System for automatically detecting and positioning foreign matters in image of digestive system department
By combining multispectral image data with a physical information neural network based on the law of light refraction, the problem of accuracy in foreign body detection and localization in gastroenterology imaging has been solved. This enables accurate reconstruction and localization of foreign bodies in complex fluid environments, improving detection accuracy and reliability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- AFFILIATED HOSPITAL OF JIANGSU UNIV
- Filing Date
- 2026-01-28
- Publication Date
- 2026-05-08
AI Technical Summary
Existing technologies make it difficult to accurately detect and locate foreign objects in gastroenterology imaging due to refractive distortion and color distortion caused by complex fluid environments. This is especially true in bile, blood, and gastric juice environments, where the accuracy and reliability of detection are extremely limited.
This method employs a physical information neural network based on the law of light refraction, combining multispectral image data with the refraction correction module to correct image distortion and a spectral enhancement module to improve visual contrast. It also extracts texture features using a gray-level co-occurrence matrix and uses a classification model to identify foreign objects and locate their contours.
It enables accurate reconstruction and localization of foreign body images in complex fluid environments, improving detection accuracy and reliability, ensuring localization stability and robustness under interference from mucus, reflective and turbid fluids, and shortening emergency diagnosis and intervention time.
Smart Images
Figure CN121987130A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the technical field of gastroenterology, and more particularly to an automatic foreign body detection and localization system in gastroenterology imaging. Background Technology
[0002] In gastroenterology, the auxiliary diagnosis of foreign objects such as fish bones, button batteries, or stones currently relies mainly on manual observation or general target detection algorithms. However, in the real clinical environment filled with bile, blood, and gastric juice, the accuracy and reliability of the detection are extremely limited.
[0003] Due to the difference in refractive index between liquids and air, partially or fully submerged foreign objects will exhibit severe geometric distortion and edge distortion in images, making it difficult for conventional algorithms to identify their true physical contours. At the same time, the strong absorption of specific spectra by turbid fluids will cause the loss of color information on the target surface, greatly reducing the visual contrast between the foreign object and the surrounding biological tissue, which can easily lead to missed or misdiagnosis. Furthermore, the dynamic visual displacement caused by the continuous peristalsis and fluid level fluctuations in the digestive tract makes it impossible for existing static detection models to achieve accurate spatial coordinate positioning, making it difficult to effectively assist doctors in emergency intervention.
[0004] Therefore, how to eliminate refractive distortion and color distortion interference caused by complex fluid environments and achieve accurate reconstruction and positioning of foreign objects has become a technical problem that urgently needs to be solved in clinical diagnosis. Summary of the Invention
[0005] This application aims to at least partially address one of the technical problems in the related art.
[0006] Therefore, one objective of this application is to propose an automatic foreign body detection and localization system in gastroenterology imaging that can eliminate refractive distortion and color distortion interference caused by complex fluid environments, and achieve accurate reconstruction and localization of foreign bodies.
[0007] To achieve the above objectives, the first aspect of this application proposes an automatic foreign body detection and localization system in gastroenterology imaging, comprising:
[0008] The data acquisition module is used to acquire multispectral image data of the digestive tract through the endoscope acquisition terminal, and to acquire real-time liquid level and depth information corresponding to the multispectral image data.
[0009] The refraction correction module is used to input the multispectral image data and the real-time liquid level depth information into a preset refraction correction model, and to process the multispectral image data using the refraction correction model. The refraction correction model integrates a physical information neural network based on the law of light refraction. The refraction correction model performs reverse calculation of the refraction path and position mapping on the multispectral image data based on the real-time liquid level depth information, and outputs restored image data.
[0010] The spectral enhancement module is used to perform adaptive spectral enhancement processing on the restored image data. It uses a preset spectral transformation matrix to redistribute the pixel values of different bands to generate enhanced image data.
