Medical image segmentation method and system, electronic equipment and storage medium

Abstract

The invention discloses a medical image segmentation method and system, electronic equipment and a storage medium, and the method comprises the steps: obtaining medical image data containing a target medical structure; the target medical structure comprises a target tissue and / or a target organ; inputting the medical image data into a backbone network to obtain a contour region and a segmentation starting point of the target medical structure; and segmenting the medical image according to the contour region and the segmentation starting point. According to the invention, the medical image range of the target medical structure is determined through the contour region and the segmentation starting point obtained by the backbone network, and the medical image range is segmented, so that the medical image is segmented efficiently, the accurate medical image corresponding to the target medical structure is obtained, and the efficiency and accuracy of clinical diagnosis are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a medical image segmentation method, system, electronic equipment and storage medium. Background Art

[0002] Cancer is a major cause of death and life expectancy around the world. As the incidence of cancer continues to increase worldwide, using existing science and technology to help doctors conduct early screening and diagnosis will help with treatment, slow the progression of the disease, and increase life expectancy. To achieve this goal, automatic recognition of organs and parts of medical images is one of the key factors in improving diagnostic efficiency. Currently, with the development of deep learning technology, some supervised neural network models based on backbone networks have been widely used in the medical field for image segmentation, but it is not easy to obtain high-quality medical image labels through these neural network models. First of all, in terms of data volume, a large amount of effective training data is required; secondly, labeling medical image data requires the labeler to have professional knowledge, that is, effective training data needs to be collected by professionals; finally, because there is no fixed standard in the industry, the observation results of professionals will still be different.

[0003] However, it is not easy to obtain high-quality medical image labels. In terms of data volume, general computer vision image datasets can easily collect more than a thousand labeled data for each category, while under the same conditions, there are only hundreds of available medical images. This is due to many reasons: in order to protect the privacy of patients, the types and quantities of open medical datasets are very limited; even if a certain amount of available data is obtained within the institution, unlike general image annotation, annotating medical image data requires the annotator to have professional knowledge; even if they have received professional training, their annotation results will still vary due to differences in individual observations, and there is no effective gold standard in the industry. Considering the reality of small number of samples and large differences, there are still many problems in applying supervised deep learning models to clinical practice. Summary of the invention

[0004] The technical problem to be solved by the present invention is to overcome the low efficiency of medical image segmentation and low accuracy of segmentation results in the prior art, and to provide a medical image segmentation method, system, electronic device and storage medium.

[0005] The present invention solves the above technical problems through the following technical solutions:

[0006] The present invention provides a medical image segmentation method, the medical image segmentation method comprising:

[0007] Acquiring medical image data containing a target medical structure; the target medical structure includes a target tissue and / or a target organ;

[0008] Inputting the medical image data into a backbone network to obtain a contour area and a segmentation starting point of the target medical structure;

[0009] The medical image is segmented according to the contour area and the segmentation starting point.

[0010] Preferably, the step of inputting the medical image data into a backbone network to obtain the contour area and segmentation starting point of the target medical structure comprises:

[0011] extracting features of the medical image data;

[0012] Performing convolution kernel dimensionality reduction on the features to obtain a contour region of the target medical structure and probability values ​​of pixels within the contour region; wherein the probability values ​​are proportional to the possibility that the pixels constitute the contour of the target medical structure;

[0013] The pixel point with the largest probability value in the contour area is determined as the segmentation starting point.

[0014] Preferably, the step of segmenting the medical image according to the contour area and the segmentation starting point comprises:

[0015] The contour area and the segmentation starting point are input into a deep reinforcement learning model to obtain a contour map of the medical structure.

[0016] Preferably, after the step of segmenting the medical image according to the contour area and the segmentation starting point, the step further includes:

[0017] The segmented medical image is processed in three-dimensional space by Kalman filtering to obtain a three-dimensional image.

