Knetis image measurement model system and method based on deep learning

Abstract

The invention discloses a knee arthritis image measurement model system and method based on deep learning, particularly relates to the technical field of image processing, and is used for solving the problem of low image measurement accuracy. According to the method, the image preprocessing technology and the deep learning segmentation method are integrated, high-precision automatic segmentation and analysis are performed on the knee joint image, the contrast ratio and the resolution ratio of the image are enhanced, meanwhile, the image quality is further improved through biased field correction and non-local mean value noise reduction, and the image segmentation is performed by adopting the U-Net model, so that the image quality is improved. A complex structure in the joint is effectively recognized, segmentation precision is improved, a segmentation strategy is dynamically adjusted according to measurement influence information generated in the segmentation process, the processing flow is optimized in real time through different regulation and control signals, diagnosis efficiency and accuracy are improved, and therefore the automation level and reliability of medical image processing are improved, and the medical image processing efficiency is improved. Efficient and accurate support is provided for clinic, and the accuracy of osteoarthritis image measurement is improved.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and more specifically, to a knee arthritis image measurement model system and method based on deep learning. Background Art

[0002] Knee arthritis (OA) is a common degenerative joint disease characterized by articular cartilage wear, bone hyperplasia, joint inflammation and dysfunction.

[0003] Deficiencies of existing technologies:

[0004] Traditional methods include manually measuring joint space width, osteophyte size and other indicators in X-rays or magnetic resonance imaging (MRI) images. These methods rely on the doctor's experience and subjective judgment and may have certain errors and inconsistencies. The use of specific image processing software to assist in the marking and measurement of joint structures has improved some efficiency and accuracy, but still requires manual intervention, especially for early minor lesions or structural changes. Doctors may ignore them due to lack of experience or limited vision, thus affecting the accuracy of early diagnosis and disease assessment. Summary of the invention

[0005] In order to overcome the above-mentioned defects of the prior art, the present invention provides a knee arthritis image measurement model system and method based on deep learning. By integrating image preprocessing technology and deep learning segmentation method, high-precision automatic segmentation and analysis of knee joint images are performed. At the same time, the image quality is further improved through offset field correction and non-local mean denoising. The U-Net model is used for image segmentation to identify complex structures within the joints. The measurement influence information generated during the segmentation process is analyzed to dynamically adjust the segmentation strategy and optimize the different control signals generated to solve the problems raised in the above-mentioned background technology.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] The knee arthritis imaging measurement model method based on deep learning includes the following steps:

[0008] Use nuclear magnetic resonance detection to acquire imaging data, and use a 3D balanced full-steady-state rapid field gradient echo sequence to make preliminary adjustments to the imaging data;

[0009] Perform offset field correction and non-local mean noise reduction preprocessing on the adjusted image data;

[0010] The preprocessed image data is used as knee joint images, and the U-Net model is constructed and used for knee joint image segmentation. The image segmentation process is analyzed to obtain the measurement influence information generated during the image segmentation process.

[0011] The image adaptation status is determined according to the generated measurement impact information and different signals are generated, and different strategy adjustments are performed according to the generated different signals.

[0012] In a preferred embodiment, the image data is preliminarily adjusted using a 3D balanced full-steady-state rapid field gradient echo sequence, and the specific process is as follows:

[0013] Adjust the scanning parameters according to the scanning area, including flip angle, repetition time, and echo time;

[0014] The desired body region was scanned using a 3D balanced full-steady-state fast-field gradient echo sequence, and image reconstruction was performed using fast Fourier transform.

[0015] In a preferred embodiment, before performing offset field correction, a field map is acquired by a dual echo sequence, and two images with different echo times are acquired;

[0016] The frequency shift is calculated based on the phase difference between the two images;

[0017] According to the frequency shift, the inhomogeneity value of the local magnetic field is calculated;

[0018] Applying correction, using the inhomogeneity value to adjust the phase of each voxel, and reconstructing the image using the corrected phase and the amplitude of the original image;

[0019] The images corrected using the offset field were processed using non-local means noise reduction;

[0020] Define a large search window around each pixel to find the area with the central pixel, and define a smaller neighborhood for each pixel within the search window to calculate the similarity between pixels;

[0021] For the central pixel and each pixel in the search window, calculate the similarity and weight of the corresponding neighborhood;

[0022] The pixel value of each pixel is updated according to the calculated weight.

