A medical image organ accurate automatic segmentation method and system

Abstract

The application provides a medical image organ accurate automatic segmentation method and system, the method comprises the following steps: step 1: extracting the outline mask information of the organ; step 2: using the organ outline mask information, based on the deep neural network to carry out target region segmentation; step 3: the target region segmented by the deep neural network is optimized.The medical image organ accurate automatic segmentation method and system of the application use the organ outline mask information as the input of the training deep neural network, which is beneficial to solve the problem of underfitting or overfitting of the network training model caused by too small foreground (target) region and too large background region in the medical image, while ensuring the segmentation accuracy, greatly optimizing the network training time and improving the segmentation speed.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of medical imaging technology, and particularly relates to a medical imaging organ precise automatic segmentation method and system. BACKGROUND

[0002] In recent years, due to irregular diet and work and rest of people, the incidence of pancreatic diseases has increased significantly year by year, especially pancreatic cancer, which is one of the highest mortality tumors, and is known as "King of Cancer". The five-year survival rate is less than 1%. Early pancreatic cancer can be treated by clinical surgery, but due to the insidious onset of the disease, there are no obvious symptoms in the early stage, which seriously affects the diagnosis of early pancreatic cancer by doctors. Usually when the patient has obvious symptoms such as weight loss and pain, it is already in the middle and late stages of pancreatic cancer, missing the best period for surgery. Even if a targeted clinical treatment plan is adopted, the overall prognosis is not ideal. Therefore, early diagnosis and treatment of pancreatic cancer have important research value and also have important significance in clinical practice.

[0003] With the great progress of medical equipment and the rapid development of imaging genomics technology, automatic detection, segmentation, classification and three-dimensional reconstruction characterization of pancreatic anatomical structure on medical images can significantly improve the early diagnosis rate of pancreatic diseases by clinicians, so as to implement more efficient and accurate treatment methods. Due to the complex anatomical structure of the pancreas, the shape and position vary greatly, there are a large number of adjacent tissues closely connected with the pancreas and similar density around the pancreas, and the gray scale contrast is not obvious. The traditional segmentation method cannot achieve accurate pancreatic segmentation only by gray scale information. At the same time, due to the imaging principle of medical images and the difference in the characteristics of the tissue itself, the formation of the image is easily affected by factors such as noise and tissue movement. These factors greatly increase the difficulty of pancreatic segmentation, and even experienced imaging physicians face great challenges in manually segmenting the pancreas. In addition, the shape and position of the pancreas vary from person to person, and even the shape and position of the same person at different times and different medical image slices are not the same. In addition, the large amount of medical image data brings heavy mental and physical burden to doctors, which may lead to missed diagnosis or misdiagnosis.

[0004] Pancreatic segmentation has experienced the development process from manual segmentation to semi-automatic segmentation and then to full-automatic segmentation. Among them, manual segmentation refers to manually labeling the region belonging to the pancreas in the pancreatic medical image according to the clinical experience and anatomical knowledge of doctors. This method is low in efficiency and accuracy, and is highly subjective, and requires sufficient professional skills.

[0005] Semi-automatic segmentation mainly processes pancreatic image based on the professional knowledge of doctors, using threshold, simple linear iterative clustering algorithm to calculate the pancreas probability map, and then using different statistical model algorithm to process the probability map to realize the final pancreatic segmentation. This method is mainly based on probability map and statistical model algorithm, which greatly improves the segmentation speed and reduces the manual intervention, but the segmentation accuracy is often not high and the robustness is poor.

[0006] Fully automatic segmentation refers to using only computer and algorithm model to segment the pancreas image without human intervention. Fully automatic method mainly includes two categories: atlas-based registration algorithm and convolutional neural network (CNN) deep method. Atlas-based registration algorithm is to first use the labeled pancreas region to construct the pancreas atlas, and then use the registration algorithm to compare the constructed pancreas atlas to segment the pancreas structure of medical image. Deep learning method mainly uses CNN model to classify medical image to achieve automatic segmentation. For general classification CNN network such as VGG (Visual Geometry Group Network) and ResNet (Residual Network), the pancreas image is first cut to form a number of pixel blocks of the same size, and then each pixel block is classified to complete the segmentation. The current mainstream pancreatic segmentation method is to use fully connected neural network (FCN) and U-Net model for direct end-to-end targeted segmentation: the three-dimensional medical image of pancreas (such as CT or nuclear magnetic resonance image) is cut into two-dimensional slices along the axial plane, coronal plane and sagittal plane, and then input into three FCN networks for training, and through multiple iterations, the final pancreatic segmentation is realized. Although this method greatly improves the segmentation accuracy, the network needs 10 hours of training time.

