Method for realizing ultrasonic guidance assisted popliteal sciatic nerve block based on U-Net and application

Abstract

The invention provides a U-Net-based method for realizing ultrasonic guidance to assist in spischial nerve block of popliteal fossa and application, and relates to the technical field of medical image processing. The method comprises the following steps: constructing a popliteal space ultrasonic data set and a U-Net model suitable for identifying a target tissue in a popliteal space ultrasonic image; the U-Net model comprises an encoder, a decoder and a jump connection structure for connecting the encoder and the decoder; wherein the encoder comprises a plurality of encoder blocks; an efficient cascade multi-scale cavity convolution ECMAC and ESAM space attention mechanism is executed in sequence in each encoder block; an encoder cross-regional feature fusion structure is arranged in the encoder; the decoder comprises a plurality of decoder blocks with an ESAM space attention mechanism; a decoder cross-regional feature fusion structure is arranged in the decoder; acquiring a popliteal space ultrasonic image through ultrasonic probe equipment; and performing segmentation identification on the acquired popliteal space ultrasonic image by using the trained U-Net model so as to extract image features of the target tissue in the popliteal space ultrasonic image. Through the improved U-net model, the accuracy of popliteal ultrasonic image segmentation identification is improved, and an analysis tool is provided for realizing ultrasonic guidance and assistance of popliteal sciatic nerve block.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a method for improving ultrasound-guided assisted popliteal sciatic nerve block based on U-Net. Background Art

[0002] The popliteal fossa is a diamond-shaped area behind the knee joint that contains many important anatomical structures, such as the sciatic nerve, popliteal artery, popliteal vein, and popliteal lymph nodes. Ultrasound examination of the popliteal fossa area can help doctors evaluate the status of these structures and perform related diagnoses or operations.

[0003] For example, the popliteal fossa area has the popliteal artery and popliteal vein, and ultrasound images can be used to evaluate the status of the blood vessels, such as checking for thrombosis, arteriosclerosis, vascular stenosis, and other problems. When there may be a mass or swollen lymph nodes in the popliteal fossa area, ultrasound can be used to evaluate the nature of these abnormal structures and help diagnose diseases such as infection and tumors. For another example, soft tissues in the popliteal fossa, such as muscles and ligaments, can also be examined by ultrasound to help diagnose injuries or lesions to the knee joint or surrounding tissues.

[0004] Popliteal sciatic nerve block (PSNB) is a common anesthesia technique widely used in lower limb surgery. In order to improve the accuracy of the block, ultrasound imaging technology is often used to guide the positioning of the nerve. However, due to the limited quality of ultrasound images (such as noise, high reflection, etc.), it is challenging to accurately identify the location of the sciatic nerve.

[0005] During anesthesia, popliteal ultrasound images can assist anesthesiologists in locating the sciatic nerve and its branches (tibial nerve and common peroneal nerve), providing precise guidance for popliteal sciatic nerve block. Nerve block under ultrasound guidance can help anesthesiologists accurately inject anesthetic drugs, thereby achieving effective local anesthesia.

[0006] In existing methods, ultrasound images are manually analyzed, which is inefficient and subjective. Traditional neural image recognition algorithms have difficulty accurately extracting neural contours in complex ultrasound backgrounds.

[0007] To this end, the present invention provides a method and application of improved ultrasound-guided popliteal sciatic nerve block based on U-Net to automatically identify and locate the sciatic nerve and muscles in the popliteal area, thereby improving the success rate and safety of nerve block, which is a technical problem that needs to be solved urgently. Summary of the invention

[0008] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a method and application of ultrasound-guided popliteal sciatic nerve block based on U-Net. The present invention can use the improved U-net model to process the popliteal ultrasound image, and then automatically identify and locate the sciatic nerve and muscles in the popliteal area, so as to improve the success rate and safety of nerve block.

[0009] In order to solve the existing technical problems, the present invention provides the following technical solutions:

[0010] A method for implementing ultrasound-guided popliteal sciatic nerve block based on U-Net, comprising:

[0011] A popliteal ultrasound dataset and a U-Net model suitable for identifying target tissues in popliteal ultrasound images are constructed; the popliteal ultrasound dataset includes a plurality of popliteal ultrasound images processed by data cropping, data labeling and data enhancement operations; the U-Net model includes an encoder, a decoder, and a jump connection structure connecting the encoder and the decoder; wherein,

[0012] The encoder includes a plurality of cascaded encoder blocks; each encoder block sequentially executes efficient cascaded multi-scale dilated convolution ECMAC and spatial attention mechanism ESAM; wherein the efficient cascaded multi-scale dilated convolution ECMAC is composed of multi-scale dilated convolution and channel attention mechanism ECA; the encoder can respectively capture the image features of the target tissue in the aforementioned popliteal fossa ultrasound image at different sizes; the corresponding cascaded encoder blocks can respectively obtain shallow feature maps of different sizes;

[0013] The encoder is provided with an encoder cross-region feature fusion structure, which can fuse shallow feature maps of different sizes obtained by the cascaded encoder blocks with deep feature maps obtained in the decoder after convolution processing;

[0014] The decoder includes a plurality of decoder blocks with ESAM spatial attention mechanism; each decoder block can correspond to an output deep feature map;

[0015] The decoder is provided with a decoder cross-region feature fusion structure, and the decoder cross-region feature fusion structure can perform feature fusion on the deep feature map output by each layer with the shallow feature map in the encoder after transposed convolution processing;

[0016] The skip connection structure can correspondingly connect the aforementioned encoder block and decoder block;

[0017] Collect popliteal ultrasound images through an ultrasound probe device;

[0018] The trained U-Net model is used to segment and identify the acquired popliteal fossa ultrasound image to extract the image features of the target tissue in the aforementioned popliteal fossa ultrasound image.

[0019] Further, the target tissue includes at least one of the biceps femoris, tibial nerve, common peroneal nerve, artery and semimembranosus muscle;

[0020] The U-Net model uses a preset proportion of popliteal fossa ultrasound images in the popliteal fossa ultrasound dataset as training samples for model training;

[0021] After completing the training, the remaining proportion of popliteal ultrasound images in the aforementioned popliteal ultrasound dataset was used as test samples for model testing, and the dice score was used as an evaluation indicator to measure the performance of the U-Net model in the ultrasound popliteal image segmentation task.

