Artificial intelligence assisted CT image quality enhancement and noise reduction method

Abstract

The invention relates to the technical field of image processing, in particular to an artificial intelligence-assisted CT image quality enhancement and noise reduction method, which comprises the following steps of S1, acquiring low-dose CT scanning image data; s2, preprocessing the acquired low-dose CT image data; s3, extracting preliminary feature information in the CT image; s4, performing adaptive estimation on noise in the image; s5, inputting the noise estimation result and the initial feature information into a deep convolutional denoising network; s6, performing quality enhancement on the denoised image based on a generative adversarial network framework; s7, reconstructing the CT image after quality enhancement; and S8, performing medical applicability detection on the final image. Through the generative adversarial network, the convolutional neural network and the medical image discrimination model, denoising and quality enhancement of the low-dose CT image are realized, the CT image with a high signal-to-noise ratio and high structure fidelity is generated, the medical applicability of the CT image is ensured, and clinical diagnosis requirements are met.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and particularly to an artificial intelligence-assisted CT image quality enhancement and noise reduction method. Background Art

[0002] With the wide application of computed tomography (CT) technology, CT images have become an indispensable and important tool in medical diagnosis, especially showing significant advantages in detecting early lesions and precise positioning. However, in order to reduce the potential harm of radiation to patients, low-dose CT scans are increasingly used. However, although low-dose CT reduces the radiation dose, it brings a decrease in the signal-to-noise ratio of the image and a reduction in image quality. Traditional image post-processing methods attempt to improve image quality through filtering or enhancement techniques, but often result in the loss of image details or the introduction of artifacts during the denoising process, making it difficult to ensure the reliability of the image in diagnosis. Existing image processing technologies are difficult to effectively improve image quality while maintaining the integrity of the image structure, especially under low-dose CT conditions, and the balance between image quality and signal-to-noise ratio has always been a technical problem.

[0003] Existing technologies mainly rely on rule design or simple convolutional neural network models. Although they can achieve a certain degree of denoising and image enhancement, when processing CT images, it is difficult to simultaneously consider the distribution characteristics of noise and the structural features of the image, resulting in the final generated images often not meeting the high standards of clinical diagnosis. In addition, there is a lack of automated means for the medical applicability detection of images, and doctors need to manually evaluate the image quality, increasing the workload. Therefore, an artificial intelligence-based CT image quality enhancement and noise reduction method is proposed to solve the problems of the quality and diagnostic accuracy of low-dose CT images. Summary of the Invention

[0004] Based on the above objectives, the present invention provides an artificial intelligence-assisted CT image quality enhancement and noise reduction method.

[0005] The artificial intelligence-assisted CT image quality enhancement and noise reduction method includes the following steps: S1: Obtain low-dose CT scan image data, where the image data includes two-dimensional tomographic images collected by a CT device under low-dose conditions; S2: Preprocess the obtained low-dose CT image data, and the preprocessing includes normalizing, artifact removal, and edge smoothing processing of the image data; S3: Based on the preprocessed image data, extract preliminary feature information in the CT image, and the feature information includes the edges, textures, and gray-scale distributions of the image; S4: Input the extracted preliminary feature information into a convolutional neural network model to adaptively estimate the noise in the image. The convolutional neural network model generates feature weights related to the image noise by learning the noise distribution characteristics in the CT image to guide the denoising operation; S5: Input the noise estimation result and the preliminary feature information into a deep convolutional denoising network. The denoising network consists of multiple convolutional and deconvolutional operations to remove the image noise; S6: Enhance the quality of the denoised image based on the generative adversarial network framework. The generative adversarial network includes a generator and a discriminator. The generator is used to generate enhanced CT images, and the discriminator is used to evaluate the similarity between the enhanced image and the standard high-dose image; S7: Reconstruct the quality-enhanced CT image to generate a final image output with high signal-to-noise ratio and high structural fidelity; S8: Perform medical applicability detection on the final image, and automatically evaluate the image through a trained medical image discrimination model.

[0006] Optionally, the specific steps of S1 include: S11: Set the low-dose scanning parameters of the CT device. The scanning parameters include the X-ray tube voltage, tube current, and exposure time. The range of the tube voltage is controlled between 80 kVp and 120 kVp, the range of the tube current is controlled between 10 mA and 50 mA, and the exposure time is 1.5 seconds; S12: Start the CT device based on the set low-dose scanning parameters, emit a low-dose X-ray beam through the X-ray tube, and receive the attenuation information after passing through the human body by the detector; S13: Convert the received attenuation information into a two-dimensional tomographic image through the reconstruction algorithm in the CT device. The reconstruction algorithm uses the filtered back-projection technique, and the specific formula is: , where, represents the reconstructed two-dimensional tomographic image, and are the coordinate positions of the image; represents the projection data received at the projection angle and the projection distance ; represents the projection angle; represents the projection path distance; represents the convolution kernel of the filtering function acting on the projection data, where is the projection path coordinate; and are the cosine and sine functions of the projection angle respectively; is the integration variable, representing the cumulative weight of the image points along different projection angles .

