A method, device, equipment and storage medium for organ image segmentation

Abstract

The present invention provides a method, device, equipment and storage medium for organ image segmentation, which belongs to the field of computing technology and can improve the accuracy of organ segmentation, especially small organs. Specifically, by referring to the shape of large organs to help small organs that are difficult to segment achieve accurate segmentation, the current organ segmentation method is improved, so that small organs can also obtain good segmentation effects even when the shape is uncertain, the number of voxels is small, and the boundaries are blurred.

Description

Technical Field

[0001] The present disclosure relates to the field of computing technology, and in particular to a method, apparatus, device and storage medium for organ image segmentation. Background Art

[0002] When doctors perform abdominal organ surgery, they need to determine the direction, position and shape of the organs in order to accurately locate the surgery. Computed tomography (CT) is an important means for doctors to obtain information about the patient's internal body. Experienced doctors can imagine the specific location of the organs based on CT images. The shapes and appearances of different small organs of different individuals vary greatly, which makes it difficult to locate small organs and difficult for computers to accurately identify them. In addition, due to the rich soft tissue in the human body, the boundary contrast of some organs is low, especially around small organs, which are more difficult to identify.

[0003] Existing techniques have difficulty demarcating boundaries, which results in poor segmentation of small organs. Therefore, the segmentation of small abdominal organs remains a challenging task. Summary of the invention

[0004] The purpose of the present disclosure is to provide a method, device, equipment and storage medium for organ image segmentation to improve the accuracy of organ segmentation in order to solve the problems existing in the prior art.

[0005] According to one aspect of the present disclosure, a method for organ image segmentation is proposed.

[0006] Get the original image data set, preprocess the original image data set,

[0007] Based on the data set, a first segmentation model is obtained by training using a deep learning method, wherein the first segmentation model is used for coarse segmentation of multiple organs.

[0008] Acquire a rough segmentation result of multiple organs based on the first segmentation model,

[0009] According to the region with more voxels of small organs in the rough segmentation result, one or more first small blocks are obtained on the image with the original resolution, and the position and size information of the first small blocks are saved,

[0010] The shape features of the small organ and the organs around it in the first small block are calculated according to the rough segmentation result, and the shape features are represented by an active appearance model.

[0011] The second segmentation model is obtained by training the shape feature changes of the first small block and the surrounding organs, and the second segmentation model is used for fine segmentation of the small organs.

[0012] Based on the second segmentation model, the small organ is finely segmented to obtain one or more second small blocks,

[0013] Based on the position and size of the first small block, the first small block is replaced with the second small block.

[0014] In some embodiments, the organ image is specifically an organ CT image.

[0015] In some embodiments, preprocessing includes randomly dividing the original image data set into a training set and a validation set, then scaling the original image to a uniform size to fit the network model, dividing the range of HU values ​​according to the window position and window width of the organ in the CT image, setting the HU value above the preset range as the upper limit, setting the HU value below the preset range as the lower limit, and normalizing the HU value.

[0016] In some embodiments, the first segmentation model and the second segmentation model are trained based on 3D-UNet.

[0017] In some embodiments, the method of obtaining a rough segmentation result of multiple organs based on the first segmentation model is specifically to first use the preprocessed data to verify the model, and then use the maximum connected labeling method to post-process the segmentation result, that is, first mark all connected areas of each organ, retain only the maximum connected area of ​​each organ, and then enlarge it to the size of the original image, so that the label and the original image correspond to each other, and obtain the rough segmentation result of multiple organs.

[0018] In some embodiments, the shape features of the small organs and the organs around them in the first small block are calculated based on the rough segmentation results. Specifically, a set of points defining the correspondence are found in each image of the first small block that has been cut out. A statistical shape model can be generated by aligning the control point set on each image and applying principal component analysis. By warping each grayscale image into an average reference image, principal component analysis can be applied to the generated texture to generate a statistical texture model. Then, each normalized intensity image is warped into a model reference frame using known correspondences, and an active appearance model is generated through training.

[0019] In some embodiments, the fine segmentation of small organs based on the second segmentation model to obtain one or more second small blocks is specifically performed by matching the active appearance model to estimate the correspondence with the model frame; using the generated correspondence to warp the normalized intensity into the model frame; using voxel probability estimation to calculate the probability image in the reference frame; using the correspondence to warp it into the image frame; and selecting the label to which each voxel belongs based on the probability value, thereby obtaining an image segmentation result.

