Spinal canal nerve anesthesia puncture guiding method based on feature enhancement

Abstract

The invention discloses a spinal canal nerve anesthesia puncture guiding method based on feature enhancement, and the method comprises the following steps: S1, constructing an original data set, carrying out the preprocessing of the original data set, and obtaining a preprocessed data set; s2, constructing a puncture guide model; introducing the spinal canal nerve ultrasonic image into a feature enhancement module to obtain feature representations of the dura mater and the ligamentum flavum; s3, inputting the extracted feature representations of the dura mater and the ligamentum flavum into a target detection module, identifying the dura mater and the ligamentum flavum in the spinal canal nerve ultrasound image, and obtaining position coordinates of the dura mater and the ligamentum flavum in the image; and S4, deploying the trained feature enhancement module and the target detection module in a background server, and detecting the positions of the dura mater and the ligamentum flavum in the acquired spinal canal nerve ultrasound image. According to the invention, the position information of the dura mater and the ligamentum flavum in the image can be fed back to an anesthetist to assist the anesthetist to perform spinal canal nerve anesthesia puncture.

Description

Technical Field

[0001] The present invention belongs to an image processing method, and particularly relates to a spinal canal nerve anesthesia puncture guiding method based on feature enhancement. Background Art

[0002] Anesthesia is a medical means, the purpose of which is to reduce the pain of patients and ensure the smooth progress of surgery. The anesthesia method of injecting local anesthetic around the spinal canal nerve trunk in the space between the dura mater and the ligamentum flavum to produce nerve conduction block in the innervated area is called epidural anesthesia, which is one of the commonly used anesthesia methods in clinical practice. It is applicable to lower limb surgeries such as femoral head replacement, knee and hip joint replacement, and open reduction and internal fixation of lower limb fractures. Epidural anesthesia can effectively relieve the pain of patients and improve the quality of surgery. After the introduction of ultrasonic imaging equipment into clinical use, especially in the field of anesthesia puncture, anesthesiologists need to undergo long-term professional ultrasonic machine operation training to proficiently master ultrasonic-guided anesthesia puncture and accurately locate the position of the spinal canal nerve.

[0003] Therefore, there is an urgent need for an auxiliary technical means to assist anesthesiologists in quickly getting started and accurately finding the position of the spinal canal nerve during ultrasonic-guided spinal canal nerve anesthesia puncture. With the rapid development of artificial intelligence, neural networks have been applied to many object detection tasks and achieved remarkable results. The object detection task based on neural networks is to use neural networks to identify a certain object in an image, frame it with a recognition box, and return the coordinate information of the object in the image. Summary of the Invention

[0004] Aiming at the technical problem that the existing spinal canal nerve anesthesia puncture guiding method cannot accurately locate the position of the spinal canal nerve, a spinal canal nerve anesthesia puncture guiding method based on feature enhancement is provided. The present invention introduces neural network technology into the spinal canal nerve anesthesia puncture task, and proposes a spinal canal nerve anesthesia puncture guiding method based on feature enhancement, which uses neural networks to automatically identify the dura mater and ligamentum flavum in the ultrasonic image of the spinal canal nerve and return the position information.

[0005] A spinal canal nerve anesthesia puncture guiding method based on feature enhancement includes the following steps:

[0006] Step S1: Construct an original data set. The original data set includes several ultrasonic images of the spinal canal nerve. Preprocess the original data set to obtain the preprocessed data set, and divide the data set into a training set and a test set;

[0007] Step S2: Construct a puncture guidance model, which includes a feature enhancement module and an object detection module; import the spinal canal nerve ultrasound images in the dataset in Step S1 into the feature enhancement module, weight the important regions of the spinal canal nerve ultrasound images, capture the dura mater and ligamentum flavum features, and suppress the information of fat and spinal cord, so as to obtain the feature representations of the dura mater and ligamentum flavum;

[0008] Step S3: Input the extracted feature representations of the dura mater and ligamentum flavum into the object detection module to identify the dura mater and ligamentum flavum in the spinal canal nerve ultrasound image, and obtain the position coordinates of the dura mater and ligamentum flavum in the image;

[0009] Step S4: Use the training set in Step S1 to train the puncture guidance model, deploy the trained feature enhancement module and object detection module on the background server, and detect the positions of the dura mater and ligamentum flavum in the collected spinal canal nerve ultrasound images.

