KR20190029173A - Method and device for classifying medical ultrasound image based on deep learning using smart device - Google Patents

Abstract

According to an embodiment of the present invention, a method for classifying a medical ultrasound image based on deep learning using a smart device can guarantee the performance in real time. To this end, the method comprises the following steps: receiving an ultrasound reflecting signal reflected by a target object through a probe to generate input data; sharing a rendering period performing an image recovery process for the input data to simultaneously perform deep neural network (DNN) learning; and classifying the body structure of an ultrasound image according to a DNN learning execution result.

Description

TECHNICAL FIELD [0001] The present invention relates to a medical ultrasound image classification method and apparatus using a smart device,

The present invention relates to a deep learning-based medical ultrasound image classification method and apparatus using a smart device, and more particularly, to a smart device having a mobile GPU, which reduces idle cores of the GPU in the ultrasound image classification, The present invention relates to a deep-running-based medical ultrasound image classification method and apparatus using a smart device capable of minimizing the size of a body part and ensuring accuracy and real-time performance of body part classification.

In recent years, the technology of diagnosing and medical imaging, which is realized by converging IT technology in medical field, is rapidly developing. Medical ultrasound, which is mainly used, is used to visualize tomographic images in real time to judge the size, structure and pathological damage of muscles, tendons, internal organs and organs of human body.

In order to implement diagnostic medical images using medical ultrasound waves, an ultrasound wave is generated by placing a probe on an object, and ultrasound waves reflected from the object are received to form an image. In other words, when an ultrasonic wave is generated, a sound wave passes through the medium in a very short time, and a reflected wave is generated when the acoustic impedance passes between two different mediums. The diagnostic medical image is generated by measuring the reflected wave and inversing the distance through the time until the reflected sound returns.

As diagnostic medical images are widely used in the medical field, there is a growing demand for a technique for generating and recognizing diagnostic medical images using ultrasound signals on a mobile device regardless of time and place through a smart device. 1 shows an example of an ultrasound medical imaging apparatus using a smartphone developed at present.

However, in the case of a smart device used for portability, when the ultrasonic signal is processed by the processing power of the internal CPU, it is difficult to provide a frame rate image useful for actual diagnosis.

In addition, there is a problem in that the size of the screen of the smart device is smaller than that of the cart-type device, and it is difficult to control parameters of all the images that can be used in the existing equipment due to lack of space on the screen.

KR 10-2012-0059740 (ultrasound system having at least one GPU, Samsung Medison Co., Ltd.) 2012.06.11.

The object of the present invention is to provide a smart device for generating medical ultrasound images, which is capable of reducing the idle core of a graphics processing unit (GPU) and enhancing the accuracy of body part discrimination and real- Based ultrasound image classification method and apparatus.

 The deep learning-based medical ultrasound image classification method using a smart device according to an embodiment of the present invention includes a step of receiving ultrasound reflected signals reflected from a target object through a probe to generate input data, And simultaneously classifying the body structure of the ultrasound image according to the result of the learning of the depth neural network.

According to an embodiment of the present invention, the step of sharing the rendering cycle and performing the depth learning of the neural network concurrently includes: a second step of performing black hole filtering on log compression data subjected to envelope detection and log compression on the ultrasound signal input to the smart device; Sharing a render cycle to concurrently operate image resizing and a first hidden layer on the log compressed data; and sharing a third render cycle to enhance the edges of the black hole filtered data by the black hole filtering, Concurrently operating a third hidden layer by sharing a fourth render cycle that performs a scan conversion on the enhanced data and concurrently operating a third hidden layer, Based on the neural net learning network, By including the step of performing, and wherein the smart device running simultaneously a body part classification to the depth-based neural network learning in the rendering cycle shared by sharing rendering cycle to execute the B- mode image restoration process having a mobile GPU.

According to one embodiment, the method further comprises a data preprocessing step for providing the data set of the depth neural network learning, wherein the preprocessing step comprises the steps of: inversely transforming the image data obtained from the probe into raw form data; And setting the transformed primitive data as a data set of the neural network learning. Thus, data of different standards obtained from different probes are converted into one standardized data and used for learning of the neural network.

According to an exemplary embodiment, concurrently executing depth learning neural network learning by sharing a rendering period includes performing the black-hole filtering, the image resizing and hidden layer 1, the edge enhancement and second hidden layer, And sharing execution by the GPU to each render cycle as independent executions.

