A self-training system for ultrasound artificial intelligence models

Abstract

The present invention provides a self-training system for an ultrasonic artificial intelligence model, which relates to the technical field of model self-training, and includes: an ultrasonic scanning device, which has a factory artificial intelligence model integrated therein; a self-training system for hospital staff to input multiple local medical images obtained by scanning the ultrasonic scanning device for annotation to obtain multiple medical annotated images; save the factory model network and multiple pre-configured model networks corresponding to the factory artificial intelligence model; call each model network according to the model building instruction input by the hospital staff to update the factory model network to obtain an updated model network, and train the updated model network according to each medical annotated image to obtain an optimized model; configure the optimized model to the corresponding ultrasonic scanning device to replace the factory artificial intelligence model. The beneficial effect is that hospital staff can easily train an artificial intelligence model that meets the personalized needs of the hospital and has high accuracy based on local medical images.

Description

Technical Field

[0001] The present invention relates to the technical field of model self-training, and in particular to a self-training system of an ultrasonic artificial intelligence model. Background Art

[0002] Ultrasound is the most commonly used inspection equipment at the grassroots level. Compared with CT, X-ray and other detection methods, ultrasound diagnosis has the characteristics of real-time scanning, non-invasive, radiation-free, low price, and is mobile and flexible in scanning. With the rapid development of AI technology, AI data analysis and deep learning technology are also working to make medical equipment smarter and more accurate.

[0003] In the prior art, a certain ultrasound-related AI function is generally provided directly to hospitals as a function of ultrasound equipment. This cannot solve the problem that when hospitals in different regions or hospitals with different patient distributions use the same ultrasound AI model, the recognition accuracy varies. Since hospitals cannot customize or modify the relevant parameters of the AI ​​function, or retrain or optimize the AI ​​model based on the ultrasound images scanned by their own hospitals, when the model recognition accuracy is low, they need to wait for the equipment manufacturer that provides the AI ​​model to optimize it. However, when equipment manufacturers optimize the AI ​​models they provide, first of all, they cannot obtain ultrasound images of various hospitals. Secondly, even if they obtain ultrasound images scanned by many different hospitals, they cannot effectively train an AI model that is suitable for all hospitals. At the same time, the hospital waits for a long time for the equipment manufacturer to provide the optimized AI model, which reduces the efficiency of the use of the equipment's AI function and thus reduces the use value of the equipment. Summary of the invention

[0004] In view of the problems existing in the prior art, the present invention provides a self-training system for an ultrasonic artificial intelligence model, comprising:

[0005] Ultrasonic scanning equipment, wherein a factory artificial intelligence model is integrated inside the ultrasonic scanning equipment;

[0006] A self-training system is connected to the ultrasonic scanning device, and the self-training system includes:

[0007] At least one data input port, for hospital staff to input a plurality of medically annotated images obtained by annotating a plurality of local medical images obtained by scanning the corresponding ultrasound scanning device, and save each of the medically annotated images as a training sample to a first storage module;

[0008] A second storage module is used to store the factory model network corresponding to the factory artificial intelligence model and multiple pre-configured model networks;

[0009] A model training module, connected to the first storage module and the second storage module respectively, for calling each of the model networks to update the factory model network to obtain an updated model network according to the model building instruction input by the hospital staff, and training the updated model network according to each of the medically annotated images to obtain an optimized model;

[0010] At least one model configuration port is connected to the model training module and is used to configure the optimized model into the corresponding ultrasonic scanning device to replace the factory artificial intelligence model.

[0011] Preferably, the self-training system also includes an image processing module, which is respectively connected to the data input port and the first storage module, and the image processing module provides multiple image processing services for pre-training processing of each of the medically annotated images, and saves each of the medically annotated images after pre-training processing as the training sample to the first storage module.

[0012] Preferably, the model network includes a plurality of basic model networks and a plurality of model sub-networks;

[0013] Then the model training module calls any of the basic model networks according to the model building instruction, and calls each of the model sub-networks to form the updated model network with the called basic model network as the base network;

[0014] Or the model training module calls each of the model sub-networks according to the model building instruction, and uses the factory model network as the basic network to build the updated model network.

[0015] Preferably, the self-training system also includes a visualization window connected to the model training module, which is used to provide a visual operation interface for the construction process of the updated model network.

