Prostate transrectal ultrasound image automatic segmentation method, device, equipment and medium

Abstract

The invention discloses an automatic segmentation method for a prostate transrectal ultrasound image, and relates to the technical field of machine learning, and the method comprises the following steps: S201, obtaining a prostate transrectal ultrasound image set of a patient; s202, creating a machine learning model for segmenting the prostate transrectal ultrasound image; s203, the obtained prostate transrectal ultrasound image set is preprocessed; s204, the machine learning model is used for automatically segmenting the preprocessed prostate transrectal ultrasound image set, and the prostate area is determined. The prostate image segmentation process is completely obtained through autonomous learning of the machine learning model, and a doctor with rich experience does not need to manually segment the prostate image; according to the prostate image segmentation method and device, the dependence on the professional level of a doctor is reduced, the burden of the doctor on diagnosing prostate diseases is greatly reduced, meanwhile, manual calibration can be carried out on the automatically segmented prostate image, the calibrated prostate image serves as an image sample to be stored in a database, and the accuracy of prostate image segmentation is further improved.

Description

Technical Field

[0001] The present invention relates to the field of machine learning technology, and in particular to a method, device, equipment and medium for automatic segmentation of prostate transrectal ultrasound images. Background Art

[0002] With the continuous development of medical technology, transrectal ultrasound has become an important means to assist doctors in diagnosing prostate diseases. Transrectal ultrasound is a technology that obtains high-definition images of the prostate by inserting an ultrasound probe into the rectum. This method is widely used to detect and evaluate prostate cancer, prostate hyperplasia and other prostate diseases. It has the advantages of being non-invasive, fast and low-cost.

[0003] In related technologies, the quality and resolution of ultrasound images have been significantly improved, allowing doctors to observe the internal structure of the prostate more clearly. However, auxiliary diagnosis through transrectal ultrasound images requires high experience of doctors and takes a lot of time. At present, many doctors lack the ability to accurately diagnose prostate diseases with the help of transrectal ultrasound images. The main reasons include image complexity, difficulty in identifying lesions, and high professional knowledge requirements for image interpretation. The introduction of automatic segmentation technology aims to solve the above problems. Automatic segmentation technology can help doctors quickly and accurately identify and mark the prostate and its internal lesions, thereby reducing diagnosis time and improving diagnostic accuracy. By using advanced image processing algorithms and machine learning technology, the automatic segmentation system can extract prostate-related information from a large number of transrectal ultrasound images and automatically complete the boundary identification of the prostate, which not only reduces the workload of doctors, but also improves the consistency and reliability of diagnosis.

[0004] However, there are still some challenges in the existing automatic segmentation methods. For example, the prostate morphology of different patients varies greatly, and the noise and artifacts in the ultrasound image may affect the accuracy of the segmentation results. In addition, the morphology of lesions such as prostate cancer is complex and changeable, which also increases the difficulty of automatic segmentation. Therefore, the development of a more accurate and robust automatic segmentation method for transrectal ultrasound images has important research value and application prospects. The present invention proposes a new automatic segmentation method, device, equipment and medium for prostate transrectal ultrasound images to solve the above problems in the prior art and improve the accuracy and efficiency of prostate disease diagnosis. Summary of the invention

[0005] The purpose of the present invention is to solve the problems in the prior art and to propose a device for automatically segmenting prostate transrectal ultrasound images.

[0006] The method for automatic segmentation of prostate transrectal ultrasound images comprises the following steps:

[0007] S201, obtaining a set of transrectal ultrasound images of the patient's prostate;

[0008] S202. Creating a machine learning model for segmenting prostate transrectal ultrasound images;

[0009] S203, preprocessing the acquired prostate transrectal ultrasound image set;

[0010] S204. Use a machine learning model to automatically segment the preprocessed prostate transrectal ultrasound image set to determine the prostate region.

[0011] Preferably, in step S202, the machine learning model is a three-dimensional convolutional neural network based on deep learning, and the machine learning model is trained according to a set of prostate transrectal ultrasound image samples in a prostate image database. The three-dimensional convolutional neural network includes an encoder network, a decoder network, a prostate detail compensation module, a three-dimensional perception module, an edge enhancement module and a multi-scale feature fusion module. The input of the three-dimensional convolutional neural network is a prostate transrectal ultrasound image, and the output is a segmented prostate area. The encoder network includes at least one group of three-dimensional convolutional layers and pooling layers, and the output of each group of three-dimensional convolutional layers and pooling layers is used as the input of the next group of three-dimensional convolutional layers and pooling layers. The decoder network includes at least one group of upsampling layers and three-dimensional convolutional layers, and the output of each group of upsampling layers and three-dimensional convolutional layers is used as the input of the next group of upsampling layers and three-dimensional convolutional layers.