[0022] Furthermore, when executing the multi-scale dilated convolution, the steps specifically include:

[0023] For the input feature map X∈R C×H×W , C, H and W are channels, height and width respectively, let the input feature map X∈R C×H×W It undergoes three convolutions with the same kernel size of 3×3 and dilation rates of 1, 2, and 5 respectively, and each convolution is followed by a BN layer and ReLU activation function.

[0024] Use residual connection to add the feature maps output after the first convolution and the second convolution by element-by-element addition to supplement the details of the image features;

[0025] The feature map obtained by adding elements step by step is then concatenated with the feature map after the third convolution to form a new feature map X1∈R 2C×H×W .

[0026] Further, the execution of the ECA channel attention mechanism includes the steps of:

[0027] For the concatenated feature map X1∈R 2C×H×W , through global average pooling, the average value of each channel is obtained:

[0028]

[0029] Among them, z c represents the average value of all pixels on the cth channel, X c，i，j It refers to the pixel value of feature map X at channel c, height i, and width j;

[0030] The average value z for each channel c ∈R C, using a one-dimensional convolution operation with a kernel size of k, we get s = Conv1D(z,k), where k is a positive integer;

[0031] The result s obtained after the convolution operation of each channel is passed through the Sigmoid activation function to obtain the corresponding attention weight α c =σ(s); where σ(·) represents the Sigmoid activation function;

[0032] The aforementioned attention weight α c Applied to the concatenated feature map X1, and reweighting each channel, we get in, is the pixel value at channel c, height i, and width j in the reweighted feature map X1.

[0033] Furthermore, the execution of the ESAM spatial attention mechanism specifically includes the steps of:

[0034] For the input feature map P∈R 2C×H×W After maximum pooling and average pooling respectively, and compression along the channel dimension, the feature map P is obtained by maximum pooling and 1×1 convolution. M ∈R 1×H×W , after average pooling and 1×1 convolution, the feature map is P A ∈R 1×H×W ; At the same time, the aforementioned input feature map P is subjected to depth-separable convolution to obtain the feature map P D ∈R 1×H×W ;

[0035] The above feature map P M ∈R 1×H×W , P A ∈R 1×H×W and P D ∈R 1×H×W Concatenate in the channel dimension to obtain the feature fusion graph F = concat(P M ,P A ,P D );

[0036] Apply a 7×7 convolution to the above feature fusion map F to obtain the feature fusion map Conv(F)∈R 1×H×W ; Fusion graph Conv(F)∈R 1×H×W Through the Sigmoid activation function, we get the attention map M = σ(Conv(F)), M∈R 1×H×W ;

[0037] Multiply the input feature map P by the aforementioned attention map M element by element to obtain the weighted feature map P′=M⊙P, where ⊙ represents element-by-element multiplication.

[0038] The weighted feature map P′ is subjected to a depth-separable convolution to obtain the feature map P * ∈R C×H×W .

[0039] Furthermore, the encoder cross-region feature fusion structure can reduce the size of the aforementioned shallow feature map while increasing the number of channels through step-by-step 3×3 convolutions until the size and number of channels of the aforementioned shallow feature map are the same as the size and number of channels of any deep feature map in the decoder, and then perform feature fusion on the aforementioned shallow feature map and the aforementioned deep feature map.

[0040] Furthermore, the decoder cross-region feature fusion structure can perform feature fusion of the deep feature map with the shallow feature map by gradually upsampling the size and the number of channels of the deep feature map while keeping the size and the number of channels consistent with any shallow feature map.

[0041] A device for implementing ultrasound-guided popliteal sciatic nerve block based on U-Net, comprising:

[0042] A data and model construction unit, used to construct a popliteal ultrasound data set and a U-Net model suitable for identifying target tissues in popliteal ultrasound images; the popliteal ultrasound data set includes multiple popliteal ultrasound images processed by data cropping, data labeling and data enhancement operations; the U-Net model includes an encoder, a decoder, and a jump connection structure connecting the aforementioned encoder and decoder; wherein the encoder includes multiple cascaded encoder blocks; each encoder block sequentially executes efficient cascaded multi-scale void convolution ECMAC and ESAM spatial attention mechanism; wherein the efficient cascaded multi-scale void convolution ECMAC is composed of multi-scale void convolution and ECA channel attention mechanism; the encoder can respectively capture the image features of the target tissues in the aforementioned popliteal ultrasound images at different sizes; The encoder blocks corresponding to the above cascade can respectively obtain shallow feature maps of different sizes; the encoder is provided with an encoder cross-region feature fusion structure, which can fuse the shallow feature maps of different sizes obtained by the above cascaded encoder blocks with the deep feature maps obtained in the decoder after convolution processing; the decoder includes a plurality of decoder blocks with ESAM spatial attention mechanism; each decoder block can output a deep feature map correspondingly; the decoder is provided with a decoder cross-region feature fusion structure, which can fuse the deep feature maps output by each layer with the shallow feature maps in the encoder after transposed convolution processing; the jump connection structure can splice the above encoder blocks with the decoder blocks accordingly;

[0043] An image acquisition unit, used for acquiring a popliteal fossa ultrasound image through an ultrasound probe device;

[0044] The image recognition unit is used to segment and recognize the collected popliteal fossa ultrasound image using the trained U-Net model to extract the image features of the target tissue in the popliteal fossa ultrasound image.

