Focus recognition model deployment method, focus recognition method, system and equipment
By building a target inference engine for lesion recognition model in the TensorRT platform, the problems of long inference time and large memory usage in the prior art are solved, and more efficient lesion recognition and lower memory consumption are achieved.
Patent Information
- Application Number
- CN202311408085.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-10-26
- Publication Date
- 2025-04-29
AI Technical Summary
In the prior art, the C++ deep learning framework or the python deep learning framework is used to deploy the lesion recognition model, resulting in the long-term model inference time and large memory occupancy.
By obtaining the model parameters of the lesion identification model and building a target inference engine in the TensorRT platform, optimizing the network structure and configuration parameters of the model, reducing unnecessary hierarchies, merging convolutional layers and normalization layers, serializing and saving target binary data.
The inference speed of lesion recognition model has been accelerated, the efficiency of lesion recognition in medical image is improved, and memory consumption is effectively reduced.
Smart Images
Figure CN119904402A_ABST
Abstract
Description
Technical Field
[0001] The present disclosure relates to the field of medical imaging technology, and in particular to a deployment method of a lesion recognition model, a lesion recognition method, a system and a device. Background Art
[0002] Medical images are images that reflect the internal structure of the human body and are one of the main bases for modern medical diagnosis. Medical image lesion recognition has gradually evolved from manual recognition to model intelligent recognition. In the prior art, C++ (a computer language) deep learning framework or python (a computer language) deep learning framework is usually used to deploy lesion recognition models. However, this method has problems such as long inference time and large memory usage. Summary of the invention
[0003] The technical problem to be solved by the present invention is to overcome the defects of the prior art that the lesion recognition model is deployed using a C++ deep learning framework or a python deep learning framework, which leads to a long model inference time and large memory usage, and to provide a deployment method for a lesion recognition model, a lesion recognition method, a system and a device.
[0004] The present invention solves the above technical problems through the following technical solutions:
[0005] The present disclosure provides a deployment method of a lesion recognition model, the deployment method comprising:
[0006] Obtaining a lesion recognition model;
[0007] deriving model parameters of the lesion recognition model;
[0008] Based on the model parameters, a target reasoning engine of the lesion recognition model is constructed in the TensorRT (a reasoning framework) platform.
[0009] Preferably, the step of constructing a target inference engine of the lesion recognition model in the TensorRT platform based on the model parameters includes:
[0010] Building an initial network structure of the lesion recognition model in the TensorRT platform;
[0011] Importing the model parameters into the initial network structure to obtain a target network structure;
[0012] Based on the target network structure and engine configuration parameters, the target inference engine is constructed.
[0013] Preferably, the engine configuration parameters include at least one of model parameter type, workspace size, and batch size.
[0014] Preferably, after the step of importing the model parameters into the initial network structure to obtain the target network structure, the step further includes:
[0015] The convolutional layer and the normalization layer in the target network structure are merged, or the convolutional layer, the normalization layer and the activation layer in the target network structure are deleted to update a new target network structure.
[0016] Preferably, after the step of constructing the target inference engine of the lesion recognition model in the TensorRT platform based on the model parameters, the step further includes:
[0017] Serializing the target inference engine to obtain target binary data;
[0018] The target binary data is saved.
[0019] The present disclosure also provides a lesion identification method, the lesion identification method comprising:
[0020] Acquire target medical images;
[0021] Inputting the target medical image into the target reasoning engine and outputting a lesion recognition result;
[0022] Among them, the target reasoning engine is obtained based on the deployment method of the above-mentioned lesion recognition model.
[0023] Preferably, after the step of inputting the target medical image into the target reasoning engine and outputting the lesion identification result, the step further includes:
[0024] A connected domain analysis is performed on the lesion recognition result, connected domains in the lesion recognition result that are smaller than a preset connected range are eliminated, and a new lesion recognition result is updated.
[0025] The present disclosure also provides a deployment system for a lesion recognition model, the deployment system comprising:
[0026] A model acquisition module, used to acquire a lesion recognition model;
[0027] A parameter export module, used for exporting model parameters of the lesion recognition model;
[0028] An engine building module is used to build a target reasoning engine of the lesion recognition model in the TensorRT platform based on the model parameters.
[0029] Preferably, the engine building module includes:
[0030] An initial network building unit, used to build the initial network structure of the lesion recognition model in the TensorRT platform;
[0031] A target network building unit, used for importing the model parameters into the initial network structure to obtain a target network structure;
[0032] The engine construction unit is used to construct the target reasoning engine based on the target network structure and engine configuration parameters.
