Intelligent optimization method for ultrasonic image quality enhancement

Abstract

The invention discloses an intelligent optimization method for ultrasonic image quality enhancement. The intelligent optimization method comprises the following steps: acquiring an original ultrasonic image through ultrasonic imaging equipment; performing multi-stage preprocessing on the obtained original ultrasonic image; extracting multi-scale features from the ultrasonic image subjected to the multi-stage preprocessing; performing intelligent scoring on the images subjected to multi-stage preprocessing and multi-scale feature extraction; dynamically optimizing the processing parameters under the guidance of the image quality score; identifying and emphatically enhancing a target area in the ultrasonic image; fusing a local optimization result with the global image; and outputting the optimized ultrasonic image and the corresponding quality scoring report. Compared with the prior art, the ultrasonic image quality enhancement method has the following advantages and effects that the ultrasonic image quality enhancement method provided by the invention has remarkable innovativeness in technology, solves the ultrasonic image quality problem existing in the prior art, provides a more efficient and accurate auxiliary tool for clinical diagnosis, and has wide application prospect and practical value.

Description

Technical Field

[0001] The present invention relates to the field of ultrasonic image technology, and in particular to an intelligent optimization method for ultrasonic image quality enhancement. Background Art

[0002] Ultrasound imaging technology, as a non-invasive, real-time, and economical medical imaging diagnostic tool, has been widely used in clinical diagnosis and treatment. However, due to the influence of sound wave propagation characteristics, tissue structure complexity, equipment performance limitations, and environmental noise during ultrasound imaging, the imaging quality often deteriorates. This quality degradation is mainly manifested in noise interference, artifact generation, insufficient contrast, and loss of details, which directly affect the accuracy and reliability of diagnosis.

[0003] First, there are three common problems with ultrasound images: noise interference, artifacts affecting imaging quality, insufficient contrast, and loss of edge information. Traditional methods for image denoising, artifact repair, and contrast enhancement mostly use fixed rule algorithms, which lack pertinence and intelligence and can easily lead to loss of details or over-processing. At the same time, these methods often lack dynamic adjustment capabilities and cannot meet complex and diverse clinical needs.

[0004] Secondly, most existing technologies process the entire image in a unified manner and lack targeted optimization of local feature areas, which easily leads to loss of local details. At the same time, there is a lack of parameter adjustment mechanism based on real-time feedback of image quality, and it is unable to adapt to different imaging environments, individual differences of patients and equipment characteristics.

[0005] With the rapid development of artificial intelligence technology, deep learning-based image processing methods are gradually being applied to the field of medical image quality enhancement. The introduction of technologies such as convolutional neural networks, generative adversarial networks, and reinforcement learning has provided new solutions for ultrasound image denoising, artifact correction, contrast enhancement, and edge feature extraction.

[0006] However, although deep learning technology has made certain progress in the field of medical imaging, its application in ultrasound image processing is still imperfect, especially in terms of real-time and personalized optimization, where there is much room for improvement. Summary of the invention

[0007] The purpose of the present invention is to provide an intelligent optimization method for ultrasound image quality enhancement to solve the problems raised in the background technology.

[0008] The above technical objectives of the present invention are achieved through the following technical solutions: An intelligent optimization method for ultrasound image quality enhancement includes the following steps: Step S1, obtaining an original ultrasound image by an ultrasound imaging device, wherein the original ultrasound image includes but is not limited to a two-dimensional ultrasound image, a three-dimensional ultrasound image, a color Doppler image or an elastic imaging image; Step S2, performing multi-level preprocessing on the acquired original ultrasound image; Step S3, extracting multi-scale features from the ultrasound image after multi-level preprocessing based on a deep learning model of a convolutional neural network; Step S4, using a quality assessment model trained by a generative adversarial network to intelligently score the images that have undergone multi-level preprocessing and multi-scale feature extraction, and generate an image quality scoring matrix, where the scoring indicators include signal-to-noise ratio, edge clarity, grayscale consistency, and contrast balance; Step S5: using reinforcement learning technology to dynamically optimize processing parameters under the guidance of image quality scores, the optimization content includes but is not limited to denoising intensity, sharpness, contrast adjustment coefficient and local enhancement weight; Step S6: using an attention mechanism-based optimization network to identify and focus on enhancing target areas in the ultrasound image, where the target areas include abnormal tissues, pathological structures, and specific anatomical parts; Step S7: By using an adaptive weight allocation mechanism, the local optimization result is integrated with the global image to ensure the overall coherence and consistency of the enhanced ultrasound image; Step S8: outputting the optimized ultrasound image and its corresponding quality score report, the report content covers parameter optimization details, quantitative analysis of quality improvement and description of key optimization areas.