[0016] Preferably, the model training module includes a rule verification unit, which is used to verify in real time whether the connection relationship between the input end of the called model sub-network and the corresponding output end of the connected basic network, as well as between the output end of the called model sub-network and the corresponding input end of the connected basic network, satisfies the preset connection rules during the process of constructing the updated model network, and give a corresponding prompt in the visualization window when any of the connection relationships does not satisfy the connection rules.

[0017] Preferably, the self-training system also includes a remote communication module for allowing the hospital staff to communicate remotely with the model training expert to assist the hospital staff in constructing the updated model network.

[0018] Preferably, the self-training system also includes a data synchronization port, which is respectively connected to the second storage module and the model training module, and is used to synchronize the factory artificial intelligence model and the corresponding optimization model to a remote server for the developer of the factory artificial intelligence model to view and optimize.

[0019] Preferably, the self-training system also includes a data receiving port for receiving the model optimization solution given by the development manufacturer based on the factory artificial intelligence model and the corresponding optimization model and fed back by the remote server, so as to assist the hospital staff in optimizing the factory artificial intelligence model.

[0020] Preferably, the self-training system is integrated into the corresponding ultrasonic scanning device, or the self-training system is configured in a host computer connected to the ultrasonic scanning device.

[0021] The above technical solution has the following advantages or beneficial effects:

[0022] 1) Through the self-training system of this technical solution, hospital staff can easily train an artificial intelligence model that meets the personalized needs of the hospital and has high accuracy based on the local medical images obtained by the ultrasound scanning equipment. Equipment manufacturers do not need to provide hospitals with trained artificial intelligence models, which effectively solves the problem that it is difficult for equipment manufacturers to collect ultrasound scanning samples from various hospitals when optimizing ultrasound artificial intelligence models. It also solves the problem that it is impossible to effectively train an ultrasound artificial intelligence model that is well applicable to all hospitals. At the same time, it reduces the time for each hospital to wait for the optimization of the ultrasound artificial intelligence model, and improves the efficiency of the use of the equipment's artificial intelligence function;

[0023] 2) Providing a visual operation interface for model construction, effectively reducing the difficulty of operation for hospital staff and reducing their dependence on professional expertise in model construction;

[0024] 3) The self-trained model can be synchronized to the remote server of the AI ​​model developer, providing a data source for the developer to optimize the model;

[0025] 4) The self-training system of this technical solution does not rely on the network environment when in use, and is suitable for use scenarios where the hospital cannot connect to the external network. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 The figure is a schematic diagram of the structure of a self-training system of an ultrasonic artificial intelligence model in a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0027] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. The present invention is not limited to this embodiment, and other embodiments may also fall within the scope of the present invention as long as they conform to the gist of the present invention.

[0028] In a preferred embodiment of the present invention, based on the above problems existing in the prior art, a self-training system for an ultrasonic artificial intelligence model is provided. Figure 1 As shown, including:

[0029] Ultrasonic scanning device 1, the ultrasonic scanning device 1 has a factory-made artificial intelligence model integrated therein;

[0030] The self-training system 2 is connected to the ultrasonic scanning device 1, and the self-training system 2 includes:

[0031] At least one data input port 21, for hospital staff to input multiple medically annotated images corresponding to multiple local medical images obtained by scanning with corresponding ultrasound scanning equipment, and save each medically annotated image as a training sample to a first storage module 22;

[0032] The second storage module 23 is used to store the factory model network corresponding to the factory artificial intelligence model and multiple pre-configured model networks;

[0033] The model training module 24 is connected to the first storage module 22 and the second storage module 23 respectively, and is used to call each model network to update the factory model network to obtain an updated model network according to the model building instruction input by the hospital staff, and train the updated model network according to each medical annotated image to obtain an optimized model;

[0034] At least one model configuration port 25 is connected to the model training module 24 and is used to configure the optimized model into the corresponding ultrasonic scanning equipment to replace the factory artificial intelligence model.

[0035] Specifically, in this embodiment, the self-training system 2 can be integrated into the corresponding ultrasonic scanning device 1. At this time, the self-training system 2 corresponds to the ultrasonic scanning device 1 one by one. However, since the self-training system 2 requires a larger storage space and the model training also has higher requirements on computing power, if the ultrasonic scanning device 1 cannot meet the self-training requirements, the self-training system 2 can also be configured in a host computer connected to the ultrasonic scanning device 1.

