Oral CT image quality enhancement method and device based on generative adversarial network

Abstract

The invention discloses an oral cavity CT image quality enhancement method and device based on a generative adversarial network, and relates to the related field of medical image informatics, and the method comprises the steps: carrying out the denoising and standardization processing of an input oral cavity CT image; performing different-scale feature extraction on the oral CT image by adopting a multi-scale convolutional neural network through a generator network, enhancing image details and contrast, and generating a newly added image; carrying out difference comparison on the newly added image through a discriminator, and feeding back improvement information to guide a generator network to carry out generated image optimization; after an image is generated through the optimized convergent generator network, pseudo-color mapping is carried out on the image, the contrast ratio of the image is improved, the visual features of the structure in the oral cavity in the image are enhanced, and a final oral cavity CT image is obtained. The technical problems of image detail loss and insufficient contrast improvement in existing oral CT image quality enhancement are solved, and the technical effect of effectively enhancing influence details and contrast improvement is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical image informatics, and particularly relates to an oral CT image quality enhancement method and device based on a generative adversarial network. BACKGROUND

[0002] The detail definition and contrast of an oral CT image directly affect oral disease diagnosis and treatment plan formulation, and is a key to guaranteeing the accuracy of clinical diagnosis and treatment. At present, the main methods to solve the problem of improving the quality of oral CT images include traditional filtering denoising technology, single-scale image enhancement algorithm, etc. These traditional methods often lead to blurred edges of the processed image and poor structure visualization effect due to the lack of effective capture of multi-scale features of the image and the difficulty in balancing the relationship between noise removal and detail preservation, which cannot meet the needs of clinical observation of fine structures.

[0003] In the related art, the oral CT image quality enhancement has the technical problems of image detail loss and insufficient contrast improvement. SUMMARY

[0004] The present application provides an oral CT image quality enhancement method and device based on a generative adversarial network, which adopts denoising and standardization preprocessing of the input oral CT image, uses a multi-scale convolutional neural network in the generator network to extract different scale features of the image, enhances the details and contrast to generate new images, compares the differences between the new images and the real images through the discriminator, feeds back the improvement information and guides the optimization of the generator, and after the generator converges, performs pseudo-color mapping and contrast adjustment on the generated images, enhances the visualization features of the oral structure, and obtains the final high-quality oral CT image. The technical problems of image detail loss and insufficient contrast improvement in the existing oral CT image quality enhancement are solved, and the technical effect of effectively enhancing the details and improving the contrast is achieved.

[0005] The present application provides an oral CT image quality enhancement method based on a generative adversarial network, which includes: denoising and standardization processing of the input oral CT image; using a multi-scale convolutional neural network in the generator network to extract different scale features of the oral CT image, and enhancing the details and contrast of the image to generate enhanced new images; comparing the differences between the generated new images through the discriminator, and feeding back the improvement information to guide the optimization of the generated images by the generator network; after the optimized and converged generator network generates the images, performing pseudo-color mapping on the images, improving the contrast of the images, enhancing the visualization features of the oral structure in the images, and obtaining the final oral CT image.

[0006] In a possible implementation, the input oral CT image is denoised and standardized, and the following processing is performed: the oral CT image is denoised by a deep learning denoising network; and the pixel value of the denoised oral CT image is normalized to a standard range, where the standard range is [0, 1] or [-1, 1].

[0007] In a possible implementation, different scale features of the oral CT image are extracted by a generator network using a multi-scale convolutional neural network, and image details and contrast are enhanced to generate an enhanced new image, and the following processing is performed: different scale features of the oral CT image are extracted by the multi-scale convolutional neural network, including global features and local detail features; the local detail features are enhanced in detail, and the contrast is improved; the global features and the local detail features are multi-scale fused and transmitted to the generator network to generate the new image.

[0008] In a possible implementation, the generated new image is compared with the real image by a discriminator, and feedback improvement information is fed back to guide the generator network to optimize the generated image, and the following processing is performed: the image generation difference is extracted by comparing the generated new image with the real image; the feedback improvement information is generated based on the image generation difference, the generator network is subjected to adversarial training, and the output quality of the generator is continuously optimized.

