WO2023153564A1 - Medical image processing method and apparatus - Google Patents

Abstract

According to an embodiment disclosed herein, a medical image processing method performed by a computing device may include the steps of: training a first model on mapping information about a cone-beam CT (CBCT) training image and a bone mineral density-related training image corresponding to the CBCT training image, and thereby inferring the bone mineral density-related training image from the CPCT training image; and training a second model to infer the bone mineral density-related training image from the CBCT training image and an initial bone mineral density-related image output from the first model.

Description

Medical image processing method and apparatus

The present disclosure relates to a method and apparatus for processing a medical image, and more specifically, to a method and apparatus for processing a Cone-Beam CT (CBCT) image.

Bone Mineral Density (BMD) measurement is a method of estimating bone mass to diagnose osteoporosis and predict future fracture risk. Even in dental implant treatment, accurate in vivo measurement of alveolar bone quality is very important in determining the primary stability of the implant. Therefore, it is necessary to diagnose whether alveolar bone mineral density (BMD) is sufficient before implant placement. In general CT images, bone mineral density is quantitatively measured through correction of HU (Hounsfield Units) values, and this method is called QCT (Quantitative CT).

Meanwhile, recently, CBCT images have been widely used for dental diagnosis, treatment, and surgical planning. CBCT images offer advantages such as low radiation dose to the patient, short acquisition time, and high resolution compared to conventional CT images. However, since the voxel value of the CBCT image is arbitrary and does not accurately provide the HU value, it is impossible to accurately measure the bone quality of the alveolar bone using CBCT.

Several studies are being conducted to resolve this HU discrepancy between conventional CT and CBCT data. Representatively, a study was conducted to analyze the relationship between voxel values of CBCT images and HU of general CT images using a bone density calibration phantom (BMD calibration phantom) containing material inserts of various CT attenuation coefficients. However, this is due to the nonuniformity of the voxel value of the CBCT image and the nonlinearity relationship between the voxel value and the HU of the CBCT image, resulting in the fact that it is impossible to correlate the voxel value of the CBCT image with the HU value. stopped

Accordingly, there is a need in the art for a technique for quantitatively and immediately measuring bone mineral density from CBCT images.

The problem to be solved is to provide a method for quantitatively and immediately measuring bone density from a CBCT image having voxel values in a non-linear relationship with bone density values, unlike general CT images. In addition to the above tasks, it may be used to achieve other tasks not specifically mentioned.

A method of processing a medical image performed by a computing device according to some embodiments of the present invention, wherein the method learns mapping information of a Cone-Beam CT (CBCT) image for learning and a bone density-related image for learning corresponding to the CBCT image for learning Thus, learning a first model to infer the bone density-related image for learning from the CBCT image for learning, and inferring the bone density-related image for learning from the initial bone density-related image output from the CBCT image for learning and the first model , training the second model.

The bone density-related image for learning may be either a quantitative CT (QCT) or a CT image.

The first model includes a first generator that learns the mapping information to infer a bone density-related image from a CBCT image, and a first generator that determines whether an input image is a real bone density-related image or a synthesized bone density-related image. It may be learned in a direction of minimizing a result of a first loss function including a discriminator and an adversarial loss function of the first model calculated based on a discriminating success probability of the first discriminator.

The first model includes a second generator that learns the mapping information and infers the CBCT image from the bone density-related image, and a second generator that determines whether an input image is a real CBCT image or a synthesized CBCT image. Further comprising a discriminator, the adversarial loss function of the first model calculated based on the discriminant success probabilities of the first discriminator and the second discriminator, and a cycle consistency loss of the first model It may be optimized in a direction of minimizing a result of the first loss function including a combination of functions.

The first model may be implemented to include a Cycle-GAN (Generative Adversarial Network) structure.

The bone density-related image includes the mapping information on the voxel intensity relationship between the CBCT image for learning and the bone density-related image for learning, and bone contrast is increased with respect to the CBCT image for learning QCT -It may be a QCT-like or CT-like image.

The second model may output a final bone density related image including bone mineral density (BMD) data corresponding to the CBCT image for training.

The second model may be optimized in a direction of minimizing a result of a second loss function including a function related to a difference between the bone density related image for training and the final bone density related image.

The function related to the difference is the mean absolute difference (MAD) of the voxel intensity difference between the training bone density related image and the final bone density related image and the structural similarity between the training bone density related image and the final bone density related image ( Structural similarity; SSIM).

The second model may be implemented to include a multi-channel U-Net structure that receives the CBCT image for learning and the initial bone density-related image through multi-channel.

The final bone density related image may be a QCT-like image or a CT-like image including the mapping information included in the initial bone density related image and anatomical structure information of the CBCT image for learning and suppressing artifacts. .

The method includes obtaining a raw CBCT image of a subject and a raw CT image corresponding to the raw CBCT image, respectively; Bone density of the raw CT image taken under conditions corresponding to the raw CT image Calibrating using a BMD calibration phantom CT image and acquiring the corrected raw CT image as a raw QCT image, removing a non-anatomical region from the raw CBCT image and the raw QCT image, The method may further include registering a raw CBCT image and the raw QCT image, acquiring the raw CBCT image as the CBCT image for training, and obtaining the raw QCT image as the bone density-related image for training. .

The method includes obtaining a raw CBCT image of a subject and a raw CT image corresponding to the raw CBCT image, respectively, removing a non-anatomical region from the raw CBCT image and the raw CT image The step of registering the raw CBCT image and the raw CT image, and acquiring the raw CBCT image as the CBCT image for training and acquiring the raw CT image as the bone density-related image for training can do.

A medical image processing method performed by a computing device according to some embodiments of the present invention, the method comprising: acquiring a Cone-Beam CT (CBCT) image; and inputting the CBCT image to a pre-learned deep learning model. and obtaining a final bone density-related image including Bone Mineral Density (BMD) data corresponding to the CBCT image.

The final bone density related image may be a QCT-like image or a CT-like image.

The method further comprises acquiring the bone density data corresponding to the CBCT image based on the final bone density associated image, wherein the acquiring the bone density data comprises: the final bone density associated image is the QCT-like image If , immediately acquiring the bone density data from the final bone density related image; or, if the final bone density related image is the CT-like image, correcting the final bone density related image to correspond to the CT-like image. The method may include generating a QCT image and acquiring the bone density data from the generated QCT image.

According to some embodiments of the present disclosure, unlike general CT images, bone density can be quantitatively and immediately measured from a CBCT image having voxel values in a non-linear relationship with bone density values.

According to some embodiments of the present disclosure, a synthesized QCT image with improved bone contrast and uniformity in the image is obtained by processing a CBCT image through a deep learning model including a combination of a Cycle-GAN structure and a multi-channel U-Net structure. can

1 is a block diagram illustrating a computing device providing a medical image processing method according to some embodiments of the present disclosure.

2 is a schematic diagram of a medical image processing method according to some embodiments of the present disclosure.

3 is a flowchart of a medical image processing method according to some embodiments of the present disclosure.

4 is a flowchart of a deep learning model learning method for providing a medical image processing method according to some embodiments of the present disclosure.

5 is a schematic diagram of a deep learning model learning method for providing a medical image processing method according to some embodiments of the present disclosure.