[0012] Preferably, the input of the prostate detail compensation module is a three-dimensional organ image, and the output is a predicted organ region, and the prostate detail compensation module includes at least a set of three-dimensional convolution residual blocks and a pooling layer.

[0013] Preferably, the three-dimensional perception module includes a space perception module and a channel perception module, and the input of the three-dimensional perception module is the output feature N of the prostate detail compensation module. i , output features As the input of the encoder network, the spatial perception module consists of a 3D convolutional layer and an activation function layer. The working mechanism of the spatial perception module is as follows:

[0014]

[0015] Among them, Conv1() represents a three-dimensional convolution layer with a convolution kernel of 1×1×1, β() represents an activation function layer, and × represents feature multiplication;

[0016] The channel perception module consists of a global maximum pooling layer, a global average pooling layer, a fully connected layer, and an activation function layer. The working mechanism of the channel perception module is as follows:

[0017]

[0018] Among them, GMP() represents the global maximum pooling layer, GAP() represents the global average pooling layer, represents feature concatenation, α() represents the fully connected layer and the activation function layer;

[0019] Finally, the feature map and The fusion is performed through feature addition operation, which is described as:

[0020]

[0021] Among them, + represents the feature addition module.

[0022] Preferably, the edge enhancement module includes N three-dimensional convolution layers with convolution kernels of 3×3×3 and one three-dimensional convolution layer with a convolution kernel of 1×1×1. The input of the edge enhancement module is the shallow features of the encoder module, and its output is used as the input of the multi-scale feature fusion module.

[0023] Preferably, the multi-scale feature fusion module is composed of multiple three-dimensional convolutional layers and activation function layers, and its input is the encoder network input feature and the decoder network input feature group. First, the multi-layer feature F is fused by splicing operation. i (i∈0, 1, 2), a 3D convolutional layer with a convolution kernel of 1×1×1 is used to reduce the channel dimension, thereby reducing the computational complexity, and then N 3D convolutional layers with a convolution kernel of 7×7×7 are used for feature extraction, and a density connection is established between the feature map and the per-pixel classifier. In order to enhance the sensitivity to edge information, the encoder input feature E1 enhanced by the edge enhancement module is also introduced. In the spatial attention structure, a 3D convolutional layer with a convolution kernel of 3×3×3 is used for feature extraction, and an activation function layer is used to limit the value of the weight score to [0, 1]. Finally, N 3D convolutional layers with a convolution kernel of 3×3×3 are used for final feature extraction to obtain the final predicted prostate area. The working mechanism of the multi-scale feature fusion module is as follows:

[0024]

[0025] F″ M =F′ M ×β(Conv3(F′ m ))

[0026]

[0027] in, represents feature concatenation, Conv1() represents a three-dimensional convolution layer with a convolution kernel of 1×1×1, represents a three-dimensional convolution layer with a convolution kernel of 7×7×7, × represents feature multiplication, β() represents the activation function layer, Represents a three-dimensional convolution layer with a convolution kernel of 3×3×3.

[0028] According to a second aspect of the embodiments of the present invention, a device for automatically segmenting a prostate transrectal ultrasound image comprises a prostate transrectal ultrasound detection device and a prostate image processing device;

[0029] The prostate transrectal ultrasound detection device comprises a transrectal ultrasound probe for acquiring a set of transrectal ultrasound images of the prostate to be identified;

[0030] The prostate image processing device includes a real-time segmentation device, an offline training server device and a prostate image database. The real-time segmentation device is used to segment the prostate area in the prostate transrectal ultrasound image set through a machine learning model. The prostate image database stores prostate transrectal ultrasound image samples, each prostate transrectal ultrasound image sample carries corresponding prostate area annotation information, and the prostate area annotation information is used to indicate the area of ​​the prostate. The offline training server device is used to perform offline training on the machine learning model after the prostate image database data is updated, and update the parameters.

