Ultrasonic image recognition method based on reinforcement learning and BI-RADS guide

Abstract

The invention relates to an ultrasonic image recognition method based on reinforcement learning and a BI-RADS guide. The method comprises the steps that firstly, an ultrasonic image is acquired and preprocessed, and a focus area annotation image is obtained; then, according to a BI-RADS guide, feature extraction is carried out on the focus area labeling image, and malignant visual features are obtained; then, an ultrasonic image recognition model is trained based on the malignant visual features, and the ultrasonic image recognition model is set based on a near-end strategy optimization algorithm or a deep Q learning algorithm; and finally, inputting a to-be-recognized ultrasonic image into the trained ultrasonic image recognition model, and determining the type of the ultrasonic image. According to the method, a reinforcement learning algorithm is combined with feature extraction of medical image data, lesions are classified according to a BI-RADS (breast image report and data system) guide, ultrasonic image data classification is automatically carried out, and the method has the remarkable advantages of improving diagnosis accuracy, reducing time and cost and reducing doctor dependence.

Description

Technical Field

[0001] The present application relates to the field of image processing technology, and in particular to an ultrasound image recognition method based on reinforcement learning and BI-RADS guidelines. Background Art

[0002] Currently, traditional ultrasound diagnosis relies primarily on manual analysis and judgment by physicians. However, this diagnostic method has several significant limitations: First, the challenge of insufficient clinical experience. Young physicians, due to their limited clinical experience, may experience errors in image recognition and lesion identification. This error often leads to suboptimal diagnostic accuracy, especially in complex cases, resulting in the risk of missed or misdiagnosed cases, impacting patient treatment outcomes and prognosis. Second, the high cost of diagnosis. Ultrasound diagnosis is relatively expensive, encompassing not only equipment maintenance and operating costs but also physician training. Experienced ultrasound specialists are particularly scarce in some regions, leading to a long and costly training cycle for physicians, increasing the overall burden of diagnostic services. Third, the diagnostic process is lengthy. Traditional manual diagnosis often requires a long time. The entire process, from image acquisition and analysis to the final diagnosis, is cumbersome, especially when large numbers of cases or images must be compared one by one. This situation makes it difficult to meet the urgent clinical need for rapid diagnosis and timely treatment, hindering the effectiveness of early disease detection and treatment.

[0003] Therefore, in the relevant technology, there is an urgent need for a method that can solve the problems of high dependence on operators and high rates of missed and misdiagnosis during breast ultrasound screening. Summary of the Invention

[0004] Based on this, it is necessary to provide an ultrasound image recognition method based on reinforcement learning and BI-RADS guidelines that can solve the problems of high operator dependence and high misdiagnosis rate during breast ultrasound screening, in response to the above technical problems.

[0005] In a first aspect, the present application provides an ultrasound image recognition method based on reinforcement learning and BI-RADS guidelines. The method comprises: Acquire and preprocess the ultrasound image to obtain an annotated image of the lesion area; Extract features from the annotated image of the lesion area according to the BI-RADS guidelines to obtain malignant visual features; Training an ultrasound image recognition model based on the malignant visual features, wherein the ultrasound image recognition model is set based on a proximal strategy optimization algorithm or a deep Q learning algorithm; The ultrasound image to be identified is input into the trained ultrasound image recognition model to determine the ultrasound image category.

[0006] Optionally, in one embodiment of the present application, the preprocessing includes: determining a malignant feature label of the ultrasound image according to the diagnosis report and pathological results; The data format of the ultrasound image is converted and image segmentation is performed.

[0007] Optionally, in one embodiment of the present application, extracting features from the lesion area annotated image according to the BI-RADS guidelines includes: Gradient variance is used to determine the edge blur feature vector; The angular eigenvector is determined using profile curvature analysis; Radon transform is used to detect linear structures and determine burr feature vectors; Determine differential leaf eigenvectors based on regional convex defects.

[0008] Optionally, in one embodiment of the present application, the training of the ultrasound image recognition model based on the malignant visual features includes: defining an agent state space based on the malignant visual features; Define the agent's action space, including benign, malignant, and requiring further inspection; Define a reward function to indicate the accuracy of ultrasound image classification.

[0009] Optionally, in one embodiment of the present application, the training of the ultrasound image recognition model based on the malignant visual features further comprises: Interact based on the current strategy to collect state, action, and reward data; Calculate the advantage function and update the strategy; Limit the policy update amplitude and repeat the iteration until the policy converges.