[0036] When the self-training system 2 is configured on a host computer, one host computer can meet the self-training needs of multiple ultrasound scanning devices 1 in the hospital, that is, the local medical images scanned by multiple ultrasound scanning devices 1, including but not limited to ultrasound scanning images and videos, are obtained. The hospital ultrasound diagnosis doctor or diagnosis expert issues an ultrasound diagnosis report based on the local medical images, and then the local medical images are annotated according to the ultrasound diagnosis report. The annotated content can be the diagnosis result or the lesion area. The annotated medical images obtained after annotation can be input into the self-training system through the corresponding data input port 21 as training samples.

[0037] In a preferred embodiment of the present invention, the self-training system 2 also includes an image processing module 26, which is respectively connected to the data input port 21 and the first storage module 22. The image processing module 26 provides multiple image processing services for pre-training processing of each medically annotated image, and saves each medically annotated image after pre-training processing as a training sample to the first storage module 22.

[0038] Specifically, in this embodiment, the above-mentioned image processing services include but are not limited to image cropping services, image rotation services, image filtering services, image brightness and chromaticity enhancement services. The above-mentioned services can be called independently, and the above-mentioned image processing services can also be a combination of the above-mentioned services, which further facilitates hospital staff to use pre-training processing of various medical annotated images.

[0039] Furthermore, the self-training system 2 is also configured with a second storage module 23 for storing the factory model network corresponding to the factory artificial intelligence model and multiple pre-configured model networks. By configuring the factory model network, it is convenient for hospital staff to update the model network based on the model network.

[0040] In a preferred embodiment of the present invention, the model network includes a plurality of basic model networks and a plurality of model sub-networks;

[0041] Then the model training module 24 calls any basic model network according to the model building instruction, and calls each model sub-network to form an updated model network with the called basic model network as the basic network;

[0042] Or the model training module 24 calls each model sub-network according to the model construction instruction, and uses the factory model network as the basic network to form an updated model network.

[0043] Specifically, in this embodiment, the above-mentioned basic model network includes but is not limited to convolutional neural networks, recurrent neural networks, and deep belief networks, and the above-mentioned model subnetworks include but are not limited to convolutional blocks, upsampling blocks, and downsampling blocks. When constructing the updated model network, it can be constructed on the basis of the existing factory model network according to demand, or the factory model network can be directly abandoned and constructed on the basis of the basic model network. It can be understood that for hospital staff with relatively high model construction expertise, it is also possible to directly construct according to each model subnetwork without calling the basic network first, providing a flexible construction scheme for the updated model network. Preferably, when constructing the updated model network on the basis of the basic network, the construction operation includes but is not limited to expanding the number of convolutional layers, pooling layers, etc. of the basic network, or deleting the number of convolutional layers and pooling layers, or adjusting the parameters of the convolutional layers and pooling layers, or replacing a certain network module in the basic network.

[0044] In a preferred embodiment of the present invention, the self-training system 2 also includes a visualization window 27, which is connected to the model training module 24 and is used to provide a visual operation interface for the construction process of the updated model network.

[0045] Specifically, in this embodiment, by providing a visualization window 27, hospital staff can construct an updated model network by dragging and dropping without editing codes, etc., thereby further reducing the professional dependence of hospital staff.

[0046] In a preferred embodiment of the present invention, the model training module 24 includes a rule verification unit 241, which is used to verify in real time whether the connection relationship between the input end of the called model sub-network and the corresponding output end of the connected basic network, and between the output end of the called model sub-network and the corresponding input end of the connected basic network, meets the preset connection rules during the process of constructing the updated model network, and give a corresponding prompt in the visualization window when any connection relationship does not meet the connection rule.

[0047] Specifically, in this embodiment, by configuring the connection rules, obvious construction errors can be effectively avoided during the construction of the updated model network, so as to effectively reduce the subsequent training time that may be occupied by the defective updated model network, thereby improving the success rate of model construction and the efficiency of model training.