[0010] Further, Step S1 includes:

[0011] Step S11: Collect spinal canal nerve ultrasound image data of different patients to obtain N spinal canal nerve ultrasound images;

[0012] Step S12: Perform rectangular box annotation on the dura mater and ligamentum flavum in the spinal canal nerve ultrasound images;

[0013] Step S13: Combine the spinal canal nerve ultrasound image samples with the corresponding annotation information to obtain a set of data;

[0014] Step S14: Repeat Steps S12 and S13 to obtain a complete dataset, and divide it into a training set and a test set.

[0015] Further, the feature enhancement module in Step S2 includes a multi-scale convolutional layer, a feature pyramid layer, and a feature weighted fusion layer;

[0016] The operation process of the feature enhancement module is specifically as follows:

[0017] Step S21: Input the spinal canal nerve ultrasound image into the multi-scale convolutional layer, and obtain the features of the dura mater and ligamentum flavum at different scales through convolutional kernels of different scales, which is expressed as:

[0018] {C1, C2, C3} = δ(Mconv(F));

[0019] where C1, C2, and C3 are the features of the dura mater and ligamentum flavum at different scales output by the multi-scale convolutional layer, which are the first image feature, the second image feature, and the third image feature in sequence; Mconv(·) represents the multi-scale convolutional kernel, δ represents the Sigmoid activation function, and F is the input spinal canal nerve ultrasound image;

[0020] Step S22: The feature pyramid layer enhances the learning ability of features of the dura mater and ligamentum flavum at different scales, further extracts features at different levels, and obtains multi-scale feature representations of the dura mater and ligamentum flavum;

[0021] Among them, the feature pyramid layer includes a P3 layer, a P2 layer, and a P1 layer connected in sequence;

[0022] Specifically, step S22 is as follows:

[0023] Step S221: Obtain the third image feature map by upsampling the third image feature, expressed as:

[0024] M3 = C3;

[0025] Among them, M3 represents the third image feature map;

[0026] Then, apply a convolution with a kernel size of 1×1 to the third image feature map to adjust the number of channels, expressed as:

[0027] P3 = Conv1(M3);

[0028] Among them, P3 represents the output of the P3 layer of the feature pyramid layer;

[0029] Step S222: Obtain the second image feature map by upsampling the second image feature, expressed as:

[0030] M2 = C2;

[0031] Among them, M2 represents the second image feature map;

[0032] Perform an upsampling pooling operation on the output of the P3 layer and splice it with the second image feature map to obtain an enhanced second image feature map, expressed as:

[0033] M2A = M2 + Uppool(P3);

[0034] Among them, Uppool(·) represents the upsampling pooling operation, and M2A represents the enhanced second image feature map;

[0035] Apply a convolution with a kernel size of 1×1 to the enhanced second image feature map to adjust the number of channels, expressed as:

[0036] P2 = Conv1(M2A);

[0037] Among them, P2 represents the output of the P2 layer of the feature pyramid layer;

[0038] Step S223: Perform an upsampling pooling operation on the output of the P2 layer and splice it with the second image feature map to obtain an enhanced first image feature map, expressed as:

[0039] M1A = M1 + Uppool(P2);

[0040] Among them, M1A represents the enhanced first image feature map;

[0041] Apply a convolution with a kernel size of 1×1 to the enhanced first image feature map for channel number adjustment, which is expressed as:

[0042] P1 = Conv1(M1A);

[0043] Among them, P1 represents the output of the P1 layer of the feature pyramid layer;

[0044] Finally, the output of the feature pyramid is P = {P1, P2, P3};

[0045] Step S23: Import P into the feature weighted fusion layer to learn the correlation between different features in the spinal canal nerve ultrasound image, highlight the features of the dura mater and ligamentum flavum, and suppress the features of fat and spinal cord;

[0046] Step S23 is specifically:

[0047] Input P1, P2, and P3 in P into the residual attention layer of the feature weighted fusion layer in sequence for channel compression. The first convolutional layer and the second convolutional layer of the residual attention layer generate compressed first multi-scale feature map and second multi-scale feature map through feature compression, which is expressed as:

[0048] A i = Compress1(P i );

[0049] B i = Compress2(P i );

[0050] i ∈ {1, 2, 3};

[0051] Among them, A i represents the first multi-scale feature map, B i represents the second multi-scale feature map, Compress1 represents the feature compression operation of the first convolutional layer, and Compress2 represents the feature compression operation of the second convolutional layer;