[0023] In a preferred embodiment, a U-Net model is constructed and used for knee joint image segmentation, and the specific process is as follows:

[0024] The processed knee joint images are cropped, normalized and data enhanced;

[0025] Adjust the number of layers, convolution kernel size, and activation function of the U-Net model according to task requirements;

[0026] Combine Dice loss and cross entropy loss to adjust the edge and category of the U-Net model;

[0027] Batch normalization and Dropout were used to adjust the internal covariate shift, and the adjusted U-Net model was used for knee joint image segmentation.

[0028] In a preferred embodiment, the image segmentation process is analyzed to obtain the measurement influence information generated in the image segmentation process. The specific process is as follows:

[0029] The measurement influencing information includes structural stability information and edge accuracy information;

[0030] The structural stability information includes the structural similarity index, and the edge accuracy information includes the multi-scale edge preservation index;

[0031] The obtained structural similarity index and multi-scale edge preservation index are calculated jointly to obtain the measurement stability coefficient.

[0032] In a preferred embodiment, the image adaptation is determined according to the generated measurement impact information and different signals are generated, and different strategies are adjusted according to the generated different signals. The specific process is as follows:

[0033] comparing the measurement stability factor to the measurement evaluation threshold;

[0034] If the measurement stability coefficient is greater than or equal to the measurement evaluation threshold, a segmentation adaptation stability signal is generated and no measurement control adjustment is performed;

[0035] If the measurement stability coefficient is less than the measurement evaluation threshold, a segmentation adaptation adjustment signal is generated, indicating that adjustment or secondary optimization is required.

[0036] The knee arthritis imaging measurement model system based on deep learning is used for the above-mentioned knee arthritis imaging measurement model method based on deep learning, including:

[0037] An image preliminary adjustment module is used to obtain image data using nuclear magnetic resonance detection, and to make preliminary adjustments to the image data using a 3D balanced full-steady-state rapid field gradient echo sequence;

[0038] A calibration module, used for performing offset field correction and non-local mean noise reduction preprocessing on the adjusted image data;

[0039] The segmentation analysis module uses the preprocessed image data as the knee joint image, constructs and uses the U-Net model to perform knee joint image segmentation, analyzes the image segmentation process, and obtains the measurement influence information generated during the image segmentation process;

[0040] The strategy adjustment module is used to determine the image adaptation status according to the generated measurement impact information and generate different signals, and perform different strategy adjustments according to the generated different signals.

[0041] Technical effects and advantages of the present invention:

[0042] The present invention integrates magnetic resonance imaging with 3D balanced full-steady-state fast field gradient echo sequence, image preprocessing technology and deep learning segmentation method to perform high-precision automatic segmentation and analysis of knee joint images, uses 3D-BFFE to enhance the contrast and resolution of the image, and further improves the image quality through offset field correction and non-local mean noise reduction. The U-Net model is used for image segmentation to effectively identify complex structures within the joint and improve the segmentation accuracy. The segmentation strategy is dynamically adjusted according to the measurement influence information generated during the segmentation process, and the processing flow is optimized in real time through different control signals, which improves the diagnostic efficiency and accuracy, thereby improving the automation level and reliability of medical image processing, providing efficient and accurate diagnostic support for clinicians, and improving the accuracy of osteoarthritis image measurement. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a flow chart of the knee arthritis image measurement model method based on deep learning of the present invention.

[0044] Figure 2 This is a schematic diagram of the structure of the knee arthritis image measurement model system based on deep learning of the present invention. DETAILED DESCRIPTION

[0045] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0046] Example 1: Figure 1 As shown, a knee arthritis imaging measurement model method based on deep learning, the method comprises:

[0047] Use nuclear magnetic resonance detection to acquire imaging data, and use a 3D balanced full-steady-state rapid field gradient echo sequence to make preliminary adjustments to the imaging data;

[0048] Perform offset field correction and non-local mean noise reduction preprocessing on the adjusted image data;

[0049] The preprocessed image data is used as knee joint images, and the U-Net model is constructed and used for knee joint image segmentation. The image segmentation process is analyzed to obtain the measurement influence information generated during the image segmentation process.