[0007] For pancreatic segmentation algorithm, most researchers currently improve the segmentation accuracy of pancreas by iterative network or multi-stage network, but at the cost of greatly increasing the time consumed by segmentation. For clinical medicine, if the segmentation result cannot be displayed in time, it becomes unrealistic to assist doctors in diagnosis and treatment.

[0008] At the same time, due to the complexity and diversity of human environment, the image segmentation process of other organs, especially internal organs, also has similar problems as pancreatic image segmentation.

[0009] Therefore, it is particularly important to design an effective, fast and accurate medical image organ segmentation scheme. SUMMARY

[0010] To solve the above problems, the present application provides a medical image organ precise automatic segmentation method, comprising:

[0011] Step 1: extracting the contour mask information of the organ;

[0012] Step 2: using the organ contour mask information, performing target region segmentation based on a deep neural network;

[0013] Step 3: performing optimization processing on the target region segmented by the deep neural network.

[0014] Further, the method comprises:

[0015] Step 11: obtaining labeled real medical image data of the organ region , using the labeled real image data , adding random noise to produce noise medical image data , wherein i represents the i-th medical image slice, represents the labeled real medical image data of the i-th medical image slice, represents the noise medical image data of the i-th medical image slice;

[0016] Step 12: using and , training a fixed point model to obtain a context target region prediction function and a pre-segmentation result : , is a parameter learned by training, which is a vector matrix used to describe the relationship between the labeled real medical image data and the noise medical image data ;

[0017] Step 13: using the context target region prediction function , respectively generating target region prediction functions in three directions of axial plane, coronal plane and sagittal plane: target region prediction function in the axial plane direction, target region prediction function in the coronal plane direction, and target region prediction function in the sagittal plane direction;

[0018] Step 14: for an input three-dimensional medical image data to be segmented , k represents the k-th medical image slice, is the image data of the k-th medical image slice, using the target region prediction functions in three directions , and This generates organ contour mask information in three directions. , and , This refers to the organ contour mask information in the axial plane direction. This provides organ contour mask information in the coronal plane. This provides organ contour mask information in the sagittal plane.

[0019] Furthermore, the deep neural network includes an encoding block and a decoding block, wherein the encoding block and the decoding block respectively adopt a dense residual U-shaped structure block.

[0020] Furthermore, the dense residual U-shaped structure block includes an input convolutional layer, a U-shaped symmetric encoder-decoder structure block, and a residual connection layer, wherein the number of layers in the U-shaped symmetric encoder-decoder structure block is L.

[0021] Furthermore, the input convolutional layer is used to convert the number of input channels to... Feature map Transform into Local feature map of the middle of each channel To extract local feature information.

[0022] Furthermore, the U-shaped symmetric encoding and decoding structure block includes L-2 upsampling modules and L-2 upsampling modules;

[0023] L-2 upsampling modules are connected in sequence to perform L-2 convolutional upsampling operations on the input data, specifically including: upsampling, convolution, batch normalization and ReLU activation;

[0024] L-2 downsampling modules are connected in sequence to perform L-2 convolutional downsampling operations on the input data, specifically including downsampling, convolution, batch normalization, and ReLU activation.

[0025] Furthermore, the method includes:

[0026] Set the number of layers L to extract feature maps with multi-scale spatial resolution.

[0027] Furthermore, the residual connection layer is used to fuse local feature maps. and contextual multi-scale feature maps .

[0028] Furthermore, deep neural networks employ a U-shaped structure, including encoding and decoding processes;

[0029] Encoding process: Execute T encoding blocks, of which T-2 dense residual U-shaped structure blocks with decreasing layer numbers are first encoded, and then 2 encoding blocks are executed for dilated convolution, batch normalization and activation;

[0030] The decoding process corresponds to the encoding process, and first performs a decoding block to perform a hole convolution, batch normalization and activation; and then performs T-2 dense residual U-shaped structure blocks with sequentially decreasing number of layers for decoding, and the number of layers of the dense residual U-shaped structure blocks of the decoding block is the same as that in the corresponding depth encoding block.

