Spinal deformation identification method and system based on image identification

Abstract

The invention relates to the technical field of spine deformation recognition, in particular to a spine deformation recognition method and system based on image recognition, and the method comprises the following steps: S1, data collection and preprocessing: obtaining multi-view and multi-time-point spine image data of the spine of a patient; s2, soft tissue shielding correction: automatically identifying and correcting interference of the soft tissue on the spine image data; s3, dynamic change prediction of the spine: performing dynamic change prediction on the spine image data subjected to soft tissue shielding correction; and S4, deformation judgment and result output: judging the deformation condition, the deformation type and the severity of the spine, and generating a visual report containing the spine deformation condition and future change prediction. According to the method, the noise in the image is effectively removed, the definition of the spine structure is enhanced, and meanwhile, the interference problem of soft tissue on the spine image is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of spinal deformation recognition, and in particular to a spinal deformation recognition method and system based on image recognition. Background Art

[0002] Spinal deformity diseases, such as scoliosis, lordosis, kyphosis, etc., are common orthopedic diseases, especially in adolescents and the elderly. Spinal deformity not only affects the patient's quality of life, but may also lead to serious respiratory, cardiovascular and other complications. In order to detect and intervene in spinal deformity at an early stage, medical imaging technologies such as X-rays, MRI and CT have been widely used in the detection and analysis of spinal morphology. Through these imaging technologies, doctors can observe the morphological changes of the spine, evaluate the condition and formulate corresponding treatment plans.

[0003] Although traditional medical imaging equipment provides a large amount of data support for spinal detection, due to the occlusion of soft tissue and the interference of image noise, the acquired spinal images are often not clear enough and it is difficult to accurately reflect the structure and subtle changes of the spine. At the same time, existing spinal analysis methods are mostly static analysis, which cannot dynamically capture the morphological change trend of the spine over time and are prone to miss early pathological signals. In addition, the existing system is inefficient in image enhancement and fine segmentation, making it difficult to provide doctors with comprehensive and accurate diagnostic information.

[0004] The present invention aims to provide a spinal deformity identification method and system based on image recognition, so as to provide doctors with more accurate and personalized diagnostic information and improve the accuracy of spinal deformity detection and the efficiency of clinical application. Summary of the invention

[0005] The present invention provides a method and system for identifying spinal deformation based on image recognition.

[0006] The method for identifying spinal deformity based on image recognition comprises the following steps:

[0007] S1, data acquisition and preprocessing: use medical imaging equipment to obtain multi-view and multi-time-point spinal image data of the patient's spine, and preprocess the acquired image data, including denoising and image enhancement;

[0008] S2, soft tissue occlusion correction: Based on the preprocessed spinal image data, the interference of soft tissue on the spinal image data is automatically identified and corrected, including:

[0009] S21, soft tissue area recognition: Based on the preprocessed spinal images, a convolutional neural network (CNN) model is used to automatically identify the soft tissue areas in the images;

[0010] S22, soft tissue segmentation and positioning: After the soft tissue area is identified, segmentation and positioning operations are performed through the image segmentation algorithm to separate the soft tissue from the spinal bones, and to clarify the location of the soft tissue and the range of the soft tissue occlusion on the spinal image;

[0011] S23, Image Correction and Bone Enhancement: Perform image correction on the blocked area and apply image enhancement technology to enhance the contour and structure of the spine bones;

[0012] S3, prediction of dynamic changes of the spine: dynamic changes of the spine image data after soft tissue occlusion correction are predicted to capture the subtle morphological changes of the spine over time and identify deformation trends in advance;

[0013] S4, deformation judgment and result output: Based on the corrected spinal image data and the predicted results of the dynamic changes of the spine, the deformation of the spine, its type and severity are judged, and a visual report containing the spinal deformation and future change predictions is generated. The report includes a 3D spinal model and deformation annotations.

[0014] Optionally, the data collection and preprocessing in S1 includes:

[0015] S11, data acquisition: Based on medical imaging equipment (such as X-ray, MRI, CT, etc.), obtain the spinal image data of the patient's spine from multiple perspectives and at multiple time points;

[0016] S12, data denoising: the collected spinal image data are denoised using the non-local means (NLM) denoising algorithm;

[0017] S13, image enhancement: The brightness and details of the spinal image are enhanced by using the adaptive histogram equalization (CLAHE) algorithm through local contrast enhancement.