[0045] A system for implementing ultrasound-guided popliteal sciatic nerve block based on U-Net, comprising:

[0046] A network node for sending and receiving popliteal ultrasound images;

[0047] An image segmentation module, used to identify the target tissue in the popliteal fossa ultrasound image and perform segmentation processing;

[0048] A system server, wherein the system server is connected to the network node and the image segmentation module;

[0049] The system server is configured to: construct a popliteal ultrasound dataset and a U-Net model suitable for identifying target tissues in popliteal ultrasound images; the popliteal ultrasound dataset includes multiple popliteal ultrasound images processed by data cropping, data labeling and data enhancement operations; the U-Net model includes an encoder, a decoder, and a jump connection structure connecting the aforementioned encoder and decoder; wherein the encoder includes multiple cascaded encoder blocks; each encoder block sequentially executes efficient cascaded multi-scale dilated convolution ECMAC and ESAM spatial attention mechanism; wherein the efficient cascaded multi-scale dilated convolution ECMAC is composed of multi-scale dilated convolution and ECA channel attention mechanism; the encoder can respectively capture the image features of the target tissues in the aforementioned popliteal ultrasound images at different sizes; the corresponding encoder blocks of the above cascade can respectively obtain shallow feature maps of different sizes; the encoder is set There is an encoder cross-region feature fusion structure, which can fuse the shallow feature maps of different sizes obtained by the above-mentioned cascaded encoder blocks with the deep feature maps obtained in the decoder after convolution processing; the decoder includes multiple decoder blocks with ESAM spatial attention mechanism; each decoder block can output a corresponding deep feature map; the decoder is provided with a decoder cross-region feature fusion structure, which can fuse the deep feature maps output by each layer with the shallow feature maps in the encoder after transposed convolution processing; the jump connection structure can correspondingly splice the above-mentioned encoder blocks and decoder blocks; the popliteal fossa ultrasound image is collected by an ultrasound probe device; the collected popliteal fossa ultrasound image is segmented and recognized using the trained U-Net model to extract the image features of the target tissue in the above-mentioned popliteal fossa ultrasound image.

[0050] A computer-readable storage medium stores a computer program, wherein the computer program, when executed by a processor, implements the implementation steps of any of the above methods.

[0051] Based on the above advantages and positive effects, the advantages of the present invention are: a U-Net model suitable for identifying target tissues in popliteal fossa ultrasound images is designed; the encoder of the U-Net model includes multiple cascaded encoder blocks; each encoder block executes efficient cascaded multi-scale dilated convolution ECMAC and spatial attention mechanism ESAM in sequence; wherein the efficient cascaded multi-scale dilated convolution ECMAC is composed of multi-scale dilated convolution and channel attention mechanism ECA; the encoder can respectively capture the image features of the target tissues in the aforementioned popliteal fossa ultrasound images at different sizes; the corresponding cascaded encoder blocks can respectively obtain shallow feature maps of different sizes; the encoder is provided with The encoder cross-region feature fusion structure can fuse the shallow feature maps of different sizes obtained by the above-mentioned cascaded encoder blocks with the deep feature maps obtained in the decoder after convolution processing; the decoder includes multiple decoder blocks with ESAM spatial attention mechanism; each decoder block can output a corresponding deep feature map; the decoder is provided with a decoder cross-region feature fusion structure, which can fuse the deep feature maps output by each layer with the shallow feature maps in the encoder after transposed convolution processing; the jump connection structure can correspondingly splice the above-mentioned encoder blocks and decoder blocks.

[0052] Furthermore, the use of multi-scale dilated convolution and ECA channel attention mechanism in the encoder enables the U-net model to capture image features of different sizes, compensates for the loss of feature information during the U-Net double-layer convolution operation, and reduces the number of parameters and computation.

[0053] Furthermore, the ESAM spatial attention mechanism is used in both the encoder and the decoder, which enables the U-Net model to better focus on important spatial locations and effectively suppress irrelevant information during the popliteal fossa ultrasound image segmentation and recognition process, thereby improving the overall performance of the U-Net model.

[0054] Furthermore, an encoder cross-region feature fusion structure and a decoder cross-region feature fusion structure are set in the encoder and decoder respectively, so that the global and local contextual information can be better captured, which helps to alleviate the problems of gradient vanishing and gradient exploding, making training more stable and efficient, thereby improving the network detection accuracy and enhancing the recognition and positioning of small target tissues.

[0055] Furthermore, the improved U-Net model has a high recognition accuracy for popliteal fossa ultrasound images and can accurately identify and locate the biceps femoris, sciatic nerve (tibial nerve and common peroneal nerve) and semimembranosus muscle. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 A schematic diagram of a flow chart provided for an embodiment of the present invention.

[0057] Figure 2 A schematic diagram of the structure of the U-Net model provided in an embodiment of the present invention.

[0058] Figure 3 A schematic diagram of the structure of an efficient cascaded multi-scale dilated convolution ECMAC provided in an embodiment of the present invention.

[0059] Figure 4 A schematic diagram of the structure of the ESAM spatial attention mechanism provided for an embodiment of the present invention.

[0060] Figure 5 A schematic diagram of a cross-level feature fusion structure of an encoder provided in an embodiment of the present invention.

[0061] Figure 6 A schematic diagram of the structure of a decoder cross-level feature fusion structure provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0062] The following is a further detailed description of a method and application of improved ultrasound-guided popliteal sciatic nerve block based on U-Net disclosed in the present invention in conjunction with the accompanying drawings and specific embodiments. It should be noted that the technical features or combinations of technical features described in the following embodiments should not be considered isolated, and they can be combined with each other to achieve better technical effects. In the drawings of the following embodiments, the same reference numerals appearing in each drawing represent the same features or components, which can be applied to different embodiments. Therefore, once an item is defined in one drawing, it does not need to be further discussed in subsequent drawings.

[0063] It should be noted that the structures, proportions, sizes, etc. illustrated in the drawings of this specification are only used to match the contents disclosed in the specification for people familiar with this technology to understand and read, and are not used to limit the limiting conditions for the implementation of the invention. Any modification of the structure, change of the proportion relationship or adjustment of the size should fall within the scope of the technical content disclosed by the invention without affecting the effects and purposes that can be achieved by the invention. The scope of the preferred embodiments of the present invention includes other implementations, in which the functions can be performed in a substantially simultaneous manner or in a reverse order according to the functions involved, which should be understood by those skilled in the art of the technical field to which the embodiments of the present invention belong.

[0064] Techniques, methods, and devices known to those of ordinary skill in the relevant art may not be discussed in detail, but where appropriate, the techniques, methods, and devices should be considered part of the authorization specification. In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limiting. Therefore, other examples of the exemplary embodiments may have different values.

[0065] Example

[0066] See also Figure 1 FIG. 1 is a flow chart of the present invention. The implementation step S100 of the method is as follows:

[0067] S101, constructing a popliteal fossa ultrasound dataset and a U-Net model suitable for identifying target tissues in popliteal fossa ultrasound images.

