Kidney stone positioning method and system based on deep learning

Abstract

The invention discloses a kidney stone positioning method and system based on deep learning, and relates to the technical field of image processing. The method comprises the following steps: receiving a gray-scale image of B ultrasound in a target kidney region as a target image; taking the target image as input of a target recognition model to obtain a recognition result of the target image; the identification result is the size and the position of the stone; mapping to the size and position of the calculus in the target kidney region according to the identification result; according to the method, the accuracy of detecting the shape of the kidney stone and positioning the kidney stone is improved through the target recognition model, the calculation cost is reduced, then the detection frequency in an operation is reduced, and reliable technical support is provided for the intelligent extracorporeal lithotripter.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to a method and system for locating kidney stones based on deep learning. Background Technology

[0002] Image processing systems in the modern visual information chain are responsible for redistributing the "energy" of the raw pixel stream acquired from imaging devices, ensuring that this visual data can be efficiently "utilized" by downstream tasks at all levels. This involves progressively reducing the high-dimensional, highly redundant raw images to the most application-friendly abstraction level (detection, segmentation, recognition, localization), and weaving this into a multi-layered, multi-scale feature network for the terminal algorithm. This provides stable, reliable, and high-precision localization for kidney stones in medical diagnosis.

[0003] The prior art (publication number: CN118762005A) discloses a method and system for detecting kidney stones based on CT slices. The method involves acquiring and preprocessing preoperative CT images; acquiring and analyzing CT slices of the kidney obtained from a CT scanning device; acquiring all regions containing kidney stones in the CT slices using a three-dimensional region growing method; filtering the regions containing kidney stones based on the volume constraints of the kidney stones; obtaining a first mask for the kidney region based on two-dimensional region growing and tracking of the CT slices, and obtaining a second mask for the kidney region based on a deep learning model; and filtering out kidney stones from the regions containing kidney stones based on the first and second masks to obtain the CT coordinates of the kidney stones.

[0004] The aforementioned patent uses a deep learning model to locate kidney stones. However, in practical applications, this model still has accuracy deficiencies in determining the shape, size, and location of kidney stones, requiring multiple determinations of the kidney stone's location during surgery, thereby increasing the overall cost. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of the current model in terms of accuracy in determining the shape, size, and location of kidney stones, which necessitates multiple determinations of the kidney stone's location during surgery, thereby increasing the overall cost. Therefore, this invention proposes a kidney stone localization method and system based on deep learning.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] First, a deep learning-based method for locating kidney stones is proposed, the method comprising:

[0008] Receive the grayscale image of the target kidney region from the ultrasound as the target image;

[0009] The target image is used as input to the target recognition model to obtain the recognition result of the target image; the recognition result is the size and location of the stone.

[0010] The size and location of the stones are mapped to the target kidney region based on the identification results;

[0011] The target recognition model is an improvement upon the YOLOv8 model, specifically including:

[0012] In the main structure, the C2f module is replaced with the target modulation module;

[0013] In the neck structure, the Upsample module is replaced with the target dynamic upsample module;

[0014] Between the skip connections of the trunk and neck structures, a target image enhancement module is added after the SPPF module.

[0015] Optionally, the principle process of the target modulation module includes:

[0016] The input features are determined as the input to the target modulation module;

[0017] The input features are normalized to obtain the first feature;

[0018] The second feature is convolved with the first feature;

[0019] Substituting the second feature into the target activation function yields the third feature;

[0020] A fourth feature is obtained by performing a fully connected operation on the input feature map;

[0021] The fifth feature is obtained by multiplying the third feature and the fourth feature;

[0022] The target feature is obtained by performing a fully connected operation on the input feature and the fifth feature.

[0023] Optionally, the principle process of the target dynamic upsampling module includes:

[0024] The input feature map is used as the input to the target dynamic upsampling module;

[0025] The input feature map is sampled to obtain sampled features;

[0026] Linear projection of the input feature map yields a first linear feature and a second linear feature;

[0027] The target linear feature is obtained by dynamically adjusting the first and second linear features.

