KR20220135683A - Apparatus for Denoising Low-Dose CT Images and Learning Apparatus and Method Therefor - Google Patents

Abstract

In accordance with the present invention, provided are a low-dose CT image noise reduction device and a learning device and a method therefor. The low-dose CT image noise reduction device includes a denoising neural network embodied as a pre-trained artificial neural network to output a denoising CT image by removing a noise of an applied low-dose CT image in accordance with a learned scheme. As the denoising neural network separately applies a denoising CT image and a normal-dose CT image, outputted from the applied low-dose CT image, to a lesion identification neural network, which is an artificial neural network pre-trained to identify images containing lesions during a learning procedure, the lesion identification neural network is trained to reduce an observation loss calculated with a difference between first and second feature maps acquired during the lesion identification procedure from the denoising CT image and the normal-dose CT image, thereby effectively reducing noise without causing a blur in the low-dose CT image or changing a pixel value.

Description

Apparatus for Denoising Low-Dose CT Images and Learning Apparatus and Method Therefor

The present invention relates to an apparatus for reducing image noise and a learning apparatus and method therefor, and to a low-dose CT image noise reduction apparatus and a learning apparatus and method for the same.

Computed tomography (CT) is a useful medical imaging technique for non-invasively acquiring high-resolution cross-sectional images of the human body to observe internal tissues such as bones, organs, and blood vessels. However, since there is a problem that radiation exposure is applied due to the use of X-rays, a low-dose irradiation method that physically reduces the X-ray radiation dose has been proposed as one of the methods to reduce the radiation dose during computed tomography.

However, when a low-dose irradiation method is used, there is a problem in that noise is increased in the CT image. As such, the noise generated in the image deteriorates the image quality of the CT image and becomes a factor that prevents the diagnosis of a tumor having a small contrast or a small size in the image.

Accordingly, recently, a technique for removing noise using an artificial neural network such as a convolutional neural network (CNN) has been mainly used. However, the noise removal technique using the artificial neural network shows excellent noise removal performance, but as a result of removing the noise, the image is blurred or the characteristics of the pixel values of the image are changed to remove the texture and fine structures of the image. It causes another problem.

Korean Patent Registration No. 10-2165915 (Registered on 10.07. 2020)

An object of the present invention is to provide a low-dose CT image noise reduction apparatus capable of reducing noise without causing blur or changing pixel values in a low-dose CT image, and a learning apparatus and method therefor.

Another object of the present invention is to provide an apparatus for reducing noise in a low-dose CT image, which is learned using a separate neural network previously learned for tumor detection, and effectively removes noise in a low-dose CT image, and a learning apparatus and method therefor.

To achieve the above object, the low-dose CT image noise reduction apparatus according to an embodiment of the present invention is implemented with a pre-trained artificial neural network, and removes the applied noise of the low-dose CT image according to the learned method to obtain a denoising CT image. and a denoising neural network to output, wherein the denoising neural network outputs a denoising CT image and a normal dose from a low-dose CT image applied to a lesion identification neural network, which is an artificial neural network pre-trained to identify an image containing a lesion in the learning process. By applying each CT image, the lesion identification neural network reduces the observation loss calculated by the difference between the first feature map and the second feature map obtained in the lesion identification process from the denoising CT image and the normal-dose CT image. is learned

The denoising neural network includes: an encoder that receives the low-dose CT image, extracts features of the low-dose CT image according to a learned method, and obtains an encoding feature map; and a decoder that receives the encoding feature map and recovers the CT image from the encoded feature map according to a learned method to obtain the denoising CT image.

The lesion identification neural network receives each of the denoising CT image and the normal dose CT image, extracts features of the denoising CT image according to a pre-learned method to obtain the first feature map, and the normal dose CT a feature extractor configured to extract features of an image to obtain the second feature map; and the first feature map and the second feature map are respectively applied, and it is determined whether predetermined lesions are detected in the first feature map and the second feature map according to a pre-learned method, and the denoising CT image and a lesion identification unit for classifying each of the normal-dose CT images into an image including a lesion and an image not including a lesion.