6 is a diagram exemplarily illustrating a deep learning model training method for providing a medical image processing method according to some embodiments of the present disclosure.

7 is a diagram illustrating bone density data measurement performance according to the medical image processing method of the present disclosure.

8 is a diagram illustrating bone density data measurement performance according to the medical image processing method of the present disclosure.

9 is a diagram illustrating bone density data measurement performance according to the medical image processing method of the present disclosure.

10 is a flowchart of a method of obtaining training data in a medical image processing method according to some embodiments of the present disclosure.

Hereinafter, with reference to the accompanying drawings, embodiments of the present disclosure will be described in detail so that those skilled in the art can easily carry out the present disclosure. However, the present disclosure may be embodied in many different forms and is not limited to the embodiments described herein. And in order to clearly describe the present disclosure in the drawings, parts irrelevant to the description are omitted, and similar reference numerals are attached to similar parts throughout the specification.

In the present disclosure, when a part "includes" a certain component, it means that it may further include other components without excluding other components unless otherwise stated. Devices constituting the network may be implemented as hardware, software, or a combination of hardware and software.

In the present disclosure, the term “image” may be used as a medical image of a subject, particularly a medical image that is a target of input/output and processing for a deep learning model in the present disclosure. For example, the image may be a CT image, a CBCT image, and/or a QCT image captured of the maxilla region of the subject. Alternatively, the image may be a synthesized QCT image that is output by inputting the CBCT image captured of the subject to a deep learning model. At this time, the output image may be an image including bone density data corresponding to the input CBCT image.

However, it is not limited thereto, and images in the present disclosure include CT images, CBCT images, QCT images, magnetic resonance imaging (MRI), ultrasound images, and endoscopic examination images of any body part of the subject. , thermography images, nuclear medicine images, and/or images output as intermediate or final products by inputting them to a deep learning model.

In the present disclosure, the term “image associated with bone density” may be an image including bone density data of a subject. For example, it may mean an image capable of measuring bone density data of a subject and/or an image usable for measuring bone density data. In particular, in the present disclosure, a bone density related image may be a general CT image or a quantitative CT (QCT) image obtained by correcting a general CT image based on a phantom for correction.

In the present disclosure, the term “model” may be used to mean a deep learning model including one or more neural network structures. In particular, in the present disclosure, a model may mean a deep learning model for obtaining data by processing a medical image. For example, the model may be a deep learning model trained to receive a CBCT image captured of a subject and output an image including bone density data for the input CBCT image. However, it is not limited to this, and the model is a deep learning model trained to receive various medical images other than CBCT images and output arbitrary data such as quantification data, corrected image data, and segmentation data for the input medical images. can

Also, in the present disclosure, a model may be implemented as a combination of one or more neural network structures, for example, a combination of a Cycle-GAN (Generative Adversarial Network) structure and a U-Net structure. However, it is not limited thereto, and in the present disclosure, the model may be implemented with only one of a Cycle-GAN structure and a U-Net structure, and in addition, CNN (Convolutional Neural Network; CNN), RNN (Recurrent Neural Network), LSTM It can be implemented using any neural network structure usable for medical image processing, such as Long Short-Term Memory (Long Short-Term Memory), Autoencoder, and Deep Residual Network (DRN).

In the present disclosure, the term “raw image” may be used as a term referring to a medical image captured of an object under examination, and in particular, a medical image in a state in which image preprocessing is not performed after being captured. For example, it may refer to a medical image that has not been subjected to image preprocessing such as noise removal, registration, cropping, and resize after being photographed with respect to the subject. In the present disclosure, a raw image may be classified into a raw CBCT image, a raw CT image, a raw QCT image, etc. according to the type of a photographed medical image, but is not limited thereto. Meanwhile, in the present disclosure, the following learning data may be obtained through pre-processing of raw images.

In the present disclosure, the term “learning data” may be used as a term meaning data used for learning the above-described model. In the present disclosure, data for learning may be composed of one or more image pairs including a CBCT image and a CT image or QCT image corresponding to the CBCT image, that is, obtained from a subject corresponding to the CBCT image or a corresponding imaging condition. .

In addition, in the present disclosure, data for learning may be composed of an image for learning obtained through pre-processing of the aforementioned raw image. For example, a raw CBCT image may be matched with a corresponding raw QCT image, and the matched raw CBCT image may be obtained as a CBCT image. Alternatively, after noise is removed from the raw CBCT image, the corresponding raw CBCT image may be matched with a corresponding raw QCT image, and the matched raw CBCT image may be obtained as a CBCT image. Similarly, for the raw QCT image obtained from the raw CT image and the corresponding phantom CT image for bone density correction, the matched raw QCT image can be obtained as a QCT image through noise removal and/or matching with the corresponding raw CBCT image. can However, it is not limited thereto.

In the present disclosure, the term “bone density data” refers to data related to bone density of a subject, and in particular, in the present disclosure, it means one of data to be acquired by inputting a medical image to a deep learning model. can do. Specifically, in the present disclosure, the term bone density data may refer to data in a form capable of quantitatively measuring bone density. For example, bone density data in the present disclosure may be a synthesized QCT image obtained by inputting a CBCT image of a subject to a deep learning model, or may be bone density related data obtained based on voxel intensity of the synthesized QCT image. there is. However, it is not limited thereto, and the bone density data may be obtained in any form of data, such as a graph, a heat map, or a table, in addition to an image.

1 is a block diagram illustrating a computing device providing a medical image processing method according to some embodiments of the present disclosure.

Referring to FIG. 1 , a computing device 10 according to the present disclosure includes one or more processors 11, a memory 12 for loading programs executed by the processors 11, and a storage 13 for storing programs and various data. ), and a communication interface 14. However, since the above-described components are not essential to implement the computing device 10 according to the present disclosure, the computing device 10 may have more or fewer components than the components listed above. For example, the computing device 10 may further include an output unit and/or an input unit (not shown), or the storage 13 may be omitted.

A program consists of a series of computer readable instructions grouped together on a functional basis and is executed by a processor. The program may include instructions that, when loaded into memory 12, cause processor 11 to perform methods/operations in accordance with various embodiments of the present disclosure. That is, the processor 11 may perform methods/operations according to various embodiments of the present disclosure by executing instructions.

The processor 11 controls the overall operation of each component of the computing device 10 . The processor 11 may include at least one of a Central Processing Unit (CPU), a Micro Processor Unit (MPU), a Micro Controller Unit (MCU), a Graphic Processing Unit (GPU), or any type of processor well known in the art of the present disclosure. can be configured to include Also, the processor 11 may perform an operation for at least one application or program for executing a method/operation according to various embodiments of the present disclosure.

For example, the processor 11 according to the present disclosure learns mapping information of a Cone-Beam CT (CBCT) image for learning and a bone density-related image for learning, and learns a first model to infer a bone density-related image for learning from the CBCT image for learning, and The second model may be trained to infer a bone density related image for learning from the CBCT image and the initial bone density related image output from the first model.

Alternatively, the processor 11 according to the present disclosure controls the communication interface 14 to be described later to receive a CBCT image for learning and a bone density-related image for learning, or a raw CBCT image and a raw CT received through the communication interface 14 Data for learning may be generated by processing an image or a raw QCT image.

