CT scanning parameter optimization method and system based on reinforcement learning

Abstract

The invention relates to the technical field of CT scanning parameter optimization, in particular to a CT scanning parameter optimization method and system based on reinforcement learning, and the method comprises the steps: obtaining basic information of a patient, scanning part information and initial CT scanning parameters; obtaining a CT image quality evaluation data set; constructing a reinforcement learning environment based on the basic information of the patient, the scanning part information and the initial CT scanning parameters; using the CT image quality evaluation data set to train a reinforcement learning model; according to the reinforcement learning model, optimized CT scanning parameters are generated; according to the method, the optimized CT scanning parameters are output and used for executing CT scanning, the system can generate customized optimal scanning parameters for each patient by taking the individual features of the patient as a part of a reinforcement learning environment, the accuracy and safety of CT examination are greatly improved, multi-target collaborative optimization is achieved, and the method is suitable for popularization and application. The optimal balance point can be found among a plurality of targets such as image quality, noise level and radiation dose.

Description

Technical Field

[0001] The present invention relates to the technical field of CT scanning parameter optimization, and more specifically, to a CT scanning parameter optimization method and system based on reinforcement learning. Background Art

[0002] Since its advent, computed tomography (CT) technology has played a huge role in the field of medical diagnosis. However, the optimization of CT scanning parameters has always been a thorny issue. Traditional methods mainly rely on the experience of radiologists and technicians to determine scanning parameters through preset scanning plans or manual adjustments. This method has obvious limitations: first, it is difficult to fully consider the individual differences of each patient, such as body shape, age and other factors; second, it is difficult to achieve the best balance between ensuring image quality and reducing radiation dose; third, with the continuous advancement of CT technology, the combination of scanning parameters has become more and more complex, and it is difficult to find the optimal solution based on manual experience alone.

[0003] In recent years, some researchers have tried to use machine learning methods to optimize CT scanning parameters. For example, some studies have proposed the use of neural networks to predict the optimal tube current settings, and some researchers have tried to use genetic algorithms to optimize multiple scanning parameters. These methods have indeed made progress in some aspects, but there are still some problems: first, most methods can only optimize a single or a few parameters, and it is difficult to achieve comprehensive parameter optimization; second, these methods usually require a large amount of labeled training data, which is often difficult to obtain in practice; third, many methods lack sufficient flexibility and are difficult to adapt to different CT equipment and clinical needs.

[0004] In addition, existing methods generally lack an effective clinical validation mechanism. Many algorithms perform well in laboratory environments, but are difficult to achieve the expected results in actual clinical applications. This is mainly because laboratory environments are difficult to fully simulate the complex and changeable situations in clinical settings, and there is a lack of a mechanism for continuous optimization and updating.

[0005] Considering the above problems, there is an urgent need for a new CT scanning parameter optimization method that can comprehensively consider multiple scanning parameters, adapt to the individual differences of different patients, achieve the best balance between image quality and radiation dose, and have good adaptability and scalability. Summary of the invention

[0006] The CT scanning parameter optimization method and system based on reinforcement learning proposed in the present invention are designed to solve the above technical problems. By introducing the reinforcement learning algorithm, the method can adaptively learn the optimal parameter adjustment strategy, overcoming the limitation of the traditional method that relies on manual experience. At the same time, the method considers the comprehensive optimization of multiple scanning parameters, which can minimize the radiation dose while ensuring image quality.

[0007] The present invention provides a method for optimizing CT scan parameters based on reinforcement learning, including:

[0008] An acquisition step, including:

[0009] Acquire the patient's basic information, scan site information, and initial CT scan parameters;

[0010] Acquire a CT image quality evaluation data set;

[0011] A processing step, including:

[0012] Based on the patient's basic information, scan site information, and initial CT scan parameters, construct a reinforcement learning environment;

[0013] Use the CT image quality evaluation data set to train a reinforcement learning model;

[0014] According to the reinforcement learning model, generate optimized CT scan parameters;

[0015] An output step, including:

[0016] Output the optimized CT scan parameters for performing a CT scan.

[0017] Preferably, the step of acquiring the CT image quality evaluation data set specifically includes:

[0018] Acquire multiple groups of CT images and their corresponding scan parameters;

[0019] Use a deep convolutional neural network to evaluate the quality of the CT images to obtain image quality indicators;

[0020] Combine the CT images, scan parameters, and image quality indicators to form CT image quality evaluation data.

