WO2020111534A2

WO2020111534A2 - Method and system for estimating brain vascular age from medical image

Info

Publication number: WO2020111534A2
Application number: PCT/KR2019/014530
Authority: WO
Inventors: 장진희; 남윤호
Original assignee: 가톨릭대학교 산학협력단
Priority date: 2018-11-27
Filing date: 2019-10-31
Publication date: 2020-06-04
Also published as: WO2020111534A3; KR20200062656A; KR102225775B1

Abstract

The present invention relates to a method and a system for estimating brain vascular age from a medical image, the method for estimating brain vascular age from a medical image comprising the steps of: collecting brain vascular images with age information of a plurality of subjects; pre-processing a brain vascular image with age information to detect the position of a middle cerebral artery (MCA) for machine learning through a convolutional neural network (CNN) having a plurality of learning layers; machine learning through the convolutional neural network (CNN) comprising a plurality of convolutional layers for calculating feature maps by using the pre-processed brain vascular image with the age information as an input value of an input layer and applying a weight to the input value; and estimating a brain vascular age through the machine-learned convolutional neural network (CNN).

Description

Method and system for estimating cerebrovascular age from medical imaging

The present invention relates to the estimation of cerebrovascular age, and more particularly, to a method and system for estimating cerebrovascular age from medical images using machine learning.

The medical image acquires the anatomical and functional information of the object for medical purposes. Medical images can be obtained in various ways from various parts of the human body. For example, various cerebrovascular images can be obtained for the diagnosis of cerebrovascular disease, and one of them is a technique called time-of-flight MR angiography (TOF MRA). This technique uses a magnetic resonance image (Magnetic resonance image) to make the signal of blood with a strong flow rate appear strong and the signal of the stationary blood weakly extract only the images of blood vessels from each section and extract the blood vessel images obtained from each section. It is a technical method to obtain images of blood vessels without using contrast agents by combining and reconstructing in two or three dimensions. TOF MRA is a blood vessel imaging technique that is widely used in the medical field because it can present cerebral vessel structures in detail with excellent image resolution. In addition, as the related technology continues to develop, complex blood flow phenomena such as vortices can be diagnosed, and recently, microvascular images can also be obtained. In addition, CT cerebral angiography, which reconstructs cerebrovascular images by obtaining a computed tomography image after injecting a contrast agent, is widely used for high spatial resolution and fast examination time, and can obtain precise cerebrovascular images and hemodynamic information. Meanwhile, aging of the brain is well known and important. Since the brain is rich in blood supply to support high metabolic activity, it is not surprising that age-dependent changes include brain parenchyma as well as brain vasculatures. Histopathological studies have shown changes in cerebral arteries with aging, and microscopic changes have been associated with large morphological changes in the arteries of the head and neck.

Attempts have been made to measure the aging observed in the brain, and some results have been successful, whereas the measurement and digitization of aging observed in the cerebral blood vessels has weaker results than the attempt. This is due to the fact that features related to aging observed in cerebrovascular imaging are not properly defined. Previously, hand-crafted features were used to measure age-related changes found in brain MRA, but there was a lack of accuracy and time-consuming discomfort. In addition, there was no single measure or index to assess the overall aging status of cerebral blood vessels.

Imaging biomarkers for brain and vascular aging if the age of the cerebral vessels can be predicted by analyzing the morphology of blood vessels shown in cerebral vascular imaging, such as predicting age from structural brain MRI images using cross-sectional population data. biomarker).

[Advanced technical literature]

(Patent Document 1) Registered Patent Publication No. 10-1620302 (2016.05.04)

The problem to be solved by the present invention was created to solve the limitations and inconveniences of structural analysis of cerebrovascular images using the above-described conventional hand-crafted features, machine learning. For example, by using deep learning to analyze changes in the shape of cerebral blood vessels appearing on cerebrovascular images, medical knowledge that can predict the cerebrovascular age by extracting knowledge-based cerebrovascular features It provides a method and system for estimating cerebrovascular age from imaging.