[0011] The foreign object detection module is used to extract texture contrast difference features of multiple candidate regions from enhanced image data; input the texture contrast difference features into a preset classification model to detect foreign objects and obtain the probability distribution of foreign object targets;
[0012] The contour localization module is used to fit the edges of foreign objects in the enhanced image data according to the probability distribution of the foreign object target, and generate a foreign object localization result with marking information.
[0013] In addition, the automated foreign body detection and localization system in gastroenterology imaging proposed in this application may also have the following additional technical features:
[0014] In one embodiment of this application, acquiring multispectral image data and real-time liquid level depth information specifically includes:
[0015] The multispectral imaging unit of the endoscope acquires multiple narrow-band images covering a predetermined wavelength range, forming multispectral image data;
[0016] The distance between the endoscope lens and the surface of the liquid in the digestive tract is measured by the depth sensor module integrated into the endoscope, providing real-time liquid level and depth information.
[0017] In one embodiment of this application, the training process of the refraction correction model includes:
[0018] Construct a training dataset, the samples of which include multispectral images acquired in a fluid environment at a known depth and their corresponding non-refractive reference images;
[0019] A neural network model is constructed, and the physical equations describing the refraction and classification of light at the fluid interface are incorporated as constraints into the loss function of the neural network model.
[0020] The neural network model incorporating physical constraints is trained using the training dataset until it can accurately predict the refraction effect based on the input depth information and output the corrected image, thus obtaining the refraction correction model.
[0021] In one embodiment of this application, the specific steps for adaptive spectral enhancement processing of the restored image data are as follows:
[0022] The estimated absorption coefficients of the target fluid for each spectral channel at the current depth are obtained by querying a predefined lookup table based on the real-time liquid level and depth information.
[0023] Based on the estimated absorption coefficient, a spectral compensation matrix is constructed;
[0024] The multispectral vectors of each pixel in the restored image data are multiplied by the spectral compensation matrix to achieve spectral reconstruction and obtain the enhanced image data.
[0025] In one embodiment of this application, the texture contrast difference features are input into a preset classification model for foreign object determination to obtain the probability distribution of the foreign object target. The specific steps are as follows:
[0026] The texture contrast difference features are input into the feature fusion layer of the classification model to obtain a fused high-dimensional feature vector.
[0027] The high-dimensional feature vector is input into the fully connected layer of the classification model for calculation;
[0028] The calculation results of the fully connected layer are processed by the Softmax function in the output layer of the classification model to output the confidence score of the candidate region belonging to the preset foreign object category.
[0029] Based on the confidence scores of all candidate regions, a probability distribution of the foreign object target is generated, which characterizes the spatial distribution probability of the foreign object.
[0030] In one embodiment of this application, the specific method for extracting texture contrast difference features of multiple candidate regions from enhanced image data includes:
[0031] The enhanced image data is subjected to multi-scale sliding window scanning to divide the multiple candidate regions;
[0032] For each candidate region, its gray-level co-occurrence matrix is calculated, and contrast, energy, and homogeneity features are extracted from the gray-level co-occurrence matrix.
[0033] The texture contrast difference feature is generated by calculating the difference between each candidate region and its surrounding neighborhood organization in terms of contrast, energy, and homogeneity features.
[0034] In one embodiment of this application, the lookup table records the estimated absorption coefficients of different fluid types at different depths for each spectral band.
[0035] In one embodiment of this application, the refractive correction module is further configured to: identify the fluid category in the current digestive tract based on the global color moment features of the multispectral image data using a clustering algorithm, and select the corresponding estimated absorption coefficient from the lookup table according to the fluid category.
[0036] The automatic foreign body detection and localization system in gastroenterology imaging according to this application achieves physical restoration of foreign body image distortion in complex fluid environments by integrating a physical information neural network based on the law of light refraction, effectively correcting geometric distortion caused by interface refraction. Combining multispectral image data with a predefined lookup table compensation mechanism, the system can dynamically reconstruct the spectral energy distribution based on real-time liquid level depth, significantly improving the visual contrast between the foreign body and surrounding tissues. Furthermore, it extracts multidimensional texture features such as contrast, energy, and homogeneity of candidate regions through a gray-level co-occurrence matrix, and utilizes active contour fitting technology to achieve precise edge capture, ensuring the clinical stability and robustness of the localization results under interference from mucus, reflective light, and turbid fluids.