The denoising neural network is

와

는 각각 디노이징 CT 영상과 정상 선량 CT 영상을 나타내고,

와

는 각각 제1 및 제2 특징맵의 픽셀값을 나타낸다. 그리고 W, H, C는 각각 제1 및 제2 특징맵의 폭, 높이 및 채널 길이를 나타내고, N은 특징맵의 개수를 나타낸다.)에 따라 계산되는 상기 관측 손실(L)이 기지정된 기준 손실 이하가 되도록 반복 학습될 수 있다.(here

Wow

denotes a denoising CT image and a normal-dose CT image, respectively,

Wow

denotes pixel values of the first and second feature maps, respectively. And W, H, and C represent the width, height, and channel length of the first and second feature maps, respectively, and N represents the number of feature maps.) It can be repeated learning so as to become the following.

A learning apparatus for a low-dose CT image noise reduction apparatus according to another embodiment of the present invention for achieving the above object is implemented as an artificial neural network and removes noise of the applied low-dose CT image according to the learned method to obtain a denoising CT image. As a learning device for a low-dose CT image noise reduction device that outputs, the denoising CT image and the normal-dose CT image output from the low-dose CT image noise reduction device are respectively applied and learned by being implemented with a pre-trained artificial neural network. a lesion identification neural network that detects whether a predetermined lesion is included in each of the denoising CT image and the normal-dose CT image according to a method, and identifies an image including the lesion and an image not including the lesion; and the lesion identification neural network calculates an observation loss as a difference between a first feature map and a second feature map obtained in a lesion identification process from the denoising CT image and the normal-dose CT image, and reduces the observation loss. A low-dose CT image noise reduction device includes a learning unit for backpropagating the observation loss.

To achieve the above object, a learning method for a low-dose CT image noise reduction apparatus according to another embodiment of the present invention is implemented with an artificial neural network to remove noise from the applied low-dose CT image according to the learned method to obtain a denoising CT image. A learning method for a low-dose CT image noise reduction apparatus that outputs The denoising CT image and the normal-dose CT image output from the low-dose CT image noise reduction device are respectively applied and learned using an artificial neural network that has been previously trained to identify an image containing a lesion by receiving a CT image. detecting whether a predetermined lesion is included in each of the denoising CT image and the normal-dose CT image according to the method to identify an image including the lesion; and calculating an observation loss with a difference between a first feature map and a second feature map obtained from the denoising CT image and the normal-dose CT image in the process of identifying the image including the lesion, and backpropagating the observation loss with the low-dose CT image noise reduction device to reduce the observation loss.

Accordingly, the low-dose CT image noise reduction apparatus and the learning apparatus and method for the same according to an embodiment of the present invention can effectively reduce noise without causing blur or changing pixel values in a low-dose CT image.

1 shows a schematic structure of a learning apparatus for a low-dose CT image noise reduction apparatus according to an embodiment of the present invention.
FIG. 2 shows an example of the structure of the denoising neural network of FIG. 1 .
3 shows an example of the structure of the lesion identification neural network of FIG. 1 .
4 shows the performance of the low-dose CT image noise reduction apparatus according to the present embodiment.
5 shows a learning method for a low-dose CT image noise reduction apparatus according to an embodiment of the present invention.

In order to fully understand the present invention, the operational advantages of the present invention, and the objects achieved by the practice of the present invention, reference should be made to the accompanying drawings illustrating preferred embodiments of the present invention and the contents described in the accompanying drawings.

Hereinafter, the present invention will be described in detail by describing preferred embodiments of the present invention with reference to the accompanying drawings. However, the present invention may be embodied in several different forms, and is not limited to the described embodiments. In addition, in order to clearly explain the present invention, parts irrelevant to the description are omitted, and the same reference numerals in the drawings indicate the same members.

Throughout the specification, when a part "includes" a certain component, it does not exclude other components unless otherwise stated, meaning that other components may be further included. In addition, terms such as "... unit", "... group", "module", and "block" described in the specification mean a unit that processes at least one function or operation, which is hardware, software, or hardware. and a combination of software.

1 shows a schematic structure of a learning apparatus for a low-dose CT image noise reduction apparatus according to an embodiment of the present invention, FIG. 2 shows an example of the denoising neural network structure of FIG. 1, and FIG. 3 is the lesion of FIG. An example of an identification neural network structure is shown.

