Ultrasonic guided pelvis puncture method based on deep learning

Abstract

The invention discloses a deep learning-based ultrasonic guide renal pelvis puncture method, which comprises the following steps of: acquiring and processing an ultrasonic image of a renal pelvis, constructing a basic data set and an enhanced data set, constructing a multi-scale deep learning model, inputting the basic data set into the multi-scale deep learning model for processing, the method comprises the steps of obtaining channel attention features and space attention features, fusing the channel attention features and the space attention features to obtain fusion features of a current layer, and processing an enhanced data set and the fusion features of the current layer to obtain pixel coordinates of a renal pelvis in an ultrasonic image for accurately positioning the position of the renal pelvis in the ultrasonic image of the renal pelvis. According to the method, the kidney features under different scales are effectively extracted by constructing the multi-scale deep learning model, the influence of the ultrasonic imaging angle and depth on the ultrasonic image quality of the pelvis is overcome, and the accuracy of pelvis puncture positioning is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical fields of intelligent ultrasonic image recognition and urology, and particularly relates to an ultrasonic-guided renal pelvis puncture method based on deep learning. Background Technique

[0002] Renal pelvis puncture is one of the commonly used diagnostic and treatment means in urology, mainly used for treating diseases such as kidney stones and hydronephrosis. After the introduction of ultrasonic equipment into the clinic, ultrasonic-guided renal pelvis puncture has become the preferred puncture method in the clinic due to its real-time performance, accuracy, and safety. Through ultrasonic guidance, doctors can directly observe the dynamic images of the kidney and surrounding tissues, accurately locate the puncture target, and reduce the occurrence of mis-puncture and complications. However, the ultrasonic-guided renal pelvis puncture technique requires high operating experience of doctors and needs long-term professional training and practical accumulation. Therefore, there is an urgent need for an intelligent auxiliary technology to help doctors quickly and accurately locate the target during the puncture process and improve the puncture success rate and safety.

[0003] The Chinese invention patent with the publication number CN 202859244 U proposes a renal pelvis puncture needle, designs the model of the renal pelvis puncture needle, and the main content is to design a renal pelvis puncture needle: the puncture needle includes structures such as an outer needle tube, an outer needle seat, an inner needle core, and an inner needle seat, and filters and grooves are designed on the side of the inner needle core to improve the safety and convenience of puncture.

[0004] The Chinese invention patent with the publication number CN 117197098 A proposes an automatic segmentation method for children's kidneys and pyelonephritis applied to computed tomography based on deep learning, which relates to the field of deep learning. The method includes the following steps: S1, local feature extraction stage; according to the image of the children's kidneys and pyelonephritis to be segmented, local features of the input image are obtained through local feature extraction; S2, global feature extraction stage; the input of the local feature extraction stage is used as the input of the global feature extraction stage to extract the global feature information of the image; S3, decoding and upsampling stage; using the output of the global feature extraction stage as the input, after feature fusion with the output of the local feature extraction stage, through the designed adaptive attention module, the model learns the important information in the fused features and filters redundant information; S4, segmentation stage; the input is converted into the segmented categories using convolution for segmentation.

[0005] The Chinese invention patent with the publication number CN117557577A proposes a method and system for renal pelvis and ureter segmentation based on non-contrast CT and deep learning, designing deep learning and renal pelvis segmentation. The method includes the following steps: S1. Obtain non-contrast CT images of the urinary system; S2. Use a Swin-UNet object detection model based on the attention mechanism to process the non-contrast CT images with the non-contrast CT images as the input; S3. The object detection model includes several layers of neural networks. In each layer of neural network, downsampling, windowing, and adjacent comparison operations are sequentially performed to obtain the feature map of each layer; S4. Starting from the smallest feature map, upsampling, windowing, and adjacent comparison operations are sequentially performed, and then connected to the high-level feature map to determine whether each window block in the high-level feature map belongs to a certain organ; S5. Repeat multiple times to obtain the positions of the bilateral kidneys in the input image and provide the relevant positions to two segmentation models; S6. The first segmentation model is used to segment the renal pelvis and the upper segment of the ureter to obtain the first segmentation result, and the second segmentation model is used to segment the middle and lower segments of the ureter to obtain the second segmentation result, and the first segmentation result and the second segmentation result are added and processed to obtain the renal pelvis structure and the ureter structure.