[0031] Further, the labeled real medical image data is used to train the final deep neural network.

[0032] Further, the target region segmentation based on the deep neural network comprises:

[0033] For an input test three-dimensional medical image data , the extracted organ mask information , and are input to the trained deep neural network for segmentation to obtain target region segmentation (output) results in three directions: axial plane direction target region , coronal plane direction target region and sagittal plane direction target region .

[0034] Among them, is the organ contour mask information in the axial plane direction, is the organ contour mask information in the coronal plane direction, is the organ contour mask information in the sagittal plane direction.

[0035] Further, the optimization processing of the target region segmented by the deep neural network comprises:

[0036] Step 31: using the labeled real medical image data , obtaining target region shape information , wherein i represents the i-th medical image slice, represents the target region shape information of the i-th medical image slice, using the target region shape information to train the RKSSM model to obtain the RKSSM learning parameter;

[0037] Step 32: for a three-dimensional medical image data to be segmented , using the RKSSM model and the probability map M obtained by training to generate the target region initial shape , wherein C k represents the target region initial shape of the k-th medical image slice; according to the Bayes rule, the optimal target region segmentation shape should satisfy the following Bayes equation:

[0038]

[0039] in, The shape of the target region segmentation Maximum likelihood estimation, It is the prior shape probability. It is under the conditions of X and M. The probability will As The initial value;

[0040] Step 33: Introduce a Gaussian mixture model and utilize a probability graph. The Bayesian equation, iteratively solving the above equation. This continues until the maximum likelihood estimate is obtained, thus yielding the optimal target region segmentation shape. ,

[0041] Where M is the probability map for target region segmentation, which is the probability map for a given input 3D medical image data to be segmented. The sigmoid function is used to segment the target region in the axial, coronal, and sagittal directions. , and Probability graph generated through processing :

[0042]

[0043] in, Indicates the first Zhang medical imaging slides, m k For the first Probability diagram of medical image slides. The first The target area in the axial, coronal, and sagittal planes of a medical imaging slice.

[0044] This invention provides a medical imaging organ precision automatic segmentation system, comprising:

[0045] The mask extraction module is used to extract the contour mask information of organs;

[0046] The region segmentation module is used to segment target regions based on deep neural networks using organ contour mask information.

[0047] The optimization processing module is used to perform optimal processing on the target region segmented by the deep neural network.

[0048] The medical image organ precise automatic segmentation method and system of the application use organ contour mask information as the input of training the deep neural network, which is beneficial to solve the underfitting or overfitting problem of the network training model caused by too small foreground (target) area and too large background area in the medical image, greatly optimizes the network training time while ensuring the segmentation accuracy, and improves the segmentation speed.

[0049] The prior art adopts a two-dimensional deep neural network structure when performing pancreas segmentation, without considering the three-dimensional spatial information of the pancreas medical image, the application uses a deep neural network DS-U-Net, processes from three directions at the same time, fully considers the spatial position relationship of the three-dimensional medical image, and is more realistic. In addition, the DS-U-Net can strengthen the boundary contour information of the pancreas and improve the segmentation accuracy.

[0050] Other features and advantages of the present application will be set forth in the following description, and in part will become apparent to those skilled in the art from the description, or can be learned by practice of the present application. The objects and other advantages of the present application can be achieved and obtained by the structures indicated in the specification, claims and drawings. BRIEF DESCRIPTION OF DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0052] Figure 1 A medical image organ precise automatic segmentation method flowchart according to an embodiment of the present application is shown;

[0053] Figure 2 A dense residual U-shaped structure block structure relationship diagram of the DS-U-Net network according to an embodiment of the present application is shown;

[0054] Figure 3 A DS-U-Net network architecture diagram according to an embodiment of the present application is shown;

[0055] Figure 4 A medical image organ precise automatic segmentation system structure diagram according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0056] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0057] The embodiments of the present application provide a medical image organ precise automatic segmentation method. The method is based on a contour (shape) mask, a deep neural network and a statistical Bayesian shape model to optimize automatic segmentation of medical image organs, and realizes precise and rapid automatic segmentation of organs.

