Image segmentation method and system for stone removal

Abstract

The invention relates to the technical field of image analysis, and discloses an image segmentation method and system for calculus removal, and the method comprises the steps: carrying out the real-time image collection of a target part based on a pre-deployed calculus image collection device, obtaining an initial calculus image, carrying out the preprocessing of the initial calculus image, and obtaining an initial calculus image; obtaining a preprocessed initial stone image; inputting the preprocessed initial stone image into a deep learning model based on a convolutional neural network to obtain a preliminary segmentation result; optimizing the preliminary segmentation result by adopting a Sobel operator in combination with closed operation to obtain an optimized segmentation result; a connected domain analysis method is adopted for the optimized segmentation result, and a complete stone area image is obtained. According to the invention, the problem of inaccurate segmentation results caused by the influence of noise on seed point selection in the prior art can be solved.

Description

Technical Field

[0001] The present invention relates to the technical field of image analysis, and in particular to an image segmentation method and system for stone removal. Background Art

[0002] With the continuous advancement of modern medical imaging technology, the diagnosis and treatment of stones have gradually shifted from traditional manual diagnosis to digital and automated computer-aided diagnosis (CAD) systems. Especially in the process of stone removal, image segmentation technology plays an important role in extracting the stone area and assisting doctors in making accurate diagnosis and treatment decisions. The image segmentation task of stones usually faces multiple challenges, mainly including blurred boundaries between stones and surrounding tissues, noise interference, and the diversity of stone shapes.

[0003] Although traditional image segmentation methods such as threshold segmentation, edge detection and region growing are effective in some simple scenarios, the segmentation effect of traditional methods is not ideal due to the background noise in the stone image, the complexity of the stone morphology and the changes in the image under different imaging conditions. With the successful application of deep learning, especially convolutional neural network (CNN) in computer vision, deep learning methods have gradually become the mainstream means of medical image segmentation. However, the existing image segmentation methods based on deep learning still have problems such as inaccurate extraction of stone regions, poor recognition ability of complex morphological stones, and high consumption of computing resources. In particular, in the extraction of edge information and the processing of feature fusion, they often rely on complex network structures, resulting in low computational efficiency.

[0004] Therefore, developing an efficient and accurate stone image segmentation method that can retain details and reduce noise interference when extracting stone areas and can cope with stones of different shapes and sizes is an urgent problem to be solved in the current field of medical image processing. Summary of the invention

[0005] In view of this, the present invention proposes an image segmentation method and system for stone removal, aiming to solve the problem in the current technology that seed point selection is affected by noise, resulting in inaccurate segmentation results.

[0006] The present invention provides an image segmentation method for stone removal, comprising: Based on the pre-deployed stone image acquisition device, real-time image acquisition is performed on the target part to obtain an initial stone image, and the initial stone image is pre-processed to obtain a pre-processed initial stone image; The preprocessed initial stone image is input into the deep learning model based on convolutional neural network to obtain the preliminary segmentation result; The preliminary segmentation results are optimized using the Sobel operator combined with the closing operation to obtain the optimized segmentation results; The connected domain analysis method is used on the optimized segmentation results to obtain a complete stone area image.

[0007] Furthermore, the preprocessing of the initial stone image to obtain the preprocessed initial stone image specifically comprises: performing noise elimination on the initial stone image using a multi-scale filtering method to obtain a denoised stone image, and performing contrast enhancement and histogram equalization processing on the denoised stone image to obtain a preprocessed initial stone image.

[0008] Furthermore, the preprocessed initial stone image is input into a deep learning model based on a convolutional neural network to obtain a preliminary segmentation result. Specifically, the preprocessed initial stone image is input into a deep learning model based on a convolutional neural network, local edge information of the preprocessed initial stone image is extracted through shallow features in the encoder, and global context information of the preprocessed initial stone image is extracted through deep features. Finally, the shallow features and the deep features are fused through jump connections and decoded by a decoder to obtain a confidence map. By setting a threshold for the confidence map, the area in the confidence map that is greater than the set threshold is automatically extracted as a seed point, and the preliminary segmentation result is obtained by combining the region growing method.