[0006] However, the above technology has the following limitations: There will be electromagnetic interference noise and tissue scattering noise when collecting renal ultrasound images of patients. The ultrasound images containing noise will affect the doctor's accurate judgment of the puncture target and increase the risk of puncture complications. The above technology does not have a solution for improving the quality of ultrasound images and it is difficult to ensure the accuracy of puncture.

[0007] There are individual differences in the renal anatomical structures of different patients, and different imaging angles and depths of ultrasound will lead to inconsistent image manifestations. The above technology does not consider these variation factors and lacks intelligent auxiliary positioning and navigation functions.

[0008] The existing renal pelvis puncture lacks a standardized and intelligent auxiliary system. To train an artificial intelligence model that can accurately identify the position of the patient's renal pelvis and plan the best puncture path, a large amount of labeled data and clinical verification are required. This method consumes a lot of manpower and material resources and there are problems with data quality and standardization. Summary of the Invention

[0009] In view of the deficiencies of the prior art, the present invention provides an ultrasound-guided renal pelvis puncture method based on deep learning to solve the problems raised in the above background technology.

[0010] To achieve the above object, the present invention provides the following technical solution: An ultrasound-guided renal pelvis puncture method based on deep learning, including the following steps: Step S1: Obtain ultrasound images of the renal pelvis; Step S2: Construct a deep denoising autoencoder; process the ultrasonic image of the renal pelvis in Step S1 through the deep denoising autoencoder to obtain a low-noise ultrasonic image of the renal pelvis; Step S3: Process the low-noise ultrasonic image of the renal pelvis in Step S2 to construct a basic dataset and an augmented dataset; Step S4: Construct a multi-scale deep learning model; input the basic dataset into the multi-scale deep learning model for processing. In the multi-scale deep learning model, the basic dataset first undergoes feature branch separation processing to obtain the features of the low-noise renal pelvis image data, and then the features of the low-noise renal pelvis image data are processed to obtain channel attention features and spatial attention features. The channel attention features and spatial attention features are fused to obtain the fused features of the current layer; Step S5: Construct a deep learning model based on self-supervised learning; the deep learning model based on self-supervised learning includes a multi-scale feature extraction module and a renal pelvis localization module. Process the augmented dataset through the multi-scale feature extraction module to obtain multi-scale features; input the multi-scale features and the fused features of the current layer obtained in Step S4 into the renal pelvis localization module for processing to obtain the pixel coordinates of the renal pelvis in the ultrasonic image; Step S6: Based on the pixel coordinates of the renal pelvis in the ultrasonic image, analyze the ultrasonic image of the renal pelvis to locate the position of the renal pelvis in the ultrasonic image of the renal pelvis.

[0011] Furthermore, for the ultrasonic image of the renal pelvis in Step S1, the specific process is as follows: Place the probe of the ultrasonic instrument on the patient's abdomen, scan horizontally along the intercostal space or the midline area of the patient's abdomen, and gradually move the probe of the ultrasonic instrument to the renal pelvis area. Detect the renal pelvis area at different angles and depths through the probe of the ultrasonic instrument to obtain the ultrasonic image of the renal pelvis.

[0012] Furthermore, for the low-noise ultrasonic image of the renal pelvis in Step S2, the specific process is as follows: Input the ultrasonic image of the renal pelvis into the encoder of the deep denoising autoencoder. After data preprocessing and operations of the first convolutional layer and the second convolutional layer in the encoder, obtain a deep feature map , expressed as: ; where represents the latent space encoding operation; represents the weight matrix in the encoder; represents the bias of the average pooling layer in the encoder; represents the ultrasonic image of the renal pelvis; Input the deep feature map into the decoder of the deep denoising autoencoder. After upsampling and operations of the first transposed convolutional layer and the second transposed convolutional layer, generate a low-noise ultrasonic image of the renal pelvis , expressed as: ; In the formula, represents the upsampling operation; represents the weight matrix in the decoder; represents the bias of the decoder average pooling layer.