[0058] The medical image organ precise automatic segmentation method will be described in detail below with the pancreas organ as an example.

[0059] The medical image organ precise automatic segmentation method mainly includes the following steps:

[0060] Step 1: Extract the contour (shape) mask information of the organ;

[0061] Step 2: Use the organ contour mask information to perform target region segmentation based on a deep neural network;

[0062] Step 3: Optimize the target region segmented by the deep neural network. Each step is described as follows:

[0063] In step 1, the purpose of extracting the organ contour mask information of the medical image is to solve the overfitting or underfitting problem caused by the imbalance between the background and foreground target region data. The embodiments of the present application extract the organ contour mask information of the medical image from three directions, namely the axial plane (Axial), the coronal plane (Coronal) and the sagittal plane (Sagittal), by using a fixed-point model. The main steps include the following:

[0064] Step 11: Obtain the labeled real medical image data of the labeled organ (pancreas) region , use these labeled real image data , add random noise to produce noise medical image data , where i represents the i-th medical image slice. represents the labeled real medical image data of the i-th medical image slice, represents the noise medical image data of the i-th medical image slice.

[0065] Step 12: Use the image data and , train the fixed-point model to obtain the context target region prediction function and pre-segmentation result : , The learned parameters are used to describe the vector matrix representing the relationship between the annotated real medical image data and noisy medical image data .

[0066] Wherein, the fixed-point model can be exemplarily implemented by referring to but not limited to the document “Quannan Li, Jingdong Wang, David Wipf, Zhuowen Tu. Fixed-Point Model for Structured Labeling. International Conference on Machine Learning (ICML 2013)”. The document is incorporated by reference in its entirety into the embodiments of the present application.

[0067] In addition to the fixed-point model method, this step can also use Regression Forests, Classification Forests based on superpixels, Deep Regression Forests, Neural Random Forests, etc.

[0068] Step 13: using the context target region prediction function , respectively, to generate the target region prediction function , and in the axial, coronal and sagittal directions. Wherein, is the axial direction target region prediction function of , is the coronal direction target region prediction function of , is the sagittal direction target region prediction function of .

[0069] Step 14: as shown in Figure 1 , for an input three-dimensional medical image data to be segmented , representing the th medical image slice, is the image data of the kth medical image slice), using the three direction target region prediction functions , and corresponding to generate three direction organ contour mask information , and , organ contour mask information for axial plane direction, organ contour mask information for coronal plane direction, organ contour mask information for sagittal plane direction.

[0070] The organ contour mask information of the organ region of interest (target region) is extracted, which is beneficial to solve the overfitting or underfitting problem of deep neural network training.

[0071] Step 2: target region segmentation based on deep neural network. The deep neural network can adopt V-Net (Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation), U-Net (Convolutional Networks for Biomedical Image Segmentation), fully convolutional neural network (Fully Convolutional Networks: FCN), densely connected convolutional neural network (Densely Connected Convolutional Networks: DenseNet), Dense V-Net, Dense U-Net, etc. The embodiment of the present application preferably adopts a kind of based on densely connected square U-Net (Dense Squared U-Net: DS-U-Net). The DS-U-Net network includes encoder Encoder (encoding block) and decoder Decoder (decoding block).

[0072] Step 21: construct (or obtain) a deep convolutional neural network model for target region segmentation, and the structure and processing process of the network model are as follows.

[0073] The encoder and decoder of the deep convolutional neural network DS-U-Net are both composed of a densely connected residual U-shaped structure block. The main operations performed by the densely connected residual U-shaped structure block include down-sampling, convolution, activation function ReLU, batch normalization and up-sampling. Taking network layer number as an example, the overall structure of the densely connected residual U-shaped structure block is shown in Figure 2 , which can effectively extract multi-scale feature information between different stage layers, mainly composed of three parts:

[0074] (1) Input convolutional layer: The number of input channels is... Feature map Transform into Local feature map of the middle of each channel This layer primarily performs convolution operations to extract local feature information, using a convolution kernel size of [size missing]. For input convolutional layers, see [link / details]. Figure 2 The first and second boxes from top to bottom on the left side of the middle section describe this.