[0007] Optionally, S2 specifically includes: S21: Normalize the acquired low-dose CT image data, adjust the gray value of the image data to a predetermined range, and make the gray values of all pixels fall within the range of 0 to 1; S22: Perform artifact removal on the normalized image. Based on the bilateral filtering algorithm, perform joint smoothing processing on the image in terms of space and pixel intensity; S23: Perform edge smoothing on the artifact-removed image. By applying Gaussian smoothing filtering technology, remove the high-frequency noise in the image.

[0008] Optionally, S3 specifically includes: S31: Extract edge features from the preprocessed image data. Specifically, adopt a gradient-based edge detection algorithm, and identify the regions where the gray values of the image change sharply by calculating the change rate of the gray values of the pixels in the image; S32: Extract texture features from the image after edge feature extraction. Use the gray-level co-occurrence matrix analysis method, and extract the local texture information in the image by statistically analyzing the co-occurrence relationship of the gray values of adjacent pixels, including the directionality, roughness, and repeatability of the texture; S33: While extracting texture features, extract the gray distribution features of the image, calculate the overall gray value distribution of the image, and obtain the distribution of bright and dark regions in the image by statistically analyzing the gray histogram of the pixels.

[0009] Optionally, S4 specifically includes: S41: Input the extracted preliminary feature information into a pre-trained convolutional neural network model. The convolutional neural network model consists of multiple convolutional layers, pooling layers, and fully connected layers, and extracts image features at different scales through convolutional operations; S42: In the convolutional layer, the convolutional kernel scans the input image features pixel by pixel and calculates the feature responses of local regions; S43: In the pooling layer, reduce the data dimension through downsampling operations, extract the image features, and at the same time retain the key noise patterns of the image; S44: After multiple convolutional and pooling processes, enter the fully connected layer to calculate the feature weights. Through adaptive learning of the feature weights, generate a feature weight matrix related to noise. The feature weight matrix is used to reflect the intensity, distribution, and position of the noise in the image; S45: Finally, by outputting the feature weight matrix, the convolutional neural network can adaptively estimate the noise in the image and generate a corresponding noise distribution model.

[0010] Optionally, S5 specifically includes: S51: Take the noise estimation result and the preliminary feature information as inputs and input them into a deep convolutional denoising network. The preliminary feature information includes the edges, textures, and gray-scale distributions of the image, while the noise estimation result represents the noise distribution at each pixel position in the image; S52: In the deep convolutional denoising network, use multi-layer convolutional operations to jointly process the noise estimation result and the preliminary feature information. The joint processing comprehensively combines the noise and feature information in the image through convolutional operations, enabling the denoising process to consider the structural features of the image and extract the joint distribution of noise and features; S53: Perform a deconvolution operation on the image features after convolutional processing to map the extracted high-dimensional features back to the spatial dimension of the original image and generate a denoised image; S54: Perform weighted fusion on the denoised image obtained by deconvolution and the preliminary feature information, and balance the denoising effect and the retention of image detail features by adjusting the weight ratio, thereby generating the final denoised image.

[0011] Optionally, the specific steps of S6 are as follows: S61: Input the denoised CT image into the generator in the generative adversarial network framework. The generator is a convolutional neural network structure that extracts and enhances the detail features in the input denoised image through convolutional and deconvolutional operations; S62: The generator extracts the feature information in the denoised image through multiple convolutional layers and gradually adjusts the details in the image; S63: Input the generated enhanced CT image into the discriminator. The discriminator compares the enhanced image with a preset standard image, calculates the difference between the two, and outputs a similarity score; S64: Based on the feedback from the discriminator, the generator will optimize the image generation process. By continuously adjusting the weights of the convolutional kernels, the similarity between the enhanced image and the standard image is maximized. Furthermore, through multiple iterations of optimization in the training of the generative adversarial network, the output gradually approaches the quality of the standard high-dose CT image.

[0012] Optionally, the specific steps of S63 are as follows: S631: Take the generated enhanced CT image and the preset standard high-dose CT image as inputs and send them into the convolutional neural network structure of the discriminator respectively. Through convolutional operations, extract the feature vectors of the enhanced image and the standard image respectively; S632: The discriminator compares the feature vectors of the enhanced image and the standard image, calculates the difference between the two, and obtains the overall difference value of the two images in the feature space by comparing the feature differences of each pixel point; S633: Calculate a similarity score based on the calculated difference value. The similarity score represents the degree of matching between the enhanced image and the standard image. The higher the score, the closer the two are in terms of details and structure.

[0013] Optionally, S7 specifically includes: S71: Use the quality-enhanced CT image as input, send it into the image reconstruction module, and reconstruct the image through the filtered back-projection technique. S72: During the reconstruction process, use a high signal-to-noise ratio optimization algorithm to optimize the reconstructed image. S73: Use a structure fidelity optimization algorithm to ensure the detail and structure fidelity of the reconstructed image. Specifically, optimize the image quality by calculating the structural similarity index of the image.