[0009] Further configuration is that the step S2 is specifically: Step S21, perform denoising processing, the spatial domain filter retains edge details and reduces noise through weighted median filtering; the frequency domain wavelet transform decomposes the noise signal at multiple scales and removes artifact interference step by step, while retaining important high-frequency features; by fusing the results of the two filtering methods, the smoothness and clarity of the image are further improved; Step S22, eliminating interference of artifacts generated during the imaging process on diagnostic accuracy by image artifact correction, while retaining key features and texture details in the original ultrasound image to the maximum extent; Step S23: Adopting an adaptive histogram equalization algorithm to enhance the contrast of the global and local areas.

[0010] Further configuration is that the step S22 is specifically: Step S221, artifact area detection; Step S222, artifact area classification; Step S223: repair artifacts.

[0011] Further configuration is that the step S221 is specifically as follows: Step S2211, using a histogram to analyze the grayscale distribution in the original ultrasound image, and calculating the local variance and global uniformity index of the grayscale value; the artifact area is usually manifested as an abnormally low grayscale or high grayscale area, and its local variance is significantly lower than that of the normal tissue area; Step S2212, extracting texture features of the image through multi-scale texture analysis based on wavelet transform; the texture details of the artifact area are less and discontinuous, and there is a significant difference in texture distribution from the normal tissue area; Step S2213: using a multi-scale edge detection technique based on the Canny algorithm to identify edge structures in the image; artifact regions usually have blurred edges and abnormal edge gradient distribution; Step S224: Based on the grayscale and texture analysis and guided by edge features, the artifact detection range is further expanded through an adaptive region growing algorithm, and an artifact region mask is generated through iterative calculation.

[0012] Further configuration is that the step S222 is specifically: Step S2221: assign three core feature weights to each artifact area, including grayscale mean weight W1, texture complexity weight W2 and edge clarity weight W3, where: W1 is calculated based on the difference between the grayscale mean of the artifact area and the surrounding tissue. The greater the difference, the higher the weight; W2 is determined by calculating the texture entropy value of the artifact area. The simpler the texture, the higher the weight; W3 is determined by the gradient strength and number of edges of the edge detection algorithm, and areas with blurred or broken edges have higher weights; Step S2222, combining the feature weights, using the weighted scoring formula S=αW1+βW2+γW3, where α, β, γ are dynamically adjusted weight coefficients, and dividing the artifact area into low impact level, medium impact level and high impact level according to the score.

[0013] Further configuration is that the step S223 is specifically: Step S2231, repairing the low-impact level artifact area; A method based on local texture interpolation is used for lightweight repair. By calculating the grayscale gradient of the artifact area and the adjacent normal tissue, a dynamic interpolation matrix is ​​generated to perform weighted compensation on the pixels in the artifact area to ensure a continuous transition between the grayscale value of the artifact after repair and the surrounding tissue. Step S2232, repairing the medium impact level artifact area; Medium-impact artifacts usually have a large area, and the grayscale and texture are abnormally significant, so a regional repair strategy based on directional wavelet transform is adopted; Directional wavelet transform is used to decompose the artifact area into sub-bands of different scales and directions; The characteristic distribution of artifact area and normal area is estimated by Gaussian mixture model; In each sub-band, the artifact pixels are gradually restored according to the hybrid model results, and the sub-bands are synthesized to reconstruct the repaired artifact area; Step S2233, repairing the high impact level artifact area; High-impact artifacts have the greatest impact on diagnosis, so a global inpainting network based on deep learning is used for pixel reconstruction; The global repair network is a generative adversarial network, including a generator and a discriminator: The generator inputs the grayscale distribution and edge mask of the artifact area, and generates the repair result by learning the correlation between the artifact area and the surrounding normal area; The discriminator is used to determine whether the restoration result has real texture characteristics, and the feedback is used to optimize the generator performance; The final output restoration result is fused with the original ultrasound image to ensure the consistency and integrity of the overall image.

[0014] Further configuration is that the step S3 is specifically: The input layer in the deep learning model receives multi-channel image input, including the original ultrasound image, the image after contrast enhancement in step S23, and the edge detection result; The middle layer in the deep learning model is a multi-scale convolutional neural network with kernel sizes of 3×3, 5×5, and 7×7 to extract local details and global information simultaneously; The output layer in the deep learning model generates a feature map for subsequent quality assessment, which contains global texture information and local structural details.

