HK40018665B - Method, apparatus, and medical system for processing image of digestive tract - Google Patents

Description

A method, apparatus, and medical system for processing images of the digestive tract.

Technical Field

This application relates to the field of artificial intelligence technology, and in particular to a method, apparatus, and medical system for processing digestive tract images.

Background Technology

With the continuous development of artificial intelligence, it is gradually being applied to medical image processing, using AI to assist doctors in diagnosis, thereby improving the speed and accuracy of diagnosis.

Currently, the general approach is to train a target detection model using a large number of medical image samples with labeled polyp locations, then use the trained target detection model to extract polyp features from the medical image samples to be processed, and finally mark the corresponding bounding boxes at the polyp locations in the medical image samples to be processed based on the identified polyp features.

However, this method can only identify polyps with obvious characteristics, and the accuracy of identifying lesion types is low.

Summary of the Invention

This application provides a method, apparatus, and medical system for processing digestive tract images to improve the accuracy of identifying lesion types.

Firstly, a method for processing digestive tract images is provided, comprising:

Acquire images of the digestive tract to be processed;

Obtain the lesion category of each lesion pixel in the digestive tract image;

Based on the lesion categories of each lesion pixel, a segmented image of the digestive tract is obtained, which is labeled with the lesion region and the lesion category. The lesion region is formed by lesion pixels of the same lesion category, and the lesion category of the lesion region is the lesion category of the pixels that form the lesion region.

Secondly, a processing apparatus for digestive tract images is provided, the processing apparatus comprising:

Acquisition unit, used to acquire images of the digestive tract to be processed;

The identification unit is used to obtain the lesion category of each lesion pixel in the digestive tract image through a trained digestive tract image segmentation model. The digestive tract image segmentation model is trained based on digestive tract image samples that are labeled with lesion regions and lesion categories.

The segmentation unit is used to obtain a segmented image of the digestive tract that is labeled with lesion regions and lesion categories based on the lesion categories of each lesion pixel. The lesion region is formed by lesion pixels of the same lesion category, and the lesion category of the lesion region is the lesion category of the pixels that form the lesion region.

In one possible implementation, the processing apparatus includes a generation unit, wherein:

The generation unit is used to, after obtaining a digestive tract segmentation image with lesion regions and lesion categories identified according to the lesion categories of each lesion pixel, superimpose the digestive tract segmentation image onto the digestive tract image to generate a digestive tract image including the identified lesion regions and lesion categories.

In one possible implementation, the processing device includes a first determining unit and a marking unit, wherein:

The first determining unit is used to determine the overlap rate of the digestive tract segmentation image with each digestive tract segmentation image in a preset number of digestive tract images before and after the digestive tract image is obtained according to the lesion category of each lesion pixel, and to determine the reliability of the digestive tract segmentation image based on the determined overlap rate.

The marking unit is used to mark the determined confidence level on the segmented image of the digestive tract.

In one possible implementation, the processing device further includes a training unit. The gastrointestinal image sample set includes a training set, a validation set, and a test set. Each gastrointestinal image sample in the gastrointestinal image sample set is labeled with a lesion region and a lesion category. The training unit is used for:

A digestive tract image segmentation model is trained based on the digestive tract image samples in the training set.

Based on the validation set, the evaluation index results for different segmentation results of the digestive tract image segmentation model under multiple sets of model parameters are determined, and multiple preferred values for the evaluation index results of different segmentation results of the digestive tract image segmentation model under multiple sets of model parameters are determined.

Based on the test set, determine the evaluation index result of the model parameter corresponding to each of the plurality of preferred values, and determine the optimal value of the evaluation index result of the model parameter corresponding to the plurality of preferred values;

The model parameters corresponding to the determined optimal values are used as the model parameters of the digestive tract image segmentation model to obtain the trained digestive tract image segmentation model.

In one possible implementation, the training unit is specifically used for:

Adjust the model parameters of the digestive tract image segmentation model according to the convergence direction of the preset error function;

The preset error function is obtained by weighting the value of the cross-entropy error function and the value of the minimized mixed error function. The minimized mixed error function is used to represent the similarity between the prediction result of the segmentation model for the lesion category of each pixel in the digestive tract image sample and the actual result for the lesion category of each pixel in the digestive tract image sample.

In one possible implementation, the digestive tract image segmentation model includes multiple digestive tract image segmentation models trained separately for digestive tract image sample sets of different image types. The processing device further includes a second determining unit, which is used to:

Before obtaining the lesion category of each lesion pixel in the digestive tract image through the trained digestive tract image segmentation model, the image type corresponding to the digestive tract image is identified.

Based on the image type corresponding to the digestive tract image, a digestive tract image segmentation model associated with the image type is determined; wherein, different image types are associated with different digestive tract image segmentation models.

Thirdly, a medical system is provided, comprising: an endoscope, an output module, and any of the processing devices discussed in the second aspect, wherein:

The endoscope is used to acquire images of the digestive tract to be processed and send them to the processing device;

The output module is used to output the segmented image of the digestive tract obtained by the processing device.

In one possible implementation, the system further includes an image filtering module, wherein:

The image filtering module is used to filter out digestive tract images that do not meet preset conditions from multiple digestive tract images acquired by the endoscope according to a trained digestive tract image filtering and recognition model, obtain digestive tract images to be processed, and send them to the processing device.

In one possible implementation, the system further includes an organ site identification module, wherein:

The organ identification module is used to obtain the digestive tract image to be processed from the image filtering module, obtain the organ identification result corresponding to the digestive tract image according to the trained organ classification and identification model, and send the organ identification result to the output module.

Fourthly, a computer device is provided, comprising:

At least one processor, and

A memory that is communicatively connected to the at least one processor;

The memory stores instructions executable by the at least one processor, which implements the method as described in any one of the first aspect and any possible implementation by executing the instructions stored in the memory.

Fifthly, a computer-readable storage medium is provided that stores computer instructions that, when executed on a computer, cause the computer to perform the method as described in any one of the first aspects and any possible embodiments.

Compared to existing technologies that use target detection models to detect polyps, this embodiment uses a digestive tract image segmentation model to determine the lesion category of each lesion pixel. Since the lesion category is determined at the pixel level, it can identify lesions even those with indistinct overall features, improving the accuracy of lesion type identification. Furthermore, because the lesion category is determined for each individual pixel, the lesion region can be accurately identified, improving the precision of lesion region identification and facilitating subsequent treatment and observation by doctors.

