CN112862805B

CN112862805B - Automatic auditory neuroma image segmentation method and system

Info

Publication number: CN112862805B
Application number: CN202110238730.5A
Authority: CN
Inventors: 刘钦源; 柴露; 张国凯
Original assignee: Tongji University
Current assignee: Tongji University
Priority date: 2021-03-04
Filing date: 2021-03-04
Publication date: 2022-12-06
Anticipated expiration: 2041-03-04
Also published as: CN112862805A

Abstract

The invention provides an auditory neuroma image automatic segmentation method and system, which utilize a pre-trained auditory neuroma image automatic segmentation model to automatically segment an acquired nuclear magnetic resonance image of an auditory neuroma patient to obtain a segmentation result, wherein the auditory neuroma image automatic segmentation model adopts a cascade structure of a preprocessing network and a segmentation network. According to the embodiment of the invention, the segmentation performance of the model is improved by constructing the preprocessing network aiming at improving the segmentation performance of the model, the limitation of a single Unet segmentation model is broken through, the current optimal automatic segmentation model of the auditory neuroma image is obtained, the automatic segmentation of the auditory neuroma image can be realized at a higher speed, and the segmentation efficiency of the auditory neuroma image is greatly improved.

Description

Automatic auditory neuroma image segmentation method and system

Technical Field

The invention belongs to the technical field of automatic image segmentation, and particularly relates to an auditory neuroma image segmentation method based on a deep neural network.

Background

Auditory neuroma is an uncommon intracranial tumor, most of middle-aged people over 30 years old have small tumor volume and are often accompanied by tinnitus, nausea and other symptoms which are easy to overlook, when the tumor volume is small, the tumor is enlarged and presses facial nerves and cerebellum, so that severe facial paralysis and limb discordance occur, and doctors take treatment opinions about the benign tumor as soon as possible.

Magnetic Resonance Imaging (MRI) is frequently used in brain diagnosis and treatment, and MR Imaging has the following advantages: mri can exhibit a wide range of image contrast; MRI can realize relatively complete soft tissue imaging; mri resolution is higher. Doctors can observe detailed details of body parts such as brains and abdomens of patients by means of MRI, which is very beneficial to diagnosis and treatment of diseases. MRI images typically use four sequences, T1ce, T2, FLAIR, different sequences may show different tissue characteristics. In the treatment of acoustic neuroma, physicians often use T2 sequence-weighted MRI images as a diagnostic basis, where the tumor area is larger and visually brighter than the surrounding tissue.

Through medical images, doctors can manually label tumor regions according to knowledge and experience of the doctors, but under the conditions of uneven medical resource distribution and continuous growth of medical image data, the problems of long calibration period and uneven calibration precision are gradually revealed by manual labeling, so that automatic labeling becomes the key point of attention of people. Because the medical image has a single style and a fuzzy boundary, the traditional semantic segmentation method has poor effect on medical image segmentation and needs manual assistance, and full-automatic labeling cannot be realized. In order to reduce the burden of doctors and improve the segmentation accuracy, computer-aided diagnosis and treatment technology based on deep learning is gradually widely applied in recent years, but the performance of a single segmentation network is improved to a limited extent, so that a new automatic segmentation method for auditory neuroma images is required to be provided to improve the segmentation performance.

Disclosure of Invention

The invention provides an automatic auditory neuroma image segmentation method and system aiming at the technical problem that the performance improvement of the existing single auditory neuroma image segmentation network is limited.

In order to achieve the technical purpose, the invention adopts the following technical scheme.

One aspect provides an automatic auditory neuroma image segmentation method, which comprises the following steps:

acquiring a nuclear magnetic resonance image of an acoustic neuroma patient;

and automatically segmenting the acquired nuclear magnetic resonance image of the acoustic neuroma patient by utilizing a pre-trained acoustic neuroma image automatic segmentation model to obtain a segmentation result, wherein the acoustic neuroma image automatic segmentation model adopts a cascade structure of a preprocessing network and a segmentation network.