[0018] Optionally, the convolutional neural network (CNN) model in S21 includes:

[0019] S211, input layer: at the input layer, receiving the pre-processed spinal image data, adjusting it to a fixed size (224×224 pixels), and performing standardization processing on the adjusted spinal image data;

[0020] S212, convolution layer: In the convolution layer, feature extraction is performed using depthwise separable convolution, including depthwise convolution and pointwise convolution. Convolution operations are performed independently on each channel, and features of multiple channels are integrated through pointwise convolution (1×1 convolution);

[0021] S213, batch normalization layer: after each convolutional layer, add a batch normalization layer;

[0022] S214, activation function: Use the Swish activation function as the activation function to enhance the nonlinear feature extraction capability of the model;

[0023] S215, pooling layer: using a hybrid pooling layer, combining maximum pooling and average pooling, to extract local features while retaining overall information;

[0024] S216, Attention Mechanism: Introduce the channel attention mechanism, generate the importance weight of each channel by calculating the global average pooling and maximum pooling results of each channel;

[0025] S217, output layer: The features extracted by the convolutional network are mapped to the output category space through the fully connected layer and the Softmax activation function. The Softmax function converts the output value into a probability distribution for predicting the classification of soft tissue and other related areas in the spinal image.

[0026] Optionally, the soft tissue segmentation and positioning in S22 includes:

[0027] S221, input layer: input the spinal image data after soft tissue recognition, including the marked soft tissue area. The input image is standardized and has a fixed size (512×512 pixels);

[0028] S222, encoder (downsampling part): The encoder using U-Net extracts high-level features of soft tissue and spine through multi-layer convolution and pooling operations;

[0029] S223, decoder (up-sampling part): the decoder gradually restores the spatial resolution of the image through up-sampling;

[0030] S224, skip connection: Use skip connection to combine high-level features of the encoder with the decoder;

[0031] S225: Segmentation output: The Softmax layer generates a probability map of each pixel belonging to soft tissue or spinal bone. The segmentation result will mark the category (soft tissue or bone) of each pixel in the image.

[0032] Optionally, the image correction and bone enhancement in S23 include:

[0033] S231, image correction: using an image correction algorithm based on gradient enhancement, performing region recognition on the input spinal image data, detecting spinal bone regions blocked by soft tissue, and enhancing the gradient information of the spinal bone regions blocked by soft tissue;

[0034] S232, bone enhancement: applying a high-frequency enhancement filter to perform image enhancement on the spinal bone region, and extracting the high-frequency part (such as the edge) of the spinal image data through the Laplacian operator.

[0035] Optionally, the prediction of dynamic changes of the spine in S3 includes:

[0036] S31, time series feature extraction: performing time series analysis on the corrected spinal image data, and capturing the time series features of the spine by extracting features from the spinal morphology data at different times;

[0037] S32, dynamic change trend prediction: Use the time series characteristics of the spine to predict the dynamic change trend and predict the future morphological change trend of the spine.

[0038] Optionally, the temporal feature extraction in S31 includes:

[0039] S311, image sequence input: for each time t, the corrected spine image I spine (t), which is input into the temporal convolutional neural network (TCN) to extract the spinal morphological features at time t;

[0040] S312, temporal feature integration: The temporal convolutional neural network (TCN) gradually extracts temporal features through multiple convolutional layers and integrates feature information at different times to capture the dynamic changes of spinal morphology.

[0041] Optionally, the dynamic change trend prediction in S32 includes:

[0042] S321, time series input and recursive processing: recursively process the historical time series features through a recursive neural network (RNN), and predict the future changes in spinal morphology by memorizing the spinal morphological features at different times;

[0043] S322, dynamic change trend prediction: The recurrent neural network (RNN) predicts the morphological changes of the spine at n moments in the future based on the hidden state and current input.

[0044] Optionally, the deformation judgment and result output in S4 includes:

[0045] S41, deformation judgment and analysis: according to the corrected spinal image data and the results of the prediction of the dynamic changes of the spine, the spinal morphology is analyzed by morphological analysis to identify the deformed area and determine the deformation type (scoliosis, lordosis, kyphosis) and its severity;

[0046] S42, generate a visualization report: Generate a visualization report of spinal deformation based on the analysis results. The report includes a 3D spinal model, deformation annotations, and future change trend predictions. The 3D model is generated through image reconstruction technology, and the deformed area is annotated with different colors and marks. The future prediction results are presented in the form of curves or morphological change diagrams.

[0047] The spinal deformity recognition system based on image recognition is used to implement the spinal deformity recognition method based on image recognition, and includes the following modules:

[0048] Data acquisition and preprocessing module: obtains multi-view and multi-time-point spinal image data of the patient's spine through medical imaging equipment, and performs denoising and image enhancement on the acquired spinal image data;

[0049] Soft tissue occlusion correction module: Based on the preprocessed spinal image data, it automatically identifies the interference of soft tissue on the spinal image and corrects the interference of the spinal image;

[0050] Spinal dynamic change prediction module: performs time series analysis on the corrected spinal image data to predict the morphological changes of the spine at future time points, captures the subtle morphological changes of the spine over time, and identifies deformation trends in advance;

[0051] Deformation judgment and result output module: Based on the corrected spinal images and the predicted results of dynamic changes of the spine, the deformation of the spine, its type and severity are judged, and a visual report containing the spinal deformation and future change predictions is generated.

