Image processing methods

Abstract

This invention provides an image processing method, including acquiring an image to be processed containing a target object; inputting the image to be processed into a convolutional layer of an image processing model to obtain an initial feature map of the image; inputting the initial feature map into a self-attention mechanism layer of an encoder, and obtaining a target feature map corresponding to the initial feature map based on the positional information of each feature in the initial feature map and the positional relationship between each feature and other features; and inputting the target feature map into the decoder for processing to obtain an object segmentation map and object label of the target object in the image to be processed. This method proposes a position-sensitive self-attention mechanism. When predicting the object segmentation map of the target object in the image to be processed, it enhances the prediction accuracy of the model through this non-local interaction based on the positional relationship between each feature in the initial feature map of the image to be processed and other features in the global domain. This method also exhibits high sensitivity when applied to esophageal cancer screening scenarios.

Description

[Technical Field]

[0001] The embodiments of the present invention relate to the field of computer technology, and in particular to an image processing method. [Previous Technology]

[0002] Esophageal cancer is a highly fatal cancer. Early detection of resectable / curable esophageal cancer has a huge potential to reduce mortality. Although some esophageal cancer screening methods have been developed, these screenings are invasive, require anesthesia, or are expensive, making them difficult to apply to the general population, and often fail to achieve satisfactory sensitivity in identifying early-stage cancer. [Summary of the Invention]

[0003] In view of the above, embodiments of the present invention provide an image processing method. One or more embodiments of the present invention simultaneously relate to an image processing apparatus, two esophageal CT image processing methods, two esophageal CT image processing apparatuses, a computing device, a computer-readable storage medium, and a computer program, to address the technical deficiencies existing in the prior art. According to a first embodiment of the present invention, an image processing method is provided, comprising: acquiring an image to be processed containing a target object; inputting the image to be processed into a convolutional layer of an image processing model to obtain an initial feature map of the image to be processed, wherein the image processing model includes an encoder and a decoder; inputting the initial feature map into a self-attention mechanism layer of the encoder; obtaining a target feature map corresponding to the initial feature map based on the positional information of each feature in the initial feature map and the positional relationship between each feature and other features; inputting the target feature map into the decoder for processing to obtain an object segmentation map and object labels for the target object in the image to be processed. According to a second embodiment of the present invention, an image processing apparatus is provided, comprising: a first feature map acquisition module configured to acquire an image to be processed containing a target object, and input the image to be processed into a convolutional layer of an image processing model to obtain an initial feature map of the image to be processed, wherein the image processing model includes an encoder and a decoder; a second feature map acquisition module configured to input the initial feature map into a self-attention mechanism layer of the encoder, and obtain a target feature map corresponding to the initial feature map based on the position information of each feature in the initial feature map and the positional relationship between each feature and other features; and a first feature map processing module configured to input the target feature map into the decoder for processing to obtain an object segmentation map and object label of the target object in the image to be processed. According to a third embodiment of the present invention, a method for processing esophageal CT images is provided, comprising: acquiring a CT image containing the esophagus; inputting the CT image into a convolutional layer of an image processing model to obtain an initial feature map of the CT image, wherein the image processing model includes an encoder and a decoder; inputting the initial feature map into a self-attention mechanism layer of the encoder; obtaining a target feature map corresponding to the initial feature map based on the positional information of each feature in the initial feature map and the positional relationship between each feature and other features; and inputting the target feature map into the decoder for processing to obtain a segmentation map of the esophagus and pathological labels in the CT image.According to a fourth embodiment of the present invention, an esophageal CT image processing apparatus is provided, comprising: a third feature map acquisition module configured to acquire a CT image containing the esophagus, input the CT image into a convolutional layer of an image processing model to obtain an initial feature map of the CT image, wherein the image processing model includes an encoder and a decoder; a fourth feature map acquisition module configured to input the initial feature map into a self-attention mechanism layer of the encoder, and obtain a target feature map corresponding to the initial feature map based on the position information of each feature in the initial feature map and the positional relationship between each feature and other features; and a second feature map processing module configured to input the target feature map into the decoder for processing to obtain a segmentation map of the esophagus and pathological labels in the CT image. According to a fifth embodiment of the present invention, an esophageal CT image processing method is provided, comprising: receiving a user's processing request for an esophageal CT image, wherein the esophageal CT image is an image of the esophagus scanned by a computed tomography (CT) scanner; inputting the esophageal CT image into an image processing model, and, if the user is determined to be a first type of user, outputting and displaying a segmentation diagram of the esophagus and pathological labels in the esophageal CT image to the first type of user; or, if the user is determined to be a second type of user, outputting and displaying pathological labels of the esophageal CT image to the second type of user; wherein the image processing model is the image processing model in the above-described image processing method. According to a sixth embodiment of the present invention, an esophageal CT image processing apparatus is provided, comprising: a request receiving module configured to receive a user's processing request for an esophageal CT image, wherein the esophageal CT image is an image of the esophagus scanned by a computed tomography (CT) scanner; a first output module configured to input the esophageal CT image into an image processing model, and, if the user is determined to be a first type of user, output and display to the first type of user a segmentation diagram of the esophagus and pathological labels in the esophageal CT image; or a second output module configured to, if the user is determined to be a second type of user, output and display to the second type of user a pathological label of the esophageal CT image; wherein the image processing model is the image processing model in the above-described image processing method. According to a seventh embodiment of the present invention, a computing device is provided, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, wherein the computer-executable instructions, when executed by the processor, implement the steps of the above-described image processing method or the esophageal CT image processing method. According to an eighth embodiment of the present invention, a computer-readable storage medium is provided, which stores computer-executable instructions that, when executed by a processor, implement the steps of the above-described image processing method or esophageal CT image processing method.According to a ninth embodiment of the present invention, a computer program is provided, wherein when the computer program is executed on a computer, the computer performs the steps of the above-described image processing method or esophageal CT image processing method. One embodiment of the present invention implements an image processing method, including acquiring an image to be processed containing a target object; inputting the image to be processed into a convolutional layer of an image processing model to obtain an initial feature map of the image to be processed, wherein the image processing model includes an encoder and a decoder; inputting the initial feature map into a self-attention mechanism layer of the encoder; obtaining a target feature map corresponding to the initial feature map based on the positional information of each feature in the initial feature map and the positional relationship between each feature and other features; and inputting the target feature map into the decoder for processing to obtain an object segmentation map and object labels for the target object in the image to be processed. Specifically, this image processing method proposes a novel position-sensitive self-attention mechanism. When predicting the object segmentation map of the target object in the image to be processed, the model enhances the prediction accuracy through non-local interaction based on the positional relationship between each feature in the initial feature map of the image to be processed and other features in the global domain. When applied to the scenario of esophageal cancer screening, this non-invasive method can accurately identify early cancer with high sensitivity.