[0021] Preferably, the deep convolutional neural network includes 5 convolutional pooling layers, 5 pooling layers, and 4 fully connected layers, where:

[0022] The activation function of the first 4 fully connected layers is the ReLU function;

[0023] The activation function of the last fully connected layer is the softmax function.

[0024] Preferably, the step of constructing the reinforcement learning environment specifically includes:

[0025] Use the patient's basic information and scan site information as the state space;

[0026] Use the adjustment range of the CT scan parameters as the action space;

[0027] Construct a reward function based on image quality, image noise, and radiation dose.

[0028] Preferably, the step of training the reinforcement learning model using the CT image quality evaluation dataset specifically includes:

[0029] Initialize the Q-value table and the Q'-value table;

[0030] Select an action based on the ε-greedy policy;

[0031] Execute the selected action to obtain a new state and a reward;

[0032] Update the Q-value table and the Q'-value table;

[0033] Repeat the above steps until the model converges.

[0034] Preferably, the step of updating the Q-value table and the Q'-value table adopts the Double Deep Q-Learning algorithm.

[0035] Preferably, it further includes a clinical verification step:

[0036] Obtain a clinical verification dataset;

[0037] Perform a CT scan using the optimized CT scan parameters to obtain a verification image;

[0038] Evaluate the quality, noise level, and radiation dose of the verification image;

[0039] According to the evaluation results, determine whether it is necessary to further optimize the reinforcement learning model.

[0040] Preferably, the step of evaluating the verification image specifically includes:

[0041] Segment the verification image using an image segmentation algorithm;

[0042] Calculate the signal-to-noise ratio (SNR) and the dose-area product (DAP) of the segmented image;

[0043] Compare the calculated SNR and DAP with a preset threshold.

[0044] Preferably, the basic patient information includes height, weight, age, and gender; the scan location information includes the head, chest, abdomen, and whole body.

[0045] A CT scan parameter optimization system based on reinforcement learning for executing the method includes:

[0046] A data acquisition module for obtaining basic patient information, scan location information, and initial CT scan parameters;

[0047] A data preprocessing module for constructing a CT image quality evaluation dataset;

[0048] A reinforcement learning module for:

[0049] Constructing a reinforcement learning environment;

[0050] Training a reinforcement learning model;

[0051] Generating optimized CT scan parameters;

[0052] A parameter output module for outputting the optimized CT scan parameters;

[0053] A clinical verification module for:

[0054] Performing a CT scan using the optimized CT scan parameters;

[0055] Evaluating the scan results;

[0056] Feeding back to the reinforcement learning module for further optimization according to the evaluation results.

[0057] The beneficial effects of the present invention are mainly reflected in the following aspects:

[0058] Firstly, it can achieve true personalized optimization. By taking the individual characteristics of patients as part of the reinforcement learning environment, the system can generate customized optimal scan parameters for each patient, greatly improving the accuracy and safety of CT examinations. Secondly, this method realizes the collaborative optimization of multiple objectives. Through a carefully designed reward function, the system can find the best balance among multiple objectives such as image quality, noise level, and radiation dose, which is difficult to achieve by traditional methods.

[0059] More importantly, the method of the present invention has the ability of continuous learning and self-optimization. By introducing a clinical verification mechanism, the system can continuously obtain feedback from actual applications and update and optimize the model accordingly. This dynamic optimization feature enables the system to continuously adapt to new clinical needs and technological progress and maintain long-term effectiveness.

[0060] In addition, the complete system solution proposed by the present invention realizes the full-process automation from data acquisition, preprocessing, parameter optimization to clinical verification through modular design. This not only greatly improves work efficiency but also reduces the possibility of human errors. The scalability and flexibility of the system also enable it to easily adapt to the needs and work processes of different hospitals.

[0061] Generally speaking, the present invention provides an innovative and comprehensive solution to the problem of CT scan parameter optimization. It can not only significantly improve the quality and safety of CT examinations but also is expected to promote the technological progress of the entire medical imaging field and bring better diagnosis and treatment experiences to patients. Brief Description of the Drawings

[0062] Figure 1 This is the system logic block diagram of the present invention.

[0063] Figure 2 This is the logic block diagram of the data preprocessing module of the present invention.

[0064] Figure 3 This is the logic block diagram of the reinforcement learning module of the present invention.

[0065] Figure 4 This is the logic block diagram of the clinical verification module of the present invention. Detailed Description of the Invention

[0066] Please refer to Figures 1-4 , the present invention provides a method and system for optimizing CT scan parameters based on reinforcement learning. This method adaptively optimizes CT scan parameters by using reinforcement learning algorithms to improve image quality, reduce radiation dose, and meet the personalized needs of different patients.