[0068] The popliteal fossa ultrasound data set includes a plurality of popliteal fossa ultrasound images processed by data cropping, data labeling and data enhancement operations (wherein the data enhancement operations include occlusion, mirroring, rotation, etc.). The popliteal fossa ultrasound image refers to an image obtained by ultrasound examination in the popliteal fossa region (i.e., the posterior knee region).

[0069] When constructing a popliteal ultrasound dataset, it is preferred that the obtained popliteal ultrasound image at the popliteal sciatic nerve is divided into multiple categories, and the target tissue contained in the popliteal ultrasound image is marked, and finally constructed by data enhancement. The target tissue includes but is not limited to at least one of the biceps femoris, tibial nerve, common peroneal nerve, artery and semimembranosus.

[0070] The popliteal ultrasound image dataset in this embodiment is derived from the practice of clinical anesthesia in a certain hospital. Popliteal ultrasound images of 500 patients are collected and sorted to construct the popliteal ultrasound image dataset, which contains about 3,000 images taken by popliteal ultrasound examination.

[0071] During the sorting process, the original popliteal ultrasound images were first renumbered to protect the privacy of the patient. Then, 1,500 popliteal ultrasound images containing the biceps femoris, tibial nerve, common peroneal nerve, artery, and semimembranosus were selected. In order to ensure the invariance of the characteristics of the tissue area and reduce the amount of calculation, this embodiment preferably crops and masks some of the original popliteal ultrasound images. For each image containing part of the target tissue, three technicians in this field segmented it according to the different anatomical structures of the target tissue and labeled it using the LabelMe tool.

[0072] Among them, it is worth mentioning that the U-Net model uses a preset proportion of popliteal fossa ultrasound images in the popliteal fossa ultrasound dataset as training samples for model training; after the training is completed, the remaining proportion of popliteal fossa ultrasound images in the aforementioned popliteal fossa ultrasound dataset are used as test samples for model testing, and the dice score is used as an evaluation indicator for measuring the performance of the U-Net model in the ultrasound popliteal fossa image segmentation task.

[0073] In this embodiment, the U-Net model preferably uses 80% of the popliteal fossa ultrasound images in the popliteal fossa ultrasound dataset as training samples for model training; after the training is completed, the remaining 20% ​​of the popliteal fossa ultrasound images in the aforementioned popliteal fossa ultrasound dataset are used as test samples for model testing.

[0074] Combination Figure 2 As shown, the U-Net model in this embodiment includes an encoder, a decoder, and a jump connection structure connecting the encoder and the decoder.

[0075] Specifically, the encoder includes multiple cascaded encoder blocks; each encoder block executes efficient cascaded multi-scale dilated convolution ECMAC and ESAM spatial attention mechanism in sequence; wherein the efficient cascaded multi-scale dilated convolution ECMAC is composed of multi-scale dilated convolution and channel attention mechanism ECA.

[0076] Specific combination Figure 3 , which is a schematic diagram of the structure of an efficient cascaded multi-scale atrous convolution ECMAC in an encoder provided by an embodiment of the present invention.

[0077] The Efficient cascaded multi-scale atrous convolutions (ECMAC) specifically uses multi-scale atrous convolutions and ECA channel attention mechanism, which enables the U-Net model to capture image features of different sizes and compensates for the loss of feature information during the U-Net double-layer convolution operation.

[0078] When performing multi-scale dilated convolution operations, combined with Figure 3 As shown, it specifically includes step S110:

[0079] S111, for the input feature map X∈R C×H×W , C, H and W are channels, height and width respectively, let the input feature map X∈R C×H×W It undergoes three convolutions with the same kernel size of 3×3 and dilation rates of 1, 2, and 5 respectively, and each convolution is processed using a BN layer and ReLU activation function.

[0080] Since multi-scale dilated convolution may produce artifacts or discontinuous features when processing boundary information, when executing step S112, a residual connection is used to solve this problem.

[0081] S112 uses residual connection to add the feature maps output after the first convolution and the second convolution by element-wise addition to supplement the details of the image features.

[0082] S113, the feature map obtained by adding elements step by step is then channel-joined with the feature map after the third convolution to form a new feature map X1∈R 2C×H×W .

[0083] The execution of the above multi-scale dilated convolution operation can improve the richness and diversity of feature maps, but in this operation, some feature maps may contribute more to the relevant information of the identified target, while other feature maps may contribute less to the relevant information of the identified target.

[0084] To this end, in this embodiment, it is preferred to combine the ECA channel attention mechanism, preferably by calculating the intensity weight of each channel to enhance the important features in the feature map (preferably the target tissue in this embodiment), and suppress unimportant features (preferably other tissues in the popliteal fossa ultrasound image except the aforementioned target tissue), thereby improving the expressive power of the feature map, enabling the model to better capture global and local information, improve the accuracy and robustness of the segmentation, and finally output a feature map of size 2C×H×W.

[0085] Specifically, the execution of the ECA channel attention mechanism includes step S120:

[0086] S121, for the concatenated feature map X1∈R 2C×H×W , through global average pooling (GAP), the spatial information of each channel is aggregated into a scalar, that is, the average value of each channel is obtained:

[0087]

[0088] Among them, z c represents the average value of all pixels on the cth channel, X c，i，j It refers to the pixel value of feature map X1 at channel c, height i, and width j.

[0089] After obtaining the aforementioned average value z c Afterwards, it is preferred to perform a one-dimensional convolution operation to capture the cross-channel dependencies. That is:

[0090] S122, average value z for each channelc ∈R C , using a one-dimensional convolution operation with a kernel size of k, we get s=Conv1D(z,k), where k is a positive integer.

[0091] Specifically, R C It refers to a C-dimensional real vector space, in which each vector element corresponds to the average value of a channel. Among them, R represents a real number set, R C represents the set of all vectors containing C real numbers. For example, when C = 3, R C It represents the three-dimensional real number space, that is, all vectors of the form (x1,x2,x3), where x1,x2,x3 are real numbers.

[0092] In this embodiment, the size k of the convolution kernel is adaptively selected to a smaller value according to the number of channels, and k=3 is preferably selected here.

[0093] S123, the result s obtained after the convolution operation of each channel is passed through the Sigmoid activation function to obtain the corresponding attention weight α c =σ(s); where σ(·) represents the Sigmoid activation function.

[0094] S124, the aforementioned attention weight α c Applied to the concatenated feature map X1, and reweighting each channel, we get in, is the pixel value at channel c, height i, and width j in the reweighted feature map X1.