Implementation Method

[0005] Numerous specific details are set forth in the following description to provide a full understanding of the invention. However, the invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention; therefore, the invention is not limited to the specific embodiments disclosed below. The terminology used in one or more embodiments of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “said,” and “the” used in one or more embodiments of the invention and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” used in one or more embodiments of the invention refers to and includes any or all possible combinations of one or more associated listed items. It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of the invention, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of the invention, and similarly, second may also be referred to as first. Depending on the context, the word “if” as used herein may be interpreted as “when” or “in response to determination.” First, the terms used in one or more embodiments of this invention are explained. nnUNet: This refers to a model trained using proprietary data. nnUNet, as an improvement on uNet, has shown excellent performance in 3D medical imaging. CT: Computed Tomography, which uses precisely collimated X-ray beams, gamma rays, ultrasound, etc., along with highly sensitive detectors, to perform a series of cross-sectional scans around a part of the body. It features fast scanning time and clear images, and can be used to examine various diseases. Esophageal cancer is a highly fatal cancer, with a low 5-year survival rate according to incomplete statistics. However, early detection of resectable / curable esophageal cancer can significantly reduce mortality. Currently, some tools exist for esophageal cancer screening. For example, in high-risk areas, endoscopic techniques, cytological techniques, and blood-based biomarkers are used for esophageal cancer screening. In some areas, the implementation of endoscopy has reduced the mortality rate of esophageal squamous cell carcinoma (ESCC). However, endoscopy requires a high level of expertise, and the false negative rate can be high. Furthermore, due to its invasiveness and high cost, it is difficult to apply to the general population. Cytological techniques can serve as a minimally invasive alternative to endoscopy, but their sensitivity and specificity remain lower than endoscopy. Blood tests offer a wide range of non-invasive examinations, but their sensitivity for detecting early esophageal cancer is only moderate. Studies suggest that blood tests need to be combined with other screening techniques, such as imaging, to optimize results.In this context, there is an urgent need for new non-invasive, low-cost, easy-to-use, and highly accurate esophageal cancer screening methods. Routine CT imaging for other clinical indications provides opportunities for disease screening without additional costs, additional examinations, or radiation exposure for patients. Previous studies have shown that abdominal and chest CT have provided value for osteoporosis screening and cardiovascular disease prediction. In terms of cancer detection, researchers have found that pancreatic cancer can be detected with high accuracy using deep learning on plain CT, which has long been considered impossible (detectable only by contrast-enhanced CT). This ability to detect cancer based on plain CT, coupled with established clinical indications such as lung cancer screening on chest CT, could be considered for population-based screening of asymptomatic adult cancers. However, detecting esophageal cancer on plain CT may be challenging. (1) Pancreatic cancer is primarily adenocarcinoma, and unlike other solid organs, esophageal cancer is primarily squamous cell carcinoma located in a long, hollow organ; therefore, it is difficult to obtain high specificity, which is a necessary condition for an ideal screening tool. (2) Importantly, early esophageal cancer tumors can be very small. Early tumor detection is extremely challenging due to the poor contrast between tumors and normal esophageal tissue in plain CT scans (such as chest CT). In fact, even on enhanced CT, human experts often require significant effort and expertise to detect early esophageal tumors, referencing other clinical information such as endoscopy, endoscopic ultrasound, and FDG-PET; even then, some small tumors remain difficult to detect on CT. To date, research on deep learning-based esophageal cancer image analysis has focused on tumor segmentation, improving segmentation accuracy through enhanced local image feature extraction / modeling or fusion of multimodal imaging. Based on this, this invention proposes a deep learning-based tool for screening esophageal cancer using plain CT. Specifically, if an esophageal tumor is present, it can be classified as either cancerous or benign. As mentioned above, the texture of local images on plain CT may be insufficient for detecting esophageal tumors. In clinical practice, overall esophageal features, such as "asymmetric esophageal wall thickening" and "esophageal wall compression," are key signs for screening esophageal cancer, especially in its early stages. On the other hand, in deep learning, each convolutional kernel can only participate in a local subset of voxels or local patterns, rather than the global context. Therefore, in this embodiment of the invention, a global attention layer is combined with positional embedding to enhance long-term dependency resolution in 3D medical image segmentation. This design can improve tumor identification capabilities, especially for early-stage tumors. Specifically, during training, a sample dataset of 741 patients was collected, including two major types of esophageal tumors (ESCC and leiomyomas) and normal esophagus, for image processing model training.On the validation test set, the image processing model achieved an AUC (Area Under Curve) of 0.990, a sensitivity of 93.0%, and a specificity of 97.5%, exceeding the average sensitivity of 75.0% and specificity of 83.8% achieved by four physicians. Specifically, this invention provides an image processing method. One or more embodiments of this invention simultaneously relate to an image processing device, two esophageal CT image processing methods, two esophageal CT image processing devices, a computing device, a computer-readable storage medium, and a computer program, which will be described in detail in the following embodiments. Referring to Figure 1, Figure 1 illustrates a specific scenario of an image processing method provided according to an embodiment of the present invention applied to esophageal cancer screening. Figure 1 includes a CT scanner 102, a terminal 104, and a server 106. In practice, the CT scanner 102 performs a CT scan on the user undergoing esophageal cancer screening, acquiring the user's esophageal CT image. The terminal 104 acquires the esophageal CT image from the CT scanner 102 and sends it to the server 106. The server 106 inputs the esophageal CT image into a pre-trained image processing model, outputs a segmented esophageal image corresponding to the CT image, and returns the segmented esophageal image to the terminal 104. The user (e.g., a doctor) on the terminal 104 can then assess the user's esophageal condition based on the segmented esophageal image, determining whether the esophagus is normal, benign, or cancerous. The image processing model can be understood as a deep learning model pre-trained using a global self-attention mechanism combined with historical esophageal CT images. Furthermore, the image processing model can not only output the segmented esophageal image but also simultaneously output object labels for the esophageal CT image based on the labels of each segment, such as benign, normal, or cancerous labels. The image processing method provided in this invention is applied to the specific scenario of esophageal cancer screening. It proposes a novel position-sensitive self-attention technique that enhances the nnUNet through non-local interaction by capturing global contextual features, thereby improving the segmentation accuracy of the image processing model. Specifically, the global contextual features of the esophagus are key features for cancer detection, especially early-stage cancer. To capture such global context, this invention proposes a novel position-sensitive self-attention technique to enhance the nnUNet through non-local interaction. In experiments, the image processing model achieved a sensitivity of 93.0% and a specificity of 97.5% in detecting esophageal tumors on 180 patients. In comparison, the average sensitivity and specificity of four physicians were 75.0% and 83.8%, respectively. For classification tasks, the image processing model achieved 17%, 31%, and 14% higher accuracy than the average physician in cancer, benign tumors, and normal tissue, respectively.Compared with existing invasive esophageal cancer screening methods, the image processing method provided in this embodiment of the invention has superior performance and is more sensitive to early-stage cancers and benign tumors. That is, the image processing method provided in this embodiment of the invention is a novel, non-invasive, low-cost, and high-precision opportunistic esophageal cancer screening method. Referring to Figure 2, Figure 2 shows a flowchart of an image processing method provided according to an embodiment of the invention, specifically including the following steps: Step 202: Obtain an image to be processed containing the target object, input the image to be processed into the convolutional layer of an image processing model, and obtain an initial feature map of the image to be processed. The image processing model includes an encoder and a decoder. Specifically, the image processing method provided in this embodiment of the invention can be applied to esophageal cancer screening scenarios, as well as other applicable scenarios, such as scenarios similar to esophageal cancer screening. The specific application scenario of this image processing method varies, and the image to be processed containing the target object also differs; for example, when this image processing method is applied to an esophageal cancer screening scenario, the image to be processed containing the target object can be understood as a CT image containing the esophagus; that is, the image to be processed containing the target object is a CT image containing the esophagus. For ease of understanding, the following embodiments will use a CT image containing the esophagus as an example. Taking a CT image containing the esophagus as an example, the process involves acquiring the image containing the target object, inputting the image into the convolutional layer of an image processing model, and obtaining the initial feature map of the image. This can be understood as acquiring a CT image containing the esophagus, inputting the CT image into the convolutional layer of a pre-trained image processing model, and obtaining the initial