[0018] Preferably, the step of acquiring medical image data containing the target medical structure comprises:

[0019] Acquire T2W images and ADC images of target medical structures;

[0020] Performing standardization processing on the T2W image and the ADC image; wherein the standardization processing includes at least one of normalization processing, resolution unification, and gray value unification;

[0021] Medical image data containing the target medical structure is output according to the standardized processing result.

[0022] Preferably, the step of extracting features of the medical image data comprises:

[0023] Acquire a T2W image, an ADC image, and an image obtained by Laplace transforming the T2W image corresponding to the medical image data;

[0024] Performing three-dimensional stitching on the T2W image, the ADC image, and an image obtained by Laplace transform of the T2W image;

[0025] The 3D stitching result is input into the convolutional neural network or the Transformer neural network and the features are output.

[0026] The present invention also provides a medical image segmentation system, the medical image segmentation system comprising:

[0027] An acquisition module, configured to acquire medical image data containing a target medical structure; the target medical structure includes a target tissue and / or a target organ;

[0028] An output module, used for inputting the medical image data into a backbone network to obtain a contour area and a segmentation starting point of the target medical structure;

[0029] The segmentation module is used to segment the medical image according to the contour area and the segmentation starting point.

[0030] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and used to run on the processor, wherein the processor implements the above-mentioned medical image segmentation method when executing the computer program.

[0031] The present invention also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the medical image segmentation method described above is implemented.

[0032] On the basis of being in accordance with the common sense in the art, the above-mentioned preferred conditions can be arbitrarily combined to obtain the preferred embodiments of the present invention.

[0033] The positive and progressive effects of the present invention are:

[0034] The present invention determines the medical image range of the target medical structure and segments it through the contour area and segmentation starting point obtained by the backbone network, thereby efficiently segmenting the medical image and obtaining a more accurate medical image corresponding to the target medical structure, thereby improving the efficiency and accuracy of clinical diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 is a flow chart of a medical image segmentation method according to Embodiment 1 of the present invention;

[0036] Figure 2 This is a flowchart of Example 1 of the medical image segmentation method according to Embodiment 1 of the present invention;

[0037] Figure 3 This is a diagram of a convolutional neural-based backbone network architecture of Example 1 of the medical image segmentation method of Embodiment 1 of the present invention;

[0038] Figure 4 This is a schematic diagram of obtaining contour areas and segmentation points in Example 1 of the medical image segmentation method according to Embodiment 1 of the present invention;

[0039] Figure 5 A T2W image containing prostate tissue according to Example 1 of the medical image segmentation method of Embodiment 1 of the present invention;

[0040] Figure 6 A T2W image including a prostate tissue segmentation starting point in Example 1 of the medical image segmentation method of Embodiment 1 of the present invention;

[0041] Figure 7 A contour area map of the prostate tissue determined by a convolutional neural backbone network in Example 1 of the medical image segmentation method of Embodiment 1 of the present invention;

[0042] Figure 8 A network architecture diagram of a deep reinforcement learning model of Example 1 of the medical image segmentation method of Embodiment 1 of the present invention;

[0043] Fig. 9 This is a schematic diagram of obtaining a contour map through deep reinforcement learning model serialization in Example 1 of the medical image segmentation method of Example 1 of the present invention;

[0044] Fig.10 This is a diagram of the Transformer backbone network structure architecture of Example 2 of the medical image segmentation method of Example 1 of the present invention;

[0045] Fig.11 This is a schematic diagram of a contour area obtained through a Transformer backbone network structure in Example 2 of the medical image segmentation method of Example 1 of the present invention;

[0046] Fig.12 is a structural diagram of a medical image segmentation system according to Embodiment 2 of the present invention;

[0047] Fig.13 This is a schematic diagram of the structure of an electronic device according to Embodiment 3 of the present invention. DETAILED DESCRIPTION

[0048] The present invention is further described below by way of examples, but the present invention is not limited to the scope of the examples.

[0049] Example 1

[0050] This embodiment provides a medical image segmentation method. Figure 1 , medical image segmentation methods include:

[0051] S1. Acquire medical image data containing a target medical structure.