[0014] Optionally, S8 specifically includes: S81: Input the finally reconstructed CT image into the trained medical image discrimination model. S82: The discrimination model extracts features from the input CT image through multiple convolutional layers and pooling layers, extracting the structural information, texture features, and gray-scale distribution in the image. S83: The discrimination model inputs the extracted image features into a preset classification algorithm through the classification layer to evaluate the diagnostic applicability of the image. S84: According to the classification result, the discrimination model outputs a medical applicability score, with the score range from 0 to 1, which is used to reflect whether the image is suitable for clinical diagnosis. If the score reaches the preset threshold, it indicates that the CT image meets the clinical diagnosis requirements.

[0015] Advantages of the present invention: In the present invention, by introducing a generative adversarial network and a convolutional neural network model, denoising and quality enhancement are performed on low-dose CT images. The generator extracts and enhances the detail features in the image through multi-layer convolutional and deconvolutional operations. The discriminator compares the generated image with the standard high-dose image and outputs a similarity score, gradually optimizing the quality of the generated CT image.

[0016] In the present invention, through a high signal-to-noise ratio optimization algorithm and a structure fidelity optimization algorithm, important features such as edges and textures of the image can be maintained while removing noise. The generated CT image has a high signal-to-noise ratio and high structure fidelity.

[0017] In the present invention, by automatically extracting image features and performing multi-dimensional evaluations on the quality, noise level, and structural integrity of the image, the discrimination model can quickly and accurately output an applicability score, ensuring that the generated CT image meets the clinical diagnosis requirements. Description of the Drawings

[0018] To more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only those of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0019] Figure 1 Schematic diagram of the CT image quality enhancement and noise reduction method according to an embodiment of the present invention; Figure 2 Schematic diagram of the process for adaptively estimating the noise in the image according to an embodiment of the present invention. Detailed implementation manners

[0020] As Figure 1 - Figure 2 shown, the artificial intelligence-assisted CT image quality enhancement and noise reduction method includes the following steps: S1: Obtain low-dose CT scan image data, where the image data includes two-dimensional tomographic images collected by the CT device under low-dose conditions; S2: Preprocess the obtained low-dose CT image data. The preprocessing includes normalizing, removing artifacts, and edge smoothing of the image data to optimize the basic quality of the image data and provide a good input basis for subsequent steps; S3: Based on the preprocessed image data, extract the preliminary feature information in the CT image. The feature information includes the edges, textures, and gray-scale distributions of the image to ensure that the image structure information is fully retained; S4: Input the extracted preliminary feature information into a convolutional neural network model to adaptively estimate the noise in the image. The convolutional neural network model learns the noise distribution characteristics in the CT image and generates feature weights related to the image noise to guide the denoising operation; S5: Input the noise estimation result and the preliminary feature information into a deep convolutional denoising network. The denoising network consists of multiple convolutional and deconvolutional operations to remove the image noise while keeping the image structure information intact; S6: Enhance the quality of the denoised image based on the generative adversarial network framework. The generative adversarial network includes a generator and a discriminator. The generator is used to generate the enhanced CT image, and the discriminator is used to evaluate the similarity between the enhanced image and the standard high-dose image to optimize the output of the generator; S7: Reconstruct the quality-enhanced CT image to generate a final image output with high signal-to-noise ratio and high structural fidelity, ensuring that the clarity and detail fidelity of the image meet the diagnostic requirements; S8: Perform medical applicability detection on the final image, and automatically evaluate the image through the trained medical image discrimination model to ensure that the generated CT image meets the requirements of clinical diagnosis.

[0021] S1 specifically includes: S11: Set the low-dose scanning parameters of the CT device. The scanning parameters include X-ray tube voltage, tube current, and exposure time. Among them, the range of tube voltage is controlled between 80 kVp and 120 kVp, the range of tube current is controlled between 10 mA and 50 mA, and the exposure time is 1.5 seconds to ensure that the CT scan is completed under low-dose conditions to reduce the impact of radiation on patients; S12: Start the CT device based on the set low-dose scanning parameters, emit a low-dose X-ray beam through the X-ray tube, and receive the attenuation information after passing through the human body through the detector; S13: Convert the received attenuation information into a two-dimensional tomographic image through the reconstruction algorithm in the CT device. The reconstruction algorithm uses the filtered back-projection technique, and the specific formula is: , where, represents the reconstructed two-dimensional tomographic image, and are the coordinate positions of the image; represents at the projection angle and the projection distance the projection data received below; represents the projection angle; represents the projection path distance; represents the convolution kernel of the filtering function acting on the projection data, where is the projection path coordinate; and are the cosine and sine functions of the projection angle respectively, used to calculate the corresponding projection direction; is the integration variable, representing the cumulative weight of the image points along different projection angles ; By performing filtering and back-projection operations on all projection data, a complete CT two-dimensional tomographic image is obtained; Through the above steps, CT scan image data can be efficiently obtained under low-dose conditions, ensuring a reduction in the impact of radiation dose on patients, while ensuring the integrity and accuracy of the image data, providing a high-quality input basis for subsequent image processing and noise reduction.