[0052] Beneficial effects of the present invention:

[0053] The present invention effectively removes noise from images and enhances the clarity of the spinal structure by using a convolutional neural network model and adaptive image processing technology. At the same time, based on soft tissue occlusion correction technology, it solves the problem of soft tissue interference with spinal images. It is particularly suitable for patients with thick soft tissue or complex muscle structure, greatly improves the visibility of spinal bones, provides a more accurate image basis for subsequent dynamic change prediction and deformation analysis, and significantly improves the efficiency and accuracy of spinal deformation detection.

[0054] The present invention, by introducing a time series convolutional neural network and a recursive neural network, can perform time series analysis on the corrected spinal image data and capture the subtle morphological changes of the spine over time. The dynamic change prediction module can predict the future morphological trend of the spine, helping doctors to identify the risk of spinal deformity in advance and intervene in the early stages of the disease, thereby effectively avoiding the development of serious deformities.

[0055] The present invention can accurately determine the deformation of the spine and its type, and quantify the severity of the deformation through the corrected spinal image data and dynamic change prediction results. The generated visualization report includes a 3D spinal model, deformation annotations and future change trends, providing doctors with intuitive and detailed diagnostic information. Through this personalized report, doctors can formulate more accurate treatment plans for patients based on specific circumstances, thereby improving the effectiveness of clinical diagnosis and treatment. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0057] Figure 1 A schematic diagram of a recognition method flow chart of an embodiment of the present invention;

[0058] Figure 2 Schematic diagram of system function modules according to an embodiment of the present invention. DETAILED DESCRIPTION

[0059] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. At the same time, it is explained here that in order to make the embodiments more detailed, the following embodiments are the best and preferred embodiments, and those skilled in the art may also adopt other alternatives to implement some known technologies; and the accompanying drawings are only for more specific description of the embodiments, and are not intended to specifically limit the present invention.

[0060] It should be noted that the references to "one embodiment", "an embodiment", "an exemplary embodiment", "some embodiments" and the like in the specification indicate that the embodiments described may include specific features, structures or characteristics, but not every embodiment may include the specific features, structures or characteristics. In addition, when a specific feature, structure or characteristic is described in conjunction with an embodiment, it should be within the knowledge of a person skilled in the art to implement such feature, structure or characteristic in conjunction with other embodiments (whether or not explicitly described).

[0061] In general, a term can be understood, at least in part, from its use in context. For example, depending, at least in part, on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in the singular sense, or can be used to describe a combination of features, structures, or characteristics in the plural sense. Additionally, the term "based on" can be understood as not necessarily intended to convey an exclusive set of factors, but can instead, depending, at least in part, on the context, allow for the presence of other factors that are not necessarily explicitly described.

[0062] like Figure 1 As shown, the method for identifying spinal deformation based on image recognition includes the following steps:

[0063] S1, data acquisition and preprocessing: use medical imaging equipment to obtain multi-view and multi-time-point spinal image data of the patient's spine, and preprocess the acquired image data, including denoising and image enhancement;

[0064] S2, Soft Tissue Occlusion Correction: Based on the pre-processed spinal image data, the soft tissue interference on the spinal image data is automatically identified and corrected to improve the clarity of the spinal bones. It is especially suitable for patients with thick soft tissue or complex muscle structure. Specifically, it includes:

[0065] S21, soft tissue area recognition: Based on the preprocessed spinal images, a convolutional neural network (CNN) model is used to automatically identify the soft tissue areas in the images;

[0066] S22, soft tissue segmentation and positioning: After the soft tissue area is identified, segmentation and positioning operations are performed through the image segmentation algorithm to separate the soft tissue from the spinal bones, and to clarify the location of the soft tissue and the range of the soft tissue occlusion on the spinal image;

[0067] S23, Image Correction and Bone Enhancement: Image correction is performed on the occluded areas to reduce the occlusion effect of soft tissue on the spinal bones. Image enhancement technology is also applied to enhance the display of the outline and structure of the spinal bones, thereby improving the visibility and analysis accuracy of the bone images.

[0068] S3, prediction of dynamic changes of the spine: dynamic changes of the spine image data after soft tissue occlusion correction are predicted to capture the subtle morphological changes of the spine over time and identify deformation trends in advance;

[0069] S4, deformation judgment and result output: Based on the corrected spinal image data and the prediction results of the dynamic changes of the spine, the spinal deformation, deformation type and severity are judged, and a visual report containing the spinal deformation and future change predictions is generated. The report includes a 3D spinal model and deformation annotations, providing doctors with detailed diagnostic information and personalized treatment recommendations;

[0070] Through the above content, the detection accuracy and efficiency of spinal deformation are effectively improved. Through precise soft tissue occlusion correction and dynamic change prediction, it is possible to capture subtle morphological changes of the spine in real time. Combined with personalized deformation judgment and visual output, it provides doctors with a more comprehensive and reliable diagnostic basis, greatly improving the diagnostic effect in clinical applications and the convenience of patient management.