[0028] The sampled features are added to the target linear features to obtain the target upsampled feature map;

[0029] The specific formula for the dynamic adjustment operation is expressed as follows:

[0030]

[0031] Where T represents the target linear feature, X1 and X2 represent the first linear feature and the second linear feature, respectively, sigmoid() represents the activation function, and linear() represents the linear function.

[0032] Optionally, the principle process of the target image enhancement module includes:

[0033] The input feature tensor is used as the input to the target image enhancement module;

[0034] The input feature tensor is subjected to width-average pooling and height-average pooling operations respectively to obtain width-pooled feature tensor and height-average pooling tensor;

[0035] The width pooling feature tensor and the height average pooling tensor are segmented to obtain each width segment tensor and each height segment tensor.

[0036] Perform depthwise convolution operations on each width segmentation tensor and each height segmentation tensor to obtain each width convolution tensor and each height convolution tensor;

[0037] A fusion operation is performed on the convolutional tensors of each width to obtain a width-fused tensor;

[0038] A fusion operation is performed on the aforementioned high-convolution tensors to obtain a high-fusion tensor;

[0039] After normalizing the width fusion tensor and the height fusion tensor respectively, element-wise multiplication is performed to obtain the first feature tensor;

[0040] The second feature tensor is obtained by performing average pooling on the first feature tensor.

[0041] Normalize the second feature tensor to obtain the third feature tensor;

[0042] Substituting the third feature tensor into a deep convolutional layer yields the query feature tensor, the key feature tensor, and the value feature tensor, respectively.

[0043] Substituting the query feature tensor, key feature tensor, and value feature tensor into the one-way self-attention mechanism yields the fourth feature tensor;

[0044] The target feature tensor is obtained by multiplying the first feature tensor and the fourth feature tensor.

[0045] Optionally, during the training process, the target recognition model uses a loss function as the regression loss for the bounding box of the detection head, and the calculation formula of the loss function includes:

[0046]

[0047] in, This represents the total loss value. This represents the basic intersection-union ratio loss value. This represents the distance focusing value, and d represents the non-monotonic focusing coefficient value. γ represents the average loss value of the crossover ratio. and The ratio, (x, y) represents the coordinates of the center point of the predicted bounding box, (x... g ,y g W represents the coordinates of the center point of the true bounding box. g and H g Wa and Ha are the width and height of the ground truth bounding box, respectively. e and π represent the natural constants, v represents the aspect ratio difference measure, and α represents the weighting coefficient, with a value range of (0,1).

[0048] A deep learning-based kidney stone localization system is proposed, comprising:

[0049] Image acquisition module: Receives a grayscale image of the target kidney region from ultrasound as the target image;

[0050] Target recognition module: Takes the target image as input to the target recognition model and obtains the recognition result of the target image; the recognition result is the size and location of the stone; the size and location of the stone in the target kidney region are mapped according to the recognition result;

[0051] Model Improvement Module: The target recognition model is an improvement upon the YOLOv8 model, specifically including:

[0052] In the main structure, the C2f module is replaced with the target modulation module; in the neck structure, the Upsample module is replaced with the target dynamic upsampling module; and between the skip connections in the main structure and the neck structure, a target image enhancement module is added after the SPPF module.

[0053] Optionally, the model improvement module is also used in the principle process of the target modulation module, which includes:

[0054] The input features are determined as the input to the target modulation module;

[0055] The input features are normalized to obtain the first feature;

[0056] The second feature is convolved with the first feature;

[0057] Substituting the second feature into the target activation function yields the third feature;

[0058] A fourth feature is obtained by performing a fully connected operation on the input feature map;

[0059] The fifth feature is obtained by multiplying the third feature and the fourth feature;

[0060] The target feature is obtained by performing a fully connected operation on the input feature and the fifth feature.

[0061] Optionally, the model improvement module includes: an input determination module, a sampling module, a linearization module, a dynamic adjustment module, and a feature addition module.