[0095] As another preferred implementation of this embodiment, the input feature map P∈R 2C×H×W , where C is the number of channels, H is the height, and W is the width, using the ESAM spatial attention mechanism.

[0096] Specific, combined Figure 4 As shown, the execution of the ESAM spatial attention mechanism specifically includes step S130:

[0097] S131, for the input feature map P∈R 2C×H×W After maximum pooling and average pooling respectively, and compression along the channel dimension, the feature map P is obtained by maximum pooling and 1×1 convolution. M ∈R 1×H×W , after average pooling and 1×1 convolution, the feature map is P A ∈R 1×H×W ; At the same time, the aforementioned input feature map P is subjected to depth-separable convolution to obtain the feature map P D ∈R 1×H×W .

[0098] It is worth noting that the depthwise separable convolution can reduce the computational complexity and the number of parameters because the depthwise separable convolution can decompose the standard convolution into depthwise convolution and pointwise convolution.

[0099] In standard convolution, the size of the convolution kernel is usually K×K×C, while depthwise convolution is a convolution operation performed separately for each input channel, with a convolution kernel size of K×K×1.

[0100] Combination Figure 4 As shown in the figure, the size of the input tensor is C×H×W. In the depthwise convolution, the convolution kernel size is k=3, and each channel is independently convolved with a 3×3 convolution operation. To facilitate subsequent image processing, the stride and padding are set to 1 to ensure that the channel, height, and width of the output image remain unchanged.

[0101] Then, point-by-point convolution is performed, which is a 1×1 convolution operation used to linearly combine the channels at each position. The output tensor size of the depthwise convolution is C×H×W, and the point-by-point convolution uses a convolution kernel of 1×1×C×C′, where C′ is the number of output channels, i.e., C in the figure. Point-by-point convolution generates an output tensor of C′×H×W by performing a 1×1 convolution on each position. This operation linearly combines the features obtained by the depthwise convolution in the channel dimension, and finally obtains the desired output feature map.

[0102] S132, the aforementioned feature map P M ∈R 1×H×W , P A ∈R 1×H×W and P D ∈R 1×H×W Concatenate in the channel dimension to obtain the feature fusion graph F = concat(P M ,P A ,P D ). At this time, the obtained feature fusion map is a feature map of size 3×H×W.

[0103] In actual operation, in order to capture more subtle local features in the feature map P and extract more useful information, the feature map P is obtained by using depthwise separable convolution while using maximum pooling and average pooling. D ∈R 1×H×W After that, the feature map P M ∈R 1×H×W , P A ∈R 1×H×W and P D ∈R 1×H×WThe concatenation is performed on the channel dimension, and the resulting feature fusion graph F = concat(P M ,P A ,P D ), thereby enhancing the richness of feature expression.

[0104] S133, use a 7×7 convolution on the above feature fusion map F to obtain the feature fusion map Conv(F)∈R 1×H×W ; Fusion graph Conv(F)∈R 1×H×W Through the Sigmoid activation function, we get the attention map M = σ(Conv(F)), M∈R 1 ×H×W .

[0105] In order to further enhance the features of those spatial positions considered important in the attention map M, steps S134 and S135 are performed.

[0106] S134, perform element-wise product on the input feature map P and the aforementioned attention map M to obtain a weighted feature map P′=M⊙P; wherein ⊙ represents element-wise multiplication.

[0107] S135, the weighted feature map P′ is subjected to a depth-separable convolution to obtain the feature map P * ∈R C×H×W For subsequent image processing.

[0108] Among them, executing step S135 to perform depth-separable convolution on the weighted feature map P′ can halve the dimension while extracting higher-level feature information, thereby optimizing the feature representation and improving the overall performance of the U-Net model.

[0109] It is also worth noting that in the encoder part of this embodiment, the encoder can respectively capture the image features of the target tissue in the aforementioned popliteal fossa ultrasound image at different sizes; the corresponding encoder blocks of the above cascade can respectively obtain shallow feature maps of different sizes.

[0110] Specific, combined Figure 2 As shown in the figure, after image processing by four encoder blocks, shallow feature maps F1, F2, F3 and F4 with sizes of 64×512×512, 128×256×256, 256×128×128 and 512×64×64 are obtained respectively.

[0111] In addition, it is worth mentioning that the encoder is also provided with an encoder cross-region feature fusion structure, which can fuse the shallow feature maps of different sizes obtained by the above-mentioned cascaded encoder blocks with the deep feature maps obtained in the decoder after convolution processing.

[0112] Specific, combined Figure 5 As shown, the encoder cross-region feature fusion structure can reduce the size of the aforementioned shallow feature map while increasing the number of channels through step-by-step 3×3 convolutions, until the size and number of channels of the aforementioned shallow feature map are the same as the size and number of channels of any deep feature map in the decoder, and then perform feature fusion on the aforementioned shallow feature map and the aforementioned deep feature map.

[0113] The strided 3x3 convolution can decompose a standard 3x3 convolution operation into two smaller convolution operations (e.g., 1x3 and 3x1 convolution) to perform, thereby reducing the amount of computation and memory usage while still effectively capturing spatial relationships.

[0114] As one of the preferred implementations of this embodiment, when the shallow feature maps, i.e., the feature maps F1, F2, F3 and F4 of the first four layers in the encoder structure, each shallow feature map is adjusted to the corresponding size of 1024×32×32 after a step-by-step 3×3 convolution operation, and then preferably added element-by-element with the deep feature map 1024×32×32 in the decoder.

[0115] Preferably, after performing element-by-element addition, the feature map obtained is preferably

[0116] Among them, F k (i,j) represents the value of the kth feature map at position (i,j).

[0117] The element-by-element addition process in the encoder's cross-region feature fusion structure effectively ensures the gradual extraction and integration of information in shallow feature maps, thereby retaining multi-scale contextual information.

[0118] In the decoder of this embodiment, multiple decoder blocks with ESAM spatial attention mechanism are included; each decoder block can correspond to the output deep feature map.

[0119] It is worth mentioning that the ESAM spatial attention mechanism is used in both the encoder downsampling part and the decoder to further enhance the feature representation ability of the model and emphasize the extraction of effective features.