[0020] According to another aspect of the present disclosure, a device for organ image segmentation is provided, comprising:

[0021] The preprocessing unit is used to obtain the original image data set and preprocess the original image data set.

[0022] A first segmentation model training unit is used to obtain a first segmentation model by training based on the data set using a deep learning method, wherein the first segmentation model is used for coarse segmentation of multiple organs.

[0023] an image rough segmentation unit, configured to obtain a rough segmentation result of multiple organs based on the first segmentation model,

[0024] A recording unit is used to divide the image of the original resolution into one or more first small blocks according to the region with more voxels of small organs in the rough segmentation result, and save the position and size information of the first small blocks,

[0025] The second segmentation model training unit is used to calculate the shape features of the small organ and the organs around it in the first small block according to the rough segmentation result, the shape features are characterized by the active appearance model, and the second segmentation model is obtained by training using the shape feature changes of the first small block and the surrounding organs, and the second segmentation model is used for fine segmentation of the small organ.

[0026] The image fine segmentation unit is used to finely segment the small organs based on the second segmentation model to obtain one or more second small blocks, and replace the first small block with the second small block based on the position and size of the first small block.

[0027] According to another aspect of the present disclosure, an electronic device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein:

[0028] When the processor executes the computer program, the steps of the above method are implemented.

[0029] According to another aspect of the present disclosure, a computer-readable storage medium for storing processor-executable instruction data is provided. When the processor-executable instruction data stored in the computer-readable storage medium is executed, it can cause the processor to implement the organ image segmentation method as described above.

[0030] The present disclosure provides a method, device, equipment and storage medium for organ image segmentation, which can improve the accuracy of organ segmentation, especially small organs. Specifically, by referring to the shape of large organs to help small organs that are difficult to segment achieve accurate segmentation, the current organ segmentation method is improved, so that small organs can also achieve good segmentation effects even when the shape is uncertain, the number of voxels is small, and the boundaries are blurred. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The technical solutions and other beneficial effects of the present invention will be made apparent by describing in detail the specific embodiments of the present invention in conjunction with the accompanying drawings.

[0032] Figure 1 A schematic flow chart of the organ image segmentation method provided in an embodiment of the present disclosure.

[0033] Figure 2 A schematic diagram of the gallbladder CT image segmentation process provided in an embodiment of the present disclosure.

[0034] Figure 3 A schematic diagram of an organ image segmentation device provided in an embodiment of the present disclosure.

[0035] Figure 4 A schematic diagram of the structure of an electronic device provided in an embodiment of the present disclosure. DETAILED DESCRIPTION

[0036] The technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the accompanying drawings. The terms "first", "second", etc. (if any) in the specification and claims of the present disclosure and the accompanying drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the objects described in this way can be interchanged where appropriate. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. In the description of the present disclosure, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined. For those of ordinary skill in the art, the specific meanings of the above terms in the present disclosure can be understood according to the specific circumstances.

[0037] Embodiment 1:

[0038] Specifically, see Figure 1 , is an organ image segmentation method provided by the present disclosure. The specific steps are as follows:

[0039] S1 obtains an original image data set and preprocesses the original image data set.

[0040] In one or more embodiments, the original image may be a CT image of an abdominal organ, and the process of preprocessing the original CT image data is specifically as follows: randomly dividing the data set into a training set and a validation set, and then scaling the image to a uniform size to accommodate the network model, and then dividing the range of HU values ​​(the unit of CT value) according to the window position and window width of the abdominal organs in the CT image, setting all values ​​above the range as the upper limit of the HU value, and setting all values ​​below the range as the lower limit of the HU value, so as to better contrast the abdominal organs, and performing normalization so that the data values ​​are [0,1].

[0041] S2 uses a deep learning method to train a first segmentation model based on the data set, and the first segmentation model is used for coarse segmentation of multiple organs.