Referring to FIG. 1 , a learning apparatus for a low-dose CT image noise reduction apparatus according to the present embodiment may include a denoising neural network 100 , a learning unit 200 , and a lesion identification neural network 300 .

Here, the denoising neural network 100 is a low-dose CT image noise reduction device of the present invention, which receives a low-dose CT image, removes noise included in the applied low-dose CT image, and outputs it.

The low-dose CT image noise reduction device according to the present embodiment can be implemented with various artificial neural networks, but here, as an example, the denoising neural network 100 is a convolutional neural network (hereinafter referred to as a convolutional neural network), in particular a residual encoder decoder convolutional neural network ( Residual Encoder-Decoder Convolutional Neural Network (hereinafter referred to as RED-CNN) is assumed to be implemented.

The RED-CNN may include an encoder 110 and a decoder 120 as shown in FIG. 2 . The encoder 110 receives a low-dose CT image as an input image, extracts features of the low-dose CT image according to a learned method, and obtains an encoding feature map. Then, the decoder 120 receives the encoding feature map and recovers the CT image from the encoded feature map according to the learned method to obtain the denoising CT image.

The encoder 110 includes a plurality of encoding layers 111 to 11n each including a convolution layer and an activation function, and the decoder 120 includes a plurality of encoding layers of the encoder 110 . Corresponding to (111 to 11n), a plurality of decoding layers 121 to 12n each including a deconvolution layer and an activation function may be included. The plurality of encoding layers 111 to 11n receive an input image or an encoding feature map output from a previously arranged encoding layer, extract features, and output an encoding feature map. In addition, the plurality of decoding layers 121 to 12n receives and restores the encoding feature map finally output from the encoder 110 or the decoding feature map output from the previously arranged decoding layer, and outputs the decoding feature map.

At this time, in the RED-CNN, at least one decoding layer among the plurality of decoding layers 121 to 12n of the decoder 120 receives the input image or the encoding feature map output from a predetermined encoding layer through a shortcut connection. can The decoding layer receiving the input image or encoding feature map through a shortcut connection is applied by concatenating with the encoding feature map finally output from the encoder 110 or the decoding feature map output from the previously arranged decoding layer can receive

And the decoded feature map finally output from the decoder 120 is output as a denoising CT image in which noise is removed from the low-dose CT image input to the denoising neural network 100 .

Here, it is assumed that the denoising neural network 100 is implemented with RED-CNN, but the denoising neural network 100 may be implemented with various artificial neural networks such as a VGG neural network.

As such, when the low-dose CT image noise reduction device is implemented as an artificial neural network, the low-dose CT image noise reduction device can be trained in advance according to the deep learning technique, but noise can be removed from the low-dose CT image. Accordingly, the learning unit 200 learns the denoising neural network 100, which is a low-dose CT image noise reduction device.

First, the learning unit 200 receives a denoising CT image that is output by inputting a low-dose CT image to the denoising neural network 100 as a learning target, and applies the applied denoising CT image to the pre-learned lesion identification neural network 300 . Accordingly, the feature map obtained in the process of the lesion identification neural network 300 identifying whether the denoising CT image is a lesion image is acquired as the first feature map. In addition, the learning unit 200 applies a normal-dose CT image corresponding to the low-dose CT image to the lesion identification neural network 300 to determine whether the lesion identification neural network 300 is a lesion image of the denoising CT image. A feature map obtained in the process of identifying ? is obtained as a second feature map.

Here, the normal-dose CT image is a CT image obtained by irradiating an X-ray dose in a predetermined normal range rather than a low dose, and may be a CT image obtained by photographing the same object as the low-dose CT image. That is, in the present invention, learning is performed using a low-dose CT image and a normal-dose CT image as a learning data set.

And the learning unit 200 calculates the loss of the denoising neural network 100 by using a perceptual loss function that minimizes the difference between the first feature map and the second feature map, and uses the calculated loss in the denoising neural network. By backpropagating to (100), the denoising neural network 100, which is a low-dose CT image noise reduction device, is trained. Here, the loss calculated by the learning unit 200 using the perceptual loss function may be referred to as an observation loss.

The lesion identification neural network 300 receives a CT image, detects whether a predetermined lesion is included in the applied CT image, and recognizes a CT image including a lesion and a CT image not including a lesion. to be.

