CT left ventricle segmentation method, device, equipment and storage medium
Through the 8-layer U-Net model based on deep neural network, the computational complexity and insufficient accuracy of left ventricular segmentation in CT images are solved, efficient and accurate left ventricular segmentation is achieved, the misdiagnosis rate is reduced, and more automated and Objective diagnostic assistance.
Patent Information
- Application Number
- CN201911038966.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2019-10-29
- Publication Date
- 2024-11-29
- Estimated Expiration
- 2039-10-29
AI Technical Summary
In the existing medical CT images, left ventricular segmentation methods have problems with computational complexity and insufficient segmentation accuracy, especially due to the lack of obvious contrast between the left ventricular region and background and the influence of noise, which makes manual segmentation time-consuming and labor-intensive and dependent on the operator's experience. Increased misdiagnosis rate.
Using the CT left ventricular segmentation method based on deep neural network, the 8-layer U-Net model is used for training. The image is encoded and decoded through the combined structure of the contraction path and the expansion path, combined with the residual structure and auxiliary path, and the image is encoded and decoded using random transformation. Enhance the training data and improve the generalization ability and robustness of the model.
It achieves higher accuracy and efficiency of left ventricular segmentation, reduces the rate of misdiagnosis, improves the accuracy and consistency of segmentation results, and can assist in the diagnosis of cardiovascular diseases more automatically and objectively.
Smart Images

Figure CN112365504B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of medical image processing, and in particular to a CT left ventricle segmentation method, apparatus, device and storage medium. Background Art
[0002] Cardiovascular disease seriously threatens human health and is characterized by high morbidity, high disability and high mortality. Early diagnosis and treatment can help improve heart function and patient prognosis and reduce mortality.
[0003] Cardiac CT Angiography (CTA) has become an important method for diagnosing cardiovascular diseases in the clinic due to its non-invasiveness and simple operation. Accurate segmentation of the left ventricle plays a vital role in the evaluation of cardiac function. Accurate segmentation of the left ventricle (LV) is a key step in quantifying clinical indicators such as ventricular volume, ejection fraction, left ventricular mass, wall thickening, and wall motion abnormalities to quantitatively analyze the overall and local functions of the heart, and then diagnose and treat cardiovascular diseases.
[0004] The left ventricle segmentation is difficult due to the unclear contrast between the left ventricle and the background area in the cardiac CT image and the presence of noise. Manual segmentation of the left ventricle is time-consuming and laborious, and the accuracy of the segmentation depends on the operator's experience and is easily affected by the operator's subjectivity, increasing the misdiagnosis rate. The computer-aided diagnosis (CAD) system can provide a more accurate, objective, automatic and efficient auxiliary diagnosis method. The segmentation of the left ventricle is an important step in designing a medical CT image CAD system for diagnosing cardiovascular diseases. Therefore, the research on the segmentation method of the left ventricle has important scientific significance and application value.
[0005] Traditional medical image segmentation methods, such as segmentation methods based on active contour models, have problems such as complex calculations and insufficient segmentation accuracy. Summary of the invention
[0006] Based on this, it is necessary to provide a segmentation method with better accuracy and higher efficiency for the problem of left ventricle segmentation.
[0007] A CT left ventricle segmentation method based on a deep neural network, comprising:
[0008] Acquire CT cardiac images;
[0009] The CT heart image is input into a deep neural network model to obtain a CT left ventricle segmentation result, and the deep neural network model is trained based on an 8-layer U-Net model.
[0010] Several optional methods are also provided below, but they are not intended to be additional limitations on the above-mentioned overall solution, but are merely further supplements or preferences. Under the premise that there are no technical or logical contradictions, each optional method can be combined with the above-mentioned overall solution separately, and multiple optional methods can also be combined.
