CN105574859A - Liver tumor segmentation method and device based on CT (Computed Tomography) image - Google Patents

Abstract

The invention provides a liver tumor segmentation method and device based on CT images. The method includes: performing Gaussian denoising on the CT image data of the liver, transforming it into standardized data with a gray mean value of 0 and a variance of 1 and performing a downsampling operation; extracting lesion slices and Normal tissue slices are divided into positive samples and negative samples; a multi-level deep convolutional neural network is constructed, the model is trained by stochastic gradient descent method, and the network model is obtained, and the rough segmented binary image and pixel classification of the tumor are obtained through the classifier The probability image of the tumor; the morphological erosion operation is performed on the rough segmented binary image of the tumor to obtain the foreground image required by the graph cut, and then the binary image of the liver is subtracted from the rough segmented binary image of the tumor and the morphological Corrosion operation is performed to obtain the background image corresponding to the normal tissue of the liver; an undirected graph is constructed, and the final segmented area of the tumor is obtained using the graph cut optimization algorithm.

Description

A liver tumor segmentation method and device based on CT images

technical field

The invention belongs to the field of medical image processing, in particular to a CT image-based liver tumor segmentation method and device.

Background technique

The liver is an important and complex functional organ to maintain the life activities of the human body. There are many types of liver lesions, and the incidence rate is high. Computed Tomography (CT) images have become one of the important routine means in clinical diagnosis and an important means of examination for liver diseases. Currently, the treatments for liver tumors mainly include tumor resection, interventional therapy, and radiotherapy, among which tumor resection is the most effective treatment. These treatment methods all require accurate knowledge of the number, location, size, and shape of the tumor before surgery, which is helpful for the formulation of a treatment plan for liver tumors. , size, shape, gray scale and texture are different, it is difficult to develop a general tumor segmentation algorithm. Manual segmentation requires anatomical knowledge and experience, and is highly subjective, requiring a lot of time and effort. Due to factors such as blurred tumor borders and large differences in performance, most liver segmentation methods cannot achieve the accuracy required by clinical practice.

The existing fully automatic liver tumor segmentation method, the main process is to manually extract features from the training data, select features, design a classifier, obtain a prediction model through supervised learning or unsupervised learning, and predict the test data according to this model. The extraction process is computationally intensive and time-consuming, and whether good features can be selected largely depends on experience and luck.

Contents of the invention

The technical problem to be solved by the present invention is to overcome the limitation of manual design and extraction of features in traditional segmentation methods.

In order to achieve the above purpose, an embodiment of the present invention provides a liver tumor segmentation method based on CT images, including: step 1, performing Gaussian denoising on the CT image data of the liver, and converting it into a gray scale with a mean value of 0 and a variance of 1 and perform downsampling operation; step 2, extract lesion slices and normal tissue slices from the gold standard image of the CT image of the liver, and divide them into positive samples and negative samples according to the labels of the central pixel points of the slices, And use the random sampling method to make the number of positive and negative samples equal; step 3, build a multi-level deep convolutional neural network, train the model through the stochastic gradient descent method, and automatically learn the tumor characteristics and liver normal tissue characteristics with supervision to obtain the network model , obtain the rough segmented binary image of the tumor and the probability image of pixel classification through the classifier; step 4, perform morphological erosion on the coarse segmented binary image of the tumor to obtain the foreground image required by the graph cut, and then divide the liver The binary image of the tumor is subtracted from the rough segmented binary image of the tumor and the morphological erosion operation is performed to obtain the background image corresponding to the normal tissue of the liver; step 5, construct an undirected graph according to the foreground image and the background image, using The graph cut optimization algorithm obtains the final segmented area of the tumor.

Further, in one embodiment, in the step 2, according to the label of the center pixel of the slice, it is divided into positive samples and negative samples respectively, including: the positive sample corresponds to the tumor slice, and the negative sample corresponds to the normal tissue slice, in order The number of positive and negative samples input to the training model is balanced, and the random sampling method is used to make the number of positive and negative samples equal.

Further, in one embodiment, in the step 3, the rough segmented binary image of the tumor and the probability image of pixel classification are obtained through the classifier, including: classifying each The pixel points are classified to obtain the probability value of belonging to the tumor or the normal tissue of the liver, and according to the classification of the probability value of the category, the rough segmented binary image of the tumor and the probability image of the pixel classification are obtained.