Alternatively, the processor 11 according to the present disclosure controls the communication interface 14 to receive a CBCT image of a subject, inputs the CBCT image to a pre-learned deep learning model stored in memory and/or storage, and calculates the final bone density. A related image may be obtained, and an output unit (not shown) may be controlled to output a final bone density related image to a user terminal. The foregoing examples are only examples in which the processor 11 controls the overall operation of each component of the computing device 10 to provide the medical image processing method according to the present disclosure, and do not limit the present disclosure.

Memory 12 stores various data, commands and/or information. Memory 12 may load one or more programs from storage 13 to execute methods/operations according to various embodiments of the present disclosure. The memory 12 may be implemented with volatile memory such as RAM, but the technical scope of the present disclosure is not limited thereto.

The storage 13 may store programs non-temporarily. The storage 13 is a non-volatile memory such as read only memory (ROM), erasable programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, or the like, a hard disk, a removable disk, or well known in the art. It may be configured to include any known type of computer-readable recording medium.

The storage 13 according to the present disclosure may store, for example, a deep learning model according to the present disclosure learned by the processor 11 described above. Alternatively, the storage 13 may store a program including instructions for training a deep learning model according to the present disclosure. However, it is not limited thereto.

The communication interface 14 according to the present disclosure may be a wired/wireless communication module. The communication interface 14 according to the present disclosure, for example, receives a CBCT image of an arbitrary subject from a server (not shown) or inputs the received CBCT image to the deep learning model to obtain a final bone density related image. It may be transmitted to a server and/or a user terminal, or may receive training data for learning the above-described deep learning model. However, it is not limited thereto.

2 is a schematic diagram of a medical image processing method according to some embodiments of the present disclosure.

Referring to FIG. 2 , according to the medical image processing method of the present disclosure, the computing device 10 inputs a CBCT image 200 to a pre-learned deep learning model 100, and includes bone density data corresponding to the CBCT image. A final bone density-related image 500 may be obtained. Accordingly, ultimately, bone mineral density can be measured immediately and quantitatively from CBCT images. This is due to the fact that the voxel intensity value of the conventional CBCT image is arbitrary and does not correctly provide the HU (Hounsfield Units) value for bone density measurement, despite advantages such as low radiation dose, short acquisition time, and high resolution. This contrasts with those that were difficult to use for measuring bone density.

3 is a flowchart of a medical image processing method according to some embodiments of the present disclosure. FIG. 3 shows the medical image processing method of the present disclosure described in FIG. 2 in detail according to the type of the final bone density related image.

Referring to FIG. 3 , first, the computing device 10 may obtain a CBCT image (S110). For example, the processor 11 may control a camera unit (not shown) to receive a CBCT image captured of the subject, or control the communication interface 14 to receive a CBCT image captured of the subject. .

Next, the computing device 10 may acquire a final bone density-related image including bone density data corresponding to the CBCT image by inputting the CBCT image to the pre-learned deep learning model (S120).

Here, the final bone density related image may be a QCT-like image or a CT-like image corresponding to the previously input CBCT image. Specifically, the type of final bone density-related image obtained by inputting a CBCT image to a pre-learned deep learning model may be different according to the type of training data used to learn the deep learning model.

For example, when a CBCT image and a corresponding QCT image are used as training data for learning a deep learning model, the deep learning model learns mapping information between the CBCT image and the QCT image, and accordingly, when the CBCT image is input, the CBCT image A QCT-like image including bone density data corresponding to may be output as a final bone density related image.

Similarly, when a CBCT image and a corresponding CT image are used as training data for learning a deep learning model, the deep learning model learns mapping information between the CBCT image and the CT image, and accordingly, when the CBCT image is input, the CBCT image A CT-like image including bone density data corresponding to may be output as a final bone density related image.

A method for learning the deep learning model 100 will be described later in detail with reference to FIGS. 4 to 6 and a method for obtaining learning data used for learning the deep learning model 100 with reference to FIG. 10 .

Next, when the final bone density related image is a QCT-like image (S130, Y), the computing device 10 may immediately acquire bone density data from the final bone density related image (S140). For example, bone density data may be immediately measured and obtained from voxel values of a final bone density related image, which is a QCT-like image.

Alternatively, if the final bone density related image is not a QCT-like image but a CT-like image (S130, N), the computing device 10 corrects the final bone density related image to generate a corresponding QCT image, and generates the generated QCT image. Bone density data can be obtained from (S150).

Specifically, when the final bone density related image is a CT-like image, instead of immediately measuring bone density data from the final bone density related image, the CT image is corrected through a phantom for bone density correction to obtain a QCT image, and a QCT image is obtained from the corresponding QCT image. data can be measured. A method of correcting a CT image using a phantom for bone density correction is a method known in the art of the present disclosure, and thus a detailed description thereof will be omitted.

A method for learning the deep learning model 100 according to the present disclosure will be described in detail with reference to FIGS. 4 to 6 below.

4 is a flowchart of a deep learning model learning method for providing a medical image processing method according to some embodiments of the present disclosure.

Referring to FIG. 4 , first, the computing device 10 learns a CBCT image for learning and mapping information of a bone density-related image for learning corresponding to the CBCT image for learning, and trains a first model to infer a bone density-related image for learning from the CBCT image for learning. It can (S210).

Here, the first model 110 is trained to infer a bone density-related image for learning corresponding to the CBCT image for learning from the CBCT image for learning, and outputs an initial bone density-related image 400 according to the input of the CBCT image.

Here, the initial bone density related image 400 may include voxel intensity mapping information between an input CBCT image for learning and a corresponding bone density related image for learning. Here, the bone density-related image for learning corresponding to the CBCT image for learning may be, for example, a QCT image or a CT image that can be obtained from the same subject and/or under the same imaging conditions as the CBCT image for learning. That is, the initial bone density-related image 400 may include information about a voxel intensity mapping (or correspondence) relationship between a CBCT image and a QCT image or CT image that can be obtained from the same subject and/or under the same imaging conditions.

As such, since the bone density distribution information of the bone density related image corresponding to the CBCT image is reflected in the initial bone density related image 400, the bone contrast (the difference in luminance between the bone and the surrounding tissue in the medical image) with respect to the CBCT image is may have a tendency to increase. That is, the initial bone density related image 400 may be a QCT-like image or a CT-like image in which bone contrast is increased with respect to the input CBCT image.

For example, when the first model 110 is trained to infer a bone density related image for learning from a CBCT image for learning using a CBCT image for learning and a bone density related image for learning, which is a corresponding QCT image, as learning data, the initial bone density related image 400 may be a QCT-like image. Similarly, when the first model 110 is trained to infer a bone density related image for learning from a CBCT image for learning by using a CBCT image for learning and a bone density related image for learning, which is a corresponding CT image, as learning data, the initial bone density related image (400 ) may be a CT-like image.

Next, the computing device 10 may train the second model to infer a bone density related image for learning from the CBCT image for training and the initial bone density related image output from the first model (S220).