Currently, various types of lesion identification neural networks for detecting various lesions from CT images have already been developed. These lesion identification neural networks may be learned differently and individually according to each lesion, but they have a common point in that they are trained to detect a lesion from a CT image. Here, as an example, it is assumed that the lesion identification neural network 300 is a neural network trained for tumor detection.

Referring to FIG. 3 , the lesion identification neural network 300 may include a feature extraction unit 310 and a lesion identification unit 320 . The feature extraction unit 310 may include a plurality of feature extraction layers each including a convolution layer and an activation function, and at least one pooling layer for pooling a feature map output from a previously disposed feature extraction layer.

Each of the plurality of feature extraction layers receives an input CT image or a feature map output from a previously arranged feature extraction layer, extracts features according to weights set by learning, and outputs a feature map. In addition, at least one pooling layer is disposed between predetermined feature extraction layers among a plurality of feature extraction layers and pools a feature map output from the directly disposed feature extraction layer to change the size of the feature map. As an example, FIG. 3 illustrates a case in which a pooling layer is respectively disposed after the fourth and eighth feature extraction layers, and it is assumed that the two pooling layers perform max pooling, respectively.

The lesion identification unit 320 receives the feature map finally output from the feature extraction unit 310 and identifies whether the lesion is included in the applied feature map according to a pre-learned method, and the applied CT image includes the lesion. It classifies whether an image is included or not. The lesion identification unit 320 may include, for example, at least one fully connected layer. In FIG. 3 , as an example, the lesion identification unit 320 is illustrated as being composed of one fully connected layer, but the lesion identification unit 320 may be composed of a plurality of fully connected layers.

As described above, in the present embodiment, the learning unit 200 applies each of the low-dose CT image and the normal-dose CT image to the lesion identification neural network 300 . And the lesion identification neural network 300 extracts the features of the low-dose CT image and acquires one of a plurality of feature maps obtained in the process of identifying whether the lesion is included as the first feature map. In addition, the learning unit 200 acquires as a second feature map one of a plurality of feature maps obtained when the lesion identification neural network 300 extracts features of a normal-dose CT image to identify whether a lesion is included. At this time, the learning unit 200 preselects one feature extraction layer among a plurality of feature extraction layers of the feature extraction unit 310 of the lesion identification neural network 300, and sets the feature map output from the selected feature extraction layer as the first feature. It can be obtained as a map and a second feature map. That is, the first feature map and the second feature map only have a difference in whether the input CT image is a denoising CT image or a normal dose CT image, and a feature map obtained through the same feature extraction process in the lesion identification neural network 300 . to be.

When the first feature map and the second feature map are obtained, the learning unit 200 calculates and obtains the observation loss (L) from the obtained first feature map and the second feature map according to Equation 1 to obtain the denoising neural network ( 100) to backpropagate.

와

는 각각 디노이징 CT 영상과 정상 선량 CT 영상을 나타내고,

와

는 각각 제1 및 제2 특징맵의 픽셀값을 나타낸다. 그리고 W, H, C는 각각 제1 및 제2 특징맵의 폭, 높이 및 채널 길이를 나타내고, N은 특징맵의 개수를 나타낸다.)(here

Wow

denotes a denoising CT image and a normal-dose CT image, respectively,

Wow

denotes pixel values of the first and second feature maps, respectively. And W, H, and C represent the width, height, and channel length of the first and second feature maps, respectively, and N represents the number of feature maps.)

The learning unit 200 checks whether the observed loss (L) obtained according to Equation 1 is less than or equal to a predetermined reference loss, and if the observation loss (L) is not less than or equal to the reference loss, denoising the low-dose CT image again in the neural network 100 ) to obtain a denoising CT image, and input the denoising CT image and the normal dose CT image to the lesion identification neural network to obtain first and second feature maps to calculate the observation loss (L). That is, the learning unit 200 may repeatedly learn the denoising neural network 100 until the observation loss L is equal to or less than the reference loss. In some cases, the learning unit 200 may be set to repeat learning a predetermined number of times regardless of the reference loss.

As described above, the denoising neural network 100, which is a low-dose CT image noise reduction device, can be implemented with various artificial neural networks, and the low-dose CT image noise reduction device implemented with the artificial neural network must be learned in advance according to the deep learning technique. do. Therefore, conventionally, artificial neural networks for removing noise were trained by setting the mean-squared-error (MSE) as a loss function. However, MSE has the problem of blurring the image because pixel-level loss ignores small changes in pixel values.