[0022] S2 specifically includes: S21: Normalize the acquired low-dose CT image data, adjust the grayscale values of the image data to a predetermined range, so that the grayscale values of all pixels are within the range of 0 to 1, thereby optimizing the image contrast and standardizing the data, ensuring the consistency between different images, and facilitating the implementation of subsequent processing steps; Normalization refers to adjusting the grayscale values of the image data to a predetermined range, specifically adjusting the grayscale value of each pixel through the formula: , where represents the grayscale value at position in the image; represents the normalized grayscale value; and are respectively the minimum and maximum grayscale values in the image; S22: Perform artifact removal on the normalized image. Based on the bilateral filtering algorithm, through the joint smoothing processing of the image in terms of space and pixel intensity, while removing artifacts, retain the important edges and structural information in the image, and ensure image clarity while removing artifacts; The specific filtering process is as follows: , where is the image after artifact removal; is the pixel value at position in the image; is the normalization factor, used to normalize the output of the filter; is the filtering window; and are the standard deviations of the spatial and pixel intensity of the filter, used to control the intensity of filtering smoothing; S23: Perform edge smoothing on the image after artifact removal. By applying Gaussian smoothing filtering technology, remove the high-frequency noise in the image, make the image edge transition natural, and at the same time keep the structural information of the image clear, so as to reduce the interference of noise in the subsequent processing process; The formula for the specific smoothing process is: , where represents the Gaussian filter; is the standard deviation of the Gaussian filter, used to control the smoothing intensity; and are the image coordinates; Through the above preprocessing steps, including normalization, artifact removal, and edge smoothing, the quality of the low-dose CT image can be effectively improved, the influence of artifacts and noise can be reduced, and the image is ensured to have higher clarity and contrast, laying a good foundation for subsequent image enhancement and noise removal operations.

[0023] S3 specifically includes: S31: Extract edge features from the preprocessed image data. Specifically, a gradient-based edge detection algorithm is used. By calculating the rate of change of pixel gray values in the image, regions with sharp changes in gray values in the image are identified. The extracted edge information is used to reflect the contours of organs and tissues in the CT image. The specific edge detection formula is: , where represents the edge intensity at position ; represents the image gray value at position ; and are the gray gradients of the image in the direction and the direction respectively. By calculating the gray gradients of each pixel, the change in gray values at the edge positions in the image is identified, thereby extracting the edge features of the image. S32: Extract texture features from the image after edge feature extraction. Using the gray-level co-occurrence matrix analysis method, by statistically analyzing the co-occurrence relationship of adjacent pixel gray values, local texture information in the image is extracted, including the directionality, roughness, and repeatability of the texture. These texture features can reflect the detailed information of tissues in the CT image. The specific calculation formula for texture features is: , where represents the texture feature value of the image; represents the joint probability between gray values and ; and represent the gray values of pixels respectively; S33: While extracting texture features, extract the gray distribution features of the image, calculate the overall gray value distribution of the image, and obtain the distribution of bright and dark regions in the image by statistically analyzing the gray histogram of pixels, so as to reflect the density differences of different tissues in the CT image. The formula for calculating the overall gray value distribution of the image is: , where represents the relative frequency of gray value in the image; represents the number of pixels with gray value ; represents the total number of pixels in the image. The above steps can accurately extract the edge, texture, and gray distribution features in the CT image by using gradient edge detection, gray-level co-occurrence matrix analysis, and gray histogram techniques. These features provide detailed information for subsequent noise estimation and denoising processing, ensuring that the image maintains high quality and structural integrity during the noise removal process.

[0024] S4 specifically includes: S41: Input the extracted preliminary feature information, including the edges, textures, and grayscale distributions of the image, into a pre-trained convolutional neural network (CNN) model. The CNN model consists of multiple convolutional layers, pooling layers, and fully connected layers. Extract image features at different scales through convolutional operations while retaining the local and global information of the image; S42: In the convolutional layer, the convolutional kernel scans the input image features pixel by pixel and calculates the feature responses in the local area. Through multiple convolutional operations, the noise patterns and structural information in the image can be captured. The weights of the convolutional kernel are optimized through a large-scale training dataset, enabling the network to adaptively learn the distribution characteristics of different types of noise in the CT image; S43: In the pooling layer, reduce the data dimension through downsampling operations to extract concise image features while retaining the key noise patterns in the image. The pooling layer can reduce the computational complexity and enhance the generalization ability of the model to avoid overfitting; S44: After multiple convolutional and pooling processes, enter the fully connected layer to calculate the feature weights. Through adaptive learning of the feature weights, a feature weight matrix related to noise is generated. The feature weight matrix is used to reflect the intensity, distribution, and location of the noise in the image; S45: Finally, by outputting the feature weight matrix, the convolutional neural network can adaptively estimate the noise in the image and generate a corresponding noise distribution model. This noise distribution model is used to guide subsequent denoising operations to ensure that while removing the noise, the structure and detail information in the image are retained.