Here, the second model 120 is trained to infer a bone density related image for learning corresponding to the CBCT image for learning from the CBCT image for learning and the initial bone density related image 400, and outputs the final bone density related image 500. The second model 120 may be connected in series with the first model 110 to form the deep learning model 100 according to the present disclosure as an integral body. However, it is not limited thereto.

The final bone density related image 500 output from the second model 120 includes voxel intensity mapping information included in the initial bone density related image 400 and anatomical structure information of the input CBCT image, and artifacts ) may be suppressed QCT-like images or CT-like images.

For example, when the second model 120 is trained to infer a bone density related image for learning corresponding to a CBCT image for learning from the initial bone density related image 400, which is a CBCT image for learning and a QCT-like image, the final bone density related image 500 is It may be a QCT-like image. Similarly, when the second model 120 is trained to infer a training bone density related image corresponding to a training CBCT image from the training CBCT image and the initial bone density related image 400 that is a CT-like image, the final bone density related image 500 may be a CT-like image.

On the other hand, the image with artifact suppression here is an image from which obstructive elements such as artifacts and scattering noise are removed, and an image with improved uniformity, which means the degree to which the HU value of each pixel of the image for a uniform object is uniform can mean

Also, as described above, the final bone density-related image 500 may be an image including bone density data corresponding to the input CBCT image. Accordingly, bone density data corresponding to the CBCT image 200 may be obtained from the final bone density related image 500 obtained by inputting the CBCT image 200 to the deep learning model 100 . For example, information on the bone density of a subject whose CBCT image has been taken may be obtained immediately and quantitatively from the final bone density related image 500, which is a QCT-like image.

5 is a schematic diagram of a deep learning model learning method for providing a medical image processing method according to some embodiments of the present disclosure. Specifically, FIG. 4 schematically shows the learning process of the first model 110 and the second model 120 of the present disclosure described above with reference to FIG. 3 .

First, the first model 110 receives an image pair of a CBCT image 300 for learning and a bone density related image 310 for learning corresponding thereto, and maps information between the CBCT image 300 for learning and the bone density related image 310 for learning learning, and based on this, it may be learned to infer the bone density related image 310 for learning from the CBCT image 300 for learning. Accordingly, the first model 110 may output an initial bone density related image 400 according to the input of the CBCT image 300 for learning.

Here, the first model 110 may be optimized based on at least a determination success probability for discriminating an actual bone density related image among the output initial bone density related image 400 and the training bone density related image 310 . For example, when the learning bone density related image 310 is a QCT image corresponding to the learning CBCT image 300 and the initial bone density related image 400 is a QCT-like image, the first model 110 is an initial bone density related image ( 400) and the bone density related image 310 for learning, the actual QCT image can be optimized based on at least the probability of success in discriminating. Similarly, when the bone density related image 310 for learning is a CT image corresponding to the CBCT image 300 for learning and the initial bone density related image 400 is a CT-like image, the first model 110 is an initial bone density related image ( 400) and the bone density-related image 310 for learning, it can be optimized based on at least the probability of success in discriminating an actual CT image.

In this case, a first loss function may be used to optimize the first model 110 . The value of the first loss function may be calculated by including a plurality of factors including a probability of success in discriminating between the initial bone density related image 400 and the bone density related image 310 for learning. Calculation of the first loss function will be described later with reference to FIG. 6 .

Next, the second model 120 receives the CBCT image 300 for learning and the initial bone density related image 400 output from the first model 110, and learns to infer the bone density related image 310 for learning from them It can be. Accordingly, the second model 120 may output a final bone density related image 500 including bone density data corresponding to the CBCT image 300 for training.

Here, the second model 120 may be optimized based on the difference (or loss) between the bone density related image 310 for training and the output final bone density related image 500 . In this case, a second loss function may be used to optimize the second model 120 . The value of the second loss function may include a function related to a difference between the bone density related image 310 for training and the output final bone density related image 500 .

Specifically, the value of the second loss function is the mean absolute difference (MAD) of the voxel intensity difference between the bone density related image 310 for training and the output final bone density related image 500, and the bone density related image for training ( 310) and structural similarity (SSIM) between the output final bone density associated image 500. Calculation of the second loss function will be described later with reference to FIG. 6 .

6 is a diagram exemplarily illustrating a deep learning model training method for providing a medical image processing method according to some embodiments of the present disclosure.

Referring to FIG. 6 , the first model 110 may be implemented to include a Cycle-GAN (Generative Adversarial Network) structure, and the second model 120 is a CBCT image for learning (multi-channel) 300) and an initial bone density related image 400 may be implemented to include a multi-channel U-Net structure.

Regarding the first model 110 first, a general GAN structure consists of a generator and discriminator trained simultaneously. The generator aims to trick the discriminator into classifying the similar image as a real image by creating a real-like image, while the discriminator aims to classify real-like images from each other.

In comparison, Cycle-GAN adds a constraint condition that the loss between the generated image and the actual image must be minimized. Accordingly, the Cycle-GAN structure further includes a generator and a discriminator that restores the generated image closer to the actual image, that is, performs an inverse transformation process, and thus consists of two generators and two discriminators. In this way, Cycle-GAN can limit the model to double by doubling the process of a general GAN by applying inverse transformation and increase the accuracy of the output image.

Accordingly, the first model 110 learns the mapping information to infer the bone density related image 400 for learning from the CBCT image 300 for learning. The second generator 112 learned to infer the CBCT image 300, the first discriminator 115 learned to determine whether the input image is an actual bone density related image or a synthesized bone density related image, and the input A second discriminator 116 trained to determine whether the image is a real CBCT image or a synthesized CBCT image may be included. However, it is not limited thereto, and the first model 110 may be composed of only the first generator 111 and the first discriminator 115 among the above-described configurations. In this case, the first model 110 may be implemented to include a GAN structure instead of a Cycle-GAN structure.

Specifically, referring to the upper left part of FIG. 6 , the first generator 111 may receive the CBCT image 300 for learning and output an initial bone density related image 400, and the second generator 113 may output the first bone density related image 400. The restored CBCT image 301 may be output by receiving the initial bone density related image 400 output from the generator 111 . Here, the restored CBCT image 301 may be an image obtained by reconstructing the initial bone density related image 400 to be close to the CBCT image 300 for learning input to the first generator 111 .

Next, referring to the lower left portion of FIG. 6 , the second generator 112 may receive the bone density related image 310 for learning and output a synthesized CBCT image 311, and the first generator 114 may output the second The synthesized CBCT image 311 output from the generator 112 may be input and a restored bone density related image 312 may be output. Here, the restored bone density related image 312 may be an image obtained by reconstructing the synthesized CBCT image 311 to be close to the bone density related image 310 for learning input to the second generator 112 .

Next, referring to the central left part of FIG. 6 , the first discriminator 115 receives an initial bone density related image 400 and/or a bone density related image 310 for learning, and determines whether the input image is an actual bone density related image. Alternatively, it may be determined whether the image is a synthesized bone density related image.

For example, when the learning bone density related image 310 is a QCT image corresponding to the learning CBCT image 300 and the initial bone density related image 400 is a QCT-like image, the first model 110 is an initial bone density related image ( 400) and the bone density related image 310 for learning are input, and it is possible to determine whether the input image is an actual QCT image or a synthesized QCT image.

