1. A method of vessel segmentation incorporating global and neighborhood information, the method comprising the steps of:

step 1, data preprocessing: carrying out skull rejection and bias field correction to generate a training sample;

step 2, constructing a vessel segmentation network Gnet

Building a network model Gnet combined with global information, training the network model by using the training sample generated in the step 1, and inputting the medical image to be segmented into the built network model so as to generate a target image in a corresponding mode;

step 3, constructing a vascular network Rnet of neighborhood information

And (2) providing a convolutional neural network Rnet based on planar twenty-six neighborhood connectivity by combining the connectivity of a blood vessel region, wherein the input of the network is the output of the Gnet constructed in the step 2, the network consists of a plurality of continuous convolutional layers and full-connection layers, and the numerical value of 1 multiplied by 1 is output through the full-connection layers after 4 times of continuous operations of convolution and maximum pooling. Training the Rnet network model by using Mean Square Error (MSE) as a loss function, wherein the Rnet network model is defined as follows:

wherein Ii and Oi represent the number of disconnected Islands in the input and output blood vessels, and the smaller the mean square error of the island number Islands is calculated by a Skimage image processing tool kit, the higher the accuracy rate of Rnet describing the number of disconnected Islands in the segmentation graph output by Gnet is;

step 4, training a segmentation network combining global and neighborhood information

The comprehensive Loss function Loss (lambda, eta) is the sum of weighted Loss functions G _ Loss of Gnet and R _ Loss of Rnet, and the formula is expressed as:

Loss(λ,η)＝λU_Loss+ηI_Loss

wherein λ and η are weight parameters given to two different loss functions of the Gnet and the Rnet, and λ + η is 1; the influence degree of the two networks on the segmentation result is adjusted by distributing different weight values of the two network loss functions, and the segmentation effect is optimized.

2. The vessel segmentation method combining global and neighborhood information as claimed in claim 1, wherein said step 2, constructing the vessel segmentation network Gnet comprises the steps of:

2.1 optimizing the jump connection process by adding a ConvLSTM network module. The ConvLSTM network optimized jump connection process can further combine the spatial information of the image to strengthen the propagation and coding of the original characteristics so as to eliminate the ambiguity generated by irrelevant noise effect in the jump connection;

2.2, optimizing a down-sampling process by adding an attention mechanism, aiming at the problem of loss of characteristic information of tiny blood vessels in the down-sampling process, adding an attention gate module, and simultaneously combining multi-scale original image input on each layer of down-sampling; the proposed down-sampling optimization fully combines the feature information of the original image, can inhibit the feature correspondence of irrelevant areas, automatically learns and distinguishes the appearance and the size of the target, and focuses on the significant features of the target blood vessel area.

3. A method as claimed in claim 1 or 2, wherein the step 4 of training the segmentation network combining the global and neighborhood information comprises the following steps:

4.1 Gnet and Rnet are trained independently and in parallel, Gnet is trained by G _ loss, Rnet is trained by R _ loss, and the two networks have no key, only the output of Gnet is used as the input of Rnet,

4.2 until R _ loss is stable, carrying out combined training on Gnet and Rnet, and freezing network parameters when the Rnet is stable during independent training; the training mode at this time is that the Gnet segmentation network is trained through a comprehensive Loss function Loss (lambda, eta), and the lambda and eta values are adjusted through experiments to obtain the optimal solution of the segmentation result.

Background

At present, cerebrovascular diseases become one of the diseases with the highest fatality rate and the lowest improvement rate in neurosurgical diseases. Whether the medical image is accurately segmented or not determines whether a doctor can provide reliable diagnosis and treatment basis in clinic. In addition, in different clinical medicine fields such as neurosurgery and cardiovascular and cerebrovascular, the segmentation and reconstruction of blood vessels are important for disease diagnosis, treatment schemes and evaluation of clinical results. Therefore, accurate and fast segmentation of blood vessels has become one of the hot spots in medical imaging research.