[0071] Data collection and preprocessing in S1 include:

[0072] S11, data acquisition: Based on medical imaging equipment (such as X-ray, MRI, CT, etc.), obtain the spinal image data of the patient's spine from multiple perspectives and at multiple time points;

[0073] S12, data denoising: The collected spinal image data is denoised using the non-local mean (NLM) denoising algorithm, which is expressed as:

[0074]

[0075] Among them, C(x) is the normalization factor, K(x,y) is the similarity weight function, Ω is the neighborhood range, and I(y) is the value of the pixel in the neighborhood;

[0076] S13, image enhancement: The adaptive histogram equalization (CLAHE) algorithm is used to enhance the brightness and details of the spine image through local contrast enhancement, which is expressed as:

[0077]

[0078] Where I(x,y) is the pixel value of the input image, min(I) and max(I) are the minimum and maximum pixel values ​​of the image, respectively, and L is the grayscale level of the image;

[0079] Through the above content, the spinal images acquired by medical imaging equipment are efficiently preprocessed to effectively remove random noise while retaining the details and clarity of the spinal structure. This not only improves the visibility and contrast of the image, but also provides a solid foundation for subsequent soft tissue identification and spinal deformation analysis, ensuring the accuracy and reliability of the diagnostic results, and improving the efficiency of clinical applications and the accuracy of patient management.

[0080] The convolutional neural network (CNN) models in S21 include:

[0081] S211, input layer: In the input layer, the preprocessed spinal image data is received, adjusted to a fixed size (224×224 pixels), and the adjusted spinal image data is standardized to ensure that the mean and variance of the input data are consistent, which is expressed as:

[0082]

[0083] Among them, I input is the normalized input image data, I original is the original preprocessed spine image data, μ is the average value of the image pixel value, and σ is the distribution of the image pixel value;

[0084] S212, convolution layer: In the convolution layer, feature extraction is performed using depthwise separable convolution, including depthwise convolution and pointwise convolution. Convolution operations are performed independently on each channel to reduce computational complexity, and features of multiple channels are integrated through pointwise convolution (1×1 convolution) to improve the expressiveness of the model, expressed as:

[0085] Depthwise Convolution:

[0086] Point-wise convolution:

[0087] in, and are the outputs of the lth layer of depthwise convolution and pointwise convolution, respectively. and are the weights of the lth layer of depth convolution and point-by-point convolution, I (l-1) is the input data of the l-1th layer, and They are the bias terms of the l-th layer depth convolution and point-by-point convolution respectively;

[0088] S213, batch normalization layer: After each convolution layer, add a batch normalization layer to improve training speed and model stability. Batch normalization reduces internal covariate shift by standardizing the output of each batch, expressed as:

[0089]

[0090] in, is the output after batch normalization, Z (l) is the convolution output of the lth layer, μ B is the mean of the batch data, indicating the average value of the output pixel values ​​of the current batch. is the variance of the batch data, indicating the dispersion of the output pixel values ​​of the current batch, ∈ is a constant used to prevent division by zero errors, and Y (l) is the batch-normalized output of the lth layer, γ is a trainable scaling parameter used to adjust the output distribution after batch normalization, and β is a trainable offset parameter used to adjust the output mean after batch normalization;

[0091] S214, activation function: Use the Swish activation function as the activation function to enhance the nonlinear feature extraction capability of the model, expressed as:

[0092] A (l) =I (l) ·σ(I (l) );

[0093] in, is the Sigmoid function used in the activation function, A (l) is the output data after activation, I (l) is the output of the lth layer of convolution;

[0094] S215, pooling layer: Use a hybrid pooling layer, combining maximum pooling and average pooling, to extract local features while retaining overall information. Maximum pooling extracts the most significant features in the image, and average pooling smoothes the feature map to reduce the risk of overfitting of the model, expressed as:

[0095]

[0096] Among them, P (l) is the output after the lth layer of pooling, α is the weight coefficient for controlling the maximum pooling and average pooling, ranging from 0 to 1, For the maximum pooling operation, take the maximum value of the pooling window in the feature map. For the average pooling operation, take the average value within the pooling window. is the activation value of the l-th layer convolution output;

[0097] S216, Attention Mechanism: The channel attention mechanism is introduced to generate the importance weight of each channel by calculating the global average pooling and maximum pooling results of each channel. This mechanism helps the model dynamically adjust the feature weights, enhance the focus on important features, and suppress irrelevant or interfering features. It is expressed as:

[0098] M c =σ(W1·avgpool(Z)+W2·maxpool(Z);

[0099] Among them, Mc is the channel attention weight, which is used to adjust the weight of the feature, σ(x) is the activation function (Sigmoid), which limits the result between 0 and 1, W1 and W2 are trainable weight matrices, which are used to adjust the result of global pooling, avgpool(Z) is the global average pooling, which takes the average value of the feature map of each channel, maxpool(Z) is the global maximum pooling, which takes the maximum value of the feature map of each channel, and Z is the input feature map;