[0031] According to a third aspect of an embodiment of the present invention, a computer device includes: a processor, a memory, an input / output device, a transceiver, and a bus system;

[0032] The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

[0033] A fourth aspect of an embodiment of the present invention is a computer-readable storage medium, wherein the storage medium stores computer program instructions, and the computer program instructions can implement the method described in any one of claims 1 to 7.

[0034] Compared with the prior art, the present invention has the following advantages:

[0035] The prostate image segmentation process in the present invention is completely obtained by autonomous learning of the machine learning model, without the need for experienced doctors to manually segment the prostate image, reducing the dependence on the doctor's professional level and greatly reducing the burden on doctors in diagnosing prostate diseases. At the same time, the automatically segmented prostate image can be manually calibrated, and the calibrated prostate image will be stored in the database as an image sample, further improving the accuracy of prostate image segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a structural schematic diagram of the prostate transrectal ultrasound image segmentation and processing system in the present invention.

[0037] Figure 2 It is a schematic diagram of the flow of the prostate transrectal ultrasound image segmentation method of the present invention.

[0038] Figure 3 Schematic diagram of the structure of the three-dimensional convolutional neural network of the machine learning model in the present invention.

[0039] Figure 4 It is a schematic diagram of the structure of the prostate detail compensation module in the present invention.

[0040] Figure 5 It is a structural schematic diagram of the three-dimensional perception module in the present invention.

[0041] Figure 6 It is a schematic diagram of the structure of the edge enhancement module in the present invention.

[0042] Figure 7 Schematic diagram of the structure of the multi-scale feature fusion module in the present invention.

[0043] Figure 8 Schematic diagram of the generator and discriminator training process of the machine learning model in the present invention.

[0044] Fig. 9 Schematic diagram of the automatic segmentation result of the prostate region by the machine learning model in the present invention.

[0045] In the figure: 101 a user to be examined, 102 a transrectal ultrasound detection device for the prostate, 1021 a transrectal ultrasound probe, 103 a prostate image processing device, 1031 a real-time segmentation device, 1032 an offline training server device, 1033 a prostate image database, 301 an encoder network, 302 a decoder network, 401 a space perception module, 402 a channel perception module. DETAILED DESCRIPTION

[0046] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the present invention is further explained below in conjunction with specific implementation methods.

[0047] Reference Figure 1-7 As shown, a method for automatic segmentation of a prostate transrectal ultrasound image comprises the following steps:

[0048] S201, obtaining a set of transrectal ultrasound images of the patient's prostate;

[0049] S202. Creating a machine learning model for segmenting prostate transrectal ultrasound images;

[0050] S203, preprocessing the acquired prostate transrectal ultrasound image set;

[0051] S204. Use a machine learning model to automatically segment the preprocessed prostate transrectal ultrasound image set to determine the prostate region.

[0052] In the above step 201, the three-dimensional prostate transrectal ultrasound image set includes at least one three-dimensional prostate transrectal ultrasound image.

[0053] In the above step 203, the preprocessing includes tensor processing and normalization processing operations.

[0054] As an embodiment of the present invention, in step S202, the machine learning model is a three-dimensional convolutional neural network based on deep learning, and the machine learning model is trained according to a set of prostate transrectal ultrasound image samples in a prostate image database 1033, and the prostate transrectal ultrasound image sample collection contains at least 100 three-dimensional prostate transrectal ultrasound image samples. The three-dimensional convolutional neural network includes an encoder network 301, a decoder network 302, a prostate detail compensation module, a three-dimensional perception module, an edge enhancement module, and a multi-scale feature fusion module. The input of the three-dimensional convolutional neural network is a prostate transrectal ultrasound image, and the output is a segmented prostate area. The encoder network 301 includes at least one group of three-dimensional convolutional layers and pooling layers, and the output of each group of three-dimensional convolutional layers and pooling layers is used as the input of the next group of three-dimensional convolutional layers and pooling layers. The decoder network 302 includes at least one group of upsampling layers and three-dimensional convolutional layers, and the output of each group of upsampling layers and three-dimensional convolutional layers is used as the input of the next group of upsampling layers and three-dimensional convolutional layers.

