A rapid detection method for medical image quality

Abstract

The present invention discloses a method for rapid detection of medical image quality. The present invention relates to the technical field of medical image processing, specifically to a method for rapid detection of medical image quality, which is particularly applicable to the automated quality assessment and real-time feedback of medical images such as CT, MRI, X-ray, and ultrasound. The advantages of the present invention are as follows: Through the optimized design of the parallel computing framework (CUDA) and the lightweight convolutional neural network (MobileNetV3-Small), the processing speed of image quality detection is significantly improved. By using the thread block and thread grid structure of CUDA, the feature extraction and classification tasks are divided into multiple parallel sub-tasks, and each sub-task is executed in an independent thread. At the same time, depthwise separable convolution and inverted residual structures are used to reduce the network calculation amount. Through the above technical means, the present invention greatly reduces the end-to-end processing delay and can meet the requirements of real-time quality inspection. Especially in the scenario of batch image processing, the detection efficiency is significantly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and specifically provides a method for rapidly detecting the quality of medical images. Background Art

[0002] Medical images play a crucial role in clinical diagnosis, and their quality directly affects the accuracy of diagnostic results. However, during the processes of image acquisition, transmission, and processing, the quality of medical images often deteriorates due to factors such as equipment performance, patient movement, and noise interference, resulting in problems such as artifacts, excessive noise, and insufficient contrast. Traditional image quality detection mainly relies on manual evaluation, which has problems such as low efficiency, strong subjectivity, and easy missed detection.

[0003] The existing technologies have certain defects. Firstly, the detection speed is slow, making it difficult to meet the requirements of real-time quality inspection, especially in the scenario of batch image processing. Secondly, the adaptability is poor. Existing methods mostly target single quality problems (such as noise or artifacts) and lack comprehensive evaluation of multi-dimensional quality indicators. Thirdly, the dynamic adjustment ability is insufficient, and it is unable to adaptively adjust the detection criteria according to image types, equipment parameters, and clinical requirements. Therefore, we propose a method for rapidly detecting the quality of medical images. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for rapidly detecting the quality of medical images.

[0005] To solve the problems raised in the above background art, the present invention provides the following technical solution: A method for rapidly detecting the quality of medical images, the detection method includes the following specific steps:

[0006] Step 1: Data acquisition. Receive the original DICOM format image data transmitted by a medical imaging device through a protocol compliant with the DICOM standard. The medical imaging device includes but is not limited to CT scanners, MRI imagers, X-ray machines, and ultrasonic diagnostic devices.

[0007] Step 2: Preprocessing:

[0008] S1. Perform gray normalization processing on the original image, and uniformly map the gray value range of the image to a preset interval [0, 1] to eliminate the influence of gray differences when images are acquired by different devices.

[0009] S2. Based on a noise suppression algorithm using wavelet transform, suppress the noise in the image.

[0010] S3. Detect the edge information in the image through the Canny edge detection algorithm, and then use a morphological closing operation (kernel size of 5×5) to process the edge information to locate the artifact area.

[0011] S4. Automatically crop the image ROI. Based on the image type and prior knowledge of the examination site, determine the ROI area and standardize and scale it to a preset size (224×224 pixels) to reduce the amount of data for subsequent processing.

[0012] Step 3: Multi-dimensional feature extraction:

[0013] S1. Extracting the clarity, contrast, and noise distribution characteristics of the image through a lightweight convolutional neural network. The lightweight convolutional neural network is a MobileNetV3-Small network. Its input layer receives the normalized image data after grayscale normalization, noise suppression, ROI automatic cropping, and normalization scaling in step 2. The network performs feature extraction through depthwise separable convolution and inverse residual structure, and the output layer generates quantized feature vectors of clarity, contrast, and noise distribution.

[0014] S2. Calculate the structural similarity index (SSIM) and peak signal-to-noise ratio (PSNR) using an image processing algorithm to obtain the image quality characteristics at the structural and pixel levels. The specific algorithm for calculating the structural similarity index (SSIM) is as follows:

[0015] Let the original image be I and the image to be evaluated be J. Divide the image into N×N sub-blocks. For each sub-block i, calculate its mean μ Ii 、μ Ji ,variance and covariance σ IiJi , calculate the SSIM value of each sub-block according to the following formula:

[0016]

[0017] Where: C1=(k1L) 2 , C2=(k2L) 2 , L is the dynamic range of the image pixel value, k1 = 0.01, k2 = 0.03, the SSIM values of all sub-blocks are averaged to obtain the structural similarity index (SSIM) of the entire image;

[0018] The algorithm for calculating the peak signal-to-noise ratio (PSNR) is as follows: First, the mean square error (MSE) between the original image I and the image to be evaluated J is calculated using the formula:

[0019]