[0037] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description
[0038] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:
[0039] Figure 1 This is a system block diagram of the automatic foreign body detection and localization system in gastroenterology imaging according to this application;
[0040] Figure 2 This is a flowchart of the training process for the refractive correction model of the automatic foreign body detection and localization system in gastroenterology imaging of this application.
[0041] Figure 3 This is a flowchart illustrating the extraction of texture contrast difference features in the automatic foreign body detection and localization system for gastroenterology imaging according to this application. Detailed Implementation
[0042] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.
[0043] The automatic foreign body detection and localization system in gastroenterology imaging according to an embodiment of this application will be described below with reference to the accompanying drawings.
[0044] like Figures 1-3 As shown in the embodiment of this application, the automatic foreign body detection and localization system in gastroenterology imaging may include:
[0045] The data acquisition module is used to acquire multispectral image data of the digestive tract through the endoscope acquisition terminal, and to acquire real-time liquid level and depth information corresponding to the multispectral image data.
[0046] In one embodiment of this application, acquiring multispectral image data and real-time liquid level depth information specifically includes:
[0047] The multispectral imaging unit of the endoscope acquires multiple narrow-band images covering a predetermined wavelength range, forming multispectral image data; among them, the narrow-band filter component at the tip of the endoscope decomposes the reflected light into... Monochromatic light of a specific wavelength (such as...) The original data of each pixel can be represented as a multidimensional vector. ,in Represents wavelength The pixel grayscale values are used to form multispectral image data.
[0048] The distance between the endoscope lens and the surface of the liquid in the digestive tract is measured by the depth sensor module integrated into the endoscope, providing real-time liquid level and depth information.
[0049] Specifically, the depth sensor module emits infrared pulse signals, which are reflected off the surface of the fluid in the digestive tract and received by the sensor. The time difference between the transmission and reception is then calculated. The specific formula is as follows:
[0050] ;in, The speed at which the signal propagates in the cavity medium. This provides real-time liquid level and depth information.
[0051] It should be noted that the distance measured in real time by the depth sensor described in the above embodiments is... This determines the initial weights for calculating the deflection angle in the subsequent refraction correction model.
[0052] The refraction correction module is used to input the multispectral image data and the real-time liquid level depth information into a preset refraction correction model, and to process the multispectral image data using the refraction correction model. The refraction correction model integrates a physical information neural network based on the law of light refraction. The refraction correction model performs reverse calculation of the refraction path and position mapping on the multispectral image data based on the real-time liquid level depth information, and outputs restored image data.
[0053] In one embodiment of this application, the refractive correction module is further configured to: identify the fluid category in the current digestive tract based on the global color moment features of the multispectral image data using a clustering algorithm, and select the corresponding estimated absorption coefficient from the lookup table according to the fluid category.
[0054] It should be noted that the global color moment feature extraction described in the above embodiments can classify the color moment feature vectors using K-means or Support Vector Machine (SVM), and then output the fluid refractive index constant. This allows for the identification of fluid categories.
[0055] In one embodiment of this application, the training process of the refraction correction model includes:
[0056] Construct a training dataset, the samples of which include multispectral images acquired in a fluid environment at a known depth and their corresponding non-refractive reference images;
[0057] It should be noted that refractive distortion is collected through a physical simulation environment or an external simulation device. Multispectral images, real-time depth and refractive reference images This constitutes a training pair;
[0058] For example, different types of simulated digestive tract fluids (such as bile and blood) are filled in a transparent tank, and standard grid plates or foreign objects are placed at different depths to acquire images using a multispectral endoscope.
[0059] A neural network model is constructed, and the physical equations describing the refraction and classification of light at the fluid interface are incorporated as constraints into the loss function of the neural network model.