Similarly, when the bone density related image 310 for learning is a CT image corresponding to the CBCT image 300 for learning and the initial bone density related image 400 is a CT-like image, the first model 110 is an initial bone density related image ( 400) and the bone density related image 310 for learning are input, and it is possible to determine whether the input image is an actual CT image or a synthesized CT image.

Meanwhile, the first discriminator 115 may output 1 when it determines that the input image is an actual bone density-related image, and outputs 0 when it determines that the input image is a synthesized bone density-related image. Accordingly, if the first discriminator 115 outputs 0 when the initial bone density-related image 400 is input and outputs 1 when the bone density-related image 310 for learning is input, it is determined that the discrimination was successful. It can be.

Meanwhile, the second discriminator 116 may receive the synthesized CBCT image 311 and/or the CBCT image 300 for learning, and determine whether the input image is an actual CBCT image or a synthesized CBCT image. Specifically, the second discriminator 116 may output 0 when it determines that the input image is an actual CBCT image, and outputs 1 when it determines that the input image is a synthesized CBCT image. Accordingly, if the second discriminator 116 outputs 1 when the synthesized CBCT image 311 is input and outputs 0 when the CBCT image 300 for learning is input, it can be determined that discrimination has succeeded. there is.

On the other hand, in order to optimize the first model 110, an adversarial loss function calculated based on the discrimination success probabilities of the above-described first discriminator 115 and the second discriminator 116 may be used. . That is, the adversarial loss function calculated based on the output of the discriminator may constitute a first loss function for optimizing the first model 110 together with a cycle consistency loss function to be described later.

The adversarial loss function constituting the first loss function may be defined as follows for each of the first discriminator 115 and the second discriminator 116 . Here, D _QCT is the first discriminator 115, D _CBCT is the second discriminator 116, I _QCT is the bone density-related image 310 for learning, I _CBCT is the CBCT image 300 for learning, and G _CBCT→QCT is the first 1 generator 111, G _QCT→CBCT refer to the second generator 112, respectively:

[Equation 1]

L _ADV (G _CBCT→QCT )=D _QCT (I _QCT ) ² +(D _QCT (G _CBCT→QCT (I _CBCT ))-1) ²

[Equation 2]

L _ADV (G _QCT→CBCT )=D _CBCT (I _CBCT ) ² +(D _CBCT (G _QCT→CBCT (I _QCT ))-1) ²

When the first discriminator 115 has the highest probability of success in discriminating, the value of the adversarial loss function (Equation 1) for the first discriminator 115 may be zero. When the second discriminator 116 has the highest probability of success in discriminating, the value of the adversarial loss function (Equation 2) for the second discriminator 116 may be zero.

Meanwhile, the first loss function for optimizing the first model 110 may include a cycle coherence loss function in addition to the above-described adversarial loss function.

The cycle coherence loss function may be a function that calculates a loss by comparing an 'image obtained by reconstructing a synthesized image close to a real image' and a previously input 'real image'. In other words, it may be a function indicating how closely the synthesized image output by inputting the real image to the generator is restored to the real image.

The cycle consistency loss function of the first model 110 is the loss (or difference) between the restored CBCT image 301 and the training CBCT image 300, and the restored bone density related image 312 and the training bone density related image ( 310) can be calculated based on the loss between

The cyclic coherence loss function constituting the first loss function may be defined as:

[Equation 3]

_LCYC =

G _QCT→CBCT (G _CBCT→QCT (I _CBCT) )-I _CBCT

+

G _CBCT→QCT (G _QCT→CBCT (I _QCT ))-I _QCT

Looking at Equation 3, the first term to the right of the equal sign is the loss between the restored CBCT image 301 and the learning CBCT image 300, and the second term is the restored bone density related image 312 and the learning bone density related image 310 indicates liver loss.

As the second generator 113 restores the initial bone density-related image 400 closer to the training CBCT image 300 input to the first generator 111, the value of the first term of Equation 3 will approach 0, similarly As the first generator 114 reconstructs the synthesized CBCT image 311 closer to the bone density-related image 310 for training input to the second generator 112, the value of the second term of Equation 3 may also approach 0. .

Taken together, the first loss function of the first model 110, as a combination of the adversarial loss function and the cyclic coherence loss function, can be defined as follows. Here, λ is a weight, and controls the relative importance of the adversarial loss function of the first model 110. For example, λ may be 10. However, it is not limited thereto, and any suitable λ value can be selected in silico or experimentally:

[Equation 4]

L _GAN =L _ADV (G _CBCT→QCT) +L _ADV (G _QCT→CBCT )+λL _CYC

The first model 110 may be optimized based on the result of a first loss function comprising a combination of an adversarial loss function and a cyclic coherence loss function. As described above, the higher the discrimination success probability of the first discriminator 115 and the second discriminator 116, and the higher the performance of the first generators 111 and 114 and the second generators 112 and 113 The higher the restoration performance (herein, in particular, the higher the restoration performance), the resultant value of the first loss function of the first model 110 may be reduced. In other words, the first model 110 may be optimized in a direction in which the result of the first loss function is minimized.

Meanwhile, in each of the generators 111, 112, 113, and 114 constituting the first model 110, convolution blocks are 7x7 and 3x3 convolutions to which batch normalization and ReLU activation are applied. It is composed of solution layers, and residual blocks may be inserted between down-sampling layers and up-sampling layers.

Each of the generators 111, 112, 113, and 114 constituting the first model 110 may be implemented as a ResNet structure including, for example, 9 residual blocks. Through the residual block included in each generator, the network learns the difference between the source and the target, and through this, the initial bone density including more accurate voxel intensity mapping information between the CBCT image 300 for training and the bone density related image 400 for training A related image 400 may be generated. However, it is not limited thereto.

Meanwhile, in each of the discriminators 115 and 116, the convolution block may be composed of 4x4 convolution layers to which batch normalization and Leaky ReLu activation are applied, and subsequent down-sampling layers. Each discriminator may be implemented as PatchGAN, which discriminates whether an image generated by a generator is genuine or not, in units of patches of a specific size, for example. For example, it can be implemented with PatchGAN of 70 x 70 patches. However, it is not limited thereto.

In summary, the first model 110 according to some embodiments of the present disclosure, that is, the Cycle-GAN structure, includes training data including a CBCT image 300 for training and a bone density related image 310 for training corresponding thereto, and a residual block Since it is implemented using , it is possible to learn voxel intensity mapping information between the CBCT image 300 for learning and the bone density-related image 310 for learning. Accordingly, the bone density distribution information of the learning bone density related image 310 corresponding to the learning CBCT image is reflected in the initial bone density related image 400, so the bone contrast increases with respect to the learning CBCT image 300 It can be.

The first model 110 as described above works complementarily with the second model 120 to be described later, ultimately resulting in improved bone contrast and uniformity from the CBCT image and a final bone density-related image containing highly accurate bone density data. (500) can be obtained.

Next, in relation to the second model 120, a general U-Net structure includes a U-shaped encoder and decoder structure. Specifically, U-Net extracts the overall context of the input image in the contracting path of the encoder, and on the other hand, in the expanding path of the decoder, the contracting path of the contracting path through skip connection. The context extracted at the same level is combined with the pixel location information. That is, localization accuracy may be increased by combining a part of the result output from the encoding region with the decoding region.