[0010] Optionally, in one embodiment of the present application, the training of the ultrasound image recognition model based on the malignant visual features further comprises: Randomly initialize the parameters of the Q network and the target network; use -The greedy strategy interacts with the environment, collects state, action, reward, and next state storage data, and stores them in the replay buffer; Calculate the loss and update the Q network; The parameters of the Q network are copied to the target network every fixed number of steps.

[0011] In a second aspect, the present application also provides an ultrasound image recognition device based on reinforcement learning and BI-RADS guidelines. The device comprises: The data collection module is used to acquire and pre-process ultrasound images to obtain annotated images of the lesion area; A feature extraction module is used to extract features from the lesion area annotated image according to the BI-RADS guidelines to obtain malignant visual features; A model training module, configured to train an ultrasound image recognition model based on the malignant visual features, wherein the ultrasound image recognition model is set based on a proximal strategy optimization algorithm or a deep Q learning algorithm; The ultrasonic image recognition module is used to input the ultrasonic image to be recognized into the trained ultrasonic image recognition model to determine the ultrasonic image category.

[0012] In a third aspect, the present application further provides a computer device comprising a memory and a processor, wherein the memory stores a computer program and the processor executes the steps of the method described in each of the above embodiments.

[0013] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in each of the above embodiments.

[0014] The ultrasound image recognition method based on reinforcement learning and BI-RADS guidelines first acquires and preprocesses an ultrasound image to obtain an annotated image of the lesion area. Feature extraction is then performed on the annotated image of the lesion area according to BI-RADS guidelines to obtain visual features of malignancy. An ultrasound image recognition model is then trained based on these visual features. The ultrasound image recognition model is based on a proximal strategy optimization algorithm or a deep Q-learning algorithm. Finally, the ultrasound image to be recognized is input into the trained ultrasound image recognition model to determine the ultrasound image category. This method, which utilizes a reinforcement learning algorithm combined with feature extraction of medical imaging data and lesion classification based on BI-RADS (Breast Imaging Reporting and Data System), automatically classifies ultrasound image data, offering significant advantages in improving diagnostic accuracy, reducing time and costs, and reducing physician reliance. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 FIG1 is a diagram showing an application environment of an ultrasound image recognition method based on reinforcement learning and BI-RADS guidelines in one embodiment; Figure 2 1 is a flowchart of an ultrasound image recognition method based on reinforcement learning and BI-RADS guidelines in one embodiment; Figure 3 Schematic diagram of a model training process in one embodiment; Figure 4 is a structural block diagram of an ultrasound image recognition device based on reinforcement learning and BI-RADS guidelines in one embodiment; Figure 5FIG. 1 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0017] The ultrasound image recognition method based on reinforcement learning and BI-RADS guidelines provided in the embodiment of the present application can be applied to Figure 1 In the application environment shown, the terminal communicates with the server through the network. The data storage system can store data that the server needs to process. The data storage system can be integrated on the server or placed on the cloud or other network servers. The terminal can be, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can be smart speakers, smart TVs, smart air conditioners, smart car devices, etc. Portable wearable devices can be smart watches, smart bracelets, head-mounted devices, etc. The server can be implemented as a standalone server or a server cluster consisting of multiple servers.

[0018] In one embodiment, Figure 2 As shown in the figure, an ultrasound image recognition method based on reinforcement learning and BI-RADS guidelines is provided. Figure 1 The following steps are used as an example to illustrate the server in the example: S201: Acquire an ultrasound image and perform preprocessing to obtain a labeled image of the lesion area.

[0019] In the embodiment of the present application, first, breast ultrasound images are collected and preprocessed to obtain labeled images of lesion areas. The labels include classification labels such as benign and malignant.

[0020] Specifically, in one embodiment of the present application, the preprocessing includes: S301: Determine the malignant feature label of the ultrasound image according to the diagnosis report and pathological results.

[0021] S303: Convert the data format of the ultrasound image and perform image segmentation.