[0075] (2) The layered U-shaped symmetric encoder-decoder structure block, also known as the intermediate structure block, contains three convolutional modules, L-2 upsampling modules, and L-2 upsampling modules. Of the three convolutional modules, the first convolutional module is used for... Intermediate feature map of each channel Perform convolution to obtain Contextual multi-scale feature maps of each channel When executing the dense residual U-shaped structure block, the first convolutional block is executed first, including convolution, batch normalization, and activation; then the upsampling module is executed, and L-2 upsampling modules are connected in sequence to process the intermediate feature maps. Continue continuously The subconvolutional downsampling operation specifically includes: downsampling + convolution + batch normalization + ReLU (Rectified Linear Unit) activation. The second convolutional module out of the three convolutional modules is used to process the intermediate feature maps after the L-2 downsampling modules. The convolution normalization operation includes convolution, batch normalization, and activation. The third convolutional module is connected to the second convolutional module and performs convolution normalization on the output data processed by the second convolutional module, specifically including convolution, batch normalization, and activation. L-2 upsampling modules are connected sequentially, performing L-2 convolutional upsampling operations on the data processed by the third convolutional module, specifically including upsampling + convolution + batch normalization + ReLU activation. Theoretically, the number of layers... The larger the layer size, the deeper the structural block, and the more pooling operations are performed. This results in a larger feature receptive field and richer local and global feature information. The number of layers can be adjusted by setting the parameter. It can extract feature maps with multi-scale spatial resolution;

[0076] (3) Residual connection layer is used to fuse local feature maps and contextual multi-scale feature maps ,that is .See Figure 2The rightmost view shows the boxes from top to bottom of layers 1 and 2. Specifically, the input data for the L-2 upsampling modules and the third convolutional module is processed by residual connection layers, which combine the local feature maps of the previous module for each module. Contextual multi-scale feature maps of the corresponding depth-based downsampling modules The data is obtained by fusion (addition). The data channel is 2M. This embodiment of the invention, by densely connecting local feature maps and multi-scale feature maps, can fully extract the contour and edge information of the image.

[0077] In another embodiment, L can also be equal to 5, that is, a 5-layer dense residual U-shaped structure block is used, where L is a positive integer and greater than 2.

[0078] The Deep Convolutional Neural Network (DS-U-Net) is primarily a U-shaped architecture consisting of encoders and decoders made up of dense residual U-shaped structural blocks. A single DS-U-Net can contain multiple layers of dense residual U-shaped structures with varying numbers of layers (L-values). The network architecture of DS-U-Net is as follows: Figure 3 As shown.

[0079] The following explanation will focus on the execution process and structure of DS-U-Net.

[0080] (1) Encoding process: The encoding process executes T encoding blocks, of which T-2 dense residual U-shaped structure blocks with decreasing layer number are first executed for encoding, and then 2 encoding blocks are executed for dilated convolution, batch normalization and activation.

[0081] For example, first execute 4 The dense residual U-shaped structure blocks of the layer (L=7,6,5,4): Encoding block 1 (L=7), encoding block 2 (L=6), encoding block 3 (L=5), encoding block 4 (L=4) are encoded, and then encoding blocks 5 and 6 are executed. These two blocks mainly use dilated convolution, batch normalization, activation and other operations to replace pooling and upsampling operations.

[0082] (2) The structure of the decoding process is similar to that of the encoding process. Decoding block 5 performs dilated convolution, batch normalization, and activation operations, corresponding to encoding block 5. Then, four operations are performed sequentially. The dense residual U-shaped structure blocks of the layer (L=4,5,6,7): Decoding block 4 (L=4), decoding block 3 (L=5), decoding block 2 (L=6), and decoding block 1 (L=7) perform the corresponding decoding operations.

[0083] The decoding process (which treats 6 as both a coding block and a decoding block) primarily generates (outputs) 6 segmentation probability maps. These 6 segmentation probability maps are then concatenated to generate a fused probability map, thus obtaining the segmentation result.

[0084] Among them, the probabilistic graph is a representation of the input 3D medical image data to be segmented. The sigmoid function is used to segment the target region in the axial, coronal, and sagittal directions. , and Probability graph generated through processing. :

[0085]

[0086] in, Indicates the first Zhang medical imaging slides, m k For the first Probability diagram of medical image slides. The first The target area in the axial, coronal, and sagittal planes of a medical imaging slice.