A method of estimating cerebrovascular age from a medical image according to the present invention for achieving the above technical problem includes: collecting cerebrovascular images with age information of a plurality of subjects; Performing pre-processing to detect the location of the middle cerebral artery (MCA) for machine learning through a convolutional neural network (CNN) having a plurality of learning layers on the cerebral blood vessel image having the age information; Consists of a plurality of convolutional layers for calculating feature maps by applying a weight to the input value using the pre-processed cerebrovascular image as an input value of the input layer Machine learning through a convolutional neural network (CNN); And estimating the age of cerebral blood vessels through the machine-learned convolutional neural network (CNN).

The pre-processing comprises interpolating cerebral blood vessel images into an isotropic space and normalizing signal strength; Detecting the position of the middle cerebral artery (MCA) of the z-axis to match the coverage of the top and bottom (z-axis) of the head of the cerebrovascular image; Cutting an image of a predetermined size around a position of the detected middle cerebral artery (MCA); And interpolating the cut-sized image slab to adjust the data size.

The method for estimating cerebrovascular age from a medical image according to the present invention, before cutting data to reduce sensitivity to spatial variation of cerebrovascular image data in the learning data set of the convolutional neural network (CNN) It is characterized by increasing data by performing shifting, rotation and flipping to reflect the diversity that may occur during medical imaging.

The method for estimating cerebrovascular age from a medical image according to the present invention may further include extracting features of the cerebral blood vessels through the machine-learned convolutional neural network (CNN). The input data of the convolutional neural network (CNN) may further include risk factors that induce cerebrovascular diseases including hypertension, diabetes, and hyperlipidemia.

The system for estimating cerebrovascular age from medical images according to the present invention for achieving the above technical problem includes: a cerebrovascular imaging DB that collects and stores a 3D TOF MRA image with age information of a subject; A pre-processing unit for performing pre-processing to detect the position of the middle cerebral artery (MCA) to machine learn a 3D cerebrovascular image with age information obtained from the cerebrovascular image DB through a convolutional neural network (CNN); An input unit for inputting age and cerebrovascular disease risk factors of images pre-processed by the pre-processing unit; A plurality of convolutional layers and a plurality of convolutional layers for calculating feature maps by applying a weight to the input value using the cerebral blood vessel image having the pre-processed age information as an input value of the input layer It comprises a plurality of pooling (pooling) layer to reduce the size of the feature map, and includes a convolutional neural network (CNN) that estimates the age of the cerebral blood vessels through regression analysis.

The pre-processing unit is an isotropic standardization unit for interpolating cerebral blood vessel images into an isotropic space and standardizing signal intensity; An MCA detection unit that detects the position of the middle cerebral artery (MCA) on the z-axis in order to align the top-down (z-axis) coverage of the cerebrovascular image; And a three-dimensional slab generating unit for cutting the image of a predetermined size based on the detected position of the middle cerebral artery (MCA), interpolating the cut-sized image (slab), adjusting the data size, and inputting machine learning. do.

The system for estimating cerebrovascular age from medical images according to the present invention moves before cutting data to reduce sensitivity to spatial variation of cerebrovascular image data during learning and testing of the convolutional neural network (CNN) ( It further includes an image enhancement unit that increases training data by performing shifting, rotation, and flipping.

The convolutional neural network (CNN) uses the TOF MRA image having the pre-processed age information as an input value of an input layer to apply a weight to the input value to calculate feature maps. hierarchy; A plurality of pooling layers that reduce the size of the feature maps of the plurality of convolution layers through sub-sampling; A full connection layer connected to the pooling layer and all activations; And a regression analysis unit for estimating the age of cerebral blood vessels through regression analysis using the output value of the complete connection layer. The input unit may further include risk factors and sex information for causing cerebrovascular diseases including hypertension, diabetes, and hyperlipidemia.

According to the method and system for estimating cerebrovascular age from medical images according to the present invention, the overall aging state of cerebral blood vessels is quantified by quantifying morphological changes in cerebral blood vessels through machine learning, for example, deep learning. A single measure or index can be provided to evaluate.

In addition, according to the present invention, brain vascular age can be predicted by extracting a knowledge-based cerebrovascular feature by analyzing changes in the shape of cerebral blood vessels appearing on cerebrovascular images using machine learning. It can be used as an imaging biomarker for aging.