[0052] The target medical structures include target tissues, target organs, etc. The target tissues include prostate tissue, breast tissue, etc.; the target organs include heart, liver, lung, stomach, kidney, etc.

[0053] S2. Input the medical image data into the backbone network to obtain the contour area and segmentation starting point of the target medical structure.

[0054] Among them, the backbone network includes convolutional neural network, Transformer (supervised target network, a neural network model based on attention mechanism), etc.

[0055] In an optional embodiment, step S2 includes:

[0056] S21. Extract features of medical image data.

[0057] When the backbone network is a convolutional neural network, ResNet152 (a deep convolutional neural network model for image classification) can be used to extract features to obtain features of different granularities at different scales, and the features at different scales are fused together through a feature pyramid to output the final features.

[0058] When the backbone network is a Transformer neural network, SwinTransformer (a deep learning model based on Transformer) can be used to extract features. The obtained features are upsampled by one layer to maintain the output dimension consistent with other backbone networks.

[0059] In an optional implementation, step S21 includes:

[0060] S211 , acquiring a T2W image corresponding to the medical image data, an ADC image, and an image obtained by Laplace transforming the T2W image.

[0061] S212 , three-dimensionally stitching the T2W image, the ADC image, and the image obtained by Laplace transform of the T2W image.

[0062] S213, input the three-dimensional stitching result into a convolutional neural network or a Transformer neural network, and output features.

[0063] In this embodiment, a three-dimensional image is synthesized through three-dimensional stitching to display the medical image more intuitively and vividly, obtain more accurate features, and thus improve the accuracy of the segmentation contour map.

[0064] S22. Perform convolution kernel dimensionality reduction on the features to obtain the contour area of ​​the target medical structure and the probability values ​​of the pixels in the contour area.

[0065] The probability value is proportional to the possibility that the pixel points constitute the contour of the target medical structure.

[0066] S23, determining the pixel point with the largest probability value in the contour area as the segmentation starting point.

[0067] S3. Segment the medical image according to the contour area and the segmentation starting point.

[0068] This embodiment determines the medical image range of the target medical structure through the contour area and segmentation starting point obtained by the backbone network and segments it, thereby efficiently segmenting the medical image and obtaining a more accurate medical image corresponding to the target medical structure, thereby improving the efficiency and accuracy of clinical diagnosis.

[0069] In an optional embodiment, step S1 includes:

[0070] S11. Acquire a T2W image and an ADC image of a target medical structure.

[0071] Among them, T2W images are images generated based on T2-weighted imaging technology, which is a type of magnetic resonance imaging; T2-weighted imaging is one of the most commonly used imaging techniques in MRI, which can provide information on tissue water content and distribution. ADC images are apparent diffusion coefficient maps, which are calculated through diffusion-weighted imaging.

[0072] S12. Performing standardization processing on the T2W image and the ADC image.

[0073] Among them, the standardization processing includes image resizing processing, normalization processing, resolution unification and grayscale value unification.

[0074] S13. Outputting medical image data containing the target medical structure according to the standardized processing result.

[0075] In this embodiment, the T2W image and the ADC image are processed by standardization to obtain medical image data that meets the preset standard, thereby improving the segmentation efficiency.

[0076] In an optional embodiment, step S3 includes:

[0077] S31. Input the contour area and the segmentation starting point into the deep reinforcement learning model to obtain a contour map of the medical structure.

[0078] The following introduces a training process of a deep reinforcement learning model. In this embodiment, it is defined as a Markov decision process (MDP). The Markov decision process has five elements (S, A, P, R, γ), S is used to characterize the state, that is, taking the current segmentation point as the center, specifying a preset range as the region of interest; A is used to characterize the action space, that is, the forward direction of the current segmentation point (including up, down, left, right, upper left, lower left, upper right, lower right, etc.); P is used to characterize the policy, that is, the way in which the reinforcement learning model selects the forward direction. In this embodiment, similar to the traditional DQN algorithm (deep Q-network, a Q-Learing algorithm based on deep learning), the forward direction is selected through a greedy policy; R is used to characterize the reward function, that is, the quality of the current segmentation result. The reward value of the current forward direction action is obtained by weighted average by calculating the change in the Dice (a set similarity measurement function) coefficient after adding the current segmentation point, the distance between the segmentation line and the actual contour, and the degree of convergence of the segmentation point set; γ is used to characterize the discount factor, that is, the cumulative impact of the current forward direction selection on the long term.