Attached Figure Description

Figure 1 is a schematic diagram of the structure of a medical system provided in an embodiment of this application;

Figure 2 is an example diagram of the process of processing digestive tract images by the medical system provided in the embodiment of this application;

Figure 3 is a schematic diagram of the structure of a digestive tract image processing device provided in an embodiment of this application;

Figure 4 is a schematic diagram of the deployment of various devices in a medical system provided in an embodiment of this application;

Figure 5 is a flowchart illustrating a method for processing digestive tract images according to an embodiment of this application;

Figure 6 is a flowchart illustrating a training model for digestive tract image segmentation provided in an embodiment of this application;

Figure 7 is a schematic diagram of the process of adjusting the model parameters of the digestive tract image segmentation model provided in the embodiment of this application;

Figure 8 is an example diagram of the actual results and predicted results provided in the embodiments of this application;

Figure 9 is an example diagram of different image types of digestive tract images provided in the embodiments of this application;

Figure 10 is a schematic flowchart of a method for determining the credibility of segmented images of the digestive tract provided in an embodiment of this application;

Figure 11 is an example diagram of determining the credibility of digestive tract segmentation images according to an embodiment of this application;

Figure 12 is an example diagram of digestive tract image to superimposed digestive tract segmentation image provided in an embodiment of this application;

Figure 13 is a structural diagram of a processing device provided in an embodiment of this application;

Figure 14 is a structural diagram of a medical system provided in an embodiment of this application;

Figure 15 is a structural diagram of a computer device provided in an embodiment of this application.

Detailed Implementation

To better understand the technical solutions provided in the embodiments of this application, a detailed description will be given below in conjunction with the accompanying drawings and specific implementation methods.

The following descriptions of some terms used in the embodiments of this application are provided to help those skilled in the art to better understand the technical solutions in the embodiments of this application.

The digestive tract is the tube connecting the mouth and anus, composed of many structures responsible for processing food. Digestive glands secrete digestive juices to digest food. The digestive tract includes the upper and lower digestive tracts. The upper digestive tract consists of the mouth, pharynx, esophagus, and stomach. The lower digestive tract includes the intestines and anus. The intestines include the small intestine, large intestine, and colon, etc.

Endoscope: Also known as an endoscope, it is a multidisciplinary tool used to explore the ducts and channels of an organism. It is typically used to observe areas not directly visible to the naked eye, allowing observation of internal structures and conditions within cavities, and enabling remote observation and manipulation. Endoscopes can include various imaging modes, each producing different medical image characteristics. Endoscopes include colonoscopes, gastroscopes, etc.

Gastrointestinal imaging: Essentially, these are images used to represent the digestive tract; they are a reproduction of the visual perception of the digestive tract. Since the digestive tract is located inside the human body, images of the digestive tract can generally be obtained using an endoscope.

Lesion: A site on an organism where disease has occurred or may occur. Lesions can spread pathogens and toxins, expanding the lesion. Lesions can be specifically classified into various categories, such as diffuse lesions based on their distribution pattern, and benign or malignant lesions based on their severity. If lesions are classified according to two different dimensions, a single lesion may belong to multiple categories; for example, a lesion may be both a diffuse lesion and a benign lesion.

Image type: Used to indicate the type of image of the digestive tract, such as imaging type and staining type. Imaging types include narrow band imaging (NBI), white light imaging, and infrared imaging. Staining types include iodine staining.

Machine Learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and instructional learning.

Convolutional Neural Network (CNN): Essentially, it is a mapping from input to output. It can learn a large number of mapping relationships between inputs and outputs without requiring any precise mathematical expressions between inputs and outputs. As long as the convolutional network is trained with known patterns, the network will have the ability to map between input-output pairs.

Deep neural networks are neural networks with at least one hidden layer. Similar to shallow neural networks, deep neural networks can also model complex nonlinear systems, but the additional layers provide a higher level of abstraction, thus improving the model's capabilities. Deep neural networks are typically feedforward neural networks, but research in areas such as language modeling has extended them to recurrent neural networks.

Gastrointestinal image segmentation model: This application uses gastrointestinal image samples labeled with lesion areas and lesion categories to train a model based on a neural network.

Encoder-decoder segmentation model: A type of segmentation model. The encoder includes convolutional layers and pooling layers, while the decoder is implemented through deconvolution operations.

Mask Region-Convolutional Neural Network (Mask-RCNN): A segmentation algorithm that includes four parts: candidate region selection, CNN feature extraction, classification and boundary regression, and mask prediction.

Image type recognition model: This application uses digestive tract image samples labeled with image types to train a model based on a deep neural network.

Gastrointestinal image screening and recognition model: This application uses gastrointestinal image samples that do not meet the preset conditions and trains the model based on a deep neural network.

Organ classification and identification model: This application uses digestive tract image samples labeled with organ types to train a model based on a deep neural network.

Overlap ratio: This can be understood as the degree of overlap between two segmented images of the digestive tract. A higher overlap ratio indicates a greater degree of overlap between the two segmented images. If the lesion regions in two segmented images of the digestive tract are of the exact same lesion type as their corresponding lesion regions, then the overlap ratio of these two segmented images is 1. The overlap ratio can be characterized by Intersection-over-Union (IoU) or Mean Intersection-over-Union (MIoU).

The design concept of the embodiments of this application will be explained below.

Statistics show that esophageal cancer, stomach cancer, and colon cancer, the three major malignant tumors of the digestive tract, have extremely high incidence and mortality rates, consistently ranking among the top types of malignant tumors. On the one hand, with the increasing aging population in my country, meeting the medical needs of the elderly is a huge social challenge. Furthermore, the number of professional doctors is limited, and the large number of patient images they handle daily can easily lead to missed or misdiagnosed cases. On the other hand, the varying skill levels of doctors and the complex symptoms of digestive tract diseases make accurate diagnosis impossible for all patients.

In recent years, machine learning (ML) has played an increasingly important role in the field of medical imaging. By assisting doctors in diagnosing medical images, ML has greatly improved the speed and accuracy of diagnosis, which is of great significance in alleviating the medical needs in my country.

In existing technologies, target detection models are generally used to detect polyp features in order to determine the location of polyps.

The inventors of this application have discovered that existing methods rely on the overall characteristics of polyps for target identification and diagnosis. While polyps are generally large, with well-defined and relatively regular shapes, allowing the target detection model to identify most polyps in the digestive tract, lesions have less distinct and more dispersed shapes. Therefore, the inventors have found that existing target detection models struggle to identify lesions directly in the digestive tract, resulting in low accuracy in classifying lesions.