[0009] In a possible implementation, the image generation difference is extracted by comparing the generated new image with the real image, and the following processing is performed: the discriminator uses a multi-scale feature judgment network to evaluate the realism and quality gap of the new image and the real image from different scales; and the generation difference information of different scales is combined to obtain the image generation difference.

[0010] In a possible implementation, the feedback improvement information is generated based on the image generation difference, the generator network is subjected to adversarial training, and the output quality of the generator is continuously optimized, and the following processing is performed: the oral CT image is subjected to local area shielding or image segmentation to construct a reconstruction training sample; and the generator network is trained using the reconstruction training sample, where the generator network learns to automatically reconstruct detail features and quality enhancement in the absence of labeled data by using a self-supervised loss function, optimizes the detail features and structure of the generated image based on reconstruction error, optimizes the output result of the generator network, and can approach the quality of the real oral CT image.

[0011] In a possible implementation, after the image generated by the convergent generator network is optimized, pseudo-color mapping is performed on the image, the contrast of the image is improved, and the visualization features of the intraoral structures in the image are enhanced, to obtain the final oral CT image. The following processing is performed: based on the gray scale distribution of different tissue structures in the CT image, corresponding colors are set for different gray scale values, and a mapping matching list of gray scale-color is established; according to the mapping matching list, pseudo-color processing is performed on the generated image, the structural boundaries and internal textures between different tissues are highlighted by using the visual difference of colors, and the details of different structures are enhanced; the brightness and contrast of the image subjected to the pseudo-color mapping are adjusted, to obtain the final oral CT image, wherein the brightness range of the image is adjusted, the contrast of each tissue structure is enlarged, and the distinguishability of the structural boundaries is enhanced.

[0012] The application also provides an oral CT image quality enhancement device based on a generative adversarial network, which comprises: an image preprocessing module for denoising and standardizing an input oral CT image; an image enhancement module for extracting different scale features of the oral CT image by a generator network using a multi-scale convolutional neural network, enhancing image details and contrast, and generating an enhanced new image; a generated image optimization module for comparing the generated new image with a discriminator, and feeding back improvement information to guide the generator network to optimize the generated image; and a pseudo-color mapping module for performing pseudo-color mapping on the image after the image generated by the convergent generator network is optimized, improving the contrast of the image, enhancing the visualization features of the intraoral structures in the image, and obtaining a final oral CT image.

[0013] The oral CT image quality enhancement method and device based on a generative adversarial network provided in the application first denoises and standardizes an input oral CT image, then extracts different scale features of the oral CT image by a generator network using a multi-scale convolutional neural network, enhances image details and contrast, generates an enhanced new image, compares the generated new image with a discriminator, feeds back improvement information to guide the generator network to optimize the generated image, and finally performs pseudo-color mapping on the image after the image generated by the convergent generator network is optimized, improves the contrast of the image, enhances the visualization features of the intraoral structures in the image, and obtains a final oral CT image. Through the above process, the method and device provided in the application achieve the technical effect of effectively enhancing the details and contrast. BRIEF DESCRIPTION OF DRAWINGS

[0014] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings of the embodiments of the present application will be briefly introduced below, and the flowcharts are used to illustrate the operations performed by the system according to the embodiments of the present application in the present application. It should be understood that the foregoing or the following operations are not necessarily performed in sequence. On the contrary, various steps can be processed in reverse order or simultaneously according to needs. At the same time, other operations can be added to these processes, or one or more steps of operations can be removed from these processes.

[0015] Figure 1 A flowchart of a method for enhancing the quality of oral CT images based on a generative adversarial network is provided for the embodiments of the present application.

[0016] Figure 2 A structural diagram of a device for enhancing the quality of oral CT images based on a generative adversarial network is provided for the embodiments of the present application.

[0017] Legend: image preprocessing module 10, image enhancement module 20, generated image optimization module 30, pseudo-color mapping module 40. DETAILED DESCRIPTION

[0018] In order to further illustrate the technical means and effects adopted by the present application to achieve the predetermined object of the present application, the specific embodiments, structures, features and effects according to the present application will be described in detail below with reference to the drawings and preferred embodiments.