[0033] Preferably, the engine configuration parameters include at least one of model parameter type, workspace size, and batch size.
[0034] Preferably, the engine building module further includes:
[0035] An updating unit is used to merge the convolutional layer and the normalization layer in the target network structure, or to delete the convolutional layer, the normalization layer and the activation layer in the target network structure to update a new target network structure.
[0036] Preferably, the deployment system further includes:
[0037] A serialization module, used to serialize the target inference engine to obtain target binary data;
[0038] A data saving module is used to save the target binary data.
[0039] The present disclosure also provides a lesion identification system, the lesion identification system comprising:
[0040] An image acquisition module, used to acquire target medical images;
[0041] A result output module, used to input the target medical image into the target reasoning engine and output a lesion recognition result;
[0042] Wherein, the target reasoning engine is obtained based on the deployment system of the above-mentioned lesion recognition model.
[0043] Preferably, the lesion identification system further comprises:
[0044] The result updating module is used to perform connected domain analysis on the lesion recognition result, eliminate the connected domains in the lesion recognition result that are smaller than a preset connected range, and update to obtain a new lesion recognition result.
[0045] The present disclosure also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and used to run on the processor, and when the processor executes the computer program, the above-mentioned lesion recognition model deployment method and / or lesion recognition method is implemented.
[0046] The present disclosure also provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the above-mentioned lesion recognition model deployment method and / or lesion recognition method is implemented.
[0047] On the basis of conforming to the common sense in the art, various preferred conditions can be arbitrarily combined to obtain various preferred embodiments of the present disclosure.
[0048] The positive progress effect of the present disclosure is: deploying the lesion recognition model based on TensorRT, accelerating the reasoning speed of the lesion recognition model, improving the efficiency of medical image lesion recognition, and effectively reducing memory consumption. BRIEF DESCRIPTION OF THE DRAWINGS
[0049] Figure 1 This is the first flow chart of the deployment method of the lesion recognition model of Example 1 of the present disclosure.
[0050] Figure 2 This is the second flow chart of the deployment method of the lesion recognition model of Example 1 of the present disclosure.
[0051] Figure 3 This is the third flow chart of the deployment method of the lesion recognition model of Example 1 of the present disclosure.
[0052] Figure 4 Schematic diagram of the combination of the convolution layer and the normalization layer of Example 1 of the present disclosure.
[0053] Figure 5 This is the fourth flow chart of the deployment method of the lesion recognition model of Example 1 of the present disclosure.
[0054] Figure 6 This is a schematic diagram of the network construction of the lesion recognition model of Example 1 of the present disclosure.
[0055] Figure 7 This is the first flow chart of the lesion identification method of embodiment 2 of the present disclosure.
[0056] Figure 8 This is the second flow chart of the lesion identification method of Embodiment 2 of the present disclosure.
[0057] Fig. 9 This is a schematic diagram of the first module of the deployment system of the lesion recognition model of Example 3 of the present disclosure.
[0058] Fig.10 This is a schematic diagram of the second module of the deployment system of the lesion recognition model of Example 3 of the present disclosure.
[0059] Fig.11 This is a schematic diagram of the first module of the lesion identification system of Example 4 of the present disclosure.
[0060] Fig.12This is a schematic diagram of the second module of the lesion identification system of Example 4 of the present disclosure.
[0061] Fig.13 This is a schematic diagram of the structure of an electronic device according to Embodiment 5 of the present disclosure. DETAILED DESCRIPTION
[0062] The present disclosure is further described below by way of examples, but the present disclosure is not limited to the scope of the examples.
[0063] Example 1
[0064] This embodiment provides a method for deploying a lesion recognition model. Figure 1 As shown, the deployment method includes the following steps:
[0065] S1. Obtain a lesion recognition model.
[0066] Obtain a trained lesion recognition model from a training framework such as PyTorch, MXNet, or TensorFlow.
[0067] S2. Export model parameters of the lesion recognition model.
[0068] Export model parameters such as weights in the lesion recognition model to a text file.
[0069] S3. Based on the model parameters, a target inference engine for the lesion recognition model is built in the TensorRT platform.
[0070] Based on the model parameters of the lesion recognition model, the corresponding target inference engine is built in the TensorRT platform to complete the deployment of the lesion recognition model. The target inference engine is a serialized and optimized TensorRT model.