[0011] Optionally, the 8-layer U-Net model includes a contraction path and an expansion path:
[0012] The contraction path includes 9 coding blocks, each coding block contains two 3×3 convolutional layers, each convolutional layer is followed by a batch normalization layer and a ReLU activation layer, and for the first 8 coding blocks, each coding block is followed by a 2×2 maximum pooling layer for downsampling;
[0013] The expansion path includes 8 decoding blocks, each of which contains two 3×3 convolutional layers, each of which is followed by a batch normalization layer and a ReLU activation layer, and each decoding block is preceded by a 2×2 deconvolutional layer for upsampling;
[0014] The feature maps of corresponding blocks in the contraction path and the expansion path are jump-connected according to the channel dimension.
[0015] Optionally, a residual structure is incorporated into each encoding block and decoding block of the 8-layer U-Net model; in the expansion path, an auxiliary path is added to the first 7 upsampling operations.
[0016] Optionally, the training image set of the 8-layer U-Net model is obtained as follows:
[0017] A number of CT heart images are acquired to form an initial image set;
[0018] During each iteration of the training process, each CT cardiac image in the initial image set is transformed N times to obtain N augmented images, N ≥ 8, and each transformation includes random rotation, random translation, random shearing and random scaling in sequence;
[0019] The initial image set and the augmented images together constitute the training image set.
[0020] Optionally, the angle range of the random rotation is -5° to 5°, the random translation includes horizontal random translation and vertical random translation performed simultaneously or sequentially, the pixel range of the horizontal random translation is -0.02×image width to 0.02×image width, the pixel range of the vertical random translation is -0.02×image height to 0.02×image height, the angle range of the random shearing is -5° to 5°, and the ratio range of the random scaling is 0.90 to 1.10.
[0021] The present application also provides a CT left ventricle segmentation device based on a deep neural network, comprising:
[0022] The first module is used to obtain CT cardiac images;
[0023] The second module is used to input the CT heart image into a deep neural network model to obtain a CT left ventricle segmentation result, and the deep neural network model is trained based on an 8-layer U-Net model.
[0024] The present application also provides a CT left ventricle segmentation device based on a deep neural network, comprising a computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor. When the computer processor executes the computer program, the CT left ventricle segmentation method based on a deep neural network is implemented.
[0025] The present application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the deep neural network-based CT left ventricle segmentation method when executing the computer program.
[0026] The present application also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a computer processor, it implements the CT left ventricle segmentation method based on deep neural network.
[0027] The CT left ventricle segmentation method based on deep neural network provided in this application has better accuracy and higher efficiency. BRIEF DESCRIPTION OF THE DRAWINGS
[0028] Figure 1 A schematic diagram of the method flow of the deep neural network model in this application;
[0029] Figure 2 This is a schematic diagram of the structure of the 8-layer U-Net model;
[0030] Figure 3 This is a schematic diagram of the left ventricle obtained by segmentation using the method provided in this application and the prototype U-Net method; wherein:
[0031] (a)-(e) are the segmentation results of 5 images in the validation set;
[0032] (f)-(g) are the segmentation results of 2 images in the test set;
[0033] Figure 4 Schematic diagram of the target area (horizontal ellipse enclosed area) obtained by Graph Cuts annotation and the target area (vertical ellipse enclosed area) obtained by segmentation using the method of the present application;
[0034] Figure 5 FIG. 1 is a schematic diagram of a computer device in one embodiment. DETAILED DESCRIPTION
[0035] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.
[0036] In order to better describe and illustrate the embodiments of the present application, reference may be made to one or more drawings, but the additional details or examples used to describe the drawings should not be considered as limiting the scope of the invention of the present application, any of the currently described embodiments or preferred methods.
[0037] It should be noted that when a component is referred to as being "connected" to another component, it may be directly connected to the other component or there may be a central component. When a component is referred to as being "disposed on" another component, it may be directly disposed on the other component or there may be a central component at the same time.
[0038] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein in the specification of this application are only for the purpose of describing specific embodiments and are not intended to limit this application.
[0039] A CT left ventricle segmentation method based on a deep neural network, comprising:
[0040] Acquire CT cardiac images;
[0041] The CT heart image is input into the deep neural network model to obtain the CT left ventricle segmentation result. The deep neural network model is trained based on the 8-layer U-Net model.