The inventors of this application further discovered that the existing method is based on the overall characteristics of the polyp for target identification. The overlap between the identified polyp outline and the actual polyp outline is not high, and the determined polyp location is not accurate enough.

The inventors of this application further discovered that the existing method usually requires separate treatment for different organs in the digestive tract, which may require doctors to make different diagnoses and treatments based on the patient's organ location, making the operation quite cumbersome.

In view of this, the inventors of this application have designed a digestive tract image processing method based on machine learning. This method first determines the lesion category to which each lesion pixel belongs using a digestive tract image segmentation model, and then obtains a segmented digestive tract image labeled with lesion regions and lesion categories based on the lesion categories of each lesion pixel. Since the processing is performed on a pixel-by-pixel basis, it can improve the accuracy of determining lesion regions and lesion categories, as well as precisely locate the position of each lesion region.

The inventors of this application further discovered that the properties of various organs in the digestive tract are relatively similar. When examining a patient's body, the digestive tract is usually examined as a whole. Therefore, the inventors discovered that a digestive tract image segmentation model that is universal for all organs in the digestive tract can be trained to segment lesions on various organs in digestive tract images. On the one hand, this will not affect the accuracy of segmentation, and on the other hand, it can avoid the need to switch between different models when examining different organs, thereby reducing the processing load for doctors or equipment.

The inventors of this application further discovered that during a doctor's diagnosis, the imaging mode of the endoscope may be switched to more clearly detect lesions. For example, a doctor may initially use white light for diagnosis, but when a suspicious lesion is found, the endoscope may be switched to NBI (Non-Intense Blasting). Because different image types exhibit significant differences in color, texture, and other details, the inventors found that using different digestive tract image segmentation models for different image types can further improve the accuracy of identifying lesion types and locations.

The inventors of this application further discovered that endoscopy continuously acquires images of the digestive tract, and the digestive tract image segmentation model processes each image to determine the lesion region and type. The inventors found that even for two identical digestive tract images of the same location, the segmentation model may produce different results, and doctors may need to monitor a large number of segmented images. Therefore, the inventors further discovered that after obtaining segmented digestive tract images, the reliability of each segmented image can be verified based on preceding and following images, and a reliability rating can be assigned to each image. This rating allows for the filtering out of images with lower reliability, while enabling doctors to make diagnoses based on images with higher reliability, thus reducing the workload for doctors and improving diagnostic efficiency and accuracy.

After introducing the design concept of the embodiments of this application, the architecture of the medical system of the embodiments of this application will be described below. Please refer to Figure 1, which shows the medical system 100 involved in the embodiments of this application. Figure 1 can also be understood as a schematic diagram of an application scenario of a digestive tract image processing method involved in the embodiments of this application.

The medical system 100 shown in Figure 1 includes an endoscope 110 and a digestive tract image processing device 120.

Specifically, the endoscope 110 can be referred to the previous discussion and will not be repeated here. The digestive tract image processing device 120 can be implemented by a device with a graphics processing unit (GPU), such as a terminal device or a server. The terminal device is a personal computer, etc. The server can be a physical server or a virtual server, etc. The digestive tract image processing device 120 includes a digestive tract image processing unit 121, which can be understood as a part of the digestive tract image processing device 120, and is mainly used for segmenting digestive tract images.

The endoscope 110 can be integrated with the digestive tract image processing device 120. For example, a corresponding processing chip can be installed in the endoscope 110 to realize the functions of the digestive tract image processing device 120. The endoscope 110 can be integrated with the digestive tract image processing device 120 or set up relatively independently. For example, the endoscope 110 can communicate with the digestive tract image processing device 120 through an interface to realize the functions of the medical system 100.

After introducing the basic structure of the medical system 100, the functions of each device in the medical system 100 will be introduced below.

Endoscope 110: Used to acquire images of the digestive tract.

Specifically, the endoscope 110 acquires numerous images of the digestive tract. These multiple images may include digestive tract images of different image types. They may also include digestive tract images of varying image quality. Furthermore, they may include images of different organ sites within the patient's digestive tract.

The digestive tract image processing device 120 is used to obtain digestive tract images acquired by the endoscope 110, and to filter the digestive tract images by a digestive tract image screening and recognition model to filter out digestive tract images that do not meet the preset conditions in order to obtain digestive tract images that meet the preset conditions.

The digestive tract image processing device 120 then uses a trained organ classification and recognition model to identify the organ parts of the digestive tract images that meet the preset conditions, identifies the organ parts corresponding to each digestive tract image, and outputs the organ part identification results.

The digestive tract image processing device 120 identifies the image type of the digestive tract image that meets preset conditions using a trained image type recognition model. The processing unit 121 within the digestive tract image processing device 120 determines a digestive tract image segmentation model associated with the image type, and obtains a segmented digestive tract image labeled with lesion regions and lesion classifications using the digestive tract image segmentation model. The processing unit 121 is also used to overlay the segmented digestive tract image onto the digestive tract image to obtain a digestive tract image including lesion regions and lesion classifications.

For example, referring to Figure 2, the digestive tract image processing device 120 determines that a, b, c, e, and d have issues such as blurring, abnormal color, or overexposure. Therefore, it determines that a, b, c, and d do not meet the preset conditions, while it determines that e meets the preset conditions.

After obtaining image 'e', the digestive tract image processing device 120 determines the organ location corresponding to 'e'. The organ location may be, for example, the stomach, esophagus, pharynx, or intestine, and outputs the identified organ location result.

The digestive tract image processing device 120 determines the image type of e. The digestive tract image processing device 120 determines a digestive tract image segmentation model associated with that image type, and obtains a segmented digestive tract image f corresponding to e based on the digestive tract image segmentation model. After processing the segmented digestive tract image f, the digestive tract image processing device 120 overlays it onto a digestive tract image and outputs a digestive tract image including the lesion region and lesion type.

After introducing the medical system 100, the structure of the digestive tract image processing device 120 will be introduced below.

Referring to Figure 3, the digestive tract image processing device 120 includes a processor 310, a memory 320, a display panel 330, and an I/O interface 340.

Specifically, the memory 320 stores program instructions, and the processor 310 executes these instructions to implement the functions of the digestive tract image processing device 120 described above. The I/O interface 340 is used to enable interaction with the endoscope 110. The display panel 330 can be used to display segmented digestive tract images or digestive tract images including lesion regions and lesion types.