[0013] Furthermore, in step S3, the basic dataset and the enhanced dataset are constructed. The specific process is as follows: Define low-noise pyelonephritis ultrasound images Generate a total number of samples of , define the index sequence ; represents the total number of low-noise pyelonephritis ultrasound image samples; represents the set of index sequences of low-noise pyelonephritis ultrasound image samples; randomly and uniformly extract 50% of the elements from the set of index sequences of low-noise pyelonephritis ultrasound image samples , denoted as the set , and the remaining elements form the set ; According to the set , select the corresponding low-noise pyelonephritis ultrasound image samples in the set of index sequences of low-noise pyelonephritis ultrasound image samples to construct the basic dataset ; According to the set , select the corresponding low-noise pyelonephritis ultrasound image samples in the set of index sequences of low-noise pyelonephritis ultrasound image samples to construct the enhanced dataset .

[0014] Furthermore, in step S4, the fused features of the current layer. The specific process is as follows: After preprocessing the low-noise pyelonephritis ultrasound image samples in the basic dataset , obtain the low-noise pyelonephritis ultrasound image after image preprocessing, and then input the low-noise pyelonephritis ultrasound image after image preprocessing into the feature branch separation to obtain the features of the low-noise pyelonephritis image data; Represent the features of the low-noise pyelonephritis image data as ; Divide into a high-resolution feature map , a medium-resolution feature map and a low-resolution feature map in three feature paths, expressed as: ; In the formula, represents the feature branch separation function; The specific steps of the expansion of the feature branch separation function, expressed as: ; ; ; ; In the formula, represents the pre - processed feature map obtained after the convolution operation on the features of the input low - noise renal pelvis image data of ; represents the convolution operation of represents the convolution operation represents the convolution operation represents the convolution operation Input the high - resolution feature map , the medium - resolution feature map and the low - resolution feature map into the variable convolution module for processing. The specific process is as follows: The variable convolution module consists of a first variable convolution module, a second variable convolution module and a third variable convolution module; Input the high - resolution feature map , the medium - resolution feature map and the low - resolution feature map into the first variable convolution module, the second variable convolution module and the third variable convolution module respectively for processing, and obtain the outputs of the first variable convolution module, the second variable convolution module and the third variable convolution module respectively, which are represented as: ; ; ; In the formula, , and respectively represent the values of the high - resolution feature map at the position , the values of the medium - resolution feature map at the position and the values of the low - resolution feature map at the position ; i and j respectively represent the abscissa and ordinate of the resolution feature map; represents the negative value of the convolution kernel radius; represents the weight value at the position in the convolution kernel; and respectively represent the vertical and horizontal positions in the convolution kernel; and represents the spatial offset learned at the position ; represents the modulation factor; represents the convolution kernel radius; Fuse the outputs of the first variable convolution module, the second variable convolution module, and the third variable convolution module through feature concatenation to obtain the output of feature concatenation, and input the output of feature concatenation into the channel attention module and the spatial attention module for processing to obtain the channel attention at the layer feature and the spatial attention at the layer feature ; Use the channel attention module to enhance the output of feature concatenation to obtain the channel attention at the layer feature , expressed as: ; In the formula, represents the activation function; represents the multi-layer perceptron; and respectively represent the average pooling and max pooling operations; represents the output of feature concatenation; is , and 's concatenated features, ; represents feature concatenation; Use the spatial attention module to enhance the output of feature concatenation to obtain the spatial attention at the layer feature , expressed as: ;

[0015] Use the feature pyramid network to fuse the channel attention at the layer feature and the spatial attention at the layer feature to obtain the fused feature of the current layer, expressed as: ; In the formula, represents the fused feature of the current layer; Conv( is the convolution operation; ; represents the fused feature of the current first layer; represents the fused feature of the current layer; Represents the fused feature of the current layer; Represents the fused feature of the +1 layer.

[0016] Furthermore, the pixel coordinates of the renal pelvis in the ultrasonic image in step S5 are as follows: Input the enhanced dataset into a deep learning model based on self-supervised learning. The deep learning model based on self-supervised learning includes a multi-scale feature extraction module and a renal pelvis localization module; train the enhanced dataset through the multi-scale feature extraction module , and then obtain multi-scale features by repeating the feature branch separation step in step S4 above ; The renal pelvis localization module consists of two fully connected layers. Input the multi-scale features and the fused feature of the current layer into the renal pelvis localization module for processing, and output the position of the renal pelvis pixel point , that is, the pixel coordinates of the renal pelvis in the ultrasonic image.