[0050] The image adaptation status is determined according to the generated measurement impact information and different signals are generated, and different strategy adjustments are performed according to the generated different signals.

[0051] Acquire knee joint image data, use magnetic resonance imaging (MRI) to acquire MRI image data of the corresponding knee joint, and use 3D balanced full-steady-state fast field gradient echo sequence (3D-BFFE) to analyze and process the acquired MRI image data. The specific analysis steps are as follows:

[0052] Scanning parameters such as flip angle, repetition time (TR), echo time (TE) etc. are adjusted according to the scanning area, and the MRI scanning process is started. A 3D balanced full-steady-state fast field gradient echo sequence is used to capture the required body area. The MRI images acquired by the scan are reconstructed and adjusted using fast Fourier transform.

[0053] The MRI images obtained by scanning are acquired and preprocessed to eliminate the image distortion caused by the uneven magnetic field, that is, offset field correction is performed. The specific steps are as follows:

[0054] Before performing offset field correction, it is necessary to first obtain a field map, which represents the change in magnetic field strength over the entire scan area. The field map can be obtained by a dual echo sequence or other specialized sequences, in which two images with different echo times (TE) are usually acquired;

[0055] The frequency offset is calculated using the phase difference between the two images. The frequency offset is directly related to the field inhomogeneity and can be expressed as: Among them, Δφ is the phase difference, and ΔTE is the difference between the two echo times;

[0056] Estimate the local magnetic field inhomogeneity. According to the frequency offset, the local magnetic field inhomogeneity value can be estimated. The calculation formula is: Where γ is the gyromagnetic ratio;

[0057] Apply the correction and use the ΔB0 value calculated above to adjust the phase of each voxel. The formula is: cor =φ org -γΔB0ΔTE, where φ org is the phase of the original image;

[0058] Reconstruct the image using the corrected phase and amplitude of the original image, reduce artifacts and distortion in the image, and improve image quality;

[0059] Offset field correction can help improve the diagnostic value of MRI images, especially when accurately measuring structures such as cartilage, as the corrected images can provide more accurate information.

[0060] In order to suppress noise and artifacts in MRI images, non-local mean denoising is used to further process the images corrected by the offset field. Non-local mean (NLM) denoising is an effective image processing technique, especially suitable for suppressing noise and artifacts in MRI images. This method relies on the search and utilization of repeated textures and structures in the image. Compared with traditional local denoising methods, NLM can better preserve image details and structures. The following are the specific steps for processing using non-local mean denoising:

[0061] Define a search window. Define a large search window W around each pixel to find areas similar to the central pixel. Define a smaller neighborhood N for each pixel within the search window to calculate the similarity between pixels. Define areas with high similarity as similar areas.

[0062] For the central pixel i and each pixel j in the search window, the similarity of the corresponding neighborhood is calculated, and the weight w(i, j) is calculated accordingly. The weight calculation formula is: In the formula, v(N i ) and v(N j ) are the grayscale value vectors of the neighborhood of pixel i and j, respectively, ∥·∥ 2，a represents the weighted Euclidean distance, h is a parameter that adjusts the filter strength (usually depends on the noise level);

[0063] Perform weight normalization and weight the contribution of each pixel according to its similarity with the central pixel, that is:

[0064] Update the pixel value, that is, use the calculated weight to update the value of each pixel. The specific formula is as follows: norm (i) = w norm (I,j)I(j), where I(j) is the intensity value of the original image at pixel j, and I norm (i) is the new intensity value of the image at pixel i after denoising.

[0065] Through the above steps, non-local mean denoising can significantly improve the quality of MRI images, especially after offset field correction, further reducing noise and artifacts and retaining important anatomical and pathological information, which is important for subsequent image analysis.

[0066] The knee joint imaging data after non-local mean denoising were processed for subsequent analysis.

[0067] It should be noted that the knee joint images are obtained based on the MRI image data involving knee joint detection scanned in historical time periods as samples, and the latest MRI image data involving knee joint detection can also be collected as knee joint image data; in practice, many MRI processing software packages such as FSL (especially its PRELUDE and FUGUE tools), SPM or AFNI provide ready-made offset field correction tools, which can perform these calculations and image reconstruction more conveniently.