[0009] Furthermore, the convolutional neural network-based deep learning model adopts an improved U-Net architecture, that is, a residual module, a parallel convolution module and an attention module are combined on the original U-Net architecture. The improved U-Net architecture includes an encoder and a decoder. The encoder includes multiple convolution modules and a pooling layer. Each of the multiple convolution modules is added with a residual module, and finally a parallel convolution module is added to form the decoder. The decoder includes a deconvolution module, a feature fusion module, a convolution module and a parallel convolution module, and the feature fusion module includes an attention module.

[0010] Furthermore, the local edge information of the preprocessed initial stone image is extracted through shallow features in the encoder, specifically: the preprocessed initial stone image is input into the encoder, and after normalization, it is subjected to a convolution operation of a 3×3 convolution kernel, and nonlinear expression ability is introduced through an activation function, and then the convolution operation of multiple 3×3 convolution kernels in the convolution module is used to extract basic edge information, texture and low-level features in the preprocessed initial stone image, and the basic edge information, texture and low-level features in the preprocessed initial stone image are normalized and then subjected to maximum pooling processing by the pooling layer, and then the local edge information is formed by splicing through the parallel convolution module; The basic edge information includes boundary lines and contour lines in the preprocessed initial stone image; The texture includes the roughness of the stone surface and the degree of concavity and convexity of the stone surface; The low-level features include grayscale contrast and color contrast between the stone and the pre-processed initial stone image background.

[0011] Furthermore, the global context information of the initial stone image preprocessed by deep feature extraction is specifically as follows: the preprocessed initial stone image is input into the encoder, normalized and then subjected to a convolution operation with a 3×3 convolution kernel, nonlinear expression capability is introduced through an activation function, and then multiple 3×3 convolution operations in multiple convolution modules are performed to extract local features in the preprocessed initial stone image, and after each convolution operation, maximum pooling is performed through a pooling layer to aggregate global context information with local features, and the global context information obtained after each convolution operation is integrated through a parallel convolution module to obtain the global context information of the preprocessed initial stone image.

[0012] Furthermore, the shallow features and deep features are finally fused through jump connections and decoded by a decoder to obtain a confidence map. The specific content is: the shallow features and deep features are input into the decoder, first passed through a deconvolution module, an inverse convolution operation is performed, and then a jump connection is performed to fuse and splice the shallow features and deep features that have passed the deconvolution module to obtain a fused feature map, and then the fused feature map is convolved, and the attention module is combined to improve the accuracy of the fused feature map. Finally, the fused feature map with improved accuracy is mapped to a single-channel confidence map through a 1×1 convolution layer.

[0013] Furthermore, the preliminary segmentation result is optimized using the Sobel operator in combination with a closing operation to obtain an optimized segmentation result. The specific content is: the Sobel operator is used to perform edge detection on the preliminary segmentation result, and the edge information of the preliminary segmentation result is extracted by calculating the gradient of the image to obtain the stone area. The stone area is then dilated to expand the stone area, and finally an erosion operation is performed to restore the original shape of the stone area and remove the noise introduced in the dilation operation, thereby finally forming an optimized segmentation result.