[0071] The TensorRT deployment solution usually uses frameworks such as onnx (open neural network exchange format) to export model weights, while this solution uses a solution that directly reads model parameters, builds the network structure, and loads model parameters.
[0072] In this solution, by directly exporting the model parameters and deploying the lesion recognition model based on TensorRT, the reasoning speed of the lesion recognition model can be accelerated, and memory consumption can be effectively reduced.
[0073] In one feasible solution, Figure 2 As shown, step S3 includes:
[0074] S301. Build the initial network structure of the lesion recognition model on the TensorRT platform;
[0075] S302. Import the model parameters into the initial network structure to obtain the target network structure;
[0076] S303. Build the target inference engine based on the target network structure and the engine configuration parameters.
[0077] Specifically, first build the initial network structure on the TensorRT platform; then import the model parameters such as weights in the text file to obtain the target network structure; finally, build the target inference engine according to the engine configuration parameters.
[0078] In this solution, by deploying the lesion recognition model on the TensorRT platform and selecting appropriate engine configuration parameters, the inference speed of the lesion recognition model can be accelerated and the memory consumption can be reduced.
[0079] In an implementable solution, the engine configuration parameters include at least one (one or more) of the model parameter type, the workspace size, and the batch size.
[0080] Specifically, the model parameter type can be selected from types such as fp32 (32-bit floating-point number), fp16 (16-bit floating-point number), int32 (32-bit integer), and int8 (8-bit integer); the workspace size is in MB and can be calculated according to the size of the model itself and the engine configuration. Too high may affect the number of applications sharing the GPU simultaneously, and too low may filter out some algorithms and create a sub-optimal engine; the batch size is 1 to n. Generally speaking, the larger the batch size, the smaller the average amount of calculation for each data, but too large a batch size may cause the video memory to be insufficient because the deployment needs to satisfy batch * data size + (plus) model size + (plus) configuration parameters < GPU video memory. In actual work, the batch size can be set according to actual conditions. For example, if two data need to be calculated simultaneously, the batch size is set to 2. Among them, n is an integer greater than or equal to 1.
[0081] In a specific example, the model parameter type, the workspace size, and the batch size can be set to fp32, 1GB, and 1 respectively.
[0082] In this solution, by optimizing and configuring parameters such as the model parameter type, the workspace size, and the batch size, the efficiency of model deployment can be improved.
[0083] In an implementable solution, as Figure 3 shown, after step S302, it further includes:
[0084] Step S304: merge the convolutional layer and the normalization layer in the target network structure, or delete the convolutional layer, the normalization layer and the activation layer in the target network structure to update and obtain a new target network structure.
[0085] Specifically, in the process of network construction, without affecting the results, the original target network structure can be further optimized, and unnecessary activation layers and normalization layers can be discarded or merged to further improve the network's reasoning efficiency.
[0086] like Figure 4 As shown, parallel convolutional layers and normalization layers can be combined.
[0087] When deleting convolutional layers, normalization layers, and activation layers, the model structure is generally evaluated first. For example, some multi-scale convolutional layers, normalization layers, and activation layers are removed, and then the result indicators are obtained through training and reasoning. The new indicators are compared with the original indicators. If the new indicators meet the preset requirements, the above convolutional layers, normalization layers, and activation layers can be removed.
[0088] In this scheme, the reasoning efficiency of the lesion recognition model can be further improved by merging the activation layer and normalization layer in the target network structure during the deployment process.
[0089] In one feasible solution, Figure 5 As shown, after step S3, the following steps are also included:
[0090] S4, serializing the target inference engine to obtain target binary data;
[0091] S5. Save the target binary data.
[0092] Specifically, the target inference engine is serialized into binary data and saved to the hard disk so that it can be deserialized to obtain the engine object and used in actual operation.
[0093] In this solution, by serializing the target inference engine, the storage and use of the engine can be facilitated.
[0094] The following takes a specific implementation as an example to illustrate the implementation process of the deployment method of the lesion recognition model provided in this embodiment.
[0095] 1. Environment and model construction
[0096] 1.1 Environment Configuration
[0097] 1.1.1cuda (universal parallel computing architecture) / cudnn (deep neural network library) and creation of virtual environment
[0098] 1.1.2 Install the corresponding version of TensorRT according to the cuda version
[0099] For example, if the local machine is cuda11.0 version, download the corresponding version of TensorRT; install the TensorRT Pythonwheel (an installation component) file, and similarly install uff (a universal file format), graphsurgeon (an image tool library), and onnx_graphsurgeon (an image tool library); activate the virtual environment and switch the path to the python file under TensorRT; system environment configuration: Advanced system environment - Environment variables - System variables - Path, add the library path of TensorRT; then use pip (package management tool) to download according to the python version to verify whether it is successful.
