Non-enhanced CT kidney stone automatic detection method based on multi-scale feature fusion

Abstract

The invention discloses a non-enhanced CT kidney stone automatic detection method based on multi-scale feature fusion, and relates to the technical field of clinical image diagnos.The method includes the steps that non-enhanced CT coronal position kidney image data are collected, de-identification processing is conducted on the collected non-enhanced CT coronal position kidney image data, and the non-enhanced CT coronal position kidney image data are obtained; processing the de-identified non-enhanced CT coronary position kidney image data through a data standardization and image enhancement method, extracting key image features by adopting a convolutional neural network, and establishing a kidney stone automatic detection model based on the extracted key image features; and finally, the final automatic kidney stone detection model is deployed to a high-performance workstation with GPU (Graphic Processing Unit) acceleration capability or is integrated with the existing PACS (Picture Archiving and Communication System) of a medical institution, so that a clinical automatic detection function is realized, automatic and efficient detection of kidney stones is realized, and the detection efficiency is improved. And powerful technical support is provided for clinical diagnosis.

Description

Technical Field

[0001] This invention relates to the field of clinical imaging diagnostic technology, specifically to an automated non-enhanced CT method for detecting kidney stones based on multi-scale feature fusion. Background Technology

[0002] Kidney stones are a common urinary system disease with a high recurrence rate. If they are not identified and intervened in a timely and accurate manner, they can easily lead to serious complications such as obstructive hydronephrosis and kidney damage, which will significantly burden patients' quality of life and the medical system.

[0003] In actual clinical imaging, kidney stone detection still faces multiple challenges: on the one hand, small stones are small and have indistinct boundaries, making them easy to be confused with adjacent tissues, leading to missed or false detections; on the other hand, the "domain shift" problem caused by differences in populations, imaging equipment, and scanning protocols across different regions is prominent, and the model is prone to performance degradation when applied across centers. In addition, anatomical distortions and density changes caused by factors such as upper urinary tract obstruction can also interfere with the accurate identification of stones. Based on this, the field of automated kidney stone detection urgently needs a technical solution that can: simultaneously capture fine-grained boundaries and global context at multiple scales, have stronger generalization ability to differences in equipment and populations, and take into account the impact of clinical noise and structural changes. Summary of the Invention

[0004] The purpose of this invention is to provide an automatic detection method for kidney stones using non-enhanced CT based on multi-scale feature fusion, which solves the problems existing in the background technology.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: The present invention provides an automatic detection method for kidney stones using non-enhanced CT based on multi-scale feature fusion, specifically including the following steps: S1. Acquire non-contrast CT coronal kidney image data, and perform de-identification processing on the acquired non-contrast CT coronal kidney image data to obtain de-identified non-contrast CT coronal kidney image data. S2, the de-identified non-enhanced CT coronal kidney image data is processed through data standardization and image enhancement methods to obtain the processed non-enhanced CT coronal kidney image data; S3, based on the processed non-enhanced CT coronal kidney image data, uses a convolutional neural network to extract key image features; S4. An automatic kidney stone detection model is established based on the extracted key image features, and the final automatic kidney stone detection model is obtained through a multi-dimensional training optimization strategy. S5 can deploy the final automated kidney stone detection model to a high-performance workstation with GPU acceleration capabilities, or integrate it with existing PACS systems in medical institutions to achieve automated clinical detection.

[0006] Preferably, the acquisition of non-contrast CT coronal renal image data and the de-identification processing of the acquired non-contrast CT coronal renal image data to obtain de-identified non-contrast CT coronal renal image data include the following steps: Non-contrast CT coronal renal imaging data were collected from different imaging devices and populations based on clinical standards and multi-center acquisition mode; The acquired non-contrast CT coronal kidney image data were de-identified to remove all privacy information that could be associated with the patient's identity, resulting in de-identified non-contrast CT coronal kidney image data.