According to one embodiment, the smart device generates an ultrasound image by performing an initialization process, a render scene process, and a graphics pipeline process on the beamformed input data, The graphics pipeline performs the image restoration process using a graphics processing unit (GPU) in an OpenGL ES environment.

According to an exemplary embodiment, the image reconstruction process may include performing envelope detection and log compression on the input data in a first render cycle, and performing gain control on the ultrasound image to adjust the overall gain of the ultrasound image. Performing the black hole filtering in the second render cycle, the edge enhancement in the third render cycle, and the digital scan conversion (DSC) in the fourth render cycle.

 The deep learning-based medical ultrasound image classification apparatus using a smart device according to an embodiment of the present invention includes an ultrasound receiver for receiving ultrasound reflection signals reflected from a target object through a probe and generating input data, And the image processing unit performs the body part classification on the basis of the deep neural network learning in the shared rendering period at the same time.

According to one embodiment, the image processing unit may share a second render cycle for performing black-hole filtering on the log-compressed data subjected to envelope detection and log compression on the input data, so as to perform image resizing for the log- Run hidden layers simultaneously.

According to an exemplary embodiment, the image processing unit concurrently executes a second hidden layer by sharing a third render cycle for enhancing edges of data in which black holes are removed by the black hole filtering.

According to one embodiment, the image processing unit concurrently executes the third hidden layer by sharing a fourth render cycle for performing the scan conversion on the enhanced data.

According to an embodiment, the image processing unit may include a smart GPU having a mobile GPU sharing a rendering period for performing an image restoration process, and performing a depth-based neural network learning based on the first hidden layer To perform body part classification simultaneously.

According to an embodiment of the present invention, the depth neural network learning classifies input data by learning a depth neural network structure based on a data set, and is a data preprocessing for providing the data set, Data by converting the data of the different specifications obtained from different probes into one standardized data by setting the inverse transformed raw data to the data set of the deep neural network learning, .

According to one embodiment, the image processing unit may perform the black hole filtering, the image resizing and the first hidden layer, the edge enhancement and the second hidden layer, the scan conversion, and the execution of the third hidden layer by the GPU as independent executions Share it in the render cycle.

According to one embodiment, the image processing unit performs an initialization process, a render scene process, and a graphics pipeline process on the beamformed input data to generate an ultrasonic image, Line performs the image restoration process using a graphics processing unit (GPU) of an OpenGL ES environment.

According to an exemplary embodiment, the image processing unit performs envelope detection and log compression in a first render cycle, and then performs gain control to adjust the overall gain of the image, black hole filtering in the second render cycle, edge enhancement in the third render cycle, The digital scan conversion is continuously performed in the fourth render cycle.

 The smart device for classifying medical ultrasound images according to an embodiment of the present invention shares a second render cycle for performing black hole filtering on log compressed data subjected to envelope detection and log compression on the input ultrasound signal, Sharing the third render cycle to concurrently operate image resizing and the first hidden layer on the data and enhance the edges of the data with the black holes removed to concurrently operate the second hidden layer, The third hidden layer is simultaneously operated by sharing a fourth render cycle for performing the scan conversion so as to share a rendering period for performing an image restoration process so that the first to third hidden layers GP, which performs body part classification simultaneously based on the in-depth neural network learning included U (Graphic processing unit).

The neural network learning according to an exemplary embodiment is a structure for classifying input data by learning a neural network structure based on a data set. In the data preprocessing for providing the data set, image data acquired from a probe is converted into raw data And sets the inverse transformed primitive data to the data set of the deep neural network learning to convert data of different specifications obtained from different probes into one normalized data, .

A GPU according to an embodiment may perform the black hole filtering and the image resizing and the first hidden layer, the edge enhancement and the second hidden layer, the scan conversion and the execution of the third hidden layer by the GPU as independent executions, Share in the cycle.

The GPU according to an exemplary embodiment of the present invention generates an ultrasound image by performing an initialization process, a render scene process, and a graphics pipeline process on the beamformed input data, Performs the image restoration process using a graphics processing unit (GPU) in an OpenGL ES environment.

According to an embodiment of the present invention, a smart device having a mobile GPU (Graphics Processing Unit) shares a rendering cycle for performing an image restoration process to simultaneously perform body part classification on the basis of a neural network learning, And the number of rendering cycles can be minimized to improve processing performance.

