CN111062955A

CN111062955A - Lung CT image data segmentation method and system

Info

Publication number: CN111062955A
Application number: CN202010188475.3A
Authority: CN
Inventors: 金烁; 赵威; 代笃伟; 侯雪雪; 徐正清; 王博; 申建虎; 张伟; 金红波; 靳博方; 潘承燕
Original assignee: Tianjin Precision Diagnosis Medical Technology Co Ltd
Current assignee: Tianjin Precision Diagnosis Medical Technology Co Ltd
Priority date: 2020-03-18
Filing date: 2020-03-18
Publication date: 2020-04-24

Abstract

The invention discloses a lung CT image data segmentation method and a lung CT image data segmentation system, and belongs to the technical field of medical image processing and artificial intelligence. The method comprises the following steps: labeling the lung contour and the target area, and performing numerical clipping and normalization processing on the image data; training and learning by utilizing a first neural network to obtain a lung contour segmentation model; extracting a lung contour, determining a lung region of interest, and performing numerical cutting and normalization processing; cutting the lung region of interest, and performing training and learning by using a second neural network to obtain a lung target region segmentation model; and segmenting the target region of the lung according to the image data and the segmentation model of the target region of the lung. The system comprises a labeling module, a first cutting normalization module, a first training learning module, a second cutting normalization module, a second training learning module and a detection module. The invention improves the processing efficiency of the lung CT image data and can rapidly segment the target area in the lung CT image data.

Description

Lung CT image data segmentation method and system

Technical Field

The invention relates to the technical field of medical image processing and artificial intelligence, in particular to a lung CT image data segmentation method and a lung CT image data segmentation system.

Background

The image is obtained by scanning a certain thickness of layer surface of human body with X-ray beam by CT equipment, receiving the X-ray transmitted through the layer surface by a detector, converting into visible light, converting into electric signal by photoelectric conversion, converting into digital signal by an analog/digital converter, and inputting into a computer for processing. CT imaging technology has been widely used in medical examinations, and especially the detection of targets in CT images has become a pre-step in the diagnosis of various diseases.

At present, the target detection in the lung CT image is mainly checked by the naked eyes of doctors, the detection method depends on the detection experience of the doctors on the target area, the time consumption for the doctors to read the lung CT image is long, the speed is low, the efficiency is low, and the judgment results of different doctors on the target area are different. Recently, with the rapid spread of new coronaviruses, the number of patients with new coronary pneumonia is increasing dramatically. Because the lung CT image of a patient infected by the new coronavirus is characterized earlier than the clinical characterization, the lung CT flat scan examination is mainly adopted. Under the new crown epidemic situation environment, medical resources are in short supply, nearly thousand patients are queued to wait for lung CT examination in an epidemic situation serious disaster area every day, because sufficient doctors do not read a large amount of CT images, and the speed of manually reading the CT images is low, the efficiency is low, the queuing waiting examination time in a patient hospital is increased, cross infection is easily caused, and the condition of an illness is delayed. Aiming at the new coronary pneumonia virus epidemic situation, a method capable of rapidly processing lung CT image data is urgently needed to replace the existing mode of manually reading lung CT image data.

Disclosure of Invention

In order to solve the problems of long time consumption, low speed, low efficiency and the like of the existing method for manually reading lung CT image data, the invention provides a lung CT image data segmentation method, which comprises the following steps:

labeling a lung contour and a target area in lung CT flat scanning image data, and performing numerical clipping and normalization processing on the image data;

using the normalized image data and the labeled lung contour data, and performing training learning by using a first neural network to obtain a lung contour segmentation model;

extracting a lung contour according to the lung contour segmentation model, taking a cube externally connected with the lung contour as a lung interesting region, and performing numerical cutting and normalization processing on the lung interesting region;

the normalized lung interesting region is cut, and the cut lung interesting region data and the labeled target region are used for training and learning by utilizing a second neural network to obtain a lung target region segmentation model;

and segmenting the target region of the lung according to the image data and the segmentation model of the target region of the lung.

The step of labeling the lung contour and the target region in the lung CT flat scan image data specifically comprises the following steps:

acquiring lung CT flat scanning image data and cleaning the image data;

and marking the lung contour and the target area in the cleaned image data by adopting a manual outlining marking method.