Further, the preprocessing network adopts a pix2pixGAN network, and the split network adopts a uet split network, wherein:

the pix2pixGAN network and the Unet segmentation network are both symmetrical network structures and respectively comprise a downsampling part and an upsampling part, the structure of the pix2pixGAN generator comprises 26 layers, and the downsampling part comprises 2 dropout layers;

the Unet splits the network into 23 layers, with the down-sampled portion containing 10 convolutional layers, 2 max pooling operations, and the up-sampled portion containing 8 convolutional layers and 4 anti-convolutional layers.

Still further, the training method of the Unet split network is as follows: acquiring a training set of a nuclear magnetic resonance image sample of a brain of a patient suffering from acoustic neuroma; and training the Unet segmentation network by taking the training set as input data and the labeled segmentation result corresponding to the image sample as target data, and selecting the Unet segmentation network with the minimum loss value as a final Unet segmentation network after multiple iterations.

Still further, the training method of the pix2pixGAN network is as follows: acquiring a training set and a verification set of a nuclear magnetic resonance image sample of a brain of an acoustic neuroma patient; acquiring a corrected image corresponding to the acoustic neuroma nuclear magnetic resonance image in the training set based on the training set and the Unet segmentation network; respectively taking the auditory neuroma nuclear magnetic resonance image in the training set and the corresponding correction image as input data and target data of a pix2pixGAN network, inputting the input data and the target data into the pix2pixGAN network, and performing repeated iterative training to obtain a generator network; and testing on the acoustic neuroma image automatic segmentation model by using the test set, and selecting a generator network with optimal test performance as a preprocessing network of the acoustic neuroma image automatic segmentation model.

Still further, the method for obtaining the corrected image corresponding to the acoustic neuroma nuclear magnetic resonance image comprises the following steps:

calculating a Dice evaluation coefficient of a nuclear magnetic resonance image of a brain of a patient with acoustic neuroma in the training set in an Unet segmentation network;

calculating partial differential of the Dice evaluation coefficient relative to the nuclear magnetic resonance image, and superposing the partial differential value as a correction value of the current nuclear magnetic resonance image; and (5) obtaining a final corrected image of the acoustic neuroma brain nuclear magnetic resonance image through multiple iterations.

In another aspect, the present invention provides an automatic segmentation system for auditory neuroma images, comprising: the system comprises a data acquisition module and an auditory neuroma image automatic segmentation module, wherein the data acquisition module is used for acquiring a nuclear magnetic resonance image of an auditory neuroma patient; the acoustic neuroma image automatic segmentation module is used for utilizing a pre-trained acoustic neuroma image automatic segmentation model to automatically segment the acquired nuclear magnetic resonance image of the acoustic neuroma patient to obtain a segmentation result, and the acoustic neuroma image automatic segmentation model adopts a cascade structure of a preprocessing network and a segmentation network

The present invention also provides a computer-readable storage medium comprising: at least one processor and a memory communicatively coupled to the at least one processor, wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to perform an automatic segmentation method of an auditory neuroma image as provided in any one of the above aspects in real time.

The invention has the following beneficial technical effects:

according to the embodiment of the invention, the segmentation performance of the model is improved by constructing the preprocessing network which aims at improving the segmentation performance of the model, the limitation of a single Unet segmentation model is broken through, the current optimal auditory neuroma image automatic segmentation model is obtained, and the performance of the neuroma image segmentation method is improved.

Meanwhile, the segmentation mask and the nuclear magnetic resonance image are fused by using the post-processing module, so that the segmentation result is more visual, a doctor can perform more accurate diagnosis on the tumor position and the tumor characteristics by means of the result, and the problems of long calibration period and uneven calibration precision of manual labeling are effectively solved. The embodiment of the invention can realize the automatic segmentation of the auditory neuroma image at a higher speed on the computer terminal, greatly improves the segmentation efficiency of the auditory neuroma image and effectively simplifies the diagnosis and treatment process.