[0042] This application uses an 8-layer U-Net model to train a deep neural network model, which can extract deeper abstract information and improve the accuracy of left ventricular segmentation results.
[0043] The training image set of the 8-layer U-Net model is obtained as follows:
[0044] A number of CT heart images are acquired to form an initial image set;
[0045] During each iteration of the training process, each CT cardiac image in the initial image set is transformed N times to obtain N augmented images, N ≥ 8, and each transformation includes random rotation, random translation, random shearing and random scaling in sequence;
[0046] The initial image set and the augmented images together constitute the training image set.
[0047] In the initial image set, the CT cardiac images are DICOM image data. The window width and window position of each CT cardiac image in the initial image set are uniformly adjusted, and the adjusted CT cardiac images are annotated using the Graph Cuts method (this application uses the GraphCuts method for annotation, making the annotated data more stable and reliable), and the initial image set and the corresponding annotation results are divided into a training set, a validation set, and a test set.
[0048] In one embodiment, the steps of acquiring the initial image set are as follows:
[0049] (1) CT left ventricle DICOM image data of 101 patients were collected. The image size was 512 × 512 pixels. 20 images were taken from 100 cases at intervals. The 20 images included 2 images without the left ventricle appearing and 2 images after the left ventricle disappeared. In addition, 280 continuous images of the remaining case were taken, for a total of 2280 images.
[0050] (2) The window width of 2280 images was uniformly adjusted to 600 HU (Hounsfield Units) and the window level to 100 HU;
[0051] (3) Convert the adjusted DICOM image data into JPG image data with a grayscale range of 0 to 255 and annotate it using the Graph Cuts method.
[0052] (4) 70 cases in the annotated results were used as the training set (1400 images), 30 cases were used as the validation set (600 images), and the remaining case was used as the test set (280 images).
[0053] When training the 8-layer U-Net model, in order to improve the generalization ability and robustness of the neural network model, the data of the initial image set is online amplified for different iteration batches, that is, a different amplified training image set is used for training in each iteration.
[0054] The amplified image is obtained by performing random transformation operations on each image in the initial image set, wherein the angle range of random rotation is -5° to 5° (positive and negative represent different rotation directions, one of which is clockwise and the other is counterclockwise), the random translation includes horizontal random translation and vertical random translation performed simultaneously or sequentially, the pixel range of horizontal random translation is -0.02×image width (in pixels) to 0.02×image width, the pixel range of vertical random translation is -0.02×image height (in pixels) to 0.02×image height (positive and negative represent opposite directions of translation, preferably random translation in horizontal and vertical directions is performed simultaneously), the angle range of random shearing is -5° to 5° (positive and negative represent opposite shearing directions), and the ratio range of random scaling is 0.90 to 1.10 (smaller than 1 is enlarged, and larger than 1 is reduced).
[0055] In one embodiment, 1400 initial training set images are augmented online, and each CT cardiac image is transformed 8 times to obtain 8 augmented images. Each transformation includes random rotation, random translation, random shearing and random scaling in sequence, and finally a training set image consisting of 12600 images is obtained.
[0056] The 8-layer U-Net model is trained using the training set images to extract neural network features and generate candidate deep neural network models. The candidate deep neural network models are verified and evaluated using the validation set. The model with the best performance is used as the final deep neural network model. The final deep neural network model is tested using the test set to evaluate its generalization and robustness.
[0057] like Figure 2 As shown, the 8-layer U-Net model includes a contraction path and an expansion path:
[0058] The contraction path includes 9 encoding blocks, each of which contains two 3×3 convolutional layers (with a step size of 1 and a zero padding of 1). A batch normalization layer (i.e., BN layer, which accelerates network convergence) and a ReLU activation layer are added after each convolutional layer. For the first 8 encoding blocks, a 2×2 maximum pooling layer is used for downsampling (with a step size of 2 and no zero padding) after each encoding block to halve the size of the feature map, and the first 3×3 convolutional layer of the next encoding block doubles the number of feature channels.