The step of performing numerical clipping and normalization processing on the image data specifically comprises the following steps:

adjusting the image resolution to a first resolution uniform value;

and cutting the numerical matrix of the image according to a preset first gray value cutting range, and carrying out normalization processing on data in the numerical matrix of the image to enable all numerical values in the numerical matrix of the cut image to be a numerical value between 0 and 1.

The step of performing numerical clipping and normalization processing on the lung region of interest specifically comprises the following steps:

adjusting the lung region-of-interest resolution to a second resolution uniform value;

and cutting the lung region-of-interest numerical matrix according to a preset second gray value cutting range, and normalizing data in the lung region-of-interest numerical matrix to change all numerical values in the cut lung region-of-interest numerical matrix into a numerical value between 0 and 1.

The first and second resolution uniform values are each set to [1mm, 1mm, 2.5mm ], the first gradation value clipping range is set to [ -200,400], and the second gradation value clipping range is set to [ -1200,600 ].

The first neural network is a 3D-Unet network, and loss functions used in the 3D-Unet network training process comprise dice-loss and Focal-loss; the second neural network is a 3D-Unet network added with an attention mechanism.

The invention also provides a lung CT image data segmentation system, which comprises:

the labeling module is used for labeling the lung contour and the target region in the lung CT flat scanning image data;

the first cutting normalization module is used for carrying out numerical cutting and normalization processing on the image data;

the first training learning module is used for performing training learning by using the image data processed by the first cutting normalization module and the lung contour data labeled by the labeling module through a first neural network to obtain a lung contour segmentation model;

the second cutting normalization module is used for extracting a lung contour according to the lung contour segmentation model obtained by the first training learning module, taking a cube externally connected with the lung contour as a lung interesting region, and performing numerical cutting and normalization processing on the lung interesting region;

the second training learning module is used for cutting the lung interesting region processed by the second cutting normalization module, and using the cut lung interesting region data and the target region marked by the marking module to perform training learning by using a second neural network to obtain a lung target region segmentation model;

and the segmentation module is used for segmenting the lung target region according to the image data and the lung target region segmentation model.

The labeling module comprises:

the acquisition unit is used for acquiring lung CT flat scanning image data;

the cleaning unit is used for cleaning the image data acquired by the acquisition unit and removing CT flat scanning image data with artifacts and/or distortion;

and the marking unit is used for marking the lung contour and the target area in the image data cleaned by the cleaning unit by adopting a manual outlining marking method.

The first clipping normalization module comprises:

the first setting unit is used for presetting a first gray value cutting range;

the first adjusting unit is used for adjusting the image resolution to a first resolution uniform value;

the first clipping unit clips the numerical matrix of the image according to a first gray value clipping range preset by the first setting unit;

the first normalization unit is used for changing all numerical values in the numerical value matrix of the image cut by the first cutting unit into a numerical value between 0 and 1.

The second clipping normalization module comprises:

the extraction unit is used for extracting the lung contour according to the lung contour segmentation model obtained by the first training learning module;

the second setting unit is used for presetting a second gray value cutting range;

the establishing unit is used for taking the lung outline external cube extracted by the extracting unit as a lung interesting region;

the second adjusting unit is used for adjusting the resolution of the region of interest of the lung determined by the establishing unit to a second resolution uniform value;

the second cutting unit is used for cutting the lung region-of-interest numerical matrix according to a second gray value cutting range preset by the second setting unit;

and the second normalization unit is used for changing all values in the lung region-of-interest value matrix cut by the second cutting unit into a value between 0 and 1.

According to the lung CT image data segmentation method and system provided by the invention, the CT image and the artificial intelligence technology are combined, and the artificial intelligence deep learning technology is utilized, so that a computer learns a large number of lung CT image characteristics in a short time, and the processing efficiency of lung CT image data is improved. The lung CT image data segmentation method and the lung CT image data segmentation system can rapidly segment the target region in the lung CT image data, further carry out three-dimensional reconstruction on the segmented target region, are convenient for visual display, can be widely applied to the examination of various lung diseases, such as pulmonary nodules, lung cancer, new coronary pneumonia and the like, and provide clear and accurate target region images for doctors.