[0120] This is because the jump connection of the original U-Net model directly concatenates the features of the encoder and the decoder, which may contain a lot of redundant or irrelevant information. In order to suppress this irrelevant information, adding the ESAM spatial attention mechanism can help the model focus on important spatial positions and enhance useful feature expressions.

[0121] In addition, the decoder is also provided with a decoder cross-region feature fusion structure, which can perform feature fusion of the deep feature map with the shallow feature map by gradually upsampling the size and the number of channels of the deep feature map while keeping the size and the number of channels consistent with any shallow feature map.

[0122] Combination Figure 6 As shown, the decoder cross-region feature fusion structure fuses deep feature maps of sizes 1024×32×32, 512×64×64, 256×128×128 and 128×256×256 with shallow feature maps of 64×512×512 by gradually upsampling.

[0123] In actual operation, transposed convolution is used to change the size of the deep feature map from smaller to larger, and the number of channels gradually decreases until it becomes 64×512×512. Then, the upsampled feature map is further processed by a 3×3 convolution layer and a ReLU activation function to extract more useful information. This operation is very helpful for identifying smaller targets such as neurons.

[0124] The specific formula is as follows: out =ReLU(Conv2d(Deconv(x))). Where x is the input tensor of shape C×H×W, Deconv is a transposed convolution with a kernel size of 4×4, a stride of 2, and a padding of 1 to ensure that the image size remains unchanged after the convolution. Conv2d uses a 3×3 convolution and a ReLU activation function to get the output image x. out .

[0125] It is worth noting that in this embodiment, the decoder cross-region feature fusion structure uses a step-by-step upsampling method, and the above operation needs to be repeated according to the tensor size of the input image. For example, the deepest feature map needs to go through the above steps 4 times to get a picture of size 64×512×512, while the feature map of the second layer only needs 1 time.

[0126] It is also worth mentioning that this embodiment introduces cross-level feature fusion in both the encoder and the decoder, so it can better capture global and local context information, which helps to alleviate the problems of gradient vanishing and gradient exploding, making training more stable and efficient, thereby improving network detection accuracy and enhancing the recognition and positioning of small targets.

[0127] In addition, the skip connection structure in this embodiment can correspondingly splice the aforementioned encoder block and decoder block.

[0128] At this point, the configuration of the U-net model is completed.

[0129] In order to verify the performance of the improved U-Net model on the custom popliteal ultrasound data set and analyze the experimental results, in actual operation, after completing the U-Net model training, it is preferred to perform steps S102 and S103.

[0130] S102, collecting a popliteal fossa ultrasound image using an ultrasound probe device.

[0131] S103, using the trained U-Net model to segment and identify the collected popliteal fossa ultrasound image, so as to extract image features of the target tissue in the popliteal fossa ultrasound image.

[0132] The experimental results show that for popliteal ultrasound images, the improved U-Net model can accurately identify the nerves (sciatic nerve, tibial nerve and common peroneal nerve) and muscles (biceps femoris and semimembranosus) and arteries in the popliteal fossa, among which the mIOU reaches above 0.89 and the Dice coefficient reaches above 0.94, indicating that the U-Net model has high reliability and very accurate recognition and positioning.

[0133] In addition, it is worth emphasizing that the U-Net model mentioned in this embodiment is aimed at the rotated popliteal fossa ultrasound image and the popliteal fossa ultrasound image under special environment, and the algorithm can also accurately segment the target position.

[0134] For other technical features, please refer to the previous embodiments and will not be described in detail here.

[0135] In addition, the present invention also provides an embodiment, providing a device for implementing ultrasound-guided assisted popliteal sciatic nerve block based on U-Net, comprising:

[0136] A data and model construction unit, used to construct a popliteal ultrasound dataset and a U-Net model suitable for identifying target tissues in popliteal ultrasound images; the popliteal ultrasound dataset includes multiple popliteal ultrasound images processed by data cropping, data labeling and data enhancement operations; the U-Net model includes an encoder, a decoder, and a jump connection structure connecting the aforementioned encoder and decoder; wherein the encoder includes multiple cascaded encoder blocks; each encoder block sequentially executes efficient cascaded multi-scale dilated convolution ECMAC and spatial attention mechanism ESAM; wherein the efficient cascaded multi-scale dilated convolution ECMAC is composed of multi-scale dilated convolution and channel attention mechanism ECA; the encoder composed of the encoder blocks can respectively capture the images of the target tissues in the aforementioned popliteal ultrasound images at different sizes. Features; the encoder blocks corresponding to the above cascade can respectively obtain shallow feature maps of different sizes; the encoder is provided with an encoder cross-region feature fusion structure, and the encoder cross-region feature fusion structure can fuse the shallow feature maps of different sizes respectively obtained by the above cascaded encoder blocks with the deep feature maps obtained in the decoder after convolution processing; the decoder includes multiple decoder blocks with ESAM spatial attention mechanism; each decoder block can output a corresponding deep feature map; the decoder is provided with a decoder cross-region feature fusion structure, and the decoder cross-region feature fusion structure can fuse the deep feature maps output by each layer with the shallow feature maps in the encoder after transposed convolution processing; the jump connection structure can correspondingly splice the above encoder blocks and decoder blocks.

[0137] The image acquisition unit is used to acquire popliteal fossa ultrasound images through an ultrasound probe device.

[0138] The image recognition unit is used to segment and recognize the collected popliteal fossa ultrasound image using the trained U-Net model to extract the image features of the target tissue in the popliteal fossa ultrasound image.

[0139] For other technical features, please refer to the previous embodiments and will not be described in detail here.

[0140] In addition, the present invention also provides an embodiment, providing a system for implementing ultrasound-guided assisted popliteal sciatic nerve block based on U-Net, including:

[0141] A network node used to send and receive popliteal ultrasound images.

[0142] The image segmentation module is used to identify the target tissue in the popliteal fossa ultrasound image and perform segmentation processing.

[0143] A system server is connected to the network node and the image segmentation module.