As one embodiment, the display panel 330 is an optional component. In this case, the digestive tract image processing device 120 may not display segmented digestive tract images or digestive tract images including lesion areas and lesion types. Instead, it may process the data corresponding to the segmented digestive tract images or digestive tract images including lesion areas and lesion types, and then display them on other imaging devices for easy viewing by doctors.

Based on the medical system 100 described in Figure 1, the following is an example illustration of the scenario deployment diagram of each device in the medical system 100.

Referring to Figure 4, the doctor inserts endoscope 110 into the patient's digestive tract, and endoscope 110 acquires images of the patient's digestive tract. Endoscope 110 sends the acquired digestive tract images to digestive tract image processing device 120, which processes the digestive tract images to generate segmented digestive tract images.

Figure 4 uses a personal computer as an example to illustrate the digestive tract image processing device 120, but in reality, the specific form of the digestive tract image processing device 120 is not limited. Figure 4 illustrates the wired communication method, but in reality, the endoscope 110 and the digestive tract image processing device 120 can communicate in any way, such as wired or wireless communication. Wireless communication methods include Bluetooth or Wi-Fi.

After introducing the medical system and equipment deployment examples involved in the embodiments of this application, the processing method of gastrointestinal images involved in the embodiments of this application will be described below. This method is executed by the gastrointestinal image processing apparatus 121 discussed above. This method involves image segmentation technology in machine learning, which will be specifically illustrated through the following embodiments.

Please refer to Figure 5, which is a flowchart of a method for processing digestive tract images according to an embodiment of this application. The method specifically includes:

S510, acquire images of the digestive tract to be processed.

The digestive tract images to be processed can be digestive tract images obtained from endoscope 110, or digestive tract images that meet preset conditions after being screened by digestive tract image processing device 120.

S520, obtain the lesion category of each lesion pixel in the digestive tract image.

After obtaining images of the digestive tract, it's possible that all pixels in the image are lesion pixels, some pixels are lesion pixels, or there are no lesion pixels. If lesion pixels are present in the digestive tract image, the lesion category corresponding to each lesion pixel can be determined. For example, the lesion category corresponding to each lesion pixel can be determined using a digestive tract image segmentation model. The digestive tract image segmentation model can be referred to the previous discussion and will not be repeated here.

In one possible embodiment, the digestive tract image to be processed is input into a trained digestive tract image segmentation model. The digestive tract image segmentation model outputs the probability of each pixel in the digestive tract image belonging to each category. The categories include background or various lesions. The lesion category with the highest category probability corresponding to the pixel can be determined as the category of the pixel, thus determining the lesion category corresponding to each lesion pixel.

S530: Based on the lesion category of each lesion pixel, obtain a segmented image of the digestive tract that is labeled with the lesion region and lesion category.

After obtaining the classification of each lesion pixel, pixels belonging to the same category can be grouped together to determine the corresponding lesion region. If pixels belonging to the same category are all clustered together, then the lesion region for that category includes one region. Pixels belonging to the same category may also be scattered across multiple regions, in which case the corresponding lesion region includes multiple regions. This process continues until each lesion region and its corresponding lesion category are obtained, which is equivalent to obtaining a segmented image of the digestive tract.

In one possible embodiment, different categories of lesion regions in the segmented images of the digestive tract are distinguished by different identifiers.

Specifically, gastrointestinal imaging may include various types of lesion areas. After identifying the corresponding lesion areas, different labels can be used to distinguish between different types of lesions. These labels may include different colors or different symbols.

In the embodiment shown in Figure 5, since the lesion category corresponding to each lesion pixel is determined on a pixel-by-pixel basis, compared with the existing technology that detects polyp location based on the overall features of the polyp, even if the shape features of the lesion are not obvious, the processing method in this embodiment can still determine the corresponding lesion region, improving the accuracy of lesion identification. Furthermore, since the lesion classification is determined for each lesion pixel, each lesion region can be determined more precisely, improving the accuracy of lesion location.

After introducing the overall idea of the processing method shown in Figure 5, the specific implementation methods involved in each step will be described below.

In embodiment S520 of this application, the lesion category of each lesion pixel can be determined by a trained digestive tract image segmentation model. Before executing S520, it is necessary to obtain a trained digestive tract image segmentation model. The training process of the digestive tract image segmentation model is illustrated below.

Please refer to Figure 6. The training process includes:

S610: Train a digestive tract image segmentation model based on digestive tract image samples in the training set.

Specifically, a suitable set of gastrointestinal image samples can be obtained from existing databases, or it can be obtained through manual annotation. The set of gastrointestinal image samples includes multiple gastrointestinal image samples, each annotated with lesion regions and lesion categories. After obtaining the set, it is divided into a training set, a validation set, and a test set. The training set is used to train the gastrointestinal image segmentation model; the validation set is used to select several preferred model parameters from the model parameters; and the test set is used to evaluate the preferred model parameters and determine the optimal model parameters.

S620, Based on the validation set, determine the evaluation index results of the digestive tract image segmentation model under different segmentation results with multiple sets of model parameters, and determine multiple optimal values of the evaluation index results of the digestive tract image segmentation model under different segmentation results with multiple sets of model parameters.

Specifically, digestive tract image segmentation models can employ encoder-decoder segmentation models or Mask-RCNN. Regardless of the model used, the training set needs to continuously adjust the model parameters during training. Then, a validation set is used to verify the evaluation metrics of the segmentation results under different model parameters, identifying several optimal values among the evaluation metrics. These optimal values can be determined based on the actual segmentation quality requirements, and each optimal value can be used to represent the value that best meets the segmentation quality requirements in this training session.

S630, determine the evaluation index result of the model parameter corresponding to each of the multiple preferred values based on the test set, until the optimal value of the evaluation index result of the model parameter corresponding to the multiple preferred values is determined;

Specifically, after obtaining the model parameters corresponding to multiple optimal values, the evaluation results of the model parameters corresponding to the multiple optimal values are determined based on the test set and preset evaluation indicators, thereby determining the optimal value. The evaluation indicators corresponding to the test set can be the same as the evaluation indicators selected for the test set.

S640, the model parameters corresponding to the determined optimal value are used as the digestive tract image segmentation model to obtain the trained digestive tract image segmentation model.

In the embodiment illustrated in Figure 6, evaluation metrics are used to validate the model parameters of the gastrointestinal image segmentation model. The optimal evaluation metric results are used as the model parameters of the trained gastrointestinal image segmentation model, ensuring the quality of the segmentation results. Furthermore, after obtaining multiple sets of optimized model parameters on the validation set, the optimal model parameters are obtained using the test set. This ensures that the final optimal model parameters have broad applicability and avoids overfitting in the trained gastrointestinal image segmentation model.