[0100] 1.2 Model parameter export
[0101] Taking the pytorch framework training model as an example, the trained model is a pytorch file. You can use a python script to read the file and save the loaded model weight parameters into a binary file.
[0102] 1.3 Building a network model
[0103] Using TensorRT's API (programming interface) to build the network layer by layer is similar to using a general training framework, such as using Pytorch or TensorFlow to build the network. Figure 6 As shown in the figure, the input image (target medical image) is subjected to multiple Max Pooling processes to reduce its resolution; then multiple Up Sampling processes are performed to increase its resolution; images with similar resolutions are subjected to Skip Connections; and the connected images are subjected to Conv+activation+normalization processes to obtain the output results. It should be noted that for the weight part, such as the convolution or normalization layer, the weight content of the binary file needs to be assigned to the TensorRT network; during the network construction process, the original network structure is further optimized without affecting the result precursor, and the unnecessary structural modules are discarded or merged to further improve the network's reasoning efficiency. During the construction process: 1. It is necessary to verify the input and output results layer by layer to ensure that the model is built correctly; 2. The loaded weights can be divided into float32, float16 and int8 (optional type). This example uses float32 type weights; different types of weights will bring different deviation ranges to the final result. Generally speaking, the inference speed of different types of weights is Int8>float16>float32, which can be provided according to actual needs.
[0104] 1.4 Build TensorRT inference engine and storage serialization engine
[0105] An engine object (i.e., the target inference engine mentioned above) is created according to the input network structure and optimization configuration; in the serialization stage, the engine object is serialized into a block of binary data and saved to the hard disk so that it can be deserialized and used in actual operation.
[0106] The optimization configuration (i.e., the above-mentioned model parameter type) includes parameter type, workspace size, and batch size, which can be optimized and adjusted according to actual conditions.
[0107] The deployment method of the lesion recognition model provided in this embodiment can speed up the reasoning speed of the lesion recognition model by directly exporting model parameters and deploying the lesion recognition model based on TensorRT, while also effectively reducing memory consumption.
[0108] Example 2
[0109] This embodiment provides a method for identifying a lesion. Figure 7 As shown, the lesion identification method comprises the following steps:
[0110] S6. Acquire target medical images.
[0111] Obtain the target medical image that needs to be identified, pre-process the image data, such as regularization, resizing, etc., and move it to the GPU (graphics processing unit) buffer.
[0112] S7, inputting the target medical image into the target reasoning engine and outputting the lesion recognition result;
[0113] Among them, the target reasoning engine is obtained based on the deployment method of the lesion recognition model in Example 1.
[0114] Read serialized binary data from the hard disk and deserialize it through the TensorRT API to obtain the target inference engine; input the target medical image into the target inference engine and output the lesion recognition result.
[0115] In this solution, lesion recognition is performed through the target inference engine deployed by TensorRT, which speeds up the reasoning speed of the lesion recognition model and improves the efficiency of lesion recognition in medical images.
[0116] In one feasible solution, Figure 8 As shown, after step S7, the following steps are also included:
[0117] S8. Performing a connected domain analysis on the lesion recognition results, eliminating connected domains in the lesion recognition results that are smaller than a preset connected range, and updating to obtain a new lesion recognition result.
[0118] That is, the smaller connected domains in the lesion recognition results are removed and the lesion recognition results are updated.
[0119] Connected domain analysis is to cluster the lesion results. 4-connected, 8-connected, and 26-connected domains can be used for analysis. Taking the 4-connected domain as an example, traverse each result. If there are lesion results of the same type above, below, left, and right, they are grouped into one cluster.
[0120] In this scheme, the accuracy of lesion identification results can be improved by screening the connected domains.
[0121] The lesion recognition method provided in this embodiment performs lesion recognition through the target inference engine deployed by TensorRT, which speeds up the reasoning speed of the lesion recognition model and improves the efficiency and accuracy of medical image lesion recognition.
[0122] Example 3
[0123] This embodiment provides a deployment system for a lesion recognition model. Fig. 9 As shown, the deployment system includes:
[0124] Model acquisition module 1, used to acquire a lesion recognition model;
[0125] A parameter export module 2 is used to export model parameters of the lesion recognition model;
[0126] The engine building module 3 is used to build a target reasoning engine of the lesion recognition model in the TensorRT platform based on the model parameters.