feature map of the CT image. In practical applications, the image processing model includes an encoder and a decoder, with the convolutional layer located in the encoder. Therefore, inputting the image into the convolutional layer of the image processing model to obtain the initial feature map of the image can be understood as inputting the CT image into the convolutional layer of the encoder to obtain the initial feature map of the CT image. Furthermore, since the initial feature map of the image to be processed after one convolution has many features, inputting this initial feature map with many features into the self-attention mechanism layer for self-attention operation would result in a large computational burden. On the other hand, if the image to be processed undergoes multiple convolutions, the features in the resulting initial feature map would be smaller, leading to inaccurate computation. Therefore, to avoid the above problems, the image to be processed can be convolved through two convolutional layers in the encoder, and then the subsequent self-attention mechanism layer operation can be performed based on the initial feature map after convolution. That is, the image to be processed is input into the two convolutional layers of the encoder of the image processing model to obtain the initial feature map of the image to be processed. Of course, the specific number of convolutional layers for feature extraction can be set according to the actual application, and this embodiment of the invention does not impose any limitation on this.Before using an image processing model to predict object segmentation maps and object labels for target objects in the image to be processed, the image processing model needs to be pre-trained to ensure the accuracy and stability of the subsequent use of the image processing model. Specifically, the training steps of the image processing model are as follows: 1. Obtain a sample image containing sample objects and determine the sample object segmentation map and sample object labels in the sample image; 2. Input the sample image into the convolutional layer of the encoder to obtain the initial sample feature map of the sample image; 3. Input the initial sample feature map into the self-attention mechanism layer of the encoder, and obtain the target sample feature map corresponding to the initial sample feature map based on the position information of each initial sample feature in the initial sample feature map and the positional relationship between each initial sample feature and other initial sample features; 4. Input the target sample feature map into the decoder for processing to obtain the predicted object segmentation map and predicted object labels for the sample objects in the sample image; 5. Train the image processing model based on the sample object segmentation map, sample object labels, and the loss function formed by the predicted object segmentation map and predicted object labels. Note that the application scenarios of the image processing model vary, and the sample images containing sample objects also differ. Continuing with the previous example, in the context of applying the image processing model to esophageal cancer screening, the sample images for this target can be understood as CT images containing the esophagus. These sample images are determined from historical 3D plain CT images of patients, including 3D plain CT images of the esophagus from normal patients, as well as 3D plain CT images of the esophagus from patients with benign or cancerous conditions. These CT images are used as sample images to train the image processing model. When the sample images are CT images of the esophagus, the sample segmentation map of each sample image can be understood as a segmentation map of each CT image of the esophagus, with each segment labeled (e.g., normal, benign, etc.). The sample object label can be understood as the pathological label of each CT image of the esophagus, such as normal, benign, or cancerous. In practical applications, the segmentation maps of each esophageal CT image are annotated by doctors based on practical experience. However, when accurate annotation is not possible on the esophageal CT image itself, annotation can be performed on the enhanced contrast images of the esophageal CT image. The segmentation map of the esophageal CT image is then determined based on the annotation results in the enhanced contrast images. The specific implementation method is as follows: Determining the sample object segmentation map and sample object label in the sample image includes: identifying at least two enhanced contrast sample images containing the sample object; and determining the sample object segmentation map and sample object label in the sample image based on the sample annotations of the sample object in the at least two enhanced contrast sample images.Taking a CT image of the esophagus as an example, the enhanced contrast image of this sample image can be understood as an image generated after focusing on suspicious areas identified on a plain CT scan of the esophagus, following intravenous injection of contrast agent. In practice, multiple enhanced contrast CT images containing the esophagus are acquired. Based on the sample annotations for the esophagus in each enhanced contrast CT image, a robust image configuration method is used to align the sample annotations for the esophagus from the enhanced contrast CT to the plain CT, thus converting them into sample segmentation maps of the sample objects in the sample image. After determining the sample segmentation map of the sample objects in each sample image, a sample object label can be determined based on each segment in the sample segmentation map. For example, if the volume of a normal segment is greater than a preset volume threshold, the sample object can be considered labeled as normal. Therefore, after acquiring the sample image, the sample object segmentation map, and the sample object labels, the image processing model is specifically trained based on the sample image, the sample object segmentation map, and the sample object labels. First, the sample image is input into the convolutional layer of the encoder to obtain the initial sample feature map of the sample image. The specific implementation of obtaining the initial sample feature map is the same as that of obtaining the initial feature map of the image to be processed in the above embodiment. For example, the initial sample feature map of the sample image can also be obtained by performing convolution through two convolutional layers of the encoder. After obtaining the initial sample feature map, it is input into the self-attention mechanism layer of the encoder. Based on the position information of each initial sample feature in the initial sample feature map and the positional relationship between each initial sample feature and other initial sample features, the target sample feature map corresponding to the initial sample feature map is obtained. This self-attention mechanism layer is a global self-attention mechanism layer, where other initial sample features can be understood as features other than their own sample features in the initial sample feature map. In practical applications, the receptive domain of ordinary uNet segmentation networks is limited, heavily relying on local text patterns rather than global context. If a segmentation model cannot correctly segment the target, it will certainly affect subsequent classification tasks. Therefore, establishing a more robust image processing model is more sensitive and important for tumors, especially early-stage tumors. Therefore, in the image processing model of the image processing method provided in the embodiments of the present invention, an architectural improvement is proposed by integrating a global self-attention layer, that is, adding a self-attention mechanism layer after each convolutional layer of the encoder to enhance the global context modeling capability of the model.Specifically, the initial sample feature map is input into the self-attention mechanism layer. When determining the target sample feature map corresponding to the initial sample feature map through the self-attention mechanism layer, the position of each initial sample feature is determined according to the coordinate grid of the initial sample feature map. Then, based on the position of each initial sample feature and its relationship with the positions of other initial sample features, the target sample feature corresponding to each initial sample feature is calculated through the first preset algorithm. Finally, based on the target sample feature, the accurate target sample feature map corresponding to the initial sample feature map can be obtained. That is, based on the positional relationship between features, each initial sample feature learns the target sample feature map composed of target sample features with global context knowledge. The specific implementation method is as follows: Obtaining the target sample feature map corresponding to the initial sample feature map based on the position information of each initial sample feature in the initial sample feature map and the positional relationship between each initial sample feature and other initial sample features includes: determining the coordinate grid of the initial sample feature map, and determining the position information of each initial sample feature in the initial sample feature map based on the coordinate grid; calculating the target sample feature corresponding to each initial sample feature using a first preset algorithm based on the position information of each initial sample feature and the positional relationship between each initial sample feature and other initial sample features; and obtaining the target sample feature map corresponding to the initial sample feature map based on the target sample feature. Wherein, the image to be processed containing the target object is a three-dimensional image, and the initial sample feature map is obtained by convolution of the three-dimensional image through a convolutional layer. Therefore, the coordinate grid of the initial sample feature map can be determined through the three-dimensional initial sample feature map, thereby determining the position of each feature (i.e., feature point) in the initial sample feature map within the coordinate grid. Continuing with the previous example, the sample image is a three-dimensional plain CT image of the esophagus. Therefore, the initial sample feature map obtained after convolution of the sample image can be used to determine the coordinate network of the initial sample feature map according to the coordinate system. The position information of each initial sample feature in the initial sample feature map can be determined according to the coordinate grid. Based on the position information of each initial sample feature and the positional relationship between each initial sample feature and other sample features, the target sample feature