Drawings

FIG. 1 is a flowchart of a method for segmenting CT image data of a lung of a patient with new coronary pneumonia according to an embodiment of the present invention;

FIG. 2 is a schematic structural diagram of a 3D-Unet neural network according to an embodiment of the present invention;

FIG. 3a is a schematic representation of a CT flat scan image of the original lung of a patient with coronary pneumonia according to an embodiment of the present invention;

FIG. 3b is a schematic diagram of the lung contour extracted in FIG. 3a using a lung contour segmentation model according to an embodiment of the present invention;

FIG. 4 is a schematic structural diagram of an Attention 3D-Unet neural network according to an embodiment of the present invention;

FIG. 5a is a schematic diagram of a CT flat scan image of the lung of another new patient with coronary pneumonia according to the present invention;

FIG. 5b is a schematic diagram of the lung target region segmentation model of FIG. 5a according to an embodiment of the present invention;

fig. 6 is a schematic structural diagram of a lung CT image data segmentation system according to an embodiment of the present invention.

Detailed Description

The technical solution of the present invention is further described below with reference to the accompanying drawings and examples.

Referring to fig. 1, the embodiment of the present invention takes CT image data of a lung of a patient with new coronary pneumonia as an example to describe a method for segmenting CT image data of the lung, including the following steps:

s101, acquiring CT flat scan image data of the lung of a new coronary pneumonia patient.

The CT flat scan image data of the lung of the new coronary pneumonia patient is acquired from the new coronary pneumonia patients of different ages in a plurality of authorized medical centers and regions. In the data acquisition process, personal information of a new coronary pneumonia patient is strictly kept secret to prevent the personal information from being leaked in the acquisition process. For example, fig. 3a is a schematic diagram of a CT flat scan image of an original lung of a new patient with coronary pneumonia according to the present embodiment.

And S102, cleaning the CT flat scan image data of the lung of the new coronary pneumonia patient.

The method for cleaning the CT flat scan image data of the lung of the new coronary pneumonia patient mainly means removing the CT flat scan image data with artifacts and/or distortion deformation so as to obtain data meeting the requirement of neural network training.

Step S103, labeling the lung contour and the target area in the CT flat scanning image data.

In specific application, a manual delineation labeling method is adopted to label data from three directions of a sagittal position, a coronal position and a columnar position respectively, original two-dimensional CT flat scan image data is referred, and a labeled three-dimensional result is subjected to three-dimensional modification, so that the labeling result is more continuous; the labeled lung contours and target region data format are saved in nrrd format.

And S104, adjusting the resolution of the CT flat-scan image to a uniform resolution value, cutting the numerical matrix of the CT flat-scan image according to a preset first gray value cutting range, and performing normalization processing on the cut data.

In this embodiment, the resolution unity value may be set to [1mm, 1mm, 2.5mm ], and the first gray value clipping range is [ -200,400 ]. After the cropping, the data in the numerical matrix of the CT flat-scan image needs to be normalized by the following formula, so that all the numerical values in the numerical matrix of the cropped CT flat-scan image become one numerical value between 0 and 1:

（1）

wherein k represents the value of the pixel point in the numerical matrix of the clipped CT flat-scan image.

And S105, using the normalized original CT flat scan image data and the labeled lung contour data, and performing training and learning by using a first neural network to obtain a lung contour segmentation model.

Because the new coronary pneumonia CT image features are all expressed in the lung, the interference of other organ tissues in the CT image is eliminated, and the accuracy of the target region can be effectively improved. Because the lung region and the surrounding tissues have obvious difference, the lung region can be extracted by using a deep learning segmentation algorithm. And training and learning by using the normalized original CT flat scan image data and the labeled lung contour data and utilizing a first neural network to obtain a lung contour segmentation model with excellent performance.

In this embodiment, the first neural network is a 3D-Unet network, and a network structure thereof is shown in fig. 2. Loss functions used in the 3D-Unet network training process include dice-loss and Focal-loss. The dice-loss is a loss function commonly used in medical image segmentation, and dice is a set similarity measurement function and is generally used for calculating the similarity of two samples, and the mathematical expression of dice-loss is as follows:

（2）

wherein the content of the first and second substances,

represents the intersection between M and N;

respectively representing the number of elements of M and N, M representing a GT-segmented image and N representing a predicted segmented image.

Focal-loss is a loss that solves the problem of unbalanced classes and differences in difficulty of classification. The segmentation task is a classification at the pixel level, so the Focal length is also valid in the segmentation task. The mathematical expression for Focal loss is as follows:

（3）

wherein gamma is a focusing parameter,p _tto predict the probability that a sample belongs to a 1,

the number of pixels of different classes.