[0048] In a preferred embodiment of the present invention, the self-training system 2 also includes a remote communication module 28 for allowing hospital staff to communicate remotely with model training experts to assist hospital staff in building an updated model network.

[0049] Specifically, in this embodiment, the remote communication module 28 enables hospital staff to receive real-time guidance from remote model training experts during the construction of the updated model network, thereby improving the efficiency of model optimization.

[0050] In a preferred embodiment of the present invention, the self-training system 2 also includes a data synchronization port 29, which is respectively connected to the second storage module 23 and the model training module 24, and is used to synchronize the factory artificial intelligence model and the corresponding optimization model to a remote server 3 for the developer of the factory artificial intelligence model to view and optimize.

[0051] In a preferred embodiment of the present invention, the self-training system 2 also includes a data receiving port 20, which is used to receive the model optimization solution given by the developer based on the factory artificial intelligence model and the corresponding optimization model fed back by the remote server, so as to assist hospital staff in optimizing the factory artificial intelligence model.

[0052] The above description is only a preferred embodiment of the present invention, and does not limit the implementation mode and protection scope of the present invention. For those skilled in the art, it should be aware that all solutions obtained by equivalent substitutions and obvious changes made using the contents of this specification and illustrations should be included in the protection scope of the present invention.

Claims

1. A self-training system for an ultrasonic artificial intelligence model, characterized in that: include: Ultrasonic scanning equipment, wherein a factory artificial intelligence model is integrated inside the ultrasonic scanning equipment; A self-training system is connected to the ultrasonic scanning device, and the self-training system includes: At least one data input port, for hospital staff to input a plurality of medically annotated images obtained by annotating a plurality of local medical images obtained by scanning the corresponding ultrasound scanning device, and save each of the medically annotated images as a training sample to a first storage module; A second storage module is used to store the factory model network corresponding to the factory artificial intelligence model and multiple pre-configured model networks; A model training module, connected to the first storage module and the second storage module respectively, for calling each of the model networks to update the factory model network to obtain an updated model network according to the model building instruction input by the hospital staff, and training the updated model network according to each of the medically annotated images to obtain an optimized model; At least one model configuration port, connected to the model training module, for configuring the optimization model into the corresponding ultrasonic scanning device to replace the factory artificial intelligence model; The model network includes multiple basic model networks and multiple model sub-networks; the model training module calls any of the basic model networks according to the model building instruction, and calls each of the model sub-networks to form the updated model network with the called basic model network as the basic network; Or the model training module calls each of the model sub-networks according to the model building instruction, and forms the updated model network based on the factory model network as the basic network; The self-training system also includes a visualization window, connected to the model training module, for providing a visualization operation interface for the construction process of the updated model network; The model training module includes a rule verification unit, which is used to verify in real time whether the connection relationship between the input end of the called model sub-network and the corresponding output end of the connected basic network, as well as between the output end of the called model sub-network and the corresponding input end of the connected basic network, meets the preset connection rules during the process of constructing the updated model network, and give a corresponding prompt in the visualization window when any of the connection relationships does not meet the connection rules.

2. The self-training system according to claim 1, characterized in that: The self-training system also includes an image processing module, which is respectively connected to the data input port and the first storage module. The image processing module provides multiple image processing services for pre-training processing of each of the medically annotated images, and saves each of the medically annotated images after pre-training processing as the training sample to the first storage module.

3. The self-training system according to claim 1, characterized in that: The self-training system also includes a remote communication module for allowing the hospital staff to communicate remotely with the model training experts to assist the hospital staff in constructing the updated model network.

4. The self-training system according to claim 1, characterized in that: The self-training system also includes a data synchronization port, which is respectively connected to the second storage module and the model training module, and is used to synchronize the factory artificial intelligence model and the corresponding optimization model to a remote server for the developer of the factory artificial intelligence model to view and optimize.

5. The self-training system according to claim 4, characterized in that: The self-training system also includes a data receiving port for receiving the model optimization solution given by the development manufacturer based on the factory artificial intelligence model and the corresponding optimization model, which is fed back by the remote server, so as to assist the hospital staff in optimizing the factory artificial intelligence model.

6. The self-training system according to claim 1, characterized in that: The self-training system is integrated into the corresponding ultrasonic scanning device, or the self-training system is configured in a host computer connected to the ultrasonic scanning device.

Images