[0060] Calculate the pixel point in Theoretical displacement at depth displacement predicted by the model The residual. The pixel offset formula derived from Snell's law:
[0061] ,in Angle of incidence The angle of refraction;
[0062] The total loss function is: ,in The corrected predicted image after the output of the neural network; A pre-acquired non-refractive reference image; As a data-driven loss term, the difference in pixel grayscale values between the predicted image and the standard image is calculated by using the L2 norm (mean squared error). This is the pixel displacement vector field predicted by the intermediate layer inside the neural network; This is the theoretical displacement; This is the Physics-informed Loss, which ensures that the displacement laws learned by the model conform to the geometric properties of optical refraction.
[0063] The neural network model incorporating physical constraints is trained using the training dataset until it can accurately predict the refraction effect based on the input depth information and output the corrected image, thus obtaining the refraction correction model.
[0064] Furthermore, and Let these be the weight coefficients. In the early stages of training, let... > At this point, the network is allowed to prioritize learning the basic features of the image, such as texture and color, so that the output image is roughly similar in outline. In the later stages of training, the amount can be gradually increased. The weights are assigned to each element. After the image is initially restored, the physical loss term is strengthened to force the network to correct local distortions that do not conform to the refraction law, thereby improving the physical accuracy of edge fitting. Finally, the group with the highest overlap of foreign object contours in the test set is selected. , The combination is used as the final preset value.
[0065] The spectral enhancement module is used to perform adaptive spectral enhancement processing on the restored image data. It uses a preset spectral transformation matrix to redistribute the pixel values of different bands to generate enhanced image data.
[0066] In one embodiment of this application, the specific steps for adaptive spectral enhancement processing of the restored image data are as follows:
[0067] The estimated absorption coefficients of the target fluid for each spectral channel at the current depth are obtained by querying a predefined lookup table based on the real-time liquid level and depth information. It should be noted that the lookup table described in the above embodiment records the estimated absorption coefficients of different fluid types for each spectral band at different depths.
[0068] It should be noted that different liquids absorb different wavelengths to varying degrees, and the missing energy for each spectral channel can be obtained through a pre-set lookup table.
[0069] Fluid type Liquid level depth (mm) 415nm (blue) absorption coefficient μ 540nm (green) absorption coefficient μ Absorption coefficient μ at 600nm (red) bile 2.0 0.45 0.12 0.05 bile 5.0 1.12 0.30 0.12 bile 10.0 2.25 0.61 0.25 blood 2.0 1.85 1.62 0.08 blood 5.0 4.62 4.05 0.20 gastric juice 5.0 0.08 0.05 0.03
[0070] Table 1 is a partial lookup table for spectral absorption compensation of gastrointestinal fluids.
[0071] Based on the estimated absorption coefficient, a spectral compensation matrix is constructed;
[0072] The multispectral vectors of each pixel in the restored image data are multiplied by the spectral compensation matrix to achieve spectral reconstruction and obtain the enhanced image data.
[0073] Specifically, for example, depth can be obtained through a depth sensor. The value is 5.2mm, and the refraction correction module identifies the current fluid type as bile. The system then locates the row corresponding to bile in the table. Since 5.2mm is not listed in the table, the system can use a linear lookup formula. Outputs the precise absorption coefficient for the next band at the current depth. The value is 1.1652. Substituting this value into the spectral compensation index model: ; Thus, the calculation is obtained The system will perform the above interpolation calculations simultaneously on 415nm (blue), 540nm (green), and 600nm (red). Assume the calculated compensation weights are as follows: The above simulation calculations have been performed. ;
[0074] Light compensation matrix At this time, Transpose to obtain the original pixel vector The light compensation rectangle is compared with the original pixel vector. Multiplication completes the spectral reconstruction of each pixel.
[0075] ;
[0076] Foreign object detection module is used to extract texture contrast difference features of multiple candidate regions from enhanced image data;
[0077] In one embodiment of this application, the specific method for extracting texture contrast difference features of multiple candidate regions from enhanced image data includes:
[0078] The enhanced image data is scanned using a multi-scale sliding window to divide the multiple candidate regions; thereby, foreign objects of different sizes can be effectively covered.