[0062] The input determination module is used to determine the input feature map as the input to the target dynamic upsampling module;

[0063] The sampling module is used to sample the input feature map to obtain sampled features;

[0064] The linear module is used to perform linear projection on the input feature map to obtain a first linear feature and a second linear feature;

[0065] The dynamic adjustment module is used to perform dynamic adjustment operations on the first linear feature and the second linear feature to obtain the target linear feature.

[0066] The feature addition module is used to add the sampled features to the target linear features to obtain the target upsampled feature map;

[0067] The specific formula for the dynamic adjustment operation is expressed as follows:

[0068]

[0069] Where T represents the target linear feature, X1 and X2 represent the first linear feature and the second linear feature, respectively, sigmoid() represents the activation function, and linear() represents the linear function.

[0070] Optionally, the model improvement module also includes the following principles in the target image enhancement module:

[0071] The input feature tensor is used as the input to the target image enhancement module;

[0072] The input feature tensor is subjected to width-average pooling and height-average pooling operations respectively to obtain width-pooled feature tensor and height-average pooling tensor;

[0073] The width pooling feature tensor and the height average pooling tensor are segmented to obtain each width segment tensor and each height segment tensor.

[0074] Perform depthwise convolution operations on each width segmentation tensor and each height segmentation tensor to obtain each width convolution tensor and each height convolution tensor;

[0075] A fusion operation is performed on the convolutional tensors of each width to obtain a width-fused tensor;

[0076] A fusion operation is performed on the aforementioned high-convolution tensors to obtain a high-fusion tensor;

[0077] After normalizing the width fusion tensor and the height fusion tensor respectively, element-wise multiplication is performed to obtain the first feature tensor;

[0078] The second feature tensor is obtained by performing average pooling on the first feature tensor.

[0079] Normalize the second feature tensor to obtain the third feature tensor;

[0080] Substituting the third feature tensor into a deep convolutional layer yields the query feature tensor, the key feature tensor, and the value feature tensor, respectively.

[0081] Substituting the query feature tensor, key feature tensor, and value feature tensor into the one-way self-attention mechanism yields the fourth feature tensor;

[0082] The target feature tensor is obtained by multiplying the first feature tensor and the fourth feature tensor.

[0083] Optionally, the target recognition module is further configured to use a loss function as the regression loss of the bounding box of the detection head during the training process of the target recognition model, wherein the calculation formula of the loss function includes:

[0084]

[0085] in, This represents the total loss value. This represents the basic intersection-union ratio loss value. This represents the distance focusing value, and d represents the non-monotonic focusing coefficient value. γ represents the average loss value of the crossover ratio. and The ratio, (x, y) represents the coordinates of the center point of the predicted bounding box, (x... g ,y g W represents the coordinates of the center point of the true bounding box. g and H g Wa and Ha are the width and height of the ground truth bounding box, respectively. e and π represent the natural constants, v represents the aspect ratio difference measure, and α represents the weighting coefficient, with a value range of (0,1).

[0086] The beneficial effects of this invention are:

[0087] This invention proposes a deep learning-based method for locating kidney stones. A target grayscale image of the target kidney region is acquired via ultrasound and designated as the target image. This target image is then used as input to a target recognition model to obtain the recognition result. The target image can be any image from various datasets. The recognition result indicates the size and location of the stone. The recognition result is sent to medical personnel for processing. The target recognition model is an improvement upon the YOLOv8 model. Specific improvements include: replacing the C2f module in the main renal structure with a target modulation module; replacing the Upsample module in the neck structure with a target dynamic upsampling module; and adding a target image enhancement module after the SPPF module between the skip connections in the main and neck structures. This invention improves the accuracy of detecting the shape and location of kidney stones through the target recognition model, reduces computational costs, and consequently reduces the frequency of intraoperative detection, providing reliable technical support for intelligent extracorporeal shock wave lithotripsy (ESWL). Attached Figure Description

[0088] Figure 1 A flowchart illustrating a deep learning-based method for locating kidney stones, provided as an embodiment of the present invention;

[0089] Figure 2 A network architecture diagram of a target recognition model provided in an embodiment of the present invention;

[0090] Figure 3 This is a flowchart illustrating the principle process of a target modulation module provided in an embodiment of the present invention.