[0087] The decoding process corresponds to the encoding process. First, dilated convolution, batch normalization, and activation are performed on one decoding block. Then, T-2 dense residual U-shaped blocks with progressively decreasing layer numbers are decoded. The dense residual U-shaped blocks in the decoding block have the same number of layers as the dense residual U-shaped blocks in the corresponding depth of the encoding block. The output of the previous decoding block is concatenated with the output of the corresponding depth of the encoding block and used as the input of the next decoding block to achieve feature reuse. T segmentation probability maps are obtained through decoding. Then, the T segmentation probability maps are concatenated to generate a fused probability map, which yields the segmentation result.

[0088] In this invention example, the DS-U-Net network model mainly includes a 6-stage encoding process, a 5-stage decoding process, and a fusion process. The advantages of this network model are: compared with other neural network models, it can extract richer multi-scale features with relatively low computational and memory consumption; furthermore, since the dense residual U-shaped structural blocks do not use any pre-trained models in image classification, the DS-U-Net network model can be easily and flexibly applied to solve other computer vision image problems.

[0089] Step 22: Utilizing labeled real medical image data Train the DS-U-Net network model to obtain a pre-trained DS-U-Net network model.

[0090] Steps 21 and 22 are performed during the initial construction or training of the deep neural network. Subsequently, the trained deep neural network can be directly obtained for target region segmentation.

[0091] Step 23: For an input test three-dimensional medical image data , using the organ mask information (data) extracted in step 1 、 and , input into the trained DS-U-Net model for segmentation to obtain the target region segmentation (output) results in three directions: axial plane direction target region , coronal plane direction target region and sagittal plane direction target region , as shown in Figure 1 . And generate the probability map M of the target region segmentation result.

[0092] The DS-U-Net network framework in the embodiment of the present application makes full use of the spatial information of the pancreas medical image, can reduce feature loss; the DS-U-Net network encourages feature reuse, avoids learning redundant feature maps, and reduces the probability of overfitting problem; the dense connection mechanism of the DS-U-Net ensures the maximum information flow between layers, simplifies network training; adding auxiliary side path (feature reuse) improves the gradient flow in the network, increases gradient propagation and stabilizes the learning process. At the same time, using this new network method can enhance the edge and boundary contour information of the pancreas and improve the segmentation accuracy.

[0093] Step 3: Optimal processing of target region segmentation. After the target region is segmented by the deep neural network, it still needs to be optimized to further improve the segmentation accuracy. The optimization process proposed in the embodiment of the present application is an important step of the accurate deep learning medical image automatic segmentation method. The implementation of the optimization processing is to use the spatial position constraint of the medical image pancreas voxel (the comprehensive probability of the three-dimensional pixel in three directions, that is, the probability map of the segmentation result), and the optimization method can be Bayesian model, Random Forests algorithm (Random Forests) and the like. The embodiment of the present application takes the Bayesian model as an example to optimize the segmentation result output by step 2, as shown in Figure 1 .

[0094] Step 31: Using the labeled real medical image data , obtain the target region shape information , wherein i represents the i-th medical image slice, represents the target region shape information of the i-th medical image slice, and the target region shape information is used to train a robust kernel statistical shape model (Robust Kernel Statistical Shape Model: RKSSM) to obtain RKSSM learning parameters;

[0095] Step 32: For a three-dimensional medical image data to be segmented Using the RKSSM and probabilistic graph obtained during training Generate the initial shape of the target region. C k This represents the initial shape of the target region in the k-th medical image slice. According to Bayes' theorem, the optimal target region segmentation shape... The following Bayesian equation should be satisfied:

[0096]

[0097] in, The shape of the target region segmentation Maximum likelihood estimation, It is the prior shape probability. It is under the conditions of X and M. The probability will As The initial value.

[0098] Step 33: Introduce a Gaussian Mixture Model and utilize a probability graph. Iteratively solve the above equations The maximum likelihood estimate is obtained, thereby yielding the optimal target region segmentation shape. In each iteration, a Gaussian mixture model is used as the cost function. The Bayesian equation is used to determine if the iteration has converged. If converged, the target region is segmented into its current shape. If the target region segmentation fails to converge, the target region segmentation shape is updated as the optimal segmentation result. Continue iterating.