[0017] Compared with the existing technology, the present invention has the following beneficial effects: (1) The present invention significantly improves the quality of kidney ultrasonic images through a deep denoising autoencoder, removes electromagnetic interference and tissue scattering noise, thereby improving the accuracy and safety of puncture path localization.

[0018] (2) By constructing a multi-scale deep learning model, the present invention can effectively extract kidney features at different scales, overcomes the influence of different ultrasonic imaging angles and depths on the quality of renal pelvis ultrasonic images, and significantly improves the accuracy of renal pelvis puncture localization.

[0019] (3) The present invention adopts a self-supervised learning method, and can achieve high-precision renal pelvis localization and puncture path planning based on a relatively small dataset, significantly reducing the costs of data collection and training of the deep denoising autoencoder. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is a flowchart of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0021] As Figure 1 shown, the present invention provides a technical solution: an ultrasonic-guided renal pelvis puncture method based on deep learning, including the following steps: Step S1: Obtain an ultrasonic image of the renal pelvis; Step S2: Construct a deep denoising autoencoder; process the ultrasonic image of the renal pelvis in step S1 through the deep denoising autoencoder to obtain a low-noise renal pelvis ultrasonic image; Step S3: Process the low-noise renal pelvis ultrasound images in Step S2 to construct a basic dataset and an enhanced dataset; Step S4: Construct a multi-scale deep learning model; input the basic dataset into the multi-scale deep learning model for processing. In the multi-scale deep learning model, the basic dataset first undergoes feature branch separation processing to obtain the features of the low-noise renal pelvis image data, and then the features of the low-noise renal pelvis image data are processed to obtain channel attention features and spatial attention features. The channel attention features and spatial attention features are fused to obtain the fusion features of the current layer; Step S5: Construct a deep learning model based on self-supervised learning; the deep learning model based on self-supervised learning includes a multi-scale feature extraction module and a renal pelvis localization module. The enhanced dataset is processed by the multi-scale feature extraction module to obtain multi-scale features; the multi-scale features and the fusion features of the current layer obtained in Step S4 are input into the renal pelvis localization module for processing to obtain the pixel coordinates of the renal pelvis in the ultrasound image; Step S6: Based on the pixel coordinates of the renal pelvis in the ultrasound image, analyze the ultrasound image of the renal pelvis to locate the position of the renal pelvis in the ultrasound image of the renal pelvis.

[0022] Among them, for the ultrasound image of the renal pelvis in Step S1, the specific process is as follows: Place the probe of the ultrasonic instrument on the patient's abdomen, scan horizontally along the intercostal space or the midline area of the abdomen, gradually move the probe of the ultrasonic instrument to the renal pelvis area, and detect the renal pelvis area at different angles and depths through the probe of the ultrasonic instrument to obtain the ultrasound image of the renal pelvis.

[0023] Among them, for the low-noise renal pelvis ultrasound image in Step S2, the specific process is as follows: The ultrasound image of the renal pelvis is vulnerable to interference from factors such as instrument accuracy, patient physiological activities, and the environment during the diagnosis of kidney diseases, resulting in noise. As a result, key details such as the fine internal structure and blood vessel texture of the kidney are blurred, affecting the doctor's accurate judgment of the lesion situation; therefore, a deep denoising autoencoder is constructed to fit the characteristics of the ultrasound image of the renal pelvis, accurately remove the noise, and provide a clear image basis for subsequent diagnosis; The encoder of the deep denoising autoencoder consists of a first convolutional layer and a second convolutional layer, The decoder of the deep denoising autoencoder consists of an upsampling layer, a first deconvolutional layer, and a second deconvolutional layer; Input the ultrasound image of the renal pelvis into the encoder, and obtain a deep feature map through the data preprocessing and the operations of the first convolutional layer and the first convolutional layer , which is expressed as: ; In the formula, represents the latent space encoding operation; denotes the weight matrix in the encoder; denotes the bias of the average pooling layer in the encoder; denotes the ultrasonic image of the renal pelvis; Input the depth feature map into the decoder, and generate a low-noise ultrasonic image of the renal pelvis through upsampling and operations of the first transposed convolutional layer and the second transposed convolutional layer , which is expressed as: ; In the formula, denotes the upsampling operation; denotes the weight matrix in the decoder; denotes the bias of the average pooling layer in the decoder.