The existing medical image segmentation methods can be divided into two categories, the first category is the traditional semi-automatic segmentation method, such as a threshold method, a tracking-based method, a clustering-based method, a model-based segmentation method and the like, but the methods not only consume a large amount of manual time intervention and operation, but also depend on professional knowledge and experience of experts seriously, and therefore, a large amount of subjective differences exist. The second category is a segmentation algorithm based on artificial intelligence, represented by deep learning. With the rise of the enthusiasm of deep learning research and its strong performance in medical image segmentation, the related research of medical image segmentation based on deep learning is rapidly growing.

In recent years, U-Net networks proposed by researchers have demonstrated excellent performance in the field of medical image segmentation. The special U-shaped structure and skip-connection operation of the medical image matching method enable the network to be combined with information of low resolution and high resolution, and characteristics of the medical image are well adapted. However, the CNN-based U-net network inevitably brings up the calculation amount, and the excessive and redundant use of the calculation resources and the model parameters causes the model to repeatedly extract similar low-level features, thereby reducing the segmentation capability to some extent. When the U-net network is applied to cerebrovascular segmentation, the phenomena of breakage of small blood vessels of the brain, non-communication with large blood vessels and the like can occur, and the accuracy of the blood vessel segmentation is greatly reduced. The combination of the importance of the cerebral vessels and the specificity of their location, the accuracy of the segmentation and the imprecision of the results create enormous obstacles and challenges for current research and clinical applications.

Disclosure of Invention

The existing method based on deep learning mainly considers the information of the image from the global perspective, so a large number of disconnected regions occur, and the boundary segmentation effect is poor. In order to improve the accuracy of cerebral vessel segmentation, the invention provides a vessel segmentation method combining global and neighborhood information, which obviously improves the accuracy of cerebral vessel segmentation in MRA image processing and can effectively solve the problem of fine vessel segmentation fracture.

The technical scheme adopted by the invention for solving the technical problems is as follows:

a method of vessel segmentation incorporating global and neighborhood information, the method comprising the steps of:

step 1 data preprocessing

The proportion of the cerebrovascular system in the intracranial volume is relatively small (about 1% -5%), the performance of the FMM depends on the fitting of a background region (namely a non-vascular region) to a great extent, the fitting dependence on the vascular region is small, and meanwhile, the large-area background reserved in the data can cause calculation burden and influence the processing time of the model; therefore, skull rejection and bias field correction are carried out to generate training samples;

step 2, constructing a vessel segmentation network Gnet

Building a network model Gnet combined with global information, training the network model by using the training sample generated in the step 1, and inputting the medical image to be segmented into the built network model so as to generate a target image in a corresponding mode;

further, in step 2, the vessel segmentation network Gnet uses a brand-new down-sampling mode on the basis of the Unet network and combines a long-term memory network with a short-term memory network at a jump junction, so that the extraction of feature information is optimized, and the loss of image information is reduced, and the specific process is as follows:

2.1, a ConvLSTM network module is added to optimize a jump connection process, the jump connection process optimized by the ConvLSTM network can further combine spatial information of images to strengthen the propagation and coding of original characteristics so as to eliminate ambiguity generated by irrelevant noise effect in jump connection and enable segmentation and positioning to be more accurate;

2.2, optimizing a down-sampling process by adding an attention mechanism, aiming at the problem of loss of characteristic information of tiny blood vessels in the down-sampling process, adding an attention gate module, and simultaneously combining multi-scale original image input on each layer of down-sampling; the provided downsampling optimization fully combines the characteristic information of the original image, can inhibit the characteristic correspondence of irrelevant areas, automatically learns and distinguishes the appearance and the size of the target, focuses on the significant characteristics of the target blood vessel area, and reduces the downsampling information loss;