[0020] Where: M and N represent the number of rows and columns of the image respectively. Then the peak signal-to-noise ratio (PSNR) is calculated according to the following formula:

[0021]

[0022] Among them, MAX1 represents the maximum possible value of pixels in the image (for an 8-bit grayscale image, MAX1 = 255);

[0023] Step Four: Dynamic Quality Assessment:

[0024] S1. Input the features extracted in Step Three into the classification model. The classification model is a hybrid model of random forest and support vector machine (SVM), and its training data includes a multi-modal medical image dataset labeled as "qualified" and "unqualified", which covers various image types (including CT tomography sequences, MRI multi-parameter imaging, X-ray plain films, and ultrasound dynamic images) and image data of different examination parts;

[0025] S2. Judge whether the image quality is qualified according to a preset dynamic threshold. The dynamic threshold is adaptively adjusted based on the image type, device parameters, and clinical diagnosis requirements. The specific adjustment method includes: loading the corresponding reference quality parameters according to the image device type, which are obtained by analyzing and statistically processing a large amount of image data collected by the same type of device, dynamically weighting the thresholds of different quality indicators (clarity, contrast, noise distribution, SSIM, and PSNR) in combination with the clinical requirements of the examination part. For example, for lung CT images, the threshold weights of clarity and noise distribution may be relatively high. Update the threshold range through online learning based on historical quality inspection data. When new quality inspection data is generated, adjust the threshold range according to the characteristics and quality assessment results of the data;

[0026] Step Five: Result Output:

[0027] S1. Generate a visual report containing quality scores, defect types, and marked abnormal areas, overlay a heat map mark of the artifact area on the original image to intuitively display the location and severity of the artifact area, and at the same time generate a PDF report containing quality scores and recommended processing suggestions ("requiring rescan", "acceptable");

[0028] S2. Real-time feedback to the imaging device and the diagnostic system through the interface, adopt a parallel computing framework to accelerate the feature extraction and classification process, optimize the allocation of computing resources and the thread scheduling of parallel computing, improve the computing efficiency, deploy a lightweight model at the imaging device end, and achieve an end-to-end processing delay ≤ 0.5 seconds to ensure that the results can be fed back in a timely manner.

[0029] As a further solution of the present invention: In Step Two, the noise suppression algorithm based on wavelet transform is specifically: Select the db4 wavelet basis function, perform 3-layer wavelet decomposition on the original image, decompose the image into sub-bands of different frequencies, and perform threshold processing on the coefficients in each sub-band. The threshold calculation formula is:

[0030]

[0031] Where: σ represents the noise standard deviation, N represents the number of sub-band coefficients. The inverse wavelet transform is performed on the processed sub-band coefficients to reconstruct the image, and the image after noise suppression is obtained.

[0032] As a further solution of the present invention: in the third step, the structure of the lightweight convolutional neural network MobileNetV3-Small network includes:

[0033] Initial convolutional layer: A 3×3 convolutional kernel is used, and the number of output channels is 16. Preliminary feature extraction is performed on the input normalized image data.

[0034] 10 inverse residual modules: Each inverse residual module includes a depthwise separable convolution, a pointwise convolution, and a linear bottleneck structure, which are used to further extract and compress features.

[0035] Squeeze-and-Excitation attention module: The channel compression ratio is 16, which performs channel weighting on the features to enhance the expression ability of important features.

[0036] Global average pooling layer, which converts the feature map into 3 feature vectors.

[0037] Fully connected layer, which outputs the quantization feature vectors of sharpness, contrast, and noise distribution.

[0038] As a further solution of the present invention: in the fourth step, the specific process of adjusting the dynamic threshold based on historical quality inspection data through online learning to update the threshold range is as follows: Establish a historical quality inspection data repository to store the image data, feature vectors, quality assessment results, and related equipment parameters and clinical diagnosis information of each quality inspection. When new quality inspection data is generated, it is compared and analyzed with the historical quality inspection data, the similarity between the feature vector of the new data and the feature vector of the historical data is calculated, and then according to the similarity and the quality assessment result, it is judged whether the threshold range needs to be adjusted. When the feature vector of the new data has a high similarity with the feature vector of the low-quality image and the quality assessment result is unqualified, the threshold of the corresponding quality index is lowered. When the feature vector of the new data has a high similarity with the feature vector of the high-quality image and the quality assessment result is qualified, the threshold of the corresponding quality index is raised. Then the threshold range is updated, and the new quality inspection data is added to the historical quality inspection data repository for subsequent threshold adjustment.