[0091] Figure 4 This is a framework diagram of a deep learning-based kidney stone localization system provided in an embodiment of the present invention. Detailed Implementation

[0092] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0093] This invention provides a deep learning-based method for locating kidney stones. See also... Figure 1 , Figure 1 A flowchart illustrating a deep learning-based method for locating kidney stones, provided as an embodiment of the present invention. The method includes the following steps:

[0094] Receive the grayscale image of the target kidney region from the ultrasound as the target image;

[0095] The target image is used as input to the target recognition model to obtain the recognition result of the target image; the recognition result is the size and location of the stone.

[0096] The size and location of the stones are mapped to the target kidney region based on the identification results;

[0097] See Figure 2 , Figure 2 This invention provides a network architecture diagram for a target recognition model, which is an improvement upon the YOLOv8 model. Specific improvements include:

[0098] In the main structure, the C2f module is replaced with the target modulation module;

[0099] In the neck structure, the Upsample module is replaced with the target dynamic upsample module;

[0100] Between the skip connections of the trunk and neck structures, a target image enhancement module is added after the SPPF module.

[0101] The kidney stone localization method based on deep learning provided in this invention improves the accuracy of detecting the shape and location of kidney stones through a target recognition model, reduces computational costs, and thus reduces the frequency of intraoperative detection, providing reliable technical support for intelligent extracorporeal lithotripsy machines.

[0102] In one implementation, see [link to implementation details]. Figure 3 , Figure 3 A flowchart illustrating the principle process of a target modulation module provided in this embodiment of the invention. The principle process of the target modulation module includes:

[0103] The input features are determined as the input to the target modulation module;

[0104] The first feature is obtained by normalizing the input features;

[0105] The second feature is convolved with the first feature;

[0106] Substituting the second feature into the target activation function yields the third feature;

[0107] A fourth feature is obtained by performing a fully connected operation on the input feature map;

[0108] The fifth feature is obtained by multiplying the third and fourth features;

[0109] The target feature is obtained by performing a fully connected operation on the input feature and the fifth feature.

[0110] In one implementation, the high-efficiency modulation module, through a parallel dual-branch design, adaptively aggregates the multi-scale morphological context of the renal pelvis and calyces while maintaining the high-resolution texture of CT. The fully connected remodulation mechanism fuses the original features and weighted responses a second time, enhancing the difference in CT values. Even if the stone and bone density overlap, the model can capture complex details more accurately.

[0111] In one implementation, the principle and process of the target dynamic upsampling module include:

[0112] The input feature map is used as the input to the target dynamic upsampling module;

[0113] The input feature map is sampled to obtain the sampled features;

[0114] Linear projection of the input feature map yields the first linear feature and the second linear feature;

[0115] The target linear feature is obtained by dynamically adjusting the first and second linear features.

[0116] The sampled features are added to the target linear features to obtain the target upsampled feature map;

[0117] The specific formula for dynamic adjustment is expressed as follows:

[0118]

[0119] Where T represents the target linear feature, X1 and X2 represent the first linear feature and the second linear feature, respectively, sigmoid() represents the activation function, and linear() represents the linear function.

[0120] In one implementation, during the kidney stone localization task, the target dynamic upsampling module adaptively fuses high- and low-level semantics to significantly improve the contrast between the stone edge and the background. The sampled features retain high-resolution texture, and the gating method reweights the two-way linear projection, automatically suppressing artifact channels such as the renal pelvis and blood vessels, enhancing the stone-specific response, improving the model's ability to manage spatial details, and increasing the overall feature map resolution.

[0121] In one implementation, the principle process of the target image enhancement module includes:

[0122] The input feature tensor is used as the input to the target image enhancement module;

[0123] The input feature tensor is subjected to width-average pooling and height-average pooling operations respectively to obtain the width-pooled feature tensor and the height-average pooling tensor.