The second model 120 according to the present disclosure may include a multi-channel input for receiving the CBCT image 300 for training and the initial bone density related image 400 output from the first model 110 . The second model 120 can be trained to infer the training bone density related image 310 from the training CBCT image 300 and the initial bone density related image 400, and includes bone density data corresponding to the training CBCT image 300 A final bone density related image 500 may be output.

The second model 120 according to the present disclosure can simultaneously learn spatial information of the CBCT image 300 for learning and the initial bone density-related image 400 corresponding to the CBCT image 300 through multi-channel input. Specifically, the second model 120 can learn voxel intensity mapping information between the CBCT image and the bone density related image included in the initial bone density related image 400 while maintaining the anatomical structure of the CBCT image 300. Accordingly, artifacts and scattering noise are suppressed, and the final bone density-related image 500 with improved uniformity can be output.

The second model 120 according to the present disclosure may be optimized based on a result of a second loss function including a function of a difference between the bone density related image 310 for learning and the final bone density related image 500. Specifically, the function for the difference between the learning bone density related image 310 and the final bone density related image 500 is the average value absolute deviation (Mean It may include at least one of Absolute Difference (MAD) and Structural Similarity (SSIM).

Here, the average absolute deviation (MAD) is defined as an average of absolute differences in intensity between the training bone density-related image 310 and the final bone density-related image 500, and can be expressed as follows. where I _QCBCT refers to the final bone density associated image 500:

[Equation 5]

L _MAD =

I _QCT -I _QCBCT

On the other hand, the structural similarity (SSIM) is a function used to measure the similarity between two images and can be expressed as follows. where μ is the mean, σ ² is the variance, and C ₁ and C ₂ correspond to variables for stabilizing the weak denominator, respectively:

[Equation 6]

Accordingly, the second loss function of the second model 120 may be defined as follows. Here, α is a weight, for example, α may be 0.6. However, it is not limited thereto, and any suitable value of α can be selected in silico or experimentally:

[Equation 7]

As such, by training the second model 120 using the second loss function including a combination of mean value absolute deviation (MAD) and structural similarity (SSIM), considering structural similarity and pixel-wise errors, more You can get faster convergence and higher accuracy.

Meanwhile, the multi-channel U-Net structure of the second model 120 according to the present disclosure includes, in addition to the above-described multi-channel input, an encoder and a decoder composed of 3 x 3 convolutional layers to which batch normalization and ReLU activation are applied, and each layer level A skip connection in may be further included. 6 shows an example of a multi-channel U-Net structure implemented to include four skip connections for each level of an encoder and a decoder. However, it is not limited thereto.

The second model 120 according to the present disclosure is implemented to include a multi-channel U-Net structure and is capable of extracting both global and local features for a spatial region of an image, which is an initial bone density related image. The final bone density-related image 500 including the mapping information included in 400 and the anatomical structure information of the CBCT image 300 for learning and suppressing artifacts can be output.

Accordingly, the deep learning model 100 according to the present disclosure including the combination of the first model 110 and the second model 120 improves bone contrast and uniformity with respect to the input CBCT image, while the input A final bone density-related image 500 including bone density data for the CBCT image may be output. Accordingly, the medical image processing method according to the present disclosure may ultimately provide a method of quantitatively and immediately measuring bone mineral density from a CBCT image.

Experimental data described later through FIGS. 7 to 9 are provided using the deep learning model 100 in which the first model 110 (Cycle-GAN) and the second model 120 (multi-channel U-Net) are combined. The difference in performance of the medical image processing method according to the present disclosure with respect to the conventional method of correcting a CBCT image using a model implemented with only one of Cycle-GAN and U-Net and/or a phantom for BMD correction is shown. .

7 to 9 are diagrams illustrating performance of measuring bone density data according to the medical image processing method of the present disclosure. Figures 7-9 show, in particular, the bone density data measurement performance for some embodiments of the present disclosure where the final bone density related image is a QCT-like image.

Here, bone density data measurement performance may mean bone density data measurement performance for each medical image. For example, as described at the beginning of the present disclosure, compared to the fact that bone mineral density can be quantitatively measured through the correction of HU values in general CT images, in the case of conventional CBCT images, non-linearity between the non-uniform characteristics of voxel values and HU values Because of this relationship, it may not be possible to measure immediately meaningful bone density data from conventional CBCT images. In this case, the performance of measuring bone density data for a conventional CBCT image can be regarded as lower than the performance of measuring bone density data for a general CT image.

First, referring to FIG. 7 , FIG. 7 is a diagram quantitatively illustrating the performance of measuring bone density data according to the medical image processing method of the present disclosure, and specifically, a final bone density related image (500, hereinafter “QCBCT image”) according to the present disclosure. This, actual QCT image (ie, ground truth), image obtained by inputting CBCT image to Cycle-GAN model (hereinafter, “CYC_CBCT image”), image obtained by inputting CBCT image to U-Net model (hereinafter, “U_CBCT image”) and a quantitative performance difference with respect to a CBCT image (hereinafter, “CAL_CBCT image”) corrected using a phantom for BMD correction.

For quantitative performance comparison, in FIG. 7 , Mean Absolute Deviation (MAD), Peak Signal to Noise Ratio (PSNR), Normalized Cross Correlation (NCC), Structural Similarity (SSIM), and Spatial Non-uniformity ( Spatial Non-Uniformity (SNU), and slope (Slope of Linear regression) values were used.

Here, the peak signal-to-noise ratio (PSNR) is defined as the logarithm of the Maximum Possible Intensity (MAX) for the Root Mean Squared Error (MSE) between the actual QCT image and the image to be compared, as follows: can be expressed as:

[Equation 8]

Here, the normalized cross-correlation (NCC) is defined as a value obtained by dividing the product of the actual QCT image by the intensity of the image to be compared by the standard deviation of each, and can be expressed as follows:

[Equation 9]

Here, spatial non-uniformity (SNU) is defined as an absolute value of a difference between a maximum value and a minimum intensity value in rectangular ROIs (Regions Of Interest) set in an image.

Here, the slope, as an index for evaluating the linearity of bone density, is a value obtained by analyzing the relationship between voxel intensity between an actual QCT image and an image to be compared through linear regression of intensity values within an image.

Peak signal-to-noise ratio (PSNR), structural similarity (SSIM), normalized cross-correlation (NCC) and slope (Slope) are higher, mean absolute deviation (MAD) and spatial non-uniformity (SNU) are lower values, bone density measurement It can be interpreted as better performance.

Referring back to FIG. 7 , the final bone density-related image (QCBCT image in FIG. 7 , see rows 3 and 9) according to the present disclosure is compared with the actual QCT image, the average absolute deviation (MAD), peak signal-to-noise ratio (PSNR) ), structural similarity (SSIM), normalized cross correlation (NCC), and slope (Slope) values, CYC_CBCT images, U_CBCT images, and CAL_CBCT regardless of the location of the captured bone (eg, maxilla or mandible) It appears to greatly outperform the video.