corresponding to each initial sample feature is calculated through the first preset algorithm. Finally, the target sample feature map is constructed based on the target sample feature. That is, after obtaining the initial sample feature map of each sample image through the convolutional layer, the target sample feature corresponding to each initial sample feature in the initial sample feature map is calculated through the first preset algorithm, and the target sample feature map can then be determined based on the target sample feature.Specifically, the implementation of the first preset algorithm is shown in Formula 1: where represents the target sample feature corresponding to each initial sample feature, N represents the coordinate grid of the initial sample feature map, P represents other sample features corresponding to o, o represents each current initial sample feature participating in the calculation in the initial sample feature map, represents the query variable at position o, represents the key variable at position o, represents the value variable at position o, and , , are all projections of the input, and are all learnable matrices. Taking the initial sample feature map as an example, each initial sample feature in the initial feature map is calculated using Formula 1, and the target sample feature map formed by the corresponding target sample feature is then ; that is, the input of the self-attention mechanism layer is X, and the output after passing through the self-attention mechanism layer is y. Through the calculation of Formula 1, each initial sample feature can learn the knowledge of other sample features and the positional relationship between other sample features. In subsequent applications, the features of the context can also be considered as correct segmentation predictions. In practical applications, the self-attention mechanism has high computational complexity when applied to visual problems (such as image processing). To reduce computational complexity, each point o (initial sample feature) in the initial sample feature map can be processed using local region calculation. The specific implementation is as follows: Obtaining the target sample feature map corresponding to the initial sample feature map based on the position information of each initial sample feature in the initial sample feature map and the positional relationship between each initial sample feature and other initial sample features includes: determining the coordinate grid of the initial sample feature map and determining the position information of each initial sample feature in the initial sample feature map based on the coordinate grid; determining the associated region of each initial sample feature and the associated sample features of the associated region based on the position information; calculating the target sample feature corresponding to each initial sample feature using a second preset algorithm based on the position information of each initial sample feature and the positional relationship between each initial sample feature and the associated sample features; and obtaining the target sample feature map corresponding to the initial sample feature map based on the target sample feature. The determination of the coordinate grid of the initial sample feature map and the position information of each initial sample feature can be found in the above embodiments, and will not be repeated here.Specifically, after determining the positional information of each initial sample feature in the initial sample feature map, the associated region of each initial sample feature and the associated sample features of that associated region can be determined based on the positional information between each initial sample feature and other sample features. For example, depending on the actual application, each initial sample feature is used as the center point, and a preset distance is extended outward. The area between this preset distance and the initial sample feature is considered as the associated region. Then, based on the positional information of each initial sample feature and the positional relationship between each initial sample feature and the associated sample features, the target sample feature corresponding to each initial sample feature is calculated using a second preset algorithm. Finally, based on the target sample feature, a target sample feature map is constructed. The specific second preset algorithm is shown in Formula 2: where m represents the local region of each o, r is a learnable parameter encoding containing positional information, represents the distance from position p to position o, and and represent the encoding of the key variable k and value variable v, respectively. Other parameter representations are shown in Formula 1. As shown in Equation 2, for each position o in the coordinate grid of the initial sample feature map, a local region is extracted in each self-attention mechanism layer. By calculating the other sample features in the local region m for each initial sample feature, the computational load can be significantly reduced. Furthermore, a learned relative position encoding term is introduced, and additional positions are embedded in the query, key, and value to capture random interactions with precise positional information. To improve the performance of the image processing model, the initial sample feature map is processed using a multi-head self-attention mechanism to obtain the final result. The final output is obtained by merging the results of each head. Finally, the samples are reshaped to the original size of the feature map. In this embodiment of the invention, to improve the computational performance and reduce the computational complexity of the self-attention mechanism, the associated sample features of the nearby local regions of each initial sample feature in the initial sample feature map are determined based on positional information. Based on the positional relationship between each initial sample feature and the associated sample features, and combined with the multi-head attention mechanism, the target sample feature corresponding to each initial sample feature is calculated. Specifically, the associated sample features of each initial sample feature are obtained by determining them in the following way: Determining the associated region of each initial sample feature and the associated sample features of the associated region based on the location information includes: determining the association distance between each initial sample feature and other initial sample features based on the location information of each initial sample feature; determining the associated region of each initial sample feature based on the association distance, and identifying each initial sample feature in the associated region as an associated sample feature. The preset distance threshold can be set according to actual applications, and this invention does not impose any limitations on it.In specific implementation, based on the location information of each initial sample feature, the association distance between each initial sample feature and other initial sample features is determined. Then, based on the relationship between the association distance and a preset association distance, the association region of each initial sample feature is determined, and each initial sample feature in each association region is identified as an associated sample feature. After determining the associated sample features of each initial sample feature in the association region, the association distance between each initial sample feature and its corresponding associated sample feature is determined. Then, based on the location information of each initial sample feature and the association distance between each initial sample feature and its corresponding associated sample feature, the target sample feature corresponding to each initial sample feature is accurately calculated using a second preset algorithm to increase the prediction accuracy of the trained image processing model. The specific implementation method is as follows: The step of calculating the target sample feature corresponding to each initial sample feature based on the location information of each initial sample feature and the positional relationship between each initial sample feature and its associated sample feature using a second preset algorithm includes: calculating the target sample feature corresponding to each initial sample feature based on the location information of each initial sample feature and the association distance between each initial sample feature and its associated sample feature using a second preset algorithm. Furthermore, since the initial sample feature map may be three-dimensional with a number of channels, and the channel count dimension is generally quite large, dimensionality reduction can be performed to further reduce the computational cost of the self-attention mechanism layer and improve its computational efficiency. The specific implementation is as follows: Before obtaining the target sample feature map corresponding to the initial sample feature map based on the position information of each initial sample feature in the initial sample feature map and the positional relationship between each initial sample feature and other initial sample features, the method further includes: reducing the dimensionality of the initial sample feature map according to a preset dimensionality reduction method to obtain a dimensionality-reduced initial sample feature map. The preset dimensionality reduction method can be set according to actual applications, and this invention does not impose any limitations on it. For example, the preset dimensionality reduction method can be understood as reducing the channel count of the three-dimensional data. Simultaneously, to increase the complexity of the image processing model and further improve the accuracy of subsequent use of the image processing model, a self-attention mechanism layer can be added after multiple convolutional layers. Specifically, the self-attention mechanism layer includes at least two layers, and each self-attention mechanism layer is placed after the convolutional layer.Taking a three-layer system with at least two self-attention mechanism layers as an example; firstly, the initial sample feature map is input into the first self-attention mechanism layer of the encoder. In the first self-attention mechanism layer, the target sample feature map corresponding to the initial sample feature map is obtained according to the above calculation method. Then, the target sample feature map is input into the next convolutional layer for convolution to obtain the convolutional feature map. This convolutional feature map is then input into the next self-attention mechanism layer for computation, outputting the corresponding target sample feature map. This process is repeated until the target sample feature map output by the last self-attention mechanism layer is obtained. Step 204: The initial feature map is input into the self-attention mechanism layer of the encoder. Based on the position information of each feature in the initial feature map and the positional relationship between each feature and other features, the target feature map corresponding to the initial feature map is obtained. Specifically, the processing of the initial feature map in the self-attention mechanism layer and the specific processing method for obtaining the target feature map can be found in the detailed explanation of the image processing model training in the above embodiment, and will not be repeated here. After obtaining the initial feature map of the image to be processed, the initial feature map is input into the self-attention mechanism layer of the encoder of the pre-processed image processing model. Based on the position information of each feature in the initial feature map and the positional relationship between each feature and other features, the target feature map corresponding to the initial feature map is obtained. The specific implementation of obtaining the target feature map corresponding to the initial feature map based on the position information of each feature in the initial feature map and the positional relationship between each feature and other features in the self-attention mechanism layer can be found in the above description and will not be repeated here. Step 206: The target feature map is input into the decoder for processing to obtain the object segmentation map and object label of the target object in the image to be processed. After obtaining the target feature map, it is input into the decoder for processing to obtain the object segmentation map and object label of the target object in the image to be processed. Specifically, obtaining the object segmentation map and object label of the target object in the image to be processed includes: obtaining the object segmentation map of the target object in the image to be processed, and the segmentation label of each segmentation block in the object segmentation map; calculating the volume of segmentation blocks with the same segmentation label, and determining the object label of the target object based on the volume of the segmentation block corresponding to each type of segmentation label. Continuing with the previous example, when the image to be processed is a CT image of the esophagus, the object segmentation map of the target object in the image can be understood as an esophageal segmentation map. The segmentation label of each segment in the esophageal segmentation map can be understood as the pathological condition of each segment, such as normal, benign, or cancerous. Then, the volume of segmented blocks with the same segmentation label is calculated, and the object label of the target object is determined based on the volume of the segmented blocks corresponding to each type of segmentation label.For example, the volumes of segments labeled "normal," "benign," and "cancer" are calculated. If the volume of the normal segment is greater than the volumes of the benign and cancerous segments, then the object label for the esophagus can be considered normal. In practical applications, the object label can also be determined based on the volume of the segments corresponding to each type of segment label, using other rules. For example, if the volume of the normal segment is greater than a preset volume threshold, the object label for the esophagus can also be considered normal. The specific implementation of determining the object label based on the segment volume can be set according to the actual application, and this embodiment of the invention does not impose any limitations on this. In one possible implementation, the three-dimensional CT image of each esophagus can be classified into three object labels, such as cancer, benign, or normal, using an image processing model. To achieve an interpretable classification, a simple, non-parametric method is used to provide the final spatial-level decision. First, a voxel graph is constructed for all predicted normal esophagus and esophageal abnormalities (cancer + benign tumors). All connected components in the graph are calculated, and only the largest connected volume is retained, discarding smaller volume segmentation results. If the 3D volume of an esophageal abnormality is less than Kmm³, it is considered normal. K is adjusted to achieve 99% specificity on the validation set. To further classify abnormal cases as cancer or benign, the largest volume category can be selected as the final classification label. Furthermore, if an image processing model is pre-trained before predicting object segmentation maps and object labels for the image to be processed according to the image processing method, the object segmentation map and object label of the target object in the image to be processed can be obtained directly from this image processing model. The specific implementation is as follows: After obtaining the image to be processed containing the target object, the process further includes: inputting the image to be processed into the image processing model to obtain the object segmentation map and object label of the target object in the image to be processed. The image processing method provided in this embodiment of the invention proposes a novel position-sensitive self-attention mechanism. When predicting the object segmentation map of a target object in the image to be processed, the model's prediction accuracy is enhanced through non-local interaction based on the positional relationship between each feature in the initial feature map of the image to be processed and other features in the global domain. Applied to esophageal cancer screening, this non-invasive method can accurately identify early-stage cancer with high sensitivity. The following description, in conjunction with Figure 3, uses the application of the image processing method provided by this invention in esophageal cancer screening as an example to further illustrate the image processing method. Figure 3 shows a flowchart of the processing procedure of an image processing method provided in an embodiment of the invention, specifically including the following steps: Step 302: Acquire esophageal CT images of historical patients and determine the esophageal segmentation map and esophageal pathology label of the esophageal CT images.Specifically, a training dataset is constructed based on esophageal CT images, esophageal segmentation maps for each esophageal CT image, and esophageal pathology labels. The image processing model is trained using this dataset. The training of the image processing model can be understood as a three-class classification problem for esophageal CT images. First, the training dataset is obtained and labeled as follows: where is the 3D plain esophageal CT image of the i-th patient, and is a voxel-level annotation label (i.e., esophageal segmentation map) with the same size as , representing the segmentation targets: background, esophageal cancer, normal, and benign tumor. To obtain the annotation labels in Y, for 3D plain esophageal CT images of patients receiving radiotherapy, the tumor volume (GTV) delineation made by the doctor on the 3D plain esophageal CT image during radiotherapy can be directly used. For 3D plain esophageal CT images of other patients, tumor annotations made by the doctor on contrast-enhanced esophageal CT images, along with reference to clinical and endoscopic reports, can be used to complete the annotation. Then, a robust image alignment method is used to align the annotated mask from the contrast-enhanced CT image to the plain CT image, and manual correction is performed during quality checks. This refers to the patient's label, specifically the esophageal pathology label corresponding to the patient's 3D plain esophageal CT image, such as esophageal cancer, benign, or normal. The esophageal pathology label can be obtained from pathology and radiology reports. In Figure 3, part a represents the 3D plain esophageal CT image, part b represents the process of aligning the enhanced esophageal CT image to the plain CT image, and part c represents the voxel-level annotation label of the 3D plain esophageal CT image, i.e., the esophageal segmentation map. Step 304: Input the esophageal CT image into the first convolutional layer of the encoder in the image processing model to obtain the first feature map of the esophageal CT image. Step 306: Input the first feature map into the second convolutional layer to obtain the second feature map, and use this second feature map as the initial feature map, inputting it into the self-attention mechanism layer after the second convolutional layer. Specifically, after the initial feature map is input (input in Figure 3) to the self-attention mechanism layer, it undergoes processing including 1*1 convolution (dimensionality reduction, Conv1*1 in Figure 3), first normalization (Norm in Figure 3), multi-head self-attention mechanism (Multi-head Attention in Figure 3), second normalization (Norm in Figure 3), multilayer perceptron (MLP), and 1*1 convolution (restoration, Conv1*1 in Figure 3), resulting in an output (output in Figure 3) of the same size as the input. In practical applications, to increase the model complexity of the self-attention mechanism layer, the same processing can be performed twice within a single self-attention mechanism layer to improve the accuracy of the trained image processing model. Step 308: The target feature map output from the final self-attention mechanism layer is input into the decoder of the image processing model to obtain the predicted esophageal segmentation map and predicted esophageal pathological label of the esophageal CT image.Step 310: Train the image processing model based on the predicted esophageal segmentation map, the predicted esophageal pathology label, and the loss function composed of the esophageal segmentation map and the esophageal pathology label to obtain the image processing model. Step 312: In practical applications, input the esophageal CT image to be predicted into the image processing model to obtain the esophageal segmentation map and esophageal pathology label of the esophageal CT image to be predicted. The esophageal pathology label includes labels such as normal, cancer, or benign. The image processing method provided in this embodiment of the invention provides a relatively convenient and simple non-contrast CT scan method for screening esophageal cancer. To better capture the overall background and detect early tumors, a position-sensitive self-attention mechanism is proposed, which enhances the convolutional layer through non-local interactions in the encoder. High sensitivity and specificity are achieved on large-scale datasets. Compared with other tools such as blood tests and endoscopy, the feasibility of using non-contrast CT scans for large-scale opportunistic screening of esophageal cancer is better. Corresponding to the above method embodiment, this invention also provides