[0014] Furthermore, the connected domain analysis method is used for the optimized segmentation results to obtain a complete stone area image. The specific content is: the optimized segmentation results are divided into regions based on pixel values, and the same pixel values ​​are classified into the same stone area, thereby converting the optimized segmentation results into a complete stone area image.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) By adopting an improved U-Net architecture combined with a residual module, a parallel convolution module and an attention module, it is possible to effectively extract shallow and deep features in the image, especially in capturing edge information and fusing multi-scale features. The use of jump connections further enhances the fusion of shallow and deep features and improves the precision of image segmentation results. (2) The Sobel operator combined with a closed operation is used to optimize the preliminary segmentation results, which can further refine the segmentation boundaries and enhance the coherence of the stone area. The combination of dilation and erosion operations effectively removes noise and artifacts, ensuring accurate extraction of the stone area. (3) The optimized segmentation results are post-processed by the connected domain analysis method to ensure that the complete stone area image can be extracted, avoiding omission or misidentification of the segmented area, and providing accurate regional positioning for subsequent stone removal and medical diagnosis. (4) Through the improved U-Net architecture, the powerful feature extraction capabilities of deep convolutional neural networks are fully utilized. At the same time, combined with lightweight network optimization, the model has effectively improved its computational efficiency while ensuring high accuracy, making it suitable for actual clinical applications. (5) It can automatically perform image acquisition, preprocessing, segmentation and post-processing, greatly reducing the need for manual intervention, reducing the workload of doctors and improving diagnostic efficiency. In addition, the system can automatically extract stone areas based on the set thresholds and seed points, further improving the automation level of the system. On the other hand, the present invention also proposes an image segmentation system for stone removal, the system comprising: The first module is configured to perform real-time image acquisition of the target part based on a pre-deployed stone image acquisition device to obtain an initial stone image, and pre-process the initial stone image to obtain a pre-processed initial stone image; The second module is configured to input the preprocessed initial stone image into a deep learning model based on a convolutional neural network to obtain a preliminary segmentation result; The third module is configured to optimize the preliminary segmentation result by using the Sobel operator combined with the closing operation to obtain an optimized segmentation result; The fourth module is configured to use the connected domain analysis method on the optimized segmentation result to obtain a complete stone area image.

[0016] It can be understood that the above-mentioned image segmentation system for stone removal has the same beneficial effects and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present invention. Moreover, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings: Figure 1 A flowchart of an image segmentation method for stone removal according to an embodiment of the present invention; Figure 2 The present invention is a structural block diagram of an image segmentation system for stone removal according to an embodiment of the present invention. DETAILED DESCRIPTION

[0018] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art. It should be noted that, in the absence of conflict, the embodiments of the present invention and the features described in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0019] See also Figure 1 As shown, an embodiment of the present invention provides an image segmentation method for stone removal, comprising: S1: Based on the pre-deployed stone image acquisition device, real-time image acquisition is performed on the target part to obtain an initial stone image, and the initial stone image is pre-processed to obtain a pre-processed initial stone image; S2: The preprocessed initial stone image is input into the deep learning model based on convolutional neural network to obtain the preliminary segmentation result; S3: The preliminary segmentation result is optimized by using the Sobel operator combined with the closing operation to obtain the optimized segmentation result; S4: The connected domain analysis method is used on the optimized segmentation results to obtain a complete stone area image.

[0020] Furthermore, the initial stone image is preprocessed to obtain the preprocessed initial stone image. Specifically, the initial stone image is subjected to noise elimination using a multi-scale filtering method to obtain a denoised stone image, and the denoised stone image is subjected to contrast enhancement and histogram equalization processing to obtain a preprocessed initial stone image.

[0021] It should be noted that the multiscale filtering method can effectively remove different types of noise (such as salt and pepper noise, artifacts, etc.) and improve image quality by analyzing images at multiple scales. By increasing the contrast of the image, the outline and details of the stone will be more prominent, thereby improving the segmentation algorithm's ability to recognize the stone area. Especially in low-contrast areas, histogram equalization can effectively enhance the details of the stone area, improve the contrast between the stone and the background, and make the stone boundary more distinct.

[0022] Furthermore, the preprocessed initial stone image is input into a deep learning model based on a convolutional neural network to obtain a preliminary segmentation result. The specific content is: the preprocessed initial stone image is input into a deep learning model based on a convolutional neural network, the local edge information of the preprocessed initial stone image is extracted through the shallow features in the encoder, and the global context information of the preprocessed initial stone image is extracted through the deep features. Finally, the shallow features and the deep features are fused through jump connections and decoded by the decoder to obtain a confidence map. By setting a threshold for the confidence map, the area in the confidence map that is greater than the set threshold is automatically extracted as a seed point, and the preliminary segmentation result is obtained by combining the region growing method.

[0023] Specifically, the confidence map is a probability map that reflects each pixel point belonging to the stone area. Each pixel value in the map represents the probability value of the location belonging to the stone area. The larger the value, the more likely it is a stone area, and the smaller the value, the background or non-stone area.