[0019] The embodiments of the present application provide a method for enhancing the quality of oral CT images based on a generative adversarial network, as shown in Figure 1 The method comprises the following steps: Step S100, denoising and standardizing the input oral CT image.

[0020] Specifically, this step is image preprocessing, and a two-stage processing architecture of denoising + standardization is adopted. First, the Gaussian noise, salt and pepper noise and particle noise generated by the device scanning in the oral CT image are removed through a deep learning model, and then the pixel values are mapped to a unified numerical range, providing standardized input for subsequent network training.

[0021] In a possible implementation, the input oral CT image is denoised and standardized, and step S100 further includes step S110 of denoising the oral CT image by using a deep learning denoising network. Specifically, an improved deep learning denoising network based on U-Net is used, and the network includes an encoder, a decoder, and a skip connection structure. The encoder gradually extracts image features and compresses the size through convolution layers and pooling layers; the decoder realizes feature upsampling by using a transposed convolution layer to restore the image resolution; and the skip connection splices the feature maps of the same scale of the encoder and the decoder feature maps to retain edge detail information. In the training process, a mean square error loss function MSE is used, an Adam optimizer is used as the optimizer, the learning rate is 0.001, and the denoised image is output after 100 iterations. For example, for a noisy tooth root canal region image, the network can learn the noise distribution characteristics to remove the noise while maintaining the fine structure of the root canal wall.

[0022] Step S120, the pixel value of the denoised oral CT image is normalized to a standard range, where the standard range is [0, 1] or [-1, 1]. Specifically, the pixel value distribution of the denoised image is counted to determine the maximum value Max and the minimum value Min, and all pixel values are mapped to the target range by using a linear normalization formula. If the standard range is [0, 1], the formula is: normalized pixel value = (original pixel value-Min) / (Max-Min); if the range is [-1, 1], the formula is: normalized pixel value = 2×(original pixel value-Min) / (Max-Min)-1. For example, the pixel value range of a certain denoised oral CT image is [50, 2000], and after normalization in the range [0, 1], the pixel value 50 is mapped to 0, the pixel value 2000 is mapped to 1, and the intermediate pixel values are linearly converted in proportion to ensure uniform image brightness distribution.

[0023] Step S200, the generator network uses a multi-scale convolutional neural network to extract features of different scales from the oral CT image, enhances the image details and contrast, and generates an enhanced new image.

[0024] Specifically, the generator network takes a multi-scale convolutional neural network CNN as the core, extracts features of different scales through a multi-branch structure, combines a detail enhancement module and a feature fusion technology, and outputs a high-quality enhanced image. The generator network as a whole adopts an encoder-decoder architecture, and embeds a multi-scale feature extraction unit and a contrast enhancement layer.

[0025] In one possible implementation, different scale features of the oral CT image are extracted by a multi-scale convolutional neural network in the generator network, and the image details and contrast are enhanced, and an enhanced new image is generated, and step S200 further includes step S210 of extracting different scale features of the oral CT image by a multi-scale convolutional neural network, including global features and local detail features. Specifically, the multi-scale CNN includes three parallel convolution branches, which use different size convolution kernels to extract features: a large-scale branch uses a 7x7 convolution kernel with a step size of 1 to extract global features such as the overall shape of the teeth and the contour of the jawbone; a medium-scale branch uses a 5x5 convolution kernel with a step size of 1 to extract local features such as the texture of the tooth surface and the periodontal membrane area; and a small-scale branch uses a 3x3 convolution kernel with a step size of 1 to extract fine features such as the root canal opening shape and the alveolar bone small defect. A batch normalization layer BN and a ReLU activation function are connected after each branch to enhance the feature expression capability. For example, for an image containing teeth and surrounding soft tissue, the large-scale branch can capture the spatial position relationship of the teeth and the jawbone, and the small-scale branch can identify the tiny cracks on the tooth surface.