According to the embodiment of the present invention, the body part classification is performed based on the in-depth neural network learning using the inverse transformed raw data irrespective of the kind of the ultrasonic acquisition probe, Data can be used as input, and problems caused by conventional image cropping can be prevented.

In addition, according to the embodiment of the present invention, even if the screen space of the smart device is insufficient compared to the cart type device or the non-specialist must judge the image, the accuracy of the body part discrimination is automatically determined by automatically executing the body part classification based on the deep- And real-time performance can be guaranteed.

1 shows an example of an ultrasound medical imaging apparatus using a smartphone developed at present.
2 is a block diagram showing a schematic configuration of a deep learning-based medical ultrasound image classification apparatus using a smart device according to an embodiment of the present invention.
Figures 3-5 are diagrams illustrating cropping execution for a typical linear array, a convex array, and a phased array, respectively.
Figure 6 shows the inverse transformed raw data set as the dataset of the neural network learning in accordance with an embodiment of the present invention.
FIGS. 7A and 7B are diagrams for explaining a rendering cycle for efficiently implementing OpenGL ES programming.
FIG. 8 illustrates a general image restoration process.
9 is a diagram for explaining a rendering cycle sharing method according to an embodiment of the present invention.
FIG. 10 shows a procedure for implementing image restoration and a neural network learning algorithm in a cascade manner.
11 is a diagram illustrating a process of using a neural network learning algorithm to classify body structures according to an embodiment of the present invention.
12 and 13 are flowcharts for explaining a deep learning-based medical ultrasound image classification method using a smart device according to an embodiment of the present invention.

Prior to describing the embodiments of the present invention, the characteristics and problems of the existing smart-device-based ultrasound medical imaging device will be introduced. In order to solve these problems, the technical means adopted by the embodiments of the present invention are sequentially Be presented.

The smart device based ultrasound medical imaging device has a problem that the screen size is smaller than that of the cart type device and the parameters of all the images that can be used in the existing equipment due to the lack of the screen space are difficult to control.

In addition, if a non-specialist who has little background knowledge on medical ultrasound images uses a smart device, it is difficult to classify the body structure due to difficulty in image judgment, and since the performance of the GPU (Graphics processing unit) of the smart device is limited, However, since the signal processing process is already complicated for realizing the real-time characteristics, there is a disadvantage in that the real-time characteristic can not be guaranteed when the signal processing process for other purposes is added.

In order to solve the problem of the conventional smart device based ultrasound medical imaging device, the present invention discloses a neural network that classifies body structures such as liver, thyroid, and gallbladder in medical ultrasound equipment.

According to the embodiment of the present invention, in order to improve the performance and accuracy of classification of the body structure through the smart device, regardless of the type of the ultrasound acquisition probe, instead of the image data, By performing the site classification, log data having the same data type can be used as input without regard to the kind of the ultrasonic acquisition probe, and problems caused by image cropping can be prevented.

Particularly, in the embodiment of the present invention, in order to implement a neural network algorithm for classifying the body structure while maintaining real-time operation characteristics of a smart device GPU, a rendering cycle sharing method for reducing the input / output time of actual data and the number of GPU idle cores Lt; / RTI > A detailed description thereof will be given below with reference to the drawings.

2 is a block diagram showing a schematic configuration of a deep learning-based medical ultrasound image classification apparatus using a smart device according to an embodiment of the present invention.

2, a deep learning-based medical ultrasound image classification apparatus 100 using a smart device according to an embodiment of the present invention includes an ultrasound receiver 110 for receiving an ultrasound reflection signal from a target object through a probe and generating input data, And an image processing unit 131 for sharing a rendering period for performing an image reconstruction process on the input data and performing a body part classification based on a depth neural network learning in a shared rendering period. A memory unit 150 for storing software algorithms and the like necessary for the ultrasound signal processing, and a display unit 170 for outputting ultrasound images.

Here, the deep neural network learning is a structure for classifying input data by learning a deep neural network structure based on a data set. For data preprocessing for providing the data set, log compression data, which is a form obtained by inversely transforming an ultrasound image having geometric information into rectangle-shaped raw data, is set as a data set of a neural network learning network. That is, the image data obtained from the probe is inversely transformed into the data of the raw form, and the inversely transformed raw form data is set as the data set of the deep neural network learning, whereby data of different specifications obtained from different probes It is converted into one standardized data and used for depth learning of neural network.