[0144] The system server is configured to: construct a popliteal ultrasound dataset and a U-Net model suitable for identifying target tissues in popliteal ultrasound images; the popliteal ultrasound dataset includes multiple popliteal ultrasound images processed by data cropping, data labeling and data enhancement operations; the U-Net model includes an encoder, a decoder, and a jump connection structure connecting the aforementioned encoder and decoder; wherein the encoder includes multiple cascaded encoder blocks; each encoder block sequentially executes efficient cascaded multi-scale dilated convolution ECMAC and spatial attention mechanism ESAM; wherein the efficient cascaded multi-scale dilated convolution ECMAC is composed of multi-scale dilated convolution and channel attention mechanism ECA; the encoder can respectively capture the image features of the target tissue in the aforementioned popliteal ultrasound image at different sizes; the corresponding encoder blocks of the above cascade can respectively obtain shallow feature maps of different sizes; the encoder is set There is an encoder cross-region feature fusion structure, which can fuse the shallow feature maps of different sizes obtained by the above-mentioned cascaded encoder blocks with the deep feature maps obtained in the decoder after convolution processing; the decoder includes multiple decoder blocks with ESAM spatial attention mechanism; each decoder block can output a corresponding deep feature map; the decoder is provided with a decoder cross-region feature fusion structure, which can fuse the deep feature maps output by each layer with the shallow feature maps in the encoder after transposed convolution processing; the jump connection structure can correspondingly splice the above-mentioned encoder blocks and decoder blocks; the popliteal fossa ultrasound image is collected by an ultrasound probe device; the collected popliteal fossa ultrasound image is segmented and recognized using the trained U-Net model to extract the image features of the target tissue in the above-mentioned popliteal fossa ultrasound image.

[0145] For other technical features, please refer to the previous embodiments and will not be repeated here.

[0146] In addition, an embodiment of the present invention also provides a computer-readable storage medium having a program stored thereon for use in the aforementioned system for improving ultrasound-guided popliteal sciatic nerve block based on U-Net. When the program is executed by a processor, it can implement any of the steps of the method for improving ultrasound-guided popliteal sciatic nerve block based on U-Net.

[0147] For other technical features, please refer to the previous embodiments and will not be repeated here.

[0148] In the above description, within the scope of the target protection of the present disclosure, the components can be selectively and operatively combined in any number. In addition, terms such as "including", "comprising" and "having" should be interpreted as inclusive or open by default, rather than exclusive or closed, unless they are explicitly defined to the contrary. All technical, scientific or other terms have the meaning understood by those skilled in the art unless they are defined to the contrary. Common terms found in dictionaries should not be interpreted too idealistically or too impractically in the context of relevant technical documents, unless the present disclosure explicitly defines them as such.

[0149] Although the exemplary aspects of the present disclosure have been described for illustrative purposes, those skilled in the art should appreciate that the above description is only a description of the preferred embodiments of the present invention and is not intended to limit the scope of the present invention. The scope of the preferred embodiments of the present invention includes other implementations in which functions may not be performed in the order in which they appear or are discussed. Any changes or modifications made by those skilled in the art based on the above disclosure are within the scope of the claims.

Claims

1. A method for implementing ultrasound-guided popliteal sciatic nerve block based on U-Net, characterized in that: include: A popliteal ultrasound dataset and a U-Net model suitable for identifying target tissues in popliteal ultrasound images are constructed; the popliteal ultrasound dataset includes a plurality of popliteal ultrasound images processed by data cropping, data labeling and data enhancement operations; the U-Net model includes an encoder, a decoder, and a jump connection structure connecting the encoder and the decoder; wherein, The encoder includes a plurality of cascaded encoder blocks; each encoder block sequentially executes efficient cascaded multi-scale dilated convolution ECMAC and spatial attention mechanism ESAM; wherein the efficient cascaded multi-scale dilated convolution ECMAC is composed of multi-scale dilated convolution and channel attention mechanism ECA; the encoder can respectively capture the image features of the target tissue in the aforementioned popliteal fossa ultrasound image at different sizes; the corresponding cascaded encoder blocks can respectively obtain shallow feature maps of different sizes; The encoder is provided with an encoder cross-region feature fusion structure, which can fuse shallow feature maps of different sizes obtained by the cascaded encoder blocks with deep feature maps obtained in the decoder after convolution processing; The decoder includes a plurality of decoder blocks with ESAM spatial attention mechanism; each decoder block can correspond to an output deep feature map; The decoder is provided with a decoder cross-region feature fusion structure, and the decoder cross-region feature fusion structure can perform feature fusion on the deep feature map output by each layer with the shallow feature map in the encoder after transposed convolution processing; The skip connection structure can correspondingly connect the aforementioned encoder block and decoder block; Collect popliteal ultrasound images through an ultrasound probe device; The trained U-Net model is used to segment and identify the acquired popliteal fossa ultrasound image to extract the image features of the target tissue in the aforementioned popliteal fossa ultrasound image.

2. The method according to claim 1, characterized in that The target tissue includes at least one of the biceps femoris, the tibial nerve, the common peroneal nerve, the artery, and the semimembranosus muscle; The U-Net model uses a preset proportion of popliteal fossa ultrasound images in the popliteal fossa ultrasound dataset as training samples for model training; After completing the training, the remaining proportion of popliteal ultrasound images in the aforementioned popliteal ultrasound dataset was used as test samples for model testing, and the dice score was used as an evaluation indicator to measure the performance of the U-Net model in the ultrasound popliteal image segmentation task.

3. The method according to claim 1, characterized in that When executing the multi-scale dilated convolution, the steps are as follows: For the input feature map X∈R C×H×W , C, H and W are channels, height and width respectively, let the input feature map X∈R C×H×W It undergoes three convolutions with the same kernel size of 3×3 and dilation rates of 1, 2, and 5 respectively, and each convolution is followed by a BN layer and ReLU activation function. Use residual connection to add the feature maps output after the first convolution and the second convolution by element-by-element addition to supplement the details of the image features; The feature map obtained by adding elements step by step is then concatenated with the feature map after the third convolution to form a new feature map X1∈R 2C×H×W .

4. The method according to claim 3, characterized in that The execution of the ECA channel attention mechanism includes the following steps: For the concatenated feature map X1∈R 2C×H×W , through global average pooling, the average value of each channel is obtained: Among them, z c represents the average value of all pixels on the cth channel, X c，i，j It refers to the pixel value of feature map X at channel c, height i, and width j; The average value z for each channel c ∈R C , using a one-dimensional convolution operation with a kernel size of k, we get s = Conv1D(z,k), where k is a positive integer; The result s obtained after the convolution operation of each channel is passed through the Sigmoid activation function to obtain the corresponding attention weight α c =σ(s); where σ(·) represents the Sigmoid activation function; the aforementioned attention weight α c Applied to the concatenated feature map X1, and reweighting each channel, we get in, is the pixel value at channel c, height i, and width j in the reweighted feature map X1.