[0052] Use A i and B i to calculate the similarity between positions and generate an attention map, which is expressed as:

[0053]

[0054] Among them, Att iAttention map is denoted as, and Softmax(·) represents the similarity calculation function. Denotes the transpose of the first multi-scale feature map;

[0055] Then perform element-wise multiplication on P i and the attention map to obtain the attention-weighted feature map P i * :

[0056] P i * = P i *Att i ;

[0057] Finally, add P i and P i * to obtain the output of the residual attention layer, denoted as:

[0058] O i = γ * P i * + P i ;

[0059] where γ represents the learnable scaling factor used to scale the effect of attention enhancement, and O i is the output of the residual attention layer; After inputting P1, P1, and P3 into the residual attention layer respectively, the residual features O1, O2, and O3 of P1, P1, and P3 are obtained respectively;

[0060] Step S24: Input O1, O2, and O3 into the feature fusion layer to obtain the fusion feature, specifically as follows:

[0061] First, input O1, O2, and O3 into the self-attention layer to obtain the attention weights θ1, θ2, and θ3 respectively; Then, multiply and add the corresponding positions of O1 and O2, O2 and O3, and O1 and O3 to obtain the self-attention feature. The specific formula is as follows:

[0062] O sum = θ1 * θ2 * O1 * O2 + θ2 * θ3 * O2 * O3 + θ1 * θ3 * O1 * O3;

[0063] where O sum represents the self-attention feature, and input it into the max-pooling layer to obtain the pooled feature Denoted as:

[0064]

[0065] where Maxpool(·) represents the max-pooling operation;

[0066] Finally, The input normalization layer performs a normalization operation to obtain the final fused feature ZO, which is also the feature representation of the dura mater and the ligamentum flavum, expressed as:

[0067]

[0068] Among them, NormLayer(·) represents the L1 norm normalization operation.

[0069] Furthermore, the object detection module includes a region proposal network, a prior box pre-selection layer, and a prediction output layer; step S3 is specifically as follows:

[0070] Step S31: Generate candidate target regions through the region proposal network;

[0071] Input the fused feature ZO obtained in step S2 into the sliding window layer, generate multiple prior boxes with different scales and ratios at each position of the fused feature ZO, and obtain a set of prior boxes RL, a total of M;

[0072] Predict the target label and the predicted target position through the first score prediction fully connected layer and the first regression prediction fully connected layer respectively, expressed as:

[0073] P fore ,P back =Softmax(W cls )*A f +b cls ;

[0074] tx, ty, tw, th=W reg *A f +b reg ;

[0075] Among them, w cls , b cls are the weights and biases of the score prediction fully connected layer, W reg , b reg are the weights and biases of the regression prediction fully connected layer, A f represents the feature at the corresponding position of this prior box on the fused feature ZO, tx, ty, tw, th represent 4 parameters of the predicted target position, indicating the displacement required for the current prior box to be fine-tuned to the desired bounding box position, P fore represents the probability distribution score of the foreground category, that is, the predicted target label, P back represents the probability distribution score of the background category;

[0076] Step S32: Input all prior boxes into the prior box pre-screening layer in sequence to obtain S high-quality prior boxes; first, obtain the foreground probability distribution scores and predicted bounding box regression values of all prior boxes, then sort all prior boxes in descending order according to the foreground probability distribution scores, and start traversing the prior box set RL from the prior box with the highest score;

[0077] The specific traversal operation is as follows:

[0078] Add the prior box with the highest score in the prior box set RL to the output result T;

[0079] Calculate the intersection over union (IoU) between this prior box and other prior boxes; remove other prior boxes with IoU greater than the preset threshold from RL;

[0080] Repeat the operation until all prior boxes are traversed, and return the final output result T;

[0081] Step S33: Input the final output result T into the prediction output layer to obtain the positions of the dura mater and ligamentum flavum in the spinal canal nerve ultrasound image. The specific process is as follows:

[0082] Step S331: Crop out the corresponding regional feature blocks on ZO according to the coordinates of each prior box in the high-quality prior box set obtained in S32, and obtain the feature region block set ZO * ;

[0083] Step S332: Input the feature blocks in ZO * into the average pooling layer respectively, and then perform feature splicing to obtain the average pooling feature map SO;

[0084] Step S333: Input the average pooling feature map SO into the dilated convolutional layer to obtain the high-dimensional feature SO * ;