[0079] The general process is to start from the segmentation starting point, determine the forward direction of the segmentation point to obtain the next segmentation point, until the segmentation is completed. The specific steps are as follows:

[0080] For the kth current segmentation point Vk, the preset range centered on it is the region of interest, and the probability value of the pixel points in the region of interest is adjusted using the deep reinforcement learning model. The next segmentation point Vk+1 is determined based on the current segmentation point and the preset segmentation step. The relative position of Vk+1 and Vk may be up, down, left, right, upper left, lower left, upper right, and lower right. The relative position of Vk+1 and Vk is determined by a greedy policy to prevent overfitting of the deep reinforcement learning model. The relative position of Vk+1 and Vk may be up, down, left, right, upper left, lower left, upper right, and lower right. The network used in the deep reinforcement learning model of this embodiment adopts a lightweight convolutional network to reduce the overall scale of the deep reinforcement learning model.

[0081] For the reward function of the deep reinforcement learning model, it is necessary to comprehensively consider the quality of the contour map and the efficiency of the segmentation contour map. The reward function of this implementation is a weighted average of three coefficients, which are the difference between the Dice coefficient of the current segmentation point and the Dice coefficient of the previous round of segmentation points after the new segmentation point is added, the shortest distance between the segmentation point set and the actual contour, and the distance between the segmentation points. In addition, in order to take into account the efficiency of the segmentation contour map, when the segmentation point set is too concentrated, the reward function will give an appropriate penalty.

[0082] In this embodiment, the contour area output from the backbone network is a relatively fuzzy area. Inputting the contour area into the deep reinforcement learning model can obtain a more accurate contour map, thereby further improving the accuracy of clinical diagnosis.

[0083] In an optional implementation, step S31 further includes:

[0084] S311, performing noise reduction and smoothing processing on the contour image through Fourier transform.

[0085] In this embodiment, by performing noise reduction and smoothing processing on the contour image, the image quality of the contour image is improved and the display effect of the image is enhanced.

[0086] In an optional embodiment, step S3 further includes:

[0087] S32, processing the segmented medical image in three-dimensional space by Kalman filtering to obtain a three-dimensional image.

[0088] In this embodiment, the segmented images of various tissues or organs of the same patient can be stacked into a complete three-dimensional image, and then post-processed in three-dimensional space through Kalman filtering. The Kalman gain is used to correct the contour map, eliminate outliers, improve the image display effect, and improve the accuracy of clinical diagnosis.

[0089] In an optional embodiment, the segmentation results can be stored in a reasonable format, such as a two-dimensional array of a two-dimensional image corresponding to the coordinates of each segmentation point, and visualized on the original medical image. The segmentation results can also be uploaded to a database for easy query and viewing by users.

[0090] Two specific examples of automatic segmentation of prostate tissue are introduced below.

[0091] Example 1

[0092] This example uses a convolutional neural network as the backbone network and a deep reinforcement learning model to automatically segment prostate tissue in the subject's magnetic resonance imaging medical images.

[0093] Figure 2 Here is the flowchart for this example:

[0094] S201. Perform data preprocessing on medical image data to obtain a training data set and a test data set.

[0095] Specifically: a magnetic resonance device is used to scan the target medical structure of the subject to be tested, a T2W image is generated, and an apparent diffusion coefficient (ADC) image is calculated by diffusion-weighted imaging. Optionally, if a computed tomography scanner is used to scan the target medical structure of the subject to be tested, a CT image is generated. The T2W image and the ADC image are normalized by Z-score (standard score) and grayscale value, and then the data resolution is unified to 384×384. The gradient filter of the computer vision library is used to obtain the edge information of the medical image data, and the collected medical image data is marked according to clinical prior knowledge. The medical image data is divided into a training data set, a validation data set, and a test data set.