[0099] By introducing an optimization scheme based on spatial location information or shape prior knowledge, the segmentation results are further refined, and missegmented pancreatic regions are eliminated or reduced, thereby improving the accuracy of organ segmentation in imaging.

[0100] The medical imaging organ precision automatic segmentation method of this invention has a strong generalization ability. In addition to segmenting the pancreas, it can also precisely segment other abdominal organs such as the liver, spleen, kidneys, and stomach.

[0101] Based on the same inventive concept, embodiments of the present invention also provide a medical imaging organ precision automatic segmentation system, such as... Figure 4 As shown, the system includes:

[0102] The mask extraction module is used to extract the contour mask information of organs;

[0103] The region segmentation module is used to segment target regions based on deep neural networks using organ contour mask information.

[0104] An optimization processing module is configured to perform optimization processing on the target region segmented by the deep neural network.

[0105] The specific implementation manners of the various modules of the medical image organ precise automatic segmentation system of the embodiments of the present application can be obtained according to any embodiment of the medical image organ precise automatic segmentation method of the present application, and will not be described again.

[0106] Although the present application has been described in detail with reference to the foregoing embodiments, it should be understood by those skilled in the art that the technical solutions recorded in the foregoing embodiments can still be modified, or some technical features can be replaced by equivalents; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A medical image organ precise automatic segmentation method, characterized in that, The method comprises the following steps: Step 1: extracting contour mask information of the organ; Specifically, the method comprises the following steps: Step 11: obtaining labeled real medical image data of labeled organ region , using the labeled real image data , adding random noise to produce noisy medical image data where i represents the i-th medical image slice, represents the labeled real medical image data of the i-th medical image slice, represents the noisy medical image data of the i-th medical image slice; Step 12: training the fixed-point model to obtain a context target region prediction function and ​​​​​​​ Step 13: Using the contextual target region prediction function to generate three target region prediction functions in axial, coronal and sagittal directions respectively: an axial direction target region prediction function , a coronal direction target region prediction function and a sagittal direction target region prediction function ; Step 14: For an input three-dimensional medical image data to be segmented , k represents the k-th medical image slice, is the image data of the k-th medical image slice, three direction organ contour mask information , and are generated by using three direction target region prediction functions , and , is the organ contour mask information in the axial plane direction, is the organ contour mask information in the coronal plane direction, is the organ contour mask information in the sagittal plane direction; Step 2: performing target region segmentation based on the deep neural network by using the contour mask information of the organ; Step 3: performing optimization processing on the target region segmented by the deep neural network.

2. The medical image organ accurate automatic segmentation method according to claim 1, wherein the deep neural network comprises an encoding block and a decoding block, and the encoding block and the decoding block respectively adopt a dense residual U-shaped structure block.

3. The medical image organ accurate automatic segmentation method according to claim 2, wherein the dense residual U-shaped structure block comprises an input convolutional layer, a U-shaped symmetric encoding-decoding structure block and a residual connection layer, and the U-shaped symmetric encoding-decoding structure block has L layers.

4. The medical image organ accurate automatic segmentation method according to claim 3, wherein the U-shaped symmetric encoding-decoding structure block comprises L-2 up-sampling modules and L-2 down-sampling modules.

5. The medical image organ accurate automatic segmentation method according to claim 4, wherein the U-shaped symmetric encoding-decoding structure block comprises L-2 up-sampling modules and L-2 down-sampling modules. The L-2 up-sampling modules are connected in sequence and used for performing L-2 convolution up-sampling operations on input data, and specifically comprise up-sampling, convolution, batch normalization and ReLU activation. The input convolutional layer is used to convert the input into a system with a number of channels. Feature map Transform into Local feature map of the middle of each channel To extract local feature information. The L-2 down-sampling modules are connected in sequence and used for performing L-2 convolution down-sampling operations on input data, and specifically comprise down-sampling, convolution, batch normalization and ReLU activation. The method comprises the following steps: Setting the number of layers L and extracting feature maps of multi-scale spatial resolution.

7. The medical image organ accurate automatic segmentation method according to claim 5, wherein the deep neural network adopts a U-shaped structure and comprises an encoding process and a decoding process.