[0022] In one embodiment of the present application, collected breast ultrasound images, which are DICOM data of breast tumors, are labeled with malignant features based on diagnostic reports and pathological findings, including blur, angulation, burrs, and microlobulation. The DICOM data is then converted to NII format, and the images are segmented using ITKSNAP. Ultrasound images contain various breast tissue structures, but the key to diagnosis lies in the lesion (which may be benign or malignant). Image segmentation can accurately separate this potential lesion from irrelevant information such as surrounding normal tissue and background noise. For example, features such as edge blur, angulation, burrs, and microlobulation are used to characterize the morphological characteristics of the lesion. Without first performing segmentation to isolate the lesion region, it is impossible to accurately calculate and evaluate these lesion-specific features. Extracting features from the entire image or the wrong region is meaningless and may even be misleading. Using the segmented lesion area as input can reduce the data volume and computational complexity of subsequent processing (such as feature extraction and reinforcement learning model training), and eliminate interference from irrelevant areas, which helps to improve the accuracy and robustness of the diagnostic model.

[0023] S203: Extract features from the annotated image of the lesion area according to the BI-RADS guidelines to obtain malignant visual features.

[0024] In an embodiment of the present application, malignant visual features in images, including blurring, angulation, burrs, and microlobulation, are extracted according to the BI-RADS (Breast Imaging-Reporting and Data System) guidelines. Specifically, malignant visual features observed in the segmented lesion region (ROI) that meet the BI-RADS definition (e.g., edge blurring, angulation, burrs, and microlobulation) are converted into a quantitative, machine-readable format. Algorithmic analysis of image data (e.g., lesion contours or texture information) is performed to determine the presence of these predefined features and generate corresponding numerical or vectorized representations.

[0025] Specifically, in one embodiment of the present application, extracting features from the lesion area annotated image according to the BI-RADS guidelines includes: S401: Determine the edge blur feature vector using gradient variance.

[0026] S403: Determine the angular feature vector using contour curvature analysis.

[0027] S405: Detect the linear structure using Radon transform to determine the burr feature vector.

[0028] S407: Determine a differential leaf eigenvector based on the regional convex defect.

[0029] In one embodiment of the present application, gradient variance is used to extract edge blur features, including gradient calculation and variance calculation, wherein the gradient of the image is calculated using an edge detection algorithm (such as the Sobel operator), and the formula is as follows:

[0030] in, is the image I at pixel position The gray value at and Respectively represent the images in and Gradient in direction.

[0031] Gradient variance is used to measure the blurriness of the edge. In the lesion area The gradient value of each pixel, is the average gradient value of the area. The gradient variance calculation formula is:

[0032] in: It is The gradient value of each pixel, is the mean of all pixel gradient values, that is, , is the total number of pixels within the lesion area.

[0033] At the same time, the angular features are extracted by analyzing the contour curvature of the lesion area, including contour extraction, curvature calculation and angle feature extraction. The edge (contour) of the lesion area is extracted by Canny edge detection or other image segmentation methods. Assume that C(t) is the point set of the edge contour, where t is the parameter on the contour. Curvature It reflects the curvature of the contour, and the calculation formula is:

[0034] in, are the coordinates of the contour points, are the parameters on the contour, is the slope of the tangent line to the contour, is the curvature of the contour.

[0035] After curvature calculation, angular features can be extracted by finding points with large curvature values. Angled features are usually edges with large angles, which are more common in malignant tumors.

[0036] Burr features usually appear as radial linear structures, which can be detected by Radon transform. Radon transform converts the image from the spatial domain to the projection space and calculates the projection at different angles. Let the original image be , the projection after Radon transformation is expressed as:

[0037] in, is the angle of projection, is the offset of the projection, is the Dirac function, which is used to calculate the projection value along a certain angle.

[0038] In the Radon transform results, burr features appear as multiple radial lines, which can be extracted by analyzing the peaks of the projections. Burr features can be represented as a set of linear structures, including angle, number, and intensity.

[0039] Microlobulation is a hallmark of malignant tumors, usually manifested as irregularities at the edge of the lesion. These features are extracted through the convex defects of the region. First, the lesion region is segmented to obtain the ROI (Region of Interest) of the lesion. Suppose the lesion region is , whose boundary point set is ,in is the index of all boundary points. After that, convex defect detection is performed by calculating the convex hull of the lesion area , and compared with the original boundary. The convex defect refers to the part of the original boundary that exceeds the convex hull. The calculation of the convex hull can be achieved by Graham scanning or other convex hull algorithms. The calculation is:

[0040] in, is the boundary point set, convex hull is the smallest convex polygon that encloses all points.

[0041] Finally, feature quantization is performed by calculating the difference between the original boundary and the convex hull (convex defect), and quantizing the area, shape and other information of the defect area into numerical features to form a differential leaf feature vector.

[0042] S205: Training an ultrasound image recognition model based on the malignant visual features, wherein the ultrasound image recognition model is set based on a proximal strategy optimization algorithm or a deep Q learning algorithm.