[0026] Step S220, the local detail features are enhanced in detail, and the contrast is improved, and the global features and the local detail features are multi-scale fused and transmitted to the generator network to generate the new image. Specifically, the detail enhancement module uses a residual connection structure to enhance the feature map extracted by the small-scale branch. First, the feature channel number is compressed by a 1x1 convolution kernel, then the 3x3 convolution kernel is used to learn the detail texture, and finally the original feature map is added to retain the basic features while enhancing the details. The contrast enhancement is realized by adaptive histogram equalization CLAHE, the clipping limit is set to 2.0, the grid size is 8x8, the gray distribution of the local feature map is adjusted, and the visibility of the dark details is improved. The feature fusion adopts a weighted summation strategy, and the weights of the global features, the medium-scale features, and the enhanced local features are set to 0.3, 0.4, and 0.3 respectively. After fusion, the channel number is unified by a 1x1 convolution kernel, and the new image is generated by the generator decoder. For example, for the image of the alveolar bone absorption area, the detail enhancement module can highlight the edge of the absorption pit, the contrast enhancement can make the trabecular structure clearer, and the fused image can present the overall shape of the alveolar bone and the local fine lesions.

[0027] Step S300, the generated new image is compared by the discriminator, and the improvement information is fed back to guide the generator network to optimize the generated image.

[0028] Specifically, the discriminator employs a multi-scale feature judgment network. By comparing the differences between newly added images and real high-quality images, it outputs the difference and converts it into a feedback signal. The generator parameters are optimized through adversarial training. The training process adopts the core idea of ​​Generative Adversarial Networks (GANs), with the generator and discriminator iteratively trained alternately.

[0029] In one possible implementation, the newly generated image is compared for differences by a discriminator, and feedback information is used to guide the generator network in optimizing the generated image. Step S300 further includes step S310, which compares the newly generated image with the real image to extract the image generation difference. Specifically, the discriminator network includes a feature extraction layer and a classification layer. The feature extraction layer uses a 4×4 convolutional kernel with a stride of 2, for a total of 5 layers, progressively extracting texture, structural, and other features of the image; the classification layer uses a fully connected layer, outputting the probability that the image is real or generated. The difference extraction is achieved by calculating the output difference of the discriminator's intermediate feature layers. The newly generated image and the real image are input into the discriminator to obtain the feature maps of each feature layer, calculate the L1 distance of corresponding pixels, sum them to obtain the single-scale difference, and then perform a weighted average of the multi-scale differences, where the weights are positively correlated with the importance of the feature layers, finally obtaining the image generation difference. For example, if the texture features of the newly generated image at the edge of the teeth differ significantly from the real image, the discriminator will capture this difference through the feature layer and output a higher difference value.

[0030] Step S320: Based on the image generation difference, feedback improvement information is generated, and adversarial training is performed on the generator network to continuously optimize the generator's output quality. Specifically, the feedback improvement information is achieved through backpropagation of the difference, converting the difference into a loss signal for the generator, and using the Adam optimizer to update the generator parameters. During adversarial training, the generator loss function uses a combination of adversarial loss and content loss. The adversarial loss uses binary cross-entropy loss (BCE) to calculate the probability loss of the generator's generated image being misclassified as real by the discriminator; the content loss uses L1 loss to calculate the pixel difference between the generated image and the real image. During training, the discriminator is fixed first to train the generator for 50 iterations, then the generator is fixed to train the discriminator for 10 iterations, alternating until the loss function converges, such as when the loss value stabilizes below 0.01. For example, when the difference is high, the feedback information guides the generator to focus on optimizing the features of the corresponding region, such as adjusting the generation parameters of the tooth edges to make the texture closer to the real image.