[0042] In one or more embodiments, the process of training a multi-organ segmentation network is specifically as follows: first load the original data list, then read the pre-trained data in sequence, and then use data enhancement to increase the amount of data. The data enhancement process is to first randomly take samples for rotation according to the patient's body amplitude, and then randomly sample and scale them according to height, fatness and thinness. Next, use the 3D-UNet segmentation network and use the divided training set to input the model for training. The network model has multiple downsampling and corresponding upsampling, each stage of which contains two 3×3×3 convolutional layers, each convolution is followed by a batch normalization function and a Relu activation function. Then there is a 2×2×2 maximum pooling layer with a stride of 2 in each dimension. The encoder and decoder use jump connections at the same stage. At the end of the decoder, the softmax function is used to output the segmentation predictions of all organs. The loss function uses a multi-classification dice loss function, defining X as the Ground Truth segmentation image and Y as the predicted segmentation image. Define N as the task category, then X n Y represents the segmented image of the nth category. The pixel value in the image is 1 for organ area and 0 for non-organ area. n Represents the predicted segmented image of the nth category. The value range of the pixels in the image is [0,1], which indicates the probability that the current pixel is an organ. The formula is as follows:

[0043]

[0044] S3 obtains a rough segmentation result of multiple organs based on the first segmentation model.

[0045] In one or more embodiments, after obtaining the rough segmentation network model, a post-processing process is performed to obtain the rough segmentation result: first, the model is verified using the pre-processed data, and then the segmentation result is post-processed using the maximum connected labeling method, that is, all connected areas of each organ are first marked, and only the maximum connected area of ​​each organ is retained. Then, it is enlarged to the size of the original image so that the label and the original image correspond, and the rough segmentation result of the abdominal multiple organs is obtained.

[0046] S4: According to the region with more voxels of small organs in the rough segmentation result, one or more first small blocks are obtained on the image with the original resolution, and the position and size information of the first small blocks are saved.

[0047] In one or more embodiments, according to the area with more small organ voxels in the rough segmentation result, small blocks, i.e., the first small blocks, are divided on the image with the original resolution, and the position and size information of the small blocks are saved in the file. The purpose of this is to make the network pay more attention to the small organ area, while saving memory consumption and improving segmentation efficiency.

[0048] S5 calculates shape features of the small organ and its surrounding organs in the first small block according to the rough segmentation result, and the shape features are represented by an active appearance model.

[0049] In one or more embodiments, the shape features of the small organ and other surrounding organs are calculated to help the boundary of the small organ to be more clearly delineated. The specific process is as follows:

[0050] First, the shape information of the organ is represented using an active appearance model. The active appearance model (AAM) is an image segmentation algorithm based on the active appearance model. In order to build a statistical model of shape and appearance, it is necessary to find a set of points in each image that define the corresponding relationship.

[0051] In the model construction phase, a statistical shape model is generated by aligning the set of control points on each image and applying principal component analysis (PCA). PCA can be applied to the generated texture by warping each grayscale image to the average reference image to generate a statistical texture model.

[0052] The shape and texture models can be combined to construct a combined statistical model of appearance, as follows:

[0053]

[0054]

[0055] Where x refers to the contour of the model, refers to the average contour of the graph, g refers to the texture of the model, refers to the average texture of the graph, and Q x and Q g represents the matrix composed of point covariance eigenvectors describing the image in the training set, and c refers to the parameters that control the shape and texture of the object. According to the parameters c of the active appearance model, a synthetic image of the modeled object can be generated, and the difference r between the synthetic image (that is, the model frame or reference frame) and the new image can be calculated. By minimizing the difference r, a fast update algorithm is derived, which can match the model with the new image in multiple iterations to obtain the correspondence between the model and the image.

[0056] S6: using the shape feature changes of the first small block and surrounding organs to train and obtain a second segmentation model, wherein the second segmentation model is used for fine segmentation of the small organ.

[0057] In one or more embodiments, during the training phase, each normalized intensity image is warped into a model reference frame using the aforementioned known correspondence. The process of warping into the model reference frame is similar to the process of obtaining the correspondence between the model and the image. The active appearance model (i.e., the model reference frame) can be matched to a new image through an algorithm, thereby obtaining a deformation of a new image converted into an active appearance model. The model reference frame refers to the AAM active appearance model, which can be expressed as the aforementioned formula (2) (3), and is a statistical model framework of shape and appearance that characterizes the shape of an organ and is constructed using information in small image blocks.