[0043] In an embodiment of the present application, an ultrasound image recognition model is trained based on malignant visual features. The ultrasound image recognition model is an intelligent agent set up based on the Proximal Policy Optimization (PPO) algorithm or the Deep Q-Learning (DQN) algorithm, which continuously optimizes and adjusts the model's behavioral decisions through interaction with the environment and feedback (such as rewards or penalties).

[0044] Specifically, in one embodiment of the present application, the training of the ultrasound image recognition model based on the malignant visual features includes: S501: defining an agent state space based on the malicious visual features.

[0045] S503: Define the agent action space, including benign, malignant, and need for further inspection.

[0046] S505: Define a reward function to indicate the accuracy of ultrasound image classification.

[0047] In one embodiment of the present application, Figure 3 As shown, the agent state space (State) is defined based on the malignant visual features, including blur, angle, burr, and differential leaf information, in the form of , where each To extract BI-RADS quantitative features, we define the action space (Action), including possible classification decisions, such as "benign", "malignant", "need further examination", etc., in the form of . At the same time, define the reward function (RewardFunction), based on the accuracy and timeliness of classification, gives different rewards or penalties. For example, accurately classified as positive samples are given positive rewards, and incorrect classifications are given negative rewards. By initializing the policy network or Q Network , using the current policy to interact with the environment, collecting (state, action, reward, next state) quadruplets, storing the experience in a replay pool (DQN) or sampled policy gradient (PPO), iteratively optimizing policy parameters to increase cumulative reward. For policy optimization problems involving complex continuous actions, the Proximal Policy Optimization (PPO) algorithm can be used, while for simple tasks in a discrete action space, the Deep Q-Learning (DQN) algorithm can be used.

[0048] In one embodiment of the present application, the training of the ultrasound image recognition model based on the malignant visual features further comprises: S601: Interact based on the current strategy and collect state, action and reward data.

[0049] S603: Calculate the advantage function and update the strategy.

[0050] S605: Limit the strategy update range and repeat the iteration until the strategy converges.

[0051] In one embodiment of the present application, the Proximal Policy Optimization (PPO) algorithm is a deep reinforcement learning algorithm that belongs to the policy gradient method and directly maximizes the cumulative reward by optimizing the policy. The training process is as follows: First, use the current policy Interact and collect state, action and reward data. After that, use temporal difference method (TD) or generalized advantage estimation (GAE) to calculate the advantage function . Then, based on the objective function Update the parameters. By limiting the policy update amplitude, the iteration is repeated until the policy converges. PPO solves the instability and low efficiency problems of the policy gradient method. Its main goal is to avoid training instability caused by drastic changes by limiting the policy update amplitude. To this end, PPO introduces a clipped probability ratio objective function, which is expressed as:

[0052]

[0053] in, is the probability ratio of the new strategy to the old strategy, Advantage Function is used to measure the quality of an action. Restricting the probability ratio to The objective function uses the min operation to ensure that the update does not deviate significantly from the old policy.

[0054] In one embodiment of the present application, the training of the ultrasound image recognition model based on the malignant visual features further comprises: S701: Randomly initialize the parameters of the Q network and the target network.

[0055] S703: Use -The greedy strategy interacts with the environment, collects state, action, reward, and next state storage data, and stores them in the replay buffer.

[0056] S705: Calculate the loss and update the Q network.

[0057] S707: Copy the parameters of the Q network to the target network every fixed number of steps.

[0058] In one embodiment of the present application, the Deep Q-Learning (DQN) algorithm is a deep reinforcement learning algorithm that belongs to the value function method and selects the optimal action by estimating the Q-value function. The training process is as follows: first, the parameters of the Q network and the target network are randomly initialized, and then, the training is performed using -The greedy strategy interacts with the environment, collects state, action, reward, and next state storage data, and stores them in the replay buffer. After that, it calculates the loss and updates the Q network. After that, the parameters of the Q network are copied to the target network every fixed number of steps. The loss function is expressed as follows:

[0059]

[0060] in, is the target Q value, are the parameters of the current Q network, is the discount factor used to calculate the present value of future rewards, are the parameters of the target network.

[0061] The Deep Q-Learning (DQN) algorithm improves training stability by introducing experience replay, a target network, and reward clipping. Specifically, experience replay stores interaction data in a replay buffer and randomly samples small batches of data for training to break data correlation. A fixed target network reduces target value fluctuations and stabilizes training. Reward clipping within a fixed range prevents exploding or vanishing gradients.