[0079] For each candidate region, its gray-level co-occurrence matrix is calculated, and contrast, energy, and homogeneity features are extracted from the gray-level co-occurrence matrix.
[0080] It should be noted that the statistics of the gray-level co-occurrence matrix are in the direction ,distance Below, grayscale and The frequency of simultaneous occurrence By calculating contrast (C), energy (E), and homogeneity (H), artifacts caused by mucus vesicles, reflective spots, and intestinal wall folds can be effectively eliminated.
[0081] Calculate each candidate region and its surrounding neighboring organizations The texture contrast difference feature is generated by the difference in the contrast, energy, and homogeneity features.
[0082] ;in, The feature values of the candidate region, The average value is the value of the surrounding neighborhood organizations; This represents the degree of inconsistency in texture between a local area and its surrounding background environment. The system sets a threshold to... Areas with abnormally high values are marked as foreign object coordinates.
[0083] The texture contrast difference features are input into a preset classification model to determine foreign objects and obtain the probability distribution of foreign object targets.
[0084] Specifically, the texture contrast difference features are input into the feature fusion layer of the classification model to obtain a fused high-dimensional feature vector;
[0085] The high-dimensional feature vector is input into the fully connected layer of the classification model for calculation;
[0086] The calculation results of the fully connected layer are processed by the Softmax function in the output layer of the classification model to output the confidence score of the candidate region belonging to the preset foreign object category. It can be understood that the multi-classification mechanism of Softmax allows the system to determine multiple foreign object types at the same time.
[0087] ;in The output feature vector, It is a fully connected layer. This is the weight matrix. This is a bias term.
[0088] Based on the confidence scores of all candidate regions, a probability distribution of the foreign object target is generated, which characterizes the spatial distribution probability of the foreign object.
[0089] Calculate which region belongs to the first Confidence score of foreign objects: ,in, Represents the confidence score. The total number of preset foreign object categories (e.g., metal, plastic, biological skeleton, etc.).
[0090] The contour localization module is used to fit the edges of foreign objects in the enhanced image data according to the probability distribution of the foreign object target, and generate a foreign object localization result with marking information.
[0091] In summary, the automated foreign body detection and localization system in gastroenterology imaging of this application, by incorporating a Physical Information Neural Network (PINN) based on the law of light refraction, achieves physical restoration of foreign body image distortion in complex fluid environments, effectively correcting geometric distortion caused by interface refraction. Combining multispectral image data with a predefined lookup table compensation mechanism, the system can dynamically reconstruct the spectral energy distribution based on real-time liquid level depth, significantly improving the visual contrast between the foreign body and surrounding tissues. Furthermore, by extracting multidimensional texture features such as contrast, energy, and homogeneity of candidate regions using a gray-level co-occurrence matrix, and employing active contour fitting technology to achieve precise edge capture, the system ensures the clinical stability and robustness of localization results under interference from mucus, reflective surfaces, and turbid fluids. By utilizing a pre-defined lookup table to transform complex physical radiation transmission modeling into efficient matrix operations, the computational load on the foreign body recognition processor is significantly reduced, ensuring sub-pixel-level localization accuracy while guaranteeing real-time response during endoscopic examinations. In summary, this system can provide doctors with high-precision spatial coordinate marking and visually enhanced images of foreign objects, significantly shortening emergency diagnosis and intervention time and effectively reducing clinical risks such as postoperative perforation.
[0092] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.
[0093] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of embodiments.