[0055] As an embodiment of the present invention, the input of the prostate detail compensation module is a three-dimensional organ image, and the output is a predicted organ region. The prostate detail compensation module includes at least a set of three-dimensional convolution residual blocks and a pooling layer. In order to enhance the perception of the prostate, the prostate detail compensation module is pre-trained on a large-scale medical dataset 3DSeg-8 to learn the characteristics of different organs. The basic assumption is that the internal layers of the convolutional neural network are not targeted at specific tasks. For example, the shallow layers of the image classification convolutional neural network are very sensitive to detail information (such as edge and texture features). Specifically, the network-based transfer learning technology is usually used in the source task T s The network is pre-trained on the source task and then the features learned from the source task are transferred to the target task T t In the final inference stage of the machine learning model, the hierarchical features of the pre-trained prostate detail compensation module are extracted and fused with the features of the encoder network 301 to learn richer and more detailed prostate features.

[0056] As an embodiment of the present invention, the three-dimensional perception module includes a space perception module 401 and a channel perception module 402. The input of the three-dimensional perception module is the output feature N of the prostate detail compensation module. i , output features As the input of the encoder network 301, since the attention mechanism is widely used in many computer vision tasks, it can effectively improve the performance of machine learning. Since the features extracted from the prostate detail compensation module contain rich prostate information, but also include some non-prostate features, in order to further enhance and improve more important features, the three-dimensional perception module adaptively filters out non-prostate features and focuses on important features by exploring the relationship between features in spatial dimensions and channel dimensions; specifically, the feature map N from Res-i (i∈1, 2, 3, 4) in the prostate detail compensation module i is input into the 3D perception module, given the input feature map N i ∈R H×W×D×C (i∈1, 2, 3, 4), first input them into the spatial perception module 401 and the channel perception module 402 respectively to generate weight scores. The spatial perception module 401 consists of a three-dimensional convolution layer with a convolution kernel of 1×1×1 and an activation function layer. The three-dimensional convolution layer is used to calculate the input feature map N i The spatial weight score W s Then, the activation function layer adjusts the calculated spatial weights so that the weight score W s Between [0,1], so we get W′ s Finally, the calculated spatial weight score W′ s With input feature N i Multiply the features to get the adaptive feature map The working mechanism of the space perception module 401 is as follows:

[0057]

[0058] Among them, Conv1() represents a three-dimensional convolution layer with a convolution kernel of 1×1×1, β() represents an activation function layer, and × represents feature multiplication;

[0059] The channel perception module 402 is composed of a global maximum pooling layer, a global average pooling layer, a fully connected layer and an activation function layer. i , the global maximum pooling layer and the global average pooling layer first compress the feature map to N along the channel dimension i ∈R H×W×D×1 Then, the feature map is input into the fully connected layer to calculate the channel weight score W c , the activation function layer adjusts the calculated channel weights to limit W cBetween [0,1], so we get W′ c ,Finally, input feature map N i and the channel weight score W′ c Calculate the adaptive feature map The working mechanism of the channel sensing module 402 is as follows:

[0060]

[0061] Among them, GMP() represents the global maximum pooling layer, GAP() represents the global average pooling layer, represents feature concatenation, α() represents the fully connected layer;

[0062] Finally, the feature map and The fusion is performed through feature addition operation, which is described as:

[0063]

[0064] Among them, + represents the feature addition module.

[0065] As an embodiment of the present invention, the edge enhancement module includes N three-dimensional convolutional layers with N convolution kernels of 3×3×3 and a three-dimensional convolutional layer with a convolution kernel of 1×1×1. The input of the edge enhancement module is the shallow features of the encoder module, and its output is used as the input of the multi-scale feature fusion module. Since the prostate in the transrectal ultrasound image of the prostate has characteristics such as blurred edges, the current method is still difficult to accurately predict the edge of the prostate. In order to enhance the network's perception of prostate edge information, the embodiment of the present invention guides the shallow features of the encoder module through the edge enhancement module to pay more attention to the edge information of the prostate. First, it is necessary to generate a label value of the three-dimensional prostate edge information. One feasible method is to generate three-dimensional prostate edge information from the label value of the prostate segmentation image through the Canny algorithm. After obtaining the label value of the three-dimensional prostate edge image, in the overall network training stage, the edge enhancement module can be trained to guide the shallow features of the encoder to learn and predict the edge information of the prostate.