[0058] For any given voxel in this frame, there is a set of probabilities p i , i = 1..n. Probability p i represents the probability that voxel i in each image is an organ. It is located inside the object and the corresponding intensity value vector k sampled in the area around the voxel i , where k i represents the CT intensity of voxel i in each image. Then, linear regression is performed to learn a function to estimate the probability of a given intensity pattern, which is repeated for each voxel near the boundary. The voxel estimator is calculated as follows:

[0059] p=f(k)=a*k+d (4)

[0060] Here, formula (4) represents a linear relationship, which is used to calculate the probability p of the point being an organ based on the CT intensity value k, thereby obtaining a probability image. p, k takes values ​​of [0,1], p = 0 means outside the organ, and p = 1 means inside the organ. Parameters a and d are two parameters that can be learned through linear regression.

[0061] Therein, voxels far from the boundary are assumed to have p=0 (outside) or p=1 (inside).

[0062] The first small block containing the small organs and the shape features of the surrounding organs are simultaneously input into the deep learning network model to train a new model that can accurately segment the small organs, namely the second segmentation model.

[0063] S7: finely segment the small organ based on the second segmentation model to obtain one or more second small blocks.

[0064] In one or more embodiments, the specific process is as follows:

[0065] First load the list of raw data, then read the first small block of the small organ in the corresponding CT image data according to the small block information saved in the file, and then we can segment it using the following steps:

[0066] 1). Use the AAM algorithm to match the correspondence between the new image and the model framework (i.e., as mentioned above, by minimizing the difference between the image and the model, an update algorithm is derived, which can match the model and the new image in iteration);

[0067] 2) Using this correspondence, the new image (i.e., the normalized intensity image to be segmented) is deformed into the AAM model frame (i.e., the active appearance model frame);

[0068] 3). Probability estimation is performed through formula (4), and the probability image of the model frame is calculated (i.e., a function is learned through linear regression to estimate the probability corresponding to the image intensity value); the probability p of the point being an organ is calculated through the CT value of the image, thereby obtaining a probability image. The probability image here is a matrix, and each point is a probability value between (0, 1), indicating the possibility of being an organ.

[0069] 4). Using this correspondence, the calculated probability image is deformed to the image frame (i.e., the probability image is distorted to the image frame, and the deformation process adopts cubic linear interpolation);

[0070] 5). The generated probability image is divided by a threshold of 0.5 (a probability image is a matrix, each point is a probability value between (0, 1), indicating the possibility of an organ. If the probability value is greater than 0.5, the probability is considered to be 1, and if it is less than 0.5, the probability is considered to be 0), and a label image is obtained, which is the segmentation result of small organs. The probability value here is a new probability image obtained after interpolation and distortion of the probability image calculated by the previous linear regression. Each voxel of the image is represented by a probability value. If the probability value is greater than or equal to 0.5, it indicates that it is the label of the organ.

[0071] S7: Based on the position and size of the first small block, the first small block is replaced by the second small block.

[0072] According to the position and size of the first small piece saved, the segmentation results of the small organs are spliced ​​onto the rough segmentation of multiple organs, replacing the original poorly segmented small organ results, thereby forming a new result for the precise segmentation of small organs.

[0073] At this point, the entire anatomically guided segmentation scheme for abdominal small organ CT images is completed.

[0074] Embodiment 2

[0075] In order to better illustrate the inventive concept of the present application, this embodiment takes the segmentation of the gallbladder organ as an example to describe the process of organ segmentation.

[0076] Among the abdominal organs, the gallbladder is a pear-shaped sac structure located behind the liver (in the gallbladder fossa of the liver), which is connected to the liver, making it difficult to divide the boundaries. Moreover, the shapes of gallbladders in different patients vary greatly. When the gallbladder of the same patient contracts and empties, the folds are tall and branched; when the gallbladder is filled, the folds are reduced and become shorter; all of these lead to poor gallbladder segmentation at present. This embodiment selects the gallbladder organ, which is more difficult to segment among the small abdominal organs, and describes in detail the implementation process of the segmentation scheme of the CT image of the small abdominal organ; the small organs here refer to organs with small size and more soft tissue, unclear boundaries, and the classic medical image segmentation algorithm cannot obtain the same accurate segmentation results as large organs. Small organs in the abdomen include gallbladder, adrenal glands, pancreas, portal vein, etc. In public data sets, the segmentation accuracy of these small organs is less than 90%, or even lower, such as the adrenal glands are only about 70%. However, other organs such as liver, kidneys, spleen, stomach and other organs have reached an accuracy of more than 95%.