[0124] The width-pooling feature tensor and the height-average pooling tensor are segmented to obtain the width segmentation tensor and the height segmentation tensor respectively.

[0125] Perform depthwise convolution operations on each width segment tensor and each height segment tensor to obtain each width convolution tensor and each height convolution tensor;

[0126] A fusion operation is performed on the convolution tensors of each width to obtain the width-fused tensor;

[0127] A fusion operation is performed on the convolution tensors of different heights to obtain a highly fused tensor;

[0128] After normalizing the width fusion tensor and the height fusion tensor respectively, element-wise multiplication is performed to obtain the first feature tensor;

[0129] The second feature tensor is obtained by performing average pooling on the first feature tensor.

[0130] Normalize the second feature tensor to obtain the third feature tensor;

[0131] Substituting the third feature tensor into the deep convolutional layer yields the query feature tensor, key feature tensor, and value feature tensor, respectively.

[0132] Substituting the query feature tensor, key feature tensor, and value feature tensor into the one-way self-attention mechanism yields the fourth feature tensor;

[0133] The target feature tensor is obtained by multiplying the first feature tensor and the fourth feature tensor.

[0134] In one implementation, in CT images of kidney stones, the target image enhancement module precisely enhances stone features through a three-level strategy: width-height separation pooling, depth convolution, and unidirectional self-attention. First, independent pooling and segmentation along the W and H axes capture the anisotropic edges of the stone and the tubular background differences. Depth convolution suppresses noise from the renal pelvis and vascular texture within a 5×5 local receptive field. The width / height fusion tensor is normalized and multiplied element-wise to generate an attention map. Subsequently, unidirectional self-attention establishes long-range dependencies along the channel dimension, automatically calibrating the weights of low-contrast layers. The second and third feature tensors are progressively compressed and activated, ultimately outputting the target feature tensor. This three-level strategy improves feature representation capabilities and enhances the performance in detecting kidney stones.

[0135] In one implementation, during the training process, the target recognition model uses a loss function as the regression loss for the bounding boxes of the detection head. The formula for calculating the loss function includes:

[0136]

[0137] in, This represents the total loss value. This represents the basic intersection-union ratio loss value. This represents the distance focusing value, and d represents the non-monotonic focusing coefficient value. γ represents the average loss value of the crossover ratio. and The ratio, (x, y) represents the coordinates of the center point of the predicted bounding box, (x... g ,y g W represents the coordinates of the center point of the true bounding box. g and H g Wa and Ha are the width and height of the ground truth bounding box, respectively. e and π represent the natural constants, v represents the aspect ratio difference measure, and α represents the weighting coefficient, with a value range of (0,1).

[0138] Based on the same inventive concept, this invention also provides a deep learning-based kidney stone localization system. See also Figure 4 , Figure 4 A framework diagram of a deep learning-based kidney stone localization system provided in this embodiment of the invention includes:

[0139] Image acquisition module: Receives a grayscale image of the target kidney region from ultrasound as the target image;

[0140] Target recognition module: Takes the target image as input to the target recognition model and obtains the recognition result of the target image; the recognition result is the size and location of the stone; the size and location of the stone in the target kidney region are mapped according to the recognition result;

[0141] Model Improvement Module: The target recognition model is an improvement upon the YOLOv8 model. Specific improvements include:

[0142] In the main structure, the C2f module is replaced with the target modulation module; in the neck structure, the Upsample module is replaced with the target dynamic upsampling module; and between the skip connections in the main structure and the neck structure, a target image enhancement module is added after the SPPF module.

[0143] The kidney stone localization system based on deep learning provided in this invention improves the accuracy of detecting the shape and location of kidney stones through a target recognition model, reduces computational costs, and thus reduces the frequency of intraoperative detection, providing reliable technical support for intelligent extracorporeal shock wave lithotripsy machines.