[0068] Use the U-Net model in deep learning to perform knee joint image segmentation. The specific steps are as follows:

[0069] The processed knee joint images are cropped, normalized (pixel values ​​are between 0 and 1), and data augmented (such as rotation, scaling, flipping, etc.) to increase the generalization ability of the model;

[0070] The U-Net model is used in the deep learning model design. U-Net is a convolutional neural network. The number of U-Net layers, convolution kernel size, activation function, etc. are adjusted according to the specific task requirements. For example, for detailed knee joint images, a deeper network or a finer convolution kernel may be required.

[0071] Specifically, the input layer is designed to accept standard MRI image sizes, such as 256x256 or 512x512 pixels, and the input size is adjusted according to the original resolution of the image and the hardware performance;

[0072] Convolutional layer configuration, for the encoder (contraction path), each encoding block contains two 3x3 convolutional layers, followed by ReLU activation function and 2x2 maximum pooling layer to reduce the feature map size. The number of channels of the convolutional layer starts from 64 and doubles in each downsampling step (64, 128, 256, 512). Doubling the number of channels can increase its learning capacity as the network goes deeper;

[0073] Decoder (extended path), i.e., each decoding block starts with a 2x2 upsampling layer to double the feature map size, followed by feature fusion using the corresponding skip connection (feature map passed from the encoder), followed by two 3x3 convolutional layers with ReLU activation, which helps the network maintain the accuracy of boundary information when reconstructing the image;

[0074] The output layer, that is, the last convolutional layer, has the same number of output channels as the number of target categories (usually 1 for binary classification problems). A 1x1 convolution is used to adapt the final output size, and then a sigmoid or softmax activation function is used to obtain the classification probability of each pixel.

[0075] Loss function selection: In order to improve the accuracy of segmentation, especially at the edges of joint areas, a combined loss function can be used, such as the combination of Dice loss and cross entropy loss. Dice loss is particularly suitable for dealing with class imbalance problems, while cross entropy loss can provide stable gradients.

[0076] Dice loss is based on the Dice coefficient, which is a statistical tool for comparing the similarity of two samples. It is particularly suitable for binary data. In medical image segmentation, the Dice coefficient can be used to calculate the similarity between the predicted segmentation and the true segmentation. The calculation expression is: Among them, Y is the true label, is the predicted label, the intersection is determined by the overlap between the prediction and the true positive sample, and the Dice loss is 1 minus the Dice coefficient, which is used in the training process: The summation is performed on all pixels together;

[0077] Cross entropy loss is a common method to measure the difference between the actual output and the predicted output. It is suitable for classification tasks and is expressed as: In the formula, M is the number of categories, Y o，c is the true label of sample o in category c, is the corresponding predicted probability;

[0078] Combining Dice loss and cross entropy, Dice loss optimizes the marginal and category imbalance problems, while cross entropy provides a stable gradient, namely: L = α*DLOSS+(1-α)*CEL, where α is a weight parameter used to adjust the relative importance of the two losses. In practical applications, α can be adjusted according to the performance on the validation set, usually between 0.5 and 0.9.

[0079] Use batch normalization and Dropout to improve the training efficiency and generalization ability of the model. Batch normalization is mainly used to solve the problem of internal covariate shift in the training process of deep learning models. It accelerates the learning process by normalizing the layer input and increases the generalization ability of the model to a certain extent:

[0080] For each mini-batch of data, calculate the mean within the batch: And the variance: Where m is the batch size, x i is a single input feature vector; the batch is normalized using the computed mean and variance: Introducing learnable parameters to restore network expressiveness: Where γ and β are learnable parameters, representing scaling and translation respectively;

[0081] Use Dropout to randomly set a portion of activation units in the network to zero during training to prevent the network from overfitting;

[0082] Choose a probability p (usually between 0.2 and 0.5), and in each training step, each unit is retained with probability p and discarded with probability 1-p;

[0083] For each activation a in the network i , use a with a i Random mask D of the same size i , where D i Each element in independently takes the value 0 or 1 (with probability p);

[0084] By applying batch normalization and Dropout in the neural network model, the performance and stability of the model on data outside the training set can be effectively improved.