That is, the bone density measurement performance of the medical image processing method according to the present disclosure is equivalent to the bone density of the conventional method of correcting a CBCT image using a model implemented with only one of Cycle-GAN and U-Net and/or a phantom for BMD correction. Compared with the measurement performance, it shows a significant difference.

Next, referring to FIG. 8, FIG. 8 is a diagram qualitatively showing the performance of measuring bone density data according to the present disclosure, and specifically, an actual QCT image for each of the maxilla and mandible, and a final bone density related image (or, in FIG. QCBCT image), CYC_CBCT image, U_CBCT image and CAL_CBCT image (see rows 1 and 3 of FIG. 8 below), and images obtained by subtracting each image from the actual QCT image (subtraction images, see rows 2 and 4 of FIG. 8 ) is shown.

As shown in the subtracted images of FIG. 8, the quality of the bone density image of the final bone density related image (see column 2) according to the present disclosure is compared with the actual QCT image (see column 1), the CYC_CBCT image (see column 3) ), shows significant improvement for U_CBCT images (see column 4) and CAL_CBCT images (see column 5). In particular, when compared with the CAL_CBCT image, in the final BMD-related image according to the present disclosure, a large bone density (voxel intensity value) difference in the tooth region found in the CAL_CBCT image and a dense bond of high BMD values are greatly reduced. can see what has happened

Next, referring to FIG. 9, FIG. 9 shows an actual QCT image, a final bone density related image according to the present disclosure (see column 1), a CYC_CBCT image (see column 2), a U_CBCT image (see column 3), and a CAL_CBCT image (see column 4). Reference) shows a linear relationship between each. Specifically, the first row (a to d) of FIG. 9 is an image obtained for the upper jaw region under the condition of 80 kVp - 8 mA, and the second row (e to h) is the image obtained for the lower jaw region under the condition of 80 kVp - 8 mA. Images acquired for the region, 3 rows (i to l) show images acquired for the upper jaw region under the condition of 90 kVp - 10 mA, and 4 rows (m to p) show images acquired for the lower jaw region under the condition of 90 kVp - 10 mA. do.

As shown in FIG. 9 , the linear relationship between the actual QCT image and the final bone density-related image according to the present disclosure has a higher slope and goodness of fit than the linear relationship between the actual QCT image and other images. Show contrast and correlation. In addition, the tendency of this linear relationship appears consistently regardless of the location of the bone to be imaged (eg, upper jaw or lower jaw) or imaging conditions (eg, 80 kVp-18 mA or 90 kVp-10 mA).

When synthesizing the above contents through FIGS. 2 to 9, the medical image processing method according to the present disclosure quantitatively evaluates bone density from CBCT images through a deep learning model including a combination of Cycle-GAN and multi-channel U-Net. It can provide a measurable method that can be measured both locally and immediately.

Specifically, the first model (ie, Cycle-GAN structure) according to the present disclosure includes mapping information between a CBCT image and a bone density-related image that is a QCT image or a CT image, and bone collation is performed for the CBCT image input to the first model. An initial bone density related image, which is a QCT-like image or a CT-like image with increased bone contrast, may be output. The second model (i.e., multi-channel U-Net structure) of the present disclosure combined with the first model receives an initial bone density related image and a CBCT image, and provides mapping information included in the initial bone density related image and anatomical analysis of the CBCT image A final bone density related image, which is a QCT-like image or a CT-like image including structural information, suppressed artifacts and increased uniformity, may be output.

Accordingly, the medical image processing method according to the present disclosure can obtain a final bone density-related image with significantly improved anatomical accuracy and quantitative accuracy in the uniformity and contrast of the bone image from the CBCT image, ultimately reducing radiation dose, It is possible to provide a method capable of quantitatively and immediately measuring bone mineral density from CBCT images while using CBCT imaging techniques having advantages such as short acquisition time and high resolution.

10 is a flowchart of a method of obtaining training data in a medical image processing method according to some embodiments of the present disclosure. Specifically, FIG. 10 shows a method of acquiring a CBCT image 300 for learning and a bone density related image 310 for learning, respectively, in order to train the deep learning model 100 according to the present disclosure as described above with reference to FIGS. 4 to 6 do.

Referring to FIG. 10 , the computing device may first acquire a raw CBCT image of a subject under examination and a raw CT image corresponding to the raw CBCT image (S310).

Here, the subject may be, for example, phantoms of human skulls, and in this case, the skull phantom may include at least one of those with and without a metal restoration that causes artifacts. However, the present invention is not limited thereto, and an object for learning data acquisition may be one or three or more phantoms for an arbitrary part.

For the subject, a raw CBCT image and a raw CT image may be respectively captured. In this case, the raw CBCT image and the raw CT image can be captured under the same conditions, such as voltage and current. However, the present invention is not limited thereto, and a raw CBCT image and a raw CT image may be captured under different conditions for the subject and then aligned through a step such as matching.

For example, a raw CT image is a voxel size of 0.469 x 0.469 x 0.5 mm ³ , dimensions of 512 x 512 pixels, depth of 16 bits, voltage of 120 (kVp), and current of 130 (mA). On the other hand, a raw CBCT image is obtained for the subject with a voxel size of 0.3 x 0.3 x 0.3 mm ³ , dimensions of 559 x 559 pixels, depth of 16 bits, voltage of 80 or 90 (kVp), and current of 8 or 10 (mA) can be taken under the condition.

In an embodiment in which a QCT image corresponding to a CBCT image for training is used as a bone density-related image for training, the computing device 10 may next obtain a raw QCT image by correcting the raw CT image (S320).

As described above, the QCT image is an image obtained through correction of the HU value in a general CT image, and in this case, a CT image of a phantom for bone density correction may be used for correction.

More specifically, a raw CT image may be corrected using a phantom CT image for bone density correction taken under conditions corresponding to the raw CT image, and the corrected raw CT image may be obtained as a raw QCT image. That is, referring to the above example, using the phantom CT image for bone density correction taken under the conditions of voltage 120 (kVp) and current 130 (mA) with respect to the bone density correction phantom, voltage 120 (kVp) and current 130 (mA) It is possible to correct the raw CT imaged under the condition of and obtain a raw QCT image. However, it is not limited thereto.

Next, the computing device 10 may remove non-anatomical elements from the raw CBCT image and the raw QCT image (S330). By removing non-anatomical regions from the learning images, it is possible to prevent adverse impacts such as a decrease in accuracy due to the non-anatomical regions in the learning process.

More specifically, a non-anatomical region may be removed from the raw CBCT image and the raw QCT image by applying a binary mask to the raw CBCT image and the raw QCT image, respectively.

Here, the binary mask image can be created using thresholding and morphological operations. Specifically, edges of anatomical regions may be extracted by applying a local range filter to each of the training CBCT image and the corresponding training QCT image forming a training image pair. Next, morphological operations of opening and flood fill can be applied to the binarized edges obtained through thresholding to remove small blobs and fill the inner regions.

The raw CBCT image and the raw QCT image may be multiplied by the intersection of two binary masks generated from the raw CBCT image and the raw QCT image through the above-described operation. Meanwhile, voxel values outside the masked area may be replaced with -1000 HU. The above is only an example of a method for removing non-anatomical regions from learning images, and the present disclosure is not limited thereto.