[0007] Preferably, the process of processing the de-identified non-contrast CT coronal kidney image data using data standardization and image enhancement methods to obtain processed non-contrast CT coronal kidney image data includes the following steps: Pixel resampling was performed on the de-identified non-contrast CT coronal kidney images to unify the spatial resolution of all images to 1 mm³. Based on the resampled non-contrast CT coronal kidney image data, the size of the non-contrast CT coronal kidney image data matrix was scaled to 224×224 pixels using bilinear interpolation, and then processed using the Z-Score normalization method to obtain standardized non-contrast CT coronal kidney image data. The Z-Score standardization formula is shown below: ; in The pixel values ​​of the standardized non-contrast CT coronal kidney image. These are the pixel values ​​of a non-contrast CT coronal kidney image after scaling. Do not specify the mean and standard deviation of all pixel values ​​for non-contrast CT coronal kidney images; Edge enhancement processing was performed on the standardized non-contrast CT coronal kidney image data, and high-frequency edge information of the non-contrast CT coronal kidney image data was extracted by the Laplacian operator. The processed non-contrast CT coronal kidney image data was then linearly superimposed on the original image.

[0008] Preferably, the extraction of key image features based on processed non-enhanced CT coronal kidney image data using a convolutional neural network includes the following steps: S31 employs ResNet, a residual-based convolutional neural network, as the backbone. Processed unenhanced CT coronal kidney images are input into the convolutional neural network, and multi-scale basic features from shallow to deep layers are extracted through layer-by-layer convolution and pooling operations. The extracted features are then integrated into a feature set. These correspond to features from shallow to deep, respectively. F1 contains rich edge and texture details, while F4 contains high-level semantic information. The feature extraction function expression is as follows: in For the first time after passing through the ResNet network Features extracted at each level For the ResNet network Feature extraction functions at each level The input is a processed non-contrast CT coronal image of the kidney; S32 introduces a void space pyramid pooling (ASPP) module to the deep feature F4, using different void ratios... Parallel convolution and global pooling are used to capture multi-scale global contextual information and output features. , ASPP output characteristics The expression is as follows: in The features are those extracted by the void space pyramid pooling module. For feature fusion function, For deep features of the input A function that performs a 1×1 convolution operation. These are the deep features of the input. Execution expansion rate The functions are 3×3 dilated convolution operations of 6, 12, and 18 respectively, used to extract features at different scales. For deep features of the input Perform global feature extraction operations; S33 introduces a pyramid pooling (PPM) module to the mid-layer feature F3, enhancing the local structural representation through multi-scale adaptive pooling and outputting features. Furthermore, a channel attention mechanism (SE Block) is introduced, which performs global average pooling on each feature map to obtain channel description vectors, and then passes them through a fully connected layer and a Sigmoid activation function to obtain channel weights; PPM output characteristics The expression is as follows: The SEBlock expression is shown below: ; in The features output after processing by the Pyramid Pooling Module (PPM) The mid-level features input into the PPM module. This is a function that processes pooling features. For upsampling operation, To perform an adaptive pooling operation with a scale of k×k on the input features, This represents the global average pooling result for the c-th channel. These represent the height and width of the feature map, respectively. Let (i,j) be the feature value of the c-th channel feature map at position (i,j). For channel weight vectors, Sigmoid activation function This is the weight matrix of the fully connected layer. For all The channel description vector is composed of This is the feature map of the c-th channel after attention weighting; S34 utilizes a feature pyramid network to perform top-down multi-level feature fusion, achieving alignment between high-level semantics and low-level details to obtain the final fused features.

[0009] Preferably, the step of establishing an automatic kidney stone detection model based on extracted key image features and obtaining the final automatic kidney stone detection model through a multi-dimensional training optimization strategy includes the following steps: The final fused features are divided into two parts: a training set and a validation set. Based on the final fusion features of the training set, a label smoothing strategy is used to optimize the cross-entropy loss function for processing, and a cosine annealing adaptive learning rate mechanism is combined to dynamically adjust the learning rate and establish a preliminary automatic detection model for kidney stones. The formula for the label smoothing strategy is shown below: ; The cosine annealing learning rate scheduling formula is shown below: ; in The target label after label smoothing. For label smoothing parameters, The final fused feature labels for the training set. For the number of categories, Let be the learning rate at step t. These are the minimum learning rate and the maximum learning rate, respectively. These are the current number of training steps and the total number of training steps, respectively. The preliminary automatic kidney stone detection model was validated using the final fusion features of the validation set. An early stopping mechanism was introduced, which terminated training when the performance of the validation set did not improve for several consecutive rounds. The model was then iteratively trained using an enhanced multi-center, multi-device CT dataset to adapt it to different scenarios, resulting in the final automatic kidney stone detection model.