[0025] The specific calculation process for the above adaptive estimation of the noise in the image is as follows: First, input the extracted preliminary feature information into the convolutional neural network model; Then, in the convolutional layer, use the convolutional kernel to perform a convolutional operation on the input feature information. Specifically: , where represents the feature value after convolution; represents the weight of the convolutional kernel at position ; represents the value of the input feature in the neighborhood ; represents the size of the convolutional kernel, which controls the range of the receptive field. Through multiple convolutional operations, the model can capture the noise patterns and structural information in the image; Next, in the pooling layer, perform a downsampling operation on the convolutional result . Use the max pooling method to calculate the maximum value in each region to reduce the data dimension. The specific formula is: , where is the pooled eigenvalue; is the size of the pooling window, which determines the area range of each sampling; the pooling layer is used to reduce the size of the feature map, retain the key features of the noise, and at the same time reduce the computational complexity; Then, the feature map after convolution and pooling is input into the fully connected layer to generate a feature weight matrix , and its calculation process is: , where is the weight function in the fully connected layer, which is used to calculate the noise feature weights; Finally, according to the feature weight matrix a noise distribution model is generated , specifically: , where, represents the noise estimate value at position ; is the noise feature weight matrix; is the pooled eigenvalue, and through the above calculation, a noise distribution model is obtained. This model is used to guide the subsequent denoising operation; the above steps can accurately capture the noise features in the CT image by using a convolutional neural network model to adaptively estimate the image noise, generate a feature weight matrix related to the noise, and the model can perform denoising operations according to the noise distribution in the image, ensuring that the noise is effectively removed while retaining the structural information of the image, improving the image quality. This adaptive denoising method based on feature weights significantly improves the accuracy and image fidelity in the denoising process.

[0026] S5 specifically includes: S51: Taking the noise estimation result and the preliminary feature information as inputs, and inputting them into the deep convolutional denoising network. The preliminary feature information includes the edges, textures, and gray-scale distributions of the image, while the noise estimation result represents the noise distribution situation at each pixel position in the image; S52: In the deep convolutional denoising network, using multiple convolutional operations to jointly process the noise estimation result and the preliminary feature information. The joint processing synthesizes the noise and feature information in the image through convolutional operations, enabling the denoising process to consider the structural features of the image and extract the joint distribution of the noise and features; S53: Performing a deconvolution operation on the image features after convolution processing, mapping the extracted high-dimensional feature map back to the spatial dimension of the original image to generate a denoised image. The deconvolution operation can restore the original resolution of the image and retain the denoising information extracted during the convolution process; S54: The denoised image obtained by deconvolution is fused with the preliminary feature information, and the weight ratio is adjusted to balance the denoising effect and the retention of image detail features, thereby generating the final denoised image; through the joint processing of the noise estimation result and the preliminary feature information in the above steps, the use of a deep convolutional denoising network can effectively remove the noise in the image and retain the structural information of the image. The finally generated denoised image has higher quality and is suitable for subsequent image quality enhancement and diagnostic applications.

[0027] The specific calculation process for removing image noise is as follows: First, the noise estimation result and the preliminary feature information are used as inputs and fed into the deep convolutional denoising network together; Then, in the deep convolutional denoising network, through multiple convolutional operations, the noise estimation result and the preliminary feature information are jointly processed. The specific form of the convolutional operation is: , where represents the denoised image features; is the convolutional kernel weight in the denoising network, which is optimized for the joint distribution of noise and features; is the noise estimation result; is the preliminary feature information; is the size of the convolutional kernel, which controls the range of the receptive field; Next, a deconvolution operation is performed on to reconstruct the denoised image. The deconvolution process maps the high-dimensional features back to the spatial dimension of the original image. The specific deconvolution operation is: , where is the pixel value of the deconvolution output image at coordinate , that is, the restored image result; is the weight of the deconvolution operation; is the output feature map of the convolutional stage of the denoising network, that is, the feature value at position , ; Finally, by fusing the deconvolution result with the preliminary feature information , the final denoised image is generated; the fusion process controls the balance between features and denoising information through the weight coefficient . The specific formula is: , where is the finally denoised image; is the fusion weight, which controls the balance between denoising information and preliminary features; is the image after deconvolution; is the preliminary feature information.