[0100] S217, output layer: The features extracted by the convolutional network are mapped to the output category space through the fully connected layer and the Softmax activation function. The Softmax function converts the output value into a probability distribution for predicting the classification of soft tissue and other related areas in the spinal image, which is expressed as:

[0101] Y output =Softmax(W fc ·A (l) +b fc );

[0102] Among them, Y output is the final classification probability distribution, which is used to judge soft tissue recognition. Softmax is the activation function, which maps the output of the fully connected layer to the probability distribution. fc is the weight matrix of the fully connected layer, A (l) is the feature vector after pooling and activation, b fc is the bias term of the fully connected layer;

[0103] Through the above content, the depthwise separable convolution reduces the computational complexity, batch normalization improves the training speed and stability, the Swish activation function enhances the nonlinear feature extraction capability, and the hybrid pooling takes into account the retention of local and global features. At the same time, the introduced channel attention mechanism enhances the model's attention to important features and dynamically adjusts the weight distribution in the feature map, thereby improving the recognition accuracy of the spinal soft tissue area and the performance of the overall model, making the model show higher efficiency and accuracy when processing complex medical images, which is suitable for soft tissue and spinal deformation detection in clinical applications.

[0104] Soft tissue segmentation and localization in S22 include:

[0105] S221, input layer: input the spinal image data after soft tissue recognition, including the marked soft tissue area. The input image is standardized and has a fixed size (512×512 pixels);

[0106] S222, encoder (downsampling part): The encoder using U-Net extracts high-level features of soft tissue and spine through multi-layer convolution and pooling operations, expressed as:

[0107]

[0108] in, is the output feature map of the l-th layer encoder, is the convolution kernel of the coding layer, is the bias term, Z (l-1) is the previous layer of the encoder (the l-1 layer), which represents the features input to the l-th layer;

[0109] S223, decoder (up-sampling part): The decoder gradually restores the spatial resolution of the image by up-sampling, which is expressed as:

[0110]

[0111] in, is the output feature map of the l-th layer decoder, is the convolution kernel of the decoding layer, is the bias term, is the output feature map of the decoder layer l-1;

[0112] S224, skip connection: Use skip connection to combine the high-level features of the encoder with the decoder to retain more spatial information and ensure that the segmentation result can be accurate to the pixel level, expressed as:

[0113]

[0114] in, represents the concatenation of the features of the encoder and decoder at layer l, is the feature map of the decoder at the layer, which is the feature representation after partial convolution of the decoder;

[0115] S225: Segmentation output: The Softmax layer generates a probability map of each pixel belonging to soft tissue or spinal bone. The segmentation result will mark the category (soft tissue or bone) of each pixel in the image, expressed as:

[0116]

[0117] Among them, Y output (x, y) represents the classification probability of the pixel position (x, y), Z(x, y) is the output value of the last layer of the network, K is the number of categories (soft tissue and bone), and Z k (x,y) is the output value of the kth class in the last layer of the network;

[0118] Through the above content, the multi-level features of the image can be effectively extracted, and the high-resolution spatial information can be retained through jump connections during the upsampling process to ensure the accuracy of the segmentation results. Through the use of the Softmax function, the classification probability of each pixel can be accurately calculated, and the fine segmentation of soft tissue and spinal bones can be achieved. It not only has a significant improvement in segmentation accuracy, but also improves the ability to process complex image structures. It is suitable for medical image processing tasks with complex tissue structures such as spinal images.

[0119] Image correction and bone enhancement in S23 include:

[0120] S231, image correction: Use the image correction algorithm based on gradient enhancement to perform regional recognition on the input spinal image data, detect the spinal bone area blocked by soft tissue, and enhance the gradient information of the spinal bone area blocked by soft tissue to improve the edge contrast to reduce the interference of soft tissue, which is expressed as:

[0121]

[0122] Among them, I corr (x,y) is the rectified image, I(x,y) is the original image, is the gradient of the image at position (x, y), and λ is the gradient enhancement coefficient, which is used to adjust the influence of the gradient;

[0123] S232, bone enhancement: Apply high-frequency enhancement filter to enhance the image of the spinal bone area, extract the high-frequency part (such as edge) of the spinal image data through Laplacian operator, and improve the clarity of the bone contour, which is expressed as:

[0124] I enh (x,y)=I corr (x,y)+α·ΔI corr (x,y);

[0125] Among them, I ebh (x,y) is the enhanced image, ΔI corr (x, y) is the Laplace operator of the corrected image, which represents the second-order derivative (high-frequency component) of the image, and α is the enhancement coefficient, which is used to adjust the effect of bone contour enhancement;

[0126] Through the above content, the interference of soft tissue is effectively reduced, and the obscured bone structure is made more clearly visible by enhancing the gradient information of the spinal bone area. Bone enhancement further improves the display effect of bone contour by processing the high-frequency part, thereby providing doctors with more accurate image references and improving the accuracy of clinical diagnosis.