[0085] Use data augmentation techniques such as random rotation, scaling, translation, and horizontal flipping to simulate various imaging conditions, improve the model's adaptability to different angles and scales, and monitor the loss on the validation set. Stop training when the loss no longer decreases significantly within a few epochs to avoid overfitting. Use a learning rate annealing strategy, such as automatically reducing the learning rate during the loss plateau period during training, to help the model adjust weights more carefully when approaching the optimal point.

[0086] K-fold cross validation was used to evaluate the performance and stability of the model, which helped ensure that the model had good generalization ability to new data. Indicators such as Dice coefficient, Jaccard index, accuracy, recall, and precision were used to comprehensively evaluate the performance of the model. In particular, accurate identification of edges is critical in osteoarthritis image segmentation.

[0087] Through the above detailed design and training strategies, the U-Net model can be effectively customized and optimized to meet the needs of knee joint image segmentation.

[0088] Analyze the image segmentation process of knee joint images, obtain the measurement influence information generated in the image segmentation process, and further analyze based on the measurement influence information to determine the image adaptation of the U-Net model in the image segmentation process. The measurement influence information includes structural stability information and edge accuracy information.

[0089] The structural stability information includes the structural similarity index, and the edge accuracy information includes the multi-scale edge preservation index;

[0090] The structural similarity index in structural stability information is an indicator to measure the visual similarity between two images. It is widely used to evaluate image quality, especially in image processing, compression and transmission. In medical image analysis, such as the segmentation task of osteoarthritis images, it evaluates the structural similarity between the segmented image and the original image. This is particularly important to ensure that key diagnostic information is not lost during the image processing process, and has an effect on the following aspects:

[0091] Image segmentation quality monitoring: The structural similarity index can be used to evaluate whether the segmentation algorithm can accurately reproduce the key structures of the original image, ensuring that key information (such as joint space, cartilage thickness, osteophyte location, etc.) is not lost or misrepresented during the segmentation process;

[0092] Enhanced reliability: High structural similarity index values ​​indicate that the segmented image is visually and structurally very close to the original image, which enhances the physician’s trust in the image segmentation and subsequent measurement results, thus supporting more accurate clinical decision making.

[0093] Overall, the structural similarity index helps ensure high quality standards for medical image analysis, which is critical for maintaining the integrity of diagnostic information and improving the accuracy of medical image processing technology. In this way, the diagnostic and therapeutic value of medical image data can be maximized, thereby improving patient outcomes.

[0094] The logic for obtaining the structural similarity index is:

[0095] Set the local window size of the image, obtain the pixel value array x of a local window in the reference image, obtain the pixel value array y of the corresponding local window in the image to be evaluated, and calculate the average brightness of the pixel value array x: The average brightness of the pixel value array x: Where N is the total number of pixel values, and the brightness value is calculated using the following expression:

[0096] Compute the contrast of an array of pixel values ​​x: Calculate the contrast of the pixel value array y: Calculate the contrast value, the calculation expression is:

[0097] The local window of the reference image is compared with the local window in the image to be evaluated to obtain the structure value. The calculation expression is: Calculate the structural similarity index, the calculation expression is:

[0098] It should be noted that the window can cover different parts of the entire image. This method takes into account the spatial correlation of the image and simulates the way the human eye perceives complexity. A window size is selected, usually 8x8, 11x11 or larger, and applied to two images. The window can be square, and a Gaussian weighted window is usually used so that the pixels in the center of the window have greater correlation than the pixels at the edge. The reference image is the original state before analysis, usually an unprocessed, ideal or expected image quality, and the image to be evaluated is an image after some processing (such as compression, enhancement, segmentation, etc.).

[0099] The multi-scale edge preservation index in edge accuracy information is an indicator for evaluating the ability of image segmentation algorithms to preserve edge features at multiple scales, taking into account the performance of segmentation algorithms at multiple levels from coarse to fine. This comprehensive evaluation helps to fully understand the ability of the algorithm to preserve key structural information of the image.