Drawings

FIG. 1 is a flow chart of an automatic auditory neuroma image segmentation method based on cascade of a preprocessing network and a Unet segmentation network of gradient descent principle according to the embodiment of the invention;

FIG. 2 is a flow chart of a specific segmentation scheme according to an embodiment of the present invention;

FIG. 3 is a two-dimensional nuclear magnetic resonance image of a brain of a patient with acoustic neuroma;

FIG. 4 is a corresponding actual labeled diagram of FIG. 3;

fig. 5 is a diagram of a network partitioning structure of the Unet according to the embodiment of the present invention;

FIG. 6 is a flowchart of the training of the Unet split network according to the embodiment of the present invention;

FIG. 7 is a diagram of a preprocessing network architecture according to an embodiment of the present invention;

FIG. 8 is a flow chart of an exemplary pre-processing network training of the present invention;

FIG. 9 is a diagram of an automated segmentation model architecture in accordance with an embodiment of the present invention;

FIG. 10 is a diagram illustrating a segmentation effect obtained by an embodiment of the present invention.

Detailed Description

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

Referring to fig. 1, an embodiment of the present invention provides an automatic segmentation method for an acoustic neuroma image, including the following steps: acquiring a brain nuclear magnetic resonance image of an acoustic neuroma patient;

and automatically segmenting the acquired nuclear magnetic resonance image by using an auditory neuroma image automatic segmentation model based on a deep learning method, wherein the auditory neuroma image automatic segmentation model adopts a cascade structure of a preprocessing network and a Unet segmentation network based on a gradient descent principle.

In particular, automated labeling of tumors can be considered as an image segmentation problem, specifically, image segmentation is a process of performing class judgment on pixels, and each pixel has only two labels of tumor and non-tumor in tumor segmentation, so that tumor segmentation can be regarded as two classes of pixels. Because the medical image has a single style and a fuzzy boundary, the traditional semantic segmentation method has poor effect on medical image segmentation and needs manual assistance, and full-automatic labeling cannot be realized. In order to reduce the burden of doctors and improve the segmentation accuracy, computer-aided diagnosis and treatment technologies based on deep learning have been widely used in recent years. The current medical image segmentation algorithm can be largely divided into a typical image segmentation algorithm and an algorithm based on a neural network, and because the medical image has a single structure and fuzzy edges, the traditional semantic segmentation method has poor effect on medical image segmentation and needs artificial assistance, and full-automatic labeling cannot be realized. In recent years, data-driven deep learning methods have been widely used in many fields. In terms of image processing, convolutional Neural Networks (CNNs) can acquire high-level features of images, and have been applied to semantic segmentation tasks aiming at understanding natural medical images. At present, the enhancement of the semantic segmentation effect based on deep learning is mainly attributed to the improvement of a network structure, which generally includes the change of network depth, width and connection mode or the proposal of a new network layer. Compared with other coding and decoding networks, the Unet symmetric network has the advantages that the undersampling operation for extracting features and the upsampling operation for mapping low-level features to complete input images are included, and the Skip Connection operation is added into the Unet, so that the combination of low-resolution information and high-resolution information is realized, the gradient disappearance condition of a coding and decoding structure is effectively eliminated while the segmentation performance of a model is improved, and a better effect is achieved in the biomedical image segmentation.

However, the improvement of the model segmentation effect by modifying the Unet network is limited, and the embodiment of the invention obtains the currently optimal automatic segmentation model of the auditory neuroma image by cascading a preprocessing network aiming at improving the model segmentation performance, as shown in fig. 2. The doctor can rely on the result to carry out comparatively accurate diagnosis to tumour position and characteristic, has promoted the efficiency of diagnosing of auditory neuroma disease by a wide margin, has simplified the flow of diagnosing.