[0024] According to the actual application requirements and image quality, a suitable threshold is set for the confidence map. This threshold is usually an empirical value or an optimal value tuned according to the training set data. After setting the threshold, the areas in the confidence map with pixel values ​​greater than the threshold will be selected as "seed points". These seed points represent pixels with a higher probability of being stone areas.

[0025] Based on the set threshold, all regions larger than the threshold are automatically extracted from the confidence map as seed points. The seed point is the starting point for region growing and has a higher probability of stone. In this way, the system can automatically identify possible stone areas from the entire image, avoiding manual labeling and intervention.

[0026] After extracting the seed point, the region growing algorithm is used to expand the pixels around the seed point. The region growing method is based on the similarity of the local area. It compares the adjacent pixels with the seed point. If the value of the adjacent pixels is also greater than the set threshold, these pixels are added to the stone area. Through continuous expansion, the region growing method gradually constructs the entire stone area.

[0027] Finally, the region obtained by the region growing algorithm is the preliminary segmentation result. In this result, the stone region has been successfully segmented from the background. The gradual expansion of region growing can ensure the coherence of the stone region and avoid mis-segmentation caused by image noise or edge blur.

[0028] It should be noted that by setting the threshold to automatically extract seed points and combining the region growing method, the entire segmentation process greatly reduces manual intervention. This automated processing method can significantly improve segmentation efficiency, reduce errors caused by human factors, and improve segmentation consistency. After setting a suitable threshold and extracting seed points, the stone area can be accurately located. The threshold setting helps the system filter out those low-probability noise areas to ensure that the segmentation results are more reliable. The extraction of seed points is based on the high probability values ​​in the confidence map, which avoids incorrect segmentation.

[0029] Furthermore, an improved U-Net architecture is adopted based on the convolutional neural network deep learning model, that is, a residual module, a parallel convolution module and an attention module are combined with the original U-Net architecture. The improved U-Net architecture includes an encoder and a decoder. The encoder includes multiple convolution modules and a pooling layer. A residual module is added to each convolution module in the multiple convolution modules, and finally a parallel convolution module is added to form the decoder. The decoder includes a deconvolution module, a feature fusion module, a convolution module and a parallel convolution module. The feature fusion module includes an attention module.

[0030] It should be noted that residual connections avoid the problem of gradient vanishing when the network is deepened, thereby accelerating the convergence process and improving the stability of training. Residual connections allow information to be transmitted between different layers, avoiding the weakening of information during multi-layer transmission, and enhancing the network's representation ability and ability to learn complex features. The residual module optimizes the flow of information through cross-layer connections, enabling the network to extract more profound features and reducing the degradation problem of the model during training. The parallel convolution module can process convolution kernels of different sizes in parallel at the same level, capturing information of more scales, so as to accurately identify stones of different sizes and shapes. The combination of multiple convolution kernels at the same level makes the network have a richer understanding of details and overall structure, thereby improving the segmentation effect. Through parallel convolution, features of different scales can be fully integrated, further improving the segmentation accuracy, especially in the processing of stone boundaries and details. The attention module can make the network pay more attention to the key information of the stone (such as edges, contours, etc.) during decoding, reduce the impact of background noise, and thus optimize the image segmentation results. By weighting important areas, the attention mechanism can dynamically adjust the focus of the network, making the network's segmentation of the stone area more accurate, especially in cases of complex morphology or poor image quality. By selectively focusing on important features, the attention module helps the network reduce computational redundancy and improve computational efficiency, while also improving the network's generalization ability and avoiding overfitting of irrelevant information.

[0031] Furthermore, the local edge information of the preprocessed initial stone image is extracted through the shallow features in the encoder, specifically: the preprocessed initial stone image is input into the encoder, and after normalization, it is subjected to a convolution operation of a 3×3 convolution kernel, and the nonlinear expression ability is introduced through the activation function, and then the convolution operation of multiple 3×3 convolution kernels in the convolution module is used to extract the basic edge information, texture and low-level features in the preprocessed initial stone image, and the basic edge information, texture and low-level features in the preprocessed initial stone image are normalized and then subjected to maximum pooling processing by the pooling layer, and then the local edge information is formed by splicing through the parallel convolution module; Basic edge information includes boundary lines and contour lines in the preprocessed initial stone image; Texture includes the roughness of the stone surface and the degree of concavity and convexity of the stone surface; The low-level features include the grayscale contrast and color contrast between the stone and the pre-processed initial stone image background.