And S106, extracting a lung contour according to the lung contour segmentation model, taking a cube externally connected with the lung contour as a lung ROI, and expanding the lung ROI.

Fig. 3b is a schematic diagram of a lung contour extracted from the CT flat scan image of the lung shown in fig. 3a by using a lung contour segmentation model according to the present embodiment. In specific applications, a suitable expansion of the lung ROI (region of interest) is required, for example: the lung ROI may be expanded by 10 pixels in X, Y and Z-axis directions, respectively. The effect of enlarging the lung ROI is that the circumscribed cube is not completely tangent to the lung contour, so that a space is left between the circumscribed cube and the lung contour, and the target region at the boundary of the lung is prevented from being missed.

And S107, adjusting the lung ROI resolution to a uniform resolution value, cutting a lung ROI numerical matrix according to a preset second gray value cutting range, and performing normalization processing on the cut data.

In this embodiment, the resolution uniform value in this step is the same as the resolution uniform value in step S104, and the second gray scale value clipping range is [ -1200,600 ]. It should be noted that: the lung ROI numerical matrix is cut, the size of the lung ROI is not changed, and the size of the numerical value in the lung ROI numerical matrix is changed; referring to tables 1 and 2, assuming that the second gray scale value clipping range is [ -1200,600], if the value of a certain pixel in the lung ROI numerical matrix is between [ -1200,600], the value of the certain pixel is unchanged; if the value of a certain pixel point in the lung ROI numerical matrix is larger than 600, the value is made to be equal to 600; if the value of a certain pixel in the lung ROI numerical matrix is less than-1200, the value is made equal to-1200.

TABLE 1

TABLE 2

Table 1 is the original lung ROI numerical matrix (representing the original lung ROI image); table 2 shows a lung ROI matrix (representing a clipped lung ROI image) clipped according to the gray scale value clipping range, wherein the bold font is the value of the changed pixel point.

After cropping, the data in the lung ROI matrix is normalized by the following formula, so that all values in the cropped lung ROI matrix become a value between 0 and 1:

（4）

wherein t represents the value of a pixel in the clipped lung ROI numerical matrix, i.e. the numerical value in Table 2. For example: the value of-1200 becomes 0 after normalization, 1 after 600 normalization, and 0.83 after 300 normalization.

And S108, randomly cutting the normalized lung ROI.

The random cropping refers to arbitrarily cropping a region in the ROI of the lung according to a preset cropping size, for example, arbitrarily cropping a region of 256 ☓ 192 ☓ 128 pixels.

And S109, training and learning by using the cut lung ROI data and the labeled target region and utilizing a second neural network to obtain a lung target region segmentation model.

In this embodiment, the Attention 3D-Unet is used as the second neural network, and the network structure is shown in FIG. 4. An Attention mechanism is added into an Attention 3D-Unet network, so that the network can learn more important area information.

And S110, segmenting a lung target region according to the CT flat scan image data of the lung of the new coronary pneumonia patient and the lung target region segmentation model.

Referring to table 3, in the present embodiment, 1000 cases of lung CT flat scan image data of new coronary pneumonia patients of different ages from multiple centers and regions are collected, 800 cases are selected as training data volume, and neural network Attention 3D-Unet training learning is performed to obtain a lung target region segmentation model; 200 cases are selected as test data quantity, and a lung target region segmentation model is input to segment the lung target region. As can be seen from Table 3, the scores of the training set loss function dice and the test set loss function dice are both close to 1, and the scores of the training set loss function dice and the test set loss function dice are close to each other, so that the accuracy and the generalization of the new coronary pneumonia patient lung target region segmentation model are verified. Meanwhile, the accuracy of the lung target region segmentation model can be verified through lung CT images: FIGS. 5a and 5b are schematic diagrams of CT flat scan images of a lung of a new coronary pneumonia patient corresponding to a certain test data, and schematic diagrams of a target region of the lung segmented by using a segmentation model of the target region of the lung, respectively; as can be seen from the comparison between fig. 5a and fig. 5b, the similarity between the target region of the lung segmented by the target region segmentation model of the lung and the target region labeled in the CT flat scan image of the lung of the original patient with new coronary pneumonia is very high, so that the accuracy of the target region segmentation model of the lung is verified again.