Further, in a preferred embodiment, the step of obtaining magnetic resonance imaging of the brain of a patient with acoustic neuroma comprises:

acquiring brain nuclear magnetic resonance imaging and labeling data of an acoustic neuroma patient under a T2 sequence;

performing format conversion on the acquired nuclear magnetic resonance image and the acquired labeling data thereof by using a library function to obtain a two-dimensional readable auditory neuroma data set;

and (3) performing data amplification by using a data enhancement method, wherein the data enhancement method comprises rotation, translation and turnover.

Specifically, in the present embodiment, the nuclear magnetic resonance images of the brain of the acoustic neuroma patient in the four sequences of T1, T1ce, T2 and FLAIR can be obtained, and since the doctor usually uses the MRI image weighted by the T2 sequence as the diagnosis basis in the treatment of the acoustic neuroma, the tumor region in the sequence has a better visual effect than the surrounding tissues, the nuclear magnetic resonance image in the T2 sequence is used as the data set.

The specific method for data enhancement in this embodiment is to use the library function provided by keras to rotate, translate and flip the image, and to use it in a generator mode for the training function.

In the automatic segmentation model for auditory neuroma images in the embodiment, the preprocessing network based on the gradient descent principle adopts a pix2pixGAN generator structure, the segmentation network adopts an Unet architecture, and the specific architecture of the network is as follows:

the generator of the pix2pixGAN structure is similar to the structure of Unet, is a symmetrical network structure, has 26 layers, and consists of a down-sampling part and an up-sampling part, wherein the down-sampling process comprises 2 dropout layers.

Specifically, a pix2pixGAN training preprocessing network with an Unet structure generator is used, and in an ideal case, the GAN structure can enable the generator to obtain a good result after multiple iterations, but due to the particularity of the GAN training process, the GAN training process is unstable and even the gradient disappears due to the segmentation network based on the Unet structure. To better train the GAN, the following modification strategy is used: try to use LeakyReLU as the activation function instead of ReLU; replace the maximum pooling (max pooling) operation with strides greater than 1; the activation function of the last layer uses tanh with an output value range of [ -1,1] instead of sigmoid; noise is introduced into the model by increasing dropout and the like, and the robustness of the model is enhanced. The authors of pix2pixGAN modified the Unet structure, following the builder structure, to obtain a builder structure for auditory neuroma image segmentation as shown in fig. 7 of the specification. In fig. 7, each rectangle represents a multi-channel signature, the number of channels being indicated on the upper side of the box and the signature size being indicated on the left side of the box. Arrows of different styles represent different calculation operations, an arrow to the right represents a convolution operation of size 4 × 4, strides is set to 2, and a LeakyReLU function is used; the arrow at the lower left indicates that deconvolution operation with an activation function of ReLU is performed, the size of a convolution kernel is 4 x 4, and the two operations can respectively modify the length and the width of data to be one half and two times of the original length and the width; the downward arrow represents a copy and join (join) operation that joins the shallow data during downsampling and the upsampled deep data in the last dimension; the left arrow indicates an up-sampling operation and a convolution operation with a size of 1, an image with the same size as the input image and each pixel value in the range of-1 to 1 is obtained through the activation function tanh, and the image is normalized to obtain the predicted segmentation result.

The Unet partitions the network into 23 layers, the downsampling process contains 10 convolutional layers, 2 max pooling operations, and the upsampling process contains 8 convolutional layers and 4 deconvolution layers.