[0085] Step S34: Object classification and bounding box regression;

[0086] The specific content of Step S34 is: Input the high-dimensional feature SO * into two parallel fully-connected layer networks, namely the second score prediction fully-connected layer and the second regression prediction fully-connected layer. The second score prediction fully-connected layer is expressed as:

[0087] Cls = Softmax(W′ cls * SO * + b′ cls );

[0088] Among them, W′ cls and b′ cls are the weight matrix and bias vector of the second score prediction fully-connected layer respectively, Softmax(·) is the activation function, and Cls represents the high-dimensional feature SO *The probability of the dura mater and the ligamentum flavum;

[0089] The second regression prediction fully-connected layer is expressed as:

[0090] Reg = W′ reg *SO * +b′ reg ;

[0091] W′ reg and b′ reg are respectively the weight matrix and the bias vector of the second regression prediction fully-connected layer. Reg represents an 8-dimensional vector, decoding the coordinates of the predicted bounding boxes of the dura mater and the ligamentum flavum, and the coordinates are the positions of the dura mater and the ligamentum flavum in the spinal canal nerve ultrasound image.

[0092] Further, deploying the trained feature enhancement module and the target detection module in the background server to detect the positions of the dura mater and the ligamentum flavum in the collected spinal canal nerve ultrasound image specifically includes:

[0093] Step S41: Collect the spinal canal nerve ultrasound image of the patient and transmit it to the background server;

[0094] Step S42: Use the trained feature enhancement module to obtain the feature representations of the dura mater and the ligamentum flavum in the spinal canal nerve ultrasound image;

[0095] Step S43: Input the feature representations of the dura mater and the ligamentum flavum into the target detection module to obtain the positions of the dura mater and the ligamentum flavum in the spinal canal nerve ultrasound image;

[0096] Step S44: Make a nerve anesthesia puncture decision according to the positions of the patient's dura mater and ligamentum flavum in the spinal canal nerve ultrasound image.

[0097] Advantages of the present invention:

[0098] 1) The feature enhancement module integrated with multiple modules designed by the present invention using multi-scale convolution and feature pyramid has stronger adaptability and can process complex image structures. The enhanced feature map represents richer detail features of the dura mater and the ligamentum flavum, making it suitable for spinal canal nerve ultrasound images.

[0099] 2) Further, the present invention uses the constructed puncture guidance model to learn the features of the spinal canal nerve ultrasound image, realizes the recognition of the dura mater and the ligamentum flavum in the spinal canal nerve ultrasound image, can feedback the position information of the dura mater and the ligamentum flavum in the image to the anesthesiologist, assist the doctor in performing spinal canal nerve anesthesia puncture, and improve the efficiency of the anesthesiologist in finding the spinal canal nerve in the spinal canal nerve ultrasound image and the accuracy of anesthesia puncture. Brief Description of the Drawings

[0100] Figure 1The flowchart of the steps of the puncture guidance method constructed for the present invention;

[0101] Figure 2 The system structure diagram of the feature enhancement module of the present invention;

[0102] Figure 3 The system structure diagram of the residual attention layer of the present invention;

[0103] Figure 4 The system structure diagram of the target detection module of the present invention. Detailed implementation manners

[0104] The following uses specific specific examples to illustrate the implementation manners of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific implementation manners. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention.

[0105] Refer to Figures 1 to 4 , A spinal canal nerve anesthesia puncture guidance method based on feature enhancement. In one example, it includes the following steps:

[0106] Step S1: Construct an original data set. The original data set includes several spinal canal nerve ultrasound images. Preprocess the original data set to obtain a preprocessed data set, and divide the data set into a training set and a test set;

[0107] Step S2: Construct a puncture guidance model. The model includes a feature enhancement module and a target detection module; Import the spinal canal nerve ultrasound images in the data set in step S1 into the feature enhancement module, weight the important regions of the spinal canal nerve ultrasound images, capture the dura mater and ligamentum flavum features and suppress the information of fat and spinal cord, so as to obtain the feature representations of the dura mater and ligamentum flavum;

[0108] Step S3: Input the extracted feature representations of the dura mater and ligamentum flavum into the target detection module, identify the dura mater and ligamentum flavum in the spinal canal nerve ultrasound image, and obtain the position coordinates of the dura mater and ligamentum flavum in the image;

[0109] Step S4: Use the training set in step S1 to train the puncture guidance model, deploy the trained feature enhancement module and target detection module on the background server, and detect the positions of the dura mater and ligamentum flavum in the collected spinal canal nerve ultrasound images.