[0096] S202. Extract features through a convolutional neural network.

[0097] Specifically: the medical image data after data preprocessing is input into ResNet152. Figure 3 The following is a diagram of the backbone network architecture based on convolutional neural networks. The rectangular box represents the network output feature map. The 384×384 below the rectangular box indicates that the length and width of the feature map are both 384; the 192×192×64 below the rectangular box indicates that the length of the feature map is 192, the width is 192, and the number of channels is 64. The 3 above indicates that the number of channels in the feature map is 3; the 64 above indicates that the number of channels in the feature map is 64; the 256 above indicates that the number of channels in the feature map is 256; the 512 above indicates that the number of channels in the feature map is 512; and the 1024 above indicates that the number of channels in the feature map is 1024. For the input image, the arrows in the first layer indicate the convolution operation, including convolution, normalization, and activation functions. The output of each convolution operation represents the features of a specific granularity at a specific scale. This example uses the feature pyramid to fuse the features at different scales together, that is, unify the feature maps of different scales by upsampling, and splice the features together. The final feature map is obtained through one-step convolution, and the size is 48×48×128.

[0098] S203: Determine the contour area and the segmentation starting point.

[0099] Specifically: for the features extracted by the backbone network, a 1×1 convolution kernel is used to reduce the dimension and obtain the contour area. Figure 4As shown in the figure, for the obtained 48×48×128 feature map, it is reduced to a 48×48×1 feature map. After the feature map is transformed into one dimension, it is mapped to a value of (0, 1) through a softmax (normalized exponential function) layer, and the probability of each value on the vector forming a contour map is expressed in the form of probability. The point with the highest probability is determined as the segmentation starting point. At the same time, the feature map after dimensionality reduction reflects the location information of the contour area. The two outputs are scaled to form the scale of the original image (384×384), and the contour area and segmentation starting point are then input into the deep reinforcement learning model. Figure 5 It is a T2W image containing prostate tissue; Figure 6 The segmentation origin (marked by an asterisk) is shown; Figure 7 The contour area of ​​the prostate tissue determined by the convolutional neural backbone network.

[0100] S204. Perform contour mapping and segmentation through a deep reinforcement learning model.

[0101] Specifically: When outlining, select the next segmentation point in a serial manner to maximize the cumulative benefit. The probability value of the pixel points in the contour area is regarded as the environment; for each state (i.e., with the current segmentation point as the center, specify the preset range as the area of ​​interest), a value function is used to calculate the benefits of all possible actions in the current state (represented by the agent), and the action with the largest benefit is selected to interact with the environment. The environment is thus advanced to the next state, and the actual benefit of the action selection is calculated. Repeat this iteration until the end condition is met and the training obtains the maximum cumulative benefit.

[0102] The deep reinforcement learning model used in this example is based on DQN. The specific network architecture diagram is as follows: Figure 8 The training of deep reinforcement learning models relies on two networks, a learning network (represented by Q) and a target network Q' (represented by Q'). The two networks are constructed in the same way, and the parameters ω' of Q' are copied from the parameters ω of Q every once in a while. The jth interaction between the agent and the environment will result in a state and take a certain action A. j , you will get a reward R j , the environment enters state s′ j . Thus we get a sample (φ(s j ), A j , R j ,φ(s′ j )) and use it as a historical interaction data. Calculate the target value based on this historical interaction data. The following is the basic calculation formula for the current target value:

[0103]

[0104] Among them, yj Used to characterize the target value, R j It is used to represent the reward obtained by the agent's interaction with the environment, j is used to represent the number of interactions between the agent and the environment, γ is used to represent the discount factor when calculating Q, which is used to adjust the weights of short-term and long-term benefits, a′ is used to represent the action that maximizes the target value, and φ(s′ j ) represents the feature vector of the next state. Using φ(s) as input in the Q network, we can get the Q value output corresponding to all actions of the Q network. j It is used to characterize the corresponding action selected in the current Q value input, and ω′ is the model parameter of the Q′ network.