[0100] When processing high-resolution medical images, it is particularly important to consider multiple scales of the image. The multi-scale edge preservation index allows researchers to understand how the algorithm handles various features from microscopic details to macroscopic structures, ensuring that the algorithm works effectively at all levels;

[0101] During algorithm development, the multi-scale edge preservation index can help adjust the algorithm's parameter settings to maximize its performance at all key levels. For example, adjusting the size and shape of the filter or applying different levels of regularization may affect the edge preservation ability of the segmentation result;

[0102] By ensuring that image segmentation algorithms can accurately identify and preserve edges and structures at multiple scales, improving performance at various scales will help improve the accuracy of subsequent image analysis and promote a comprehensive understanding and evaluation of the comprehensive performance of segmentation algorithms.

[0103] The multi-scale edge preservation index is obtained as follows:

[0104] Get the original image and the segmented image, convert the original image and the segmented image into grayscale images, and use Gaussian pyramid downsampling to generate multi-scale images. Each level is reduced to half of the previous level. Get images of multiple scales and get a scale image set: I l ={I1,I2,……,I m}, m is a positive integer, the edge of the image is extracted at each scale, and the Sobel operator is used to calculate the edges in the horizontal X and vertical Y directions respectively. Calculate the edge amplitude of the original image and the segmented image. The calculation expression of the edge amplitude is: Similarly, the edge amplitude YSE of the original image is obtained l , the segmentation edge amplitude FSE of the segmented imagel , calculate the edge similarity between two images: Calculate the multi-scale edge preservation index, the calculation expression is:

[0105] Both the structural similarity index and the multi-scale edge preservation index are important tools for evaluating image quality and segmentation accuracy. In the image segmentation process, especially in the field of medical image analysis, by ensuring that the automatic segmentation algorithm can accurately maintain edges and structures at different scales, the manual correction and re-examination required by doctors when using these automated tools is reduced, thereby improving the efficiency of the diagnostic process.

[0106] The obtained structural similarity index and multi-scale edge preservation index are calculated jointly to obtain the measurement stability coefficient, which is expressed as follows: In the formula, C x To measure the stability coefficient, α and β are the preset proportional coefficients of the structural similarity index JGX and the multi-scale edge preservation index DCD, and both α and β are greater than 0.

[0107] It should be noted that the size of the preset proportional coefficient is a specific value obtained by quantizing each parameter. In order to facilitate subsequent comparison, the size of the coefficient depends on the amount of sample data and the preliminary setting of the corresponding preset proportional coefficient for each set of sample data by technical personnel in this field. It is not unique, as long as it does not affect the proportional relationship between the parameter and the quantized value, just like the proportional relationship between the multi-scale edge preservation index and the measurement stability coefficient.

[0108] The smaller the structural similarity index and the smaller the multi-scale edge preservation index, that is, the smaller the performance value of the measurement stability coefficient, the low measurement stability coefficient indicates that the segmentation algorithm performs poorly in restoring the key visual and structural details of the image, which means that the algorithm may fail to accurately capture important features in the image, such as edges, textures, and other structural details, which are crucial information in medical image analysis;

[0109] By comprehensively considering the structural similarity index and the multi-scale edge preservation index, measuring the stability coefficient can more comprehensively evaluate the image adaptation during the segmentation process.

[0110] When the structural similarity index and multi-scale edge preservation index are high, the high measurement stability coefficient indicates that the image segmentation algorithm can very accurately maintain the structural details and edge information of the image, which means that the algorithm can accurately reproduce important anatomical structures such as boundaries, textures and other important markers when processing images; a stable measurement coefficient also means that the results obtained from different images, different patients, and even on different devices have high consistency and repeatability. A higher measurement stability coefficient is proof of the excellent performance of the image segmentation algorithm.