[0010] Preferably, the step of deploying the final automated kidney stone detection model to a high-performance workstation with GPU acceleration capabilities, or integrating it with an existing PACS in a medical institution to achieve automated clinical detection includes the following steps: The finally trained automatic kidney stone detection model was set up in a compatible environment on a high-performance workstation with GPU acceleration capabilities, and the appropriate GPU driver, CUDA and inference framework were installed. The system receives non-enhanced CT images from a PACS system or a local terminal, performs image processing consistent with the training phase, and then inputs the images into the deployed model. Through GPU parallel computing, it achieves rapid inference and automatically determines whether stones are present. The test results are sent back to the workstation in a structured format, supporting batch image parallel processing to meet high-throughput requirements. Doctors can directly view auxiliary results during the image reading process.

[0011] This embodiment also discloses a system for automatic detection of kidney stones using non-enhanced CT based on multi-scale feature fusion, characterized in that it includes: a data preprocessing module, a multi-scale feature extraction module, a model training and optimization module, a model deployment and integration module, and a data management and storage module; The data preprocessing module is used to perform privacy protection and standardization processing on the acquired non-enhanced CT coronal kidney images, providing data for subsequent feature extraction; The multi-scale feature extraction module is used to extract multi-dimensional features from local details to global context in images based on convolutional neural networks; The model training and optimization module is used to build a kidney stone detection model based on the extracted multi-dimensional features, and optimize the model performance through multi-dimensional strategies to ensure detection accuracy and generalization ability. The model deployment and integration module is used to deploy the trained kidney stone detection model to the clinical environment, enabling automated detection and result feedback of kidney stones via non-enhanced CT. The data management and storage module is responsible for the acquisition, storage, and dataset partitioning of multi-center CT image data, ensuring data security and availability.

[0012] The beneficial effects of this invention are as follows: (1) This invention acquires non-enhanced CT coronal kidney image data and performs de-identification processing on the acquired non-enhanced CT coronal kidney image data to obtain de-identified non-enhanced CT coronal kidney image data. Then, it processes the de-identified non-enhanced CT coronal kidney image data through data standardization and image enhancement methods to obtain processed non-enhanced CT coronal kidney image data. Based on the processed non-enhanced CT coronal kidney image data, it uses a convolutional neural network to extract key image features. At the same time, it establishes an automatic kidney stone detection model based on the extracted key image features and obtains the final automatic kidney stone detection model through a multi-dimensional training optimization strategy. Finally, it deploys the final automatic kidney stone detection model to a high-performance workstation with GPU acceleration capability or integrates it with the existing PACS of medical institutions to realize the clinical automatic detection function, realize the automated and efficient detection of kidney stones, and provide strong technical support for clinical diagnosis.

[0013] (2) This invention fully utilizes the multi-scale context and local detail information of non-enhanced CT images, and achieves multi-level feature fusion and adaptive enhancement by introducing cavity space pyramid pooling, pyramid pooling and channel attention mechanisms. While ensuring high detection accuracy, it effectively improves the ability to identify small-volume, vaguely defined stones, and has good cross-device and cross-population generalization performance, which can realize efficient and reliable automated screening and assisted diagnosis in clinical environments. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the automatic detection method for kidney stones using non-enhanced CT based on multi-scale feature fusion according to the present invention.

[0015] Figure 2 This is a framework diagram of the automatic kidney stone detection model based on multi-scale feature fusion convolutional neural network of the present invention. Detailed Implementation

[0016] 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.