an image processing device embodiment. Figure 4 shows a schematic diagram of the structure of an image processing device provided in one embodiment of the invention. As shown in Figure 4, the device includes: a first feature map acquisition module 402, configured to acquire an image to be processed containing a target object, input the image to be processed into a convolutional layer of an image processing model to obtain an initial feature map of the image to be processed, wherein the image processing model includes an encoder and a decoder; a second feature map acquisition module 404, configured to input the initial feature map into the self-attention mechanism layer of the encoder, and obtain a target feature map corresponding to the initial feature map based on the position information of each feature in the initial feature map and the positional relationship between each feature and other features; and a first feature map processing module 406, configured to input the target feature map into the decoder for processing to obtain an object segmentation map and object label of the target object in the image to be processed. Optionally, the apparatus includes a model training module, wherein the model training module is configured to: acquire a sample image containing sample objects, and determine a sample object segmentation map and sample object labels in the sample image; input the sample image into the convolutional layer of the encoder to obtain an initial sample feature map of the sample image; input the initial sample feature map into the self-attention mechanism layer of the encoder, and obtain a target sample feature map corresponding to the initial sample feature map based on the position information of each initial sample feature in the initial sample feature map and the positional relationship between each initial sample feature and other initial sample features; input the target sample feature map into the decoder for processing to obtain a predicted object segmentation map and predicted object labels of the sample objects in the sample image; and train the image processing model based on the sample object segmentation map and sample object labels, and a loss function formed by the predicted object segmentation map and predicted object labels.Optionally, the model training module is further configured to: determine the coordinate grid of the initial sample feature map, and determine the position information of each initial sample feature in the initial sample feature map according to the coordinate grid; calculate the target sample feature corresponding to each initial sample feature using a first preset algorithm based on the position information of each initial sample feature and the positional relationship between each initial sample feature and other initial sample features; and obtain the target sample feature map corresponding to the initial sample feature map based on the target sample feature. Optionally, the model training module is further configured to: determine the coordinate grid of the initial sample feature map, and determine the position information of each initial sample feature in the initial sample feature map according to the coordinate grid; determine the associated region of each initial sample feature and the associated sample feature of the associated region based on the position information; calculate the target sample feature corresponding to each initial sample feature using a second preset algorithm based on the position information of each initial sample feature and the positional relationship between each initial sample feature and the associated sample feature; and obtain the target sample feature map corresponding to the initial sample feature map based on the target sample feature. Optionally, the model training module is further configured to: determine the association distance between each initial sample feature and other initial sample features based on the location information of each initial sample feature; determine the association region of each initial sample feature based on the association distance, and determine each initial sample feature in the association region as an associated sample feature. Optionally, the model training module is further configured to: calculate the target sample feature corresponding to each initial sample feature using a second preset algorithm based on the location information of each initial sample feature and the association distance between each initial sample feature and the associated sample feature. Optionally, the device further includes: a dimensionality reduction module configured to reduce the dimensionality of the initial sample feature map according to a preset dimensionality reduction method to obtain a dimensionality-reduced initial sample feature map. Optionally, the self-attention mechanism layer includes at least two layers, and each self-attention mechanism layer is set after the convolutional layer. Optionally, the device further includes: a model processing module configured to input the image to be processed into the image processing model to obtain an object segmentation map and object labels of the target object in the image to be processed. Optionally, the first feature map processing module 406 is further configured to: obtain an object segmentation map of the target object in the image to be processed, and a segmentation label for each segmentation block in the object segmentation map; calculate the volume of segmentation blocks with the same segmentation label, and determine the object label of the target object based on the volume of the segmentation block corresponding to each type of segmentation label.Optionally, the model training module is further configured to: determine at least two enhanced contrast sample images containing the sample object; and determine a sample object segmentation map and a sample object label in the sample images based on the sample annotations of the sample objects in the at least two enhanced contrast sample images. Optionally, the image to be processed containing the target object is a CT image containing the esophagus. The image processing device provided in this embodiment of the invention proposes a novel position-sensitive self-attention mechanism. When predicting the object segmentation map of the target object in the image to be processed, it enhances the prediction accuracy of the model through this non-local interaction based on the positional relationship between each feature in the initial feature map of the image to be processed and other features in the global domain. Applied to the scenario of esophageal cancer screening, it can accurately identify early cancer in a non-invasive manner with high sensitivity. The above is an illustrative scheme of an image processing device according to this embodiment. It should be noted that the technical solution of this image processing device and the technical solution of the above-described image processing method belong to the same concept. Details not described in detail in the technical solution of the image processing device can be referred to in the description of the technical solution of the above-described image processing method. Referring to Figure 5, Figure 5 shows a flowchart of an esophageal CT image processing method according to an embodiment of the present invention, specifically including the following steps: Step 502: Acquire a CT image containing the esophagus, input the CT image into the convolutional layer of an image processing model to obtain an initial feature map of the CT image, wherein the image processing model includes an encoder and a decoder; Step 504: Input the initial feature map into the self-attention mechanism layer of the encoder, and obtain a target feature map corresponding to the initial feature map based on the position information of each feature in the initial feature map and the positional relationship between each feature and other features; Step 506: Input the target feature map into the decoder for processing to obtain a segmentation map of the esophagus and pathological labels in the CT image. The esophageal CT image processing method provided by this embodiment of the present invention provides a relatively convenient and simple non-contrast CT scanning method for screening esophageal cancer. In order to better capture the global background and detect early tumors, a position-sensitive self-attention mechanism is proposed, which enhances the convolutional layer through non-local interactions in the encoder. High sensitivity and specificity are achieved on large-scale datasets. Compared to other tools such as blood tests and endoscopy, the feasibility of using non-contrast CT scans for large-scale opportunistic screening of esophageal cancer is relatively good. The above is an illustrative scheme of an esophageal CT image processing method according to this embodiment. It should be noted that the technical solution of this esophageal CT image processing method belongs to the same concept as the image processing method described above. Details not described in detail in the technical solution of the esophageal CT image processing method can be found in the description of the image processing method described above.Referring to Figure 6, Figure 6 shows a schematic diagram of an esophageal CT image processing device according to an embodiment of the present invention. The device includes: a third feature map acquisition module 602, configured to acquire CT images containing the esophagus, input the CT images into the convolutional layer of an image processing model to obtain an initial feature map of the CT images, wherein the image processing model includes an encoder and a decoder; a fourth feature map acquisition module 604, configured to input the initial feature map into the self-attention mechanism layer of the encoder, and obtain a target feature map corresponding to the initial feature map based on the position information of each feature in the initial feature map and the positional relationship between each feature and other features; and a second feature map processing module 606, configured to input the target feature map into the decoder for processing to obtain a segmentation map of the esophagus and pathological labels in the CT image. The esophageal CT image processing device provided by this embodiment of the present invention provides a relatively convenient and simple non-contrast CT scanning method for screening esophageal cancer. In order to better capture the global background and detect early tumors, a position-sensitive self-attention mechanism is proposed, which enhances the convolutional layer through non-local interactions in the encoder. High sensitivity and specificity are achieved on large-scale datasets. Compared to other tools such as blood tests and endoscopy, the feasibility of using non-contrast CT scans for large-scale opportunistic esophageal cancer screening is better. The above is a schematic scheme of an esophageal CT image processing device according to this embodiment. It should be noted that the technical solution of this esophageal CT image processing device and the technical solution of the above-described esophageal CT image processing method belong to the same concept. Details not described in detail in the technical solution of the esophageal CT image processing device can be found in the description of the technical solution of the above-described esophageal CT image processing method. Referring to Figure 7, Figure 7 shows a flowchart of another esophageal CT image processing method provided by an embodiment of the present invention, specifically including the following steps. Step 702: Receive a user's processing request for an esophageal CT image, wherein the esophageal CT image is an image of the esophagus scanned by a computed tomography (CT) scanner; Step 704: Input the esophageal CT image into an image processing model. If the user is determined to be a first-type user, output and display to the first-type user a segmentation diagram of the esophagus and pathological labels in the esophageal CT image; or if the user is determined to be a second-type user, output and display to the second-type user the pathological labels of the esophageal CT image; wherein, the image processing model is the image processing model in the above-described image processing method. Specifically, the user can be understood as any type of user, such as a doctor or a non-doctoral member of the general population; and the results displayed to the user by this esophageal CT image processing method will differ depending on the user type.If the first type of user is a doctor, the esophageal segmentation diagram and pathology label from the esophageal CT image can be shown to the doctor. The doctor can then analyze the patient's condition based on this diagram and label. For the second type of user, a general user (non-medical professional), the pathology label can be shown directly. In practice, if the user is a doctor, they can directly obtain the esophageal CT image from the CT scan, input it into a computer terminal, and send it to the server. The server can then use the image processing method described above to obtain the segmentation diagram and pathology label from the esophageal CT image and return it to the computer terminal for display. If the user is a general user, they can input an image of the esophagus scanned by a CT scanner into the computer terminal, or input an image of the esophagus scanned by a CT scanner obtained through other means. The computer terminal then sends the esophageal CT image to the server, which can then use the image processing method described above to obtain the pathology label from the esophageal CT image and return it to the computer terminal for display. The esophageal CT image processing method provided in this invention offers a relatively convenient and simple non-contrast CT scanning method for screening esophageal cancer. To better capture the overall background and detect early tumors, a location-sensitive self-attention mechanism is proposed, enhancing the convolutional layers through non-local interactions in the encoder. High sensitivity and specificity are achieved on large-scale datasets. Compared to other tools such as blood tests and endoscopy, the feasibility of using non-contrast CT scanning for large-scale opportunistic esophageal cancer screening is better. Furthermore, based on user interaction requests, the image processing model can return different prediction results to users according to their type, improving the versatility of the method implementation. Referring to Figure 8, Figure 8 shows a schematic diagram of another esophageal CT image processing device according to an embodiment of the present invention. The device includes: a request receiving module 802, configured to receive a user's processing request for an esophageal CT image, wherein the esophageal CT image is an image of the esophagus scanned by a computed tomography (CT) scanner; and a first output module 804, configured to input the esophageal CT image into an image processing model, and, if the user is determined to be a first type of user, output and display to the first type of user a segmentation diagram of the esophagus and pathological labels in the esophageal CT image, or, if the user is determined to be a second type of user, output and display to the second type of user a pathological label of the esophageal CT image; wherein the image processing model is the image processing model in the above-described image processing method. The esophageal CT image processing device provided by this embodiment of the present invention provides a relatively convenient and simple non-contrast CT scanning method for screening esophageal cancer.To better capture the overall background and detect early-stage tumors, a location-sensitive self-attention mechanism is proposed, which enhances the convolutional layer through non-local interactions in the encoder. High sensitivity and specificity are achieved on large-scale datasets. Compared to other tools such as blood tests and endoscopy, the feasibility of using non-contrast CT scans for large-scale opportunistic esophageal cancer screening is better. Furthermore, based on user interaction requests, the image processing model can return different prediction results to users according to their type, improving the versatility of the method implementation. Figure 9 shows a block diagram of a computing device 900 according to an embodiment of the present invention. The components of the computing device 900 include, but are not limited to, a memory 910 and a processor 920. The processor 920 is connected to the memory 910 via a bus 930, and a database 950 is used to store data. The computing device 900 also includes an access device 940, which enables the computing device 900 to communicate via one or more networks 960. Examples of these networks include Public Switched Telephone Network (PSTN), Local Area Network (LAN), Wide Area Network (WAN), Personal Area Network (PAN), or combinations of communication networks such as the Internet. Access device 940 may include one or more of any type of wired or wireless network interface (e.g., Network Interface Card (NIC)), such as IEEE 802.11 Wireless Local Area Network (WLAN) wireless interface, Wi-MAX interface, Ethernet interface, Universal Serial Bus (USB) interface, Cellular Network interface, Bluetooth interface, Near Field Communication (NFC) interface, and so on. In one embodiment of the invention, the aforementioned components of computing device 900, as well as other components not shown in FIG. 9, may also be interconnected, for example, via a bus. It should be understood that the block diagram of the computing device structure shown in FIG. 9 is merely for illustrative purposes and is not intended to limit the scope of the invention. Those skilled in the art can add or replace other components as needed. The computing device 900 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablets, personal digital assistants, laptops, notebooks, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or PCs. The computing device 900 can also be a mobile or stationary server. The processor 920 is configured to execute computer-executable instructions, which, when executed by the processor, implement the steps of the aforementioned image processing method or esophageal CT image processing method. The above is an illustrative scheme of a computing device according to this embodiment.It should be noted that the technical solution of this computing device belongs to the same concept as the aforementioned image processing method or esophageal CT image processing method. Details not described in detail in the technical solution of the computing device can be found in the descriptions of the aforementioned image processing method or esophageal CT image processing method. An embodiment of the present invention also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of the aforementioned image processing method or esophageal CT image processing method. The above is an illustrative scheme of a computer-readable storage medium according to this embodiment. It should be noted that the technical solution of this storage medium belongs to the same concept as the aforementioned image processing method or esophageal CT image processing method. Details not described in detail in the technical solution of the storage medium can be found in the descriptions of the aforementioned image processing method or esophageal CT image processing method. An embodiment of the present invention also provides a computer program, wherein, when the computer program is executed in a computer, it causes the computer to perform the steps of the aforementioned image processing method or esophageal CT image processing method. The above is an illustrative scheme of a computer program according to this embodiment. It should be noted that the technical solution of this computer program belongs to the same concept as the aforementioned image processing method or esophageal CT image processing method. Details not described in detail in the computer program's technical solution can be found in the descriptions of the aforementioned image processing method or esophageal CT image processing method. Specific embodiments of the present invention have been described above. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims can be performed in a different order than in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multiplexing and parallel processing are also possible or may be advantageous. The computer instructions include computer program code, which can be in raw code form, object code form, executable file, or some intermediate form, etc. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium may be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments of the present invention are not limited to the described order of actions, because according to the embodiments of the present invention, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments of the present invention. In the above embodiments, the descriptions of each embodiment have different focuses; for parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments. The preferred embodiments of the present invention disclosed above are only used to help illustrate the present invention. Optional embodiments do not describe all details exhaustively, nor do they limit the invention to only the specific implementation methods described. Obviously, many modifications and variations can be made according to the content of the embodiments of the present invention. The present invention selects and specifically describes these embodiments to better explain the principles and practical applications of the embodiments of the present invention, so that those skilled in the art can better understand and utilize the present invention. The present invention is only limited by the scope of the claims and their full scope and equivalents. [Simplified Explanation of the Diagram]