[0067] First, the method of the present invention includes an acquisition step, a processing step, and an output step. In the acquisition step, the system acquires patient basic information, scan site information, initial CT scan parameters, and a CT image quality evaluation data set. Patient basic information may include age, gender, height, weight, etc., and this information is crucial for determining appropriate scan parameters. For example, for pediatric patients, a lower radiation dose is usually required. Scan site information may include head, chest, abdomen, etc., and the scanning requirements for different sites are also different. Initial CT scan parameters are usually set based on empirical values but may not be optimal.

[0068] The acquisition of the CT image quality evaluation data set is an important innovation point of the present invention. This data set contains a large number of CT image samples, and each sample has corresponding scan parameters and image quality scores. This data set provides valuable training data for the subsequent reinforcement learning process.

[0069] In the processing step, the present invention first constructs a reinforcement learning environment based on the acquired information. This environment simulates the CT scan process, where the state space includes patient information and current scan parameters, and the action space is the possible parameter adjustments. The reward function is designed based on factors such as image quality, noise level, and radiation dose to balance these conflicting objectives.

[0070] Next, the system trains a reinforcement learning model using a CT image quality evaluation dataset. The present invention adopts an advanced Double Deep Q-Learning algorithm, which reduces the overestimation problem and improves the learning stability by maintaining two Q-value networks. During the training process, the model continuously tries different parameter combinations and updates its policy according to the obtained rewards, gradually learning the optimal parameter adjustment strategy.

[0071] Preferably, in an embodiment of the present invention, the update formula of the reinforcement learning model can be expressed as:

[0072]

[0073] Where Q(s t , a t ) represents the Q-value of taking action a t in state s t , α is the learning rate, r t is the immediate reward, γ is the discount factor, and Q′ is the target network.

[0074] After training is completed, the reinforcement learning model can generate optimized CT scan parameters according to the input patient information and initial parameters. These parameters are output in the output step and used to guide the actual CT scan process.

[0075] The method of the present invention further improves the accuracy of parameter optimization by using a deep convolutional neural network to evaluate the quality of CT images. The network includes 5 convolutional pooling layers, 5 pooling layers and 4 fully connected layers, with a complex and precise structure. The first 4 fully connected layers use the ReLU activation function, which helps to introduce non-linearity and alleviate the vanishing gradient problem. The last fully connected layer uses the softmax function, which is suitable for multi-class image quality scoring.

[0076] This network structure can effectively extract the features of CT images and give an accurate quality assessment based on these features. For example, for a CT image of 512x512 pixels, the network may first extract low-level features such as edges and textures through the convolutional layer, and then identify high-level features such as organ contours and lesion areas at deeper levels, and finally give a quality score by integrating this information.

[0077] Preferably, in an embodiment of the present invention, the forward propagation process of the convolutional neural network can be expressed as:

[0078] h l = f(W l * h l-1 + b l ),

[0079] Where h lrepresents the output of the l-th layer, W l and b l are the weight and bias of this layer respectively, * represents the convolution operation, and f is the activation function.

[0080] In this way, the method of the present invention can automatically optimize CT scan parameters, adapt to the individual differences of different patients, and minimize the radiation dose while ensuring image quality. This not only improves the safety and effectiveness of CT examinations, but also enhances the work efficiency of the radiology department. In a preferred embodiment of the present invention, constructing a reinforcement learning environment is a key step. This environment simulates the CT scan process and provides a platform for the reinforcement learning algorithm to interact and learn. Specifically, the patient's basic information and scan site information are used as part of the state space. This information may include the patient's age, gender, body mass index (BMI), etc., as well as the specific scan site such as the head, chest, or abdomen. Each piece of information may affect the optimal scan parameters. For example, for patients with a high BMI, a higher tube current may be required to ensure image quality; while for pediatric patients, special attention needs to be paid to reducing the radiation dose.

[0081] The adjustment range of CT scan parameters constitutes the action space. These parameters usually include tube voltage (kV), tube current (mA), rotation time, collimation width, etc. Each parameter has its reasonable adjustment range. For example, the tube voltage may be adjusted between 80 - 140 kV, and the tube current may vary between 10 - 500 mA. The method of the present invention allows for fine-grained adjustment of parameters within these ranges to find the optimal combination.