Specifically, the dropout operation is increased to enhance the robustness of the model, resulting in a specific structure as shown in fig. 5 of the specification. Each rectangle represents a multi-channel feature map, the number of channels being indicated on the right side of the box, the feature map size being indicated on the top side of the box, and the size being indicated for the box above it if not indicated. Different patterns of arrows represent different computational operations, where the solid downward arrows represent convolution operations with a convolution kernel size of 3 × 3, and the activation function is ReLU (Rectified Linear Unit); the arrow to the left indicates the maximum pooling with the size of 2 × 2, the arrow to the right indicates the deconvolution operation with the activation function of ReLU is performed, the convolution kernel size is 2 × 2, and the two operations can respectively modify the length and width of the data to be one half and two times of the original length and width; the downward open dotted arrow represents a copy and join (join) operation that joins the shallow data during downsampling and the upsampled deep data in the last dimension; the downward hollow solid arrow indicates a convolution operation with the size of 1, the activation function Sigmoid maps the value to the [0,1] interval, and a gray level map with the same resolution as the input image and the pixel value in the interval of 0 to 1 is obtained, and the map is the segmentation result of the Unet prediction.

Further, the step of performing automatic segmentation of the auditory neuroma image by using the cascade of the preprocessing network based on the gradient descent principle and the Unet segmentation network specifically comprises:

and training by using the divided given samples to obtain an auditory neuroma automatic segmentation model.

Inputting the readable nuclear magnetic resonance image of the auditory neuroma patient into the automatic segmentation model to obtain the focus image segmentation result of the auditory neuroma.

Further, the step of training the obtained automatic segmentation model of the acoustic neuroma image includes:

a given sample is divided into a training set and a test set.

Specifically, the original data set includes brain magnetic resonance image data of 200 patients and tumor labeling results thereof, and the data of each patient is composed of a plurality of two-dimensional slices. Among various sequences of magnetic resonance image data, the contrast between a tumor region and surrounding tissues under a T2 sequence is higher, so 900 two-dimensional MRI images and labeled data thereof under the T2 sequence are selected as a data set. To facilitate reading and processing of the images, the MRI images in the dicom format were processed using the pydicom and simpletick software libraries and converted to jpeg format, resulting in brain MRI images of size 640 × 640, as shown in fig. 3 of the specification. Meanwhile, corresponding marking data are represented by a mask image, and as shown in the figure 4 of the specification, the size of the mask image is consistent with that of the brain MRI image.

And inputting the training set into a Unet network for training to obtain an Unet self-segmentation module aiming at the auditory neuroma image.

And inputting the training set and the corresponding correction graph into a pix2pixGAN network for training to obtain a preprocessing module for correcting the image of the acoustic neuroma.

Further, the specific steps of training the Unet network include:

and taking the nuclear magnetic resonance image of the brain of the acoustic neuroma patient in the training set as input data, and taking a corresponding labeling segmentation result of the image as target data.

And training by using the samples, and selecting a model with the minimum loss value as an Unet segmentation module after multiple iterations.

Specifically, as shown in fig. 6 of the specification, an acoustic neuroma segmentation model training process based on the unnet includes firstly dividing an original data set into training data and test data, performing data enhancement on the training set, performing segmentation network training by using newly generated data, comparing a loss value of the training set of the iteration with a loss value of a current optimal model after each iteration is completed, updating the optimal loss value if the loss value is smaller than the optimal loss value, saving a result of the iteration as the optimal model, and directly performing the next iteration if the loss value is larger than the optimal loss value.

Further, the specific step of inputting the training set and the corresponding correction map into a pix2pixGAN network for training to obtain the preprocessing module includes:

acquiring a correction image corresponding to the acoustic neuroma nuclear magnetic resonance image in the training set;

inputting the acoustic neuroma nuclear magnetic resonance images in the training set and the corresponding correction images into a pix2pixGAN network for training to obtain a trained generator and a trained discriminator;

and extracting a generator structure as a preprocessing module of the acoustic neuroma image segmentation model.