Next, the computing device 10 may match the raw CBCT image and the raw QCT image (S340).

Specifically, a raw QCT image may be matched to a raw CBCT image by paired-point registration. At this time, a plurality of landmarks may be set for matching. 6, including, for example, the vertex on the lateral incisors, the buccal cusps of the first premolars, and the distobuccal cusps of the first molars. Dog landmarks may be set. However, it is not limited thereto.

Meanwhile, the computing device may additionally perform a step of cropping the matched raw CBCT image and the raw QCT image into an image of an arbitrary size, and then resizing the matched raw CBCT image into an image of a preset size. For example, from the matched low CBCT image and low QCT image, a 559 x 559 x 200 pixel image centered on the maxillary region can be cropped and then resized to a 256 x 256 x 200 pixel image. .

Next, the computing device 10 may acquire the matched raw CBCT image as a CBCT image for learning, and acquire the matched raw QCT image as a bone density-related image for learning (S350).

Alternatively, when cropping and/or resizing are additionally performed on the matched raw CBCT image and raw QCT image as described above, the cropped and/or resized raw CBCT image and raw QCT image are the CBCT image for learning and the bone density for learning, respectively. It can be obtained as a related image.

In contrast, in an embodiment in which a CT image corresponding to a CBCT image for learning is used as a bone density-related image for learning, the computing device 10 removes a non-anatomical region from the acquired raw CBCT image and raw CT (S360), The raw CBCT image from which the non-anatomical region has been removed and the raw CT image may be matched (S370), and the registered CBCT image may be acquired as a CBCT image for learning, and the registered raw CT image may be obtained as a bone density-related image for learning ( S380). Since the specific method of removing the non-anatomical region for each image and the specific method of matching each image with each other have been described above, detailed descriptions are omitted here to avoid redundant description.

The acquired CBCT images for learning and the corresponding BMD-related images for learning may form a pair of training images together, and a set of one or more pairs of training images may constitute training data according to the present disclosure. The learning data obtained in this way may be used for learning the deep learning model 100 according to the present disclosure described above through FIGS. 4 to 6 .

The embodiments of the present disclosure described above are not implemented only through devices and methods, and may be implemented through a program that realizes functions corresponding to the configuration of the embodiments of the present disclosure or a recording medium on which the program is recorded.

Although the embodiments of the present disclosure have been described in detail above, the scope of the present disclosure is not limited thereto, and various modifications and improvements of those skilled in the art using the basic concepts of the present disclosure defined in the following claims are also included in the present disclosure. that fall within the scope of the right.

Claims (16)

1. A medical image processing method performed by a computing device,

   Learning a first model to infer the bone density-related image for learning from the CBCT image for learning by learning mapping information of a Cone-Beam CT (CBCT) image for learning and a bone density-related image for learning corresponding to the CBCT image for learning; and

   Learning a second model to infer the bone density related image for learning from the CBCT image for learning and the initial bone density related image output from the first model,

   Medical image processing method.
2. In claim 1,

   The bone density related image for learning,

   Either quantitative CT (QCT) or CT images,

   Medical image processing method.
3. In claim 1,

   The first model,

   A first generator for inferring a bone density-related image from a CBCT image by learning the mapping information, and a first discriminator for determining whether an input image is an actual bone density-related image or a synthesized bone density-related image,

   Learning in the direction of minimizing the result of the first loss function including the adversarial loss function of the first model calculated based on the discrimination success probability of the first discriminator,

   Medical image processing method.
4. In paragraph 3,

   The first model,

   A second generator for inferring the CBCT image from the bone density-related image by learning the mapping information, and a second discriminator for determining whether an input image is an actual CBCT image or a synthesized CBCT image,

   Including a combination of the adversarial loss function of the first model and the cycle consistency loss function of the first model calculated based on the discrimination success probabilities of the first discriminator and the second discriminator Optimized in the direction of minimizing the result of the first loss function,

   Medical image processing method.
5. In claim 1,

   The first model,

   Implemented to include a Cycle-GAN (Generative Adversarial Network) structure,

   Medical image processing method.
6. In claim 1,

   The bone density related image,

   Includes the mapping information on the voxel intensity relationship between the learning CBCT image and the learning bone density related image, and has increased bone contrast with respect to the learning CBCT image QCT-like ) or CT-like images,

   Medical image processing method.
7. In claim 1,

   The second model,

   Outputting a final bone density related image including Bone Mineral Density (BMD) data corresponding to the CBCT image for learning,

   Medical image processing method.
8. In paragraph 7,

   The second model,

   Optimized in the direction of minimizing the result of a second loss function including a function related to the difference between the learning bone density related image and the final bone density related image,

   Medical image processing method.
9. In paragraph 8,

   The function related to the difference is the mean absolute difference (MAD) of the voxel intensity difference between the training bone density related image and the final bone density related image and the structural similarity between the training bone density related image and the final bone density related image ( Structural similarity; SSIM), including at least one of

   Medical image processing method.
10. In claim 1,

    The second model,

    Implemented to include a multi-channel U-Net structure that receives the CBCT image for learning and the initial bone density related image through multi-channel,

    Medical image processing method.
11. In paragraph 7,

    The final bone density related image,

    QCT-like image or CT-like image including the mapping information included in the initial bone density related image and anatomical structure information of the CBCT image for learning and artifacts suppressed,

    Medical image processing method.
12. In claim 1,

    Obtaining a raw CBCT image of a subject and a raw CT image corresponding to the raw CBCT image; A phantom for bone density correction (BMD) photographed under conditions corresponding to the raw CT image Calibration phantom) Calibration using a CT image, and obtaining the corrected raw CT image as a raw QCT image;

    removing non-anatomical regions from the raw CBCT image and the raw QCT image;

    registering the raw CBCT image and the raw QCT image; and

    Acquiring the raw CBCT image as the CBCT image for learning and acquiring the raw QCT image as the bone density related image for learning,

    Medical image processing method.
13. In claim 1,

    Obtaining a raw CBCT image of a subject under examination and a raw CT image corresponding to the raw CBCT image, respectively;

    removing a non-anatomical region from the raw CBCT image and the raw CT image;

    registering the raw CBCT image and the raw CT image; and

    Acquiring the raw CBCT image as the CBCT image for learning and acquiring the raw CT image as the bone density-related image for learning,

    Medical image processing method.
14. A medical image processing method performed by a computing device,

    Obtaining a Cone-Beam CT (CBCT) image, and

    Acquiring a final bone density-related image including bone mineral density (BMD) data corresponding to the CBCT image by inputting the CBCT image to a pre-learned deep learning model,

    Medical image processing method.
15. In claim 14,

    The final bone density related image,

    QCT-like images or CT-like images,

    Medical image processing method.
16. In paragraph 15,

    The method,

    Acquiring the bone density data corresponding to the CBCT image based on the final bone density related image;

    Obtaining the bone density data,

    If the final bone density related image is the QCT-like image, immediately acquiring the bone density data from the final bone density related image; or

    If the final bone density related image is the CT-like image, generating a QCT image corresponding to the CT-like image by correcting the final bone density related image, and obtaining the bone density data from the generated QCT image including,

    Medical image processing method.

Images