[0039] As a further solution of the present invention: In the fourth step, the training method of the hybrid model of random forest and support vector machine (SVM) is as follows: The multi-modal medical image dataset labeled as "qualified" and "unqualified" is divided into a training set, a validation set and a test set according to 60%-70%:15%-20%:15%-20%. The training set is used to train the random forest model, 100-200 decision trees are constructed, and classification is performed according to the voting results of the decision trees. The training set is used to train the support vector machine (SVM) model, the radial basis kernel function (gamma = 0.1, C = 1.0) is selected, and the optimal classification hyperplane is solved through the sequential minimal optimization algorithm. Then, the outputs of the random forest model and the support vector machine (SVM) model are fused, and the final classification result is determined by weighted voting. By adjusting the weights on the validation set, the classification accuracy of the hybrid model is optimized. Then, the test set is used to evaluate the trained hybrid model, and evaluation metrics such as the accuracy, recall rate, and F1 value of the model are calculated.

[0040] As a further solution of the present invention: In the fifth step, the specific implementation method of accelerating the feature extraction and classification process using a parallel computing framework is as follows: The computing tasks in the feature extraction and classification process are divided into multiple parallel subtasks, each subtask is executed in an independent thread, the CUDA thread block size is configured as 256, the thread grid size is 1024, thread resources are allocated, different subtasks are assigned to different thread blocks and threads, and then the data is processed in blocks. The image data and feature vector data are divided into multiple sub-blocks, and each thread processes the data of one sub-block to improve the parallelism of data processing.

[0041] As a further solution of the present invention: In the fifth step, the visualization report is presented in the following way: A heat map marker of the artifact area is superimposed on the original image, generated based on the Grad-CAM algorithm. The color depth of the heat map represents the severity of the artifact area, and the darker the color, the more severe the artifact. A PDF report including the quality score and recommended processing suggestions (requiring re-scanning, acceptable) is generated. The PDF report also includes the basic information of the image and the specific values of multi-dimensional features.

[0042] As a further solution of the present invention: the detection method further includes a multi-modal data fusion step: when the image contains multiple sequences, each sequence is respectively pre-processed, multi-dimensionally feature-extracted, and dynamically quality-evaluated according to steps two to four to obtain the quality score of each sequence. Then, according to the importance of different sequences in clinical diagnosis, corresponding weights are assigned to the quality scores of each sequence, and the quality evaluation results of different sequences are weighted and fused to obtain a comprehensive quality score. Then, in combination with the patient's basic information, including age, weight, gender, etc., the quality score is corrected, and according to the potential relationship between the patient's basic information and the image quality, the comprehensive quality score is adjusted by establishing a regression model.

[0043] As a further solution of the present invention: during the process of data acquisition, processing, and storage of the detection method, encryption technology is used to encrypt the image data, and the encryption algorithm adopts the AES-256 encryption algorithm that complies with industry standards to prevent data leakage. Identity authentication and permission management are carried out on users accessing the image data, and only authorized medical staff and technical personnel can access and process the image data.

[0044] Adopting the above technical solutions, compared with the prior art, the beneficial effects of the present invention are as follows:

[0045] 1. Through the optimized design of the parallel computing framework (CUDA) and the lightweight convolutional neural network (MobileNetV3-Small), the present invention significantly improves the processing speed of image quality detection. By using the thread block and thread grid structure of CUDA, the feature extraction and classification tasks are divided into multiple parallel sub-tasks, and each sub-task is executed in an independent thread. At the same time, the depthwise separable convolution and inverted residual structure are used to reduce the network calculation amount. Through the above technical means, the present invention greatly reduces the end-to-end processing delay and can meet the requirements of real-time quality inspection. Especially in the scenario of batch image processing, the detection efficiency is significantly improved;

[0046] 2. By combining deep learning with traditional image processing algorithms, the present invention realizes the comprehensive evaluation of multi-dimensional quality indicators. The clarity, contrast, and noise distribution characteristics of the image are extracted through a lightweight convolutional neural network, and in combination with the calculation of the structural similarity index (SSIM) and the peak signal-to-noise ratio (PSNR), the quality of the image at the structural and pixel levels is comprehensively evaluated. In addition, a hybrid model of random forest and support vector machine (SVM) is used for classification, and the outputs of the two models are fused by means of weighted voting, further improving the classification accuracy;

[0047] 3. Through the dynamic threshold adjustment mechanism, the present invention realizes adaptive detection for different device types and clinical requirements. It loads corresponding reference quality parameters according to the type of imaging device (such as CT, MRI, etc.), and dynamically weights the thresholds of different quality indicators in combination with the clinical requirements of the examination site (such as lungs, bones, etc.). In addition, based on historical quality inspection data, the threshold range is updated through online learning. When new quality inspection data is generated, the threshold range is adjusted according to the characteristics of the data and the quality assessment results, enabling it to adapt to different application scenarios and improving the flexibility and practicality of detection.