Further, the step of acquiring a corrected image of the acoustic neuroma nuclear magnetic resonance image of the training set specifically includes:

calculating a Dice evaluation coefficient of a nuclear magnetic resonance image of a brain of a patient with the central acoustic neuroma in the Unet segmentation model;

calculating the partial differential of the Dice coefficient relative to the nuclear magnetic resonance image, and overlapping the partial differential value as a correction value of the current nuclear magnetic resonance image;

and (5) obtaining the final correction result of the acoustic neuroma brain nuclear magnetic resonance image through multiple iterations.

Specifically, each pixel of the input image I is labeled as two categories, namely 0 and 1, wherein 0 represents that the pixel belongs to a non-tumor region, and 1 represents that the pixel belongs to a tumor region, and since the activation function of the last layer of the Unet segmentation network is a sigmoid function with an output range of [0,1], the prediction output result of the model is a gray image with each pixel value in the range of [0,1], and the prediction result that the tumor region is white and the non-tumor region is black can be obtained after normalization. Let T represent the real annotation data corresponding to the image I to be segmented, the objective of the segmentation model is to make P = T, the objective can be measured by a dice coefficient, and the bigger the dice coefficient is, the closer P and T are, therefore, the objective of the preprocessing algorithm of the embodiment of the invention is to find a delta I, so that the bigger the dice coefficient is, the better the dice coefficient is. And (4) taking iteration as iteration times, performing iteration times on the image to be segmented in the training data to obtain an updated image I ', and obtaining a new image I' with a segmentation effect superior to that of the original image after the iteration times reach a certain value. Although the training data can correct the input data by using the annotation image T, the annotation image is unknown to the test data, and the image cannot be updated by using the corresponding annotation data T, so that a function is required to realize the mapping from the original image I to the new image I'. Because the style conversion performance of the pix2pixGAN structure between pictures is excellent, the pix2pixGAN structure is used for training the mapping between I's and I ', the training data of the network comprises domain nA and domain nB, wherein the domain nA is an image to be segmented in a segmented network training set, the domain nB is an image set obtained by updating data in the domain nA through the preprocessing algorithm, the domain nA corresponds to the image in the domain nB one by one, and the trained generator network can realize the mapping from an original image I to a new image I '.

Further, the specific steps of training the pix2pixGAN network include:

respectively taking the auditory neuroma nuclear magnetic resonance image in the training set and the correction result thereof as input data and target data of a pix2pixGAN network, and carrying out repeated iterative training to obtain a generator and a discriminator;

testing on an auditory neuroma image automatic segmentation model based on cascade of a generator network of a gradient descent principle and an Unet segmentation network by using a test data set, and selecting the generator network with the optimal test performance as a preprocessing module of the auditory neuroma image automatic segmentation model;

as shown in fig. 9, the trained preprocessing module is cascaded with the Unet segmentation network to obtain an automatic segmentation model of the auditory neuroma image.

Specifically, as shown in fig. 8, a preprocessing network training process based on pix2pixGAN is performed, where training data of a preprocessing network is first constructed, an image to be segmented in a segmentation network training set is used as domainA, and a set obtained by updating data in the domainA is used as domainB. And then, training for generating an antagonistic network by using the domainA and the domainB, comparing the MAE value of the iteration on the training set with the current optimal model MAE value after each iteration is finished, if the MAE value is smaller than the optimal MAE value, updating the optimal MAE value and storing the result of the iteration as the optimal model, and if the MAE value is larger than the optimal MAE value, directly performing the next iteration. And obtaining updated test data after multiple iterations, and segmenting a new image by using the Unet network to obtain a final segmentation result.

The nuclear magnetic resonance images of 270 auditory neuroma patients under the T2 sequence are segmented by adopting the method of the specific embodiment, and the Dice coefficient, the Jaccard coefficient, the sensitivity and the specificity of the method are respectively 0.8750, 0.7988, 0.9006 and 0.9992 through evaluation, so that the segmentation performance of the visible model is high, and the reliability of the model is high.