[0110] Further, step S1 includes:

[0111] Step S11: Collect spinal canal nerve ultrasound image data of different patients to obtain N spinal canal nerve ultrasound images;

[0112] Step S12: Perform rectangular box annotation on the dura mater and ligamentum flavum in the spinal canal nerve ultrasound image;

[0113] Step S13: Combine the spinal canal nerve ultrasound image sample with the corresponding annotation information to obtain a set of data;

[0114] Step S14: Repeat Steps S12 and S13 to obtain a complete data set, and divide it into a training set and a test set.

[0115] Further, in an example, the feature enhancement module in Step S2 includes a multi-scale convolutional layer, a feature pyramid layer, and a feature weighted fusion layer;

[0116] The operation process of the feature enhancement module is specifically as follows:

[0117] Step S21: Input the spinal canal nerve ultrasound image into the multi-scale convolutional layer, and obtain the features of the dura mater and ligamentum flavum at different scales through convolutional kernels of different scales, expressed as:

[0118] {C1, C2, C3} = δ(Mconv(F));

[0119] where C1, C2, and C3 are the features of the dura mater and ligamentum flavum at different scales output by the multi-scale convolutional layer, which are the first image feature, the second image feature, and the third image feature in sequence; Mconv(·) represents the multi-scale convolutional kernel, δ represents the Sigmoid activation function, and F is the input spinal canal nerve ultrasound image;

[0120] Step S22: The feature pyramid layer enhances the learning ability of the features of the dura mater and ligamentum flavum at different scales, further extracts features at different levels, and obtains the multi-scale feature representation of the dura mater and ligamentum flavum;

[0121] where the feature pyramid layer includes a P3 layer, a P2 layer, and a P1 layer connected in sequence;

[0122] Step S22 is specifically as follows:

[0123] Step S221: Obtain the third image feature map by upsampling the third image feature, expressed as:

[0124] M3 = C3;

[0125] where M3 represents the third image feature map;

[0126] Then apply a convolution with a convolution kernel size of 1×1 to the third image feature map to adjust the number of channels, expressed as:

[0127] P3 = Conv1(M3);

[0128] Among them, P3 represents the output of the P3 layer of the feature pyramid layer;

[0129] Step S222: Obtain the second image feature map by upsampling the second image feature, expressed as:

[0130] M2 = C2;

[0131] Among them, M2 represents the second image feature map;

[0132] Perform an upsampling pooling operation on the output of the P3 layer and splice it with the second image feature map to obtain an enhanced second image feature map, expressed as:

[0133] M2A = M2 + Uppool(P3);

[0134] Among them, Uppool(·) represents the upsampling pooling operation, and M2A represents the enhanced second image feature map;

[0135] Apply a convolution with a kernel size of 1×1 to the enhanced second image feature map to adjust the number of channels, expressed as:

[0136] P2 = Conv1(M2A);

[0137] Among them, P2 represents the output of the P2 layer of the feature pyramid layer;

[0138] Step S223: Perform an upsampling pooling operation on the output of the P2 layer and splice it with the second image feature map to obtain an enhanced first image feature map, expressed as:

[0139] M1A = M1 + Uppool(P2);

[0140] Among them, M1A represents the enhanced first image feature map;

[0141] Apply a convolution with a kernel size of 1×1 to the enhanced first image feature map to adjust the number of channels, expressed as:

[0142] P1 = Conv1(M1A);

[0143] Among them, P1 represents the output of the P1 layer of the feature pyramid layer;

[0144] Finally, the output of the feature pyramid obtained is P = {P1, P2, P3};

[0145] Step S23: Import P into the feature weighted fusion layer to learn the correlation between different features in the spinal canal nerve ultrasound image, highlight the features of the dura mater and ligamentum flavum, and suppress the features of fat and spinal cord;

[0146] Specifically, step S23 is as follows:

[0147] The P1, P2, and P3 in P are successively input into the residual attention layer of the feature weighted fusion layer for channel compression. The first convolutional layer and the second convolutional layer of the residual attention layer respectively generate the compressed first multi-scale feature map and the second multi-scale feature map through feature compression, which are expressed as:

[0148] A i = Compress1(P i );

[0149] B i = Compress2(P i );

[0150] i ∈ {1, 2, 3};