[0062] S207: Input the ultrasound image to be recognized into the trained ultrasound image recognition model to determine the category of the ultrasound image.

[0063] In an embodiment of the present application, the ultrasound image to be identified is input into a trained ultrasound image recognition model, and the intelligent agent automatically classifies it as benign or malignant, generates a classification report, and displays the results to provide a basis for further diagnosis and treatment.

[0064] In the aforementioned ultrasound image recognition method based on reinforcement learning and BI-RADS guidelines, an ultrasound image is first acquired and preprocessed to obtain an annotated image of the lesion area. Feature extraction is then performed on the annotated image of the lesion area according to BI-RADS guidelines to obtain visual features of malignancy. An ultrasound image recognition model is then trained based on these visual features of malignancy. The ultrasound image recognition model is configured using a proximal strategy optimization algorithm or a deep Q-learning algorithm. Finally, the ultrasound image to be recognized is input into the trained ultrasound image recognition model to determine the ultrasound image category. In other words, the use of a reinforcement learning algorithm combined with feature extraction of medical imaging data and lesion classification according to BI-RADS (Breast Imaging Reporting and Data System) allows for automated ultrasound image data classification, offering significant advantages in improving diagnostic accuracy, reducing time and costs, and reducing physician reliance.

[0065] It should be understood that, although the various steps in the flowcharts involved in the various embodiments described above are displayed in sequence according to the instructions of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the various embodiments described above can include multiple steps or multiple stages, and these steps or stages are not necessarily executed and completed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of steps or stages in other steps.

[0066] Based on the same inventive concept, the embodiments of the present application also provide an ultrasound image recognition device based on reinforcement learning and BI-RADS guidelines for implementing the ultrasound image recognition method based on reinforcement learning and BI-RADS guidelines. The implementation solution provided by the device is similar to the implementation solution described in the above method. Therefore, the specific limitations in one or more embodiments of the ultrasound image recognition device based on reinforcement learning and BI-RADS guidelines provided below can be found in the above limitations of the ultrasound image recognition method based on reinforcement learning and BI-RADS guidelines, and will not be repeated here.

[0067] In one embodiment, Figure 4 As shown, an ultrasound image recognition device 400 based on reinforcement learning and BI-RADS guidelines is provided, comprising: a data collection module 401, a feature extraction module 403, a model training module 405 and an ultrasound image recognition module 407, wherein: The data collection module 401 is used to acquire and pre-process ultrasound images to obtain an annotated image of the lesion area.

[0068] The feature extraction module 403 is used to extract features from the lesion area annotated image according to the BI-RADS guidelines to obtain malignant visual features.

[0069] The model training module 405 is used to train an ultrasound image recognition model based on the malignant visual features, where the ultrasound image recognition model is set based on a proximal strategy optimization algorithm or a deep Q learning algorithm.

[0070] The ultrasound image recognition module 407 is used to input the ultrasound image to be recognized into the trained ultrasound image recognition model to determine the category of the ultrasound image.

[0071] In one embodiment of the present application, the data collection module is further configured to: determining a malignant feature label of the ultrasound image according to the diagnosis report and pathological results; The data format of the ultrasound image is converted and image segmentation is performed.

[0072] In one embodiment of the present application, the feature extraction module is further configured to: Gradient variance is used to determine the edge blur feature vector; The angular eigenvector is determined using profile curvature analysis; Radon transform is used to detect linear structures and determine burr feature vectors; Determine differential leaf eigenvectors based on regional convex defects.

[0073] In one embodiment of the present application, the model training module is further used to: defining an agent state space based on the malignant visual features; Define the agent's action space, including benign, malignant, and requiring further inspection; Define a reward function to indicate the accuracy of ultrasound image classification.

[0074] In one embodiment of the present application, the model training module is further used to: Interact based on the current strategy to collect state, action, and reward data; Calculate the advantage function and update the strategy; Limit the policy update amplitude and repeat the iteration until the policy converges.

[0075] In one embodiment of the present application, the model training module is further used to: Randomly initialize the parameters of the Q network and the target network; use -The greedy strategy interacts with the environment, collects state, action, reward, and next state storage data, and stores them in the replay buffer; Calculate the loss and update the Q network; The parameters of the Q network are copied to the target network every fixed number of steps.

[0076] Each module in the aforementioned ultrasound image recognition device based on reinforcement learning and BI-RADS guidelines can be implemented in whole or in part through software, hardware, or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in the computer device in the form of software, so that the processor can call and execute the corresponding operations of each module.