[0048] 4. The present invention is compatible with various medical imaging types such as CT, MRI, X-ray, and ultrasound, and is applicable to hospital imaging departments, telemedicine, and AI-assisted diagnosis systems. Through the multi-modal data fusion step, the quality assessment results of different sequences (such as T1 / T2 weighted images of MRI) are weighted and fused, and the comprehensive quality score is corrected in combination with the patient's basic information (such as age, weight). In addition, the present invention uses the AES-256 encryption algorithm to encrypt the imaging data, and conducts identity authentication and permission management for users accessing the imaging data to ensure data security. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a schematic flow chart of the method steps in the embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0050] The following further describes the specific embodiments of the present invention with reference to the drawings. It should be noted here that the description of these embodiments is used to help understand the present invention, but does not constitute a limitation to the present invention.

[0051] In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0052] Please refer to the attached Figure 1 , A rapid medical imaging quality detection method of the present invention, the detection method includes the following specific steps:

[0053] Step 1: Data acquisition, receiving the original DICOM format imaging data transmitted by the medical imaging device through the DICOM standard protocol. The medical imaging device includes but is not limited to CT scanners, MRI imagers, X-ray machines, and ultrasonic diagnostic devices;

[0054] Step 2: Preprocessing:

[0055] S1. Perform grayscale normalization processing on the original image, and uniformly map the grayscale value range of the image to the preset interval [0, 1] to eliminate the influence of grayscale differences when the images are collected by different devices;

[0056] S2. Apply the noise suppression algorithm based on wavelet transform to suppress the noise in the image;

[0057] S3. Detect the edge information in the image through the Canny edge detection algorithm, and then use the morphological closing operation (kernel size 5×5) to process the edge information to locate the artifact area;

[0058] S4. Automatically crop the ROI of the image. According to the type of the image and the prior knowledge of the examination part, determine the ROI area and scale it to the preset size (224×224 pixels) to reduce the data volume for subsequent processing;

[0059] Step 3. Multi-dimensional feature extraction:

[0060] S1. Extract the clarity, contrast and noise distribution features of the image through a lightweight convolutional neural network. The lightweight convolutional neural network is the MobileNetV3-Small network. Its input layer receives the normalized image data after gray normalization, noise suppression, ROI automatic cropping and standardization scaling in Step 2. The network internally performs feature extraction through depthwise separable convolution and inverted residual structures, and the output layer generates the quantization feature vectors of clarity, contrast and noise distribution;

[0061] S2. Combine the image processing algorithm to calculate the structural similarity index (SSIM) and the peak signal-to-noise ratio (PSNR) to obtain the quality features of the image at the structural and pixel levels. The algorithm for calculating the structural similarity index (SSIM) is as follows:

[0062] Let the original image be I and the image to be evaluated be J. Divide the image into N×N sub-blocks. For each sub-block i, calculate its mean μ Ii 、μ Ji , variance and covariance σ IiJi . Calculate the SSIM value of each sub-block according to the following formula:

[0063]

[0064] where: C1 = (k1L) 2 , C2 = (k2L) 2 , L is the dynamic range of the image pixel values, k1 = 0.01, k2 = 0.03. Average the SSIM values of all sub-blocks to obtain the structural similarity index (SSIM) of the entire image;

[0065] The algorithm for calculating the peak signal-to-noise ratio (PSNR) is as follows: First, calculate the mean square error (MSE) of the original image I and the image to be evaluated J. The formula is:

[0066]

[0067] Wherein: M and N respectively represent the number of rows and columns of the image, and then the peak signal-to-noise ratio (PSNR) is calculated according to the following formula:

[0068]

[0069] Where MAX1 represents the maximum possible value of the pixels in the image (for 8-bit grayscale images, MAX1 = 255);

[0070] Step Four: Dynamic Quality Assessment:

[0071] S1. Input the features extracted in Step Three into the classification model. The classification model is a hybrid model of random forest and support vector machine (SVM). Its training data includes a multi-modal medical image dataset labeled as "qualified" and "unqualified", which covers various image types (including CT tomography sequences, MRI multi-parametric imaging, X-ray plain films, and ultrasound dynamic images) and image data of different examination sites;

[0072] S2. Judge whether the image quality is qualified according to the preset dynamic threshold. The dynamic threshold is adaptively adjusted based on the image type, device parameters, and clinical diagnosis requirements. The specific adjustment method includes: loading the corresponding benchmark quality parameters according to the image device type. The benchmark quality parameters are obtained by analyzing and statistically processing a large amount of image data collected by the same type of devices, dynamically weighting the thresholds of different quality indicators (clarity, contrast, noise distribution, SSIM, and PSNR) in combination with the clinical requirements of the examination site. For example, for lung CT images, the threshold weights of clarity and noise distribution may be relatively high, and the threshold range is updated through online learning based on historical quality inspection data. When new quality inspection data is generated, the threshold range is adjusted according to the characteristics and quality assessment results of the data;