[0111] According to historical data and clinical needs, one or more thresholds are set to distinguish different performance levels of the measurement stability coefficient. For example, the threshold can be set to 0.8, indicating that the measurement stability coefficient above this value represents excellent image segmentation performance, while the value below this value may require further investigation and adjustment;

[0112] Compare the generated measurement stability coefficient with the measurement evaluation threshold, generate different measurement control signals, and adjust the corresponding control strategy according to the generated measurement control signals;

[0113] After obtaining the measurement stability coefficient, the measurement stability coefficient is compared with the measurement evaluation threshold;

[0114] If the measurement stability coefficient is greater than or equal to the measurement evaluation threshold, a segmentation adaptation stability signal is generated, indicating that the current segmentation algorithm performs well and does not need to be adjusted;

[0115] If the measurement stability coefficient is less than the measurement evaluation threshold, a segmentation adaptation adjustment signal is generated, indicating that the segmentation performance is insufficient, prompting the need for adjustment or secondary optimization;

[0116] Therefore, the generation of the segmentation adaptation adjustment signal is a warning that the segmentation results need to be reviewed, areas that can be improved need to be identified, and related parameters (such as the sensitivity of edge detection, image preprocessing steps, etc.) need to be fine-tuned, or the built-in parameters of the U-Net model used for segmentation need to be adjusted and further analyzed. It may be necessary to replace the algorithm, adopt more advanced image processing technology, or increase the amount of training data for retraining to reduce the probability of unclear recognition in subsequent osteoarthritis imaging measurements;

[0117] The adjusted algorithm is re-evaluated for measurement stability to monitor the improvement effect and adjust the strategy based on the new measurement results to form a feedback loop of continuous improvement;

[0118] This process not only ensures the high quality of osteoarthritis image segmentation, but also, through continuous monitoring and adjustment, can gradually improve the performance of the image analysis algorithm, thereby achieving more accurate diagnosis and evaluation.

[0119] It should be noted that the setting of the measurement evaluation threshold can be determined according to specific scenarios and requirements, and is usually adjusted and optimized based on historical data, performance indicators and other factors.

[0120] The present invention integrates magnetic resonance imaging with 3D balanced full-steady-state fast field gradient echo sequence, image preprocessing technology and deep learning segmentation method to perform high-precision automatic segmentation and analysis of knee joint images, uses 3D-BFFE to enhance the contrast and resolution of the image, and further improves the image quality through offset field correction and non-local mean noise reduction. The U-Net model is used for image segmentation to effectively identify complex structures within the joint and improve the segmentation accuracy. The segmentation strategy is dynamically adjusted according to the measurement influence information generated during the segmentation process, and the processing flow is optimized in real time through different control signals, which improves the diagnostic efficiency and accuracy, thereby improving the automation level and reliability of medical image processing, providing efficient and accurate diagnostic support for clinicians, and improving the accuracy of osteoarthritis image measurement.

[0121] Embodiment 2, this embodiment is a system embodiment of embodiment 1, which is used to implement the knee arthritis image measurement model method based on deep learning introduced in embodiment 1, such as Figure 2 As shown, specifically including:

[0122] An image preliminary adjustment module is used to obtain image data using nuclear magnetic resonance detection, and to make preliminary adjustments to the image data using a 3D balanced full-steady-state rapid field gradient echo sequence;

[0123] A calibration module, used for performing offset field correction and non-local mean noise reduction preprocessing on the adjusted image data;

[0124] The segmentation analysis module uses the preprocessed image data as the knee joint image, constructs and uses the U-Net model to perform knee joint image segmentation, analyzes the image segmentation process, and obtains the measurement influence information generated during the image segmentation process;

[0125] The strategy adjustment module is used to determine the image adaptation status according to the generated measurement impact information and generate different signals, and perform different strategy adjustments according to the generated different signals.

[0126] The above formulas are all dimensionless and calculated numerically. Specific dimension removal can be achieved by various means such as standardization, which will not be elaborated here. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0127] The above embodiments can be implemented in whole or in part by software, hardware, firmware or any other combination. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from one website site, computer, server or data center to another website site, computer, server or data center by wired or wireless (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that contains one or more available media sets. The available medium can be a magnetic medium (e.g., a floppy disk, an ATA hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state ATA hard disk.