In one possible embodiment, in order to ensure the accuracy of the trained digestive tract image segmentation model, the digestive tract image sample set may include digestive tract image samples of various organs and parts of the digestive tract, as well as digestive tract image samples corresponding to various lesions.

In one possible embodiment, during the training of the digestive tract image segmentation model in S610, the specific adjustment of parameters is described in Figure 7. Figure 7 shows a flowchart of adjusting model parameters. In this embodiment, the training set is used to determine the value of the preset error function corresponding to the digestive tract segmentation model under different model parameters; the validation set is used to determine the evaluation index results corresponding to the digestive tract segmentation model under different model parameters; the test set is used to determine the optimal model parameters; and during the training process, the model parameters of the digestive tract segmentation model are adjusted in the convergence direction of the preset error function until the optimal evaluation index result is determined.

The preset error function is used to represent the error value between the actual results labeled in the training set and the prediction results of the digestive tract segmentation model under different model parameters. There are several ways to construct the preset error function, and examples are given below.

One preset error function is:

L _total =L _dice +λ*L _ce (1)

In equation (1), L _{_dice} represents minimizing the Dice Similarity Coefficient (dice) error function, and L _{_ce} represents the cross-entropy error function. λ represents the weight value used to balance the dice error function and the cross-entropy error function, and this weight value is a constant.

L _dice can be represented as:

In Equation (2), A represents the set of all pixels in the digestive tract segmentation model that predict a certain lesion category in a digestive tract image sample, and B represents the set of all pixels in the actual result that belong to the same lesion category. Minimizing the mixture error function represents the similarity between the predicted result and the actual result.

For example, please refer to Figure 8. Figure 8(1) includes three digestive tract image samples: a, b, and c. Figure 8(1) represents the category to which each pixel belongs in the actual result of the digestive tract image sample. Each small rectangle in Figure 8(1) represents a pixel. In Figure 8(1), "1" indicates that the pixel belongs to the lesion category, and "0" indicates that the pixel does not belong to the lesion category. Figure 8(2) represents the output result of the corresponding digestive tract image sample after passing through the digestive tract image segmentation model. The data in each small rectangle in Figure 8(2) represents the probability that each pixel belongs to the lesion category.

Taking the digestive tract image sample shown in Figure 8a as an example, A can be represented as:

B can be represented as: From this, we can obtain the value of L _dice corresponding to the digestive tract imaging sample a.

The cross-entropy error function _Lce can be expressed as:

L _ce =∑(y _i log(p _i )+(1-y _i )log(1-p _i )) (3)

In Equation (3), _yi represents the category to which each pixel in the actual result of the digestive tract image sample belongs, and pi represents the probability of each pixel in the digestive tract image sample output by the digestive tract image segmentation model belonging to the category. The cross-entropy error function is used to compare the predicted pixel probability distribution (across all classes) with the actual probability distribution of all pixels.

After the digestive tract image segmentation model processes each digestive tract image sample, the value of the preset error function corresponding to the digestive tract segmentation model under the model parameters can be calculated, for example, using formulas (1), (2), and (3) mentioned above. After obtaining the value of the preset error function, the model parameters of the digestive tract image segmentation model are adjusted in the direction of convergence of the preset error function.

The digestive tract image samples from the validation set are input into the digestive tract image segmentation model under the model parameters. Based on the prediction results and actual results of the digestive tract image segmentation model, the evaluation index results corresponding to the digestive tract segmentation model under the model parameters are calculated using the evaluation index.

After obtaining multiple sets of model parameters and corresponding evaluation index results, the model parameters corresponding to the optimal evaluation index results can be selected as the model parameters of the digestive tract image segmentation model.

For example, the evaluation metric could be the Mean Intersection over Union (MIoU):

In equation (4), k represents the total number of lesions corresponding to all categories, p_{_ii} represents the total number of pixels of true category i predicted as category i, and p_{_ij} represents the total number of pixels of true category i predicted as category j. MIoU is used to represent the overlap rate between the true and predicted results in the gastrointestinal imaging sample.

After obtaining the trained digestive tract image segmentation model, step 502 can be executed directly based on the trained digestive tract image segmentation model.

Alternatively, in one possible embodiment, since images of different image types differ significantly, in order to improve the accuracy of the digestive tract image segmentation model in segmenting digestive tract images, the digestive tract image segmentation model in this embodiment is trained separately for digestive tract image sample sets of different image types.

Specifically, a digestive tract image segmentation model for that image type is trained using a sample set of digestive tract images of the same image type, resulting in digestive tract image segmentation models for different image types. In other words, the digestive tract image segmentation models associated with different image types are different. This difference can be understood as the digestive tract image segmentation models corresponding to different image types not being entirely identical; it could be that the network structures of the corresponding digestive tract image segmentation models are not completely identical, or that the model parameters of the digestive tract image segmentation models are different. After obtaining the digestive tract image segmentation models corresponding to different image types, the correlation between these models can be preserved.

For example, different image types may correspond to the same digestive tract image segmentation model, but because they are trained separately, the model parameters of the digestive tract image segmentation models for different image types are different. For example, the specific relationship between image type and the model parameters of a given digestive tract image segmentation model is shown in Table 1 below.

Table 1

Furthermore, if the digestive tract image segmentation model is trained separately using digestive tract image sample sets of different image types, before executing S520, it is necessary to first determine the digestive tract image segmentation model associated with the image type of the digestive tract image.

Specifically, the image type corresponding to the digestive tract image can be determined using the image type recognition model discussed earlier. The image type can be referred to in the previous discussion and will not be repeated here. After identifying the image type corresponding to the digestive tract image, the digestive tract image segmentation model corresponding to that image type can be determined based on the image type and the correlation discussed earlier.

For example, referring to Figure 9, we can determine that the image types of the digestive tract images shown in Figure 9 a, b, and c are white light, NBI, and iodine staining, respectively. Continuing with the correlation shown in Table 1, we can determine that the model parameters used for a, b, and c in Figure 9 are the first set of model parameters, the second set of model parameters, and the third set of model parameters, respectively.