[0151] Among them, A i represents the first multi-scale feature map, B i represents the second multi-scale feature map, Compress1 represents the feature compression operation of the first convolutional layer, and Compress2 represents the feature compression operation of the second convolutional layer;

[0152] Using A i and B i calculate the similarity between positions to generate an attention map, which is expressed as:

[0153]

[0154] Among them, Att i represents the attention map, Softmax(·) represents the similarity calculation function, represents the transpose of the first multi-scale feature map;

[0155] Then perform element-wise multiplication on P i and the attention map to obtain the attention-weighted feature map P i * :

[0156] P i * = P i * Att i ;

[0157] Finally, add P i and P i * to obtain the output of the residual attention layer, which is expressed as:

[0158] O i = γ * P i * + P i ;

[0159] Among them, γ represents a learnable scaling factor used to scale the effect of attention enhancement, and O I is the output of the residual attention layer; after inputting P1, P1, and P3 into the residual attention layer respectively, the residual features O1, O2, and O3 of P1, P1, and P3 are obtained respectively;

[0160] Step S24: Input O1, O2, and O3 into the feature fusion layer to obtain the fused feature, specifically as follows:

[0161] First, input O1, O2, and O3 into the self-attention layer to obtain the attention weights θ1, θ2, and θ3 respectively; then multiply and add the corresponding positions of O1 and O2, O2 and O3, and O1 and O3 to obtain the self-attention feature. The specific formula is as follows:

[0162] O sum = θ1 * θ2 * O1 * O2 + θ1 * θ3 * O2 * O3 + θ1 * θ3 * O1 * O3;

[0163] Among them, O sum represents the self-attention feature, and input it into the max-pooling layer to obtain the pooled feature which is expressed as:

[0164]

[0165] Among them, Maxpool(·) represents the max-pooling operation;

[0166] Finally, input

[0167]

[0168] into the normalization layer for normalization operation to obtain the final fused feature ZO, which is also the feature representation of the dura mater and the ligamentum flavum, and is expressed as:

[0169] Furthermore, in one example, the object detection module includes a region generation network, a prior box pre-screening layer, and a prediction output layer; step S3 is specifically as follows:

[0170] Step S31: Generate candidate target regions through the region generation network;

[0171] Input the fused feature ZO obtained in step S2 into the sliding window layer, and generate multiple prior boxes with different scales and ratios at each position of the fused feature ZO to obtain a set of prior boxes RL, with a total of M;

[0172] Predict the target label and the predicted target position through the first score prediction fully connected layer and the first regression prediction fully connected layer respectively, and it is expressed as:

[0173] P fore , P back = Softmax(W cls ) * A f + b cls ;

[0174] tx, ty, tw, th = W reg * A f + b reg ;

[0175] Among them, W cls , b cls are the weights and biases of the fully connected layer for score prediction, W reg , b reg are the weights and biases of the fully connected layer for regression prediction, A f represents the feature at the corresponding position of this prior box on the fused feature ZO, tx, ty, tw, th represent 4 parameters of the predicted target position, indicating the displacement required for the current prior box to be fine-tuned to the desired bounding box position, P fore represents the probability distribution scores of the foreground classes, that is, the predicted target labels, P back represents the probability distribution scores of the background classes;

[0176] Step S32: Input all prior boxes into the prior box pre-screening layer in sequence to obtain S high-quality prior boxes; first obtain the foreground probability distribution scores and the predicted bounding box regression values of all prior boxes, then sort all prior boxes in descending order according to the foreground probability distribution scores, and start traversing the prior box set RL from the prior box with the highest score;

[0177] The traversal operation is specifically as follows:

[0178] Add the prior box with the highest score in the prior box set RL to the output result T;

[0179] Calculate the intersection over union (IoU) of this prior box with other prior boxes; remove other prior boxes with IoU greater than the preset threshold from RL;

[0180] Repeat the operation until all prior boxes are traversed, and return the final output result T;

[0181] Step S33: Input the final output result T into the prediction output layer to obtain the positions of the dura mater and ligamentum flavum in the spinal canal nerve ultrasound image. The specific process is as follows:

[0182] Step S331: Crop out the corresponding regional feature blocks on ZO according to the coordinates of each prior box in the high-quality prior box set obtained in S32 to obtain the feature region block set ZO * ;

[0183] Step S332: Input ZO* The feature blocks in

[0184] Step S333: Input the average pooling feature map SO into the dilated convolutional layer to obtain the high-dimensional feature SO * ;