5. The method according to claim 1, characterized in that The execution of the ESAM spatial attention mechanism specifically includes the following steps: For the input feature map P∈R 2C×H×W After maximum pooling and average pooling respectively, and compression along the channel dimension, the feature map P is obtained by maximum pooling and 1×1 convolution. M ∈R 1×H×W , after average pooling and 1×1 convolution, the feature map is P A ∈R 1 ×H×W ; At the same time, the aforementioned input feature map P is subjected to depth-separable convolution to obtain the feature map P D ∈R 1×H×W ; The above feature map P M ∈R 1×H×W , P A ∈R 1×H×W and P D ∈R 1×H×W Concatenate in the channel dimension to obtain the feature fusion graph F = concat(P M ,P A ,P D ); Apply a 7×7 convolution to the above feature fusion map F to obtain the feature fusion map Conv(F)∈R 1×H×W ; The feature fusion graph Conv(F)∈R 1×H×W Through the Sigmoid activation function, we get the attention map M = σ(Conv(F)), M∈R 1×H×W ; Multiply the input feature map P by the aforementioned attention map M element by element to obtain the weighted feature map P′=M⊙P, where ⊙ represents element-by-element multiplication. The weighted feature map P′ is subjected to a depth-separable convolution to obtain the feature map P * ∈R C×H×W .

6. The method according to claim 1, characterized in that The encoder cross-region feature fusion structure can reduce the size of the aforementioned shallow feature map while increasing the number of channels through step-by-step 3×3 convolutions until the size and number of channels of the aforementioned shallow feature map are the same as the size and number of channels of any deep feature map in the decoder, and then perform feature fusion on the aforementioned shallow feature map and the aforementioned deep feature map.

7. The method according to claim 1, characterized in that The decoder cross-region feature fusion structure can perform feature fusion of the deep feature map with the shallow feature map by gradually upsampling the size and the number of channels of the deep feature map while keeping the size and the number of channels consistent with any shallow feature map.

8. A device for implementing ultrasound-guided popliteal sciatic nerve block based on U-Net according to any one of claims 1 to 7, characterized in that include: A data and model construction unit, used to construct a popliteal ultrasound data set and a U-Net model suitable for identifying target tissues in popliteal ultrasound images; the popliteal ultrasound data set includes multiple popliteal ultrasound images processed by data cropping, data labeling and data enhancement operations; the U-Net model includes an encoder, a decoder, and a jump connection structure connecting the aforementioned encoder and decoder; wherein the encoder includes multiple cascaded encoder blocks; each encoder block sequentially executes efficient cascaded multi-scale dilated convolution ECMAC and spatial attention mechanism ESAM; wherein the efficient cascaded multi-scale dilated convolution ECMAC is composed of multi-scale dilated convolution and channel attention mechanism ECA; the encoder can respectively capture the image features of the target tissues in the aforementioned popliteal ultrasound images at different sizes; The encoder blocks corresponding to the above cascade can respectively obtain shallow feature maps of different sizes; the encoder is provided with an encoder cross-region feature fusion structure, which can fuse the shallow feature maps of different sizes obtained by the above cascaded encoder blocks with the deep feature maps obtained in the decoder after convolution processing; the decoder includes a plurality of decoder blocks with ESAM spatial attention mechanism; each decoder block can output a deep feature map correspondingly; the decoder is provided with a decoder cross-region feature fusion structure, which can fuse the deep feature maps output by each layer with the shallow feature maps in the encoder after transposed convolution processing; the jump connection structure can splice the above encoder blocks with the decoder blocks accordingly; An image acquisition unit, used for acquiring a popliteal fossa ultrasound image through an ultrasound probe device; The image recognition unit is used to segment and recognize the collected popliteal fossa ultrasound image using the trained U-Net model to extract the image features of the target tissue in the popliteal fossa ultrasound image.

9. A system for implementing ultrasound-guided popliteal sciatic nerve block based on U-Net according to any one of claims 1 to 7, characterized in that include: A network node for sending and receiving popliteal ultrasound images; An image segmentation module, used to identify the target tissue in the popliteal fossa ultrasound image and perform segmentation processing; A system server, wherein the system server is connected to the network node and the image segmentation module; The system server is configured to: construct a popliteal ultrasound dataset and a U-Net model suitable for identifying target tissues in popliteal ultrasound images; the popliteal ultrasound dataset includes multiple popliteal ultrasound images processed by data cropping, data labeling and data enhancement operations; the U-Net model includes an encoder, a decoder, and a jump connection structure connecting the aforementioned encoder and decoder; wherein the encoder includes multiple cascaded encoder blocks; each encoder block sequentially executes efficient cascaded multi-scale dilated convolution ECMAC and spatial attention mechanism ESAM; wherein the efficient cascaded multi-scale dilated convolution ECMAC is composed of multi-scale dilated convolution and channel attention mechanism ECA; the encoder can respectively capture the image features of the target tissue in the aforementioned popliteal ultrasound image at different sizes; the corresponding encoder blocks of the above cascade can respectively obtain shallow feature maps of different sizes; the encoder is set There is an encoder cross-region feature fusion structure, which can fuse the shallow feature maps of different sizes obtained by the above-mentioned cascaded encoder blocks with the deep feature maps obtained in the decoder after convolution processing; the decoder includes multiple decoder blocks with ESAM spatial attention mechanism; each decoder block can output a corresponding deep feature map; the decoder is provided with a decoder cross-region feature fusion structure, which can fuse the deep feature maps output by each layer with the shallow feature maps in the encoder after transposed convolution processing; the jump connection structure can correspondingly splice the above-mentioned encoder blocks and decoder blocks; the popliteal fossa ultrasound image is collected by an ultrasound probe device; the collected popliteal fossa ultrasound image is segmented and recognized using the trained U-Net model to extract the image features of the target tissue in the above-mentioned popliteal fossa ultrasound image.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method steps described in any one of claims 1 to 7 are implemented.

Images