[0024] Among them, in step S3, constructing the basic dataset and the augmented dataset, the specific process is as follows: Define the total number of samples for generating the low-noise ultrasonic image of the renal pelvis as , and define the index sequence ; denotes the total number of samples of the low-noise ultrasonic image of the renal pelvis; denotes the set of index sequences of the low-noise ultrasonic image samples of the renal pelvis; randomly select 50% of the elements from the set of index sequences of the low-noise ultrasonic image samples of the renal pelvis, denoted as set , and the remaining elements form set ; ; According to set , select the corresponding low-noise ultrasonic image samples of the renal pelvis from the set of index sequences of the low-noise ultrasonic image samples of the renal pelvis to construct the basic dataset ; According to set , select the corresponding low-noise ultrasonic image samples of the renal pelvis from the set of index sequences of the low-noise ultrasonic image samples of the renal pelvis to construct the augmented dataset ; In the basic dataset , all low-noise ultrasonic image samples of the renal pelvis contain clear manual annotation information to ensure that the deep learning model can learn specific pathological features during the initial training process; in the augmented dataset , process images with different angles, depths, and pathological states through data augmentation techniques. The augmented dataset does not contain manual annotation and is used for semi-supervised learning tasks to increase the generalization ability of the multi-scale deep learning model; the data augmentation techniques include operations such as rotation, translation, mirror inversion, and contrast adjustment to simulate the changes of kidney ultrasonic images obtained in real scenarios under various different conditions.