[0066] As an embodiment of the present invention, the multi-scale feature fusion module is composed of multiple three-dimensional convolutional layers and activation function layers, and its input is the encoder network 301 input feature and the decoder network 302 input feature group. First, the multi-layer feature F is fused by splicing operation. i(i∈0, 1, 2), a 3D convolutional layer with a convolution kernel of 1×1×1 is used to reduce the channel dimension, thereby reducing the computational complexity, and then N 3D convolutional layers with a convolution kernel of 7×7×7 are used for feature extraction, and a density connection is established between the feature map and the per-pixel classifier. In order to enhance the sensitivity to edge information, the encoder input feature E1 enhanced by the edge enhancement module is also introduced. In the spatial attention structure, a 3D convolutional layer with a convolution kernel of 3×3×3 is used for feature extraction, and an activation function layer is used to limit the value of the weight score to [0, 1]. Finally, N 3D convolutional layers with a convolution kernel of 3×3×3 are used for final feature extraction to obtain the final predicted prostate area. The working mechanism of the multi-scale feature fusion module is as follows:

[0067]

[0068] F″ M =F′ M ×β(Conv3(F′ M ))

[0069]

[0070] in, represents feature concatenation, Conv1() represents a three-dimensional convolution layer with a convolution kernel of 1×1×1, represents a three-dimensional convolution layer with a convolution kernel of 7×7×7, × represents feature multiplication, β() represents the activation function layer, Represents a three-dimensional convolutional layer with a convolution kernel of 3×3×3, where N is greater than or equal to 1.

[0071] As an embodiment of the present invention, a prostate transrectal ultrasound image automatic segmentation device includes a prostate transrectal ultrasound detection device 102 and a prostate image processing device 103;

[0072] The prostate transrectal ultrasound detection device 102 includes a transrectal ultrasound probe 1021, which may be a medical terminal device with an image acquisition function, and is used to obtain a set of prostate transrectal ultrasound images to be identified;

[0073] The prostate image processing device 103 includes a real-time segmentation device 1031, an offline training server device 1032 and a prostate image database 1033. The prostate image processing device 103 may be a server with image storage and processing functions. The real-time segmentation device 1031 is used to segment the prostate region in a set of prostate transrectal ultrasound images through a machine learning model. The prostate image database 1033 stores prostate transrectal ultrasound image samples. Each prostate transrectal ultrasound image sample carries corresponding prostate region annotation information. The prostate region annotation information is used to indicate the region of the prostate. The offline training server device 1032 is used to perform offline training on the machine learning model after the data in the prostate image database 1033 is updated, and to update the parameters.

[0074] When the doctor 104 performs a prostate diagnosis on the user 101 to be examined through the prostate transrectal ultrasound detection device 102, the real-time segmentation device 1031 will make a prediction based on the prostate transrectal ultrasound image obtained by the prostate transrectal ultrasound detection device 102, and output the segmented prostate result to provide to the doctor for auxiliary diagnosis. When the experienced doctor 104 believes that the prostate area predicted by the real-time segmentation device 1031 is inaccurate, he can manually intervene to recalibrate the prostate area, and after obtaining the patient's informed consent, the calibrated prostate annotated image and the prostate transrectal ultrasound image obtained by the prostate transrectal ultrasound detection device 102 will be stored in the prostate image database 1033. After reaching the data update threshold, the offline training server device 1032 will start to fine-tune the model and re-update the network parameters of the real-time segmentation device 1031.

[0075] As an embodiment of the present invention, a computer device includes: a processor, a memory, an input and output device, a transceiver, and a bus system;

[0076] The processor is configured to call the instructions stored in the memory to execute the above method.

[0077] As an embodiment of the present invention, a computer-readable storage medium is provided, wherein computer program instructions are stored in the storage medium, and the computer program instructions can implement the above method.

[0078] like Figure 8 As shown, the present invention also provides a training method for a machine learning model, comprising:

[0079] 1. First, it is necessary to obtain a set of three-dimensional prostate transrectal ultrasound images in the prostate transrectal ultrasound database 1033.

[0080] 2. Create and initialize the generator model and the discriminator model. The discriminator model can be any discriminator structure containing a three-dimensional convolutional neural network. One feasible solution is patchGAN.

[0081] 3. Create and initialize the dataset image reading module, optimizer module and loss function module of the training model. The dataset image reading module can be any module with a three-dimensional image data reading function, and the optimizer module can be any deep learning optimizer. One feasible solution is the Adam optimizer.

[0082] 4. Preprocess the acquired three-dimensional prostate transrectal ultrasound images, and the preprocessing process includes tensor quantization, normalization, image inversion and slicing operations.