[0028] S6 specifically includes: S61: Input the denoised CT image into the generator in the generative adversarial network (GAN) framework. The generator is a convolutional neural network structure. It performs convolution and deconvolution operations on the input denoised image to extract and enhance the detailed features in the image. The output of the generator is the enhanced CT image, which retains the detailed information of the original image and improves the overall image quality at the same time; S62: The generator extracts the feature information in the denoised image through multiple convolutional layers and gradually adjusts the details in the image. The weights of the convolutional kernels are optimized according to a large-scale training dataset, which can identify and amplify the key image features during the generation process, thereby generating a clearer and more structurally complete CT image; S63: Input the generated enhanced CT image into the discriminator. The discriminator compares the enhanced image with a preset standard image, calculates the difference between the two, and outputs a similarity score. This score reflects the degree of closeness of the enhanced image to the standard high-dose CT image in multiple dimensions, ensuring that the generated image meets the quality standards required for clinical applications; S64: Based on the feedback of the discriminator, the generator will optimize the image generation process. By continuously adjusting the weights of the convolutional kernels, it maximizes the similarity between the enhanced image and the standard image. Furthermore, the training of the generative adversarial network is optimized through multiple iterations, and the output gradually approaches the quality of the standard high-dose CT image. Through the collaborative work of the generator and discriminator based on the generative adversarial network, the quality of the denoised CT image can be effectively enhanced. The enhanced image generated by the generator undergoes a similarity evaluation by the discriminator to achieve the quality effect of high-quality CT images. This method not only improves the details and contrast of CT images but also ensures the overall structural fidelity of the images through the feedback of the discriminator, and is applicable to the generation of high-quality CT images for clinical diagnosis.

[0029] S63 specifically includes: S631: Take the generated enhanced CT image and a preset standard high-dose CT image as inputs and send them into the convolutional neural network structure of the discriminator respectively. Through convolution operations, the feature vectors of the enhanced image and the standard image are extracted respectively, which are used to describe the detailed features and overall structural information in the image; S632: The discriminator compares the feature vectors of the enhanced image and the standard image, calculates the difference between the two, and obtains the overall difference value of the two images in the feature space by comparing the feature differences of each pixel point; S633: According to the calculated difference value, calculate the similarity score. The similarity score represents the matching degree between the enhanced image and the standard image. The higher the score, the closer the two are in terms of details and structure.

[0030] The specific steps for calculating the difference between the two in S63 are as follows: First, take the generated enhanced CT image and the preset standard high-dose CT image as inputs and send them into the convolutional neural network structure of the discriminator. After the two images are processed through multiple convolutional layers, their feature vectors are respectively extracted, which are and ; Then, by calculating the difference between the feature vectors of the generated image and the standard image, define the difference function between them. The specific difference formula is: where, is the feature difference degree between the enhanced image and the standard image; represents the feature vector of the enhanced image; represents the feature vector of the standard image; is the image coordinate; Next, according to the difference degree , calculate the similarity score of the image. The similarity score is used to evaluate the matching degree of the two images. The calculation formula is: where, represents the similarity score between the enhanced image and the standard image, and its value range is from 0 to 1. The closer it is to 1, the more similar the images are; is the maximum difference degree of all training images in the system and is used for normalization processing; The above steps can accurately evaluate the image quality by comparing the feature vectors of the enhanced image and the standard high-dose CT image and using the difference degree function to calculate the difference between the two, and output the similarity score. This scoring mechanism can provide feedback for the generative adversarial network, making the generated enhanced CT image gradually approach the quality of the standard image and ensuring that the fidelity of the image structure and detail features meets the clinical requirements.

[0031] S7 specifically includes: S71: Take the CT image with enhanced quality as the input, send it into the image reconstruction module, and reconstruct the image through the filtered back projection (FBP) technology; S72: During the reconstruction process, use the high signal-to-noise ratio optimization algorithm to optimize the reconstructed image. This optimization algorithm calculates the signal-to-noise ratio SNR of the image. The specific calculation formula is: where, represents the power of the signal; represents the power of the noise; The algorithm adjusts the reconstruction parameters according to the SNR value to ensure that the noise interference is minimized as much as possible during the image reconstruction process; S73: Use the structural fidelity optimization algorithm to ensure the details and structural fidelity of the reconstructed image. Specifically, optimize the image quality by calculating the structural similarity index (SSIM) of the image. The formula for the similarity index is as follows: , where and are the means of images and ; and are the variances of images and ; is the covariance between images and ; and are constants to avoid division by zero. Through the filtered back-projection algorithm, high signal-to-noise ratio optimization, and structural fidelity optimization algorithm, the above steps can achieve the accurate reconstruction of the quality-enhanced CT image and output the final image with high signal-to-noise ratio and high structural fidelity.

[0032] When the SNR value increases, it indicates that the effective signal significantly exceeds the noise power at the power level; in the solution of the present invention, when SNR≥30dB, it is regarded as a high signal-to-noise ratio; Reaching this threshold can ensure: The error between the CT pixel gray value and the true linear attenuation coefficient is controlled within 3%; The contrast attenuation of the 0.5mm line pair resolution target does not exceed 5%.

[0033] Therefore, in the present invention, the high signal-to-noise ratio is sufficient to ensure that the details, textures, and density gradients in the image can be presented without bias due to the advantage of the effective signal over the noise.