[0015] Further configuration is that the step S4 is specifically: The discriminator is used to distinguish low-quality images from optimized high-quality images; The generator dynamically generates optimized images and iteratively optimizes them in the feedback loop with the discriminator; The image quality score is dynamically calculated by fusing the judgment result of the discriminator with the loss function of the generator, and finally generates a comprehensive score of multiple indicators such as signal-to-noise ratio, clarity, and contrast.

[0016] Further configuration is that the step S6 is specifically: Through the channel attention mechanism, important features related to the target area in the key channels are identified; through the spatial attention mechanism, the enhancement weight of each pixel is dynamically allocated to highlight abnormal tissues, lesion areas and key anatomical structures; the attention mask is used to generate a local enhanced image, which is combined with the global optimized image through a feature fusion network.

[0017] Further configuration is that the step S7 is specifically: According to the image quality score matrix, the fusion weight of the local enhanced area and the global image is dynamically adjusted; The fusion strategy is optimized by genetic algorithm, so that the optimized image can take into account both local clarity and global smoothness as a whole; Combined with regional reconstruction technology, gradient correction is performed on over-enhanced areas to avoid image artifacts or texture breaks.

[0018] In view of the shortcomings of the existing technology, the present invention proposes an intelligent optimization method for ultrasound image quality enhancement. By combining deep learning, reinforcement learning and multi-dimensional image processing technology, the quality of ultrasound images is significantly improved. The invention has the following innovative features: High-precision design for artifact detection and repair: The deep learning model is combined with multi-scale feature extraction and dynamic artifact grading methods to significantly improve the accuracy of artifact detection. A multi-level repair strategy is adopted for different types of artifacts to ensure the grayscale consistency and texture integrity of the artifact repair area.

[0019] Dynamic parameter adaptive optimization: Through reinforcement learning technology, denoising, enhancement and restoration parameters are adjusted in real time, so that the processing process is dynamically optimized according to image quality feedback to meet the needs of different imaging scenarios.

[0020] Focused optimization of local areas: Introducing the attention mechanism to locally enhance abnormal areas and key diagnostic parts in the image, highlighting the edge features and detail information of abnormal tissues, and further improving the accuracy of diagnosis.

[0021] Multi-task processing framework: This method implements ultrasound image denoising, artifact correction, contrast enhancement, and edge feature extraction in a unified framework, ensuring that each functional module works together to improve processing efficiency.

[0022] Real-time and scalability: By optimizing the algorithm structure and computing resource allocation, real-time processing of ultrasound image quality enhancement is achieved, and the output characteristics of different imaging devices are compatible, which is convenient for clinical promotion.

[0023] In summary, the ultrasound image quality enhancement method proposed in the present invention is significantly innovative in technology, solves the ultrasound image quality problems existing in the prior art, provides a more efficient and accurate auxiliary tool for clinical diagnosis, and has broad application prospects and practical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a schematic diagram of the process of an embodiment; Figure 2 Schematic diagram of data comparison of original ultrasound images and optimized ultrasound images of the thyroid gland of 20 patients in the embodiment. DETAILED DESCRIPTION

[0025] The present invention is further described in detail below in conjunction with the accompanying drawings.

[0026] As attached Figure 1 As shown; This embodiment discloses an intelligent optimization method for enhancing ultrasound image quality, comprising the following steps: Step S1, obtaining an original ultrasound image by an ultrasound imaging device, wherein the original ultrasound image includes but is not limited to a two-dimensional ultrasound image, a three-dimensional ultrasound image, a color Doppler image or an elastic imaging image; Step S2, performing multi-level preprocessing on the acquired original ultrasound image; Step S3, extracting multi-scale features from the ultrasound image after multi-level preprocessing based on a deep learning model of a convolutional neural network; Step S4, using a quality assessment model trained by a generative adversarial network to intelligently score the images that have undergone multi-level preprocessing and multi-scale feature extraction, and generate an image quality scoring matrix, where the scoring indicators include signal-to-noise ratio, edge clarity, grayscale consistency, and contrast balance; Step S5: using reinforcement learning technology to dynamically optimize processing parameters under the guidance of image quality scores, the optimization content includes but is not limited to denoising intensity, sharpness, contrast adjustment coefficient and local enhancement weight; Step S6: using an attention mechanism-based optimization network to identify and focus on enhancing target areas in the ultrasound image, where the target areas include abnormal tissues, pathological structures, and specific anatomical parts; Step S7: By using an adaptive weight allocation mechanism, the local optimization result is integrated with the global image to ensure the overall coherence and consistency of the enhanced ultrasound image; Step S8: outputting the optimized ultrasound image and its corresponding quality score report, the report content covers parameter optimization details, quantitative analysis of quality improvement and description of key optimization areas.