As an alternative to solving the image blurring problem, a method of performing learning by extracting image features and setting a perceptual loss function that minimizes the feature difference between images has been proposed. And here, the pre-trained VGG neural network was mainly used as a network for extracting image features. However, in the case of the existing VGG neural network, it was learned based on RGB images, not CT images, and RGB images have different properties from CT images. Extract features that are not. Therefore, the low-dose CT image noise reduction device learned by applying the perceptual loss function based on the features extracted from the VGG neural network has different characteristics from the CT image of the noise-reduced denoising CT image to detect lesions, the original purpose of the CT image. There is a problem that makes it difficult.

Accordingly, the learning apparatus for a low-dose CT image noise reduction apparatus according to the present invention acquires the first and second feature maps using the pre-trained lesion detection neural network 300 to detect a lesion in a CT image, and performs a perceptual loss function. By learning the low-dose CT image noise reduction device using the difference between the first and second feature maps calculated by In addition, the accuracy of the noise-removed low-dose CT image can be maintained by preventing the characteristics of the low-dose CT image from being changed at the pixel level.

4 shows the performance of the low-dose CT image noise reduction apparatus according to the present embodiment.

In FIG. 4, (a) is an input image, a low-dose CT image, containing noise, and (b) is a denoising CT image from which noise has been removed by a low-dose CT image noise reduction device learned by applying MSE loss. , (c) shows the denoising CT image from which noise has been removed by the low-dose CT image noise reduction device learned based on the feature map obtained from the VGG neural network. And (d) shows the denoising CT image from which the noise is removed by the low-dose CT image noise reduction device learned by the learning device for the low-dose CT image noise reduction device, and (e) is the normal-dose CT image that is the ground truth. show the image.

4, it can be seen that the low-dose CT image of (a) before noise removal contains a lot of noise, and the denoising CT image of (b) has much reduced noise compared to (a), but instead of the image It can be seen that there is a large amount of blur in the area and the details are not clear. And in the case of (c), as the VGG neural network learned based on the RGB image was used for learning, the pixel characteristics were changed, and compared to the normal-dose CT image in (e), the contrast of some pixels was increased and the brightness became brighter. can However, as shown in (d), it can be seen that the low-dose CT image noise reduction apparatus learned as a learning apparatus for the low-dose CT image noise reduction apparatus of this embodiment was obtained very similarly to the normal-dose CT image, which is the truth value.

5 shows a learning method for a low-dose CT image noise reduction apparatus according to an embodiment of the present invention.

1 to 3, the learning method for the low-dose CT image noise reduction device of FIG. 5 is described. First, a low-dose CT image is input to the low-dose CT image noise reduction device implemented with an artificial neural network (S10). Here, an artificial neural network implemented as a low-dose CT image noise reduction device is not limited. That is, various artificial neural networks can be used as a low-dose CT image noise reduction device.

The low-dose CT image noise reduction apparatus obtains a denoising CT image that is output by removing noise from the input low-dose CT image according to the method learned so far (S20).

When the denoising CT image is obtained, the denoising CT image is received and applied as an input to the lesion identification neural network, which is an artificial neural network previously trained to identify the image including the lesion (S30). And in the process of identifying the image containing the lesion by detecting whether a predetermined lesion is included in the denoising CT image according to the learning method of the lesion identification neural network, the feature map generated by extracting the features of the denoising CT image is first generated It is acquired as a feature map (S40).

Meanwhile, a normal-dose CT image obtained in advance as a truth value corresponding to the low-dose CT image is applied as an input to the lesion identification neural network (S50). And in the process of identifying the image containing the lesion by detecting whether a predetermined lesion is included in the normal-dose CT image according to the learned method of the lesion identification neural network, a feature map generated by extracting the features of the normal-dose CT image is generated as a second It is acquired as a feature map (S60).