[0105] The loss function in this example uses the mean square error, and the formula is as follows:

[0106]

[0107] Among them, m is used to represent the number of samples sampled from the historical interaction data each time the loss function is calculated. The model parameter ω of the Q network is updated through the gradient back propagation of the neural network, and the model parameter ω′ corresponds to ω.

[0108] Actions (i.e., the direction of travel) on a two-dimensional plane can be divided into eight types: up, down, left, right, upper left, lower left, upper right, lower right, and a stop action. This example sets each step length to 3 pixels. Therefore, when the current segmentation point and action of the contour are determined, the next segmentation point can be determined. In order to effectively depict and segment the contour, this example proposes a multi-granularity comprehensive benefit to balance the local and overall state changes brought about by each step of the action. At the finest granularity, the distance between the segmentation point corresponding to the current action and the contour area is evaluated. At the medium granularity, the dispersion of the last 10 segmentation points is evaluated. At the coarse granularity, the Dice coefficient between the depicted contour map and the actual contour is evaluated. The scores at the three granularities are weighted to obtain the comprehensive benefit of the current action. Fig. 9 It shows a schematic diagram of how the segmentation contour map is finally drawn, starting from the segmentation starting point, through iterative interaction between the benefits and the environment.

[0109] S205: De-noising, smoothing and post-processing the segmentation results.

[0110] Specifically, the data with significant fluctuations are mapped to the high-frequency domain in two-dimensional space through Fourier transform, and the image data with low fluctuations is filtered through a low-pass filter and remapped to the time domain to generate a smooth contour map boundary. Then all images of the same patient are stacked into a complete three-dimensional image, and post-processed in three-dimensional space using Kalman filtering, from top to bottom and bottom to top of the three-dimensional image, and outliers are removed based on the difference between the current contour and the predicted contour using Kalman gain to obtain the final segmentation result.

[0111] Example 2

[0112] The difference between Example 2 and Example 1 is that Example 1 uses a convolutional neural network to extract features, while Example 2 uses a Transformer backbone network to extract features.

[0113] Specifically: the medical image data after data preprocessing is input into Swin Transformer. Fig.10 The following is a diagram of the Transformer backbone network structure. The rectangular box represents a Transformer stage. The 96×96×48 in the upper left corner of the rectangular box indicates that the input length is 96, the width is 96, and the number of channels is 48; the 96×96×C in the upper left corner of the rectangular box indicates that the input length is 96, the width is 96, and the number of channels is C. In this example, C is 128. The 384×384 below the rectangular box indicates that the length and width of the feature map are both 384; the 12×12 below the rectangular box indicates that the length and width of the feature map are both 12; and the 48×48 below the rectangular box indicates that the length and width of the feature map are both 48. Each Transformer stage will reduce the resolution of the input feature map and expand the receptive field layer by layer. Unlike convolutional neural networks, the image will be cut into fixed-size blocks (4×4 blocks in this example) before input, and each block will be added with relative position encoding information and input into the network. This example uses the encoding layer of SwinTransformer as the backbone network to extract features. The final feature map is obtained through one-step convolution and upsampling. The size of the map is 48×48×128.

[0114] Fig.11 A schematic diagram of the contour area obtained through the Transformer backbone network.

[0115] Example 2

[0116] This embodiment provides a medical image segmentation system. Fig.12 , the medical image segmentation system includes:

[0117] The acquisition module 1 is used to acquire medical image data containing a target medical structure. The target medical structure includes a target tissue and / or a target organ.

[0118] Output module 2 is used to input medical image data into the backbone network to obtain the contour area and segmentation starting point of the target medical structure.

[0119] The segmentation module 3 is used to segment the medical image according to the contour area and the segmentation starting point.