[0073] Step Five: Result Output:

[0074] S1. Generate a visual report containing quality scores, defect types, and marked abnormal regions, and overlay a heat map marker of the artifact region on the original image to intuitively display the location and severity of the artifact region. At the same time, generate a PDF report containing quality scores and recommended processing suggestions ("requiring rescan", "acceptable");

[0075] S2. Real-time feedback to the imaging device and the diagnostic system through the interface. Adopt a parallel computing framework to accelerate the feature extraction and classification process, optimize the allocation of computing resources and the thread scheduling of parallel computing, improve the computing efficiency, and deploy a lightweight model at the imaging device side to achieve an end-to-end processing delay ≤ 0.5 seconds to ensure that the results can be fed back in a timely manner.

[0076] In one embodiment of the present invention: In step two, the noise suppression algorithm based on wavelet transform is specifically as follows: Select the db4 wavelet basis function, perform 3-layer wavelet decomposition on the original image, decompose the image into sub-bands of different frequencies, perform threshold processing on the coefficients in each sub-band, and the threshold calculation formula is:

[0077]

[0078] Where: σ represents the noise standard deviation, N represents the number of sub-band coefficients, perform inverse wavelet transform on the processed sub-band coefficients, reconstruct the image, and obtain the image after noise suppression.

[0079] In one embodiment of the present invention: In step three, the structure of the lightweight convolutional neural network MobileNetV3-Small network includes:

[0080] Initial convolutional layer: Use a 3×3 convolutional kernel, with the number of output channels being 16, and perform preliminary feature extraction on the input normalized image data;

[0081] 10 inverse residual modules: Each inverse residual module contains depthwise separable convolution, pointwise convolution, and a linear bottleneck structure, which are used to further extract and compress features;

[0082] Squeeze-and-Excitation attention module: The channel compression ratio is 16, perform channel weighting on the features, and enhance the expression ability of important features;

[0083] Global average pooling layer, convert the feature map into 3 feature vectors;

[0084] Fully connected layer, output the quantization feature vectors of sharpness, contrast, and noise distribution.

[0085] In one embodiment of the present invention: In step four, the specific process of adjusting the dynamic threshold based on historical quality inspection data through online learning to update the threshold range is as follows: Establish a historical quality inspection data repository to store the image data, feature vectors, quality assessment results, and related equipment parameters and clinical diagnosis information of each quality inspection. When new quality inspection data is generated, compare and analyze it with the historical quality inspection data, calculate the similarity between the feature vector of the new data and the feature vectors of the historical data, and then judge whether it is necessary to adjust the threshold range according to the similarity and the quality assessment result. When the feature vector of the new data has a high similarity with the feature vector of the low-quality image and the quality assessment result is unqualified, lower the threshold of the corresponding quality index. When the feature vector of the new data has a high similarity with the feature vector of the high-quality image and the quality assessment result is qualified, raise the threshold of the corresponding quality index. Then update the threshold range and add the new quality inspection data to the historical quality inspection data repository for subsequent threshold adjustment.

[0086] In one embodiment of the present invention: In step four, the training method of the hybrid model of random forest and support vector machine (SVM) is as follows: The multimodal medical image dataset labeled as "qualified" and "unqualified" is divided into a training set, a validation set, and a test set according to 60%-70%:15%-20%:15%-20%. The training set is used to train the random forest model, constructing 100-200 decision trees and classifying according to the voting results of the decision trees. The training set is used to train the support vector machine (SVM) model, selecting the radial basis kernel function (gamma = 0.1, C = 1.0), and solving the optimal classification hyperplane through the sequential minimal optimization algorithm. Then, the outputs of the random forest model and the support vector machine (SVM) model are fused, and the final classification result is determined by weighted voting. By adjusting the weights on the validation set, the classification accuracy of the hybrid model is optimized. Then, the test set is used to evaluate the trained hybrid model, and evaluation metrics such as the accuracy, recall rate, and F1 value of the model are calculated.

[0087] In one embodiment of the present invention: In step five, the specific implementation manner of accelerating the feature extraction and classification process using a parallel computing framework is as follows: The computational tasks in the feature extraction and classification process are divided into multiple parallel subtasks, and each subtask is executed in an independent thread. The CUDA thread block size is configured to be 256, and the thread grid size is 1024. Thread resources are allocated, and different subtasks are assigned to different thread blocks and threads. Then, the data is processed in chunks. The image data and feature vector data are divided into multiple sub-chunks, and each thread processes the data of one sub-chunk to improve the parallelism of data processing.