[0034] S8 specifically includes: S81: Input the reconstructed final CT image into the trained medical image discrimination model. The discrimination model is a convolutional neural network model based on deep learning. Through learning a large amount of medical image data, the model can automatically identify the features of the CT image and evaluate whether it meets the clinical diagnosis criteria; S82: The discrimination model extracts features from the input CT image through multiple convolutional layers and pooling layers, extracting the structural information, texture features, and gray distribution in the image for comparison with the preset medical standards; S83: The discrimination model inputs the extracted image features into the preset classification algorithm through the classification layer to evaluate the diagnostic applicability of the image. The classification algorithm scores based on multiple dimensions such as image quality, noise level, structural integrity, and detail fidelity, specifically including evaluation criteria such as signal-to-noise ratio and structural similarity index; S84: According to the classification result, the discrimination model outputs a medical applicability score ranging from 0 to 1, which is used to reflect whether the image is suitable for clinical diagnosis. If the score reaches the preset threshold, it indicates that the CT image meets the requirements of clinical diagnosis.

[0035] The present invention covers any alternatives, modifications, equivalent methods and solutions made on the essence and scope of the present invention. For the public to have a thorough understanding of the present invention, specific details are described in detail in the following preferred embodiments of the present invention. However, those skilled in the art can fully understand the present invention without the description of these details. In addition, well-known methods, processes, procedures, components and circuits, etc. are not described in detail to avoid unnecessary confusion to the essence of the present invention.

[0036] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. An artificial intelligence-assisted CT image quality enhancement and noise reduction method, characterized in that: The following steps are involved: S1: Acquire low-dose CT scan image data, where the image data includes a two-dimensional tomographic image acquired by a CT device under low-dose conditions; S2: preprocessing the acquired low-dose CT image data, wherein the preprocessing includes normalizing the image data, removing artifacts, and performing edge smoothing processing; S3: extracting preliminary feature information from the CT image based on the preprocessed image data, wherein the feature information includes edge, texture and grayscale distribution of the image; S4: inputting the extracted preliminary feature information into a convolutional neural network model to adaptively estimate the noise in the image, wherein the convolutional neural network model generates feature weights related to the image noise by learning the noise distribution characteristics in the CT image to guide the denoising operation; S5: inputting the noise estimation result and the preliminary feature information into a deep convolutional denoising network, wherein the denoising network is composed of multiple layers of convolution and deconvolution operations to remove image noise; S6: enhancing the quality of the denoised image based on a generative adversarial network framework, wherein the generative adversarial network includes a generator and a discriminator, wherein the generator is used to generate an enhanced CT image, and the discriminator is used to evaluate the similarity between the enhanced image and the standard high-dose image; S7: reconstructing the quality-enhanced CT image to generate a final image output with high signal-to-noise ratio and high structural fidelity; S8: Perform medical applicability detection on the final image and automatically evaluate the image using the trained medical image discrimination model.

2. The artificial intelligence-assisted CT image quality enhancement and noise reduction method according to claim 1, characterized in that: The S1 specifically includes: S11: Setting low-dose scanning parameters of the CT device, the scanning parameters including X-ray tube voltage, tube current and exposure time, wherein the tube voltage is controlled within a range of 80 kVp to 120 kVp, the tube current is controlled within a range of 10 mA to 50 mA, and the exposure time is 1.5 seconds; S12: starting the CT device based on the set low-dose scanning parameters, emitting a low-dose X-ray beam through the X-ray tube, and receiving attenuation information after passing through the human body through the detector; S13: Convert the received attenuation information into a two-dimensional tomographic image through a reconstruction algorithm in the CT device. The reconstruction algorithm adopts a filtered back projection technique. The specific formula is: ,in, represents the reconstructed two-dimensional tomographic image, and is the coordinate position of the image; Indicates the projection angle and projection distance The received projection data; Indicates the projection angle; represents the projected path distance; Represents the convolution kernel of the filter function acting on the projection data, where is the projection path coordinate; and are the cosine and sine functions of the projection angle, respectively; is the integral variable, indicating the angles along different projections The accumulated weights of the image points.

3. The artificial intelligence-assisted CT image quality enhancement and noise reduction method according to claim 1, characterized in that: The S2 specifically includes: S21: performing normalization processing on the acquired low-dose CT image data, adjusting the grayscale value of the image data to a predetermined range, so that the grayscale value of all pixels is between 0 and 1; S22: performing artifact removal on the normalized image by performing joint smoothing of the image space and pixel intensity based on a bilateral filtering algorithm; S23: performing edge smoothing on the image after artifact removal, and removing high-frequency noise in the image by applying Gaussian smoothing filtering technology.

4. The artificial intelligence-assisted CT image quality enhancement and noise reduction method according to claim 1, characterized in that: The S3 specifically includes: S31: extracting edge features from the preprocessed image data, specifically using a gradient-based edge detection algorithm to identify areas in the image where grayscale values ​​change sharply by calculating the rate of change of pixel grayscale values ​​in the image; S32: extracting texture features from the image after edge feature extraction, using gray level co-occurrence matrix analysis method, by counting the co-occurrence relationship of adjacent pixel gray values, extracting local texture information in the image, including texture directionality, roughness and repeatability; S33: While extracting the texture features, the grayscale distribution features of the image are extracted, the overall grayscale value distribution of the image is calculated, and the distribution of bright and dark areas in the image is obtained by statistically analyzing the grayscale histogram of the pixels.

