indicating a femoral region in which the femoral head region and the trochanter region are separately extracted in a single region including the femoral head region and the trochanter region.

15. An image processing device according to claim 12 or 13, wherein the first acquisition means acquires information indicating the femoral head region and information indicating the trochanter region in the single region separately by rule-based processing.

16. The image processing device described in claim 1, wherein the trained model is a trained model that integrates a first trained model obtained using training data including information indicating a neck region of interest including the femoral neck, and a second trained model obtained using training data including information indicating a neck region of interest including the femoral neck.

17. The image processing device according to claim 16, wherein the acquisition means acquires a neck region of interest in the input image using a position and angle of the neck region of interest based on an output from the trained model.

18. An image processing device according to any one of claims 1 to 17, wherein the cervical region of interest in the input image is a target region for bone density measurement.

19. An image processing device as claimed in any one of claims 1 to 18, further comprising: a display control means for causing a display unit to display a cervical region of interest in the input image; and an operation unit that allows an operator to operate a correction of the cervical region of interest in the input image based on the display on the display unit.

20. An image processing device according to any one of claims 1 to 18, further comprising a display control means for displaying a warning on a display unit when the acquired cervical region of interest is determined to be abnormal.

21. An image processing device according to any one of claims 1 to 20, wherein the input image is either an energy subtraction image, at least one of two images acquired using radiation of two types of tube voltages, at least one of two images acquired using a sensor having two types of materials with different radiation absorption rates, or a simple radiation image.

22. The image processing apparatus of claim 21, wherein one of the two types of materials includes CsI and the other of the two types of materials includes GOS.

23. An image processing device according to any one of claims 1 to 22, wherein the femur region includes a proximal femur region.

24. An image processing device according to any one of claims 1 to 23, wherein the acquisition means extracts a reference region including the measurement object from the medical image, and uses an image showing the reference region as the input image.

25. A radiography system comprising: an image processing device according to any one of claims 1 to 24; a radiation generating device that generates radiation; and a radiation detecting device that detects the radiation and generates a radiographic image.

26. A method of operating an image processing device that processes medical images, comprising: obtaining a neck region of interest including the femoral neck in an input image by using an input image as input data for a trained model obtained using training data including information indicating a femur region in an image of the femur and training data including information indicating a neck region of interest including the femoral neck.

27. A program which, when executed by a computer, causes the computer to carry out the method of operating an image processing apparatus according to claim 26.

Images