[0185] Step S34: Object classification and bounding box regression;

[0186] Specifically, Step S34 is as follows: Input the high-dimensional feature SO * into two parallel fully-connected layer networks, namely the second score prediction fully-connected layer and the second regression prediction fully-connected layer. The second score prediction fully-connected layer is expressed as:

[0187] Cls = Softmax(W′ cls *SO * + b′ cls );

[0188] where W′ cls and b′ cls are the weight matrix and bias vector of the second score prediction fully-connected layer respectively, Softmax(·) is the activation function, and Cls represents the probability that the high-dimensional feature SO * is the dura mater and ligamentum flavum;

[0189] The second regression prediction fully-connected layer is expressed as:

[0190] Reg = W′ reg *SO * + b′ reg ;

[0191] W′ reg and b′ reg are the weight matrix and bias vector of the second regression prediction fully-connected layer respectively. Reg represents an 8-dimensional vector, decoding the coordinates of the predicted rectangular boxes of the dura mater and ligamentum flavum, and the coordinates are the positions of the dura mater and ligamentum flavum in the spinal canal nerve ultrasound image.

[0192] Furthermore, in an example, in Step S4, deploying the trained feature enhancement module and object detection module to the background server to detect the positions of the dura mater and ligamentum flavum in the collected spinal canal nerve ultrasound images is specifically as follows:

[0193] Step S41: Collect the spinal canal nerve ultrasound image of the patient and transmit it to the background server;

[0194] Step S42: Use the trained feature enhancement module to obtain the feature representations of the dura mater and ligamentum flavum in the spinal canal nerve ultrasound image;

[0195] Step S43: Input the feature representations of the dura mater and ligamentum flavum into the target detection module to obtain the positions of the dura mater and ligamentum flavum in the spinal canal nerve ultrasound image;

[0196] Step S44: Make a nerve anesthesia puncture decision based on the positions of the patient's dura mater and ligamentum flavum in the spinal canal nerve ultrasound image.

[0197] In an example, a comparison of the results with the deep convolutional neural network, VGG19, and YOLOV8 in the prior art is provided, and the evaluation metrics are mAP50 and mAP50:95.

[0198] The results are shown in Table 1. The model proposed by the present invention achieves the maximum values in both mAP50 and mAP50:95, indicating that the present model has high accuracy. The method of the present invention is superior to the deep convolutional neural network, which shows that the feature enhancement module and target detection module proposed by the present invention are effective in the identification of the dura mater and ligamentum flavum in the spinal canal nerve ultrasound image, and thus can assist doctors in judging the position of the spinal canal nerve.

[0199] Table 1 Quantitative analysis of the test dataset on different models

[0200] model mAP50 mAP50:95 deep convolutional neural network 52.4 56.1 VGG19 53.8 57.0 YOLOV7 57.4 61.5 puncture guidance model 62.2 66.4

[0201] The present invention has been described in detail with reference to the embodiments accompanied by the drawings. Those of ordinary skill in the art can make various variations of the present invention according to the above description. Therefore, some details in the embodiments should not constitute a limitation to the present invention, and the present invention will take the scope defined by the appended claims as the protection scope.

Claims

1. A method for guiding spinal nerve anesthesia puncture based on feature enhancement, characterized in that: The following steps are involved: Step S1: constructing an original data set, the original data set includes a number of spinal canal nerve ultrasound images, preprocessing the original data set, obtaining a preprocessed data set, and dividing the data set into a training set and a test set; Step S2: constructing a puncture guidance model, which includes a feature enhancement module and a target detection module; importing the spinal nerve ultrasound image of the data set in step S1 into the feature enhancement module, weighting the important areas of the spinal nerve ultrasound image, capturing the features of the dura mater and the yellow ligament and suppressing the information of fat and spinal cord, thereby obtaining the feature representation of the dura mater and the yellow ligament; Step S3: inputting the extracted feature representations of the dura mater and the yellow ligament into the target detection module, identifying the dura mater and the yellow ligament in the spinal canal nerve ultrasound image, and obtaining the position coordinates of the dura mater and the yellow ligament in the image; Step S4: Use the training set of step S1 to train the puncture guidance model, deploy the trained feature enhancement module and target detection module on the background server, and detect the positions of the dura mater and the yellow ligament in the collected spinal canal nerve ultrasound image.