[0127] In one feasible solution, Fig.10 As shown, the engine building module 3 includes:
[0128] An initial network building unit 301 is used to build an initial network structure of a lesion recognition model in the TensorRT platform;
[0129] A target network building unit 302 is used to import model parameters into the initial network structure to obtain a target network structure;
[0130] The engine construction unit 303 is used to construct a target reasoning engine based on the target network structure and engine configuration parameters.
[0131] In an implementable solution, the engine configuration parameters include at least one of a model parameter type, a workspace size, and a batch size.
[0132] In an implementable solution, the engine building module further includes:
[0133] The updating unit 304 is used to merge the convolutional layer and the normalization layer in the target network structure, or delete the convolutional layer, the normalization layer and the activation layer in the target network structure to update and obtain a new target network structure.
[0134] In one feasible solution, the deployment system further includes:
[0135] A serialization module 4 is used to serialize the target inference engine to obtain target binary data;
[0136] The data storage module 5 is used to store the target binary data.
[0137] Since the deployment system of the lesion identification model provided in this embodiment is based on the same principle as the deployment method of the lesion identification model provided in Example 1, they will not be described in detail here.
[0138] The deployment system of the lesion recognition model provided in this embodiment can speed up the reasoning speed of the lesion recognition model by directly exporting model parameters and deploying the lesion recognition model based on TensorRT, while also effectively reducing memory consumption.
[0139] Example 4
[0140] The present disclosure also provides a lesion identification system, such as Fig.11 As shown, the lesion recognition system includes:
[0141] An image acquisition module 6, used to acquire target medical images;
[0142] A result output module 7, used to input the target medical image into the target reasoning engine and output the lesion recognition result;
[0143] Among them, the target reasoning engine is obtained based on the deployment method of the above-mentioned lesion recognition model.
[0144] In one feasible solution, Fig.12 As shown, the lesion recognition system also includes:
[0145] The result updating module 8 is used to perform connected domain analysis on the lesion recognition result, eliminate the connected domains in the lesion recognition result that are smaller than a preset connected range, and update to obtain a new lesion recognition result.
[0146] Since the lesion identification system provided in this embodiment has the same principle as the lesion identification method provided in Example 2, they will not be described in detail here.
[0147] The lesion recognition method provided in this embodiment performs lesion recognition through the target inference engine deployed by TensorRT, which speeds up the reasoning speed of the lesion recognition model and improves the efficiency and accuracy of medical image lesion recognition.
[0148] Example 5
[0149] This embodiment provides an electronic device, Fig.13 The electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the deployment method of the lesion recognition model of Example 1 or the lesion recognition method of Example 2 when executing the program. Fig.13 The electronic device 30 shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present disclosure.
[0150] like Fig.13 As shown, the electronic device 30 may be in the form of a general-purpose computing device, for example, it may be a server device. The components of the electronic device 30 may include, but are not limited to: at least one processor 31, at least one memory 32, and a bus 33 connecting different system components (including the memory 32 and the processor 31).
[0151] The bus 33 includes a data bus, an address bus, and a control bus.
[0152] The memory 32 may include a volatile memory, such as a random access memory (RAM) 321 and / or a cache memory 322 , and may further include a read-only memory (ROM) 323 .
[0153] The memory 32 may also include a program / utility 325 having a set (at least one) of program modules 324, such program modules 324 including but not limited to: an operating system, one or more application programs, other program modules, and program data, each of which or some combination may include an implementation of a network environment.
[0154] The processor 31 executes various functional applications and data processing by running the computer program stored in the memory 32, such as the deployment method of the lesion recognition model of Example 1 or the lesion recognition method of Example 2 of the present disclosure.
[0155] The electronic device 30 may also communicate with one or more external devices 34 (e.g., keyboards, pointing devices, etc.). Such communication may be performed via an input / output (I / O) interface 35. Furthermore, the model generating device 30 may also communicate with one or more networks (e.g., a local area network (LAN), a wide area network (WAN), and / or a public network, such as the Internet) via a network adapter 36. Fig.13As shown, the network adapter 36 communicates with other modules of the model-generated device 30 via the bus 33. It should be understood that, although not shown in the figure, other hardware and / or software modules can be used in conjunction with the model-generated device 30, including but not limited to: microcode, device drivers, redundant processors, external disk drive arrays, RAID (disk array) systems, tape drives, and data backup storage systems, etc.