In addition, according to the present invention, knowledge-based cerebrovascular characteristics and risk factor information of a patient can be added to provide cerebrovascular information that cannot be previously measured. In addition, by using new cerebrovascular information to check the health status of each individual and utilize it to promote health, early diagnosis and prevention of specific diseases can be prevented.

1 is a block diagram showing the technical idea of the present invention.

2 shows an example of a change in morphological characteristics of brain blood vessels with age.

Figure 3 is a block diagram showing an embodiment of the configuration of the cerebrovascular age estimation system from a medical image according to the present invention.

4 is a block diagram showing a more specific configuration of the pre-processing unit 320.

5 shows an embodiment of the pre-processing unit 320 as an MRA image.

6 shows an example of the image enhancement unit 330.

7 shows an example of a convolutional neural network (CNN, 350) used as a component of a cerebrovascular age estimation system from a medical image according to the present invention.

8 shows an example of an implementation of a convolutional neural network (CNN, 350) used as a component of a cerebrovascular age estimation system from a medical image according to the present invention.

9 is a flowchart illustrating an embodiment of a method for estimating cerebrovascular age from a medical image according to the present invention.

10 shows the pretreatment process in more detail.

11 shows an example of providing services such as risk and lifestyle advice for risk factors using the cerebrovascular age estimated by the system for estimating cerebrovascular age from the medical image according to the present invention.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. The configurations shown in the embodiments and drawings described in this specification are only preferred embodiments of the present invention, and do not represent all of the technical spirit of the present invention, and various equivalents that can replace them at the time of this application It should be understood that there may be and variations.

Figure 1 is a block diagram showing the technical idea of the present invention, as a brain blood vessel image, machine learning through a convolutional neural network (Convolution Neural Network), for example, deep learning, deep learning Indicates that it is estimated. FIG. 2 shows an example of changes in the morphological characteristics of cerebral blood vessels according to age, and it can be seen that the morphological characteristics of the cerebral blood vessels shown in the brain MRA image change with age.

Referring to FIG. 1, a cerebral blood vessel shown in an MRA image through a convolutional neural network 120 using a MRA (Magnetic Resonance Angiography) image input through the input unit 110 as a training data set The age of the cerebral blood vessels is estimated by deep learning the change of the morphological characteristics of the (130).

Figure 3 is a block diagram showing an embodiment of the configuration of the cerebral vascular age estimation system from the medical image according to the present invention, the cerebral vascular age estimation system from the medical image according to an embodiment of the present invention is a cerebrovascular image DB ( 310, a pre-processing unit 320, an input unit 340, and a convolutional neural network 350, and may further include an image enhancement unit 330.

The cerebrovascular image DB 310 may be constructed as a database for collecting and storing cerebrovascular images, for example, three-dimensional TOF MRA images, with age information of a subject.

The pre-processing unit 320 performs machine learning through the convolutional neural network (CNN, 350), for example, a cerebral blood vessel image having the age information obtained from the cerebrovascular imaging DB 310, for example, a 3D TOF MRA image. Pre-treatment is performed for deep learning. 4 is a block diagram showing a more specific configuration of the pre-processing unit 320, the pre-processing unit 320 includes an isotropic standardization unit 410, an MCA detection unit 420 and a SLAB generation unit 430. Referring to FIG. 4, the isotropic standardization unit 410 interpolates a cerebrovascular image, for example, a TOP MRA image into an isotropic space and normalizes signal strength.

The MCA detection unit 420 detects the position of the middle cerebral artery (MCA) on the z-axis in order to match the coverage in the upper and lower range (z-axis) of the brain vessel image, for example, the 3D TOF MRA image.

The SLAB generation unit 430 cuts an image of a predetermined size based on the detected position of the middle cerebral artery (MCA) and interpolates the cut-sized image (slab) to reduce the data size.

5 shows an embodiment of the pre-processing unit 320 as an MRA image.