[0077] Figure 2 This is a flow chart of the gallbladder CT image segmentation solution based on anatomical guidance provided in this embodiment. This embodiment is based on the BTCV challenge dataset and can be used to solve the problem of inaccurate gallbladder segmentation in abdominal multi-organ segmentation. The fine segmentation model for the gallbladder can be completed with less computing resources. Specifically, it includes the following steps:

[0078] (1) Preprocessing the original data. The specific process of preprocessing the original CT image data is as follows: remove 2 cases without gallbladder, take 28 complete labeled data, randomly select training set and validation set according to 4:1, and then scale the image from n×512×512 to 128×256×256. The threshold range with more voxels is selected by calculation, that is, the HU value of the CT image is kept in the range of [-350,350], the part outside the range is set to 350, and the part below the range is set to -350, so as to better contrast the abdominal organs, and perform normalization processing so that the data value is [0,1];

[0079] (2) The medical image segmentation method 3D-UNet was used to pre-train the abdominal multi-organ network model. The model parameters were selected by adjusting the parameters to perform four downsampling and upsampling. The training was performed for 1000 epochs with an initial learning rate of 1e-4. The learning rate began to decay after 900 epochs.

[0080] (3) Obtain a rough segmentation result and use post-processing to reduce the discrete area. From the results, it can be seen that organs with high density and stable shape, such as the liver and spleen, have achieved good segmentation effects, while the effect of the gallbladder is poor and there is room for improvement.

[0081] (4) According to the rough segmentation results, a small block with more gallbladder voxels is taken, and the size is selected to be an appropriate size of 112*112*112 that can wrap the gallbladder, and the center point and size of the small block are saved in the file.

[0082] (5) Based on the rough segmentation results, the shape features of the gallbladder and its surrounding organs in the small block are constructed and trained. Here, through anatomical prior knowledge, it can be found that the small block containing the gallbladder also contains organs such as the liver, spleen, and pancreas. For these large organs around the gallbladder, since the pre-trained multi-organ segmentation network has achieved good segmentation results, it can be used to guide the neural network to strengthen the learning of the boundary and shape of the gallbladder.

[0083] Here, in order to build a statistical model of the shape and appearance of the gallbladder and surrounding organs, it is necessary to find a set of points that define correspondences in each of the cut-out images. By aligning the set of control points on each image and applying principal component analysis, a statistical shape model can be generated. PCA can be applied to the generated texture to generate a statistical texture model by warping each grayscale image into an average reference image. Each normalized intensity image is then warped into the model reference frame using known correspondences to generate an active appearance model through training.

[0084] (6) Use the shape feature changes of the gallbladder and surrounding organs to train the gallbladder segmentation network model. The specific operations are as follows:

[0085] A global search is first performed using the AAMs of all structures; a local search is then performed using the AAM of each individual structure to optimize the match; and finally segmentation is performed by combining the generated probability images into a single labeled image.

[0086] As the input image size becomes smaller, the resolution is greatly improved, and the learning efficiency of the model is improved. When the model converges, it means that the network has learned the voxel-based semantic segmentation results of the gallbladder and the shape feature information of each organ in the small block.

[0087] (7) The test set is verified according to the model in step (6) to obtain the segmentation result of the gallbladder, and then the gallbladder labels that are not well segmented in the multi-organ segmentation result are replaced to generate the multi-organ segmentation result corresponding to the original image. It can be seen from the visualization effect and evaluation indicators that the segmentation effect of the gallbladder has been significantly improved.

[0088] Embodiment three:

[0089] To achieve the above purpose, this embodiment provides an organ image segmentation device 300, see Figure 3Specifically, the device 300 includes: a preprocessing unit 301 , a first segmentation model training unit 302 , an image coarse segmentation unit 303 , a recording unit 304 , a second segmentation model training unit 305 , and an image fine segmentation unit 306 .

[0090] The preprocessing unit 301 is used to obtain the original image data set and preprocess the original image data set.

[0091] The first segmentation model training unit 302 is used to obtain a first segmentation model by training based on the data set using a deep learning method, wherein the first segmentation model is used for coarse segmentation of multiple organs.