[0088] In one embodiment of the present invention: In step five, the visualization report is presented in the following way: A heatmap marker of the artifact area is superimposed on the original image, generated based on the Grad-CAM algorithm. The color depth of the heatmap represents the severity of the artifact area, and the darker the color, the more severe the artifact. A PDF report containing the quality score and recommended processing suggestions (requiring rescan, acceptable) is generated. The PDF report also includes the basic information of the image and the specific values of multi-dimensional features.

[0089] In one embodiment of the present invention: The detection method further includes a multi-modal data fusion step: When the image contains multiple sequences, each sequence is preprocessed, multi-dimensionally feature-extracted, and dynamically quality-assessed according to steps two to four respectively to obtain the quality score of each sequence. Then, according to the importance of different sequences in clinical diagnosis, corresponding weights are assigned to the quality scores of each sequence, and the quality assessment results of different sequences are weighted and fused to obtain a comprehensive quality score. Then, in combination with the patient's basic information, including age, weight, gender, etc., the quality score is corrected, and according to the potential relationship between the patient's basic information and the image quality, the comprehensive quality score is adjusted by establishing a regression model.

[0090] In one embodiment of the present invention: During the data acquisition, processing, and storage of the detection method, encryption technology is used to encrypt the image data, and the encryption algorithm uses the AES-256 encryption algorithm that complies with industry standards to prevent data leakage. Identity authentication and permission management are performed on users accessing the image data, and only authorized medical staff and technical personnel can access and process the image data.

[0091] Example 1. Please refer to the appendix Figure 1 , Data acquisition:

[0092] Retrieve DICOM images (resolution 512×512 pixels, slice thickness 1mm) of 1000 chest CT scans from the hospital's PACS system, among which 800 are used for training and 200 are used for testing;

[0093] Preprocessing:

[0094] Gray normalization: Linearly map the image pixel values to the range [0,1];

[0095] Noise suppression: Perform 3-layer wavelet decomposition using the db4 wavelet basis function, and perform soft threshold processing on each sub-band coefficient. The threshold calculation formula is:

[0096]

[0097] Set the coefficients less than 33.6 to zero, and the noise level is reduced by ≥30%;

[0098] Artifact area localization: Detect the edge information in the image through the Canny edge detection algorithm, and then use morphological closing operation (kernel size 5×5) to locate the artifact area;

[0099] ROI automatic cropping: According to the prior knowledge of lung CT, crop the ROI area and scale it to 224×224 pixels;

[0100] Multi-dimensional feature extraction:

[0101] Extract the clarity, contrast, and noise distribution features using the MobileNetV3-Small network. The network input size is 224×224, and the output is a 3D feature vector (clarity 0.92, contrast 0.85, noise distribution 0.12);

[0102] Calculate the SSIM value:

[0103]

[0104] Calculate the PSNR value:

[0105]

[0106] Dynamic quality assessment:

[0107] Input the extracted features into a hybrid model of random forest and SVM, and the output quality score is 0.92;

[0108] Judge whether the image is qualified according to the dynamic threshold (such as clarity ≥ 0.85, PSNR ≥ 30dB). Since PSNR = 27dB < 30dB, it is determined to be unqualified.

[0109] Result output:

[0110] Generate a visualization report and overlay a heat map of the artifact area on the original image (red indicates severe artifacts, and green indicates mild artifacts);

[0111] Accelerate the processing through the CUDA parallel computing framework to achieve an end-to-end latency ≤ 0.5 seconds.

[0112] Example 2. Please refer to the appendix Figure 1 , Data acquisition:

[0113] Obtain 500 knee MRI images (T1 / T2 weighted images, resolution 256×256), of which 350 are used for training and 150 are used for testing;

[0114] Preprocessing:

[0115] Use the U-Net network to locate the knee joint area and crop it to 128×128 pixels;

[0116] Perform histogram equalization on each MRI sequence;

[0117] Feature extraction and fusion:

[0118] Calculate the image entropy, gradient magnitude histogram (GMH), and local binary pattern (LBP) texture features;

[0119] The image entropy is 4.5, GMH is 0.75, and LBP is 0.68;

[0120] Perform weighted fusion on the features of T1 and T2 weighted images (T1 weight = 0.6, T2 weight = 0.4);

[0121] T1 weighted image feature vector: [4.5, 0.75, 0.68]

[0122] T2 weighted image feature vector: [4.2, 0.70, 0.65]

[0123] Feature vector after weighted fusion: [4.38, 0.73, 0.67]

[0124] Quality assessment:

[0125] Use a support vector machine (SVM) for classification, and the output quality score is 0.88;

[0126] Combine the patient's age (over 60 years old) to correct the score. Let the correction coefficient be 1.1, then the final score is: 0.88×1.1 = 0.97;

[0127] Result output:

[0128] Generate a PDF report, including the quality score, recommended treatment suggestions, and basic image information.