Further, in one embodiment, in the step 5, an undirected graph is constructed according to the foreground image and the background image, and a graph cut optimization algorithm is used to obtain the final segmented region of the tumor, including: according to the foreground image and the background image An undirected graph is constructed from the image, and the energy function is optimized to the minimum using the maximum flow/minimum cut algorithm, and all pixels are divided into targets or backgrounds, which are marked as 1 and 0, respectively, to obtain the final segmented area of the tumor.

In order to achieve the above object, an embodiment of the present invention also provides a liver tumor segmentation device based on CT images, including: an image preprocessing module, which is used to perform Gaussian denoising on the CT image data of the liver, and convert it into gray mean 0, normalized data with a variance of 1 and down-sampling operation; the sample acquisition module is used to extract lesion slices and normal tissue slices from the gold standard image of the CT image of the liver, and divide them according to the label of the center pixel point of the slice. Divide it into positive samples and negative samples, and use random sampling to make the number of positive and negative samples equal; the model training module is used to build a multi-level deep convolutional neural network, train the model through the stochastic gradient descent method, and supervise automatic learning The characteristics of the tumor and the characteristics of the normal liver tissue are obtained to obtain the network model, and the rough segmentation binary image of the tumor and the probability image of the pixel classification are obtained through the classifier; the erosion operation module is used to perform morphological erosion on the rough segmentation binary image of the tumor Operation, to obtain the foreground image required by the graph cut, and then subtract the binary image of the liver from the rough segmented binary image of the tumor and perform the morphological erosion operation to obtain the background image corresponding to the normal tissue of the liver; segmented area generation A module for constructing an undirected graph according to the foreground image and the background image, and using a graph cut optimization algorithm to obtain the final segmented region of the tumor.

Further, in one embodiment, the sample collection module divides them into positive samples and negative samples respectively according to the label of the pixel point in the center of the slice, specifically including: the positive sample corresponds to the tumor slice, and the negative sample corresponds to the normal tissue slice, in order to make The number of positive and negative samples input to the training model is balanced, and the random sampling method is used to make the number of positive and negative samples equal.

Further, in one embodiment, the model training module obtains the rough segmented binary image of the tumor and the probability image of pixel classification through a classifier, which specifically includes: classifying each pixel in the image through the classifier of the last layer in the network The points are classified to obtain the probability value of whether they belong to the tumor or the normal tissue of the liver. According to the classification of the probability value of the category, the rough segmentation binary image of the tumor and the probability image of the pixel classification are obtained.

Further, in one embodiment, the segmented region generation module constructs an undirected graph according to the foreground image and the background image, and uses a graph cut optimization algorithm to obtain the final segmented region of the tumor, which specifically includes: according to the foreground image and the background image An undirected graph is constructed from the image, and the energy function is optimized to the minimum using the maximum flow/minimum cut algorithm, and all pixels are divided into targets or backgrounds, which are marked as 1 and 0, respectively, to obtain the final segmented area of the tumor.

The present invention proposes a fully automatic method and device for segmenting liver tumors based on CT images, which automatically learns features through a deep learning model and extracts richer essential features in the data set, which is more reliable than manually designed and extracted features. The tumor segmentation results are optimized through the graph cut method, making the final segmentation more accurate and robust; moreover, the entire segmentation process does not require any manual intervention, which can provide accurate tumor information for the diagnosis and treatment of liver tumors, and is helpful to improve the quality of liver tumors. The success rate of surgery.

Description of drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only These are some embodiments of the present invention. For those skilled in the art, other drawings can also be obtained according to these drawings on the premise of not paying creative efforts.

Fig. 1 is the method flowchart of the liver tumor segmentation method based on CT image according to the embodiment of the present invention;

2 is a schematic diagram of a multi-level deep convolutional neural network constructed in an embodiment of the present invention;

FIG. 3 is a schematic structural diagram of an apparatus for segmenting liver tumors based on CT images according to an embodiment of the present invention.

detailed description

The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some, not all, embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

The invention discloses a fully automatic liver tumor segmentation method based on CT images, the main process of which is: performing denoising, standardization, and down-sampling preprocessing operations on a series of original CT liver images; training images and test images Separately extract the positive and negative samples required for training and the samples required to predict the label for testing; train the deep convolutional neural network model according to the training data, and automatically extract features such as tumor edges and textures; use two types of classifiers to obtain rough segmentation regions of liver tumors And the probability image; then use the graph cut method to optimize the rough segmentation results to obtain the final tumor area.