For the same location, an endoscope can acquire multiple images of the digestive tract within a preset time period. Theoretically, the resulting segmented images of the digestive tract from the same location should be identical. However, due to various reasons, the determined segmented images may differ significantly. Too many segmented images increase the workload for doctors and may even lead to misdiagnosis. Therefore, in one possible embodiment, the reliability of each segmented image is determined. The following example illustrates the method for determining the reliability of the segmented images.

Please refer to Figure 10. The specific methods for determining credibility include:

S1001, determine the overlap rate of each digestive tract segmentation image in the digestive tract segmentation image and the digestive tract segmentation image of the preset digestive tract images before and after the digestive tract image.

S1002, Based on the determined overlap rate, determine the reliability of the digestive tract segmentation image, and mark the determined reliability on the digestive tract segmentation image.

In this embodiment, the reliability of the digestive tract segmentation image is determined based on a preset number of digestive tract segmentation images before and after the digestive tract segmentation image, and the reliability is marked on the digestive tract segmentation image, so that doctors can make judgments based on the digestive tract segmentation images with high reliability.

After introducing the overall concept of the embodiments of this application, the specific method of executing S1001 is described below:

When the endoscope 110 acquires images of the digestive tract, it typically does so continuously. Within a preset time period, the endoscope 110 acquires multiple identical images of the same digestive tract region. Therefore, the reliability of a digestive tract segmentation image can be determined based on the number of pre- and post-preset digestive tract images. The pre- and post-preset number can be understood as the number of digestive tract images preceding and following the first image. For the first digestive tract image acquired by the endoscope 110, since the first digestive tract image does not have preceding images, it can be processed using the number of post-preset digestive tract images relative to the first image. After segmenting multiple digestive tract images using the processing method described in Figure 5, multiple digestive tract segmentation images corresponding to the multiple images can be obtained.

The overlap rate between this digestive tract image and each of the other digestive tract segmented images is determined. A higher overlap rate indicates a lower probability of missegmentation, while a lower overlap rate indicates a higher probability of missegmentation. Other digestive tract segmented images can be understood as digestive tract segmented images other than this one.

For example, the digestive tract segmentation image to be determined is c, and the multiple digestive tract segmentation images corresponding to the preset number of digestive tract images before and after include a, b, d and e, as shown in Figure 11.

One method for determining the overlap rate is as follows:

When determining the overlap rate of a segmented image of a digestive tract, the MIoU mentioned earlier can be used to characterize the overlap rate between a segmented image of a digestive tract and other segmented images of the digestive tract. The following is a detailed explanation.

Where TP represents the true value, specifically the intersection of this digestive tract segmentation image with each of the other digestive tract segmentation images; FP represents the false positive value, specifically the portion of each digestive tract segmentation image minus TP; FN represents the false negative value, specifically the portion of this digestive tract segmentation image minus TP. k represents the total number of lesion categories.

One method for determining the overlap rate is as follows:

The overlap ratio (IoU) can be used to characterize the overlap rate. The specific calculation formula is as follows:

TP, FP, and FN can be referred to the previous discussion, and will not be repeated here.

For example, continuing to refer to Figure 11, we obtain the overlap rate m1 between c and a, the overlap rate m2 between c and b, the overlap rate m3 between c and d, and the overlap rate m4 between c and e in Figure 11.

After obtaining the overlap rate between the segmented digestive tract image and each other segmented digestive tract images, multiple overlap rates can be obtained. Executing S1002 allows the reliability of the segmented digestive tract image to be determined based on these multiple overlap rates. The execution method of S1002 will be explained below.

Method 1:

The number of overlap rates greater than a preset threshold is determined. If the determined number is greater than the preset number, the credibility of the digestive tract segmentation image is determined to be high. If the determined number is less than or equal to the preset number, the credibility of the digestive tract segmentation image is determined to be low.

Specifically, this method evaluates the reliability of a digestive tract segmentation image based on the number of overlap values exceeding a preset threshold, determining whether the image is high or low. This method only requires identifying the overlap values that meet the threshold, making the determination process simple. Furthermore, subsequent labeling of high or low reliability facilitates later screening and review.

For example, referring to Figure 11, with a preset number of 3, we determine that m1, m2, and m3 are greater than a preset threshold, and thus determine that the credibility of the digestive tract segmentation image is high.

Method 2:

The average of multiple overlap rates is determined, and the average value is used as the confidence level of the digestive tract segmentation image.

In this method, the average value is used as the reliability of the digestive tract segmentation image, which makes it easier to compare the reliability of each digestive tract segmentation image.

For example, continuing to refer to Figure 11, determine the average values of m1, m2, m4 and m3, and then determine the average value of the digestive tract segmentation image.

After determining the reliability of the digestive tract segmentation image, the reliability of the digestive tract segmentation image can be marked on the digestive tract segmentation image to facilitate doctors' review of each digestive tract segmentation image later.

In one possible embodiment, after determining the confidence level, digestive tract segmentation images with a confidence level less than or equal to a preset confidence level can be deleted, and digestive tract segmentation images with a confidence level greater than the preset confidence level can be output, so that all digestive tract segmentation images viewed by doctors are digestive tract segmentation images with high confidence level.

After obtaining segmented images of the digestive tract, or after obtaining segmented images of the digestive tract labeled with confidence level, in this embodiment of the application, the segmented images of the digestive tract can also be overlaid.

Specifically, overlay can be understood as integrating the effective information from two images. It can be further understood as superimposing the lesion area and lesion type in the segmented digestive tract image onto the corresponding area in the segmented digestive tract image to obtain a digestive tract image that includes both the lesion area and lesion type.

For example, please refer to Figure 12. In Figure 12, a represents the digestive tract image acquired by endoscopy, b represents a schematic diagram after magnifying the lesion in a, and c represents the lesion area obtained after segmenting the lesion area shown in b using an encoder-decoder segmentation model. By superimposing c onto a, a segmented digestive tract image including the lesion area and lesion type is obtained as shown in Figure d.

Based on the same inventive concept, this application provides a processing apparatus for digestive tract images. Referring to FIG13, the processing apparatus 121 includes:

Acquisition unit 1301 is used to acquire images of the digestive tract to be processed;

The identification unit 1302 is used to obtain the lesion category of each lesion pixel in the digestive tract image through the trained digestive tract image segmentation model. The digestive tract image segmentation model is trained based on digestive tract image samples that are labeled with lesion regions and lesion categories.

The segmentation unit 1303 is used to obtain a digestive tract segmentation image labeled with lesion region and lesion category based on the lesion category of each lesion pixel; wherein the lesion region is formed by lesion pixels of the same lesion category, and the lesion category of the lesion region is the lesion category of the pixels that form the lesion region.