[0032] Furthermore, the global context information of the preprocessed initial stone image is extracted through deep features. Specifically, the preprocessed initial stone image is input into the encoder, normalized and then subjected to a convolution operation with a 3×3 convolution kernel. Nonlinear expression capability is introduced through an activation function, and then multiple 3×3 convolution operations in multiple convolution modules are performed to extract local features in the preprocessed initial stone image. After each convolution operation, the maximum pooling process is performed through the pooling layer to aggregate the local features into the global context information. The global context information obtained after each convolution operation is integrated through a parallel convolution module to obtain the global context information of the preprocessed initial stone image.

[0033] Furthermore, the shallow features and deep features are fused through jump connections and decoded by the decoder to obtain a confidence map. The specific content is: the shallow features and deep features are input into the decoder, first pass through the deconvolution module, perform inverse convolution operation, and then perform jump connections to fuse and splice the shallow features and deep features that have passed the deconvolution module to obtain a fused feature map, and then perform convolution operation on the fused feature map. At the same time, the attention module is used to improve the accuracy of the fused feature map. Finally, the fused feature map with improved accuracy is mapped to a single-channel confidence map through a 1×1 convolution layer.

[0034] Furthermore, the preliminary segmentation results are optimized using the Sobel operator combined with a closing operation to obtain an optimized segmentation result. The specific contents are as follows: the Sobel operator is used to perform edge detection on the preliminary segmentation results, and the edge information of the preliminary segmentation results is extracted by calculating the gradient of the image to obtain the stone area. Subsequently, a dilation operation is performed on the stone area to expand the stone area, and finally an corrosion operation is performed to restore the original shape of the stone area and remove the noise introduced in the dilation operation, thereby finally forming an optimized segmentation result.

[0035] Specifically, the Sobel operator is used to perform edge detection on the preliminary segmentation results. The Sobel operator can extract edge information of local areas in the image by calculating the gradient of each pixel in the image. When processing stone images, the Sobel operator can effectively capture the boundary between the stone and the surrounding tissue, thereby clarifying the location and morphological characteristics of the stone area. The Sobel operator can reveal the changes in edge intensity in the image through horizontal and vertical convolution operations, thereby helping to distinguish stones from the background.

[0036] The stone area extracted by the Sobel operator is dilated to expand the stone area. The dilation operation fills the small gaps or discontinuous parts in the stone area by expanding the pixels in the neighborhood around the pixel to ensure the continuity of the stone area. Dilation can make up for some details that may be lost during edge detection, making the stone area more complete and eliminating small-scale missed detection or missing parts caused by image noise.

[0037] After the dilation operation, the erosion operation is performed. The erosion operation is the reverse process of dilation. It restores the original shape of the stone area by shrinking the boundary of the dilated area. The erosion operation can remove the noise and artifacts that may be introduced by the dilation operation, especially those irrelevant pixels that are mistakenly extended to the background area during dilation. Through erosion, the shape of the stone area is restored to a state closer to the real state, and at the same time, small-scale interference caused by noise or misoperation can be removed, improving the accuracy of the segmentation result.

[0038] After edge detection, dilation and erosion operations using the Sobel operator, the optimized segmentation result is finally obtained. At this point, the stone area is more accurately extracted, avoiding artifacts or missed segmentation that may exist in the initial segmentation results. The boundaries of the stone area are clearer and the shape is more in line with the actual situation.

[0039] Furthermore, the connected domain analysis method is used on the optimized segmentation results to obtain a complete stone area image. The specific content is: the optimized segmentation results are divided into regions based on pixel values, and the same pixel values ​​are classified into the same stone area, thereby converting the optimized segmentation results into a complete stone area image.

[0040] Specifically, for the optimized segmentation results, the process of connected domain analysis includes the following steps: (1) Image scanning: Scan the optimized segmentation results, traversing pixel by pixel from left to right and from top to bottom. When encountering a pixel with a value of 1 (stone area), start to detect whether it has been marked as part of a connected domain. 0 represents a non-stone area and 1 represents a stone area, which is explained in the above confidence map and will not be explained in detail here.