5. The artificial intelligence-assisted CT image quality enhancement and noise reduction method according to claim 1, characterized in that: The S4 specifically includes: S41: inputting the extracted preliminary feature information into a pre-trained convolutional neural network model, wherein the convolutional neural network model is composed of multiple convolutional layers, pooling layers, and fully connected layers, and extracting image features at different scales through convolution operations; S42: In the convolution layer, the convolution kernel scans the input image features pixel by pixel and calculates the feature response of the local area; S43: In the pooling layer, the data dimension is reduced by downsampling operation to extract image features while retaining the key noise pattern of the image; S44: After multi-layer convolution and pooling processing, enter the fully connected layer to calculate feature weights, and generate a feature weight matrix related to noise by adaptively learning the feature weights. The feature weight matrix is ​​used to reflect the intensity, distribution and position of noise in the image; S45: Finally, by outputting the feature weight matrix, the convolutional neural network can adaptively estimate the noise in the image and generate a corresponding noise distribution model.

6. The artificial intelligence-assisted CT image quality enhancement and noise reduction method according to claim 5, characterized in that: The S5 specifically includes: S51: inputting the noise estimation result and preliminary feature information as input to a deep convolutional denoising network, wherein the preliminary feature information includes the edge, texture and grayscale distribution of the image, and the noise estimation result represents the noise distribution of each pixel position in the image; S52: In the deep convolutional denoising network, a multi-layer convolution operation is used to jointly process the noise estimation result and the preliminary feature information, wherein the joint processing integrates the noise and feature information in the image through the convolution operation, so that the denoising process can consider the structural characteristics of the image and extract the joint distribution of the noise and the feature; S53: performing a deconvolution operation on the image features after the convolution processing, mapping the extracted high-dimensional features back to the spatial dimension of the original image, and generating a denoised image; S54: weight-fusing the denoised image obtained by deconvolution with the preliminary feature information, and balancing the denoising effect and the retention of image detail features by adjusting the weight ratio, thereby generating a final denoised image.

7. The artificial intelligence-assisted CT image quality enhancement and noise reduction method according to claim 1, characterized in that: The S6 specifically includes: S61: inputting the denoised CT image into a generator in a generative adversarial network framework, wherein the generator is a convolutional neural network structure, which extracts and enhances detail features in the image by performing convolution and deconvolution operations on the input denoised image; S62: The generator extracts feature information from the denoised image through multiple convolutional layers and gradually adjusts the details in the image; S63: inputting the generated enhanced CT image into a discriminator, wherein the discriminator compares the enhanced image with the preset standard image, calculates the difference between the two, and outputs a similarity score; S64: Based on the feedback from the discriminator, the generator will optimize the image generation process by continuously adjusting the weights of the convolution kernel to maximize the similarity between the enhanced image and the standard image. Then, the training of the generative adversarial network is optimized through multiple iterations, and the output gradually approaches the quality of the standard high-dose CT image.

8. The artificial intelligence-assisted CT image quality enhancement and noise reduction method according to claim 7, characterized in that: The S63 specifically includes: S631: The generated enhanced CT image and the preset standard high-dose CT image are sent as input to the convolutional neural network structure of the discriminator, and the feature vectors of the enhanced image and the standard image are extracted respectively through the convolution operation; S632: The discriminator compares the feature vectors of the enhanced image and the standard image, calculates the difference between the two, and obtains the overall difference value of the two images in the feature space by comparing the feature difference of each pixel; S633: Calculate a similarity score based on the calculated difference value, where the similarity score indicates the degree of matching between the enhanced image and the standard image. The higher the score, the closer the two are in details and structure.

9. The artificial intelligence-assisted CT image quality enhancement and noise reduction method according to claim 1, characterized in that: The S7 specifically includes: S71: sending the quality-enhanced CT image as input to an image reconstruction module, and reconstructing the image by using a filtered back projection technique; S72: During the reconstruction process, a high signal-to-noise ratio optimization algorithm is used to optimize the reconstructed image; S73: Use a structural fidelity optimization algorithm to ensure the details and structural fidelity of the reconstructed image. Specifically, the image quality is optimized by calculating the structural similarity index of the image.

10. The artificial intelligence-assisted CT image quality enhancement and noise reduction method according to claim 1, characterized in that: The S8 specifically includes: S81: inputting the reconstructed final CT image into the trained medical image discrimination model; S82: The discriminant model extracts features from the input CT image through multiple convolutional layers and pooling layers to extract structural information, texture features, and grayscale distribution in the image; S83: The discriminant model inputs the extracted image features into a preset classification algorithm through the classification layer to evaluate the diagnostic suitability of the image; S84: Based on the classification results, the discriminant model outputs a medical suitability score ranging from 0 to 1, which is used to reflect whether the image is suitable for clinical diagnosis. If the score reaches a preset threshold, it indicates that the CT image meets the clinical diagnosis requirements.

Images