In one possible embodiment, the processing device 121 includes a generation unit 1304, wherein:

The generation unit 1304 is used to, after obtaining a digestive tract segmentation image with lesion regions and lesion types identified according to the lesion types of each lesion pixel, superimpose the digestive tract segmentation image onto the digestive tract image to generate a digestive tract image including the identified lesion regions and lesion types.

In one possible embodiment, the processing device 121 includes a first determining unit 1305 and a marking unit 1306, wherein:

The first determining unit 1305 is used to determine the overlap rate of the digestive tract segmentation image with each digestive tract segmentation image in the digestive tract segmentation image with the lesion region and lesion category identified according to the lesion category of each lesion pixel, and to determine the credibility of the digestive tract segmentation image based on the determined overlap rate.

The labeling unit 1306 is used to mark the determined confidence level on the segmented image of the digestive tract.

In one possible embodiment, the processing device 121 further includes a training unit 1307, wherein the digestive tract image sample set includes a training set, a validation set, and a test set; each digestive tract image sample in the digestive tract image sample set is labeled with a lesion region and a lesion category, and the training unit 1307 is used for:

A digestive tract image segmentation model is trained based on digestive tract image samples in the training set.

Based on the validation set, the evaluation index results of the digestive tract image segmentation model under different segmentation results under multiple sets of model parameters are determined, and multiple optimal values of the evaluation index results of the digestive tract image segmentation model under different sets of model parameters are determined.

Based on the test set, determine the evaluation index results of the model parameters corresponding to each of the multiple optimal values, until the optimal value of the evaluation index results of the model parameters corresponding to the multiple optimal values is determined;

The model parameters corresponding to the determined optimal values are used as different parameters for the digestive tract image segmentation model to obtain the trained digestive tract image segmentation model.

In one possible embodiment, the training unit 1307 is specifically used for:

Adjust the model parameters of the digestive tract image segmentation model according to the convergence direction of the preset error function;

The preset error function is obtained by weighting the cross-entropy error function and the minimized mixed error function. The minimized mixed error function is used to represent the similarity between the segmentation model's prediction of the lesion category of each pixel in the sample digestive tract image and the actual result of the lesion category of each pixel in the sample digestive tract image.

In one possible embodiment, the digestive tract image segmentation model includes multiple digestive tract image segmentation models trained separately for digestive tract image sample sets of different image types. The processing device 121 further includes a second determining unit 1308, which is used for:

Before obtaining the lesion category of each lesion pixel in the digestive tract image through the trained digestive tract image segmentation model, the image type of the digestive tract image to be processed is identified, and the digestive tract image segmentation model associated with the image type is determined according to the image type; different image types are associated with different digestive tract image segmentation models.

As one embodiment, the generation unit 1304, the first determination unit 1305, the labeling unit 1306, the training unit 1307, and the second determination unit 1308 in FIG13 are optional units.

Based on the same inventive concept, this application provides a medical system 100. Referring to Figure 14, the medical system 100 includes the endoscope 110 discussed above, the output module 1401, and the processing device 121 discussed above. The processing device 121 can be referred to the content discussed in Figure 13 above, and will not be repeated here. Wherein:

Endoscope 110 is used to acquire images of the digestive tract to be processed and send them to processing device 121;

Output module 1401 is used to output segmented images of the digestive tract obtained by the processing device.

In one possible embodiment, the digestive tract image processing device 120 in the medical system 100 further includes an image screening module 1402, wherein:

The image filtering module 1402 is used to filter out digestive tract images that do not meet preset conditions from multiple digestive tract images acquired by endoscopy based on a trained digestive tract image filtering and recognition model, obtain digestive tract images to be processed, and output them to the image type recognition module.

Specifically, the image filtering module 1402 is used to obtain gastrointestinal images from the endoscope 110. Through a pre-trained gastrointestinal image filtering and recognition module, images that do not meet preset conditions are filtered out from multiple gastrointestinal images. This reduces subsequent processing workload and prevents these low-quality images from affecting the judgment results. The gastrointestinal image filtering and recognition model can be referred to the previous discussion and will not be repeated here. Preset conditions include, for example, the resolution of the gastrointestinal image being within a preset resolution range, the saturation of the gastrointestinal image being within a preset saturation range, or the brightness of the gastrointestinal image being within a preset brightness range.

In one possible embodiment, the digestive tract image processing device 120 in the medical system 100 further includes an organ location recognition module 1403, wherein:

The organ part recognition module 1403 is used to obtain the digestive tract image to be processed from the image screening module, and obtain the organ recognition result corresponding to the digestive tract image according to the trained organ classification and recognition model, and send the organ recognition result to the output module 1401.

The organ identification module 1403 is used to obtain the remaining digestive tract images after screening from the image screening module 1402, and to identify the organs corresponding to each digestive tract image according to the pre-trained digestive tract image organ classification and recognition model, and output the organ identification results to identify the corresponding organs, so as to facilitate doctors to make subsequent diagnoses for different organs.

As one embodiment, after obtaining the organ identification result, the organ part identification module 1403 can also send the digestive tract image labeled with the organ identification result to the processing device 121. After processing the digestive tract image labeled with the organ identification result, the processing device 121 outputs a digestive tract segmentation image carrying the organ identification result, lesion area, and lesion type.

As one embodiment, the image screening module 1402 and the organ part recognition module 1403 in the medical system 100 are optional modules.

Based on the same inventive concept, this application provides a computer device 1500, as shown in FIG15, which includes a processor 1501 and a memory 1502.

The processor 1501 can be a central processing unit (CPU) or a digital processing unit, etc. This embodiment does not limit the specific connection medium between the memory 1502 and the processor 1501. In Figure 15, the memory 1502 and the processor 1501 are connected via a bus 1503, which is represented by a thick line. The connection methods between other components are only illustrative and not intended to be limiting. The bus 1503 can be an address bus, a data bus, a control bus, etc. For ease of illustration, only one thick line is used in Figure 15, but this does not indicate that there is only one bus or one type of bus.

Memory 1502 may be volatile memory, such as random-access memory (RAM); memory 1502 may also be non-volatile memory, such as read-only memory, flash memory, hard disk drive (HDD), or solid-state drive (SSD); or memory 1502 may be any other medium capable of carrying or storing desired program code in the form of instructions or data structures, and accessible by a computer, but is not limited thereto. Memory 1502 may be a combination of the above-described memories.