[0027] Wherein, step S2 is specifically as follows: Step S21, perform denoising processing, the spatial domain filter retains edge details and reduces noise through weighted median filtering; the frequency domain wavelet transform decomposes the noise signal at multiple scales and removes artifact interference step by step, while retaining important high-frequency features; by fusing the results of the two filtering methods, the smoothness and clarity of the image are further improved; Step S22, eliminating interference of artifacts generated during the imaging process on diagnostic accuracy by image artifact correction, while retaining key features and texture details in the original ultrasound image to the maximum extent; Step S23: Adopting an adaptive histogram equalization algorithm to enhance the contrast of the global and local areas.

[0028] Wherein, step S22 is specifically as follows: Step S221, artifact area detection; Step S222, artifact area classification; Step S223: repair artifacts.

[0029] Wherein, step S221 is specifically as follows: Step S2211, using a histogram to analyze the grayscale distribution in the original ultrasound image, and calculating the local variance and global uniformity index of the grayscale value; the artifact area is usually manifested as an abnormally low grayscale or high grayscale area, and its local variance is significantly lower than that of the normal tissue area; Step S2212, extracting texture features of the image through multi-scale texture analysis based on wavelet transform; the texture details of the artifact area are less and discontinuous, and there is a significant difference in texture distribution from the normal tissue area; Step S2213: using a multi-scale edge detection technique based on the Canny algorithm to identify edge structures in the image; artifact regions usually have blurred edges and abnormal edge gradient distribution; Step S224: Based on the grayscale and texture analysis and guided by edge features, the artifact detection range is further expanded through an adaptive region growing algorithm, and an artifact region mask is generated through iterative calculation.

[0030] Wherein, step S222 is specifically as follows: Step S2221: assign three core feature weights to each artifact area, including grayscale mean weight W1, texture complexity weight W2 and edge clarity weight W3, where: W1 is calculated based on the difference between the grayscale mean of the artifact area and the surrounding tissue. The greater the difference, the higher the weight; W2 is determined by calculating the texture entropy value of the artifact area. The simpler the texture, the higher the weight; W3 is determined by the gradient strength and number of edges of the edge detection algorithm, and areas with blurred or broken edges have higher weights; Step S2222, combined with the feature weights, use the weighted scoring formula S=αW1+βW2+γW3, where α, β, γ are dynamically adjusted weight coefficients, and divide the artifact area into low impact level (0 to 0.3 and excluding 0.3), medium impact level (0.3 to 0.7 and including 0.3 and 0.7) and high impact level (0.7 to 1.0 and excluding 0.7) according to the score.

[0031] Wherein, step S223 is specifically as follows: Step S2231, repairing the low-impact level artifact area; A method based on local texture interpolation is used for lightweight repair. By calculating the grayscale gradient of the artifact area and the adjacent normal tissue, a dynamic interpolation matrix is ​​generated to perform weighted compensation on the pixels in the artifact area to ensure a continuous transition between the grayscale value of the artifact after repair and the surrounding tissue. Step S2232, repairing the medium impact level artifact area; Medium-impact artifacts usually have a large area, and the grayscale and texture are abnormally significant, so a regional repair strategy based on directional wavelet transform is adopted; Directional wavelet transform is used to decompose the artifact area into sub-bands of different scales and directions; The characteristic distribution of artifact area and normal area is estimated by Gaussian mixture model; In each sub-band, the artifact pixels are gradually restored according to the hybrid model results, and the sub-bands are synthesized to reconstruct the repaired artifact area; Step S2233, repairing the high impact level artifact area; High-impact artifacts have the greatest impact on diagnosis, so a global inpainting network based on deep learning is used for pixel reconstruction; The global repair network is a generative adversarial network, including a generator and a discriminator: The generator inputs the grayscale distribution and edge mask of the artifact area, and generates the repair result by learning the correlation between the artifact area and the surrounding normal area; The discriminator is used to determine whether the restoration result has real texture characteristics, and the feedback is used to optimize the generator performance; The final output restoration result is fused with the original ultrasound image to ensure the consistency and integrity of the overall image.