[0127] The prediction of dynamic changes of the spine in S3 includes:

[0128] S31, time series feature extraction: performing time series analysis on the corrected spinal image data, and capturing the time series features of the spine by extracting features from the spinal morphology data at different times;

[0129] S32, dynamic change trend prediction: Use the time series characteristics of the spine to predict the dynamic change trend and predict the future morphological change trend of the spine;

[0130] Through the above content, the tiny morphological changes of the spine over time can be effectively captured, providing real-time monitoring and prediction of future morphological changes. By analyzing spinal images at different time points, potential deformation trends can be identified in advance, helping doctors intervene in spinal problems at an early stage and avoid the occurrence of serious deformities. It not only improves the accuracy of diagnosis, but also provides data support for the formulation of personalized treatment plans, allowing patients to receive more accurate prevention and treatment.

[0131] The temporal feature extraction in S31 includes:

[0132] S311, image sequence input: for each time t, the corrected spine image I spine(t), which is input into the temporal convolutional neural network (TCN) to extract the spine morphological features at time t, expressed as:

[0133] F time (t) = W tcn *I spine (t)+b tcn ;

[0134] Among them, F time (t) is the spinal time series feature extracted at time t, I spine (t) is the corrected spine image at time t, W tcn is the weight of the TCN convolution kernel, b tcn is the bias term;

[0135] S312, temporal feature integration: The temporal convolutional neural network (TCN) gradually extracts temporal features through multiple convolutional layers and integrates feature information at different times to capture the dynamic changes of spinal morphology, which is expressed as:

[0136]

[0137] Among them, F combined is the integrated spinal temporal characteristics, F time (t) is the time series feature at the tth moment, and T is the total number of time steps;

[0138] Through the above content, it is possible to capture the subtle morphological changes of the spine over time. TCN processes long-time span data through one-dimensional convolution, extracts the features of each moment and integrates them into complete time series features. It not only maintains the order of the time series, but also captures long-term time series dependencies through convolution, which provides an important basis for the prediction of dynamic changes of the spine and improves the accuracy and timeliness of the prediction.

[0139] Dynamic change trend prediction in S32 includes:

[0140] S321, time series input and recursive processing: recursively process the historical time series features through the recursive neural network (RNN), and predict the future changes in spinal morphology by memorizing the spinal morphology features at different times, which is expressed as:

[0141] h t =σ(W h ·F time (t)+U h ·h t-1 +b h );

[0142] Among them, h t is the hidden state at the current time t, F time (t) is the time series feature of the current moment, ht-1 is the hidden state at the previous moment, W h is the input weight matrix, U h is the hidden layer weight matrix, b h is the bias term, σ is the ReLU activation function;

[0143] S322, dynamic change trend prediction: The recurrent neural network (RNN) predicts the morphological changes of the spine at n moments in the future based on the hidden state and current input, expressed as:

[0144] P spine (t+n)=W o ·h t +b o ;

[0145] Among them, P spine (t+n) is the predicted spinal morphology at the future time t+n, W o is the output layer weight matrix, h t is the hidden state at the current moment, b o is the bias term of the output layer;

[0146] Through the above content, the dynamic change pattern of the time series characteristics of the spine can be captured, and the spinal morphology at future moments can be predicted. The recursive structure of RNN can effectively handle the dependencies in the time series and capture the subtle characteristics of the spine changing over time. By recursively processing the time series characteristics at different moments, RNN can predict the future trend of spinal morphology changes, help doctors identify potential deformation risks in advance, and provide support for personalized treatment.

[0147] Deformation judgment and result output in S4 include:

[0148] S41, deformation judgment and analysis: Based on the corrected spinal image data and the results of the prediction of the dynamic changes of the spine, the spinal morphology is analyzed using morphological analysis to identify the deformed area and determine the deformation type (scoliosis, lordosis, kyphosis) and its severity, expressed as:

[0149] D spine (t) = f shape (I corr (t),P spine (t+n));

[0150] Among them, D spine (t) is the spinal deformation result at time t, f shape It is a morphological analysis function used to judge the deformation of the spine. corr (t) is the rectified spine image, P spine (t+n) is the prediction result of the spine at the future time t+n;

[0151] Morphological analysis function f shape Specifically include:

[0152] Curvature calculation: Spinal morphology is quantified by curvature, which is used to describe the severity of spinal curvature and is expressed as:

[0153]

[0154] Among them, κ(s) is the curvature at the arc length s, x(s) and y(s) are the horizontal and vertical coordinates of the spine in two-dimensional space, x'(s) and y'(s) are the first-order derivatives of x(s) and y(s) with respect to the arc length s, representing the rate of change of the spinal curve, that is, the tangent direction of the curve, x”(s) and y”(s) are the second-order derivatives of x(s) and y(s) with respect to the arc length s, representing the rate of change of the spinal curvature, which is used to calculate the curvature of the spine;

[0155] Angle calculation: For scoliosis deformation, the Cobb angle is used as a quantitative measurement indicator. The severity of the curvature is quantified by measuring the angle between two spinal inclination angles, expressed as:

[0156]

[0157] Among them, θ cobb is the Cobb angle, which is used to evaluate the curvature of the spine, and Δy and Δx are the coordinate differences of the two end points of the spine;

[0158] Deformation identification: By comparing the difference between the corrected spine image and the normal spine model, the deformed area is identified using a morphological matching algorithm, which is expressed as:

[0159] D spine (t)=∫|I corr (t)-I normal | 2 ds;

[0160] Among them, D spine (t) is the degree of deformation at time t, I cor r(t) is the corrected spine image, I normal It is the benchmark model of the normal spine;

[0161] S42, generating a visualization report: generating a visualization report of spinal deformation based on the analysis results, the report including a 3D spinal model, deformation annotations, and future change trend predictions, the 3D model is generated by image reconstruction technology, the deformed area is annotated with different colors and marks, and the future prediction results are presented in the form of curves or morphological change diagrams;

[0162] Through the above content, the deformation of the spine and its type can be accurately judged, and the severity of the deformation can be quantified. The generated visualization report not only includes the current deformation of the spine, but also shows the future change trend. The reconstruction and deformation annotation of the 3D model help doctors understand the patient's spinal status more intuitively and provide strong support for subsequent treatment.

[0163] like Figure 2 As shown, the spinal deformity recognition system based on image recognition is used to implement the above-mentioned spinal deformity recognition method based on image recognition, and includes the following modules:

[0164] Data acquisition and preprocessing module: obtains multi-view and multi-time-point spinal image data of the patient's spine through medical imaging equipment, and performs denoising and image enhancement on the acquired spinal image data;

[0165] Soft tissue occlusion correction module: Based on the preprocessed spinal image data, it automatically identifies the interference of soft tissue on the spinal image and corrects the interference of the spinal image;

[0166] Spinal dynamic change prediction module: performs time series analysis on the corrected spinal image data to predict the morphological changes of the spine at future time points, captures the subtle morphological changes of the spine over time, and identifies deformation trends in advance;

[0167] Deformation judgment and result output module: Based on the corrected spinal images and the predicted results of dynamic changes of the spine, the deformation of the spine, its type and severity are judged, and a visual report containing the spinal deformation and future change predictions is generated.

[0168] The present invention covers any substitution, modification, equivalent method and scheme made on the essence and scope of the present invention. In order to make the public have a thorough understanding of the present invention, specific details are described in detail in the following preferred embodiments of the present invention, but those skilled in the art can fully understand the present invention without the description of these details. In addition, in order to avoid unnecessary confusion about the essence of the present invention, well-known methods, processes, procedures, components and circuits are not described in detail.

[0169] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for identifying spinal deformity based on image recognition, characterized in that: The following steps are involved: S1, data acquisition and preprocessing: use medical imaging equipment to obtain multi-view and multi-time-point spinal image data of the patient's spine, and preprocess the acquired image data, including denoising and image enhancement; S2, soft tissue occlusion correction: Based on the preprocessed spinal image data, the interference of soft tissue on the spinal image data is automatically identified and corrected, including: S21, soft tissue area recognition: Based on the preprocessed spinal image, a convolutional neural network model is used to automatically identify the soft tissue area in the image; S22, soft tissue segmentation and positioning: After the soft tissue area is identified, segmentation and positioning operations are performed through the image segmentation algorithm to separate the soft tissue from the spinal bones, and to clarify the location of the soft tissue and the range of the soft tissue occlusion on the spinal image; S23, Image Correction and Bone Enhancement: Perform image correction on the blocked area and apply image enhancement technology to enhance the contour and structure of the spine bones; S3, prediction of dynamic changes of the spine: dynamic changes of the spine image data after soft tissue occlusion correction are predicted to capture the subtle morphological changes of the spine over time and identify deformation trends in advance; S4, deformation judgment and result output: Based on the corrected spinal image data and the predicted results of the dynamic changes of the spine, the deformation of the spine, its type and severity are judged, and a visual report containing the spinal deformation and future change predictions is generated. The report includes a 3D spinal model and deformation annotations.

2. The method for identifying spinal deformity based on image recognition according to claim 1, characterized in that: The data collection and preprocessing in S1 include: S11, data acquisition: Based on medical imaging equipment, obtain the spinal image data of the patient's spine from multiple perspectives and at multiple time points; S12, data denoising: using a non-local mean denoising algorithm to denoise the collected spinal image data; S13, Image enhancement: Adaptive histogram equalization algorithm is used to enhance the brightness and details of the spinal image through local contrast enhancement.