[0092] The image rough segmentation unit 303 is used to obtain a rough segmentation result of multiple organs based on the first segmentation model.

[0093] The recording unit 304 is used to divide the image of the original resolution into one or more first small blocks according to the region with more voxels of small organs in the rough segmentation result, and save the position and size information of the first small blocks.

[0094] The second segmentation model training unit 305 is used to calculate the shape features of the small organ and the organs around it in the first small block according to the rough segmentation result. The shape features are characterized by the active appearance model. The second segmentation model is obtained by training the shape feature changes of the first small block and the surrounding organs. The second segmentation model is used for fine segmentation of the small organ.

[0095] The image fine segmentation unit 306 is configured to finely segment the small organs based on the second segmentation model to obtain one or more second small blocks, and replace the first small blocks with the second small blocks based on the positions and sizes of the first small blocks.

[0096] Embodiment 4:

[0097] Accordingly, an embodiment of the present application further provides an electronic device, which may be a terminal or a server. Figure 4 As shown, Figure 4 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application.

[0098] The electronic device 400 includes a processor 401 having one or more processing cores, a memory 402 having one or more computer-readable storage media, and a computer program stored in the memory 402 and executable on the processor. The processor 401 is electrically connected to the memory 402. Those skilled in the art will appreciate that the electronic device structure shown in the figure does not limit the electronic device, and may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently.

[0099] The processor 401 is the control center of the electronic device 400. It uses various interfaces and lines to connect various parts of the entire electronic device 400, and executes various functions of the electronic device 400 and processes data by running or loading software programs (computer programs) and / or units stored in the memory 402, and calling data stored in the memory 402, thereby monitoring the electronic device 400 as a whole.

[0100] In an embodiment of the present application, the processor 401 in the electronic device 400 will load the instruction data corresponding to the process of one or more applications into the memory 402 according to the method steps of the aforementioned embodiment, and the processor 401 will run the application stored in the memory 402 to realize various functions.

[0101] The specific implementation of the above operations can be found in the previous embodiments, which will not be described in detail here.

[0102] Optional, such as Figure 4 As shown, the electronic device 400 further includes: an organ image segmentation unit 403, a communication unit 404, an input unit 405 and a power supply 406. The processor 401 is electrically connected to the organ image segmentation unit 403, the communication unit 404, the input unit 405 and the power supply 406 respectively. Those skilled in the art can understand that Figure 4 The electronic device structure shown in the figure does not constitute a limitation of the electronic device, and may include more or less components than shown in the figure, or combine certain components, or arrange the components differently.

[0103] The organ image segmentation unit 403 can be used to implement the segmentation process of the organ image. The specific implementation of the segmentation process of the organ image can be referred to the previous embodiment, which will not be repeated here.

[0104] The communication side 404 may be used to communicate with other devices.

[0105] The input unit 405 may be used to receive input numbers, character information or user feature information (such as fingerprint, iris, facial information, etc.), and generate keyboard, mouse, joystick, optical or trackball signal input related to user settings and function control.

[0106] The power supply 406 is used to supply power to various components of the electronic device 400. Optionally, the power supply 406 can be logically connected to the processor 401 through a power management system, so that the power management system can manage charging, discharging, and power consumption. The power supply 406 can also include one or more DC or AC power supplies, recharging systems, power failure detection circuits, power converters or inverters, power status indicators, and other arbitrary components.

[0107] Embodiment five:

[0108] To achieve the above objectives, this embodiment proposes a computer-readable storage medium for storing processor-executable instruction data. When the processor-executable instruction data stored in the computer-readable storage medium is executed, it can cause the processor to implement the organ image segmentation method as described above.

[0109] In the above embodiments, specific embodiments are used to illustrate the principles and implementation methods of the present disclosure. The descriptions of each embodiment have their own focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments. The descriptions of the above embodiments are only used to help understand the technical solutions and core ideas of the present disclosure; ordinary technicians in this field should understand that they can still modify the technical solutions recorded in the above embodiments, or replace some of the technical features therein with equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present disclosure.