[0032] Wherein, step S3 is specifically as follows: The input layer in the deep learning model receives multi-channel image input, including the original ultrasound image, the image after contrast enhancement in step S23, and the edge detection result; The middle layer in the deep learning model is a multi-scale convolutional neural network with kernel sizes of 3×3, 5×5, and 7×7 to extract local details and global information simultaneously; The output layer in the deep learning model generates a feature map for subsequent quality assessment, which contains global texture information and local structural details.

[0033] Wherein, step S4 is specifically as follows: The discriminator is used to distinguish low-quality images from optimized high-quality images; The generator dynamically generates optimized images and iteratively optimizes them in the feedback loop with the discriminator; The image quality score is dynamically calculated by fusing the judgment result of the discriminator with the loss function of the generator, and finally generates a comprehensive score of multiple indicators such as signal-to-noise ratio, clarity, and contrast.

[0034] Wherein, step S6 is specifically as follows: Through the channel attention mechanism, important features related to the target area in the key channels are identified; through the spatial attention mechanism, the enhancement weight of each pixel is dynamically allocated to highlight abnormal tissues, lesion areas and key anatomical structures; the attention mask is used to generate a local enhanced image, which is combined with the global optimized image through a feature fusion network.

[0035] Wherein, step S7 is specifically as follows: According to the image quality score matrix, the fusion weight of the local enhanced area and the global image is dynamically adjusted; The fusion strategy is optimized by genetic algorithm, so that the optimized image can take into account both local clarity and global smoothness as a whole; Combined with regional reconstruction technology, gradient correction is performed on over-enhanced areas to avoid image artifacts or texture breaks.

[0036] Application Examples Example 1: Quality enhancement of abdominal ultrasound images Step S1, collecting abdominal ultrasound images of the patient, the image types including two-dimensional grayscale images and color Doppler images, for diagnosing liver cysts or liver cirrhosis.

[0037] There are a lot of speckle noise and artifacts in the original ultrasound image, the contrast of the target area (liver lesion) is low, and the texture details are blurred.

[0038] Step S2: Multi-stage preprocessing a) Apply the adaptive dual-domain filtering algorithm: Spatial domain: Use weighted median filtering to retain edge details of the lesion area and reduce noise interference.

[0039] Frequency domain: Use wavelet transform to decompose the noise signal and remove artifact interference step by step.

[0040] Output results: Speckle noise is greatly reduced, and the liver edge and lesions are initially highlighted.

[0041] b) Image artifact correction Identify image artifacts that may affect the diagnosis.

[0042] A multi-scale artifact reconstruction algorithm is used to dynamically repair artifact areas, and an adaptive weighted interpolation method is used to restore texture consistency.

[0043] Output result: The artifact area is smoothed and compensated, and the liver contour is clearer.

[0044] c) Dynamic contrast adjustment An adaptive histogram equalization algorithm is used to enhance the contrast of local liver areas, making the grayscale difference between lesions and healthy tissues more obvious.

[0045] Output result: The brightness and contrast of the lesion area are improved, and the detailed texture gradually emerges.

[0046] Step S3: extracting key features of the liver region based on a multi-scale convolutional neural network: Input: Multi-channel input including original image, enhanced image and edge detection results.

[0047] Network structure: The convolution kernel sizes are 3×3, 5×5, and 7×7, which extract local edge details, texture features, and global grayscale distribution information, respectively.

[0048] Output: Generate a set of multi-scale feature maps for subsequent image quality assessment and enhancement.

[0049] Step S4: Quality assessment module based on generative adversarial network: Generator: Dynamically optimizes the input image to generate the reference image required for quality scoring.

[0050] Discriminator: Compares low-quality images with optimized images and provides scores such as signal-to-noise ratio, clarity, and contrast.

[0051] Quality score matrix: contains global signal-to-noise ratio (PSNR), structural similarity (SSIM), and contrast enhancement coefficient.

[0052] Output result: Evaluate the quality score of the original image as a basis for optimizing the processing parameters.

[0053] Step S5: Use the deep Q learning algorithm to build a dynamic parameter optimization model: Reward function: The improvement in image quality score is used as a reward to guide the optimization direction.

[0054] Dynamic adjustment: optimize denoising strength, sharpening level, contrast adjustment coefficient and local enhancement weight.

[0055] Output results: The automatically optimized parameter configuration is used to guide subsequent refinement enhancement and global fusion.

[0056] Step S6: Optimize the network based on the attention mechanism to focus on identifying and enhancing the lesion area in the liver: Channel attention mechanism: Dynamically adjust the weights of different feature channels to enhance high-frequency features related to lesions.

[0057] Spatial attention mechanism: Dynamically assign enhancement weights to each pixel to highlight the edges and textures of lesions.