According to the embodiment of the invention, the segmentation performance of the model is improved by constructing the preprocessing network which aims at improving the segmentation performance of the model, the limitation of a single Unet segmentation model is broken through, and the current optimal automatic segmentation model of the auditory neuroma image is obtained. Meanwhile, the segmentation mask and the nuclear magnetic resonance image are fused by using the post-processing module, so that the segmentation result is more visual, a doctor can perform more accurate diagnosis on the tumor position and the tumor characteristics by means of the result, and the problems of long calibration period and uneven calibration precision of manual labeling are effectively solved. The embodiment of the invention can realize the automatic segmentation of the auditory neuroma image at a higher speed on the computer terminal, greatly improves the diagnosis and treatment efficiency of the auditory neuroma disease and effectively simplifies the diagnosis and treatment process.

Corresponding to the above provided automatic segmentation method for auditory neuroma images, the invention also provides an automatic segmentation system for auditory neuroma images, comprising: the system comprises a data acquisition module and an auditory neuroma image automatic segmentation module, wherein the data acquisition module is used for acquiring a nuclear magnetic resonance image of an auditory neuroma patient; the automatic auditory neuroma image segmentation module is used for automatically segmenting the acquired nuclear magnetic resonance image of the auditory neuroma patient by utilizing a pre-trained automatic auditory neuroma image segmentation model to obtain a segmentation result, and the automatic auditory neuroma image segmentation model adopts a cascade structure of a preprocessing network and a segmentation network.

Further, the preprocessing network is used for correcting the brain nuclear magnetic resonance image, and the segmentation network is used for dividing the focus region of the corrected brain nuclear magnetic resonance image to obtain a mask image of the tumor region.

Still further, the automatic segmentation system further comprises a data post-processing module, wherein the data post-processing module is used for fusing the mask image and the corresponding brain nuclear magnetic resonance image, so that the prediction segmentation result of the model is more visual.

It should be noted that, as those skilled in the art can clearly understand that, for convenience and brevity of description, the specific working processes of the system, the device and the unit described in the present invention may refer to the corresponding processes in the foregoing method, and therefore, the details are not repeated in the original application document.

As will be appreciated by one skilled in the art, embodiments of the present application may be provided as a method, system, or computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application 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, and the like) having computer-usable program code embodied therein.

The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the application. It will be understood that each flow and/or block of the flow diagrams and/or block diagrams, and combinations of flows and/or blocks in the flow diagrams and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks. These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

While the preferred embodiments of the present application have been described, additional variations and modifications in those embodiments may occur to those skilled in the art once they learn of the basic inventive concepts. Therefore, it is intended that the appended claims be interpreted as including the preferred embodiment and all changes and modifications that fall within the scope of the present application.

It will be apparent to those skilled in the art that various changes and modifications may be made in the present application without departing from the spirit and scope of the application. Thus, if such modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is intended to include such modifications and variations as well.

Claims

1. An automatic auditory neuroma image segmentation method is characterized by comprising the following steps:

acquiring a nuclear magnetic resonance image of an acoustic neuroma patient;

the method comprises the steps that a pre-trained auditory neuroma image automatic segmentation model is used for conducting automatic segmentation on an acquired nuclear magnetic resonance image of an auditory neuroma patient to obtain a segmentation result, the auditory neuroma image automatic segmentation model adopts a cascade structure of a preprocessing network and a segmentation network, the preprocessing network adopts a pix2pixGAN network, and the segmentation network adopts a Unet segmentation network;

the training method of the pix2pixGAN network comprises the following steps: acquiring a training set and a verification set of a nuclear magnetic resonance image sample of a brain of a patient suffering from acoustic neuroma; acquiring a corrected image corresponding to the acoustic neuroma nuclear magnetic resonance image in the training set based on the training set and the Unet segmentation network; respectively taking the acoustic neuroma nuclear magnetic resonance image in the training set and the corresponding corrected image as input data and target data of a pix2pixGAN network, inputting the input data and the target data into the pix2pixGAN network, and performing iterative training for multiple times to obtain a generator network; testing on the acoustic neuroma image automatic segmentation model by using a test set, and selecting a generator network with optimal test performance as a preprocessing network of the acoustic neuroma image automatic segmentation model;