Since machine learning, for example deep learning, relies on large volumes of data, a large number of normal subjects and a wide age distribution are required. Referring to FIG. 5, in an embodiment of the present invention, in order to construct an appropriate database, MRA data from cerebral blood vessel imaging with age from specific hospital databases (Our DB, 511) and public databases CASILab (513) and IXI (515) Collected. In the embodiment of the present invention, a total of 1192 3D TOF MRA images were collected.

Reference number 512 denotes the number of MRA data according to the age distribution of a specific hospital database (Our DB, 511), and

reference numbers

514 and 516 denote MRA data according to the age distribution of public databases CASILab(513) and IXI(515). It shows the number. Because one public DB (IXI, 515) and one hospital DB (Our DB, 511) contain a large number of subjects with a uniform age distribution, the public DB (IXI, 515) and hospital DB (Our DB, 511) was used for training and verification. The pooled data were also divided into training (n = 652) and validation (n = 94) sets while maintaining the age distribution.

Since the imaging protocols and image characteristics may be different for the three

databases

511, 513, and 515, pre-processing is required before machine learning. First, in the embodiment of the present invention, the 3D TOF image is interpolated into a 0.5 mm isotropic space for the MRA DB 510 including three

databases

511, 513, and 515, and signal strength is normalized. (510)

After the interpolation and normalization, the position of the middle cerebral artery (MCA) on the z-axis is estimated as shown in Fig. B 520 to match the coverage of the MRA image data in the upper and lower range (z-axis) direction. After estimating the location of the MCA, a 6.4 cm thick slab (SLAB, 1.7 cm to 4.7 cm above) is extracted. The central part of the slab (256x256) is cut off and the SLAB (256x256x128), cut off due to limitations in the computational environment, is interpolated into a 1.0mm isotropic space (128x128x64) (Figure C, 530).

The input unit 340 inputs risk factors for cerebrovascular disease including gender, age and hypertension, diabetes, and hyperlipidemia of the image preprocessed in the preprocessing unit 320.

The convolutional neural network (CNN, 350) applies feature maps by applying a weight to the input value using a brain blood vessel image with pre-processed age information, for example, a TOF MRA image as an input value of an input layer. It comprises a plurality of convolutional layers to calculate and a plurality of pooling layers to reduce the size of the feature maps of the plurality of convolutional layers, and the age of cerebral blood vessels is estimated through regression analysis.

The image augmentation unit 330 shifts the MRA image before cutting the data to reduce sensitivity to spatial variation of cerebrovascular image data of the data set in the course of learning and testing of the convolutional neural network (CNN, 350). , Rotation and flipping to increase training data. 6 illustrates an example of the image enhancement unit 330, and shows that 652 images can be increased to 20,000 or more images by performing shifting, rotation, and flipping.

FIG. 7 shows an example of a convolutional neural network (CNN, 350) used as a component of a cerebrovascular age estimation system from a medical image according to the present invention, a plurality of convolutional layers (720, 740), a plurality of pooling It comprises a (pooling) layer (730, 750), a fully connected (fully connected) layer 760 and a regression (regression) analysis unit 770.

The plurality of

convolutional layers

720 and 740 is a feature map by applying a weight to the input value using a TOF MRA image 710 as an input value of an input layer as an example of a cerebrovascular image having pre-processed age information. (feature maps). The plurality of pooling

layers

730 and 750 reduce the size of the feature maps of the plurality of

convolutional layers

720 and 740 through sub-sampling.

The fully connected layer 760 is connected to all activations of the pooling layer 750. The regression analysis unit 770 estimates the age of the cerebral blood vessels through regression analysis using the output value of the complete connection layer 760.

8 shows an example of an implementation of a convolutional neural network (CNN, 350) used as a component of a cerebrovascular age estimation system from a medical image according to the present invention. Referring to FIG. 8, as a scheme of 3D CNN, a representative image of an activation map of each layer is shown. The 3D CNN architecture has five weight layers (not shown), three 3D convolutional layers 810, 820, and 830, and two fully connected layers 840. In this model, a 3x3x3 convolution kernel 850 with a stride of 1 and a 2x2x2 maximum pooling with a stride of 1 (860) were used for all layers. After each convolutional layer, a rectified linear unit (ReLU, 870) was applied. In the course of training, the mean square error was used as a loss function in regression (880) and the batch size was 8.