[0058] Target region attention mechanism: Use segmentation masks to guide the network to focus on liver lesions and reduce background interference.

[0059] Output results: The texture, edge clarity and contrast of the lesion area are significantly improved, making it easier for doctors to accurately identify lesions.

[0060] Step S7: Use an adaptive weighted fusion algorithm to fuse the local optimization result with the global image: Fusion strategy: Dynamically adjust the weights of the local enhanced area and the global image according to the image quality score matrix.

[0061] Avoid over-enhancement: Perform gradient correction on over-enhanced areas to ensure image coherence.

[0062] Output result: The overall image has both clarity of local details and global smoothness and consistency.

[0063] Step S8: Output content: Optimized high-quality ultrasound images in supported formats: DICOM, JPEG, and PNG.

[0064] Image quality rating report, including: Quality comparison before and after optimization; Parameter adjustment records; Enhanced details of the lesion area.

[0065] Supported functions: Real-time optimized output meets clinical real-time needs through GPU acceleration technology.

[0066] Remote image transmission and cloud storage facilitate sharing and collaboration among multiple institutions.

[0067] Experimental results: Original image: There is a lot of noise interference, and the lesion area is blurred and difficult to distinguish.

[0068] Optimized images: Significant noise suppression, clear edges and contrast enhancement of liver lesion areas, and clear texture details facilitate the diagnosis of liver cysts and their size, location, and morphology.

[0069] This embodiment demonstrates how to improve the quality of ultrasound images through intelligent optimization methods, especially achieving remarkable results in lesion identification and detail enhancement, providing reliable support for ultrasound image diagnosis.

[0070] Example 2: Quality enhancement of thyroid ultrasound images In thyroid ultrasound examination, due to the influence of speckle noise and artifacts, small lesions (such as nodules or tiny calcifications) are often difficult to identify, which poses a challenge to doctors' diagnosis. The present invention removes noise, improves the contrast and clarity of small thyroid lesions, and retains the detailed information of normal thyroid tissue.

[0071] Step S1: Acquire thyroid two-dimensional grayscale ultrasound images using ultrasound equipment, including thyroid image data of 20 patients.

[0072] The image resolution is 800 × 600 pixels and includes thyroid tissue, surrounding muscles, and small nodule areas.

[0073] Step S2: Multi-stage preprocessing The original image is processed by adaptive dual-domain filtering: Spatial domain: Use weighted median filtering to reduce speckle noise.

[0074] Frequency domain: Decompose and suppress high-frequency noise based on wavelet transform, and retain tissue boundary details.

[0075] Output results: The speckle noise is significantly reduced and the preliminary outline of the thyroid nodule is preserved.

[0076] Artifact Correction: Fixed texture discontinuities caused by artifacts.

[0077] Output results: The false signals caused by artifacts in the image are greatly reduced, and the thyroid tissue texture is more coherent.

[0078] Dynamic contrast adjustment: Adaptive histogram equalization algorithm is used to enhance the local contrast of the image and highlight the thyroid nodule area.

[0079] Output result: The brightness of the nodule area forms a significant contrast with the background.

[0080] Step S3: Use the trained convolutional neural network to extract multi-scale features: Input: original image, enhanced image and edge detection result.

[0081] Middle layer: Extract texture, edge and grayscale distribution features through multi-scale convolution kernels (3×3, 5×5, 7×7).

[0082] Output: A set of feature maps containing details of the target area (nodule edge, texture, and grayscale) for subsequent quality assessment and local enhancement.

[0083] Results: The network accurately captured the edge features of the thyroid nodule region while reducing the interference of irrelevant background information.

[0084] Step S4: Use the generative adversarial network to score the image quality: The discriminator evaluates the quality difference between the images before and after optimization.

[0085] Rating indicators: Peak signal-to-noise ratio (PSNR): 22.1 before optimization and increased to 34.7 after optimization.

[0086] Structural similarity (SSIM): 0.68 before optimization and increased to 0.89 after optimization.

[0087] Contrast enhancement factor: 1.2 before optimization and 2.6 after optimization.

[0088] Output results: The quality score shows that the clarity and contrast of the optimized image are significantly improved.

[0089] Step S6: Optimize the network through the attention mechanism to enhance the nodule area: Channel attention: Strengthen the weights of high-frequency channels related to nodule features.

[0090] Spatial attention: identifying nodule regions and dynamically assigning enhancement weights.

[0091] Output result: The boundary of the nodule is clearer and the texture details around the nodule are restored.