[0082] The design of the reward function is another innovation point of the present invention. This function is constructed based on image quality, image noise, and radiation dose, aiming to find the best balance among these potentially conflicting goals.

[0083] Preferably, the reward function can be expressed as:

[0084] R = w 1 ·Q - w 2 ·N - w 3 ·D,

[0085] where Q represents the image quality score, N represents the noise level, D represents the radiation dose, and w 1 、w 2 、w 3 are weight coefficients. These weights can be adjusted according to specific clinical needs. For example, for some diagnostic tasks that require high clarity, the value of w 1 can be increased; while for pediatric patients, the value of w 3 may be increased to pay more attention to reducing the radiation dose.

[0086] In the method of the present invention, the training process of the reinforcement learning model adopts advanced algorithms and strategies. Initializing the Q-value table and the Q'-value table is the first step of training. These two tables store the expected cumulative rewards for taking various actions in each state. The Q-value table is used to select actions, while the Q'-value table is used to calculate the target Q-value, which is the core idea of the Double Deep Q-Learning algorithm.

[0087] Action selection adopts the ε-greedy strategy, which is an effective method for exploring and exploiting trade-offs. In this strategy, the agent selects the currently considered optimal action with a probability of 1 - ε (exploitation), and randomly selects an action with a probability of ε (exploration). The value of ε usually starts from a relatively high value (such as 0.9) and gradually decreases as the training progresses, and may eventually drop to a very small value (such as 0.1). This can ensure that the agent fully explores the environment in the initial stage of training and makes more use of the learned knowledge in the later stage.

[0088] After executing the selected action, the system obtains a new state and a reward. The new state may reflect the change of the scanning parameters, while the reward reflects the effect of this parameter adjustment. For example, if the parameter adjustment results in an improvement in image quality while reducing the radiation dose, a positive reward will be obtained; otherwise, a negative reward may be obtained. Updating the Q-value table and the Q-value table is the core step in the training process. The Double Deep Q-Learning algorithm adopted by the present invention can effectively reduce the overestimation problem in Q-learning. The specific update formula is as follows:

[0089] Q(s t ,a t )=Q(s t ,a t )+α[r t +γQ′(s t+1 ,argmax a Q(s t+1 ,a))-Q(s t ,a t )],

[0090] where s t and a t are the current state and action respectively, r t is the immediate reward, γ is the discount factor (usually taken as 0.9 - 0.99), and α is the learning rate (which can start from 0.1 and gradually decrease as the training progresses). The key to this formula is to use the Q-network to select actions and the Q'-network to evaluate the value of this action, thereby reducing the possibility of over-optimistic estimation.

[0091] Preferably, in the embodiments of the present invention, both the Q-network and the Q'-network are implemented using deep neural networks. The network structure may include multiple fully connected layers, followed by a ReLU activation function and a Batch Normalization layer after each layer. For example, a possible network structure is:

[0092] Input layer (state_dim) -> FC(256) -> ReLU -> BN -> FC(128) -> ReLU -> BN -> FC(64) -> ReLU -> BN -> Output layer (action_dim)

[0093] Among them, state_dim is the dimension of the state space, and action_dim is the dimension of the action space. This network structure can effectively extract state features and estimate Q-values.

[0094] By repeatedly performing the above steps, the reinforcement learning model can gradually learn the optimal parameter adjustment strategy in different states. This process usually requires a large number of simulation scan experiments, and it may take hundreds of thousands or even millions of iterations to converge to a stable strategy. However, once the training is completed, the model can quickly generate optimized scan parameters for each new patient, greatly improving the efficiency and quality of CT examinations. The method of the present invention is not only innovative in theory but also ensures its effectiveness and safety in practical applications by introducing a clinical verification step. This step is a key link in applying artificial intelligence technology to the medical field, reflecting the high attention of the present invention to patient safety.

[0095] In the clinical verification stage, it is first necessary to obtain a dedicated clinical verification dataset. This dataset usually contains various types of patient cases, covering various age groups, body types, and pathological conditions. For example, it may include normal-sized adults, obese patients, pediatric patients, and patients with different diseases (such as pneumonia, tumors, etc.). This diversity helps to comprehensively evaluate the performance of the optimization algorithm.

[0096] Preferably, in the embodiments of the present invention, the size of the clinical verification dataset may be between 100 and 1000 cases, depending on the available resources and the required statistical significance level. Each case contains the patient's basic information, original scan parameters, and an image quality score evaluated by an experienced radiologist.