TABLE 3

According to the lung CT image data segmentation method provided by the embodiment of the invention, the CT image and the artificial intelligence technology are combined, and the artificial intelligence deep learning technology is utilized, so that a computer learns a large number of lung CT image characteristics in a short time, and the processing efficiency of lung CT image data is improved. The lung CT image data segmentation method provided by the invention can rapidly segment the target region in the lung CT image data, further carry out three-dimensional reconstruction on the segmented target region, is convenient for visual display, can be widely applied to the examination of various lung diseases, such as pulmonary nodules, lung cancer, new coronary pneumonia and the like, and provides a clearer and more accurate target region image for doctors.

Referring to fig. 6, an embodiment of the present invention further provides a lung CT image data segmentation system, including:

the first training learning module is used for performing training learning by using the image data processed by the first cutting normalization module and the labeled lung contour data of the labeling module through a first neural network to obtain a lung contour segmentation model;

Wherein, the labeling module further comprises:

the acquisition unit is used for acquiring lung CT flat scanning image data;

Wherein the first clipping normalization module further comprises:

the first setting unit is used for presetting a first gray value cutting range;

the first clipping unit clips a numerical matrix of the image according to a first gray value clipping range preset by the first setting unit;

the first normalization unit is used for changing all values in the value matrix of the image clipped by the first clipping unit into a value between 0 and 1.

Wherein the second clipping normalization module further comprises:

the establishing unit is used for taking the external cube of the lung contour extracted by the extracting unit as the interested area of the lung;

the second adjusting unit is used for adjusting the resolution of the lung region of interest determined by the establishing unit to a second resolution uniform value;

According to the lung CT image data segmentation method and system provided by the embodiment of the invention, the CT image and the artificial intelligence technology are combined, and the artificial intelligence deep learning technology is utilized, so that a computer learns a large number of lung CT image characteristics in a short time, and the processing efficiency of lung CT image data is improved. The lung CT image data segmentation method and the lung CT image data segmentation system can rapidly segment the target region in the lung CT image data, further carry out three-dimensional reconstruction on the segmented target region, are convenient for visual display, can be widely applied to the examination of various lung diseases, such as pulmonary nodules, lung cancer, new coronary pneumonia and the like, and provide clear and accurate target region images for doctors.

In practical applications, each functional module and each unit involved in this embodiment may be implemented by a computer program running on computer hardware, and the program may be stored in a computer-readable storage medium, and when executed, may include the flow of the embodiments of the methods described above. Wherein, the hardware refers to a server or a desktop computer, a notebook computer, etc. containing one or more processors and storage media; the storage medium can be a magnetic disk, an optical disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), or the like; the computer program is implemented in a computer language not limited to C, C + +, or the like.

The above-mentioned embodiments are intended to illustrate the objects, technical solutions and advantages of the present invention in further detail, and it should be understood that the above-mentioned embodiments are only exemplary embodiments of the present invention, and are not intended to limit the present invention, and any modifications, equivalents, improvements and the like made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A lung CT image data segmentation method is characterized by comprising the following steps:

2. The method of claim 1, wherein the step of labeling the lung contour and the target region in the lung CT flat scan image data comprises:

acquiring lung CT flat scanning image data and cleaning the image data;

3. The method of segmenting lung CT image data as set forth in claim 1, wherein the step of numerically clipping and normalizing the image data includes:

adjusting the image resolution to a first resolution uniform value;

4. The method of segmenting lung CT image data as set forth in claim 3, wherein the step of numerically cropping and normalizing the lung region of interest includes:

5. The method of segmenting lung CT image data as in claim 4, wherein the first and second uniform resolution values are each set to [1mm, 1mm, 2.5mm ], the first gray value clipping range is set to [ -200,400], and the second gray value clipping range is set to [ -1200,600 ].

6. The pulmonary CT image data segmentation method of claim 1, wherein the first neural network is a 3D-Unet network, and the loss functions used in the 3D-Unet network training process include dice-loss and Focal-loss; the second neural network is a 3D-Unet network added with an attention mechanism.

7. A pulmonary CT image data segmentation system, comprising:

8. The pulmonary CT image data segmentation system of claim 7, wherein the labeling module comprises:

the acquisition unit is used for acquiring lung CT flat scanning image data;

9. The pulmonary CT image data segmentation system of claim 8, wherein the first crop normalization module comprises:

the first setting unit is used for presetting a first gray value cutting range;

10. The pulmonary CT image data segmentation system of claim 9, wherein the second crop normalization module comprises: