Gastrointestinal stromal tumor segmentation method, device and medium

Abstract

The invention discloses a gastrointestinal stromal tumor segmentation method and device and a medium, and is applied to the technical field of medical image.The method comprises the steps that a stomach ultrasound image is obtained, and the stomach ultrasound image comprises a gastrointestinal stromal tumor area; inputting the stomach ultrasound image into a pre-trained segmentation model to obtain a contour map of a gastrointestinal stromal tumor region; in the segmentation model, spatial domain feature extraction is carried out on the stomach ultrasound image through a spatial domain processing module, and low-level features and spatial features are obtained; performing frequency domain feature extraction based on the spatial features through a frequency domain processing module to obtain frequency domain features; fusing the spatial features and the frequency domain features through a fusion module to obtain fused features; and performing segmentation processing through an output module according to the low-level features and the fusion features to obtain a contour map of the gastrointestinal stromal tumor region. According to the method, the segmentation precision of the gastrointestinal stromal tumor can be effectively improved.

Description

Technical Field

[0001] The present application relates to the field of medical imaging technology, and in particular to a method, device, and medium for segmenting gastrointestinal stromal tumors. Background Art

[0002] Gastrointestinal stromal tumors (GISTs) are a common type of tumor that originates from the mesenchymal tissue of the gastrointestinal tract. In related technologies, existing image segmentation models are often used to segment GIST regions from gastric ultrasound images, thereby achieving the GIST segmentation task.

[0003] However, GISTs often have characteristics such as blurred boundaries, uneven scales, and variable positions in ultrasound images. Existing image segmentation models are mostly fully convolutional neural networks (FCNs), U-Nets, and multi-scale structures that introduce dilated convolutions. These models are limited by spatial domain convolution operations and suffer from problems such as inaccurate edge segmentation and missed detection of small lesions when facing GIST segmentation tasks, resulting in low GIST segmentation accuracy. Summary of the Invention

[0004] The embodiments of the present application provide a gastrointestinal stromal tumor segmentation method, device, and medium for improving the segmentation accuracy of gastrointestinal stromal tumors.

[0005] In one aspect, an embodiment of the present application provides a method for segmenting gastrointestinal stromal tumors, comprising the following steps: Acquiring a stomach ultrasound image, wherein the stomach ultrasound image includes a gastrointestinal stromal tumor region; Inputting the stomach ultrasound image into a pre-trained segmentation model to obtain a contour map of the gastrointestinal stromal tumor area; Wherein, the segmentation model includes: A spatial domain processing module, configured to extract spatial domain features from the gastric ultrasound image to obtain low-level features and spatial features; A frequency domain processing module, configured to extract frequency domain features based on the spatial features to obtain frequency domain features; A fusion module, configured to fuse the spatial features and the frequency domain features to obtain fused features; An output module is used to perform segmentation processing based on the low-level features and the fusion features to obtain a contour map of the gastrointestinal stromal tumor area.

[0006] Optionally, in one embodiment, the spatial features include a first spatial feature and a second spatial feature; the extracting spatial domain features from the gastric ultrasound image to obtain low-level features and spatial features includes: extracting spatial domain features from the gastric ultrasound image through a deep convolutional neural network, and extracting the first spatial feature from the deep convolutional neural network. The convolution features output by the convolution layer are determined as the low-level features, and the first convolutional neural network in the deep convolutional neural network is used as the low-level features. The convolution feature output by the convolution layer is determined as the first spatial feature. ; Perform spatial domain feature extraction on the first spatial feature through the hollow spatial pyramid pooling module to obtain the second spatial feature.

[0007] Optionally, in one embodiment, the spatial feature includes a first spatial feature; the frequency domain feature extraction based on the spatial feature to obtain the frequency domain feature includes: performing a two-dimensional discrete Fourier transform on the first spatial feature to obtain a spectrum feature; performing a centering operation on the spectrum feature to obtain the centralized spectrum feature; performing convolution optimization on the centralized spectrum feature to obtain the optimized spectrum feature; and performing frequency domain enhancement on the optimized spectrum feature to obtain the frequency domain feature.

[0008] Optionally, in one embodiment, the spatial feature includes a second spatial feature; and the fusing of the spatial feature and the frequency domain feature to obtain the fused feature includes: using the second spatial feature as a query and the frequency domain feature as a key and value, performing an attention operation on the frequency domain feature and the second spatial feature to obtain the fused feature.

[0009] Optionally, in one embodiment, the segmentation processing based on the low-level features and the fusion features to obtain a contour map of the gastrointestinal stromal tumor area includes: decoding the low-level features and the fusion features to obtain a probability value of each pixel in the gastric ultrasound image belonging to a foreground category and a background category as a probability set of the gastric ultrasound image; and segmenting the target area and the background based on the probability set to obtain a contour map of the gastrointestinal stromal tumor area.

[0010] Optionally, in one embodiment, the segmentation of the target area and the background based on the probability set to obtain a contour map of the gastrointestinal stromal tumor area includes: clustering each pixel of the gastric ultrasound image based on the probability set to obtain pixels belonging to the foreground category and pixels belonging to the background category; determining the numerical value of the pixels belonging to the foreground category to be one, and determining the numerical value of the pixels belonging to the background category to be zero, to obtain a contour map of the gastrointestinal stromal tumor area.

[0011] Optionally, in one embodiment, the method further includes the following steps: during the training of the segmentation model, updating the parameters of the segmentation model using a combined loss function; wherein the combined loss function includes a cross entropy loss function, a Dice loss function and a Hausdorff distance loss function.

[0012] Optionally, in one embodiment, the method further comprises the following steps: preprocessing the stomach ultrasound image; wherein the preprocessing includes size adjustment, resampling, grayscale, channel adjustment and standardization.

[0013] On the other hand, an embodiment of the present application provides a gastrointestinal stromal tumor segmentation device, comprising: an acquisition module, configured to acquire a stomach ultrasound image, wherein the stomach ultrasound image includes a gastrointestinal stromal tumor region; a segmentation module, configured to input the stomach ultrasound image into a pre-trained segmentation model to obtain a contour map of the gastrointestinal stromal tumor region; Wherein, the segmentation model includes: A spatial domain processing module, configured to extract spatial domain features from the gastric ultrasound image to obtain low-level features and spatial features; A frequency domain processing module, configured to extract frequency domain features based on the spatial features to obtain frequency domain features; A fusion module, configured to fuse the spatial features and the frequency domain features to obtain fused features; An output module is used to perform segmentation processing based on the low-level features and the fusion features to obtain a contour map of the gastrointestinal stromal tumor area.

[0014] On the other hand, an embodiment of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are used to enable a computer to execute the above-mentioned gastrointestinal stromal tumor segmentation method.

[0015] According to a gastrointestinal stromal tumor segmentation method, device, and medium provided in an embodiment of the present application, a stomach ultrasound image is acquired, the stomach ultrasound image including the gastrointestinal stromal tumor region; the stomach ultrasound image is input into a pre-trained segmentation model to obtain a contour map of the gastrointestinal stromal tumor region; in the segmentation model, spatial domain features are extracted from the stomach ultrasound image through a spatial domain processing module to obtain low-level features and spatial features; frequency domain features are extracted based on spatial features through a frequency domain processing module to obtain frequency domain features; spatial features and frequency domain features are fused through a fusion module to obtain fused features; segmentation processing is performed through an output module based on the low-level features and fused features to obtain a contour map of the gastrointestinal stromal tumor region. According to the technical solution of the embodiment of the present application, the segmentation accuracy of gastrointestinal stromal tumors can be effectively improved.

[0016] Other features and advantages of the present application will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present application. The purposes and other advantages of the present application can be achieved and obtained through the structures particularly pointed out in the description, claims and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1A This is an example diagram of a gastrointestinal stromal tumor region of a stomach ultrasound image provided in this application; Figure 1B is another example diagram of a gastrointestinal stromal tumor region of a stomach ultrasound image provided in this application; Figure 2 This is a flow chart of a gastrointestinal stromal tumor segmentation method provided in this application; Figure 3 It is a structural diagram of the segmentation model provided by this application; Figure 4 This is an example diagram of DCNN provided by this application. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0019] The present application is further described below in conjunction with the accompanying drawings and specific embodiments. The described embodiments should not be considered as limiting the present application. All other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0020] In the following description, reference is made to “some embodiments”, which describes a subset of all possible embodiments, but it will be understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this application pertains. The terms used herein are for the purpose of describing the embodiments of this application only and are not intended to limit this application.

[0022] Gastrointestinal stromal tumors (GISTs) are a common type of tumor originating from the mesenchymal tissue of the gastrointestinal tract. Their early clinical symptoms are often atypical, and diagnosis usually relies on imaging examinations. In clinical practice, ultrasound examinations have become an important means of screening and diagnosing GISTs due to their non-invasive and real-time advantages. Compared with common medical image segmentation tasks such as fundus blood vessels and brain tumors, the automated segmentation of GISTs in ultrasound images still faces significant challenges, mainly reflected in the following two aspects: First, the boundaries of gastrointestinal stromal tumors are often unclear. Figure 1AAs shown in the figure, although the boundaries of GISTs can sometimes be imaged relatively clearly in computed tomography (CT) or magnetic resonance imaging (MRI) images, in ultrasound images, due to the limitations of the imaging mechanism and significant noise interference, the edges of GIST lesions usually show characteristics such as close grayscale and blurred structure, making them difficult to distinguish from surrounding tissues.

[0023] Second, the size and location of GIST lesions vary greatly. Figure 1B As shown in Figure 2, GISTs vary widely in size, ranging from small lesions of a few millimeters to large masses of tens of centimeters. Furthermore, GISTs can occur throughout the gastrointestinal tract, including the stomach, intestines, omentum, mesentery, and even the retroperitoneum, resulting in a high degree of uncertainty.

[0024] In recent years, deep learning has become the mainstream technology for medical image segmentation. Common representatives include fully convolutional neural networks, U-Net, and multi-scale structures that introduce dilated convolutions. These methods perform well in image segmentation tasks with clear images and clear structures. Therefore, related technologies use existing image segmentation models to perform the task of gastrointestinal stromal tumor segmentation. However, as mentioned above, gastrointestinal stromal tumors have characteristics such as blurred boundaries, uneven scales, and variable positions in ultrasound images. Existing image segmentation models still suffer from problems such as inaccurate edge segmentation and missed detection of small lesions when faced with such images. The reason is that most of these models are based on spatial domain convolution operations and rely only on local relationships between pixels for feature extraction. When the grayscale changes at the boundaries are not obvious or there is a lack of clear texture, the performance of these models will significantly decline.

[0025] To this end, embodiments of the present application provide a method, apparatus, and medium for segmenting gastrointestinal stromal tumors, aiming to improve the segmentation accuracy of gastrointestinal stromal tumors.

[0026] It should be noted that in each specific embodiment of the present disclosure, when it involves the need to perform relevant processing based on relevant data such as gastric ultrasound images, the subject's permission or consent will be obtained first, and the collection, use and processing of such data will comply with relevant laws, regulations and standards. In addition, when the embodiment of the present disclosure needs to obtain relevant data such as gastric ultrasound images, the subject's separate permission or consent will be obtained through a pop-up window or by jumping to a confirmation page. After clearly obtaining the subject's separate permission or consent, the necessary relevant data such as gastric ultrasound images for the normal operation of the embodiment of the present disclosure will be obtained.

[0027] It should be emphasized that the gastric ultrasound images (i.e., the images to be segmented) used in the embodiments of this application are pre-stored image data in a database, not image data acquired in real time from a patient. That is, the embodiments of this application do not target living human subjects. Furthermore, the final output of the embodiments of this application is the lesion region (i.e., the GIST region) in the gastric ultrasound image, which is provided to physicians for diagnostic reference and does not constitute a final diagnosis. The final diagnosis of GIST still requires the physician's judgment and decision. In other words, the embodiments of this application do not directly aim to obtain a GIST diagnosis. Therefore, the embodiments of this application do not constitute a method for diagnosing or treating a disease.

[0028] First, a gastrointestinal stromal tumor segmentation method provided by an embodiment of the present application will be described in detail below with reference to the accompanying drawings.

[0029] The gastrointestinal stromal tumor segmentation method provided in the embodiments of the present application can be applied in a terminal or a server, or can be software running on a terminal or server. The terminal can be, but is not limited to, a tablet computer, laptop computer, or desktop computer. The server can be a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. Furthermore, the server can be, but is not limited to, a node server in a blockchain network. Blockchain is a novel application model for computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms.

[0030] Reference Figure 2 and Figure 3 , Figure 2 This is a flow chart of a gastrointestinal stromal tumor segmentation method provided in this application. Figure 3 is a schematic diagram of the segmentation model provided in this application. The gastrointestinal stromal tumor segmentation method may include the following steps S101-S102: S101, acquiring a stomach ultrasound image, where the stomach ultrasound image includes a gastrointestinal stromal tumor area; S102, inputting the stomach ultrasound image into a pre-trained segmentation model to obtain a contour map of the gastrointestinal stromal tumor area; Among them, the segmentation model includes a spatial domain processing module, a frequency domain processing module, a fusion module and an output module. The spatial domain processing module is used to extract spatial domain features of gastric ultrasound images to obtain low-level features and spatial features; the frequency domain processing module is used to extract frequency domain features based on spatial features to obtain frequency domain features; the fusion module is used to fuse spatial features and frequency domain features to obtain fused features; the output module is used to perform segmentation processing based on low-level features and fused features to obtain a contour map of the gastrointestinal stromal tumor area.

[0031] In an embodiment of the present application, first, a gastric ultrasound image is obtained through a preset database, which pre-stores a plurality of gastric ultrasound images. The gastric ultrasound image refers to an ultrasound image to be measured and associated with the stomach, which includes a gastrointestinal stromal tumor region, that is, a lesion region to be segmented. Then, the gastric ultrasound image is input into a pre-trained segmentation model. The segmentation model refers to a neural network model trained by using a plurality of preset gastric ultrasound image samples and label information corresponding to each gastric ultrasound image sample, and the label information refers to a contour map of the gastrointestinal stromal tumor region in the gastric ultrasound image sample. The segmentation model can be used to segment the gastrointestinal stromal tumor region from the gastric ultrasound image to obtain a contour map, which is a binary mask map.

[0032] Specifically, in the segmentation model, the gastric ultrasound image is input into the spatial domain processing module, which performs spatial domain feature extraction on the input, obtaining low-level features and spatial features. Low-level features are simple, low-level features, such as the distinct edges, outlines, and colors of GISTs, while spatial features are high-level spatial domain features, such as the semantic and spatial domain information of GISTs. The goal here is to extract multi-scale, multi-level GIST features from a spatial domain perspective, capturing both the semantic and structural information of gastric ultrasound images.

[0033] Next, the spatial features are input into the frequency domain processing module, which extracts the frequency domain features from the input to obtain frequency domain features. Frequency domain features characterize the detailed edge information of GISTs in gastric ultrasound images. The goal here is to further extract detailed GIST features, particularly edge information, from a frequency domain perspective, building on the spatial domain feature extraction, thereby enhancing its ability to express potential edges and other aspects.

[0034] Afterwards, the frequency domain features and spatial features are input into the fusion module together, and the feature information of these two dimensions is fused into fusion features through the fusion module. The purpose is to integrate the expression of gastrointestinal stromal tumor in the spatial domain and the expression in the frequency domain, so that the information of the two feature domains complement each other, while retaining the relevant semantic information of gastrointestinal stromal tumor and enabling the segmentation model to have stronger edge recognition ability.

[0035] Finally, the low-level features and fused features are fed into the output module, which performs segmentation processing on the input to produce a contour map of the GIST region. The goal here is to integrate simple low-level features with complex fused features and perform segmentation operations to extract more valuable GIST features and, based on these, restore the GIST region in gastric ultrasound images.

[0036] In summary, the embodiment of the present application introduces a frequency domain feature extraction operation based on the spatial domain convolution operation, and combines it with the decoding operation of the output module to realize the segmentation processing of the gastrointestinal stromal tumor area in the gastric ultrasound image. It can accurately capture the gastrointestinal stromal tumor features of various scales, various positions, and blurred boundaries in the gastric ultrasound image, enhance the segmentation model's ability to model the edges and structural details of the gastrointestinal stromal tumor in the gastric ultrasound image, thereby effectively improving the segmentation accuracy of the gastrointestinal stromal tumor.

[0037] The specific implementation of the above segmentation model will be further explained below.

[0038] In some embodiments, reference Figure 3 and Figure 4 , Figure 4 is a structural diagram of the DCNN provided in this application. The above-mentioned spatial features may include a first spatial feature and a second spatial feature. When the above-mentioned spatial domain processing module is used to extract spatial domain features from a gastric ultrasound image to obtain low-level features and spatial features, it can be specifically used to perform the following operations: The spatial domain features of gastric ultrasound images are extracted by deep convolutional neural network. The convolution features output by the convolution layer are determined as low-level features, and the first The convolution feature output by the convolution layer is determined as the first spatial feature. ; The spatial domain feature is extracted from the first spatial feature through the void spatial pyramid pooling module to obtain the second spatial feature.

[0039] In this embodiment, spatial domain feature extraction can be divided into the following two processes with a progressive relationship: The first step is to introduce a deep convolutional neural network (DCNN) to extract spatial features from gastric ultrasound images. Figure 4As shown, a DCNN can include multiple sequentially connected convolutional layers (Conv), each of which performs a convolution operation on the input data to achieve feature extraction. The convolution kernels of each convolution layer can be the same or different, and the same applies to their step sizes. These parameters can be flexibly set based on actual conditions. Furthermore, the number of convolutional layers in a DCNN can also be flexibly set based on actual conditions, for example, six convolutional layers can be used, but this is not limited to this.

[0040] In DCNN, the depth of convolution features output by different convolution layers is different. Shallow convolution layers can capture simple structural features. As the number of network layers increases, deeper convolution layers can capture higher-scale gastrointestinal stromal tumor features. Therefore, DCNN is divided into shallow and deep layers. The shallow layer includes the first convolution layer to the second convolution layer. layers of convolutional layers, the deep layers include Convolutional layers up to layer of convolutional layers, is the number of convolutional layers in DCNN, and the last convolutional layer of the shallow layer (i.e. The convolution features output by the first convolution layer are determined as low-level features, and the last convolution layer (i.e., the The convolution feature output by the convolution layer is determined as the first spatial feature. and It can be set flexibly according to actual conditions, for example is 2, is 6, but not limited thereto.

[0041] In this way, shallow convolutional layers can extract low-level GIST features in gastric ultrasound images, such as edges, contours, and colors. As the number of network layers in the DCNN increases, deeper convolutional layers can gradually capture higher-level semantic information and spatial features in gastric ultrasound images that are highly relevant to GIST. This approach conforms to the hierarchical structure of GIST features from simple to complex, allowing the extraction of spatial feature information (i.e., spatial features) of GIST and capturing semantic and structural information in gastric ultrasound images that is highly relevant to GIST.

[0042] In the second step, the Atrous Spatial Pyramid Pooling (ASPP) module is introduced to further extract spatial features from the first spatial features. The ASPP module applies convolution operations in parallel on multiple convolution kernels with different dilation rates to further extract multi-scale contextual information from the first spatial features. This allows the ASPP module to capture more abstract semantic information and spatial features in gastric ultrasound images that are highly relevant to GIST. This represents global information, contains less detailed information, and has a deeper hierarchy than the first spatial features output by DCNN.

[0043] In this way, through the convolution operation of the ASPP module at different expansion rates, it is possible to obtain multi-scale feature representations of gastrointestinal stromal tumors, especially abstract semantic information and concrete spatial features, thereby improving the segmentation model's sensitivity to gastrointestinal stromal tumors of different sizes and adapting to the large scale variations of gastrointestinal stromal tumor lesions.

[0044] It can be seen that this embodiment first performs hierarchical feature extraction through DCNN to obtain low-level features and first spatial features. The low-level features indicate structural information in gastric ultrasound images that is highly correlated with gastrointestinal stromal tumors, while the first spatial features indicate semantic information and spatial features in gastric ultrasound images that are highly correlated with gastrointestinal stromal tumors. Subsequently, the ASPP module is used to further extract abstract semantic information and concrete spatial features of gastrointestinal stromal tumors from the first spatial features, thereby enhancing the semantic expression of gastrointestinal stromal tumor features and improving the segmentation model's sensitivity to gastrointestinal stromal tumors of different sizes. In this way, gastrointestinal stromal tumor features of various scales, various positions, and blurred boundaries in gastric ultrasound images can be accurately captured from a spatial domain perspective, and the segmentation model's ability to model the edges and structural details of gastrointestinal stromal tumors in gastric ultrasound images can be enhanced, thereby effectively improving the segmentation accuracy of gastrointestinal stromal tumors.

[0045] In some embodiments, reference Figure 3 The spatial feature may include a first spatial feature; when the frequency domain processing module is used to extract the frequency domain feature based on the spatial feature to obtain the frequency domain feature, it may be specifically used to perform the following operations: Performing a two-dimensional discrete Fourier transform on the first spatial feature to obtain a spectrum feature; Perform a centralization operation on the spectrum features to obtain the centralized spectrum features; Perform convolution optimization on the centralized spectrum features to obtain optimized spectrum features; The optimized spectrum features are enhanced in the frequency domain to obtain frequency domain features.

[0046] In this embodiment, a two-dimensional discrete Fourier transform (2D-DFT) is first performed on the first spatial feature to convert it from the spatial domain to the frequency domain, thereby obtaining a complex spectral feature. 2D-DFT is a prior art and will not be described in detail here. It should be noted that compared to the second spatial feature, the first spatial feature has more detailed information (local information), and frequency domain processing relies on this detailed information to be implemented. Therefore, the first spatial feature is selected as the input for frequency domain processing.

[0047] The spectral features are then centered to obtain a centered spectral feature. Specifically, the amplitude of the spectral feature is calculated and logarithmically compressed to obtain a compressed spectral feature, thereby stabilizing the dynamic range of the spectral feature. The zero-frequency component of the compressed spectral feature is then moved from four angles to the center of the spectral feature through spectral centering to generate a centered spectral feature. This centered spectral feature characterizes the frequency composition information of the gastric ultrasound image, effectively capturing the frequency condition of the gastric ultrasound image.

[0048] Afterwards, convolution optimization is performed on the centralized spectral features to obtain optimized spectral features, thereby effectively enhancing the expression of gastrointestinal stromal tumor features in the frequency domain. The specific operations here are as follows: 1. Channel Adjustment: Channel adjustment is performed on the centered spectral features to obtain channel-adjusted spectral features. For example, a 1×1 convolutional layer is used to adjust the channels of the centered spectral features, but this is not limited to this. This operation aims to remove redundant information in the frequency domain and effectively focus on key structural information in the frequency domain.

[0049] 2. Normalization and Activation: The spectral features after channel adjustment are normalized and nonlinearly activated to obtain the activated spectral features. This operation is intended to ensure the stability of the distribution of frequency domain feature information, so that the segmentation model can focus on the relative distribution pattern and structural characteristics of the spectrum, which are highly correlated with the gastrointestinal stromal tumor region, and reduce the focus on the absolute intensity in the spectrum, thereby improving the generalization ability of the segmentation model when segmenting the gastrointestinal stromal tumor region. Among them, the activation function can be flexibly set according to the actual situation, for example, it can be a ReLU function, a Sigmoid function, etc., but is not limited to this.

[0050] 3. Noise Suppression: De-noise the activated spectral features to obtain denoised spectral features. For example, a 3×3 pooling layer is used to de-noise the activated spectral features. The pooling layer can perform average pooling, but is not limited to this. This operation aims to smooth out high-frequency noise in the frequency domain information, allowing the segmentation model to focus more on the essential frequency domain structure of the gastric ultrasound image.

[0051] Fourth, regularization: Regularize the denoised spectral features to obtain optimized spectral features. For example, a dropout layer can be applied to randomly inactivate some neurons to achieve regularization, but this is not limited to this. This operation can effectively prevent overfitting of the segmentation model and further improve its generalization ability on unseen gastric ultrasound image data.

[0052] Finally, the optimized spectrum features are enhanced in the frequency domain to obtain the frequency domain features, as follows: 1. Construct a two-dimensional Gaussian high-pass mask: First, dynamically generate a Gaussian high-pass filter mask based on the size of the optimized spectral features. The specific method is to create equally spaced coordinate axes from -0.5 to 0.5 to form a coordinate plane ,Right now and are equally spaced from -0.5 to 0.5, and are based on Generate a Gaussian mask where , This is a hyperparameter whose value can be flexibly set based on the actual situation, for example, 0.2, but is not limited to this. This mask generation exploits the properties of a two-dimensional Gaussian function, which behaves as a low-pass filter in the frequency domain. Therefore, a Gaussian high-pass filter can be obtained by converting a 1.0-Gaussian low-pass filter.

[0053] Second, suppress low-frequency components while preserving edge details: The optimized spectral features are element-wise multiplied with a two-dimensional Gaussian high-pass mask to obtain high-frequency features. Because the two-dimensional Gaussian high-pass mask suppresses low-frequency regions while preserving high-frequency regions, this element-wise multiplication effectively filters out high-frequency components from the frequency domain feature information. These high-frequency components directly correspond to the edge details of gastrointestinal stromal tumors in gastric ultrasound images.

[0054] 3. Feature Enhancement: To further emphasize and highlight the edge information of gastrointestinal stromal tumors in gastric ultrasound images, the optimized spectral features are superimposed with high-frequency features. Because high-frequency features retain high frequencies and suppress low frequencies, they need to be added to the optimized spectral features to obtain high-frequency enhanced features, namely frequency domain features. This operation significantly increases the weight and contribution of high-frequency components in the overall frequency domain feature information, achieving the purpose of high-frequency enhancement and amplification. In the fused features, the frequency domain patterns related to the edges of gastrointestinal stromal tumors in gastric ultrasound images are strengthened, making them more influential in the subsequent feature fusion and decoding process.

[0055] It can be seen that this embodiment can effectively capture the high-frequency information in the gastric ultrasound image through the above-mentioned frequency domain feature extraction operation, and by enhancing the high-frequency features of the gastric ultrasound image, it makes it easier for the segmentation model to focus on the edge details of the gastrointestinal stromal tumor in the gastric ultrasound image, further extract the detailed boundary features of the gastrointestinal stromal tumor from the frequency domain perspective, and enhance its expression ability in potential edges and other aspects, thereby improving the sensitivity of the segmentation model to the boundaries and morphology of the gastrointestinal stromal tumor, thereby improving the segmentation accuracy of the gastrointestinal stromal tumor.

[0056] In some embodiments, reference Figure 3 , the spatial feature may include a second spatial feature; when the fusion module is used to fuse the spatial feature and the frequency domain feature to obtain the fusion feature, it can be specifically used to perform the following operations: Taking the second spatial feature as the query and the frequency domain feature as the key and value, the attention operation is performed on the frequency domain feature and the second spatial feature to obtain the fused feature.

[0057] In this implementation, in order to fuse spatial and frequency domain information and overcome the introduction of redundant information and noise, an attention mechanism fusion is introduced to achieve selective attention to frequency domain features guided by spatial domain information. The specific approach is as follows: 1. Feature Mapping: Map the second spatial features to queries and the frequency domain features to keys and values. The mapping method can be flexibly configured based on actual conditions. For example, a 1×1 convolution can be used to map the second spatial features to queries and the frequency domain features to keys and values. Another example is the multiplication of the second spatial features with a preset query mapping matrix as the query, the multiplication of the frequency domain features with a preset key mapping matrix as the key, and the multiplication of the frequency domain features with a preset value mapping matrix as the value, but is not limited to these methods.

[0058] Second, attention mechanism fusion: A self-attention mechanism is applied to the frequency domain features and the second spatial features to obtain a fused feature, thereby implementing an attention query operation. The self-attention mechanism achieves a deep fusion between the spatial and frequency domain expressions of gastrointestinal stromal tumors, enabling information from the two feature domains to complement each other. Specifically, the frequency domain information compensates for the spatial domain's lack of ability to express boundary information, and the spatial domain information compensates for the frequency domain's lack of ability to express semantic information. This enables the segmentation model to fully utilize both the semantic information in the spatial domain and the edge details in the frequency domain. Based on the semantic information in the spatial domain, it selectively focuses on the joint boundary regions in the frequency domain. This fused feature preserves the relevant semantic information of gastrointestinal stromal tumors while enabling the segmentation model to have stronger edge recognition capabilities. Optionally, before outputting the fused feature, further feature extraction can be performed on the fused feature to enhance the expression of gastrointestinal stromal tumor features. For example, a 1×1 convolutional layer is used to process the fused feature before outputting it, but this is not limited to this.

[0059] It can be seen that this embodiment fuses the gastrointestinal stromal tumor feature information in the spatial domain and the gastrointestinal stromal tumor feature information in the frequency domain through the attention mechanism. The attention mechanism can automatically learn and emphasize the most relevant areas in the gastric ultrasound image and suppress the interference of irrelevant areas. This can enable the second spatial feature to focus more on the gastrointestinal stromal tumor area and enhance the segmentation model's ability to model the edges and structural details of the gastrointestinal stromal tumor in the gastric ultrasound image. While retaining the relevant semantic information of the gastrointestinal stromal tumor, the segmentation model is enabled to have stronger edge recognition capabilities, thereby effectively improving the segmentation accuracy of the gastrointestinal stromal tumor.

[0060] In some embodiments, reference Figure 3 When the above output module is used to perform segmentation processing based on low-level features and fusion features to obtain a contour map of the gastrointestinal stromal tumor area, it can be used to perform the following operations: Decoding the low-level features and the fused features to obtain a probability value of each pixel in the gastric ultrasound image belonging to a foreground category or a background category as a probability set of the gastric ultrasound image; The target area and background are segmented based on the probability set to obtain the contour map of the gastrointestinal stromal tumor area.

[0061] In this embodiment, first, a convolution operation is performed on the low-level features to enhance the expression of the low-level features, and the low-level features after convolution are obtained, and the low-level features after convolution and the fusion features output by the fusion module are fused into the target segmentation features, thereby realizing the integration of the input before segmentation. Among them, the convolution operation can be set to convolution, batch normalization and activation in sequence, and the kernel size and activation function of the convolution can be flexibly set according to the actual situation. For example, the convolution kernel is 1×1, and the activation function uses the ReLU function, but it is not limited to this. Here, the low-level features retain simple structural information such as edges, contours, and colors, while the fusion features retain the relevant semantic information and detailed edge information of gastrointestinal stromal tumors. By fusing these two features, the multi-dimensional gastrointestinal stromal tumor features can be effectively integrated, thereby further capturing more valuable gastrointestinal stromal tumor features in gastric ultrasound images, and enhancing the segmentation model's ability to model the edges and structural details of gastrointestinal stromal tumors in gastric ultrasound images.

[0062] Next, DeepLabV3+'s multi-level decoder decodes the target segmentation features, obtaining the probability of each pixel in the gastric ultrasound image belonging to the foreground or background category, and integrating these probability values ​​into a probability set for the gastric ultrasound image. The decoder's role here is to gradually upsample the target segmentation features to the same spatial resolution as the input gastric ultrasound image. Through decoding, the foreground or background value of each pixel can be determined based on the target segmentation features, thereby facilitating the subsequent generation of segmentation results.

[0063] It should be understood that the multi-level decoder of DeepLabV3+ is an existing decoder, which generally includes a residual block, a normalization layer and an upsampling layer. The original DeepLabV3+ will fuse low-level features and fusion features. The fusion method is the same as the previous step, which will not be repeated here.

[0064] Finally, since the decoder outputs the pixel value of each pixel belonging to the foreground or background, in order to achieve more adaptive binarization of the segmentation results, it is necessary to segment the target area and the background based on a probability set, thereby obtaining a contour map of the gastrointestinal stromal tumor area, which is the final segmentation result, thereby improving the segmentation accuracy of gastrointestinal stromal tumors.

[0065] In some embodiments, segmenting the target region and the background based on the probability set to obtain a contour map of the gastrointestinal stromal tumor region may include: Based on the probability set, each pixel of the gastric ultrasound image is clustered to obtain pixels belonging to the foreground category and pixels belonging to the background category; The values ​​of pixels belonging to the foreground category are determined to be one, and the values ​​of pixels belonging to the background category are determined to be zero, thereby obtaining a contour map of the gastrointestinal stromal tumor region.

[0066] In this embodiment, in order to achieve more adaptive segmentation result binarization, a clustering algorithm is introduced to overcome the problem of unstable results caused by fixed thresholds. The specific approach is as follows: First, for each pixel of the stomach ultrasound image, its two-channel probability value is extracted, which includes the foreground probability and background probability , thus forming a two-dimensional vector . By traversing multiple pixels of the gastric ultrasound image, a two-dimensional vector of multiple pixels in the gastric ultrasound image can be obtained. Then, the two-dimensional vector of multiple pixels in the gastric ultrasound image is used as the input of the K-means clustering algorithm, and clustering is performed through the K-means clustering algorithm, wherein the number of categories of the K-means clustering algorithm is set to 2, corresponding to the foreground category and the background category respectively. After the clustering is completed, each pixel of the gastric ultrasound image is divided into the foreground category or the background category, thereby obtaining pixels belonging to the foreground category and pixels belonging to the background category. Finally, according to the features (pixels) of the two cluster centers, the clusters are automatically identified, the values ​​of the pixels belonging to the foreground category are determined to be one, and the values ​​of the pixels belonging to the background category are determined to be zero, thereby generating a contour map of the gastrointestinal stromal tumor area in a binary form.

[0067] In some embodiments, the above method may further include the following steps: During the training of the segmentation model, the parameters of the segmentation model are updated using a combination loss function; wherein the combination loss function may include a cross entropy loss function, a Dice loss function, and a Hausdorff distance loss function.

[0068] In this embodiment, during the training of the segmentation model, a combined loss function (CombinedLoss) is used to update the parameters of the segmentation model. The update method can be flexibly set according to the actual situation, for example, using the gradient descent method to achieve the update, but is not limited to this. The combined loss function here can include the following: 1. Cross Entropy Loss: This function measures the pixel-level classification accuracy of the segmentation model in the GIST segmentation task. It encourages the segmentation model to correctly predict the category of each pixel in the gastric ultrasound image. Category prediction refers to identifying whether it is foreground or background.

[0069] 2. Dice Loss Function (DiceLoss): Solve the category imbalance problem in the gastrointestinal stromal tumor segmentation task and directly optimize the overlap.

[0070] 3. Hausdorff Distance Loss: constrains the maximum error between the predicted boundary and the true boundary, thereby improving boundary accuracy.

[0071] During each training round, the segmentation model is first forward-propagated using the gastric ultrasound image samples from the current training round to obtain the segmentation results (i.e., predictions) for the gastric ultrasound image samples in the current training round. Then, the segmentation results and label information of the gastric ultrasound image samples from the current training round are used to calculate the specific values ​​of the three loss functions described above. Finally, the specific values ​​of the three loss functions are weighted and summed to obtain the total loss function of the segmentation model in the current training round, and the parameters of the segmentation model are updated accordingly. The weights of the three loss functions can be flexibly set based on actual conditions. For example, a weight of 1 for the cross-entropy loss function, a weight of 2 for the Dice loss function, and a weight of 0.0005 for the Hausdorff distance loss function are suitable, but are not limited to these.

[0072] In this way, the above-mentioned combined loss function can optimize the segmentation performance of the segmentation model from multiple dimensions, especially dealing with class imbalance and boundary details, and comprehensively optimizing the segmentation accuracy and boundary matching of the segmentation model.

[0073] In some embodiments, the above method may further include the following steps: The gastric ultrasound image is preprocessed; wherein the preprocessing includes size adjustment, resampling, grayscale, channel adjustment and standardization.

[0074] In this embodiment, before inputting the gastric ultrasound image into the segmentation model, the gastric ultrasound image can be preprocessed to ensure its quality and to ensure that its size conforms to the input size of the segmentation model, thereby ensuring the segmentation accuracy of the segmentation model. Specifically, the gastric ultrasound image is first resized to a specified resolution and resampled using bilinear interpolation. Then, the gastric ultrasound image is converted to a grayscale image and three channels are replicated to meet the channel number requirements of the segmentation model. After the gastric ultrasound image is converted into a tensor, it is normalized to obtain a standardized gastric ultrasound image as input to the segmentation model. The specified resolution, normalized mean, and normalized standard deviation can all be flexibly set according to actual conditions. For example, the mean used is [0.485, 0.456, 0.406] and the standard deviation is [0.229, 0.224, 0.225], but this is not limited to these.

[0075] In some embodiments, the above method may further include the following steps: The gastric ultrasound image samples used to train the segmentation model are preprocessed; the preprocessing includes data augmentation, size adjustment, resampling, grayscale conversion, channel adjustment, and normalization.

[0076] In this embodiment, when training the segmentation model, in order to ensure data diversity, the gastric ultrasound image samples are first enhanced, which includes random rotation (for example, within a range of ±15°), horizontal flipping, vertical flipping and other spatial transformation operations. At the same time, elastic deformation, Gaussian noise and brightness perturbation are introduced to simulate the changing characteristics of gastric ultrasound images under different acquisition conditions. The enhanced gastric ultrasound image samples are then resized, resampled, grayscaled, channel adjusted and normalized. For details, please refer to the description of the preprocessing of gastric ultrasound images in the aforementioned embodiment, which will not be repeated here. In this way, the quality of the gastric ultrasound image samples can be improved, and the size of the gastric ultrasound image samples can be guaranteed to meet the input size of the segmentation model, thereby ensuring the training accuracy of the segmentation accuracy.

[0077] To facilitate understanding of the above-mentioned gastrointestinal stromal tumor segmentation method of the present application, the actual application scenario of the above-mentioned gastrointestinal stromal tumor segmentation method of the present application is used as an example.

[0078] This application scenario includes two aspects: model principles, model training, and real-time application, as follows: 1. Model Principle: This application example constructs a DeepLabV3+ segmentation model that integrates frequency domain information. Its core approach is to combine semantic representation in the spatial domain with edge perception in the frequency domain to enhance the recognition of GIST edges in gastric ultrasound images, thereby effectively improving GIST segmentation accuracy. This combined architecture, through the introduction of a specially designed frequency domain processing module and a cross-modal attention fusion mechanism, aims to overcome the limitations of traditional methods in addressing the fuzzy boundaries and multi-scale lesions of GISTs.

[0079] Reference Figure 3 ,The segmentation model includes spatial domain processing module, frequency domain processing module, fusion module and output module. The specific modules are as follows: The spatial domain processing module is equipped with DCNN and ASPP modules. Figure 4 As shown in the figure, the DCNN includes six convolutional layers. The DCNN is used to extract spatial features from gastric ultrasound images. The output of the second convolutional layer is determined as the low-level feature, and the output of the sixth convolutional layer is determined as the first spatial feature. The ASPP module is used to extract spatial features from the first spatial feature to obtain the second spatial feature.

[0080] The frequency domain processing module is used to perform the following operations: first, perform 2D-DFT on the first spatial feature output by DCNN to obtain a spectrum feature, and perform a centering operation on it to obtain a centralized spectrum feature; then, perform convolution optimization on the centralized spectrum feature to obtain an optimized spectrum feature; finally, perform frequency domain enhancement on the optimized spectrum feature to obtain a frequency domain feature.

[0081] The fusion module is used to perform the following operations: using the second spatial feature output by the ASPP module as Q and the frequency domain features as K and V, performing a self-attention operation to obtain a fused feature, and performing a convolution operation before outputting the fused feature.

[0082] The output module is used to perform the following operations: first, convolution operation is performed on the low-level features to obtain the convolved low-level features, and the convolved low-level features and the fusion features output by the fusion module are spliced ​​into the target segmentation features in the channel dimension. Then, the target segmentation features are decoded using the multi-level decoder of DeepLabV3+ to obtain the probability value of each pixel in the gastric ultrasound image belonging to the foreground category or the background category, and clustering is performed based on this probability value combined with the K-Means clustering algorithm to obtain pixels belonging to the foreground category and pixels belonging to the background category. Finally, the numerical value of the pixels belonging to the foreground category is determined to be one, and the numerical value of the pixels belonging to the background category is determined to be zero, and the contour map of the gastrointestinal stromal tumor area is obtained.

[0083] 2. Model training: The first step is training initialization: fix the random seeds of Python, NumPy, PyTorch, etc. to ensure the reproducibility of the experiment; automatically detect the training device and give priority to the Graphics Processing Unit (GPU).

[0084] The second step, data loading, involves obtaining multiple preset gastric ultrasound image samples and the corresponding label information for each gastric ultrasound image sample. The label information refers to the outline of the gastrointestinal stromal tumor area in the gastric ultrasound image sample, and constructing a dataset based on this. The dataset is divided into a training set and a test set.

[0085] The third step is the design of the loss function: a combined loss function is used, including the cross entropy loss function, the Dice loss function and the Hausdorff distance loss function. The weight value of the cross entropy loss function is 1, the weight value of the Dice loss function is 2, and the weight value of the Hausdorff distance loss function is 0.0005.

[0086] Step 4: Optimizer and training schedule: Use RMSprop optimizer with a learning rate of , momentum is 0, and weight decay is Each training cycle consists of two phases: training and validation. During the training phase, the CutMix data augmentation strategy is used (with a probability of 0.5). The learning rate is dynamically adjusted using a cosine annealing strategy. After each round of validation, if the segmentation model performs better than the best historical result, the current model weights are saved.

[0087] The trained segmentation model is obtained through the above training and applied to real-time applications.

[0088] 3. Real-time application: The stomach ultrasound image to be tested is obtained through a preset database and input into the trained segmentation model, and the contour map of the gastrointestinal stromal tumor area in the stomach ultrasound image is obtained through segmentation of the segmentation model.

[0089] In addition, an embodiment of the present application provides a gastrointestinal stromal tumor segmentation device, which may include: an acquisition module, configured to acquire a stomach ultrasound image, wherein the stomach ultrasound image includes a gastrointestinal stromal tumor region; A segmentation module is used to input the gastric ultrasound image into a pre-trained segmentation model to obtain a contour map of the gastrointestinal stromal tumor area; Among them, the segmentation model includes a spatial domain processing module, a frequency domain processing module, a fusion module and an output module. The spatial domain processing module is used to extract spatial domain features of gastric ultrasound images to obtain low-level features and spatial features; the frequency domain processing module is used to extract frequency domain features based on spatial features to obtain frequency domain features; the fusion module is used to fuse spatial features and frequency domain features to obtain fused features; the output module is used to perform segmentation processing based on low-level features and fused features to obtain a contour map of the gastrointestinal stromal tumor area.

[0090] The contents of the above method embodiments are all applicable to the present device embodiments. The functions specifically implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0091] Finally, an embodiment of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are used to enable a computer to execute the above-mentioned gastrointestinal stromal tumor segmentation method.

[0092] The contents of the above method embodiments are all applicable to the present storage medium embodiment. The functions specifically implemented by the present storage medium embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0093] In some optional embodiments, the function / operation mentioned in the block diagram may not occur in the order mentioned in the operation diagram. For example, depending on the function / operation involved, the two boxes shown in succession can actually be executed substantially simultaneously or the boxes can sometimes be executed in reverse order. In addition, the embodiments presented and described in the flow chart of the present application are provided in an exemplary manner for the purpose of providing a more comprehensive understanding of the technology. The disclosed method is not limited to the operations and logical flows presented herein. Optional embodiments are contemplated in which the order of the various operations is changed and the sub-operations described as a part of a larger operation are performed independently.

[0094] In addition, although the present application is described in the context of functional modules, it should be understood that, unless otherwise stated, one or more of the functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in separate physical devices or software modules. It is also understood that a detailed discussion of the actual implementation of each module is not necessary for understanding the present application. More specifically, given the properties, functions, and internal relationships of the various functional modules in the devices disclosed herein, the actual implementation of the module will be understood within the routine skills of an engineer. Therefore, a person skilled in the art can implement the present application as set forth in the claims using ordinary techniques without undue experimentation. It is also understood that the specific concepts disclosed are merely illustrative and are not intended to limit the scope of the present application, which is determined by the full scope of the appended claims and their equivalents.

[0095] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several programs for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, and other media that can store program code.

[0096] The logic and / or steps represented in a flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable programs for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, a program execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can retrieve and execute a program from a program execution system, apparatus, or device). For purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, a program execution system, apparatus, or device.

[0097] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or more wires (electronic devices), a portable computer disk cartridge (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium and then editing, interpreting, or processing it in another suitable manner as necessary, and then storing it in a computer memory.

[0098] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable program execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used: a discrete logic circuit having logic gate circuits for implementing logical functions on data signals, an application-specific integrated circuit having suitable combinational logic gate circuits, a programmable gate array (PGA), a field-programmable gate array (FPGA), etc.

[0099] In the above description of this specification, reference to the terms "one embodiment / example," "another embodiment / example," or "certain embodiments / examples" means that the specific features, structures, materials, or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in any appropriate manner in any one or more embodiments or examples.

[0100] Although the embodiments of the present application have been shown and described, those skilled in the art will appreciate that various changes, modifications, substitutions, and variations may be made to the embodiments without departing from the principles and intent of the present application, and that the scope of the present application is defined by the claims and their equivalents.

[0101] The above is a specific description of the preferred implementation of the present application, but the present application is not limited to the embodiments. Those skilled in the art may make various equivalent modifications or substitutions without violating the spirit of the present application. These equivalent modifications or substitutions are all included in the scope defined by the claims of the present application.

Claims

1. A gastrointestinal stromal tumor segmentation method, characterized in that: The following steps are involved: Acquiring a stomach ultrasound image, wherein the stomach ultrasound image includes a gastrointestinal stromal tumor region; Inputting the stomach ultrasound image into a pre-trained segmentation model to obtain a contour map of the gastrointestinal stromal tumor area; Wherein, the segmentation model includes: A spatial domain processing module, configured to extract spatial domain features from the gastric ultrasound image to obtain low-level features and spatial features; A frequency domain processing module, configured to extract frequency domain features based on the spatial features to obtain frequency domain features; A fusion module, configured to fuse the spatial features and the frequency domain features to obtain fused features; An output module is used to perform segmentation processing based on the low-level features and the fusion features to obtain a contour map of the gastrointestinal stromal tumor area.

2. The method according to claim 1, characterized in that The spatial features include a first spatial feature and a second spatial feature; and the spatial domain feature extraction is performed on the stomach ultrasound image to obtain low-level features and spatial features, including: The spatial domain features of the gastric ultrasound image are extracted by a deep convolutional neural network, and the first The convolution features output by the convolution layer are determined as the low-level features, and the first convolutional neural network in the deep convolutional neural network is used as the low-level features. The convolution feature output by the convolution layer is determined as the first spatial feature. ; Performing spatial domain feature extraction on the first spatial feature through a dilated spatial pyramid pooling module to obtain the second spatial feature.

3. The method according to claim 1, characterized in that The spatial feature includes a first spatial feature; and extracting a frequency domain feature based on the spatial feature to obtain a frequency domain feature includes: Performing a two-dimensional discrete Fourier transform on the first spatial feature to obtain a frequency spectrum feature; Performing a centering operation on the spectrum feature to obtain the centralized spectrum feature; Performing convolution optimization on the centralized spectrum characteristics to obtain optimized spectrum characteristics; The optimized frequency spectrum features are enhanced in the frequency domain to obtain the frequency domain features.

4. The method according to claim 1, wherein The spatial feature includes a second spatial feature; and fusing the spatial feature with the frequency domain feature to obtain a fused feature includes: The second spatial feature is used as a query, and the frequency domain feature is used as a key and a value, and an attention operation is performed on the frequency domain feature and the second spatial feature to obtain the fused feature.

5. The method according to claim 1, wherein The segmentation process is performed based on the low-level features and the fusion features to obtain a contour map of the gastrointestinal stromal tumor region, including: Decoding the low-level features and the fused features to obtain a probability value of each pixel in the stomach ultrasound image belonging to a foreground category and a background category as a probability set of the stomach ultrasound image; The target region and the background are segmented based on the probability set to obtain a contour map of the gastrointestinal stromal tumor region.

6. The method according to claim 5, characterized in that The segmenting of the target region and the background based on the probability set to obtain a contour map of the gastrointestinal stromal tumor region includes: Clustering each pixel of the stomach ultrasound image based on the probability set to obtain pixels belonging to the foreground category and pixels belonging to the background category; The values ​​of pixels belonging to the foreground category are determined to be one, and the values ​​of pixels belonging to the background category are determined to be zero, thereby obtaining a contour map of the gastrointestinal stromal tumor region.

7. The method according to claim 1, characterized in that The method further comprises the following steps: During the training of the segmentation model, updating the parameters of the segmentation model using a combined loss function; The combined loss function includes a cross entropy loss function, a Dice loss function, and a Hausdorff distance loss function.

8. The method according to claim 1, characterized in that The method further comprises the following steps: The stomach ultrasound image is preprocessed; wherein the preprocessing includes size adjustment, resampling, grayscale, channel adjustment and standardization.

9. A gastrointestinal stromal tumor segmentation device, characterized in that: include: an acquisition module, configured to acquire a stomach ultrasound image, wherein the stomach ultrasound image includes a gastrointestinal stromal tumor region; a segmentation module, configured to input the stomach ultrasound image into a pre-trained segmentation model to obtain a contour map of the gastrointestinal stromal tumor region; Wherein, the segmentation model includes: A spatial domain processing module, configured to extract spatial domain features from the gastric ultrasound image to obtain low-level features and spatial features; A frequency domain processing module, configured to extract frequency domain features based on the spatial features to obtain frequency domain features; A fusion module, configured to fuse the spatial features and the frequency domain features to obtain fused features; An output module is used to perform segmentation processing based on the low-level features and the fusion features to obtain a contour map of the gastrointestinal stromal tumor area.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are used to enable a computer to execute the gastrointestinal stromal tumor segmentation method according to any one of claims 1 to 8.

Images