[0097] Performing an actual scan using the optimized CT scan parameters is the core step in the verification process. It should be particularly noted here that although the parameters are optimized by the AI algorithm, the actual operation still needs to be carried out under the supervision of professional medical staff. During the scan, the technician will closely monitor the operation status of the equipment and the patient's reaction to ensure the safety of the entire process.

[0098] After obtaining the verified images, the next step is to evaluate the quality, noise level, and radiation dose of these images. This evaluation process combines objective quantitative analysis and subjective expert evaluation. In a preferred embodiment of the present invention, the evaluation process may include the following aspects:

[0099] 1. Image quality assessment: Objective metrics such as the Structural Similarity Index (SSIM) or Peak Signal-to-Noise Ratio (PSNR) are used. At the same time, blinded scoring is performed by multiple experienced radiologists, and the scoring criteria may include contrast, clarity, degree of artifacts, etc.

[0100] 2. Noise level analysis: The noise level can be quantified by calculating the standard deviation of uniform regions in the image. For example, in CT images, regions of uniform substances such as air or water are selected, and the standard deviation of their CT values is calculated.

[0101] 3. Radiation dose assessment: It is mainly measured by the Dose-Length Product (DLP) or the effective dose. The unit of DLP is mGy·cm, while the unit of the effective dose is mSv.

[0102] During the evaluation process, the image segmentation algorithm plays an important role. The present invention uses advanced deep learning segmentation algorithms such as U-Net or Mask R-CNN, which can accurately segment the anatomical structures of interest. This not only helps to calculate the image quality metrics more accurately but also facilitates subsequent diagnostic analysis.

[0103] After segmentation, the system calculates key image quality metrics such as the Signal-to-Noise Ratio (SNR) and the Dose-Area Product (DAP). SNR is usually calculated by the following formula:

[0104]

[0105] where μ signal is the average signal intensity of the region of interest, and σ background is the standard deviation of the background region.

[0106] DAP reflects the total radiation dose received by the patient, and the calculation formula is:

[0107] DAP = D x A,

[0108] where D is the average absorbed dose and A is the area of the radiation field.

[0109] These calculated metrics are compared with preset thresholds. For example, for chest CT scans, the following thresholds may be set:

[0110] SNR > 5 (considering the low-density characteristics of the lungs);

[0111] DAP < 300 mGy·cm 2 (Based on the ALARA principle: the lower the dose, the better)

[0112] The setting of these thresholds needs to be based on a large amount of clinical experience and relevant radiation protection guidelines. If the calculated indicators exceed these thresholds, the system will issue a warning, indicating that it may be necessary to further optimize the scanning parameters.

[0113] Based on the evaluation results, the method of the present invention can determine whether it is necessary to further optimize the reinforcement learning model. If the verification results show that the optimized parameters significantly reduce the radiation dose while ensuring the image quality, then the current model can be considered to perform well. However, if it is found that the results for certain types of patients or certain specific scanning sites are not satisfactory, then these cases need to be collected, added to the training dataset, and the model needs to be retrained and optimized.

[0114] This dynamic optimization mechanism enables the method of the present invention to continuously adapt to new clinical needs and technological advancements. For example, if a hospital introduces new CT equipment or a new scanning protocol appears clinically, only relevant data needs to be collected to quickly update and optimize the model, ensuring that the system can always generate optimal scanning parameters.

[0115] Finally, the present invention also proposes a complete CT scanning parameter optimization system based on reinforcement learning. The system includes a data acquisition module 1, a data preprocessing module 2, a reinforcement learning module 3, a parameter output module 4, and a clinical verification module 5. This modular design makes the system have good scalability and flexibility.

[0116] The data acquisition module 1 is responsible for obtaining the patient's basic information, scanning site information, and initial CT scanning parameters. This module may be docked with the hospital's information system to automatically obtain the patient's electronic medical record information. At the same time, it may also include a user interface that allows technicians to manually enter some special requirements or information.

[0117] The main task of the data preprocessing module 2 is to construct a CT image quality evaluation dataset. This module may include multiple sub-modules, such as image denoising, normalization, feature extraction, etc. For example, it may use wavelet transform or deep learning methods to denoise the original CT images, and then perform image enhancement through methods such as histogram equalization.

[0118] The reinforcement learning module 3 is the core of the entire system. It is responsible for constructing the reinforcement learning environment, training the model, and generating optimized CT scanning parameters. This module may require a large amount of computing resources, so GPU acceleration or distributed computing technology may need to be considered during actual deployment.