FIGS. 3 to 5 are diagrams showing cropping execution for a general linear array, a convex array, and a phased array, respectively, and FIG. 6 is a diagram illustrating the inverse transformed raw data set as a dataset of the neural network learning according to an embodiment of the present invention / RTI >

Generally, a cropping process is performed to remove an unnecessary portion such as patient information unnecessary for ultrasound region determination in an ultrasound image. The ultrasound image is divided into the form of an ultrasound acquisition probe, which can be largely divided into a linear, convex, and phased array.

In the case of a linear image, the ultrasound image is cut out into a rectangular shape as shown in Fig. 3, and the convex or phased array has two types of full cropping shown in Fig. 4 or max cropping cropping the image. In the case of the full-cropping method, since the maximum size of the portion including the ultrasonic medical information is cut into the extended rectangle, data without the actual ultrasonic information remains. On the other hand, in the case of the max cropping, there is a risk that significant feature information may be lost in the original medical image because the image is cut out in a rectangular shape of the maximum size among the portions having the ultrasound information.

Therefore, in the embodiment of the present invention, as shown in FIG. 6, the ultrasound image 620 having the geometric information irrespective of the type of the ultrasound acquisition probe is converted into the rectangular-shaped raw data 630, Is used as the data set of the deep neural network learning, the log data having the same data type can be used as input regardless of the type of the ultrasonic acquisition probe, and the problem caused by the image cropping can be prevented.

That is, the present invention includes a process of inversely transforming image data acquired from a probe in a preprocessing step of learning of a neural network into raw data, and setting a data set of inverse-transformed raw data as a data set of a neural network learning Data of different standards obtained from different probes can be converted into one standardized data and used for learning of the neural network.

FIGS. 7A and 7B are diagrams for explaining a rendering cycle for efficiently implementing OpenGL ES programming. A rendering cycle is a unit in which a series of parallel signal processing is performed on all inputs. For example, to add 5 to all input data in Figure 7A, the GPU process performs a series of operations that add five. As shown in FIG. 7B, the render cycle requires physical input / output times of "Input texture upload", "Texture to frame buffer", and "Frame buffer to texture" in addition to "Processing".

If too little work is allocated to "Processing" here, the idle core of the GPU increases and processing performance degrades. Therefore, in order to minimize such idle time, it is necessary to minimize the number of rendering cycles.

The image processing unit 130 generates an ultrasound image by performing an initialization process, a render scene process, and a graphics pipeline process on the beamformed input data, For example, the image restoration process is performed using a graphics processing unit (GPU) in an OpenGL ES 3.0 environment.

FIG. 8 illustrates a general image restoration process. Referring to FIG. 8, in a general image reconstruction process, a DC component is removed from beamformed input data, a digital time gain compensation (TGC) is performed, a quadrature demodulation and a ratio 2 are performed, 8, the envelope detection and the log compression are performed in the first render cycle, and thereafter, the gain control for adjusting the overall gain of the image, the black hole filtering in the second render cycle, the edge enhancement in the third render cycle, In the render cycle, digital scan conversion was performed continuously.

In particular, the present invention intends to classify a body region based on a depth neural network learning by sharing a render cycle for performing the image restoration process described above. This will be described in detail with reference to FIG. FIG. 9 is a view for explaining a rendering period sharing method according to an embodiment of the present invention. Referring to FIG. 9, a neuron layer (neuron layer) for connecting an input layer and an output layer (Hidden layer).

As shown in FIG. 9, in order to share a rendering period for performing an image restoration process, the image processing unit 130 performs a black-hole filtering process on the log-compressed data on which envelope detection and log compression are performed, Sharing the two render cycles to simultaneously execute image resizing and the first hidden layer for the log compressed data in the second render cycle.

Then, the second hidden layer is concurrently executed in the third render cycle by sharing a third render cycle for enhancing the edge of the data in which black holes are removed by black hole filtering.

Then, by sharing the fourth render cycle for performing the scan conversion on the enhanced data of the edge and concurrently executing the third hidden layer in the fourth render cycle, the first hidden layer to the third hidden layer The body part classification can be performed based on the neural network learning network. Based on such a neural network learning network, body structures such as liver, thyroid gland and gallbladder can be classified as shown in FIG.