[0004] [Figure 1] is a schematic diagram of a specific scenario of an image processing method provided by an embodiment of the present invention applied to esophageal cancer screening; [Figure 2] is a flowchart of an image processing method provided by an embodiment of the present invention; [Figure 3] is a flowchart of the processing process of an image processing method provided by an embodiment of the present invention; [Figure 4] is a structural schematic diagram of an image processing device provided by an embodiment of the present invention; [Figure 5] is a flowchart of an esophageal CT image processing method provided by an embodiment of the present invention; [Figure 6] is a structural schematic diagram of an esophageal CT image processing device provided by an embodiment of the present invention; [Figure 7] is a flowchart of another esophageal CT image processing method provided by an embodiment of the present invention; [Figure 8] is a structural schematic diagram of another esophageal CT image processing device provided by an embodiment of the present invention; [Figure 9] is a structural block diagram of a computing device provided by an embodiment of the present invention.

Claims

1. An image processing method, comprising: The process involves: acquiring an image containing the target object; inputting the image into a convolutional layer of an image processing model to obtain an initial feature map; the image processing model including an encoder and a decoder; inputting the initial feature map into the self-attention mechanism layer of the encoder to determine the coordinate grid of the initial feature map, and determining the position information of each feature in the initial feature map based on the coordinate grid; obtaining the target feature map corresponding to the initial feature map based on the position information of each feature and the positional relationship between each feature and other features; and inputting the target feature map into the decoder for processing to obtain an object segmentation map and object label of the target object in the image to be processed.

2. According to the image processing method described in claim 1, the training steps of the image processing model are as follows: Acquire a sample image containing sample objects, and determine the sample object segmentation map and sample object labels in the sample image; Input the sample image into the convolutional layer of the encoder to obtain an initial sample feature map of the sample image; Input the initial sample feature map into the self-attention mechanism layer of the encoder to determine the coordinate grid of the initial sample feature map, and determine the position information of each initial sample feature in the initial sample feature map based on the coordinate grid; Obtain the target sample feature map corresponding to the initial sample feature map based on the position information of each initial sample feature and the positional relationship between each initial sample feature and other initial sample features; Input the target sample feature map into the decoder for processing to obtain a predicted object segmentation map and predicted object labels for the sample objects in the sample image; Train the image processing model based on the sample object segmentation map and sample object labels, and the loss function formed by the predicted object segmentation map and predicted object labels.

3. The image processing method according to claim 2, wherein obtaining a target sample feature map corresponding to the initial sample feature map based on the position information of each initial sample feature in the initial sample feature map and the positional relationship between each initial sample feature and other initial sample features, includes: Based on the location information of each initial sample feature and the positional relationship between each initial sample feature and other initial sample features, the target sample feature corresponding to each initial sample feature is calculated using a first preset algorithm; and the target sample feature map corresponding to the initial sample feature map is obtained based on the target sample feature.

4. The image processing method according to claim 2, wherein obtaining a target sample feature map corresponding to the initial sample feature map based on the position information of each initial sample feature in the initial sample feature map and the positional relationship between each initial sample feature and other initial sample features, includes: Based on this location information, determine the associated region of each initial sample feature and the associated sample features of that associated region; Based on the location information of each initial sample feature and the positional relationship between each initial sample feature and the associated sample feature, the target sample feature corresponding to each initial sample feature is calculated using a second preset algorithm; based on the target sample feature, the target sample feature map corresponding to the initial sample feature map is obtained.

5. The image processing method according to claim 4, wherein determining the associated region of each initial sample feature and the associated sample features of the associated region based on the location information includes: Based on the location information of each initial sample feature, determine the correlation distance between each initial sample feature and other initial sample features; The association region of each initial sample feature is determined based on the association distance, and each initial sample feature in the association region is determined as an associated sample feature.

6. The image processing method according to claim 4, wherein the target sample feature corresponding to each initial sample feature is calculated by a second preset algorithm based on the position information of each initial sample feature and the positional relationship between each initial sample feature and the associated sample feature, including: Based on the location information of each initial sample feature and the correlation distance between each initial sample feature and the associated sample feature, the target sample feature corresponding to each initial sample feature is calculated by the second preset algorithm.

7. The image processing method according to claim 2, before obtaining the target sample feature map corresponding to the initial sample feature map based on the position information of each initial sample feature in the initial sample feature map and the positional relationship between each initial sample feature and other initial sample features, further includes: The initial sample feature map is reduced in dimensionality according to the preset dimensionality reduction method to obtain the dimensionality-reduced initial sample feature map.

8. The image processing method according to claim 2, wherein the self-attention mechanism layer comprises at least two layers, and each self-attention mechanism layer is disposed after the convolutional layer.

9. The image processing method according to claim 2, further comprising, after acquiring the image of the target object to be processed: The image to be processed is input into the image processing model to obtain the object segmentation map and object label of the target object in the image to be processed.

10. The image processing method according to claim 1 or 9, wherein obtaining the object segmentation map and object label of the target object in the image to be processed includes: Obtain the object segmentation map of the target object in the image to be processed, and the segmentation label of each segmentation block in the object segmentation map; Calculate the volume of the segment blocks with the same segmentation label, and determine the object label of the target object based on the volume of the segment blocks corresponding to each type of segmentation label.

Images