For example, in the convolution layer 810, when a convolution operation is performed using a 3x3x3 convolution kernel 850 having a stride of 1 for 3D input data 805 of 128x128x64, a plurality of 128x128x64 Feature maps can be constructed. A subsampling map of 64x64x32 may be constructed by applying 2x2x2 maximum pooling 860 with a stride length of 1 from a feature map of each 128x128x64 in the convolution layer 810. In embodiments of the invention, activation can be seen primarily in vascular voxels. Network parameters including learning rate and number of epochs can be adjusted empirically to improve performance for validation through repeated experiments.

9 is a flowchart illustrating an embodiment of a method for estimating cerebrovascular age from a medical image according to the present invention. 3, 4, 7 and 9, an embodiment of a method for estimating cerebrovascular age from a medical image according to the present invention will be described. First, a medical image with age information of a plurality of subjects (subject), for example, TOF MRA images are collected to build a cerebrovascular image DB 310 (step S910).

The pre-processing unit 320 performs pre-processing to detect the position of the middle cerebral artery (MCA) in order to machine-learn a cerebral blood vessel image with age information through a convolutional neural network (CNN). It shows the process in more detail. Referring to FIG. 10, the isotropic standardization unit 410 interpolates a cerebrovascular image, for example, a TOF MRA image into an isotropic space and normalizes signal strength. (Step S1010) The MCA detection unit 420 heads the TOF MRA image The position of the middle cerebral artery (MCA) of the z-axis is detected in order to match the coverage in the upper and lower range (z-axis). (Step S1020) The SLAB generation unit 430 detects the position of the detected middle cerebral artery (MCA). The image of a predetermined size is cut out at the center. (Step S1030) Then, the size of the data is reduced by interpolating the cut image of the predetermined size (Step S1040).

When a cerebrovascular image, for example, a TOF MRA image is pre-processed, the CNN unit 350 applies the TOF MRA image having the pre-processed age information as an input value of an input layer and applies a weight to the input value. Machine learning is performed through a convolutional neural network (CNN) including a plurality of convolutional layers for calculating feature maps (step S930). The regression analysis unit of the CNN unit 350 (step S930) 770) estimates the age of cerebral blood vessels through the machine-learned convolutional neural network (CNN) (step S940).

Meanwhile, in step S920, after pre-processing a cerebrovascular image, for example, a TOF MRA image, data to reduce sensitivity to spatial variation of cerebrovascular image data in the learning data set of the convolutional neural network (CNN). The data can be increased by performing shifting, rotation, and flipping before cutting. And the characteristics of the cerebral blood vessels can be extracted through the machine-learned convolutional neural network (CNN).

Meanwhile, the input data of the convolutional neural network (CNN) may further include a risk factor that causes cerebrovascular diseases including hypertension, diabetes, and hyperlipidemia. 11 shows an example of providing an analysis of cerebrovascular age and risk factors using a system for estimating cerebrovascular age from a medical image according to the present invention. Referring to FIG. 11, a 3D cerebral blood vessel image, a patient's gender and age, and risk factors for cerebrovascular disease such as hypertension, diabetes, hyperlipidemia, and cardiovascular disease are input to estimate cerebral vascular age, and estimated cerebrovascular disease By using the information to provide aging measurement results and analysis results in graphs and comments, it can be used for personal health check and health promotion. Through this, you can improve your lifestyle and diagnose and prevent certain diseases early.

The present invention can be embodied as computer readable code on a computer readable recording medium (including all devices having information processing functions). The computer-readable recording medium includes all kinds of recording devices in which data readable by a computer system are stored. Examples of computer-readable recording devices include ROM, RAM, CD-ROM, magnetic tape, floppy disks, and optical data storage devices. Also, in this specification, the “unit” may be a hardware component such as a processor or circuit, and/or a software component executed by a hardware component such as a processor.

The present invention has been described with reference to the embodiments shown in the drawings, but these are merely exemplary, and those skilled in the art will understand that various modifications and other equivalent embodiments are possible therefrom. Therefore, the true technical protection scope of the present invention should be determined by the technical spirit of the appended claims.