[0017] Example 1 Please see Figure 1 This embodiment discloses an automatic detection method for kidney stones using non-enhanced CT based on multi-scale feature fusion, specifically including the following steps: S1. Acquire non-contrast CT coronal kidney image data, and perform de-identification processing on the acquired non-contrast CT coronal kidney image data to obtain de-identified non-contrast CT coronal kidney image data. The acquisition of non-contrast CT coronal kidney image data and the de-identification processing of the acquired non-contrast CT coronal kidney image data to obtain de-identified non-contrast CT coronal kidney image data include the following steps: Non-contrast CT coronal renal imaging data were collected from different imaging devices and populations based on clinical standards and multi-center acquisition mode; The acquired non-contrast CT coronal kidney image data were de-identified to remove all privacy information that could be associated with the patient's identity, resulting in de-identified non-contrast CT coronal kidney image data.

[0018] S2, the de-identified non-enhanced CT coronal kidney image data is processed through data standardization and image enhancement methods to obtain the processed non-enhanced CT coronal kidney image data; The process of processing the de-identified non-contrast CT coronal kidney image data using data standardization and image enhancement methods to obtain processed non-contrast CT coronal kidney image data includes the following steps: Pixel resampling was performed on the de-identified non-contrast CT coronal kidney images to unify the spatial resolution of all images to 1 mm³. Based on the resampled non-contrast CT coronal kidney image data, the size of the non-contrast CT coronal kidney image data matrix was scaled to 224×224 pixels using bilinear interpolation, and then processed using the Z-Score normalization method to obtain standardized non-contrast CT coronal kidney image data. The Z-Score standardization formula is shown below: ; in The pixel values ​​of the standardized non-contrast CT coronal kidney image. These are the pixel values ​​of a non-contrast CT coronal kidney image after scaling. These represent the mean and standard deviation of all pixel values ​​in a non-contrast CT coronal kidney image. Edge enhancement processing was performed on the standardized non-contrast CT coronal kidney image data, and high-frequency edge information of the non-contrast CT coronal kidney image data was extracted by the Laplacian operator. The processed non-contrast CT coronal kidney image data was then linearly superimposed on the original image.

[0019] S3, based on the processed non-enhanced CT coronal kidney image data, uses a convolutional neural network to extract key image features; The extraction of key image features based on the processed non-enhanced CT coronal kidney image data using a convolutional neural network includes the following steps: S31 employs ResNet, a residual-based convolutional neural network, as the backbone. Processed unenhanced CT coronal kidney images are input into the convolutional neural network, and multi-scale basic features from shallow to deep layers are extracted through layer-by-layer convolution and pooling operations. The extracted features are then integrated into a feature set. These correspond to features from shallow to deep, respectively. F1 contains rich edge and texture details, while F4 contains high-level semantic information. The feature extraction function expression is as follows: ; in For the first time after passing through the ResNet network Features extracted at each level For the ResNet network Feature extraction functions at each level The input is a processed non-contrast CT coronal image of the kidney; S32 introduces a void space pyramid pooling (ASPP) module to the deep feature F4, using different void ratios... Parallel convolution and global pooling are used to capture multi-scale global contextual information and output features. , ASPP output characteristics The expression is as follows: ; in The features are those extracted by the void space pyramid pooling module. For feature fusion function, For deep features of the input A function that performs a 1×1 convolution operation. These are the deep features of the input. Execution expansion rate The functions are 3×3 dilated convolution operations of 6, 12, and 18 respectively, used to extract features at different scales. For deep features of the input Perform global feature extraction.

[0020] S33 introduces a pyramid pooling (PPM) module to the mid-layer feature F3, enhancing the local structural representation through multi-scale adaptive pooling and outputting features. Furthermore, a channel attention mechanism (SE Block) is introduced, which performs global average pooling on each feature map to obtain channel description vectors, and then passes them through a fully connected layer and a Sigmoid activation function to obtain channel weights; PPM output characteristics The expression is as follows:

[0021] The SE Block expression is shown below: ; in The features output after processing by the Pyramid Pooling Module (PPM) The mid-level features input into the PPM module. This is a function that processes pooling features. For upsampling operation, To perform an adaptive pooling operation with a scale of k×k on the input features, This represents the global average pooling result for the c-th channel. These represent the height and width of the feature map, respectively. Let (i,j) be the feature value of the c-th channel feature map at position (i,j). For channel weight vectors, It is the Sigmoid activation function. This is the weight matrix of the fully connected layer. For all The channel description vector is composed of This is the feature map of the c-th channel after attention weighting; S34 utilizes a feature pyramid network to perform top-down multi-level feature fusion, achieving alignment between high-level semantics and low-level details to obtain the final fused features.