[0031] In one possible implementation, the generated new image is compared with the real image to extract the image generation difference. Step S310 further includes step S311, where the discriminator uses a multi-scale feature judgment network to evaluate the realism and quality gap between the new image and the real image at different scales. Specifically, the multi-scale feature judgment network contains three scale branches, which process images at different resolutions: a high-resolution branch (original size) is used to evaluate the realism of details, such as tooth surface texture and root canal details; a medium-resolution branch (half the original size) is used to evaluate the realism of local structures, such as the connection area between teeth and gums; and a low-resolution branch (quarter the original size) is used to evaluate the realism of global structures, such as the overall shape of the jawbone. The feature extraction layer of each branch uses the same convolutional structure, and the classification layer outputs the realism probability at that scale. The quality gap is evaluated by calculating the difference between the generated probability and the real probability at each scale; the larger the difference, the greater the quality gap. For example, if the generation probability of the root canal morphology in the new image is 0.3 at the high-resolution branch, while it is 0.9 in the real image, it indicates a large difference in detail quality at that scale.

[0032] Step S312: Combine the generation difference information at different scales to obtain the image generation difference quantity. Specifically, a weighted fusion strategy is used to combine multi-scale difference information. First, a weight is assigned to each scale: the high-resolution branch weight is set to 0.5 to highlight detail differences, the medium-resolution branch weight is set to 0.3 to balance local and global differences, and the low-resolution branch weight is set to 0.2 to focus on the overall structure. Then, the difference value at each scale, i.e., the absolute difference between the generation probability and the true probability, is calculated, multiplied by the corresponding weight, and summed to obtain the final image generation difference quantity.

[0033] In one possible implementation, feedback improvement information is generated based on the image generation difference, and adversarial training is performed on the generator network to continuously optimize the generator's output quality. Step S320 further includes step S321, which involves performing local region occlusion or image segmentation on the oral CT image to construct reconstruction training samples. Specifically, local region occlusion adopts a random rectangular occlusion strategy, randomly selecting 10-20 rectangular regions on the oral CT image, with a size of 10%-15% of the image, and setting the pixel value of the occluded region to 0 (black) or 255 (white). Image segmentation uses a segmentation model based on Mask R-CNN, with pre-trained weights using the COCO dataset, fine-tuned for the oral CT image to segment key regions such as teeth, gums, and jawbone, and using the segmented regions as reconstruction targets. The reconstruction training samples consist of an input image containing occlusion / segmentation information and a real complete image. For example, the root canal region of a tooth image is occluded to construct a sample pair of occluded image and complete root canal image.

[0034] Step S322: The generator network is trained using the reconstructed training samples. Through a self-supervised loss function, the generator network learns to automatically reconstruct detailed features and enhance quality even with missing label data. Based on reconstruction errors, it optimizes the detailed features and structure of the generated image, thus improving the output quality of the generator network to approximate real oral CT images. Specifically, the self-supervised loss function uses a combination of reconstruction loss and perceptual loss. The reconstruction loss uses L2 loss to calculate the squared pixel difference between the generator's reconstructed image and the real complete image, forcing the generator to learn the complete structure. The perceptual loss extracts features through a pre-trained VGG16 network, calculates the L1 distance between the generated image and the real image in the feature space, and optimizes the texture and detail perception quality of the image. During training, the generator inputs an image containing occlusion / segmentation and outputs a reconstructed image, updating parameters through backpropagation using the self-supervised loss function. For example, for training samples with occluded root canal regions, the generator learns the typical morphology and texture of root canals, gradually optimizing the reconstruction effect to ensure the generated root canal structure is highly consistent with the real image while preserving the detailed features of the surrounding alveolar bone.

[0035] In step S400, after generating images through an optimized and converged generator network, pseudo-color mapping is performed on the images to improve their contrast and enhance the visualization features of intraoral structures, thereby obtaining the final oral CT image.

[0036] Specifically, this step is a crucial part of image post-processing. It converts grayscale images into color images through pseudo-color mapping, and combines brightness and contrast adjustments to highlight the differences in different tissue structures, thereby improving the visualization effect of clinical diagnosis.

