CN110969622A

CN110969622A - Image processing method and system for assisting pneumonia diagnosis

Info

Publication number: CN110969622A
Application number: CN202010126181.8A
Authority: CN
Inventors: 曾凯; 何健; 郭桐; 冯亚崇; 杨奎; 吴小页
Original assignee: Nanjing Anke Medical Technology Co Ltd
Current assignee: Nanjing Anke Medical Technology Co Ltd
Priority date: 2020-02-28
Filing date: 2020-02-28
Publication date: 2020-04-07
Anticipated expiration: 2040-02-28
Also published as: CN110969622B

Abstract

The invention relates to the technical field of medical imaging, and discloses an image processing method and system for assisting pneumonia diagnosis, wherein the method comprises the following steps: obtaining a chest CT initial image of a thin layer; segmenting the initial image to extract a lung region image; processing the lung region image to obtain an image with the blood vessels in the lung removed; detecting the image without the pulmonary blood vessels by adopting a neural network, outputting the image with an inflammation marker, and training the neural network, wherein the training comprises pre-training and accurate training by adopting different training data sets and training objects; and segmenting a focus area in the image with the inflammation mark, and carrying out image visualization on the focus area to obtain a final image for a doctor to look up. Based on various image processing methods and neural network detection technologies, the neural network training problem of limited clinical pneumonia data can be solved, stable diagnosis accuracy is obtained, and clinical diagnosis efficiency is improved.

Description

Image processing method and system for assisting pneumonia diagnosis

Technical Field

The invention relates to the technical field of medical imaging, in particular to an image processing method and system for assisting pneumonia diagnosis.

Background

Pneumonia is an acute respiratory infectious disease affecting lung respiration, and is easy to infect, and the death rate is high, so that high requirements are put on a rapid detection and diagnosis technology of pneumonia. The pneumonia is generally detected and diagnosed by matching physical signs and chest X-ray at present, but the potential cause of the pneumonia cannot be easily determined by the method, because no standard can clearly distinguish bacterial pneumonia from non-bacterial pneumonia. Because of the low contrast of the X-ray film, the aliasing of the tissue results is serious, and the pneumonia may not be seen from the X-ray film in the early stage, especially under the condition of dehydration, if other lung medical history of obesity exists, the pneumonia is difficult to judge. Therefore, more and more doctors now use chest CT to diagnose pneumonia. Chest CT usually makes a fast helical scan of the upper abdomen of the patient's chest, and then reconstructs a thin CT image, and the doctor diagnoses the infection level of the patient's lungs through the hundreds of tomographic CT images.

Especially for unknown viral pneumonia, when the virus is infected by large-scale people, the virus mutation speed is very fast, and the workload of hospital radiology departments and clinicians is very large. In the early stage, the traditional diagnosis method of the viral pneumonia mostly adopts a kit based on viral nucleic acid detection, each kit has strong specificity aiming at different viruses, only has response to the viruses sensitive to the kit, and the test method is complex, consumes long time and has low diagnosis efficiency. Due to the sensitivity of CT detection, more and more hospitals recently begin to use the CT detection result as an important auxiliary diagnostic standard, because many patients who use the kit to detect negativity can also see the lesion on the CT image of the chest, such as the frosted glass-like lesion image, and fig. 2 is a clinical CT image of new coronary pneumonia.

In recent years, image-aided techniques based on neural networks are used in the field of aided diagnosis of CT images, such as aided diagnosis of diseases such as cerebral hemorrhage and pulmonary nodules, but most of the conventional methods based on neural networks are trained based on massive data. Therefore, for newly appearing viral pneumonia, such as new coronavirus pneumonia, the available clinical data are very limited, and the traditional method is not favorable for processing.

Disclosure of Invention

The technical purpose is as follows: in order to solve the technical problems, the invention provides an image processing method and system for assisting pneumonia diagnosis, which can obtain stable diagnosis accuracy and improve clinical diagnosis efficiency in time.