The processor 1501 is used to execute the methods involved in the various devices in the embodiments shown in Figures 5 to 12 when calling the computer program stored in the memory 1502.

Based on the same inventive concept, embodiments of this application provide a computer-readable storage medium storing computer instructions that, when executed on a computer, cause the computer to perform the methods involved in the various devices in the embodiments shown in Figures 5 to 12.

Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims (12)

1. A method for processing digestive tract images, characterized in that it includes: Acquire images of the digestive tract to be processed; Identify the image type corresponding to the digestive tract image; the image type is used to characterize the imaging type or staining type of the digestive tract image; Based on the image type corresponding to the digestive tract image, a digestive tract image segmentation model associated with the image type is determined; wherein, different image types are associated with different digestive tract image segmentation models, and the digestive tract image segmentation model includes multiple digestive tract image segmentation models trained separately for digestive tract image sample sets of different image types. The lesion category of each lesion pixel in the digestive tract image is obtained by determining the digestive tract image segmentation model; Based on the lesion categories of each lesion pixel, a segmented image of the digestive tract is obtained, which is labeled with the lesion region and the lesion category. The lesion region is formed by lesion pixels of the same lesion category, and the lesion category of the lesion region is the lesion category of the pixels that form the lesion region. 2. The method as described in claim 1, characterized in that, after obtaining a segmented digestive tract image labeled with lesion regions and lesion categories based on the determined lesion categories of each lesion pixel, the method further includes: The segmented images of the digestive tract are superimposed onto the digestive tract images to generate digestive tract images that include the identified lesion regions and lesion types. 3. The method as described in claim 1, characterized in that, after obtaining a segmented image of the digestive tract labeled with lesion regions and lesion categories based on the determined lesion categories of each lesion pixel, the method further includes: Determine the overlap rate between the digestive tract segmentation image and each digestive tract segmentation image in a preset number of digestive tract images before and after the digestive tract image; Based on the determined overlap rate, the reliability of the segmented digestive tract image is determined, and the determined reliability is marked on the segmented digestive tract image. 4. The method according to any one of claims 1-3, characterized in that the digestive tract image sample set includes a training set, a validation set, and a test set, wherein each digestive tract image sample in the digestive tract image sample set is labeled with a lesion region and a lesion category, and the digestive tract image segmentation model is trained by the following method: A digestive tract image segmentation model is trained based on the digestive tract image samples in the training set. Based on the validation set, the evaluation index results for different segmentation results of the digestive tract image segmentation model under multiple sets of model parameters are determined, and multiple preferred values for the evaluation index results of different segmentation results of the digestive tract image segmentation model under multiple sets of model parameters are determined. Based on the test set, determine the evaluation index result of the model parameter corresponding to each of the plurality of preferred values, and determine the optimal value of the evaluation index result of the model parameter corresponding to the plurality of preferred values; The model parameters corresponding to the determined optimal values are used as the model parameters of the digestive tract image segmentation model to obtain the trained digestive tract image segmentation model. 5. The method as described in claim 4, characterized in that training the digestive tract image segmentation model based on the training set includes: Adjust the model parameters of the digestive tract image segmentation model according to the convergence direction of the preset error function; The preset error function is obtained by weighting the cross-entropy error function and the minimized mixed error function. The minimized mixed error function is used to represent the similarity between the segmentation model's prediction result for the lesion category of each pixel in the digestive tract image sample and the actual result for the lesion category of each pixel in the digestive tract image sample. 6. A processing apparatus for digestive tract images, characterized in that the processing apparatus comprises: Acquisition unit, used to acquire images of the digestive tract to be processed; The identification unit is used to identify the image type corresponding to the digestive tract image; determine the digestive tract image segmentation model associated with the image type according to the image type; and obtain the lesion category of each lesion pixel in the digestive tract image through the determined digestive tract image segmentation model; wherein, the image type is used to characterize the imaging type or staining type of the digestive tract image; different image types are associated with different digestive tract image segmentation models, and the digestive tract image segmentation model includes multiple digestive tract image segmentation models trained separately for digestive tract image sample sets of different image types, wherein each digestive tract image sample in the digestive tract image sample set is labeled with lesion region and lesion category; The segmentation unit is used to obtain a segmented image of the digestive tract that is labeled with lesion regions and lesion categories based on the lesion categories of each lesion pixel. The lesion region is formed by lesion pixels of the same lesion category, and the lesion category of the lesion region is the lesion category of the pixels that form the lesion region. 7. The processing apparatus as claimed in claim 6, characterized in that the processing apparatus further comprises a training unit, the digestive tract image sample set including a training set, a validation set, and a test set, the training unit being used for: A digestive tract image segmentation model is trained based on the digestive tract image samples in the training set. Based on the validation set, the evaluation index results for different segmentation results of the digestive tract image segmentation model under multiple sets of model parameters are determined, and multiple preferred values for the evaluation index results of different segmentation results of the digestive tract image segmentation model under multiple sets of model parameters are determined. The evaluation index result of the model parameter corresponding to each of the plurality of preferred values is determined based on the test set, until the optimal value of the evaluation index result of the model parameter corresponding to the plurality of preferred values is determined; The model parameters corresponding to the determined optimal values are used as the model parameters of the digestive tract image segmentation model to obtain the trained digestive tract image segmentation model. 8. A medical system, characterized in that it comprises: an endoscope, an output module, and a processing device as described in claim 6 or 7, wherein: The endoscope is used to acquire images of the digestive tract to be processed and send them to the processing device; The output module is used to output the segmented image of the digestive tract obtained by the processing device. 9. The medical system as described in claim 8, wherein the medical system further comprises an image screening module, wherein: The image filtering module is used to filter out digestive tract images that do not meet preset conditions from multiple digestive tract images acquired by the endoscope according to a trained digestive tract image filtering and recognition model, obtain digestive tract images to be processed, and send them to the processing device. 10. The medical system as described in claim 9, wherein the medical system further comprises an organ site identification module, wherein: The organ identification module is used to obtain the digestive tract image to be processed from the image filtering module, obtain the organ identification result corresponding to the digestive tract image according to the trained organ classification and identification model, and send the organ identification result to the output module. 11. A computer device, characterized in that it comprises: At least one processor, and A memory that is communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor, and the at least one processor implements the method as described in any one of claims 1-5 by executing the instructions stored in the memory. 12. A computer-readable storage medium, characterized in that the computer-readable storage medium stores computer instructions that, when executed on a computer, cause the computer to perform the method as described in any one of claims 1-5.