[0083] 5. The generator model automatically segments the prostate region based on the input three-dimensional prostate transrectal ultrasound image.

[0084] 6. The discriminator model makes a judgment based on the output of the generator model to determine whether the prostate segmented by the generator model is real or fake.

[0085] 7. The loss function is calculated through forward propagation and back propagation to reduce the loss function value, and the generator model and the discriminator model are trained and converged.

[0086] 8. Convergence judgment is made on the loss function. If the loss function converges, the final model parameters are saved. If not, the training of the generator and discriminator models is continued. In deep learning tasks, the loss function plays a vital role in the network model training process. Designing a reasonable loss function can not only speed up the convergence speed of model training, but also improve the prediction accuracy of the model. In an embodiment of the present invention, a reasonable loss function can be used to train the generator model and the discriminator model. The following is a feasible loss function design, as shown in the following formula:

[0087] minL(D)=[D(x,y)-l r ] 2 +[D(x,G(x))-l f ] 2

[0088] Where G() represents the generator network, D() represents the discriminator network, x and y represent the input three-dimensional prostate transrectal ultrasound image and the corresponding label, respectively, l r and l f represent the true prostate label and false prostate matrix label with constants of 1 and 0, respectively;

[0089] The objective function for training the generator model is as follows:

[0090]

[0091] Among them, l t is a matrix with constant elements equal to 1, y e and Represent the predicted prostate edge map and the corresponding prostate edge label map, respectively. dice represents the Dice loss function, α and β represent hyperparameters for controlling the weight of the loss function. According to experimental results, when α and β are set to 1 and 0.5 respectively, the machine learning model of the embodiment of the present invention achieves the best prediction performance.

[0092] 9. After the loss function converges, the parameters of the generator model are saved as the machine learning segmentation model in the real-time segmentation device 1031.

[0093] like Fig. 9 As shown, the first column is the three-dimensional prostate transrectal ultrasound image in the prostate transrectal ultrasound database, and the second column is the result of automatic prostate segmentation by the machine learning model based on the three-dimensional prostate transrectal ultrasound image in the first column.

[0094] In summary, the prostate image segmentation process in the present invention is completely obtained by autonomous learning of the machine learning model, without the need for experienced doctors to manually segment the prostate image, which reduces the dependence on the professional level of doctors and greatly reduces the burden on doctors in diagnosing prostate diseases. At the same time, the automatically segmented prostate image can be manually calibrated, and the calibrated prostate image will be stored in the database as an image sample, further improving the accuracy of prostate image segmentation.

[0095] It is known from common technical knowledge that the present invention can be implemented by other embodiments that do not deviate from its spirit or essential features. Therefore, the above disclosed embodiments are only illustrative in all respects and are not exclusive. All changes within the scope of the present invention or within the scope equivalent to the present invention are included in the present invention.

Claims

1. A method for automatic segmentation of prostate transrectal ultrasound images, characterized by: The following steps are involved: S201, obtaining a set of transrectal ultrasound images of the patient's prostate; S202. Creating a machine learning model for segmenting prostate transrectal ultrasound images; S203, preprocessing the acquired prostate transrectal ultrasound image set; S204. Use a machine learning model to automatically segment the preprocessed prostate transrectal ultrasound image set to determine the prostate region.

2. The method for automatic segmentation of prostate transrectal ultrasound images according to claim 1, characterized in that: In step S202, the machine learning model is a three-dimensional convolutional neural network based on deep learning. The machine learning model is trained based on a set of prostate transrectal ultrasound image samples in a prostate image database (1033). The three-dimensional convolutional neural network includes an encoder network (301), a decoder network (302), a prostate detail compensation module, a three-dimensional perception module, an edge enhancement module and a multi-scale feature fusion module. The input of the three-dimensional convolutional neural network is a prostate transrectal ultrasound image, and the output is a segmented prostate region. The encoder network (301) includes at least one set of three-dimensional convolutional layers and pooling layers, and the output of each set of three-dimensional convolutional layers and pooling layers serves as the input of the next set of three-dimensional convolutional layers and pooling layers. The decoder network (302) includes at least one set of upsampling layers and three-dimensional convolutional layers, and the output of each set of upsampling layers and three-dimensional convolutional layers serves as the input of the next set of upsampling layers and three-dimensional convolutional layers.