The technical scheme is as follows: in order to achieve the technical purpose, the invention adopts the following technical scheme:

an image processing method for assisting pneumonia diagnosis, characterized by sequentially performing the steps of:

A. carrying out low-dose CT scanning on a patient to obtain a thin-layer chest CT initial image;

B. segmenting the initial image to extract a lung region image;

C. processing the lung region image to obtain an image with the blood vessels in the lung removed;

D. detecting the image with the removed pulmonary blood vessels by adopting a neural network, and outputting an image with an inflammation marker;

E. dividing a focus area in the image with the inflammation mark, and carrying out image visualization on the focus area to obtain a final image for a doctor to look up;

the neural network comprises a trunk network, a convolutional layer and an activation layer, the training of the neural network comprises pre-training and accurate training, the pre-training selects a lung disease data set with focus symptoms similar to pneumonia to be diagnosed to a preset size to train the trunk network, and the accurate training adopts a pneumonia data set with pneumonia type to be diagnosed to train the convolutional layer and the activation layer.

Preferably, the pulmonary disease dataset is selected from a training set of lung nodule CT data using LUNA.

Preferably, the ratio of the number of samples in the pneumonia data set of the pneumonia type to be diagnosed to the number of samples in the pulmonary disease data set is one tenth or less.

Preferably, the neural network is a backbone network of VGG, google-net or unet.

Preferably, an image segmentation method based on a region growing algorithm and threshold segmentation is adopted in the step B.

Preferably, an image segmentation method based on blood vessel characteristics is adopted in the step C.

The invention also discloses a system for assisting pneumonia diagnosis, which is characterized in that: comprising a processor and a memory for storing a program for performing any of the methods of the invention, the memory being communicatively coupled to the processor for executing the program.

Has the advantages that: due to the adoption of the technical scheme, the invention has the following beneficial effects:

based on various image processing methods and neural network detection technologies, the invention trains the main network by adopting a lung disease data set similar to the pneumonia focus of the type to be diagnosed, trains the convolution layer and the activation layer outside the main network by using less pneumonia data, can solve the neural network training problem of limited clinical pneumonia data, obtains stable diagnosis accuracy, can calculate the pneumonia area and identify the pneumonia area in the output image, and improves the clinical diagnosis efficiency.

Drawings

FIG. 1 is a flow chart of an image detection method for assisting diagnosis of new coronavirus pneumonia according to the present invention;

FIG. 2 is a clinical CT image of neocoronavirus pneumonia;

FIG. 3 is a flow chart of an image detection method for assisting pneumonia diagnosis according to the present invention;

FIG. 4 is a flow chart of the pre-training of the neural network of the present invention;

FIG. 5 is a flow chart of the precise training of the neural network of the present invention;

FIG. 6 is a schematic diagram of the neural network of the present invention outputting a final image with pneumonia markers;

FIG. 7 is a flow chart of a region growing method of the present invention.

Detailed Description

The present invention will be further described with reference to the accompanying drawings.

As shown in fig. 1 to 7, the present invention provides an image detection method for assisting pneumonia diagnosis, such as for diagnosing a novel coronary pneumonia, comprising the steps of:

A. carrying out low-dose CT scanning on a patient to obtain thin-layer chest CT image data;

B. segmenting the CT image, and extracting an image of a lung region;

C. processing the image of the lung region to remove blood vessels of the lung;

D. detecting the images without the pulmonary blood vessels by using a neural network, and judging the similarity of the areas and the new coronary pneumonia lesions;

E. and segmenting a focus area in the image with the inflammation mark, and carrying out image visualization on the area for a doctor to consult.

In the invention, the image extraction of the lung region can be completed by using a region growing + threshold segmentation method in the step B. Since the density of lung regions is much lower than that of normal human tissue, usually around-800, and the CT value of normal tissue is around 0, these lung regions can be segmented by using region growing methods, such as watershed algorithm. As shown in fig. 7, the basic idea of region growing is to group regions with similarity features together to form a final region, and the algorithm mainly considers the relationship between a pixel point in an image and a pixel point in the neighborhood thereof. Firstly, determining a seed point, then increasing the region according to a certain similarity measure, gradually generating the region with a certain similarity, and merging adjacent pixels or regions with the similarity until no combinable region or pixel exists. The region growing method needs to solve three problems: 1. selecting seed points; 2. determining a growth criterion; 3. conditions or rules are specified for growth to terminate. Specifically, in the present invention, since the region growing is sensitive to noise and is likely to cause over-segmentation or under-segmentation, the image is first subjected to a smoothing filtering process, such as gaussian filtering. And (3) automatically selecting seed points, wherein the CT value of the lung region is about-800, the lung parenchyma is generally positioned in the central region of the CT image visual field, and the seed points are quickly selected through the limitation of the position and the CT value. And selecting a three-dimensional 26 neighborhood points for judgment according to the growth criterion, wherein the gradient of the neighborhood points and the seed points is smaller than a set threshold (such as 50) and smaller than a threshold set by the tissue (such as-600), the neighborhood points and the seed points are considered as new seed points, and the growth is continued from the new seed points until the growth is finished.