[0077] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 5 As shown. The computer device includes a processor, memory, a communication interface, a display screen, and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal via wired or wireless communication. The wireless communication can be achieved via Wi-Fi, a mobile cellular network, NFC (near-field communication), or other technologies. When executed by the processor, the computer program implements an ultrasound image recognition method based on reinforcement learning and BI-RADS guidelines. The display screen of the computer device can be a liquid crystal display or an electronic ink display. The input device of the computer device can be a touch layer covering the display screen, or keys, a trackball, or a touchpad provided on the computer device housing, or an external keyboard, touchpad, or mouse.

[0078] Those skilled in the art will understand that Figure 5 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0079] In one embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor implements the steps in the above method embodiments when executing the computer program.

[0080] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0081] In one embodiment, a computer program product is provided, including a computer program, which implements the steps in the above method embodiments when executed by a processor.

[0082] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0083] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned embodiments. In particular, any reference to memory, database, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The databases involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchains. The processors involved in the various embodiments provided herein may be, but are not limited to, general-purpose processors, central processing units (CPUs), graphics processing units (GPUs), digital signal processors (DSPs), programmable logic devices (PLDs), data processing logic devices based on quantum computing, and the like.

[0084] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0085] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.

Claims

1. An ultrasound image recognition method based on reinforcement learning and BI-RADS guidelines, characterized in that: The method comprises: Acquire and preprocess the ultrasound image to obtain an annotated image of the lesion area; Extract features from the annotated image of the lesion area according to the BI-RADS guidelines to obtain malignant visual features; Training an ultrasound image recognition model based on the malignant visual features, wherein the ultrasound image recognition model is set based on a proximal strategy optimization algorithm or a deep Q learning algorithm; The ultrasound image to be identified is input into the trained ultrasound image recognition model to determine the ultrasound image category.

2. The ultrasound image recognition method based on reinforcement learning and BI-RADS guidelines according to claim 1, characterized in that: The pretreatment includes: determining a malignant feature label of the ultrasound image according to the diagnosis report and pathological results; The data format of the ultrasound image is converted and image segmentation is performed.

3. The ultrasound image recognition method based on reinforcement learning and BI-RADS guidelines according to claim 1, characterized in that: The feature extraction of the lesion area annotated image according to the BI-RADS guidelines includes: Gradient variance is used to determine the edge blur feature vector; The angular eigenvector is determined using profile curvature analysis; Radon transform is used to detect linear structures and determine burr feature vectors; Determine differential leaf eigenvectors based on regional convex defects.

4. The ultrasound image recognition method based on reinforcement learning and BI-RADS guidelines according to claim 1, characterized in that: The training of the ultrasound image recognition model based on the malignant visual features comprises: defining an agent state space based on the malignant visual features; Define the agent's action space, including benign, malignant, and requiring further inspection; Define a reward function to indicate the accuracy of ultrasound image classification.

5. The ultrasound image recognition method based on reinforcement learning and BI-RADS guidelines according to claim 4, characterized in that: The training of the ultrasound image recognition model based on the malignant visual features further comprises: Interact based on the current strategy to collect state, action, and reward data; Calculate the advantage function and update the strategy; Limit the policy update amplitude and repeat the iteration until the policy converges.

6. The ultrasound image recognition method based on reinforcement learning and BI-RADS guidelines according to claim 4, characterized in that: The training of the ultrasound image recognition model based on the malignant visual features further comprises: Randomly initialize the parameters of the Q network and the target network; use -The greedy strategy interacts with the environment, collects state, action, reward, and next state storage data, and stores them in the replay buffer; Calculate the loss and update the Q network; The parameters of the Q network are copied to the target network every fixed number of steps.

7. An ultrasound image recognition device based on reinforcement learning and BI-RADS guidelines, characterized in that: The device comprises: The data collection module is used to acquire and pre-process ultrasound images to obtain annotated images of the lesion area; A feature extraction module is used to extract features from the lesion area annotated image according to the BI-RADS guidelines to obtain malignant visual features; A model training module, configured to train an ultrasound image recognition model based on the malignant visual features, wherein the ultrasound image recognition model is set based on a proximal strategy optimization algorithm or a deep Q learning algorithm; The ultrasonic image recognition module is used to input the ultrasonic image to be recognized into the trained ultrasonic image recognition model to determine the ultrasonic image category.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images