[0156] It should be noted that although several units / modules or sub-units / modules of the electronic device are mentioned in the above detailed description, this division is merely exemplary and not mandatory. In fact, according to the embodiments of the present disclosure, the features and functions of two or more units / modules described above can be embodied in one unit / module. Conversely, the features and functions of one unit / module described above can be further divided into multiple units / modules to be embodied.
[0157] Example 6
[0158] This embodiment provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the deployment method of the lesion recognition model of Example 1 or the lesion recognition method of Example 2 is implemented.
[0159] The readable storage medium may include but is not limited to: a portable disk, a hard disk, a random access memory, a read-only memory, an erasable programmable read-only memory, an optical storage device, a magnetic storage device or any suitable combination of the above.
[0160] In a possible implementation, the present disclosure can also be implemented in the form of a program product, which includes a program code. When the program product is run on a terminal device, the program code is used to enable the terminal device to execute the deployment method of the lesion identification model of Example 1 or the lesion identification method of Example 2.
[0161] Among them, the program code for executing the present disclosure can be written in any combination of one or more programming languages, and the program code can be executed completely on the user device, partially on the user device, as a separate software package, partially on the user device and partially on a remote device, or completely on the remote device.
[0162] Although the specific embodiments of the present disclosure are described above, those skilled in the art should understand that this is only an example, and the protection scope of the present disclosure is defined by the appended claims. Those skilled in the art may make various changes or modifications to these embodiments without departing from the principles and essence of the present disclosure, but these changes and modifications all fall within the protection scope of the present disclosure.
Claims
1. A method for deploying a lesion recognition model, characterized in that: The deployment method includes: Obtaining a lesion recognition model; deriving model parameters of the lesion recognition model; Based on the model parameters, a target inference engine of the lesion recognition model is constructed in the TensorRT platform.
2. The deployment method of the lesion recognition model as claimed in claim 1, characterized in that: The step of constructing a target inference engine of the lesion recognition model in the TensorRT platform based on the model parameters includes: Building an initial network structure of the lesion recognition model in the TensorRT platform; Importing the model parameters into the initial network structure to obtain a target network structure; Based on the target network structure and engine configuration parameters, the target inference engine is constructed.
3. The deployment method of the lesion recognition model as claimed in claim 2, characterized in that: The engine configuration parameters include at least one of a model parameter type, a workspace size, and a batch size.
4. The deployment method of the lesion recognition model as claimed in claim 3, characterized in that: After the step of importing the model parameters into the initial network structure to obtain the target network structure, the following step is further included: The convolutional layer and the normalization layer in the target network structure are merged, or the convolutional layer, the normalization layer and the activation layer in the target network structure are deleted to update a new target network structure.
5. The method for deploying a lesion recognition model as claimed in claim 1, characterized in that: After the step of constructing the target inference engine of the lesion recognition model in the TensorRT platform based on the model parameters, the step further includes: Serializing the target inference engine to obtain target binary data; The target binary data is saved.
6. A method for identifying a lesion, characterized in that: The lesion identification method comprises: Acquire target medical images; Inputting the target medical image into the target reasoning engine and outputting a lesion recognition result; Wherein, the target reasoning engine is obtained based on the deployment method of the lesion recognition model described in any one of claims 1-5.
7. The method for identifying lesions according to claim 6, characterized in that: After the step of inputting the target medical image into the target reasoning engine and outputting the lesion identification result, the following steps are further included: A connected domain analysis is performed on the lesion recognition result, connected domains in the lesion recognition result that are smaller than a preset connected range are eliminated, and a new lesion recognition result is updated.
8. A deployment system for a lesion recognition model, characterized in that: The deployment system comprises: A model acquisition module, used to acquire a lesion recognition model; A parameter export module, used for exporting model parameters of the lesion recognition model; An engine building module is used to build a target reasoning engine of the lesion recognition model in the TensorRT platform based on the model parameters.
9. A lesion identification system, characterized in that: The lesion recognition system comprises: An image acquisition module, used to acquire target medical images; A result output module, used to input the target medical image into the target reasoning engine and output a lesion recognition result; Wherein, the target reasoning engine is obtained based on the deployment system of the lesion recognition model described in claim 8.
10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and used to run on the processor, characterized in that: When the processor executes the computer program, the deployment method of the lesion identification model described in any one of claims 1 to 5 and / or the lesion identification method described in claim 6 or 7 are implemented.