When the first and second feature maps are obtained, the obtained first and second feature maps are substituted in Equation 1 to calculate the observation loss (L) and backpropagated to a low-dose CT image noise reduction device to reduce low-dose CT image noise The device is trained (S70). And it is determined whether the calculated observed loss (L) is less than or equal to a predetermined reference loss (S80). If it is determined that the observed loss L is less than or equal to the reference loss or the number of repetition learning reaches the predetermined reference number, learning is terminated (S90). However, when the observation loss (L) exceeds the reference loss, the low-dose CT image is again input to the low-dose CT image noise reduction device to perform repeated learning.

The method according to the present invention may be implemented as a computer program stored in a medium for execution by a computer. Here, the computer-readable medium may be any available medium that can be accessed by a computer, and may include all computer storage media. Computer storage media includes both volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data, and read dedicated memory), RAM (Random Access Memory), CD (Compact Disk)-ROM, DVD (Digital Video Disk)-ROM, magnetic tape, floppy disk, optical data storage, and the like.

Although the present invention has been described with reference to the embodiment shown in the drawings, which is only exemplary, those skilled in the art will understand that various modifications and equivalent other embodiments are possible therefrom.

Accordingly, the true technical protection scope of the present invention should be defined by the technical spirit of the appended claims.

100: denoising neural network 110: encoder
120: decoder 200: learning unit
300: lesion identification neural network 310: feature extraction unit
320: lesion identification unit

Claims (14)

It is implemented as a pre-learned artificial neural network, and includes a denoising neural network that outputs a denoising CT image by removing the noise of the applied low-dose CT image according to the learned method,
The denoising neural network is
In the learning process, the denoising CT image and the normal-dose CT image output from the low-dose CT image applied to the lesion identification neural network, which is a pre-trained artificial neural network to identify the image containing the lesion, are respectively applied, and the lesion identification neural network is A low-dose CT image noise reduction device trained to reduce an observation loss calculated as a difference between a first feature map and a second feature map obtained in the lesion identification process from the easing CT image and the normal-dose CT image. The method of claim 1, wherein the denoising neural network
an encoder that receives the low-dose CT image and extracts features of the low-dose CT image according to a learned method to obtain an encoding feature map; and
and a decoder receiving the encoding feature map and recovering the CT image from the encoded feature map according to a learned method to obtain the denoising CT image. According to claim 1, wherein the lesion identification neural network
The denoising CT image and the normal-dose CT image are respectively applied, and the features of the denoising CT image are extracted according to a pre-learned method to obtain the first feature map, and the features of the normal-dose CT image are extracted a feature extracting unit to obtain the second feature map; and
The first feature map and the second feature map are respectively applied, and it is determined whether predetermined lesions are detected in the first feature map and the second feature map according to a pre-learned method, and the denoising CT image and the A low-dose CT image noise reduction apparatus comprising a lesion identification unit for dividing each of the normal-dose CT images into an image including a lesion and an image not including a lesion. The method of claim 1, wherein the denoising neural network
formula

(here

Wow

denotes a denoising CT image and a normal-dose CT image, respectively,

Wow

denotes pixel values of the first and second feature maps, respectively. And W, H, and C represent the width, height, and channel length of the first and second feature maps, respectively, and N represents the number of feature maps.)
A low-dose CT image noise reduction device that is repeatedly learned so that the observed loss (L) calculated according to the reference loss becomes less than or equal to a predetermined reference loss. The method of claim 2, wherein the denoising neural network
The encoder includes a plurality of convolutional layers, and the decoder includes a plurality of deconvolution layers, so that the output of at least one predetermined convolutional layer among the plurality of convolutional layers is the plurality of deconvolutional layers. A low-dose CT image noise reduction device implemented as a Residual Encoder-Decoder Convolutional Neural Network (RED-CNN) applied to at least one predetermined deconvolutional layer among The method of claim 3, wherein the lesion identification neural network
A low-dose CT image noise reduction device, which is a tumor identification neural network that receives a CT image and is trained to detect and classify the presence of a tumor in the applied CT image. A learning device for a low-dose CT image noise reduction device that is implemented with an artificial neural network and outputs a denoising CT image by removing noise from the applied low-dose CT image according to a learned method,
Implemented with a pre-learned artificial neural network, the denoising CT image and the normal-dose CT image output from the low-dose CT image noise reduction device are respectively applied, and the denoising CT image and the normal-dose CT image are respectively applied according to the learned method a lesion identification neural network that detects whether a predetermined lesion is included in the lesion and identifies an image with and without a lesion; and
The lesion identification neural network calculates an observation loss as a difference between a first feature map and a second feature map obtained in the lesion identification process from the denoising CT image and the normal-dose CT image, and the low dose so that the observation loss is reduced A learning apparatus for a low-dose CT image noise reduction apparatus comprising a learning unit for backpropagating the observation loss to a CT image noise reduction apparatus. The method of claim 7, wherein the lesion identification neural network
The denoising CT image and the normal-dose CT image are respectively applied, and the features of the denoising CT image are extracted according to a pre-learned method to obtain the first feature map, and the features of the normal-dose CT image are extracted a feature extracting unit to obtain the second feature map; and
The first feature map and the second feature map are respectively applied, and it is determined whether predetermined lesions are detected in the first feature map and the second feature map according to a pre-learned method, and the denoising CT image and the A learning apparatus for a low-dose CT image noise reduction apparatus comprising a lesion identification unit for classifying each of the normal-dose CT images into an image including a lesion and an image not including a lesion. The method of claim 7, wherein the learning unit
The observed loss (L) is expressed by the equation