[0129] Example 3. Please refer to the appendix Figure 1 , Multimodal data fusion:

[0130] Perform quality assessment on the T1 / T2 weighted images of MRI respectively, and perform weighted fusion scoring (T1 weight = 0.6, T2 weight = 0.4);

[0131] T1 weighted image score: 0.92, T2 weighted image score: 0.85;

[0132] Weighted fusion score: 0.92×0.6 + 0.85×0.4 = 0.89;

[0133] Combine the patient's basic information (age = 65 years old, weight = 70 kg) to correct the comprehensive score through a linear regression model. The correction coefficient is 1.05, then the final score is: 0.89×1.05 = 0.93;

[0134] Dynamic threshold adjustment:

[0135] Establish a historical quality inspection data repository to store the image data, feature vectors, and evaluation results of each quality inspection. When new data is generated, calculate its cosine similarity with the historical data as 0.95, and the score of similar images in the historical data is 0.90, then dynamically adjust the threshold to:

[0136] New threshold = 0.90×0.95 = 0.855.

[0137] According to the content of the above embodiments, through the multi-step preprocessing, multi-dimensional feature extraction, dynamic quality assessment of medical image data, and the optimized design of the real-time feedback mechanism, the efficiency and accuracy of image quality detection have been significantly improved. Through the synergistic effect of the parallel computing framework (CUDA) and the lightweight convolutional neural network (MobileNetV3-Small), the end-to-end processing delay is achieved to be ≤0.5 seconds, meeting the requirements of real-time quality inspection. By combining deep learning with traditional image processing algorithms, the clarity, contrast, noise distribution, and structural similarity of images are comprehensively evaluated, further improving the accuracy and reliability of detection. Through the dynamic threshold adjustment mechanism, adaptive detection for different device types and clinical needs is realized, significantly improving the flexibility and practicality of detection. In addition, through multi-modal data fusion and data encryption technology, the universality and data security of the system are ensured. The present invention has significant technical advantages and broad application prospects, and can be widely applied to hospital imaging departments, telemedicine, and AI-assisted diagnosis systems, providing an efficient, accurate, and safe solution for medical image quality detection.

[0138] Although the present invention is disclosed above in a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, any modification, equivalent change, and modification made to the above embodiments based on the technical essence of the present invention without departing from the technical solution of the present invention shall fall within the protection scope defined by the claims of the present invention.

Claims

1. A rapid detection method for medical image quality, characterized in that The detection method includes the following specific steps: Step 1, data acquisition: Receive the original DICOM format image data transmitted by the medical imaging device through the DICOM standard protocol. Step 2, preprocessing: S1. Perform gray normalization processing on the original image; S2. Based on the noise suppression algorithm of wavelet transform, perform noise suppression on the image; S3. Detect the edge information in the image through the Canny edge detection algorithm, and then use the morphological closing operation to process the edge information to locate the artifact area; S4. Automatically crop the ROI of the image, determine the ROI area, and scale it to the preset size; Step 3, multi-dimensional feature extraction: S1. Extract the clarity, contrast, and noise distribution features of the image through the MobileNetV3-Small lightweight convolutional neural network; S2. Combine the image processing algorithm to calculate the structural similarity index and the peak signal-to-noise ratio. The algorithm for calculating the structural similarity index is specifically: Let the original image be , and the image to be evaluated be . Divide the image into sub - blocks. For each sub - block , calculate its mean , , variance , , and covariance . Calculate the SSIM value of each sub - block according to the following formula: The structure of the lightweight convolutional neural network MobileNetV3-Small network includes: Initial convolutional layer: Use a 3×3 convolutional kernel, and the number of output channels is 16; 10 inverse residual modules: Each inverse residual module contains a depthwise separable convolution, a pointwise convolution, and a linear bottleneck structure; Squeeze-and-Excitation attention module: The channel compression ratio is 16; Global average pooling layer, convert the feature map into 3 feature vectors; Fully connected layer, output the quantization feature vectors of clarity, contrast, and noise distribution; In the formula for calculating the SSIM value of each sub-block: , , is the dynamic range of the image pixel values, = 0.01, = 0.

03. The values of all sub-blocks are averaged to obtain the structural similarity index of the entire image; The algorithm for calculating the peak signal-to-noise ratio is specifically as follows: First, calculate the original image and the image to be evaluated The mean square error of, the formula is: Wherein: and respectively represent the number of rows and columns of the image, and then calculate the peak signal-to-noise ratio according to the following formula: Among them, represents the maximum possible value of the pixels in the image; Step 4, dynamic quality assessment: S1. Input the features extracted in Step 3 into the classification model, and the classification model is a hybrid model of a random forest and a support vector machine (SVM); S2. Judge whether the image quality is qualified according to the preset dynamic threshold; Step 5, result output: S1. Generate a visual report including quality scores, defect types, and abnormal area markings; S2. Real-time feedback to the imaging device and the diagnostic system through the interface, and use a parallel computing framework to accelerate the feature extraction and classification process.