[0037] In one possible implementation, after generating images through an optimized and convergent generator network, pseudo-color mapping is performed on the images to improve their contrast and enhance the visualization features of intraoral structures, thus obtaining the final oral CT image. Step S400 further includes step S410, which assigns corresponding colors to different grayscale values ​​based on the grayscale distribution of different oral tissue structures in the CT image, establishing a grayscale-color mapping matching list. Specifically, a large number of high-quality oral CT images are collected, and key structures such as teeth, gingiva, jawbone, root canals, and lesions are extracted using image segmentation technology. The grayscale value distribution range of each structure is statistically analyzed, such as teeth 1500-2000 HU, gingiva -100-200 HU, jawbone 300-1000 HU, root canal 800-1200 HU, and carious teeth 500-800 HU. Based on the HSV color space, corresponding colors are set. For example, teeth are set to white (H=0°, S=0%, V=90%), gums to pink (H=350°, S=50%, V=80%), jawbone to light yellow (H=45°, S=30%, V=85%), root canals to light blue (H=210°, S=60%, V=70%), and cavities to brown (H=30°, S=70%, V=60%). Finally, the HSV color values ​​are converted to the commonly used RGB format in computer image processing using the standard HSV→RGB conversion formula, establishing a mapping matching list of grayscale value ranges to RGB colors. For example, grayscale 1500-2000 corresponds to RGB(255,255,255) (white), and -100-200 corresponds to RGB(204,102,153) (pink), etc.

[0038] Step S420: The generated image is processed using pseudo-color according to the mapping matching list. This utilizes visual color differences to highlight structural boundaries and internal textures between different tissues, enhancing the details of different structures. Specifically, a pixel-by-pixel mapping strategy is adopted. Each pixel of the generated image is traversed, its grayscale value is read, and the grayscale value range-RGB color mapping matching list is queried to find the corresponding RGB color value, which directly replaces the original grayscale pixel value. For pixels whose grayscale values ​​are at the intersection of two ranges, such as 900 HU (between 800-1200 HU for root canals and 300-1000 HU for jawbones), linear interpolation is used to calculate the RGB color, allowing for natural color transitions between different tissues, avoiding color block fragmentation, and enhancing the image's visualization effect. After pseudo-color processing, the boundaries between different tissues, such as the junction of teeth and gums, become clearer due to RGB color differences, and internal textures, such as the trabecular structure of the jawbone, are naturally presented through color gradients.

[0039] Step S430 involves adjusting the brightness and contrast of the pseudo-color mapped image to obtain the final oral CT image. This adjustment enhances the contrast of various tissue structures and improves the distinguishability of structural boundaries by adjusting the image's brightness range. Specifically, the brightness adjustment employs a linear stretching strategy, statistically analyzing the brightness value distribution of the pseudo-color image to determine the minimum brightness value L. min and maximum value L max The formula is: Adjusted brightness = (Original brightness - L) min ) × [Target brightness range / (L) max -L min The target brightness lower limit is set to map the brightness to the range of [10, 245] to avoid overexposure or underexposure. Contrast adjustment uses gamma correction, which determines the gamma value based on the image's brightness histogram. For images with low brightness, the gamma value is set to 0.8 to enhance the contrast in dark areas; for images with high brightness, the gamma value is set to 1.2 to enhance the contrast in bright areas, ultimately outputting high-quality oral CT images.

[0040] This application employs techniques such as denoising and standardization preprocessing of the input oral CT images, extracting features at different scales using a multi-scale convolutional neural network in the generator network to enhance details and contrast and generate new images, comparing the differences between the new images and the real images using a discriminator to provide feedback and guide the generator optimization, and performing pseudo-color mapping and contrast adjustment on the generated images after the generator converges to enhance the visualization features of oral structures and obtain the final high-quality oral CT images. These techniques solve the technical problems of image detail loss and insufficient contrast enhancement in existing oral CT image quality enhancement methods, achieving an effective enhancement of details and contrast.

[0041] In the above text, refer to Figure 1 A method for enhancing the quality of oral CT images based on generative adversarial networks according to embodiments of the present invention is described in detail. Next, reference will be made to... Figure 2 A dental CT image quality enhancement device based on generative adversarial networks according to an embodiment of the present invention is described.

[0042] The oral CT image quality enhancement device based on generative adversarial networks according to embodiments of the present invention addresses the technical problems of image detail loss and insufficient contrast improvement in existing oral CT image quality enhancement methods, achieving effective enhancement of image detail and contrast. The oral CT image quality enhancement device based on generative adversarial networks includes: an image preprocessing module 10, an image enhancement module 20, a generated image optimization module 30, and a pseudo-color mapping module 40.