[0120] In an alternative embodiment, see Fig.12 , the medical image segmentation system also includes:

[0121] The feature extraction module 4 is used to extract features of medical image data.

[0122] The output module 2 is also used to perform convolution kernel dimensionality reduction on the features to obtain the contour area of ​​the target medical structure and the probability values ​​of the pixels in the contour area.

[0123] The probability value is proportional to the possibility that the pixel points constitute the contour of the target medical structure.

[0124] The determination module 5 is used to determine the pixel point with the largest probability value in the contour area as the segmentation starting point.

[0125] In an optional embodiment, the segmentation module 3 is further used to input the contour area and the segmentation starting point into the deep reinforcement learning model to obtain a contour map of the medical structure.

[0126] In an alternative embodiment, see Fig.12 , the medical image segmentation system also includes:

[0127] The processing module 6 is used to process the segmented medical image in three-dimensional space by using Kalman filtering to obtain a three-dimensional image.

[0128] In an optional embodiment, the acquisition module 1 is further used to acquire a T2W image and an ADC image of the target medical structure.

[0129] The processing module 6 is further used to perform standardization processing on the T2W image and the ADC image, wherein the standardization processing includes at least one of normalization processing, resolution unification and gray value unification.

[0130] The output module 2 is further used to output medical image data containing the target medical structure according to the standardized processing result.

[0131] In an optional implementation, the acquisition module 1 is further used to acquire a T2W image, an ADC image, and an image obtained by Laplace transforming the T2W image corresponding to the medical image data.

[0132] See also Fig.12 , the medical image segmentation system also includes:

[0133] The three-dimensional stitching module 7 is used to perform three-dimensional stitching on the T2W image, the ADC image and the image obtained by Laplace transform of the T2W image.

[0134] The output module 2 is also used to input the three-dimensional stitching result into the convolutional neural network or the Transformer neural network and output features.

[0135] It should be noted that the implementation principles and technical effects of each module of this embodiment can refer to the corresponding parts of Example 1 and will not be repeated here.

[0136] Example 3

[0137] This embodiment provides an electronic device, Fig.13 The electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the medical image segmentation method of the first embodiment is implemented. Fig.13 The electronic device 30 shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present invention.

[0138] like Fig.13 As shown, the electronic device 30 may be in the form of a general-purpose computing device, for example, it may be a server device. The components of the electronic device 30 may include, but are not limited to: at least one processor 31, at least one memory 32, and a bus 33 connecting different system components (including the memory 32 and the processor 31).

[0139] The bus 33 includes a data bus, an address bus, and a control bus.

[0140] The memory 32 may include a volatile memory, such as a random access memory (RAM) 321 and / or a cache memory 322 , and may further include a read-only memory (ROM) 323 .

[0141] The memory 32 may also include a program / utility 325 having a set (at least one) of program modules 324, such program modules 324 including but not limited to: an operating system, one or more application programs, other program modules, and program data, each of which or some combination may include an implementation of a network environment.

[0142] The processor 31 executes various functional applications and data processing by running the computer programs stored in the memory 32, such as the medical image segmentation method of embodiment 1 of the present invention.

[0143] The electronic device 30 may also communicate with one or more external devices 34 (e.g., keyboards, pointing devices, etc.). Such communication may be performed via an input / output (I / O) interface 35. Furthermore, the model generating device 30 may also communicate with one or more networks (e.g., a local area network (LAN), a wide area network (WAN), and / or a public network, such as the Internet) via a network adapter 36. Fig.13As shown, the network adapter 36 communicates with other modules of the model-generated device 30 via the bus 33. It should be understood that, although not shown in the figure, other hardware and / or software modules can be used in conjunction with the model-generated device 30, including but not limited to: microcode, device drivers, redundant processors, external disk drive arrays, RAID (disk array) systems, tape drives, and data backup storage systems, etc.

[0144] It should be noted that although several units / modules or sub-units / modules of the electronic device are mentioned in the above detailed description, this division is merely exemplary and not mandatory. In fact, according to an embodiment of the present invention, the features and functions of two or more units / modules described above can be embodied in one unit / module. Conversely, the features and functions of one unit / module described above can be further divided into multiple units / modules to be embodied.