Here, the blackhole filtering, the image resizing, and the first hidden layer, the edge enhancement, the second hidden layer, the scan conversion, and the third hidden layer can share a render cycle because execution by the GPU is independent of each other . It is obvious that the first to third hidden layers are exemplified as the steps of the deep layer neural network learning, and the number of hidden layers is not limited thereto.

In this way, the image processing unit 130 of the present invention can perform the body part classification simultaneously on the basis of the deep neural network learning network in the shared rendering period by sharing the rendering period in which the smart device having the mobile GPU performs the image restoration process have.

As described above, according to the embodiment of the present invention, since the signal processing speed is faster than that of the cascade render cycle structure in which blocks are sequentially executed as shown in FIG. 10, real-time performance can be guaranteed.

As described above, according to the embodiment of the present invention, the idle core of the GPU can be reduced by simultaneously performing the body part classification on the basis of the depth neural network learning by sharing the rendering period for performing the image restoration process in the smart device having the mobile GPU The number of rendering cycles can be minimized and the processing performance can be improved.

In addition, regardless of the type of the ultrasound acquisition probe, the body part classification is performed based on the in-depth neural network learning using the inverse transformed raw data, so that the log data having the same data type can be used as input without depending on the type of the ultrasound acquisition probe. The problem caused by conventional image cropping can be prevented.

In addition, even if the screen space of the smart device is insufficient compared to the cart type device or the non-specialist must judge the image, it is possible to increase the accuracy of the body part discrimination by automatically executing the body part classification based on the deep neural network learning, can do.

12 is a flowchart illustrating a deep learning-based medical ultrasound image classification method using a smart device according to an embodiment of the present invention.

First, in a depth learning-based medical ultrasound image classification method using a smart device according to an embodiment of the present invention, a GPU receives an ultrasound reflection signal reflected from a target object through a probe and generates input data (S110).

Next, a plurality of hidden layers constituting the neural network learning are simultaneously executed by sharing a rendering period for performing an image restoration process on the input data (S120).

Next, a step of classifying the body structure of the ultrasound image according to the hidden layer execution result is included (S130). In other words, the body structure of the ultrasound image is classified according to the results of the deep neural network learning.

Here, the step of executing a plurality of hidden layers constituting the SNR learning by sharing the rendering cycle of step S120 will be described in detail with reference to FIG.

Referring to FIG. 13, a rendering sharing technique according to an embodiment of the present invention includes a second rendering cycle for performing black-hole filtering on log-compressed data subjected to envelope detection and log compression on an ultrasound signal input to a smart device And simultaneously executes image resizing and the first hidden layer on the log compressed data (S121).

Next, the second hidden layer is concurrently executed (step S122) by sharing a third render cycle for enhancing the edges of the data in which black holes are removed by black hole filtering.

Next, the third hidden layer is simultaneously executed by sharing the fourth render cycle for performing the scan conversion on the enhanced data of the edge (S123).

Next, the body part classification is performed based on the neural network learning network including the first to third hidden layers with respect to the log-compressed data (S124).

Accordingly, the rendering sharing technique according to the embodiment of the present invention can simultaneously perform the body part classification on the basis of the deep neural network learning in the shared rendering period by sharing the rendering period in which the image restoration process is performed by the smart device having the mobile GPU .

On the other hand, in the step of creating a neural network learning network, a data preprocessing step for providing a data set of a neural network learning network is executed. The preprocessing step includes the steps of inversely transforming the image data obtained from the probe into raw data and setting the inverse transformed raw data to a data set of the neural network learning, The acquired data of different specifications are converted into one standardized data and used for the learning of the neural network.

Therefore, according to the present invention, the body part classification is performed on the basis of the in-depth neural network learning using the inverse transformed raw data irrespective of the type of the ultrasonic acquisition probe, so that the log data having the same data form can be input regardless of the kind of the ultrasonic acquisition probe Can be used, and problems caused by image cropping can be prevented.

Here, the blackhole filtering and image resizing and the first hidden layer, the edge enhancement and the second hidden layer, the scan conversion, and the third hidden layer can share the render cycle because the execution by the GPU is independent of each other.

As described above, according to the embodiment of the present invention, since the signal processing speed is faster than that of the cascade render cycle structure in which blocks are sequentially executed as shown in FIG. 10, real-time performance can be guaranteed.