[0059] The expansion path includes 8 decoding blocks, each of which contains two 3×3 convolutional layers. A BN layer and a ReLU activation layer are also added after each convolutional layer. A 2×2 deconvolution layer (with a step size of 2 and no zero padding) is used before each decoding block to upsample the feature map size to double the size of the feature map, and the first 3×3 convolutional layer of the next decoding block is used to halve the number of feature channels.
[0060] The feature maps of corresponding blocks in the contraction path and the expansion path are jump-connected according to the channel dimension.
[0061] In the last decoding block, a 1×1 convolutional layer (with stride 1 and no zero padding) is used to map the feature vector to a one-dimensional space.
[0062] The contraction path is used to encode the input image and obtain context information, and the expansion path is used to decode and repair image details and spatial dimensions, generate pixel-level label output, and achieve accurate positioning. In the positioning operation, a jump connection in the form of a stacking operation is adopted to connect the shallow features of each layer in the contraction path with the deep features of the corresponding layer in the expansion path obtained by upsampling operation. Subsequently, the continuous convolutional layers learn based on this information to alleviate the gradient diffusion problem caused by the increase in the number of network layers.
[0063] A residual structure is incorporated into each encoding block and decoding block of the 8-layer U-Net model, adding the input of the first 3×3 convolutional layer to the output of the second 3×3 convolutional layer to better optimize the deepened network.
[0064] The number of feature channels of the 9 encoding blocks is 8, 16, 32, 64, 128, 256, 512, 1024, and 2048. The entire encoder part contains 27 convolutional layers.
[0065] In the expansion path, the first 7 upsampling operations are added with auxiliary paths ( Figure 2 In the figure, aux1 to aux7 are 7 auxiliary paths. Deep supervision is performed through auxiliary paths, so that deep feature maps can also be supervised to better restore details and improve segmentation accuracy. The current feature map is restored to the same size as the original image after multiple upsampling operations and 1×1 convolution layers, and the feature vector is mapped to a one-dimensional space. The loss values output by these 7 auxiliary paths and the loss value output by the last decoding block are weighted and summed from deep to shallow by 0.02, 0.04, 0.06, 0.08, 0.1, 0.2, 0.5, and 1 respectively. The entire decoder part contains a total of 68 convolution layers.
[0066] In one embodiment, if Figure 1 As shown in the figure, the training process of the 8-layer U-Net model includes:
[0067] Step 1: When training the deep neural network model, use MSRA initialization to initialize the weights of the neural network, read 1400 training sets and 600 test sets, and perform zero mean and unit variance normalization operations on them. Set the learning rate learning rate = 2e -4, batch size = 64 (4-GPU data parallelism), iteration batch epochs = 100. Note that the current iteration batch epoch = 0, and go to step 2;
[0068] Step 2: Perform 8-fold online data amplification on the 1,400 initial training set images to obtain a training set of 12,600 images;
[0069] Step 3: randomly divide the 12,600 training set images into 197 batches. Except for the last batch which contains only 56 images, the remaining batches contain 64 images. Set the current batch to 0 and go to step 4.
[0070] Step 4, input the current batch into the 8-U-Net network model;
[0071] Step 5: Use the binary cross entropy (BCE) loss function to estimate the inconsistency between the predicted value and the true value of the 8-U-Net network model, and use the back propagation (BP) algorithm to update the network parameters layer by layer through the Adam optimization function. Update batch = batch + 1. If batch < 197, go to step 4, otherwise go to step 6;
[0072] Step 6: Generate an alternative deep neural network model, input 600 validation sets for validation evaluation, and if the evaluation result is better, save the current candidate deep neural network model. Update epoch = epoch + 1, if epoch < 100, go to step 2, otherwise go to step 7;
[0073] Step 7: After the training is completed, the best candidate deep neural network model is saved as the deep neural network model for segmenting the CT left ventricle; it is tested using the test set to evaluate its generalization and robustness.