Claims

Collecting cerebrovascular images with age information of a plurality of subjects;

Performing pre-processing to detect the location of the middle cerebral artery (MCA) for machine learning through a convolutional neural network (CNN) having a plurality of learning layers on the cerebral blood vessel image having the age information;

Consists of a plurality of convolutional layers for calculating feature maps by applying a weight to the input value using the pre-processed cerebrovascular image as an input value of the input layer Machine learning through a convolutional neural network (CNN); And

And estimating the age of cerebral blood vessels through the machine-learned convolutional neural network (CNN).
The method of claim 1, wherein the pre-treatment

Interpolating cerebrovascular images into isotropic space and normalizing signal strength;

Detecting the position of the middle cerebral artery (MCA) of the z-axis to match the coverage of the top and bottom (z-axis) of the head of the cerebrovascular image;

Cutting an image of a predetermined size around a position of the detected middle cerebral artery (MCA); And

And interpolating the cut-sized image (slab) to adjust the data size.
According to claim 1,

In order to reduce sensitivity to spatial variation of cerebrovascular image data in the learning data set of the convolutional neural network (CNN), shifting is performed to reflect diversity that may occur during medical imaging before cutting the data, Method of estimating cerebral blood vessel age from medical images, characterized in that the data is increased by performing rotation and flipping.
According to claim 1,

Method for estimating cerebral blood vessel age from a medical image, characterized in that it further comprises the step of extracting the features of the cerebral blood vessels through the machine learning convolutional neural network (CNN).
According to claim 1, The input data of the convolutional neural network (CNN)

Method for estimating age of cerebrovascular from medical images, characterized in that it further comprises a risk factor for inducing cerebrovascular disease, including hypertension, diabetes, and hyperlipidemia.
A cerebrovascular image DB that collects and stores a 3D TOF MRA image with age information of a subject;

A pre-processing unit for performing pre-processing to detect the position of the middle cerebral artery (MCA) to machine learn a 3D cerebrovascular image with age information obtained from the cerebrovascular image DB through a convolutional neural network (CNN);

An input unit for inputting age and cerebrovascular disease risk factors of images pre-processed by the pre-processing unit;

A plurality of convolutional layers and a plurality of convolutional layers for calculating feature maps by applying a weight to the input value using the cerebrovascular image having the pre-processed age information as an input value of the input layer It comprises a plurality of pooling (pooling) layer to reduce the size of the feature map, including a convolutional neural network (CNN) that estimates the age of the cerebral blood vessels through regression analysis, to estimate the cerebrovascular age from medical images system.
The method of claim 6, wherein the pre-processing unit

An isotropic standardization unit for interpolating cerebrovascular images into isotropic space and standardizing signal intensity;

An MCA detection unit that detects the position of the middle cerebral artery (MCA) on the z-axis in order to align the top-down (z-axis) coverage of the cerebrovascular image; And

It cuts the image of a predetermined size around the position of the detected middle cerebral artery (MCA), interpolates the cut-off image of the predetermined size (slab), adjusts the data size, and includes a 3D Slab generator for machine learning input A system for estimating cerebrovascular age from medical images.
The method of claim 6,

Training is performed by shifting, rotating and flipping before cutting data to reduce sensitivity to spatial variation of cerebrovascular image data in the course of learning and testing the convolutional neural network (CNN). A system for estimating cerebrovascular age from medical images, further comprising an image enhancement unit for increasing data.
The method of claim 6, wherein the convolutional neural network (CNN)

A plurality of convolutional layers for calculating feature maps by applying a weight to the input value using the TOF MRA image having the pre-processed age information as an input value of an input layer;

A plurality of pooling layers that reduce the size of the feature maps of the plurality of convolution layers through sub-sampling;

A full connection layer connected to the pooling layer and all activations; And

And a regression analysis unit for estimating the age of cerebral blood vessels through regression analysis using the output value of the complete connection layer.
The method of claim 1, wherein the input unit

A system for estimating cerebrovascular age from medical images, further comprising risk factors and gender information that induce cerebrovascular diseases including hypertension, diabetes, and hyperlipidemia.