[0043] The image preprocessing module 10 is used to denoise and standardize the input oral CT images; the image enhancement module 20 is used to extract features of different scales from the oral CT images using a multi-scale convolutional neural network through a generator network, and enhance the image details and contrast to generate enhanced new images; the generated image optimization module 30 is used to perform difference comparison on the generated new images through a discriminator, and provide feedback improvement information to guide the generator network to optimize the generated images; the pseudo-color mapping module 40 is used to perform pseudo-color mapping on the images generated by the optimized and converged generator network, to improve the contrast of the images, enhance the visualization features of the oral cavity structures in the images, and obtain the final oral CT images.

[0044] The detailed description of the specific configuration of the image preprocessing module 10 is explained as follows: As mentioned above, the image preprocessing module 10 can further include: a denoising unit for denoising the oral CT image through a deep learning denoising network; and a normalization unit for normalizing the pixel values ​​of the denoised oral CT image to a standard range, wherein the standard range is [0,1] or [-1,1].

[0045] The image enhancement module 20 is described in detail below: As mentioned above, the generator network uses a multi-scale convolutional neural network to extract features at different scales from the oral CT images, enhances image details and contrast, and generates enhanced new images. The image enhancement module 20 may further include: a feature extraction unit at different scales for extracting features at different scales from the oral CT images using a multi-scale convolutional neural network, including global features and local detail features; and a new image generation unit for enhancing the local detail features and improving contrast, and fusing the global features and local detail features at multiple scales and transmitting them to the generator network to generate the new images.

[0046] The specific configuration of the image generation optimization module 30 is described in detail below: As mentioned above, the newly generated images are compared for differences by a discriminator, and feedback improvement information is provided to guide the generator network in optimizing the generated images. The image generation optimization module 30 may further include: an image generation difference extraction unit for comparing the newly generated images with real images to extract the image generation difference; and an adversarial training unit for generating feedback improvement information based on the image generation difference, performing adversarial training on the generator network, and continuously optimizing the output quality of the generator.

[0047] The image generation difference extraction unit can further include: an evaluation subunit for the discriminator to use a multi-scale feature judgment network to evaluate the realism and quality difference between the new image and the real image at different scales; and a combination subunit for combining the generation difference information at different scales to obtain the image generation difference.

[0048] Specifically, based on the image generation difference, feedback improvement information is generated to perform adversarial training on the generator network, continuously optimizing the generator's output quality. The adversarial training unit may further include: a reconstruction training sample construction subunit for performing local region occlusion or image segmentation on oral CT images to construct reconstruction training samples; and a training subunit for training the generator network using the reconstruction training samples. Through a self-supervised loss function, the generator network learns to automatically reconstruct detailed features and enhance quality even in the absence of labeled data. Based on the reconstruction error, it optimizes the detailed features and structure of the generated images, thereby optimizing the generator network's output results and achieving quality close to that of real oral CT images.

[0049] The specific configuration of the pseudo-color mapping module 40 is described in detail below: As mentioned above, after generating an image through an optimized and converged generator network, pseudo-color mapping is performed on the image to improve its contrast and enhance the visualization features of intraoral structures, thereby obtaining the final oral CT image. The pseudo-color mapping module 40 may further include: a mapping matching list establishment unit for setting corresponding colors for different gray values ​​based on the gray distribution of different oral tissue structures in the CT image, and establishing a gray-color mapping matching list; a pseudo-color processing unit for performing pseudo-color processing on the generated image according to the mapping matching list, using the visual differences in color to highlight the structural boundaries and internal textures between different tissues, and enhancing the details of different structures; and a contrast adjustment unit for adjusting the brightness and contrast of the pseudo-color mapped image to obtain the final oral CT image, wherein adjusting the brightness range of the image amplifies the contrast of each tissue structure and enhances the distinguishability of structural boundaries.

[0050] The oral CT image quality enhancement device based on generative adversarial networks provided in the embodiments of the present invention can execute the oral CT image quality enhancement method based on generative adversarial networks provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the method.