[0074] Figure 3 The diagram of the left ventricle obtained by segmentation using the method provided in this application and the prototype U-Net method; (a)-(e) are the segmentation results of 5 images in the validation set; (f)-(g) are the segmentation results of 2 images in the test set. The consistency between the segmentation results and the comparison standard (GT) is evaluated using four indicators: Dice similarity coefficient (DSC), precision (Precision) and sensitivity (Sensitivity). Figure 4 As shown, Ω Seg Represents the target area obtained by segmentation, which is equal to TP+FP, Ω GT represents the standard target area, which is equal to TP+FN. TP represents Ω Seg The number of pixels correctly segmented into the positive class, TN represents non-Ω SegThe number of pixels correctly segmented as negative class in , FP represents Ω Seg The number of pixels that are incorrectly segmented as positive, FN represents non-Ω Seg The number of pixels that are incorrectly segmented as negative class.
[0075] DSC (Formula (I)) measures the spatial overlap between the segmented target area and the standard target area, Precision (Formula (II)) measures the proportion of positive pixels in the segmented target area, and Sensitivity (Formula (III)) measures the proportion of correctly segmented pixels in the standard target area. The evaluation values of the three indicators are all between 0 and 1, and the larger the value, the better the Ω Seg With Ω GT The higher the consistency between them, the better the segmentation result.
[0076] DSC=2·(Ω Seg ∩Ω GT ) / (Ω Seg +Ω GT )=2·TP / (2·TP+FP+FN) Formula (I)
[0077] Precision=(Ω Seg ∩Ω GT ) / Ω Seg =TP / (TP+FP) Formula (II)
[0078] Sensitivity=(Ω Seg ∩Ω GT ) / Ω GT =TP / (TP+FN) Formula (III)
[0079] Table 1
[0080]
[0081] Table 2
[0082]
[0083] Tables 1 and 2 show the evaluation results (mean ± standard deviation) of the validation set and the test set, respectively.
[0084] Depend on Figure 3 As can be seen from Tables 1 and 2, for CT left ventricle segmentation, the 8-layer U-Net of the present application method has higher segmentation accuracy and robustness than the prototype U-Net.
[0085] In one embodiment, a CT left ventricle segmentation device based on a deep neural network is provided, comprising:
[0086] The first module is used to obtain CT cardiac images;
[0087] The second module is used to input the CT heart image into a deep neural network model to obtain a CT left ventricle segmentation result, and the deep neural network model is trained based on an 8-layer U-Net model.
[0088] For the functional limitations of each module, please refer to the limitations of the CT left ventricle segmentation method based on deep neural network mentioned above, which will not be repeated here. Each module in the above-mentioned CT left ventricle segmentation device based on deep neural network can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.
[0089] The deep neural network-based CT left ventricle segmentation device provided in this embodiment can be configured at a remote end to obtain CT cardiac images through a remote terminal connected thereto. Alternatively, the device of this embodiment itself can be configured at a terminal (such as a computer or medical detection equipment used by a user) to directly obtain CT cardiac images through a CT image acquisition device.
[0090] In one embodiment, a CT left ventricle segmentation device based on a deep neural network is provided, comprising a computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor, wherein when the computer processor executes the computer program, the CT left ventricle segmentation method based on a deep neural network is implemented.
[0091] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 5 As shown. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a CT left ventricle segmentation method based on a deep neural network is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covered on the display screen, or a button, trackball or touchpad set on the computer device housing, or an external keyboard, touchpad or mouse, etc.
[0092] Those skilled in the art will understand that Figure 5 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.
[0093] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the deep neural network-based CT left ventricle segmentation method when executing the computer program.
[0094] The present application also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a computer processor, it implements the CT left ventricle segmentation method based on deep neural network.
[0095] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).
[0096] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0097] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.