3. The method for automatic segmentation of prostate transrectal ultrasound images according to claim 2, characterized in that: The input of the prostate detail compensation module is a three-dimensional organ image, and the output is a predicted organ region. The prostate detail compensation module includes at least a set of three-dimensional convolution residual blocks and a pooling layer.

4. The method for automatic segmentation of prostate transrectal ultrasound images according to claim 2, characterized in that: The three-dimensional perception module comprises a space perception module (401) and a channel perception module (402), and the input of the three-dimensional perception module is the output feature N of the prostate detail compensation module. i , output features As the input of the encoder network (301), the spatial perception module (401) is composed of a three-dimensional convolution layer and an activation function layer. The working mechanism of the spatial perception module (401) is as follows: Among them, Conv1() represents a three-dimensional convolution layer with a convolution kernel of 1×1×1, β() represents an activation function layer, and × represents feature multiplication; The channel perception module (402) is composed of a global maximum pooling layer, a global average pooling layer, a fully connected layer and an activation function layer. The working mechanism of the channel perception module (402) is as follows: Among them, GMP() represents the global maximum pooling layer, GAP() represents the global average pooling layer, represents feature concatenation, α() represents the fully connected layer and the activation function layer; Finally, the feature map and The fusion is performed through feature addition operation, which is described as: Among them, + represents the feature addition module.

5. The method for automatic segmentation of prostate transrectal ultrasound images according to claim 2, characterized in that: The edge enhancement module includes N three-dimensional convolution layers with convolution kernels of 3×3×3 and one three-dimensional convolution layer with convolution kernels of 1×1×1. The input of the edge enhancement module is the shallow features of the encoder module, and its output is used as the input of the multi-scale feature fusion module.

6. The method for automatic segmentation of prostate transrectal ultrasound images according to claim 2, characterized in that: The multi-scale feature fusion module is composed of multiple three-dimensional convolutional layers and activation function layers, and its input is the input feature of the encoder network (301) and the input feature group of the decoder network (302). First, the multi-layer feature F is fused by splicing operation. i (i∈0, 1, 2), a 3D convolutional layer with a convolution kernel of 1×1×1 is used to reduce the channel dimension, thereby reducing the computational complexity, and then N 3D convolutional layers with a convolution kernel of 7×7×7 are used for feature extraction, and a density connection is established between the feature map and the per-pixel classifier. In order to enhance the sensitivity to edge information, the encoder input feature E1 enhanced by the edge enhancement module is also introduced. In the spatial attention structure, a 3D convolutional layer with a convolution kernel of 3×3×3 is used for feature extraction, and an activation function layer is used to limit the value of the weight score to [0, 1]. Finally, N 3D convolutional layers with a convolution kernel of 3×3×3 are used for final feature extraction to obtain the final predicted prostate area. The working mechanism of the multi-scale feature fusion module is as follows: F″ M =F′ M ×β(Conv3(F′ M )) in, represents feature concatenation, Conv1() represents a three-dimensional convolution layer with a convolution kernel of 1×1×1, represents a three-dimensional convolution layer with a convolution kernel of 7×7×7, × represents feature multiplication, β() represents the activation function layer, Represents a three-dimensional convolution layer with a convolution kernel of 3×3×3.

7. An automatic segmentation device for prostate transrectal ultrasound images, characterized in that: It comprises a prostate transrectal ultrasound detection device (102) and a prostate image processing device (103); The prostate transrectal ultrasound detection device (102) comprises a transrectal ultrasound probe (1021) for acquiring a set of transrectal ultrasound images of the prostate to be identified; The prostate image processing device (103) includes a real-time segmentation device (1031), an offline training server device (1032) and a prostate image database (1033). The real-time segmentation device (1031) is used to segment the prostate region in a set of prostate transrectal ultrasound images through a machine learning model. The prostate image database (1033) stores prostate transrectal ultrasound image samples, each prostate transrectal ultrasound image sample carries corresponding prostate region annotation information, and the prostate region annotation information is used to indicate the region of the prostate. The offline training server device (1032) is used to perform offline training on the machine learning model after the data in the prostate image database (1033) is updated, and to update parameters.

8. A computer device, characterized in that: include: Processors, memories, input and output devices, transceivers, and bus systems; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that: The storage medium stores computer program instructions, and the computer program instructions can implement the method according to any one of claims 1 to 7.

Images