3. The method for identifying spinal deformity based on image recognition according to claim 1, characterized in that: The convolutional neural network model in S21 includes: S211, input layer: at the input layer, receiving the pre-processed spinal image data, adjusting it to a fixed size, and performing standardization processing on the adjusted spinal image data; S212, convolution layer: In the convolution layer, feature extraction is performed using depthwise separable convolution, including depthwise convolution and pointwise convolution. Convolution operations are performed independently on each channel, and features of multiple channels are integrated through pointwise convolution. S213, batch normalization layer: after each convolutional layer, add a batch normalization layer; S214, activation function: Use the Swish activation function as the activation function to enhance the nonlinear feature extraction capability of the model; S215, pooling layer: using a hybrid pooling layer, combining maximum pooling and average pooling, to extract local features while retaining overall information; S216, Attention Mechanism: Introduce the channel attention mechanism, generate the importance weight of each channel by calculating the global average pooling and maximum pooling results of each channel; S217, output layer: The features extracted by the convolutional network are mapped to the output category space through the fully connected layer and the Softmax activation function. The Softmax function converts the output value into a probability distribution for predicting the classification of soft tissues and related areas in spinal images.

4. The method for identifying spinal deformity based on image recognition according to claim 3, characterized in that: The soft tissue segmentation and positioning in S22 includes: S221, input layer: input spinal image data after soft tissue recognition, including the marked soft tissue area, and the input image is standardized and has a fixed size; S222, encoder: The encoder using U-Net extracts high-level features of soft tissue and spine through multi-layer convolution and pooling operations; S223, decoder: the decoder gradually restores the spatial resolution of the image by upsampling; S224, skip connection: Use skip connection to combine high-level features of the encoder with the decoder; S225: Segmentation output: The Softmax layer generates a probability map of each pixel belonging to soft tissue or spinal bone. The segmentation result will mark the category of each pixel in the image.

5. The method for identifying spinal deformity based on image recognition according to claim 4, characterized in that: The image correction and bone enhancement in S23 include: S231, image correction: using an image correction algorithm based on gradient enhancement, performing region recognition on the input spinal image data, detecting spinal bone regions blocked by soft tissue, and enhancing the gradient information of the spinal bone regions blocked by soft tissue; S232, bone enhancement: applying a high-frequency enhancement filter to perform image enhancement on the spinal bone region, and extracting the high-frequency portion of the spinal image data through the Laplace operator.

6. The method for identifying spinal deformity based on image recognition according to claim 1, characterized in that: The prediction of dynamic changes of the spine in S3 includes: S31, time series feature extraction: performing time series analysis on the corrected spinal image data, and capturing the time series features of the spine by extracting features from the spinal morphology data at different times; S32, dynamic change trend prediction: Use the time series characteristics of the spine to predict the dynamic change trend and predict the future morphological change trend of the spine.

7. The method for identifying spinal deformity based on image recognition according to claim 6, characterized in that: The temporal feature extraction in S31 includes: S311, image sequence input: for each time t, the corrected spine image I spine (t), input it into the time series convolutional neural network to extract the spine morphological features at time t; S312, temporal feature integration: The temporal convolutional neural network gradually extracts temporal features through multiple convolutional layers and integrates feature information at different times to capture the dynamic changes of spinal morphology.

8. The method for identifying spinal deformity based on image recognition according to claim 7, characterized in that: The dynamic change trend prediction in S32 includes: S321, time series input and recursive processing: recursively process the historical time series features through a recursive neural network, and predict the future changes in spinal morphology by memorizing the spinal morphological features at different times; S322, dynamic change trend prediction: The recurrent neural network predicts the morphological changes of the spine at n moments in the future based on the hidden state and current input.

9. The method for identifying spinal deformity based on image recognition according to claim 8, characterized in that: The deformation judgment and result output in S4 include: S41, deformation judgment and analysis: based on the corrected spinal image data and the results of the prediction of the dynamic changes of the spine, the spinal morphology is analyzed using morphological analysis to identify the deformed area and determine the type of deformation and its severity; S42, generate a visualization report: Generate a visualization report of spinal deformation based on the analysis results. The report includes a 3D spinal model, deformation annotations, and future change trend predictions. The 3D model is generated through image reconstruction technology, and the deformed area is annotated with different colors and marks. The future prediction results are presented in the form of curves or morphological change diagrams.

10. A system for identifying spinal deformity based on image recognition, used to implement the method for identifying spinal deformity based on image recognition as claimed in any one of claims 1 to 9, characterized in that: Includes the following modules: Data acquisition and preprocessing module: obtains multi-view and multi-time-point spinal image data of the patient's spine through medical imaging equipment, and performs denoising and image enhancement on the acquired spinal image data; Soft tissue occlusion correction module: Based on the preprocessed spinal image data, it automatically identifies the interference of soft tissue on the spinal image and corrects the interference of the spinal image; Spinal dynamic change prediction module: performs time series analysis on the corrected spinal image data to predict the morphological changes of the spine at future time points, captures the subtle morphological changes of the spine over time, and identifies deformation trends in advance; Deformation judgment and result output module: Based on the corrected spinal images and the predicted results of dynamic changes of the spine, the deformation of the spine, its type and severity are judged, and a visual report containing the spinal deformation and future change predictions is generated.

Images