As described above, according to the embodiment of the present invention, the idle core of the GPU can be reduced by simultaneously performing the body part classification on the basis of the depth neural network learning by sharing the rendering period for performing the image restoration process in the smart device having the mobile GPU The number of rendering cycles can be minimized and the processing performance can be improved.

In addition, regardless of the type of the ultrasound acquisition probe, the body part classification is performed based on the in-depth neural network learning using the inverse transformed raw data, so that the log data having the same data type can be used as input without depending on the type of the ultrasound acquisition probe. Problems caused by image cropping can be prevented.

In addition, even if the screen space of the smart device is insufficient compared to the cart type device or the non-specialist must judge the image, it is possible to increase the accuracy of the body part discrimination by automatically executing the body part classification based on the deep neural network learning, can do.

Meanwhile, the embodiments of the present invention can be embodied as computer readable codes on a computer readable recording medium. A computer-readable recording medium includes all kinds of recording apparatuses in which data that can be read by a computer system is stored.

Examples of the computer-readable recording medium include ROM, RAM, CD-ROM, magnetic tape, floppy disk, optical data storage, and the like. In addition, the computer-readable recording medium may be distributed over network-connected computer systems so that computer readable codes can be stored and executed in a distributed manner. In addition, functional programs, codes, and code segments for implementing the present invention can be easily deduced by programmers skilled in the art to which the present invention belongs.

The present invention has been described above with reference to various embodiments. It will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined by the appended claims. Therefore, the disclosed embodiments should be considered in an illustrative rather than a restrictive sense. The scope of the present invention is defined by the appended claims rather than by the foregoing description, and all differences within the scope of equivalents thereof should be construed as being included in the present invention.

100: Deep running based ultrasound image classification device using smart device
110: ultrasound receiver 130: GPU
131: image processing unit 150: memory unit
170:

Claims (20)