FIG. 1 is a flow chart of a liver tumor segmentation method based on CT images according to an embodiment of the present invention. As shown in Figure 1, it includes: Step S101, Gaussian denoising is performed on the CT image data of the liver, and it is converted into standardized data with a gray mean value of 0 and a variance of 1 and a downsampling operation; Step S102, from the described Extract lesion slices and normal tissue slices from the gold standard image of CT images of the liver, divide them into positive samples and negative samples according to the labels of the center pixel points of the slices, and use random sampling to make the number of positive and negative samples equal; step S103 , build a multi-level deep convolutional neural network (as shown in Figure 2), train the model through the stochastic gradient descent method, and automatically learn the characteristics of tumors and normal liver tissues with supervision to obtain the network model, and obtain the roughness of the tumor through the classifier. Segment the binary image and the probability image of pixel classification; step S104, perform a morphological erosion operation on the coarsely segmented binary image of the tumor to obtain the foreground image required by the graph cut, and then combine the binary image of the liver with the rough segmented image of the tumor Segment the binary image for subtraction and morphological erosion to obtain the background image corresponding to the normal tissue of the liver; step S105, construct an undirected graph based on the foreground image and the background image, and use the graph cut optimization algorithm to obtain the final image of the tumor. Divide the area.

In this embodiment, in the step S101, the CT image data should be preprocessed firstly, firstly, Gaussian denoising is performed on the CT image data of the liver; secondly, normalization processing is performed to convert it into grayscale Data with a mean value of 0 and a variance of 1; third, perform a downsampling operation.

In this embodiment, in the step S102, according to the label of the pixel point in the center of the slice, it is divided into positive samples and negative samples, including: positive samples correspond to tumor slices, negative samples correspond to normal tissue slices, and the actual neutral samples The number is far greater than the number of positive samples. In order to balance the number of positive and negative samples input by the training model, random sampling is used to make the number of positive and negative samples equal.

In this embodiment, in the step S103, the coarse segmented binary image of the tumor and the probability image of pixel classification are obtained by the classifier, including: classifying each pixel in the image by the classifier of the last layer in the network Classify to obtain the probability value of whether it belongs to the tumor or the normal liver tissue, and classify according to the magnitude of the probability value of the category to obtain the rough segmentation binary image and the probability image of the pixel classification of the tumor.

In this embodiment, the constructed deep convolutional neural network structure includes a convolutional layer, a downsampling layer, a fully connected layer and a softmax classification layer. Among them, the convolutional layer obtains image features through convolution operations, such as tumor edge, texture and other features; the downsampling layer subsamples the obtained feature image to reduce the amount of data processing while maintaining useful feature information; the fully connected layer uses To capture the complex relationship between output features; softmax is a multi-class classifier, here is used for two types of classification, tumor and normal liver tissue, its output is a conditional probability value, between 0 and 1.

The transfer function in the network adopts a linear correction transfer function: max(0,x). If the output value of the neuron node in the network is less than zero, it is set to 0 through the transfer function, and if it is greater than zero, it remains unchanged. It can maintain the model's Sparsity allows a multilayer neural network to learn faster.

In addition, in this embodiment, the stochastic gradient descent method is used to minimize the loss function To obtain the network parameters, that is, the weight value θ _i corresponding to the connection between the neuron nodes in the network, the stochastic gradient descent method is to update θ _i through each sample, ${\mathrm{\θ}}_j^{\′}={\mathrm{\θ}}_j+({\mathrm{the\; y}}^i-h_{\mathrm{\θ}}(x^i\left)\right)x_j^i.$

In this embodiment, in the step S105, an undirected graph is constructed according to the foreground image and the background image, and a graph cut optimization algorithm is used to obtain the final segmented region of the tumor, including: constructing an undirected graph according to the foreground image and the background image. Directed graph, using the maximum flow/minimum cut algorithm to optimize the energy function to achieve the minimum, divide all pixels into target or background, and mark them as 1 and 0, respectively, to obtain the final segmented area of the tumor.

Corresponding to the above method, as shown in FIG. 3 , it is a schematic structural diagram of an apparatus for segmenting liver tumors based on CT images according to an embodiment of the present invention. The device in this embodiment includes: an image preprocessing module 101, which is used to perform Gaussian denoising on the CT image data of the liver, convert it into normalized data with a gray mean value of 0, and a variance of 1, and perform a downsampling operation; sample collection Module 102, for extracting lesion slices and normal tissue slices from the gold standard image of the CT image of the liver, dividing them into positive samples and negative samples according to the labels of the central pixel points of the slices, and adopting a random sampling method such that The number of positive and negative samples is equal; the model training module 103 is used to construct a multi-level deep convolutional neural network, train the model through the stochastic gradient descent method, and automatically learn tumor characteristics and liver normal tissue characteristics with supervision to obtain a network model. The device obtains the rough segmentation binary image of the tumor and the probability image of pixel classification; the erosion operation module 104 is used to perform a morphological erosion operation on the rough segmentation binary image of the tumor to obtain the foreground image required by the graph cut, and then The binary image of the liver is subtracted from the roughly segmented binary image of the tumor and the morphological erosion operation is performed to obtain a background image corresponding to the normal tissue of the liver; the segmented area generation module 105 is used to Construct an undirected graph, and use the graph cut optimization algorithm to obtain the final segmented region of the tumor.