[0022] S4. An automatic kidney stone detection model is established based on the extracted key image features, and the final automatic kidney stone detection model is obtained through a multi-dimensional training optimization strategy. The process of establishing an automatic kidney stone detection model based on extracted key image features and obtaining the final automatic kidney stone detection model through a multi-dimensional training and optimization strategy includes the following steps: The final fused features are divided into two parts: a training set and a validation set. Based on the final fusion features of the training set, a label smoothing strategy is used to optimize the cross-entropy loss function for processing, and a cosine annealing adaptive learning rate mechanism is combined to dynamically adjust the learning rate and establish a preliminary automatic detection model for kidney stones. The formula for the label smoothing strategy is shown below: ; The cosine annealing learning rate scheduling formula is shown below: ; in The target label after label smoothing. For label smoothing parameters, The final fused feature labels for the training set. For the number of categories, Let be the learning rate at step t. These are the minimum learning rate and the maximum learning rate, respectively. These are the current number of training steps and the total number of training steps, respectively. The preliminary automatic kidney stone detection model was validated using the final fusion features of the validation set. An early stopping mechanism was introduced, which terminated training when the performance of the validation set did not improve for several consecutive rounds. The model was then iteratively trained using an enhanced multi-center, multi-device CT dataset to adapt it to different scenarios, resulting in the final automatic kidney stone detection model.

[0023] S5 can deploy the final automated kidney stone detection model to a high-performance workstation with GPU acceleration capabilities, or integrate it with existing PACS systems in medical institutions to achieve automated clinical detection.

[0024] The process of deploying the final automated kidney stone detection model to a high-performance workstation with GPU acceleration capabilities, or integrating it with existing PACS systems in medical institutions to achieve automated clinical detection, includes the following steps: The finally trained automatic kidney stone detection model was set up in a compatible environment on a high-performance workstation with GPU acceleration capabilities, and the appropriate GPU driver, CUDA and inference framework were installed. The system receives non-enhanced CT images from a PACS system or a local terminal, performs image processing consistent with the training phase, and then inputs the images into the deployed model. Through GPU parallel computing, it achieves rapid inference and automatically determines whether stones are present. The test results are sent back to the workstation in a structured format, supporting batch image parallel processing to meet high-throughput requirements. Doctors can directly view auxiliary results during the image reading process.

[0025] Example 2 This embodiment also discloses a system for automatic detection of kidney stones using non-enhanced CT based on multi-scale feature fusion, characterized in that it includes: a data preprocessing module, a multi-scale feature extraction module, a model training and optimization module, a model deployment and integration module, and a data management and storage module; The data preprocessing module is used to perform privacy protection and standardization processing on the acquired non-enhanced CT coronal kidney images, providing data for subsequent feature extraction; The multi-scale feature extraction module is used to extract multi-dimensional features from local details to global context in images based on convolutional neural networks; The model training and optimization module is used to build a kidney stone detection model based on the extracted multi-dimensional features, and optimize the model performance through multi-dimensional strategies to ensure detection accuracy and generalization ability. The model deployment and integration module is used to deploy the trained kidney stone detection model to the clinical environment, enabling automated detection and result feedback of kidney stones via non-enhanced CT. The data management and storage module is responsible for the acquisition, storage, and dataset partitioning of multi-center CT image data, ensuring data security and availability.

[0026] It should be noted that the above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined by the present invention, they should all fall within the protection scope of the present invention.