6. The medical image organ precise automatic segmentation method according to claim 5, characterized in that, 8. The medical image organ accurate automatic segmentation method according to any one of claims 2-7, wherein the deep neural network adopts a U-shaped structure and comprises an encoding process and a decoding process. The encoding process comprises performing T encoding blocks, wherein T-2 dense residual U-shaped structure blocks with layers decreasing in sequence are first performed for encoding, and then 2 encoding blocks are performed for performing dilated convolution, batch normalization and activation. The decoding process corresponds to the encoding process, wherein dilated convolution, batch normalization and activation are first performed by 1 decoding block, and then T-2 dense residual U-shaped structure blocks with layers decreasing in sequence are performed for decoding, and the dense residual U-shaped structure blocks of the decoding block have the same number of layers as the corresponding deep encoding block. The residual connection layer is used to fuse the local feature maps and the contextual multi-scale feature maps .

9. The medical image organ accurate automatic segmentation method according to claim 8, wherein the target region segmentation based on the deep neural network comprises the following steps: The optimization processing on the target region segmented by the deep neural network comprises the following steps: The method comprises the following steps: The mask extraction module is used for extracting contour mask information of the organ; Specifically, the method comprises the following steps: Utilizing labeled real medical image data to train a final deep neural network.

10. The medical image organ precise automatic segmentation method according to claim 8, characterized in that, The region segmentation module is used for performing target region segmentation based on the deep neural network by using the contour mask information of the organ; For an input test three-dimensional medical image data , extracted organ mask information , and , input to a trained deep neural network for segmentation to obtain target region segmentation (output) results in three directions: axial plane direction target region , coronal plane direction target region and sagittal plane direction target region ; wherein organ contour mask information in the axial plane direction, organ contour mask information in the coronal plane direction, organ contour mask information in the sagittal plane direction.

11. The medical image organ precise automatic segmentation method according to any one of claims 1-7, characterized in that, The optimization processing module is used for performing optimization processing on the target region segmented by the deep neural network. Step 31: obtaining target region shape information by using the labeled real medical image data , wherein i represents the i-th medical image slice, , wherein i represents the i-th medical image slice, , wherein i represents the i-th medical image slice, training the RKSSM model to obtain RKSSM learning parameters; Step 32: For a three-dimensional medical image data to be segmented , using the RKSSM model and the probability map M obtained by training, an initial shape of the target region is generated , wherein C k represents the initial shape of the target region of the k-th medical image slice; according to the Bayes rule, the optimal target region segmentation shape should satisfy the following Bayes equation: wherein, is a maximum likelihood estimate of the target region segmentation shape is a prior shape probability, is a probability of under the condition of (X, M), and is taken as an initial value of ​​​ Step 33: Introduce Gaussian Mixture Model, simultaneously utilize Bayesian equation of probability graph , iteratively solve above equation , until maximum likelihood estimation is obtained, thus obtaining optimal target region segmentation shape , Where M is the probability map for target region segmentation, which is the probability map for a given input 3D medical image data to be segmented. The sigmoid function is used to segment the target region in the axial, coronal, and sagittal directions. , and Probability graph generated through processing : wherein, represents the medical image slice, m k represents the probability map of the medical image slice, m target region of the medical image slice m in the axial plane, the coronal plane and the sagittal plane, respectively.

12. A medical image organ precise automatic segmentation system, characterized in that, ​ ​ ​ Step 11: obtaining labeled real medical image data of labeled organ region , using the labeled real image data , adding random noise to produce noisy medical image data , where i represents the i-th medical image slice, represents the labeled real medical image data of the i-th medical image slice, represents the noisy medical image data of the i-th medical image slice; Step 12: training the fixed-point model to obtain a context target region prediction function and and a pre-segmentation result : , is a learned parameter for describing the relationship between the annotated real medical image data and the noisy medical image data .​ Step 13: Using the contextual target region prediction function to generate three target region prediction functions in axial, coronal and sagittal directions respectively: an axial direction target region prediction function , a coronal direction target region prediction function and a sagittal direction target region prediction function ; Step 14: For an input 3D medical image data to be segmented , k represents the k-th medical image slice, is the image data of the k-th medical image slice, three direction organ contour mask information 、 and are generated correspondingly by using three direction target region prediction functions 、 and , is the organ contour mask information in the axial plane direction, is the organ contour mask information in the coronal plane direction, is the organ contour mask information in the sagittal plane direction; ​ ​

Images