Claims

1. A method for segmenting an organ image, characterized in that: Get the original image data set, preprocess the original image data set, Based on the data set, a first segmentation model is obtained by training using a deep learning method, wherein the first segmentation model is used for coarse segmentation of multiple organs. Acquire a rough segmentation result of multiple organs based on the first segmentation model, According to the region with more voxels of small organs in the rough segmentation result, one or more first small blocks are obtained on the image with the original resolution, and the position and size information of the first small blocks are saved, The shape features of the small organ and the organs around it in the first small block are calculated according to the rough segmentation result, and the shape features are represented by an active appearance model. The second segmentation model is obtained by training the shape feature changes of the first small block and the surrounding organs, and the second segmentation model is used for fine segmentation of the small organs. Based on the second segmentation model, the small organ is finely segmented to obtain one or more second small blocks, Based on the position and size of the first small block, replacing the first small block with the second small block; The shape features of the small organs and the organs around them in the first small block are calculated based on the rough segmentation result. Specifically, a set of points defining the corresponding relationship is found in each image of the first small block that has been cut out, and a statistical shape model is generated by aligning the control point set on each image and applying principal component analysis. The statistical texture model is generated by warping each grayscale image into an average reference image, thereby applying principal component analysis to the generated texture. Then, each normalized intensity image is warped into a model reference frame using a known corresponding relationship, and an active appearance model is generated through training.

2. The method according to claim 1, characterized in that The organ image is specifically an organ CT image.

3. The method according to claim 2, characterized in that The preprocessing includes randomly dividing the original image data set into a training set and a validation set, then scaling the original image to a uniform size to adapt to the network model, dividing the range of HU values ​​according to the window position and window width of the organ in the CT image, setting the HU value above the preset range as the upper limit, setting the HU value below the preset range as the lower limit, and normalizing the HU value.

4. The method according to claim 1, characterized in that: The first segmentation model and the second segmentation model are obtained based on 3D-UNet training.

5. The method according to claim 1, characterized in that The method of obtaining the rough segmentation results of multiple organs based on the first segmentation model is specifically as follows: first, the preprocessed data is used to verify the model, and then the segmentation results are post-processed using the maximum connected labeling method, that is, all connected areas of each organ are first marked, and only the maximum connected area of ​​each organ is retained, and then it is enlarged to the size of the original image, so that the label and the original image correspond to each other, and the rough segmentation results of multiple organs are obtained.

6. The method according to claim 1, characterized in that The fine segmentation of the small organ based on the second segmentation model to obtain one or more second small blocks is specifically, matching the active appearance model to estimate the correspondence with the model frame; using the generated correspondence to warp the normalized intensity into the model frame; using the voxel probability estimation to calculate the probability image in the reference frame; Use the correspondences that are warped into the image frame; The label to which each voxel belongs is selected based on the probability value to obtain the image segmentation result.

7. A device for organ image segmentation, characterized in that: The device comprises, The preprocessing unit is used to obtain the original image data set and preprocess the original image data set. A first segmentation model training unit is used to obtain a first segmentation model by training based on the data set using a deep learning method, wherein the first segmentation model is used for coarse segmentation of multiple organs. an image rough segmentation unit, configured to obtain a rough segmentation result of multiple organs based on the first segmentation model, A recording unit is used to divide the image of the original resolution into one or more first small blocks according to the region with more voxels of small organs in the rough segmentation result, and save the position and size information of the first small blocks, The second segmentation model training unit is used to calculate the shape features of the small organ and the organs around it in the first small block according to the rough segmentation result, the shape features are characterized by the active appearance model, and the second segmentation model is obtained by training using the shape feature changes of the first small block and the surrounding organs, and the second segmentation model is used for fine segmentation of the small organ. an image fine segmentation unit, configured to obtain one or more second small blocks by finely segmenting the small organs based on the second segmentation model, and replace the first small blocks with the second small blocks based on the positions and sizes of the first small blocks; The shape features of the small organs and the organs around them in the first small block are calculated based on the rough segmentation results. Specifically, a set of points defining the corresponding relationship is found in each image of the first small block that has been cut out, and a statistical shape model is generated by aligning the control point set on each image and applying principal component analysis. The statistical texture model is generated by warping each grayscale image into an average reference image and applying principal component analysis to the generated texture. Then, each normalized intensity image is warped into a model reference frame using a known corresponding relationship, and an active appearance model is generated through training.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium for storing processor-executable instruction data, wherein the processor-executable instruction data stored in the computer-readable storage medium can cause a processor to implement the method according to any one of claims 1 to 6 when being executed.

Images