Claims

1. An automated detection method for kidney stones using non-enhanced CT based on multi-scale feature fusion, characterized in that, Includes the following steps: S1. Acquire non-contrast CT coronal kidney image data, and perform de-identification processing on the acquired non-contrast CT coronal kidney image data to obtain de-identified non-contrast CT coronal kidney image data. S2, the de-identified non-enhanced CT coronal kidney image data is processed through data standardization and image enhancement methods to obtain the processed non-enhanced CT coronal kidney image data; S3, based on the processed non-enhanced CT coronal kidney image data, uses a convolutional neural network to extract key image features; S4. An automatic kidney stone detection model is established based on the extracted key image features, and the final automatic kidney stone detection model is obtained through a multi-dimensional training optimization strategy. S5 can deploy the final automated kidney stone detection model to a high-performance workstation with GPU acceleration capabilities, or integrate it with existing PACS systems in medical institutions to achieve automated clinical detection.

2. The automatic detection method for kidney stones based on multi-scale feature fusion in non-enhanced CT according to claim 1, characterized in that, The acquisition of non-contrast CT coronal kidney image data and the de-identification processing of the acquired non-contrast CT coronal kidney image data to obtain de-identified non-contrast CT coronal kidney image data include the following steps: Non-contrast CT coronal renal imaging data were collected from different imaging devices and populations based on clinical standards and multi-center acquisition mode; The acquired non-contrast CT coronal kidney image data were de-identified to remove all privacy information that could be associated with the patient's identity, resulting in de-identified non-contrast CT coronal kidney image data.

3. The automatic detection method for kidney stones based on multi-scale feature fusion in non-enhanced CT according to claim 1, characterized in that, The process of processing the de-identified non-contrast CT coronal kidney image data using data standardization and image enhancement methods to obtain processed non-contrast CT coronal kidney image data includes the following steps: Pixel resampling was performed on the de-identified non-contrast CT coronal kidney images to unify the spatial resolution of all images to 1 mm³. Based on the resampled non-contrast CT coronal kidney image data, the size of the non-contrast CT coronal kidney image data matrix was scaled to 224×224 pixels using bilinear interpolation, and then processed using the Z-Score normalization method to obtain standardized non-contrast CT coronal kidney image data. The Z-Score standardization formula is shown below: ； in The pixel values ​​of the standardized non-contrast CT coronal kidney image. These are the pixel values ​​of a non-contrast CT coronal kidney image after scaling. Do not specify the mean and standard deviation of all pixel values ​​for non-contrast CT coronal kidney images; Edge enhancement processing was performed on the standardized non-contrast CT coronal kidney image data, and high-frequency edge information of the non-contrast CT coronal kidney image data was extracted by the Laplacian operator. The processed non-contrast CT coronal kidney image data was then linearly superimposed on the original image.

4. The automatic detection method for kidney stones based on multi-scale feature fusion in non-enhanced CT according to claim 1, characterized in that, The extraction of key image features based on the processed non-enhanced CT coronal kidney image data using a convolutional neural network includes the following steps: S31 employs ResNet, a residual-based convolutional neural network, as the backbone. Processed unenhanced CT coronal kidney images are input into the convolutional neural network, and multi-scale basic features from shallow to deep layers are extracted through layer-by-layer convolution and pooling operations. The extracted features are then integrated into a feature set. These correspond to features from shallow to deep, respectively. F1 contains rich edge and texture details, while F4 contains high-level semantic information. The feature extraction function expression is as follows: in For the first time after passing through the ResNet network Features extracted at each level For the ResNet network Feature extraction functions at each level The input is a processed non-contrast CT coronal image of the kidney; S32 introduces a void space pyramid pooling (ASPP) module to the deep feature F4, using different void ratios... Parallel convolution and global pooling are used to capture multi-scale global contextual information and output features. , ASPP output characteristics The expression is as follows: in The features are those extracted by the void space pyramid pooling module. For feature fusion function, For deep features of the input A function that performs a 1×1 convolution operation. These are the deep features of the input. Execution expansion rate The functions are 3×3 dilated convolution operations of 6, 12, and 18 respectively, used to extract features at different scales. For deep features of the input Perform global feature extraction operations; S33 introduces a pyramid pooling (PPM) module to the mid-layer feature F3, enhancing the local structural representation through multi-scale adaptive pooling and outputting features. Furthermore, a channel attention mechanism (SE Block) is introduced, which performs global average pooling on each feature map to obtain channel description vectors, and then passes them through a fully connected layer and a Sigmoid activation function to obtain channel weights; PPM output characteristics The expression is as follows: The SE Block expression is shown below: ； in The features output after processing by the Pyramid Pooling Module (PPM) The mid-level features input into the PPM module. This is a function that processes pooling features. For upsampling operation, To perform an adaptive pooling operation with a scale of k×k on the input features, This represents the global average pooling result for the c-th channel. These represent the height and width of the feature map, respectively. Let (i,j) be the feature value of the c-th channel feature map at position (i,j). This is the channel weight vector. Sigmoid activation function This is the weight matrix of the fully connected layer. For all The channel description vector is composed of This is the feature map of the c-th channel after attention weighting; S34 utilizes a feature pyramid network to perform top-down multi-level feature fusion, achieving alignment between high-level semantics and low-level details to obtain the final fused features.