[0094] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. An automatic foreign body detection and localization system in gastroenterology imaging, characterized in that, include: The data acquisition module is used to acquire multispectral image data of the digestive tract through the endoscope acquisition terminal, and to acquire real-time liquid level and depth information corresponding to the multispectral image data. The refraction correction module is used to input the multispectral image data and the real-time liquid level depth information into a preset refraction correction model, and to process the multispectral image data using the refraction correction model. The refraction correction model integrates a physical information neural network based on the law of light refraction. The refraction correction model performs reverse calculation of the refraction path and position mapping on the multispectral image data based on the real-time liquid level depth information, and outputs restored image data. The spectral enhancement module is used to perform adaptive spectral enhancement processing on the restored image data. It uses a preset spectral transformation matrix to redistribute the pixel values of different bands to generate enhanced image data. The foreign object detection module is used to extract texture contrast difference features of multiple candidate regions from enhanced image data; input the texture contrast difference features into a preset classification model to detect foreign objects and obtain the probability distribution of foreign object targets; The contour localization module is used to fit the edges of foreign objects in the enhanced image data according to the probability distribution of the foreign object target, and generate a foreign object localization result with marking information.
2. The automatic foreign body detection and localization system in gastroenterology imaging according to claim 1, characterized in that, The acquisition of multispectral image data and real-time liquid level depth information specifically includes: The multispectral imaging unit of the endoscope acquires multiple narrow-band images covering a predetermined wavelength range, forming multispectral image data; The distance between the endoscope lens and the surface of the liquid in the digestive tract is measured by the depth sensor module integrated into the endoscope, providing real-time liquid level and depth information.
3. The automatic foreign body detection and localization system in gastroenterology imaging according to claim 1, characterized in that, The training process of the refraction correction model includes: Construct a training dataset, the samples of which include multispectral images acquired in a fluid environment at a known depth and their corresponding non-refractive reference images; A neural network model is constructed, and the physical equations describing the refraction and classification of light at the fluid interface are incorporated as constraints into the loss function of the neural network model. The neural network model incorporating physical constraints is trained using the training dataset until it can accurately predict the refraction effect based on the input depth information and output the corrected image, thus obtaining the refraction correction model.
4. The automatic foreign body detection and localization system in gastroenterology imaging according to claim 1, characterized in that, The specific steps for adaptive spectral enhancement processing of the restored image data are as follows: The estimated absorption coefficients of the target fluid for each spectral channel at the current depth are obtained by querying a predefined lookup table based on the real-time liquid level depth information. Based on the estimated absorption coefficient, a spectral compensation matrix is constructed; The multispectral vectors of each pixel in the restored image data are multiplied by the spectral compensation matrix to achieve spectral reconstruction and obtain the enhanced image data.
5. The automatic foreign body detection and localization system in gastroenterology imaging according to claim 1, characterized in that, The texture contrast difference features are input into a preset classification model for foreign object detection to obtain the probability distribution of the foreign object target. The specific steps are as follows: The texture contrast difference features are input into the feature fusion layer of the classification model to obtain a fused high-dimensional feature vector. The high-dimensional feature vector is input into the fully connected layer of the classification model for calculation; The calculation results of the fully connected layer are processed by the Softmax function in the output layer of the classification model to output the confidence score of the candidate region belonging to the preset foreign object category. Based on the confidence scores of all candidate regions, a probability distribution of the foreign object target is generated, which characterizes the spatial distribution probability of the foreign object.
6. The automatic foreign body detection and localization system in gastroenterology imaging according to claim 1, characterized in that, The specific method for extracting texture contrast difference features of multiple candidate regions from enhanced image data includes: The enhanced image data is subjected to multi-scale sliding window scanning to divide the multiple candidate regions; For each candidate region, its gray-level co-occurrence matrix is calculated, and contrast, energy, and homogeneity features are extracted from the gray-level co-occurrence matrix. The texture contrast difference feature is generated by calculating the difference between each candidate region and its surrounding neighborhood organization in terms of contrast, energy, and homogeneity features.
7. The automatic foreign body detection and localization system in gastroenterology imaging according to claim 4, characterized in that, The lookup table records the estimated absorption coefficients of different fluid types at different depths for each spectral band.
8. The automatic foreign body detection and localization system in gastroenterology imaging according to claim 4, characterized in that, The refractive correction module is also used to: identify the fluid category in the current digestive tract based on the global color moment features of the multispectral image data through a clustering algorithm, and select the corresponding estimated absorption coefficient from the lookup table according to the fluid category.