[0144] It should be noted that, in this document, terms such as “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0145] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

Claims

1. A deep learning-based kidney stone localization method, characterized in that, The method comprises: receiving a gray image of an ultrasound in a target kidney region as a target image; inputting the target image into a target recognition model to obtain a recognition result of the target image; the recognition result is the size and position of the stone; mapping the size and position of the stone in the target kidney region according to the recognition result; the target recognition model is improved based on the YOLOv8 model, and the specific improvements include: replacing the C2f module in the backbone structure with a target modulation module; replacing the Upsample module in the neck structure with a target dynamic up-sampling module; adding a target image enhancement module after the SPPF module between the jump connection of the backbone structure and the neck structure.

2. The deep learning-based kidney stone localization method of claim 1, wherein, The principle process of the target modulation module includes: determining the input feature as the input of the target modulation module; normalizing the input feature to obtain a first feature; performing convolution operation on the first feature to obtain a second feature; substituting the second feature into a target activation function to obtain a third feature; performing full connection operation on the input feature map to obtain a fourth feature; multiplying the third feature and the fourth feature to obtain a fifth feature; performing full connection operation on the input feature and the fifth feature to obtain a target feature.

3. The deep learning-based kidney stone localization method of claim 1, wherein, The principle process of the target dynamic up-sampling module includes: determining the input feature map as the input of the target dynamic up-sampling module; sampling the input feature map to obtain a sampling feature; performing linear projection on the input feature map to obtain a first linear feature and a second linear feature; performing dynamic adjustment operation on the first linear feature and the second linear feature to obtain a target linear feature; adding the sampling feature and the target linear feature to obtain a target up-sampling feature map; The specific formula of the dynamic adjustment operation is: wherein, T represents the target linear feature, X1 and X2 represent the first linear feature and the second linear feature respectively, sigmoid() represents an activation function, and linear() represents a linear function.

4. The deep learning-based kidney stone localization method of claim 1, wherein, The principle process of the target image enhancement module includes: determining the input feature tensor as the input of the target image enhancement module; performing width average pooling operation and height average pooling operation on the input feature tensor to obtain a width pooling feature tensor and a height average pooling tensor; performing segmentation operation on the width pooling feature tensor and the height average pooling tensor to obtain each width segmentation tensor and each height segmentation tensor; performing deep convolution operation on each width segmentation tensor and each height segmentation tensor to obtain each width convolution tensor and each height convolution tensor; performing fusion operation on each width convolution tensor to obtain a width fusion tensor; performing fusion operation on each height convolution tensor to obtain a height fusion tensor; performing normalization on the width fusion tensor and the height fusion tensor respectively, and then performing element-level multiplication to obtain a first feature tensor; performing average pooling operation on the first feature tensor to obtain a second feature tensor; performing normalization operation on the second feature tensor to obtain a third feature tensor; substituting the third feature tensor into a deep convolution layer to obtain a query feature tensor, a key feature tensor and a value feature tensor respectively; The query feature tensor, the key feature tensor and the value feature tensor are substituted into the one-way self-attention mechanism to obtain a fourth feature tensor; The first feature tensor and the fourth feature tensor are multiplied to obtain a target feature tensor.

5. The deep learning-based kidney stone localization method of claim 1, wherein, In the training process of the target recognition model, a loss function is used as a regression loss of a bounding box of a detection head, and a calculation formula of the loss function comprises: wherein, represents the total loss value, represents the base IoU loss value, represents the distance focusing value, d represents the non-monotonic focusing coefficient value, represents the IoU average loss value, γ represents and the ratio of the two, (x, y) represents the predicted box center point coordinates, (x g ,y g ) represents the real box center point coordinates, W g and H g are the width and height of the real box, Wa and Ha are the width and height of the predicted box, e and π represent natural constants, v represents the width-height ratio difference measure value, and α represents a weight coefficient, with a value range of (0, 1).