[0025] Among them, the specific process of the fused feature of the current layer in step S4 is as follows: For the low-noise renal pelvis ultrasound image samples in the basic dataset After image preprocessing, the low-noise renal pelvis ultrasound image after image preprocessing is obtained. Then, the low-noise renal pelvis ultrasound image after image preprocessing is input into the feature branch separation to obtain the features of the low-noise renal pelvis image data; The features of the low-noise renal pelvis image data are represented as ; is divided into a high-resolution feature map , a medium-resolution feature map and a low-resolution feature map Three feature paths, which respectively capture the detailed information, structural information and global semantic information of the low-noise renal pelvis ultrasound image, are represented as: ; In the formula, represents the feature branch separation function; is the core operation of feature branch separation; The specific steps of the expansion of the feature branch separation function are represented as: ; ; ; ; In the formula, represents the preprocessed feature map obtained after the of the features of the input low-noise renal pelvis image data after the convolution operation; represents the convolution operation; represents the convolution operation; represents the convolution operation; represents the convolution operation; The high-resolution feature map , the medium-resolution feature map and the low-resolution feature map are input into the deformable convolution module for processing. The specific process is as follows: The deformable convolution module consists of a first deformable convolution module, a second deformable convolution module and a third deformable convolution module; The high-resolution feature map , the medium-resolution feature map and the low-resolution feature map They are respectively input into the first deformable convolution module, the second deformable convolution module, and the third deformable convolution module for processing, and the outputs of the first deformable convolution module, the second deformable convolution module, and the third deformable convolution module are respectively obtained, expressed as: ; ; ; In the formula, , and respectively represent the values of the high-resolution feature map at the position , the values of the medium-resolution feature map at the position , and the values of the low-resolution feature map at the position ; i and j respectively represent the abscissa and ordinate of the resolution feature map; represents the negative value of the convolution kernel radius; represents the weight value at the position in the convolution kernel; and respectively represent the vertical and horizontal positions in the convolution kernel; and represent the spatially learned offsets at the position ; represents the modulation factor; represents the convolution kernel radius; The outputs of the first deformable convolution module, the second deformable convolution module, and the third deformable convolution module are fused through feature concatenation to obtain the output of feature concatenation. The output of feature concatenation is input into the channel attention module and the spatial attention module for processing to obtain the channel attention feature and the spatial attention feature ; The channel attention module is used to enhance the output of feature concatenation to obtain the channel attention layer feature , expressed as: ; In the formula, represents the activation function; represents the multi-layer perceptron; and respectively represent the average pooling and max pooling operations; represents the input feature map; is , and 's concatenated feature, ; Indicates feature splicing; Due to the uneven spatial distribution of the quality of low-noise renal pelvis ultrasound images, there are differences in the clarity and reliability of different regions of low-noise renal pelvis ultrasound images, and it is necessary to adaptively adjust the attention. Therefore, the present invention uses a spatial attention module to enhance the output of feature splicing to obtain the layer features of spatial attention , which is expressed as: ; Using a Feature Pyramid Network (FPN) to fuse the layer features of channel attention and the layer features of spatial attention to obtain the fused features of the current layer, which is expressed as: ; In the formula, represents the fused features of the current layer; Conv( is a convolution operation used to smooth the fused features; ; represents the fused features of the current first layer; represents the fused features of the current layer; represents the fused features of the current layer; represents the fused features of the +1 layer; In the task of feature extraction of low-noise renal pelvis ultrasound images, since it is necessary to simultaneously ensure the accurate segmentation of the puncture area, the accurate estimation of the path depth, the clear positioning of the tissue boundary, and the prominent display of key anatomical structures, a single loss function is difficult to comprehensively guide the multi-scale deep learning model to learn these complex feature expressions; therefore, a multi-component loss function including a segmentation loss function, a depth estimation loss function, a contour perception loss function, and an attention guidance loss function is designed. Through the synergistic effect and dynamic weight adjustment of each loss term, the multi-scale deep learning model is guided to simultaneously focus on multiple clinical key targets during training, so as to obtain a more comprehensive and robust feature representation, improve the accuracy and safety of the puncture operation, and input the fused features of the current layer into the segmentation loss, depth estimation loss, contour perception loss, and attention guidance loss for processing respectively, and obtain the outputs of the segmentation loss, depth estimation loss, contour perception loss, and attention guidance loss respectively. Add the outputs of the segmentation loss, depth estimation loss, contour perception loss, and attention guidance loss to obtain the total loss function, which is expressed as: ; In the formula, represents the total loss function; represents the output of the segmentation loss; represents the output of the depth estimation loss; represents the output of the contour perception loss; represents the output of the attention guidance loss; , , , respectively represent the output weight parameters of the segmentation loss, the output weight parameters of the depth estimation loss, the output weight parameters of the contour perception loss, and the output weight parameters of the attention guidance loss.

[0026] Among them, the pixel coordinates of the renal pelvis in the ultrasound image in step S5, the specific process is as follows: Input the enhanced dataset into the multi-scale feature extraction module, and train the enhanced dataset through the multi-scale feature extraction module , and then obtain multi-scale features by repeating the feature branch separation step in the above step S4 , providing support for subsequent positioning tasks; expressed as: ; In the formula, represents the multi-scale feature at the position element value; represents the pixel value of the input renal pelvis ultrasound image at the position ; The renal pelvis positioning module consists of two fully connected layers, and the multi-scale feature and the fusion feature of the current layer are input into the renal pelvis positioning module for processing, and the position of the renal pelvis pixel point is output , that is, the pixel coordinates of the renal pelvis in the ultrasound image.

[0027] Among them, the deep learning model based on self-supervised learning includes an image reconstruction task, a contrast learning task, and a context prediction task. Among them, the image reconstruction task is completed through an autoencoder structure, and the input low-noise renal pelvis ultrasound image is reconstructed through the image reconstruction task to obtain the reconstructed input low-noise renal pelvis ultrasound image; the contrast learning task trains the deep learning model based on self-supervised learning to maintain feature consistency by inputting low-noise renal pelvis ultrasound images with different view enhancements, and obtains the trained deep learning model based on self-supervised learning; the context prediction task captures the structural features of the low-noise renal pelvis ultrasound image by predicting the relative positions of local regions of the reconstructed input low-noise renal pelvis ultrasound image; In the fine-tuning stage, based on the structural features of the low-noise renal pelvis ultrasound image, from the enhanced dataset Randomly select 10% of the data from the ; In the formula, represents the self-supervised localization loss function; represents the total number of positions; represents the weight of different pixel positions; represents the position of the -th pixel point in the set of predicted renal pelvis pixel points; represents the position of the -th pixel point in the set of true renal pelvis pixel points; The deep learning model based on self-supervised learning is trained using the self-supervised localization loss function. The self-supervised localization loss function includes an image reconstruction task loss function, a contrastive learning task loss function, and a context prediction task loss function. The self-supervised localization loss function can effectively learn the feature patterns in the low-noise renal pelvis ultrasound images and provide high-quality feature representations for subsequent localization tasks. The loss calculation formula is expressed as: ; In the formula, represents the total loss function; represents the image reconstruction task loss function; represents the contrastive learning task loss function; represents the context prediction task loss function.