Claims
1. A CT left ventricle segmentation method based on deep neural network, characterized in that: include: Acquire CT cardiac images; Inputting the CT heart image into a deep neural network model to obtain a CT left ventricle segmentation result, wherein the deep neural network model is trained based on an 8-layer U-Net model; The 8-layer U-Net model includes a contraction path and an expansion path. In the expansion path, the first 7 upsampling operations are added with an auxiliary path. The auxiliary path is used for deep supervision so that the deep feature map is supervised, including: restoring the current feature map to the same size as the CT heart image through multiple upsampling operations and convolution layers, mapping the feature vector to a one-dimensional space, and performing weighted summation of the loss values output by the 7 auxiliary paths and the loss value output by the last decoding block of the expansion path from deep to shallow.
2. The CT left ventricle segmentation method based on deep neural network according to claim 1, characterized in that: The contraction path includes 9 coding blocks, each coding block contains two 3×3 convolutional layers, each convolutional layer is followed by a batch normalization layer and a ReLU activation layer, and for the first 8 coding blocks, each coding block is followed by a 2×2 maximum pooling layer for downsampling; The expansion path includes 8 decoding blocks, each of which contains two 3×3 convolutional layers, each of which is followed by a batch normalization layer and a ReLU activation layer, and each decoding block is preceded by a 2×2 deconvolutional layer for upsampling; The feature maps of corresponding blocks in the contraction path and the expansion path are jump-connected according to the channel dimension.
3. The CT left ventricle segmentation method based on deep neural network according to claim 2, characterized in that: A residual structure is incorporated into each encoding block and decoding block of the 8-layer U-Net model.
4. The CT left ventricle segmentation method based on deep neural network according to claim 1, characterized in that: The training image set of the 8-layer U-Net model is obtained as follows: Several CT cardiac images are acquired, and after uniformly adjusting the window width and window position, they are annotated using the Graph Cuts method to form an initial image set; During each iteration of the training process, each CT cardiac image in the initial image set is transformed N times to obtain N augmented images, N ≥ 8, and each transformation includes random rotation, random translation, random shearing and random scaling in sequence; The initial image set and the augmented images together constitute the training image set.
5. The CT left ventricle segmentation method based on deep neural network according to claim 4, characterized in that: The angle range of the random rotation is -5°~5°, the random translation includes horizontal random translation and vertical random translation performed simultaneously or sequentially, the pixel range of horizontal random translation is -0.02×image width~0.02×image width, the pixel range of vertical random translation is -0.02×image height~0.02×image height, the angle range of random shearing is -5°~5°, and the ratio range of random scaling is 0.90 to 1.
10.
6. A CT left ventricle segmentation device based on deep neural network, characterized in that: include: The first module is used to obtain CT cardiac images; The second module is used to input the CT heart image into a deep neural network model to obtain a CT left ventricle segmentation result. The deep neural network model is trained based on an 8-layer U-Net model. The 8-layer U-Net model includes a contraction path and an expansion path. In the expansion path, the first 7 upsampling operations are added with an auxiliary path. The auxiliary path is used for deep supervision so that the deep feature map is supervised, including: restoring the current feature map to the same size as the CT heart image through multiple upsampling operations and convolution layers, mapping the feature vector to a one-dimensional space, and performing weighted summation on the loss values output by the 7 auxiliary paths and the loss value output by the last decoding block of the expansion path from deep to shallow.
7. A CT left ventricle segmentation device based on deep neural network, characterized in that: The method comprises a computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor, wherein when the computer processor executes the computer program, the deep neural network-based CT left ventricle segmentation method as described in any one of claims 1 to 5 is implemented.
8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the CT left ventricle segmentation method based on deep neural network described in any one of claims 1 to 5 is implemented.
9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a computer processor, the deep neural network-based CT left ventricle segmentation method as described in any one of claims 1 to 5 is implemented.
Citation Information
Patent Citations
Semantic segmentation method, semantic segmentation device, semantic segmentation system and storage medium
CN108876793A
A choroidal segmentation method for OCT images based on improved U-net network
CN109509178A