In this embodiment, the sample collection module 102 divides them into positive samples and negative samples according to the labels of the pixels in the center of the slice, specifically including: the positive samples correspond to tumor slices, and the negative samples correspond to normal tissue slices. In order to make the training model The number of input positive and negative samples is balanced, and the random sampling method is used to make the number of positive and negative samples equal.

In this embodiment, the model training module 103 obtains the rough segmented binary image of the tumor and the probability image of pixel classification through the classifier, which specifically includes: classifying each pixel in the image through the classifier of the last layer in the network Classify to obtain the probability value of whether it belongs to the tumor or the normal liver tissue, and classify according to the magnitude of the probability value of the category to obtain the rough segmentation binary image and the probability image of the pixel classification of the tumor.

In this embodiment, the segmented region generating module 105 constructs an undirected graph according to the foreground image and the background image, and obtains the final segmented region of the tumor using a graph cut optimization algorithm, which specifically includes: constructing a graph based on the foreground image and the background image For an undirected graph, the maximum flow/minimum cut algorithm is used to optimize the energy function to the minimum, and all pixels are divided into target or background, which are marked as 1 and 0, respectively, to obtain the final segmented area of the tumor.

The method of the present invention has passed multiple sets of CT3D image tests, and experiments show that the method has high accuracy and robustness.

The method and device for liver tumor segmentation based on CT images of the present invention automatically learns features through deep learning models, extracts richer essential features in the data set, and has better separability than manual design and extraction features. The method optimizes the results of tumor segmentation, making the final segmentation more accurate and robust; moreover, the whole segmentation process does not require any manual intervention, and can provide accurate tumor information for the diagnosis and treatment of liver tumors, which helps to improve the success rate of liver tumor surgery.

Those skilled in the art should understand that the embodiments of the present invention may be provided as methods, systems, or computer program products. Accordingly, the present invention can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) having computer-usable program code embodied therein.

The present invention is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It should be understood that each procedure and/or block in the flowchart and/or block diagram, and a combination of procedures and/or blocks in the flowchart and/or block diagram can be realized by computer program instructions. These computer program instructions may be provided to a general purpose computer, special purpose computer, embedded processor, or processor of other programmable data processing equipment to produce a machine such that the instructions executed by the processor of the computer or other programmable data processing equipment produce a An apparatus for realizing the functions specified in one or more procedures of the flowchart and/or one or more blocks of the block diagram.

These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing apparatus to operate in a specific manner, such that the instructions stored in the computer-readable memory produce an article of manufacture comprising instruction means, the instructions The device realizes the function specified in one or more procedures of the flowchart and/or one or more blocks of the block diagram.

These computer program instructions can also be loaded onto a computer or other programmable data processing device, causing a series of operational steps to be performed on the computer or other programmable device to produce a computer-implemented process, thereby The instructions provide steps for implementing the functions specified in the flow chart or blocks of the flowchart and/or the block or blocks of the block diagrams.

In the present invention, specific examples have been applied to explain the principles and implementation methods of the present invention, and the descriptions of the above examples are only used to help understand the method of the present invention and its core idea; meanwhile, for those of ordinary skill in the art, according to this The idea of the invention will have changes in the specific implementation and scope of application. To sum up, the contents of this specification should not be construed as limiting the present invention.

Claims (8)

1. based on a liver neoplasm dividing method for CT image, it is characterized in that, described method comprises:

Step 1, carries out Gauss except making an uproar to the CT view data of liver, and being translated into gray average is 0, and variance is the standardized data of 1 and carries out down-sampling operation;

Step 2, pathology section and normal tissue sections is extracted from the goldstandard image of the CT image of described liver, label according to centre of slice pixel is divided into positive sample and negative sample respectively, and adopts the method for stochastic sampling to make positive and negative sample size equal;

Step 3, build multi-level degree of depth convolutional neural networks, by stochastic gradient descent method training pattern, have automatic learning tumoral character and the liver normal structure feature of supervision, obtain network model, obtain the coarse segmentation bianry image of tumour and the probabilistic image of pixel classifications by sorter;

Step 4, morphological erosion operation is carried out to the coarse segmentation bianry image of described tumour, acquisition figure cuts required foreground image, again the bianry image of liver done phase reducing with the coarse segmentation bianry image of tumour and carry out morphological erosion operation, obtaining the background image corresponding to liver normal structure;

Step 5, build non-directed graph according to described foreground image and background image, use figure cuts the final cut zone that optimized algorithm obtains tumour.