[0051] Although this application makes various references to certain modules in the system according to the embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy distinction between each other and are not used to limit the scope of protection of this invention.

[0052] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for enhancing oral CT image quality based on a generative adversarial network, characterized in that, The method comprises the following steps: performing denoising and standardization processing on the input oral CT image, including: performing denoising processing on the oral CT image by using a deep learning denoising network; normalizing the pixel values of the denoised oral CT image to a standard range, wherein the standard range is [0, 1] or [-1, 1]. extracting different scale features of the oral CT image by using a multi-scale convolutional neural network in the generator network, and enhancing image details and contrast to generate an enhanced new image, including: 2.The oral CT image quality enhancement method based on a generative adversarial network according to claim 1, wherein, extracting different scale features of the oral CT image by using a multi-scale convolutional neural network, including global features and local detail features; performing detail feature enhancement on the local detail features, improving the contrast, and performing multi-scale fusion of the global features and the local detail features to the generator network to generate the new image. comparing the generated new image with the real image, extracting the image generation difference, and generating feedback improvement information based on the image generation difference to perform adversarial training on the generator network to continuously optimize the output quality of the generator, including: 3.The oral CT image quality enhancement method based on a generative adversarial network according to claim 1, wherein, extracting the image generation difference by comparing the generated new image with the real image, including: the discriminator uses a multi-scale feature judgment network to evaluate the realism and quality gap of the new image and the real image from different scales; combining the generation difference information of different scales to obtain the image generation difference. 4.The oral CT image quality enhancement method based on a generative adversarial network according to claim 3, characterized in that, based on the image generation difference, generating feedback improvement information to perform adversarial training on the generator network to continuously optimize the output quality of the generator, including: performing local region shielding or image segmentation on the oral CT image to construct a reconstruction training sample; training the generator network using the reconstruction training sample, wherein the generator network learns to automatically reconstruct detail features and quality enhancement under the condition of lacking label data by using a self-supervised loss function, optimizes the detail features and structure of the generated image based on the reconstruction error, optimizes the output result of the generator network, and can approach the quality of the real oral CT image. 5.The oral CT image quality enhancement method based on a generative adversarial network according to claim 4, characterized in that, after generating the image by the optimized and converged generator network, performing pseudo-color mapping on the image to improve the contrast of the image and enhance the visualization features of the structures in the oral cavity in the image to obtain the final oral CT image, including: based on the gray scale distribution of different tissue structures in the CT image, setting corresponding colors for different gray scale values, and establishing a gray scale-color mapping matching list. ​ 6.The oral CT image quality enhancement method based on a generative adversarial network according to claim 4, wherein, ​ ​ ​ 7.The oral CT image quality enhancement method based on a generative adversarial network according to claim 1, wherein, ​ ​ According to the mapping matching list, the generated image is subjected to pseudo-color processing, the structural boundaries and internal textures between different tissues are highlighted by using visual differences of colors, and details of different structures are enhanced; The image subjected to the pseudo-color mapping is subjected to brightness and contrast adjustment, and a final oral CT image is obtained, wherein the brightness range of the image is adjusted, the contrast of each tissue structure is enlarged, and the distinguishability of the structural boundaries is enhanced.

8. The oral CT image quality enhancement device based on a generative adversarial network, characterized by, The device is used to implement the oral CT image quality enhancement method based on the generative adversarial network according to any one of claims 1-7, and the device comprises: An image preprocessing module is configured to perform denoising and standardization processing on the input oral CT image; An image enhancement module is configured to perform feature extraction of different scales on the oral CT image by using a multi-scale convolutional neural network of a generator network, and to enhance image details and contrast, thereby generating an enhanced new image; A generated image optimization module is configured to compare the generated new image with a discriminator, and to feed back improvement information to guide the generator network to optimize the generated image; A pseudo-color mapping module is configured to perform pseudo-color mapping on the generated image after the optimization of the generator network, to improve the contrast of the image, to enhance the visual features of the structures in the oral cavity in the image, and to obtain a final oral CT image.

Images