[0028] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for ultrasound-guided renal pelvic puncture based on deep learning, characterized in that: The following steps are involved: Step S1: obtaining an ultrasound image of the renal pelvis; Step S2: construct a deep denoising autoencoder; process the ultrasound image of the renal pelvis in step S1 by the deep denoising autoencoder to obtain a low-noise renal pelvis ultrasound image; Step S3: Process the low-noise renal pelvis ultrasound image in step S2 to construct a basic data set and an enhanced data set; Step S4: construct a multi-scale deep learning model; The basic data set is input into the multi-scale deep learning model for processing. The basic data set is first processed by feature branch separation in the multi-scale deep learning model to obtain the features of the low-noise renal pelvis image data. The features of the low-noise renal pelvis image data are then processed to obtain channel attention features and spatial attention features. The channel attention features and the spatial attention features are fused to obtain the fusion features of the current layer. Step S5: constructing a deep learning model based on self-supervised learning; The deep learning model based on self-supervised learning includes a multi-scale feature extraction module and a renal pelvis positioning module. The multi-scale feature extraction module processes the enhanced data set to obtain multi-scale features. The multi-scale features and the fusion features of the current layer obtained in step S4 are input into the renal pelvis positioning module for processing to obtain the pixel coordinates of the renal pelvis in the ultrasound image. Step S6: Analyze the ultrasound image of the renal pelvis based on the pixel coordinates of the renal pelvis in the ultrasound image, and locate the position of the renal pelvis in the ultrasound image of the renal pelvis.

2. The method for ultrasound-guided renal pelvic puncture based on deep learning according to claim 1, characterized in that: The specific process of the ultrasound image of the renal pelvis in step S1 is as follows: Place the ultrasound instrument probe on the patient's abdomen, scan horizontally along the patient's abdominal intercostal space or abdominal midline area, gradually move the ultrasound instrument probe to the renal pelvis area, and use the ultrasound instrument probe to detect different angles and depths of the renal pelvis area to obtain an ultrasound image of the renal pelvis.

3. The method for ultrasound-guided renal pelvic puncture based on deep learning according to claim 2, characterized in that: In step S2, the low-noise renal pelvis ultrasound image is obtained by: The ultrasound image of the renal pelvis is input into the encoder of the deep denoising autoencoder, and the deep feature map is obtained after the encoder's data preprocessing and the first and second convolutional layer operations. , expressed as: ； In the formula, represents the latent space encoding operation; represents the weight matrix in the encoder; represents the bias of the encoder average pooling layer; Ultrasound image showing the renal pelvis; The depth feature map The input is sent to the decoder of the deep denoising autoencoder, and a low-noise renal pelvis ultrasound image is generated after upsampling and the first and second deconvolution layers. , expressed as: ； In the formula, Represents an upsampling operation; represents the weight matrix in the decoder; Represents the bias of the decoder average pooling layer.

4. The method for ultrasound-guided renal pelvic puncture based on deep learning according to claim 3, characterized in that: In step S3, the basic data set and the enhanced data set are constructed. The specific process is as follows: Defining low-noise renal pelvic ultrasound images The total number of samples generated is , define the index sequence ; represents the total number of low-noise renal pelvis ultrasound image samples; Represents a set of index sequences of low-noise renal pelvis ultrasound image samples; Randomly select a set of index sequences of low-noise renal pelvis ultrasound image samples 50% of the elements are uniformly extracted from the set , the remaining elements form the set ; Based on the collection , select the corresponding low-noise renal pelvis ultrasound image samples from the index sequence set of low-noise renal pelvis ultrasound image samples to construct a basic data set ; Based on the collection , select the corresponding low-noise renal pelvis ultrasound image samples from the index sequence set of low-noise renal pelvis ultrasound image samples to construct an enhanced data set .