[0128] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0129] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0130] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0131] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, and may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0132] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0133] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A knee arthritis image measurement model method based on deep learning, characterized in that: The steps include: Use nuclear magnetic resonance detection to acquire imaging data, and use a 3D balanced full-steady-state rapid field gradient echo sequence to make preliminary adjustments to the imaging data; Perform offset field correction and non-local mean noise reduction preprocessing on the adjusted image data; The preprocessed image data is used as knee joint images, and the U-Net model is constructed and used for knee joint image segmentation. The image segmentation process is analyzed to obtain the measurement influence information generated during the image segmentation process. The image adaptation status is determined according to the generated measurement impact information and different signals are generated, and different strategy adjustments are performed according to the generated different signals.

2. The knee arthritis image measurement model method based on deep learning according to claim 1, characterized in that: The image data was initially adjusted using a 3D balanced full-steady-state fast field gradient echo sequence. The specific process is as follows: Adjust the scanning parameters according to the scanning area, including flip angle, repetition time, and echo time; The desired body region was scanned using a 3D balanced full-steady-state fast-field gradient echo sequence, and image reconstruction was performed using fast Fourier transform.

3. The knee arthritis image measurement model method based on deep learning according to claim 2 is characterized by: The adjusted image data is preprocessed with offset field correction and non-local mean noise reduction. The specific process is as follows: Before performing offset field correction, a field map was acquired by a dual-echo sequence, and two images with different echo times were collected; The frequency shift is calculated based on the phase difference between the two images; Based on the frequency shift, the inhomogeneity value of the local magnetic field can be calculated; Applying correction, using the inhomogeneity value to adjust the phase of each voxel, and reconstructing the image using the corrected phase and the amplitude of the original image; The images corrected using the offset field were processed using non-local means noise reduction; Define a large search window around each pixel to find the area with the central pixel, and define a neighborhood for each pixel within the search window to calculate the similarity between pixels; For the central pixel and each pixel in the search window, calculate the similarity and weight of the corresponding neighborhood; The pixel value of each pixel is updated according to the calculated weight.

4. The knee arthritis image measurement model method based on deep learning according to claim 3 is characterized by: Construct and use the U-Net model for knee joint image segmentation. The specific process is as follows: The processed knee joint images are cropped, normalized and data enhanced; Adjust the number of layers, convolution kernel size, and activation function of the U-Net model according to task requirements; Combine Dice loss and cross entropy loss to adjust the edge and category of the U-Net model; Batch normalization and Dropout were used to adjust the internal covariate shift, and the adjusted U-Net model was used for knee joint image segmentation.

5. The knee arthritis image measurement model method based on deep learning according to claim 4 is characterized in that: The image segmentation process is analyzed to obtain the measurement impact information generated during the image segmentation process. The specific process is as follows: The measurement influencing information includes structural stability information and edge accuracy information; The structural stability information includes the structural similarity index, and the edge accuracy information includes the multi-scale edge preservation index; The obtained structural similarity index and multi-scale edge preservation index are calculated jointly to obtain the measurement stability coefficient.

6. The knee arthritis image measurement model method based on deep learning according to claim 5, characterized in that: The image adaptation is determined based on the generated measurement impact information and different signals are generated. Different strategies are adjusted based on the generated different signals. The specific process is as follows: comparing the measurement stability factor to the measurement evaluation threshold; If the measurement stability coefficient is greater than or equal to the measurement evaluation threshold, a segmentation adaptation stability signal is generated and no measurement control adjustment is performed; If the measurement stability coefficient is less than the measurement evaluation threshold, a segmentation adaptation adjustment signal is generated, indicating that adjustment or secondary optimization is required.

7. A knee arthritis image measurement model system based on deep learning, used to implement the knee arthritis image measurement model method based on deep learning according to any one of claims 1 to 6, characterized in that: include: An image preliminary adjustment module is used to obtain image data using nuclear magnetic resonance detection, and to make preliminary adjustments to the image data using a 3D balanced full-steady-state rapid field gradient echo sequence; A calibration module, used for performing offset field correction and non-local mean noise reduction preprocessing on the adjusted image data; The segmentation analysis module uses the preprocessed image data as knee joint images, builds and uses the U-Net model to perform knee joint image segmentation, analyzes the image segmentation process, and obtains the measurement influence information generated during the image segmentation process; The strategy adjustment module is used to determine the image adaptation status according to the generated measurement impact information and generate different signals, and perform different strategy adjustments according to the generated different signals.

Images