6. A deep learning based kidney stone localization system, comprising: The system comprises: An image acquisition module: receiving a gray image of a target kidney region by B-ultrasound as a target image; A target recognition module: taking the target image as an input of a target recognition model to obtain a recognition result of the target image; the recognition result is the size and position of the stone; and the size and position of the stone in the target kidney region are mapped according to the recognition result; A model improvement module: the target recognition model is obtained by improving a YOLOv8 model, and the specific improvement comprises: In the main structure, the C2f module is replaced by a target modulation module; in the neck structure, the Upsample module is replaced by a target dynamic upsampling module; and between the jump connection of the main structure and the neck structure, a target image enhancement module is added after the SPPF module.

7. The deep learning-based kidney stone localization system of claim 6, wherein, The principle process of the target modulation module comprises: determining an input feature as an input of the target modulation module; performing a normalization operation on the input feature to obtain a first feature; performing a convolution operation on the first feature to obtain a second feature; substituting the second feature into a target activation function to obtain a third feature; performing a full connection operation on the input feature map to obtain a fourth feature; multiplying the third feature and the fourth feature to obtain a fifth feature; performing a full connection operation on the input feature and the fifth feature to obtain a target feature.

8. The deep learning-based kidney stone localization system of claim 6, wherein, The model improvement module comprises an input determination module, a sampling module, a linear module, a dynamic adjustment module and a feature addition module: The input determination module is configured to determine an input feature map as an input of the target dynamic upsampling module; The sampling module is configured to sample the input feature map to obtain a sampling feature; The linear module is configured to perform linear projection on the input feature map to obtain a first linear feature and a second linear feature; The dynamic adjustment module is configured to perform a dynamic adjustment operation on the first linear feature and the second linear feature to obtain a target linear feature; The feature addition module is configured to add the sampling feature and the target linear feature to obtain a target upsampled feature map; The specific formula of the dynamic adjustment operation is: wherein, T represents the target linear feature, X1 and X2 represent the first linear feature and the second linear feature respectively, sigmoid( ) represents an activation function, and linear( ) represents a linear function.

9. The deep learning-based kidney stone localization system of claim 6, wherein, The principle process of the target image enhancement module comprises: determining an input feature tensor as an input of the target image enhancement module; performing a width average pooling operation and a height average pooling operation on the input feature tensor to obtain a width pooling feature tensor and a height average pooling tensor; performing a segmentation operation on the width pooling feature tensor and the height average pooling tensor to obtain each width segmentation tensor and each height segmentation tensor; performing a depth convolution operation on the width segmentation tensors and the height segmentation tensors to obtain width convolution tensors and height convolution tensors; performing a fusion operation on the width convolution tensors to obtain a width fusion tensor; performing a fusion operation on the height convolution tensors to obtain a height fusion tensor; performing element-level multiplication on the width fusion tensor and the height fusion tensor after normalization to obtain a first feature tensor; performing an average pooling operation on the first feature tensor to obtain a second feature tensor; performing a normalization operation on the second feature tensor to obtain a third feature tensor; substituting the third feature tensor into a depth convolution layer to obtain a query feature tensor, a key feature tensor, and a value feature tensor; substituting the query feature tensor, the key feature tensor, and the value feature tensor into a one-way self-attention mechanism to obtain a fourth feature tensor; multiplying the first feature tensor and the fourth feature tensor to obtain a target feature tensor.

10. The deep learning based kidney stone localization system of claim 6, wherein, The target recognition module is further configured to, in a training process of the target recognition model, use a loss function as a regression loss of a bounding box of a detection head, and a calculation formula of the loss function includes: ​ wherein, represents the total loss value, represents the base IoU loss value, represents the distance focusing value, d represents the non-monotonic focusing coefficient value, represents the IoU average loss value, γ represents and the ratio value, (x, y) represents the predicted box center point coordinates, (x g ,y g ) represents the real box center point coordinates, W g and H g are the width and height of the real box, Wa and Ha are the width and height of the predicted box, e and π represent natural constants, v represents the width-height ratio difference measure value, and α represents a weight coefficient, with a value range of (0, 1).

Images