[0119] The parameter output module 4 transmits the optimized parameters to the CT device. This module needs to consider the compatibility issues of CT devices of different brands and models, and may need to develop corresponding interfaces or protocols.

[0120] The clinical verification module 5 is responsible for evaluating and providing feedback on the optimization results. This module may require a dedicated database to store the verification results and may include some statistical analysis tools for generating verification reports.

[0121] Through this modular design, the system of the present invention can flexibly adapt to the needs and work processes of different hospitals. For example, for some advanced research hospitals, more emphasis may be placed on the customizability of the reinforcement learning module, allowing researchers to adjust algorithm parameters or try new learning strategies. For some primary hospitals, the system may focus more on usability and stability, providing more preset solutions and automated functions.

[0122] Generally speaking, the CT scan parameter optimization method and system based on reinforcement learning proposed by the present invention provide an innovative solution for improving the efficiency and safety of CT scans by combining advanced artificial intelligence technologies and rich clinical experience. It can not only adapt to the individual differences of different patients, but also continuously self-optimize according to clinical feedback, and is expected to play an important role in the future medical imaging field.

[0123] It should be noted that the above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the principles of the present invention shall be included in the protection scope of the present invention.

Claims

1. A CT scanning parameter optimization method based on reinforcement learning, characterized in that: include: The acquisition steps include: Obtain basic patient information, scan site information, and initial CT scan parameters; Obtain CT image quality assessment dataset; Processing steps include: Based on the basic information of the patient, the scanning part information and the initial CT scanning parameters, a reinforcement learning environment is constructed; Using the CT image quality assessment dataset to train a reinforcement learning model; generating optimized CT scanning parameters according to the reinforcement learning model; Output steps include: The optimized CT scanning parameters are output for performing CT scanning.

2. The method according to claim 1, characterized in that The step of obtaining a CT image quality assessment data set specifically includes: Acquire multiple groups of CT images and their corresponding scanning parameters; Using a deep convolutional neural network to evaluate the quality of the CT image to obtain an image quality index; The CT image, scanning parameters and image quality indicators are combined to form CT image quality evaluation data.

3. The method according to claim 2, characterized in that The deep convolutional neural network includes 5 convolutional pooling layers, 5 pooling layers and 4 fully connected layers, where: The activation function of the first four fully connected layers is the ReLU function; The activation function of the last fully connected layer is the softmax function.

4. The method according to claim 1, characterized in that: The steps of constructing the reinforcement learning environment specifically include: The basic information of the patient and the information of the scanned part are used as the state space; The adjustment range of CT scanning parameters is used as the action space; Construct a reward function based on image quality, image noise, and radiation dose.

5. The method according to claim 1, characterized in that The step of training the reinforcement learning model using the CT image quality assessment data set specifically includes: Initialize the Q value table and Q' value table; Select actions based on the ε-greedy strategy; Execute the selected action and obtain the new state and reward; Update the Q value table and Q' value table; Repeat the above steps until the model converges.

6. The method according to claim 5, characterized in that The step of updating the Q value table and the Q' value table adopts the Double Deep Q-Learning algorithm.

7. The method according to claim 1, characterized in that Also includes clinical validation steps: Acquisition of clinical validation datasets; Performing CT scanning using the optimized CT scanning parameters to obtain a verification image; evaluating the quality, noise level, and radiation dose of the validation images; Based on the evaluation results, determine whether the reinforcement learning model needs to be further optimized.

8. The method according to claim 7, characterized in that The step of evaluating and verifying the image specifically includes: Use image segmentation algorithm to segment the verification image; Calculate the signal-to-noise ratio (SNR) and dose-area product (DAP) of the segmented image; The calculated SNR and DAP are compared with preset thresholds.

9. The method according to claim 1, characterized in that: The patient's basic information includes height, weight, age and gender; the scanned part information includes head, chest, abdomen and whole body.

10. A CT scanning parameter optimization system based on reinforcement learning for executing the method according to any one of claims 1 to 9, characterized in that: include: Data acquisition module, used to obtain basic patient information, scan site information and initial CT scan parameters; Data preprocessing module, used to construct CT image quality evaluation data set; Reinforcement learning modules for: Building a reinforcement learning environment; Training reinforcement learning models; Generate optimized CT scanning parameters; A parameter output module, used for outputting the optimized CT scanning parameters; Clinically validated modules for: Performing CT scanning using the optimized CT scanning parameters; Evaluate scan results; The evaluation results are fed back to the reinforcement learning module for further optimization.

Images