[0092] Step S7: Apply an adaptive weighted fusion algorithm to integrate the locally enhanced nodule region with the global image: Ensure the smoothness of the entire image while highlighting the details of the nodule area.

[0093] Results: The overall image consistency was good, and the nodule area was significantly enhanced without over-processing.

[0094] Step S8: Output content: Export optimized thyroid ultrasound images in DICOM and PNG formats.

[0095] Automatically generate quality reports: Comparison chart before and after optimization; Image quality scores (PSNR, SSIM, contrast enhancement factor, etc.); Detailed description of the nodular enhancing area.

[0096] The experimental results are as attached Figure 2 As shown; Small nodules that were difficult to distinguish in the original images have clear edges and enhanced textures in the optimized images, which significantly improves the doctor's diagnostic efficiency and accuracy.

[0097] The optimized images are used by doctors to diagnose malignant thyroid nodules (such as thyroid cancer), showing higher diagnostic sensitivity.

[0098] Based on the thyroid ultrasound images of 20 patients, the optimized effect was quantitatively verified and has significant clinical application value.

[0099] This specific embodiment is merely an explanation of the present invention and is not a limitation of the present invention. After reading this specification, those skilled in the art may make non-creative modifications to the present embodiment as needed. However, as long as they are within the scope of the claims of the present invention, they are protected by the patent law.

Claims

1. An intelligent optimization method for ultrasound image quality enhancement, characterized in that: The following steps are included: Step S1, obtaining an original ultrasound image by an ultrasound imaging device, wherein the original ultrasound image includes but is not limited to a two-dimensional ultrasound image, a three-dimensional ultrasound image, a color Doppler image or an elastic imaging image; Step S2, performing multi-level preprocessing on the acquired original ultrasound image; Step S3, extracting multi-scale features from the ultrasound image after multi-level preprocessing based on a deep learning model of a convolutional neural network; Step S4, using a quality assessment model trained by a generative adversarial network to intelligently score the images that have undergone multi-level preprocessing and multi-scale feature extraction, and generate an image quality scoring matrix, where the scoring indicators include signal-to-noise ratio, edge clarity, grayscale consistency, and contrast balance; Step S5: using reinforcement learning technology to dynamically optimize processing parameters under the guidance of image quality scores, the optimization content includes but is not limited to denoising intensity, sharpness, contrast adjustment coefficient and local enhancement weight; Step S6: using an attention mechanism-based optimization network to identify and focus on enhancing target areas in the ultrasound image, where the target areas include abnormal tissues, pathological structures, and specific anatomical parts; Step S7: By using an adaptive weight allocation mechanism, the local optimization result is integrated with the global image to ensure the overall coherence and consistency of the enhanced ultrasound image; Step S8: outputting the optimized ultrasound image and its corresponding quality score report, the report content covers parameter optimization details, quantitative analysis of quality improvement and description of key optimization areas.

2. The intelligent optimization method for ultrasonic image quality enhancement according to claim 1, characterized in that: The step S2 is specifically as follows: Step S21, perform denoising processing, the spatial domain filter retains edge details and reduces noise through weighted median filtering; the frequency domain wavelet transform decomposes the noise signal at multiple scales and removes artifact interference step by step, while retaining important high-frequency features; by fusing the results of the two filtering methods, the smoothness and clarity of the image are further improved; Step S22, eliminating interference of artifacts generated during the imaging process on diagnostic accuracy by image artifact correction, while retaining key features and texture details in the original ultrasound image to the maximum extent; Step S23: Adopting an adaptive histogram equalization algorithm to enhance the contrast of the global and local areas.

3. The intelligent optimization method for ultrasonic image quality enhancement according to claim 2, characterized in that: The step S22 is specifically as follows: Step S221, artifact area detection; Step S222, artifact area classification; Step S223: repair artifacts.

4. The intelligent optimization method for ultrasonic image quality enhancement according to claim 3, characterized in that: The step S221 is specifically as follows: Step S2211, using a histogram to analyze the grayscale distribution in the original ultrasound image, and calculating the local variance and global uniformity index of the grayscale value; the artifact area is usually manifested as an abnormally low grayscale or high grayscale area, and its local variance is significantly lower than that of the normal tissue area; Step S2212, extracting texture features of the image through multi-scale texture analysis based on wavelet transform; the texture details of the artifact area are less and discontinuous, and there is a significant difference in texture distribution from the normal tissue area; Step S2213: using a multi-scale edge detection technique based on the Canny algorithm to identify edge structures in the image; artifact regions usually have blurred edges and abnormal edge gradient distribution; Step S224: Based on the grayscale and texture analysis and guided by edge features, the artifact detection range is further expanded through an adaptive region growing algorithm, and an artifact region mask is generated through iterative calculation.