2. the liver neoplasm dividing method based on CT image according to claim 1, is characterized in that, in described step 2, the label according to centre of slice pixel is divided into positive sample and negative sample respectively, comprising:

The corresponding tumor biopsy of positive sample, the corresponding normal tissue sections of negative sample, the positive and negative sample size inputted to make training pattern is balanced, adopts the method for stochastic sampling to make positive and negative sample size equal.

3. the liver neoplasm dividing method based on CT image according to claim 1, is characterized in that, in described step 3, obtains the coarse segmentation bianry image of tumour and the probabilistic image of pixel classifications, comprising by sorter:

By the sorter of one deck last in network, pixel each in image is classified, obtain belonging to the probable value that tumour still belongs to liver normal structure, according to the magnitude classification of the probable value of generic, obtain the coarse segmentation bianry image of described tumour and the probabilistic image of pixel classifications.

4. the liver neoplasm dividing method based on CT image according to claim 1, is characterized in that, in described step 5, build non-directed graph according to described foreground image and background image, use figure cuts the final cut zone that optimized algorithm obtains tumour, comprising:

Build non-directed graph according to described foreground image and background image, utilize max-flow/minimal cut algorithm optimization energy function to make it reach minimum, whole pixels is divided into target or background, be labeled as 1 and 0 respectively, obtain the final cut zone of tumour.

5. based on a liver neoplasm segmenting device for CT image, it is characterized in that, described device comprises:

Image pre-processing module, for carrying out Gauss except making an uproar to the CT view data of liver, being translated into gray average is 0, and variance is the standardized data of 1 and carries out down-sampling operation;

Sample collection module, pathology section and normal tissue sections is extracted in goldstandard image for the CT image from described liver, label according to centre of slice pixel is divided into positive sample and negative sample respectively, and adopts the method for stochastic sampling to make positive and negative sample size equal;

Model training module, for building multi-level degree of depth convolutional neural networks, by stochastic gradient descent method training pattern, there are automatic learning tumoral character and the liver normal structure feature of supervision, obtain network model, obtain the coarse segmentation bianry image of tumour and the probabilistic image of pixel classifications by sorter;

Etching operation module, for carrying out morphological erosion operation to the coarse segmentation bianry image of described tumour, acquisition figure cuts required foreground image, again the bianry image of liver done phase reducing with the coarse segmentation bianry image of tumour and carry out morphological erosion operation, obtaining the background image corresponding to liver normal structure;

Cut zone generation module, for building non-directed graph according to described foreground image and background image, use figure cuts the final cut zone that optimized algorithm obtains tumour.

6. the liver neoplasm segmenting device based on CT image according to claim 5, is characterized in that, described sample collection module is divided into positive sample and negative sample respectively according to the label of centre of slice pixel, specifically comprises:

The corresponding tumor biopsy of positive sample, the corresponding normal tissue sections of negative sample, the positive and negative sample size inputted to make training pattern is balanced, adopts the method for stochastic sampling to make positive and negative sample size equal.

7. the liver neoplasm segmenting device based on CT image according to claim 5, is characterized in that, described model training module obtains the coarse segmentation bianry image of tumour and the probabilistic image of pixel classifications by sorter, specifically comprises:

By the sorter of one deck last in network, pixel each in image is classified, obtain belonging to the probable value that tumour still belongs to liver normal structure, according to the magnitude classification of the probable value of generic, obtain the coarse segmentation bianry image of described tumour and the probabilistic image of pixel classifications.

8. the liver neoplasm segmenting device based on CT image according to claim 5, it is characterized in that, described cut zone generation module builds non-directed graph according to described foreground image and background image, and use figure cuts the final cut zone that optimized algorithm obtains tumour, specifically comprises:

Build non-directed graph according to described foreground image and background image, utilize max-flow/minimal cut algorithm optimization energy function to make it reach minimum, whole pixels is divided into target or background, be labeled as 1 and 0 respectively, obtain the final cut zone of tumour.