[0041] (2) Marking connected domains: Use the depth-first search (DFS) or breadth-first search (BFS) algorithm to traverse the neighboring pixels around the pixel and determine all the pixels connected to it. All pixels connected to the pixel are marked as the same connected domain.

[0042] (3) Connected domain grouping: Assign a unique label (connected domain number) to each independent stone region. Whenever a new stone region is found, assign it a new number and continue scanning the image.

[0043] (4) Region screening: Screen all detected connected regions and remove possible small noise regions. Remove regions smaller than a certain area threshold, which are likely to be noise or small artifacts.

[0044] (5) Output of connected domain labeling map: For all the selected connected regions, a labeling map or segmentation map is generated, in which each connected region is assigned a unique label. These labels correspond to different stone regions in the image.

[0045] Finally, a complete image of the stone area is formed.

[0046] It should be noted that the connected domain analysis method effectively eliminates the problem of breaks or omissions in the segmentation process by classifying areas with the same pixel values ​​into a connected area, ensuring that the stone area is completely extracted. This is especially important for dealing with stones with complex shapes and unclear boundaries. It can ensure that the stone area in the image is fully captured, thereby avoiding the risk of missed or mis-segmentation. The connected domain analysis method can refine the optimized segmentation results, especially when dealing with stones with complex shapes and different sizes, and can ensure that small stones and irregularly shaped stones can also be correctly identified. Through precise area division, it is possible to avoid the impact of noise or artifacts on the segmentation quality and ensure that the extracted stone area conforms to the actual anatomical characteristics.

[0047] See also Figure 2 As shown, an embodiment of the present invention further provides an image segmentation system for stone removal, the system comprising: The first module is configured to perform real-time image acquisition of the target part based on a pre-deployed stone image acquisition device to obtain an initial stone image, and pre-process the initial stone image to obtain a pre-processed initial stone image; The second module is configured to input the preprocessed initial stone image into a deep learning model based on a convolutional neural network to obtain a preliminary segmentation result; The third module is configured to optimize the preliminary segmentation result by using the Sobel operator combined with the closing operation to obtain an optimized segmentation result; The fourth module is configured to use the connected domain analysis method on the optimized segmentation result to obtain a complete stone area image.

[0048] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. An image segmentation method for stone removal, characterized in that: include: Based on the pre-deployed stone image acquisition device, real-time image acquisition is performed on the target part to obtain an initial stone image, and the initial stone image is pre-processed to obtain a pre-processed initial stone image; The preprocessed initial stone image is input into the deep learning model based on convolutional neural network to obtain the preliminary segmentation result; The preliminary segmentation results are optimized using the Sobel operator combined with the closing operation to obtain the optimized segmentation results; The connected domain analysis method is used on the optimized segmentation results to obtain a complete stone area image.

2. The image segmentation method for stone removal according to claim 1, characterized in that: The specific content of preprocessing the initial stone image to obtain the preprocessed initial stone image is: using a multi-scale filtering method to eliminate noise from the initial stone image to obtain a denoised stone image, and performing contrast enhancement and histogram equalization processing on the denoised stone image to obtain a preprocessed initial stone image.

3. The image segmentation method for stone removal according to claim 2, characterized in that: The method comprises inputting the preprocessed initial stone image into a deep learning model based on a convolutional neural network to obtain a preliminary segmentation result. Specifically, the preprocessed initial stone image is input into a deep learning model based on a convolutional neural network, extracting local edge information of the preprocessed initial stone image through shallow features in an encoder, extracting global context information of the preprocessed initial stone image through deep features, and finally fusing shallow features with deep features through jump connections and decoding them through a decoder to obtain a confidence map. By setting a threshold for the confidence map, the area in the confidence map that is greater than the set threshold is automatically extracted as a seed point, and the preliminary segmentation result is obtained by combining the region growing method.