the method for acquiring the corrected image corresponding to the acoustic neuroma nuclear magnetic resonance image comprises the following steps:

calculating partial differential of the Dice evaluation coefficient relative to the nuclear magnetic resonance image, and superposing the partial differential value as a correction value of the current nuclear magnetic resonance image; and (5) obtaining a final correction image of the acoustic neuroma brain nuclear magnetic resonance image through multiple iterations.

2. The method for automatically segmenting an auditory neuroma image according to claim 1,

3. The method for automatically segmenting auditory neuroma images according to claim 2, wherein the training method of the Unet segmentation network comprises: acquiring a training set of a nuclear magnetic resonance image sample of a brain of an acoustic neuroma patient; and training the Unet segmentation network by taking the training set as input data and the labeled segmentation result corresponding to the image sample as target data, and selecting the Unet segmentation network with the minimum loss value as a final Unet segmentation network after multiple iterations.

4. The method for automatically segmenting an auditory neuroma image according to claim 1, wherein acquiring magnetic resonance imaging of the brain of an auditory neuroma patient comprises: acquiring nuclear magnetic resonance imaging and labeling data of a brain of an acoustic neuroma patient under a T2 sequence; performing format conversion on the acquired nuclear magnetic resonance image and the labeling data thereof by using a library function to obtain a two-dimensional readable auditory neuroma data set; data amplification was performed using a data enhancement method.

5. An automatic auditory neuroma image segmentation system, comprising: the system comprises a data acquisition module and an auditory neuroma image automatic segmentation module, wherein the data acquisition module is used for acquiring a nuclear magnetic resonance image of an auditory neuroma patient; the automatic auditory neuroma image segmentation module is used for automatically segmenting an acquired nuclear magnetic resonance image of an auditory neuroma patient by utilizing a pre-trained automatic auditory neuroma image segmentation model to obtain a segmentation result, the automatic auditory neuroma image segmentation model adopts a cascade structure of a preprocessing network and a segmentation network, the preprocessing network adopts a pix2pixGAN network, and the segmentation network adopts a Unet segmentation network;

the training method of the pix2pixGAN network comprises the following steps: acquiring a training set and a verification set of a nuclear magnetic resonance image sample of a brain of a patient suffering from acoustic neuroma; acquiring a corrected image corresponding to the acoustic neuroma nuclear magnetic resonance image in the training set based on the training set and the Unet segmentation network; respectively taking the auditory neuroma nuclear magnetic resonance image in the training set and the corresponding correction image as input data and target data of a pix2pixGAN network, inputting the input data and the target data into the pix2pixGAN network, and performing repeated iterative training to obtain a generator network; testing on the automatic acoustic neuroma image segmentation model by using a test set, and selecting a generator network with optimal test performance as a pretreatment network of the automatic acoustic neuroma image segmentation model; the method for acquiring the corrected image corresponding to the acoustic neuroma nuclear magnetic resonance image comprises the following steps:

6. The automatic auditory neuroma image segmentation system of claim 5 wherein the preprocessing network is configured to modify the brain nuclear magnetic resonance image and the segmentation network is configured to segment the lesion region of the modified brain nuclear magnetic resonance image to obtain a mask image of the tumor region.

7. The automatic auditory neuroma image segmentation system according to claim 5 further comprising a data post-processing module for fusing the mask image with the corresponding brain NMR image to make the predicted segmentation result of the automatic auditory neuroma image segmentation module more intuitive.

8. A computer-readable storage medium, comprising: at least one processor and a memory communicatively coupled to the at least one processor, wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method for automated segmentation of an auditory neuroma image as claimed in any one of claims 1 to 4.