2. A rapid detection method for medical image quality according to claim 1, characterized in that: In the second step, the gray value range of the image is uniformly mapped to the preset interval [0,1]. The noise suppression algorithm based on wavelet transform is specifically: Select the db4 wavelet basis function, perform 3-layer wavelet decomposition on the original image, decompose the image into sub-bands of different frequencies, and perform threshold processing on the coefficients in each sub-band. The threshold calculation formula is: Wherein: represents the noise standard deviation, represents the number of subband coefficients. The inverse wavelet transform is performed on the processed subband coefficients to reconstruct the image, and the image after noise suppression is obtained.

3. A rapid medical image quality detection method according to claim 1, characterized in that: In the fourth step, the specific process of updating the threshold range of the dynamic threshold through online learning based on historical quality inspection data is as follows: Establish a historical quality inspection data repository to store the image data, feature vectors, quality assessment results, and related equipment parameters and clinical diagnosis information of each quality inspection. When new quality inspection data is generated, it is compared and analyzed with the historical quality inspection data, the similarity between the feature vector of the new data and the feature vector of the historical data is calculated, and then according to the similarity and the quality assessment result, it is judged whether the threshold range needs to be adjusted. When the feature vector of the new data has a high similarity with the feature vector of the low-quality image and the quality assessment result is unqualified, the threshold of the corresponding quality index is lowered. When the feature vector of the new data has a high similarity with the feature vector of the high-quality image and the quality assessment result is qualified, the threshold of the corresponding quality index is raised. Then the threshold range is updated, and the new quality inspection data is added to the historical quality inspection data repository.

4. A rapid detection method for medical image quality according to claim 1, characterized in that: In the fourth step, the training method of the hybrid model of random forest and support vector machine (SVM) is as follows: The multi-modal medical image dataset labeled as "qualified" and "unqualified" is divided into a training set, a validation set, and a test set according to 60%-70%:15%-20%:15%-20%. The training set is used to train the random forest model, 100-200 decision trees are constructed, and classification is performed according to the voting results of the decision trees. The training set is used to train the support vector machine (SVM) model, the radial basis kernel function is selected, and the optimal classification hyperplane is solved through the sequential minimal optimization algorithm. Then the outputs of the random forest model and the support vector machine (SVM) model are fused, and the final classification result is determined by weighted voting. Then the test set is used to evaluate the trained hybrid model, and the evaluation indexes of the model are calculated.

5. A rapid detection method for medical image quality according to claim 1, characterized in that: In the fifth step, the specific implementation method of accelerating the feature extraction and classification process by using a parallel computing framework is as follows: The computing tasks in the feature extraction and classification process are divided into multiple parallel subtasks, each subtask is executed in an independent thread, the CUDA thread block size is configured to be 256, the thread grid size is 1024, thread resources are allocated, different subtasks are assigned to different thread blocks and threads, and then the data is processed in chunks. The image data and feature vector data are divided into multiple sub-chunks, and each thread processes the data of one sub-chunk.

6. The rapid detection method for medical image quality according to claim 1, wherein: In the fifth step, the visualization report is presented in the following way: A heat map marker of the artifact area is superimposed on the original image, generated based on the Grad-CAM algorithm. The color depth of the heat map represents the severity of the artifact area, and the darker the color, the more severe the artifact. A PDF report containing the quality score and recommended treatment suggestions is generated. The PDF report also includes the basic information of the image and the specific numerical values of multi-dimensional features.

7. A rapid detection method for medical image quality according to claim 1, characterized in that: The detection method further includes a multi-modal data fusion step: when the image contains multiple sequences, each sequence is preprocessed, multi-dimensionally feature-extracted, and dynamically quality-assessed according to steps two to four to obtain the quality score of each sequence. Then, according to the importance of different sequences in clinical diagnosis, corresponding weights are assigned to the quality scores of each sequence, and the quality assessment results of different sequences are weighted and fused to obtain a comprehensive quality score. Then, in combination with the patient's basic information, the quality score is corrected, and according to the potential relationship between the patient's basic information and the image quality, the comprehensive quality score is adjusted by establishing a regression model.

8. A rapid detection method for medical image quality according to claim 1, characterized in that: During the processes of data acquisition, processing, and storage of the detection method, encryption technology is used to encrypt the image data. The encryption algorithm uses the AES-256 encryption algorithm that complies with industry standards, and identity authentication and permission management are performed on the users accessing the image data. Only authorized medical staff and technical personnel can access and process the image data.

Images