5. The intelligent optimization method for ultrasonic image quality enhancement according to claim 3, characterized in that: The step S222 is specifically as follows: Step S2221: assign three core feature weights to each artifact area, including grayscale mean weight W1, texture complexity weight W2 and edge clarity weight W3, where: W1 is calculated based on the difference between the grayscale mean of the artifact area and the surrounding tissue. The greater the difference, the higher the weight; W2 is determined by calculating the texture entropy value of the artifact area. The simpler the texture, the higher the weight; W3 is determined by the gradient strength and number of edges of the edge detection algorithm, and areas with blurred or broken edges have higher weights; Step S2222, combining the feature weights, using the weighted scoring formula S=αW1+βW2+γW3, where α, β, γ are dynamically adjusted weight coefficients, and dividing the artifact area into low impact level, medium impact level and high impact level according to the score.

6. The intelligent optimization method for ultrasonic image quality enhancement according to claim 3, characterized in that: The step S223 is specifically as follows: Step S2231, repairing the low-impact level artifact area; A method based on local texture interpolation is used for lightweight repair. By calculating the grayscale gradient of the artifact area and the adjacent normal tissue, a dynamic interpolation matrix is ​​generated to perform weighted compensation on the pixels in the artifact area to ensure a continuous transition between the grayscale value of the artifact after repair and the surrounding tissue. Step S2232, repairing the medium impact level artifact area; Medium-impact artifacts usually have a large area, and the grayscale and texture are abnormally significant, so a regional repair strategy based on directional wavelet transform is adopted; Directional wavelet transform is used to decompose the artifact area into sub-bands of different scales and directions; The characteristic distribution of artifact area and normal area is estimated by Gaussian mixture model; In each sub-band, the artifact pixels are gradually restored according to the hybrid model results, and the sub-bands are synthesized to reconstruct the repaired artifact area; Step S2233, repairing the high impact level artifact area; High-impact artifacts have the greatest impact on diagnosis, so a global inpainting network based on deep learning is used for pixel reconstruction; The global repair network is a generative adversarial network, including a generator and a discriminator: The generator inputs the grayscale distribution and edge mask of the artifact area, and generates the repair result by learning the correlation between the artifact area and the surrounding normal area; The discriminator is used to determine whether the restoration result has real texture characteristics, and the feedback is used to optimize the generator performance; The final output restoration result is fused with the original ultrasound image to ensure the consistency and integrity of the overall image.

7. The intelligent optimization method for ultrasonic image quality enhancement according to claim 2, characterized in that: The step S3 is specifically as follows: The input layer in the deep learning model receives multi-channel image input, including the original ultrasound image, the image after contrast enhancement in step S23, and the edge detection result; The middle layer in the deep learning model is a multi-scale convolutional neural network with kernel sizes of 3×3, 5×5, and 7×7 to extract local details and global information simultaneously; The output layer in the deep learning model generates a feature map for subsequent quality assessment, which contains global texture information and local structural details.

8. The intelligent optimization method for ultrasound image quality enhancement according to claim 1, characterized in that: The step S4 is specifically as follows: The discriminator is used to distinguish low-quality images from optimized high-quality images; The generator dynamically generates optimized images and iteratively optimizes them in the feedback loop with the discriminator; The image quality score is dynamically calculated by fusing the judgment result of the discriminator with the loss function of the generator, and finally generates a comprehensive score of multiple indicators such as signal-to-noise ratio, clarity, and contrast.

9. The intelligent optimization method for ultrasonic image quality enhancement according to claim 1, characterized in that: The step S6 is specifically as follows: Through the channel attention mechanism, important features related to the target area in the key channels are identified; through the spatial attention mechanism, the enhancement weight of each pixel is dynamically allocated to highlight abnormal tissues, lesion areas and key anatomical structures; the attention mask is used to generate a local enhanced image, which is combined with the global optimized image through a feature fusion network.

10. The intelligent optimization method for ultrasound image quality enhancement according to claim 1, characterized in that: The step S7 is specifically as follows: According to the image quality score matrix, the fusion weight of the local enhanced area and the global image is dynamically adjusted; The fusion strategy is optimized by genetic algorithm, so that the optimized image can take into account both local clarity and global smoothness as a whole; Combined with regional reconstruction technology, gradient correction is performed on over-enhanced areas to avoid image artifacts or texture breaks.

Images