step 3, constructing a vascular network Rnet of neighborhood information

In order to solve the phenomena that the tiny blood vessels divided by the network are broken and are not communicated with the big blood vessels, the invention combines the connectivity of the blood vessel region and provides a convolutional neural network Rnet based on the connectivity of twenty-six neighborhoods on a plane; the input of the network is the output of Gnet constructed in the step 2, the network is composed of a plurality of continuous convolution layers and full connection layers, and the value of 1 multiplied by 1 is finally output through the full connection layers after 4 times of continuous operations of convolution and maximum pooling. Training the Rnet network model by using Mean Square Error (MSE) as a loss function, wherein the Rnet network model is defined as follows:

wherein Ii and Oi represent the island number Islands of the disconnected blood vessels of input and output, the calculation is carried out by a Skimage image processing tool kit, the mean square error MSE of the island number Islands is smaller, and Rnet describes that the accuracy rate of the island number Islands of the disconnected blood vessels in a segmentation graph output by Gnet is higher;

step 4, training a segmentation network combining global and neighborhood information

The comprehensive Loss function Loss (lambda, eta) is the sum of weighted Loss functions G _ Loss of Gnet and R _ Loss of Rnet, and the formula is expressed as:

Loss(λ,η)＝λU_Loss+ηI_Loss

wherein λ and η are weight parameters given to two different loss functions of the Gnet and the Rnet, and λ + η is 1; by distributing different weight values of two network loss functions to adjust the influence degree of the two networks on the segmentation result and optimize the segmentation effect, the invention determines that the network weight distribution is most reasonable and the segmentation effect is best when lambda is 0.4 and eta is 0.6 after experiments.

Further, in the step 4, training the segmentation network combined with global and neighborhood information, and performing comprehensive training to optimize the segmentation result by combining Gnet with vascular connected domain information predicted by Rnet, and reduce the fracture after vascular segmentation, the specific training steps are as follows:

4.1 Gnet and Rnet were trained independently and in parallel. Gnet is trained by G _ loss, Rnet is trained by R _ loss, and at the moment, the two networks have no key, and only the output of Gnet is used as the input of Rnet;

4.2 until R _ loss is stable, carrying out combined training on Gnet and Rnet, and freezing network parameters when the Rnet is stable during independent training; the training mode at this time is that the Gnet segmentation network is trained through a comprehensive Loss function Loss (lambda, eta), and the lambda and eta values are adjusted through experiments to obtain the optimal solution of the segmentation result.

According to the invention, firstly, a brand-new down-sampling mode is used on the basis of the Unet network, and a long-time memory network is combined at a jump connection position, so that the extraction of characteristic information is optimized, and the loss of image information is reduced; secondly, the invention provides a parallel network for calculating neighborhood information by taking the network provided in the previous step as a basic segmentation frame, and trains a segmentation network by taking a loss function of the network as a penalty item of the basic segmentation network.

The invention has the beneficial effects that: the extraction of the characteristic information is optimized, and the loss of the image information is reduced; and the condition that the segmented blood vessel is broken is reduced by combining neighborhood connected information, and the performance of blood vessel segmentation is effectively improved.

Drawings

FIG. 1 is a schematic flow diagram of an embodiment of the present invention.

FIG. 2 is a diagram of Gnet and Rnet network models in the scheme of the present invention.

Fig. 3 is a comprehensive training diagram in the scheme of the invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in detail below with reference to the accompanying drawings.

Referring to fig. 1 to 3, a blood vessel segmentation method combining global and neighborhood information, which can optimize the extraction of feature information and reduce the loss of image information, includes the following steps:

step 1 data preprocessing

The cerebrovascular system typically accounts for a relatively small proportion of the intracranial volume (about 1% to 5%), with FMM performance being largely dependent on fitting to background regions (i.e., non-vascular regions) and less dependent on fitting within vascular regions. Meanwhile, the large-area background reserved in the data can cause calculation burden and influence the processing time of the model; therefore, skull rejection and bias field correction are carried out to generate training samples;

step 2, constructing a vessel segmentation network Gnet

Building a novel network model Gnet combined with global information, training the network model by using the training sample generated in the step 1, and inputting the medical image to be segmented into the built network model so as to generate a target image in a corresponding mode;

further, in step 2, the vessel segmentation network Gnet uses a brand-new down-sampling mode on the basis of the Unet network and combines a long-term memory network with a short-term memory network at a jump junction, so that the extraction of feature information is optimized, and the loss of image information is reduced, and the specific process is as follows:

2.1, a ConvLSTM network module is added to optimize a jump connection process, the jump connection process optimized by the ConvLSTM network can further combine spatial information of images to strengthen the propagation and coding of original characteristics so as to eliminate ambiguity generated by irrelevant noise effect in jump connection and enable segmentation and positioning to be more accurate;

2.2, an attention gate mechanism is added to optimize a down-sampling process, aiming at the problem of loss of micro blood vessel characteristic information in the down-sampling process, an attention gate module is added, and simultaneously multi-scale original image input is combined in each layer of down-sampling, so that the proposed down-sampling optimization fully combines the characteristic information of the original image, the characteristic correspondence of irrelevant areas can be inhibited, the appearance and the size of a target can be automatically learned and distinguished, the obvious characteristics of the target blood vessel area are focused, and the loss of down-sampling information is reduced;

step 3, constructing a vascular network Rnet of neighborhood information

In order to solve the phenomena of breakage of small blood vessels and non-communication with large blood vessels divided by a network, the invention combines the connectivity of a blood vessel region, provides a convolutional neural network Rnet based on planar twenty-six neighborhood connectivity, the input of the network is the output of the Gnet constructed in the step 2, the network consists of a plurality of continuous convolutional layers and full connection layers, and the numerical value of 1 multiplied by 1 is output through the full connection layers after 4 times of convolution and maximum pooling continuous operation. Training the Rnet network model by using Mean Square Error (MSE) as a loss function, wherein the Rnet network model is defined as follows:

wherein Ii and Oi represent the island number Islands of the disconnected blood vessels of input and output, the calculation is carried out by a Skimage image processing tool kit, the mean square error MSE of the island number Islands is smaller, and Rnet describes that the accuracy rate of the island number Islands of the disconnected blood vessels in a segmentation graph output by Gnet is higher;

step 4, training a segmentation network combining global and neighborhood information

The comprehensive Loss function Loss (lambda, eta) is the sum of weighted Loss functions G _ Loss of Gnet and R _ Loss of Rnet, and the formula is expressed as:

Loss(λ,η)＝λU_Loss+ηI_Loss

the method comprises the steps of obtaining a Gnet loss function and a Rnet loss function, wherein lambda and eta are weight parameters given to the Gnet loss function and the Rnet loss function, lambda + eta is 1, the influence degree of the two networks on a segmentation result is adjusted by distributing different weight values of the two network loss functions, and the segmentation effect is optimized.

Further, in the step 4, training the segmentation network combined with global and neighborhood information, and performing comprehensive training to optimize the segmentation result by combining Gnet with the vascular communication domain information predicted by Rnet, so as to reduce the fracture after vascular segmentation; the specific training steps are as follows:

4.1 independently training Gnet and Rnet in parallel, wherein Gnet is trained by G _ loss, Rnet is trained by R _ loss, and the two networks have no key, but the output of Gnet is used as the input of Rnet;

4.2 until R _ Loss is stable, carrying out combined training on Gnet and Rnet, freezing network parameters when the Rnet network is stable during independent training, wherein the training mode is that the Gnet segmentation network is trained through a comprehensive Loss function Loss (lambda, eta), and regulating lambda and eta values through experiments to obtain an optimal solution of the segmentation result.

The embodiments described in this specification are merely illustrative of implementations of the inventive concepts, which are intended for purposes of illustration only. The scope of the present invention should not be construed as being limited to the particular forms set forth in the examples, but rather as being defined by the claims and the equivalents thereof which can occur to those skilled in the art upon consideration of the present inventive concept.