Receiving ultrasound reflected signals reflected from a target object through a probe to generate input data;
Performing a Deep Neural Network (DNN) concurrently by sharing a rendering cycle for performing an image restoration process on the input data; And
And classifying the body structure of the ultrasound image according to the execution result of the deep NN learning. The method according to claim 1,
Wherein the step of sharing the rendering period to simultaneously perform the deep-
A second render cycle for performing black-hole filtering on the log-compressed data on which envelope detection and log compression are performed on the ultrasound signal input to the smart device is shared so that image resizing and the first hidden layer are simultaneously operated on the log- step;
Concurrently operating a second hidden layer by sharing a third render cycle that enhances the edge of data from which black holes have been removed by the black hole filtering;
Concurrently operating a third hidden layer by sharing a fourth render cycle that performs scan conversion on the enhanced data; And
And performing body part classification on the basis of a neural network learning network including the first hidden layer to the third hidden layer with respect to the log compressed data,
Wherein the smart device having a mobile GPU shares a rendering period for performing the image reconstruction process to simultaneously perform body part classification on the basis of the deep neural network learning in the shared rendering period, Image classification method. 3. The method of claim 2,
Further comprising a data preprocessing step for providing a data set of the deep-network neural network learning,
Converting the image data obtained from the probe into data of a raw form;
And setting the inversely transformed raw form data as a data set of the deep neural network learning, thereby converting data of different specifications obtained from different probes into one normalized data, Deep learning based ultrasound image classification method using smart device. 3. The method of claim 2,
Wherein the step of sharing the rendering period to simultaneously perform the neural network learning comprises:
Wherein the smart hole filtering and the image resizing and sharing of the first hidden layer, the edge enhancement and the second hidden layer, the scan conversion and the execution of the third hidden layer by the GPU, Deep Learning Based Ultrasonic Image Classification Method Using. The method according to claim 1,
The smart device generates an ultrasound image by performing an initialization process, a render scene process, and a graphics pipeline process on the beamformed input data, and the graphics pipeline includes an OpenGL ES And performing the image reconstruction process using a graphics processing unit (GPU) of the environment. The method according to claim 1,
The image restoration process includes:
Performing envelope detection and log compression on the input data in a first render cycle; And
A gain control in order to adjust the overall gain of the ultrasound image; the black hole filtering in the second render cycle; the edge enhancement in the third render cycle; and the digital scan conversion (DSC) The method of claim 1, further comprising the step of: A computer-readable recording medium on which a program for executing the method according to any one of claims 1 to 6 is recorded. An ultrasound receiver for receiving ultrasound reflection signals reflected from a target object through a probe and generating input data; And
And a video processor for simultaneously performing a body region classification based on a Deep Neural Network (DNN) in the shared rendering period by sharing a rendering period for performing an image restoration process on the input data, Deep learning based ultrasound image classification system using. 9. The method of claim 8,
Wherein the image processing unit comprises:
Sharing a second render cycle for performing black-hole filtering on log-compressed data on which envelope detection and log compression are performed on the input data to simultaneously perform image resizing and a first hidden layer on the log-compressed data, Deep learning based ultrasound image classification device for medical use. 10. The method of claim 9,
Wherein the image processing unit comprises:
And a third render cycle for enhancing an edge of data in which black holes are removed by the black hole filtering, thereby simultaneously operating the second hidden layer. 11. The method of claim 10,
Wherein the image processing unit comprises:
Wherein the third hidden layer is simultaneously operated by sharing a fourth render cycle for performing scan conversion on the enhanced data of the edge. 10. The method of claim 9,
Wherein the image processing unit comprises:
The smart device having the mobile GPU shares the rendering period for performing the image reconstruction process and simultaneously performs the body part classification on the basis of the depth neural network learning including the first to third hidden layers in the shared rendering period Deep learning based ultrasound image classification device using smart device. 10. The method of claim 9,
The deep-
It is a structure that classifies the input data by learning the deep neural network structure based on the data set,
The data preprocessing for providing the data set includes the steps of inverting the image data acquired from the probe into raw type data and setting the inversely transformed raw type data to the data set of the depth learning neural network, Wherein the data of the different standards acquired from the ultrasound diagnostic apparatus is converted into one standardized data and used for the learning of the neural network. 12. The method according to any one of claims 9 to 11,
Wherein the image processing unit comprises:
Wherein the smart hole filtering and the image resizing and sharing of the first hidden layer, the edge enhancement and the second hidden layer, the scan conversion and the execution of the third hidden layer by the GPU, Deep learning based ultrasound image classification system using. 9. The method of claim 8,
Wherein the image processing unit comprises:
An initialization process, a render scene process, and a graphics pipeline process are performed on the beamformed input data to generate an ultrasound image, and the graphics pipeline is subjected to graphics processing in an OpenGL ES environment And performing the image reconstruction process using a device (GPU). 9. The method of claim 8,
Wherein the image processing unit comprises:
Performing grayscale detection and log compression in the first render cycle for the image reconstruction process, performing gain control for adjusting the overall gain of the image, black-hole filtering in the second render cycle, edge enhancement in the third render cycle, Deep learning based medical ultrasound image classification device using smart device that continuously performs digital scan conversion in the render cycle. Sharing a second render cycle for performing black-hole filtering on log compression data on which envelope detection and log compression are performed on the input ultrasound signal to simultaneously operate image resizing and the first hidden layer on the log-
Sharing the third render cycle that enhances the edges of the data from which the black holes have been removed to simultaneously operate the second hidden layer,
Sharing the fourth render cycle for performing the scan conversion on the data with the enhanced edge so as to simultaneously operate the third hidden layer and sharing a rendering period for performing the image restoration process, A smart device for classifying a medical ultrasound image, comprising a GPU (Graphic Processing Unit) for simultaneously performing a body part classification based on a depth neural network learning including a hidden layer to a third hidden layer. 18. The method of claim 17,
The deep-
It is a structure that classifies input data by learning DNN (Deep Neural Network) based on dataset,
The data preprocessing for providing the data set includes the steps of inverting the image data acquired from the probe into raw type data and setting the inversely transformed raw type data to the data set of the depth learning neural network, And converting the data of the different standards acquired by the ultrasound imaging device into one standardized data and using the data for the learning of the neural network. 18. The method of claim 17,
The GPU includes:
Wherein the black hole filtering and the image resizing and the sharing of the first hidden layer, the edge enhancement and the second hidden layer, the scan conversion, and the third hidden layer by the GPU to each render cycle as independent executions, Smart devices that classify images. 18. The method of claim 17,
The GPU includes:
An initialization process, a render scene process, and a graphics pipeline process are performed on the beamformed input data to generate an ultrasound image, and the graphics pipeline is subjected to graphics processing in an OpenGL ES environment Wherein the image restoration process is performed using a device (GPU).

Images