In the invention, the step C processes the image of the lung region to remove the blood vessels in the lung, because the pathological change region of pneumonia is mainly in the alveolar region, the CT value of the region is increased, but the CT value of the blood vessels is higher and is easily confused with the pneumonia region, so that the detection accuracy can be increased by removing the blood vessels in advance. Step C performs segmentation using a vessel feature-based method, such as the Hessian (Hessian) matrix method. By calculating a partial differential matrix for each pixel neighborhood, the likelihood of whether the current pixel belongs to a vascular structure is calculated from this matrix. The following were used:

where f denotes three-dimensional CT data, h (f) denotes a Hessian matrix of f, and h (f) is obtained by calculating two-dimensional partial derivatives of f in each direction. x, y, z represent three directions of f. Eigenvalues through linear algebra theory

Decomposition, this matrix corresponding to three eigenvalues

By comparing these three feature values, the likelihood of whether the corresponding pixel represents a vascular structure can be calculated. Table 1 shows the relationship between feature values and pixels.

TABLE 1

L represents

Very small (e.g. in<1) H represents

Is relatively large (e.g. such as>10)。

In the present invention, step D can detect the lung image with the blood vessels in the lung removed by using a neural network. For example, two steps of pre-training + precision training can be adopted to realize the training problem of limited data.

a. For pre-training, a data set with some similarity to pneumonia may be used for training, such as a training data set for lung nodules. The network generally adopts multilayer convolutional neural networks, such as backbone networks of VGG, google-net, etc., the input of training is medical CT images, and the segmentation result of the current images is output. The loss function may employ a dice loss function. In the whole training process, the weight of the whole network is optimized and updated by traversing the whole data set by using a gradient descent method.

The definition of the Dice loss function is as follows:

wherein Out _ img _ label represents the result output by the neural network, and true _ img _ label represents the real mark corresponding to the current image.

b. For precision training, since the backbone network has learned a large number of features from the data set of lung nodules, training based on a small amount of pneumonia data (e.g., CT data for a small amount of new coronary pneumonia) here updates only the convolutional layer and the activation layer outside the backbone network. The number of samples in the pneumonia data set for the type of pneumonia to be diagnosed is one tenth or less compared with the number of samples in the pulmonary disease data set. The precisely trained loss function is the same as the pre-trained loss function.

The above description is only of the preferred embodiments of the present invention, and it should be noted that: it will be apparent to those skilled in the art that various modifications and adaptations can be made without departing from the principles of the invention and these are intended to be within the scope of the invention.

Claims

1. An image processing method for assisting pneumonia diagnosis, characterized by sequentially performing the steps of:

B. segmenting the initial image to extract a lung region image;

the neural network comprises a trunk network, a convolutional layer and an activation layer, the training of the neural network comprises pre-training and accurate training, the pre-training selects a lung disease data set with focus symptoms similar to pneumonia to be diagnosed to a preset size, the trunk network is trained, and the accurate training trains the convolutional layer and the activation layer by adopting the pneumonia data set of pneumonia type to be diagnosed.

2. The image processing method for assisting pneumonia diagnosis according to claim 1, characterized in that: the pulmonary disease dataset is selected from a training set of lung nodule CT data using LUNA.

3. The image processing method for assisting pneumonia diagnosis according to claim 1, characterized in that: the ratio of the number of samples in the pneumonia data set of the pneumonia type to be diagnosed to the number of samples in the pulmonary disease data set is less than one tenth.

4. The image processing method for assisting pneumonia diagnosis according to claim 1, characterized in that: the neural network adopts a backbone network of VGG, google-net or unet.

5. The image processing method for assisting pneumonia diagnosis according to claim 1, characterized in that: and B, adopting an image segmentation method based on a region growing algorithm and threshold segmentation.

6. The image processing method for assisting pneumonia diagnosis according to claim 1, characterized in that: and C, adopting an image segmentation method based on the blood vessel characteristics.

7. A system for assisting pneumonia diagnosis implementing the method of any one of claims 1 to 6, characterized by: comprising a processor and a memory, said memory storing a program for executing said method, the memory being communicatively coupled to the processor, the processor being configured to execute said program.