4. The image segmentation method for stone removal according to claim 3, characterized in that: The convolutional neural network-based deep learning model adopts an improved U-Net architecture, that is, a residual module, a parallel convolution module and an attention module are combined on the original U-Net architecture. The improved U-Net architecture includes an encoder and a decoder. The encoder includes multiple convolution modules and a pooling layer. Each of the multiple convolution modules is added with a residual module, and finally a parallel convolution module is added to form the decoder. The decoder includes a deconvolution module, a feature fusion module, a convolution module and a parallel convolution module, and the feature fusion module includes an attention module.

5. The image segmentation method for stone removal according to claim 4, characterized in that: The local edge information of the preprocessed initial stone image is extracted through shallow features in the encoder, specifically comprising: inputting the preprocessed initial stone image into the encoder, performing a convolution operation of a 3×3 convolution kernel after normalization, introducing nonlinear expression capability through an activation function, and then performing a convolution operation of multiple 3×3 convolution kernels in a convolution module to extract basic edge information, texture and low-level features in the preprocessed initial stone image, normalizing the basic edge information, texture and low-level features in the preprocessed initial stone image, performing a maximum pooling process through a pooling layer, and then performing splicing through a parallel convolution module to form local edge information; The basic edge information includes boundary lines and contour lines in the preprocessed initial stone image; The texture includes the roughness of the stone surface and the degree of concavity and convexity of the stone surface; The low-level features include grayscale contrast and color contrast between the stone and the pre-processed initial stone image background.

6. The image segmentation method for stone removal according to claim 5, characterized in that: The global context information of the preprocessed initial stone image is extracted through deep features. Specifically, the preprocessed initial stone image is input into the encoder, normalized and then subjected to a convolution operation with a 3×3 convolution kernel, nonlinear expression capability is introduced through an activation function, and then local features in the preprocessed initial stone image are extracted through multiple 3×3 convolution operations in multiple convolution modules. After each convolution operation, maximum pooling is performed through a pooling layer to aggregate global context information with local features, and the global context information obtained after each convolution operation is integrated through a parallel convolution module to obtain the global context information of the preprocessed initial stone image.

7. The image segmentation method for stone removal according to claim 6, characterized in that: The confidence map is obtained by finally fusing the shallow features and the deep features through a jump connection and decoding them through a decoder. Specifically, the shallow features and the deep features are input into the decoder, first passed through a deconvolution module, an inverse convolution operation is performed, and then a jump connection is performed to fuse and splice the shallow features and the deep features that have passed through the deconvolution module to obtain a fused feature map, and then the fused feature map is convolved, and the attention module is combined to improve the accuracy of the fused feature map. Finally, the fused feature map with improved accuracy is mapped to a single-channel confidence map through a 1×1 convolution layer.

8. The image segmentation method for stone removal according to claim 7, characterized in that: The preliminary segmentation result is optimized by using the Sobel operator in combination with a closing operation to obtain an optimized segmentation result. Specifically, the Sobel operator is used to perform edge detection on the preliminary segmentation result, and the edge information of the preliminary segmentation result is extracted by calculating the gradient of the image to obtain the stone area. Subsequently, a dilation operation is performed on the stone area to expand the stone area, and finally an corrosion operation is performed to restore the original shape of the stone area, and the noise introduced in the dilation operation is removed, thereby finally forming an optimized segmentation result.

9. The image segmentation method for stone removal according to claim 8, characterized in that: The connected domain analysis method is used for the optimized segmentation results to obtain a complete stone area image. The specific content is: the optimized segmentation results are divided into regions based on pixel values, and the same pixel values ​​are classified into the same stone area, thereby converting the optimized segmentation results into a complete stone area image.

10. An image segmentation system for stone removal, implemented in the image segmentation method for stone removal according to any one of claims 1 to 9, characterized in that: The system comprises: The first module is configured to perform real-time image acquisition of the target part based on a pre-deployed stone image acquisition device to obtain an initial stone image, and pre-process the initial stone image to obtain a pre-processed initial stone image; The second module is configured to input the preprocessed initial stone image into a deep learning model based on a convolutional neural network to obtain a preliminary segmentation result; The third module is configured to optimize the preliminary segmentation result by using the Sobel operator combined with the closing operation to obtain an optimized segmentation result; The fourth module is configured to use the connected domain analysis method on the optimized segmentation result to obtain a complete stone area image.

Images