5. The automatic detection method for kidney stones based on multi-scale feature fusion in non-enhanced CT according to claim 1, characterized in that, The process of establishing an automatic kidney stone detection model based on extracted key image features and obtaining the final automatic kidney stone detection model through a multi-dimensional training and optimization strategy includes the following steps: The final fused features are divided into two parts: a training set and a validation set. Based on the final fusion features of the training set, a label smoothing strategy is used to optimize the cross-entropy loss function for processing, and a cosine annealing adaptive learning rate mechanism is combined to dynamically adjust the learning rate and establish a preliminary automatic detection model for kidney stones. The formula for the label smoothing strategy is shown below: ； The cosine annealing learning rate scheduling formula is shown below: ； in The target label after label smoothing. For label smoothing parameters, The final fused feature labels for the training set. For the number of categories, Let be the learning rate at step t. These are the minimum learning rate and the maximum learning rate, respectively. These are the current number of training steps and the total number of training steps, respectively. The preliminary automatic kidney stone detection model was validated using the final fusion features of the validation set. An early stopping mechanism was introduced, which terminated training when the performance of the validation set did not improve for several consecutive rounds. The model was then iteratively trained using an enhanced multi-center, multi-device CT dataset to adapt it to different scenarios, resulting in the final automatic kidney stone detection model.

6. The automatic detection method for kidney stones based on multi-scale feature fusion in non-enhanced CT according to claim 1, characterized in that, The process of deploying the final automated kidney stone detection model to a high-performance workstation with GPU acceleration capabilities, or integrating it with existing PACS systems in medical institutions to achieve automated clinical detection, includes the following steps: The finally trained automatic kidney stone detection model was set up in a compatible environment on a high-performance workstation with GPU acceleration capabilities, and the appropriate GPU driver, CUDA and inference framework were installed. The system receives non-enhanced CT images from a PACS system or a local terminal, performs image processing consistent with the training phase, and then inputs the images into the deployed model. Through GPU parallel computing, it achieves rapid inference and automatically determines whether stones are present. The test results are sent back to the workstation in a structured format, supporting batch image parallel processing to meet high-throughput requirements. Doctors can directly view auxiliary results during the image reading process.

7. A system for implementing the automatic detection method for kidney stones based on multi-scale feature fusion using non-enhanced CT as described in any one of claims 1-6, characterized in that, include: The module includes a data preprocessing module, a multi-scale feature extraction module, a model training and optimization module, a model deployment and integration module, and a data management and storage module. The data preprocessing module is used to perform privacy protection and standardization processing on the acquired non-enhanced CT coronal kidney images, providing data for subsequent feature extraction; The multi-scale feature extraction module is used to extract multi-dimensional features from local details to global context in images based on convolutional neural networks; The model training and optimization module is used to build a kidney stone detection model based on the extracted multi-dimensional features, and optimize the model performance through multi-dimensional strategies to ensure detection accuracy and generalization ability. The model deployment and integration module is used to deploy the trained kidney stone detection model to the clinical environment, enabling automated detection and result feedback of kidney stones via non-enhanced CT. The data management and storage module is responsible for the acquisition, storage, and dataset partitioning of multi-center CT image data, ensuring data security and availability.

Images