2. The feature-enhanced spinal nerve anesthesia puncture guidance method according to claim 1, characterized in that: Step S1 includes: Step S11: collecting spinal canal nerve ultrasound image data of different patients to obtain N spinal canal nerve ultrasound images; Step S12: marking the dura mater and ligamentum flavum in the spinal canal nerve ultrasound image with rectangular frames; Step S13: merging the spinal canal nerve ultrasound image sample and the corresponding annotation information to obtain a set of data; Step S14: Repeat steps S12 and S13 to obtain a complete data set and divide it into a training set and a test set.

3. The feature-enhanced spinal nerve anesthesia puncture guidance method according to claim 2, characterized in that: The feature enhancement module in step S2 includes a multi-scale convolution layer, a feature pyramid layer, and a feature weighted fusion layer; The specific operation process of the feature enhancement module is as follows: Step S21: inputting the spinal nerve ultrasound image into the multi-scale convolution layer, and obtaining the features of the dura mater and ligamentum flavum at different scales through convolution kernels of different scales; Step S22: The feature pyramid layer enhances the learning ability of the features of the dura mater and the ligamentum flavum at different scales, further extracts features at different levels, and obtains multi-scale feature representations of the dura mater and the ligamentum flavum; Step S23: importing the multi-scale feature representation into the feature weighted fusion layer, learning the correlation between different features in the spinal nerve ultrasound image, highlighting the features of the dura mater and the yellow ligament, suppressing the features of fat and spinal cord, and thus obtaining residual features; Step S24: Input the residual feature into the feature fusion layer to obtain the fusion feature.

4. The feature-enhanced spinal nerve anesthesia puncture guidance method according to claim 3, characterized in that: The target detection module includes a region generation network, a priori box pre-screening layer and a prediction output layer; step S3 is specifically as follows: Step S31: Generate candidate target regions through a region generation network; The fused features obtained in step S2 are input into the sliding window layer, and a plurality of prior frames of different scales and proportions are generated at each position of the fused features to obtain a set of prior frames; The target label and target position are predicted respectively through the first score prediction fully connected layer and the first regression prediction fully connected layer; Step S32: input all the prior frames into the prior frame pre-screening layer in sequence to obtain S high-quality prior frames; first obtain the foreground probability distribution scores and predicted bounding box regression values ​​of all the prior frames, then sort all the prior frames in descending order according to the foreground probability distribution scores, and traverse the prior frame set starting from the prior frame with the highest score; The traversal operations are as follows: Add the prior box with the highest score in the prior box set to the output result; Calculate the intersection-and-union ratio of the prior frame and other prior frames; remove other prior frames whose intersection-and-union ratio is greater than a preset threshold; Repeat the operation until all prior boxes are traversed and the final output result T is returned; Step S33: input the final output result into the prediction output layer to obtain the positions of the dura mater and the yellow ligament in the spinal canal nerve ultrasound image. The specific process is as follows: Step S331: extracting the corresponding regional feature block from the coordinates of each priori frame in the high-quality priori frame set obtained in S32 to obtain a feature regional block set; Step S332: input the feature blocks in the average pooling layer respectively and then perform feature concatenation to obtain an average pooling feature map; Step S333: input the average pooling feature map into the hole convolution layer to obtain high-dimensional features; Step S34: target classification and bounding box regression; Step S34 is specifically as follows: inputting the high-dimensional features into two parallel fully connected layer networks, namely the second score prediction fully connected layer and the second regression prediction fully connected layer; obtaining the position coordinates of the dura mater and the yellow ligament in the image.

5. The feature-enhanced spinal nerve anesthesia puncture guidance method according to claim 4, characterized in that: In step S4, the trained feature enhancement module and target detection module are deployed on the background server to detect the positions of the dura mater and the ligamentum flavum in the collected spinal canal nerve ultrasound image, specifically: Step S41: collecting the patient's spinal canal nerve ultrasound image and transmitting it to the backend server; Step S42: using the trained feature enhancement module to obtain feature representations of the dura mater and ligamentum flavum of the spinal canal nerve ultrasound image; Step S43: inputting the feature representation of the dura mater and the ligamentum flavum into the target detection module to obtain the positions of the dura mater and the ligamentum flavum in the spinal canal nerve ultrasound image; Step S44: making a decision on nerve anesthesia puncture according to the positions of the patient's dura mater and ligamentum flavum in the spinal canal nerve ultrasound image.

Images