[0145] Example 4

[0146] This embodiment provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the medical image segmentation method of Embodiment 1 is implemented.

[0147] The readable storage medium may include but is not limited to: a portable disk, a hard disk, a random access memory, a read-only memory, an erasable programmable read-only memory, an optical storage device, a magnetic storage device or any suitable combination of the above.

[0148] In a possible implementation manner, the present invention may also be implemented in the form of a program product, which includes a program code. When the program product is run on a terminal device, the program code is used to enable the terminal device to execute the medical image segmentation method of embodiment 1.

[0149] Among them, the program code for executing the present invention can be written in any combination of one or more programming languages, and the program code can be executed completely on the user device, partially on the user device, as an independent software package, partially on the user device and partially on a remote device, or completely on the remote device.

[0150] Although the specific embodiments of the present invention are described above, it should be understood by those skilled in the art that this is only for illustration and the protection scope of the present invention is defined by the appended claims. Those skilled in the art may make various changes or modifications to these embodiments without departing from the principles and essence of the present invention, but these changes and modifications all fall within the protection scope of the present invention.

Claims

1. A medical image segmentation method, characterized in that: The medical image segmentation method comprises: Acquiring medical image data containing a target medical structure; the target medical structure includes a target tissue and / or a target organ; Inputting the medical image data into a backbone network to obtain a contour area and a segmentation starting point of the target medical structure; The medical image is segmented according to the contour area and the segmentation starting point.

2. The medical image segmentation method according to claim 1, characterized in that: The step of inputting the medical image data into a backbone network to obtain the contour area and segmentation starting point of the target medical structure comprises: extracting features of the medical image data; Perform convolution kernel dimensionality reduction on the features to obtain a contour region of the target medical structure and probability values ​​of pixels within the contour region; wherein the probability values ​​are proportional to the possibility that the pixels constitute the contour of the target medical structure; The pixel point with the largest probability value in the contour area is determined as the segmentation starting point.

3. The medical image segmentation method according to claim 1, characterized in that: The step of segmenting the medical image according to the contour area and the segmentation starting point comprises: The contour area and the segmentation starting point are input into a deep reinforcement learning model to obtain a contour map of the medical structure.

4. The medical image segmentation method according to claim 1, characterized in that: After the step of segmenting the medical image according to the contour area and the segmentation starting point, the following step is further included: The segmented medical image is processed in three-dimensional space by Kalman filtering to obtain a three-dimensional image.

5. The medical image segmentation method according to claim 1, characterized in that: The step of acquiring medical image data containing the target medical structure comprises: Acquire T2W images and ADC images of target medical structures; Performing standardization processing on the T2W image and the ADC image; wherein the standardization processing includes at least one of normalization processing, resolution unification, and gray value unification; Medical image data containing the target medical structure is output according to the standardized processing result.

6. The medical image segmentation method according to claim 2, characterized in that: The step of extracting features of the medical image data comprises: Acquire a T2W image, an ADC image, and an image obtained by Laplace transforming the T2W image corresponding to the medical image data; Performing three-dimensional stitching on the T2W image, the ADC image, and an image obtained by Laplace transform of the T2W image; The 3D stitching result is input into the convolutional neural network or the Transformer neural network and the features are output.

7. A medical image segmentation system, characterized in that: The medical image segmentation system comprises: An acquisition module, configured to acquire medical image data containing a target medical structure; the target medical structure includes a target tissue and / or a target organ; An output module, used for inputting the medical image data into a backbone network to obtain a contour area and a segmentation starting point of the target medical structure; The segmentation module is used to segment the medical image according to the contour area and the segmentation starting point.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and used to run on the processor, characterized in that: When the processor executes the computer program, the medical image segmentation method according to any one of claims 1 to 6 is implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the medical image segmentation method according to any one of claims 1 to 6 is implemented.

Images