(here

Wow

denotes a denoising CT image and a normal-dose CT image, respectively,

Wow

denotes pixel values of the first and second feature maps, respectively. And W, H, and C represent the width, height, and channel length of the first and second feature maps, respectively, and N represents the number of feature maps.)
A learning apparatus for a low-dose CT image noise reduction apparatus that calculates and backpropagates according to the method and repeatedly learns the low-dose CT image noise reduction apparatus so that the observed loss is less than or equal to a predetermined reference loss. The apparatus of claim 7, wherein the low-dose CT image noise reduction device is
an encoder that receives the low-dose CT image and extracts features of the low-dose CT image according to a learned method to obtain an encoding feature map; and
and a decoder that receives the encoding feature map and recovers the CT image from the encoded feature map according to a learned method to obtain the denoising CT image. A learning method for a low-dose CT image noise reduction device that outputs a denoising CT image by removing noise from a low-dose CT image implemented with an artificial neural network according to a learned method,
acquiring the denoising CT image by applying the low-dose CT image to the low-dose CT image noise reduction device;
The denoising CT image and the normal-dose CT image output from the low-dose CT image noise reduction device are respectively applied and learned using an artificial neural network that has been previously trained to identify an image containing a lesion by receiving a CT image. detecting whether a predetermined lesion is included in each of the denoising CT image and the normal-dose CT image according to the method to identify an image including the lesion; and
In the process of performing the step of identifying the image including the lesion, the observation loss is calculated by the difference between the first and second feature maps obtained from the denoising CT image and the normal-dose CT image, and the observation A learning method for a low-dose CT image noise reduction apparatus comprising the step of backpropagating the observation loss to the low-dose CT image noise reduction apparatus to reduce the loss. The method of claim 11, wherein the step of identifying the image including the lesion comprises:
The denoising CT image and the normal-dose CT image are respectively applied, and the features of the denoising CT image are extracted according to a pre-learned method to obtain the first feature map, and the features of the normal-dose CT image are extracted to obtain the second feature map; and
The first feature map and the second feature map are respectively applied, and it is determined whether predetermined lesions are detected in the first feature map and the second feature map according to a pre-learned method, and the denoising CT image and the and dividing each of the normal-dose CT images into an image including a lesion and an image not including a lesion. The method of claim 11, wherein the step of backpropagating
The observed loss (L) is expressed by the equation

(here

Wow

denotes a denoising CT image and a normal-dose CT image, respectively,

Wow

denotes pixel values of the first and second feature maps, respectively. And W, H, and C represent the width, height, and channel length of the first and second feature maps, respectively, and N represents the number of feature maps.)
A learning method for a low-dose CT image noise reduction device that calculates and backpropagates according to The method of claim 11, wherein the acquiring of the denoising CT image comprises:
applying the low-dose CT image to the low-dose CT image noise reduction device;
obtaining an encoding feature map by extracting features of the low-dose CT image according to the learned method; and
A learning method for a low-dose CT image noise reduction apparatus comprising the step of recovering a CT image from the encoding feature map according to a learned method to obtain the denoising CT image.

Images