5. The method for ultrasound-guided renal pelvic puncture based on deep learning according to claim 4, characterized in that: The fusion features of the current layer in step S4 are as follows: The basic dataset After the medium and low noise renal pelvis ultrasound image samples are subjected to image preprocessing, a low noise renal pelvis ultrasound image after image preprocessing is obtained, and then the low noise renal pelvis ultrasound image after image preprocessing is input into the feature branch separation to obtain the features of the low noise renal pelvis image data; The features of low-noise renal pelvis image data are expressed as ;Will Divided into high-resolution feature maps , medium resolution feature map and low-resolution feature maps Three characteristic paths, represented as: ； In the formula, represents the characteristic branch separation function; The specific steps of feature branch separation function expansion are expressed as: ； ； ； ； In the formula, Represents the characteristics of the input low-noise renal pelvis image data of The preprocessed feature map obtained after the convolution operation; express Convolution operation; express Convolution operation; express Convolution operation; express Convolution operation; The high-resolution feature map , medium resolution feature map and low-resolution feature maps The input is processed in the variable convolution module. The specific process is as follows: The variable convolution module is composed of a first variable convolution module, a second variable convolution module and a third variable convolution module; The high-resolution feature map , medium resolution feature map and low-resolution feature maps The first variable convolution module, the second variable convolution module, and the third variable convolution module are respectively input for processing, and the output of the first variable convolution module, the output of the second variable convolution module, and the output of the third variable convolution module are respectively obtained, which are expressed as: ； ； ； In the formula, , and Represent high-resolution feature maps In Location Values ​​at, medium-resolution feature maps In Location The value at and the low-resolution feature map In Location The value at; i and j represent the abscissa and ordinate of the resolution feature map respectively; Represents the negative value of the convolution kernel radius; Represents the position in the convolution kernel The weight value of and Respectively represent the vertical and horizontal positions in the convolution kernel; and Indicates at location The learned spatial offset; represents the modulation factor; Represents the convolution kernel radius; The output of the first variable convolution module, the output of the second variable convolution module, and the output of the third variable convolution module are fused through feature splicing to obtain the output of feature splicing, and the output of feature splicing is input into the channel attention module and the spatial attention module for processing to obtain the channel attention module. Layer Features and spatial attention Layer Features ; The channel attention module is used to enhance the output of feature concatenation and obtain the channel attention Layer Features , expressed as: ； In the formula, represents the activation function; represents a multilayer perceptron; and Represent average pooling and maximum pooling operations respectively; Represents the output of feature concatenation; for , and The splicing features, ; Represents feature splicing; The spatial attention module is used to enhance the output of feature splicing and obtain the spatial attention Layer Features , expressed as: ； Using feature pyramid network to focus on channel Layer Features and spatial attention Layer Features Fusion is performed to obtain the fusion features of the current layer, which is expressed as: ； In the formula, Represents the fusion features of the current layer; Conv( is the convolution operation; ; Represents the fusion features of the current first layer; Indicates the current Layer fusion features; Indicates the current Layer fusion features; Indicates +1 layer of fused features.

6. The method for ultrasound-guided renal pelvic puncture based on deep learning according to claim 5, characterized in that: The pixel coordinates of the renal pelvis in the ultrasound image in step S5 are specifically obtained as follows: The dataset will be enhanced The input is a deep learning model based on self-supervised learning, which includes a multi-scale feature extraction module and a renal pelvis positioning module; the enhanced dataset is trained by the multi-scale feature extraction module , and then repeat the feature branch separation step in step S4 to obtain the multi-scale feature ; The renal pelvis localization module consists of two fully connected layers, which integrate the multi-scale features And the fusion features of the current layer Input to the renal pelvis positioning module for processing, output the location of the renal pelvis pixel , which is the pixel coordinate of the renal pelvis in the ultrasound image.

Images