Medical image processing system and processing method

Abstract

The medical image processing system according to the present invention comprises: a communication unit for receiving a captured image captured by a capsule endoscope; and a processing unit that generates a readable image from the photographed image by using a lesion model result learned by a lesion classification method, the processing unit applies the photographed image to a pre-learned abnormal lesion classification model to classify an image including a preset lesion, and then the image including the preset lesion is stored in the abnormal lesion classification model. And classifying the classified images into lesion images and normal images again, screening representative images representing the lesion images from the classified lesion images, and generating the readable image from the selected representative images.

Description

TECHNICAL FIELD

[0001] The present application relates to a medical image processing system and a processing method, and more particularly, to a medical image processing system and a processing method from a capsule endoscope. BACKGROUND

[0002] Generally, a capsule endoscope is used to diagnose various small intestinal diseases. In the diagnosis of small intestinal diseases, a lesion is diagnosed based on a medical image obtained from a capsule endoscope inserted into the body.

[0003] A conventional capsule endoscope-related technology has been disclosed in "Republic of Korea Patent Gazette No. 10-1969982 (Capsule endoscope device, magnetic controller, and capsule endoscope system, April 11, 2019)". The disclosed invention can acquire a medical image through a capsule endoscope inserted into the body and diagnose a lesion based on the medical image.

[0004] In the diagnosis of such a lesion, an abnormal lesion is detected from a medical image, and a readable image is generated. Also, a clinician judges a lesion through the readable image. In the process of diagnosing a lesion, the clinician studies numerous video frames included in the readable image. In particular, it is known that the time for a specialist to interpret a lesion through a video frame is 30 minutes to 2 hours or more depending on the proficiency. Ultimately, in diagnosing a lesion, the time for generating a readable image and the time for interpreting a lesion are required. Therefore, it has a disadvantage in that a long time is required to diagnose a lesion. SUMMARY

[0005] PROBLEMS TO BE SOLVED BY THE INVENTION

[0006] The present application aims to provide a medical image processing system and a processing method that shorten the generation time of a readable image and reduce the number of video frames of a readable image provided to a clinician.

[0007] SOLUTION TO THE PROBLEM

[0008] According to the present application, the processing system of a medical image includes a communication unit that receives a photographed image photographed by a capsule endoscope, and a processing unit that generates a readable image from the photographed image using a lesion model result learned through a classification method of a lesion, the processing unit applying the photographed image to a pre-learned abnormal lesion classification model, classifying an image including a predetermined lesion, then classifying the classified image into a lesion image and a normal image, selecting a representative image representing the lesion image from the classified lesion image, and generating the selected representative image as the readable image.

[0009] The abnormal lesion classification model can include a YOLO (You Only Look Once) series model.

[0010] The abnormal lesion classification model can include at least one of YOLO v4 and YOLO v8.

[0011] In the learning data label of the abnormal lesion classification model, an expert can confirm the lesion existing in the learning data, and mark the lesion in the text and the lesion existing in the learning data by applying a preset marking method.

[0012] The marking of the lesion can identify a quadrilateral or a polygon around the lesion.

[0013] The taken image includes at least one of a hemorrhagic lesion, an inflammatory lesion, a vascular lesion, and a polyp lesion. The abnormal lesion classification model can classify the taken image into the lesion image and the normal image in a binary classification manner regardless of the type of the lesion in the re-classification process.

[0014] The abnormal lesion classification model can apply a bounding box to the lesion included in the lesion image.

[0015] The abnormal lesion classification model calculates a threshold value for each lesion including at least one of the hemorrhagic lesion, the inflammatory lesion, the vascular lesion, and the polyp lesion by statistics. When the taken image is provided, the lesion image and the normal image are classified in a binary classification manner based on the threshold value for each lesion.

[0016] In the screening of the representative image, a pre-learned video frame processing model is applied. The video frame processing model can extract features of color and texture of a plurality of lesion images, and compare similarity between images.

[0017] In analyzing the similarity, the similarity can be analyzed based on Bhattacharyya distance.

[0018] In the analysis of the similarity, HSV transformation is performed on a plurality of lesion images for mutual comparison, the number of bins of color and hue values is adjusted, and a histogram of color values (H-Histogram) and a histogram of hue values (S-Histogram) are calculated.

[0019] In the screening of the representative image, the standard image and the subsequent image are sequentially compared. When a subsequent image having a different similarity from the standard image appears, the subsequent image having the different similarity is used as the standard image, and the standard image and the subsequent image are sequentially compared to screen the standard image as the representative image.

[0020] When the representative images are screened, the number of images between a pair of representative images can be limited to n.

[0021] When the representative images are screened, the standard image and the subsequent image are compared in sequence, and the subsequent image n images after the standard image is selected as the standard image, and the standard image is selected as the representative image.

[0022] In addition, according to the present application, the medical image processing method includes the steps of receiving a captured image captured by a capsule endoscope; applying the captured image to a pre-learned abnormal lesion classification model to classify images including a predetermined lesion, and then classifying the classified images into lesion images and normal images; selecting a representative image representing the lesion images from the classified lesion images; and generating the selected representative image as a readable image.

[0023] Effects of the Invention

[0024] According to the present application, the medical image processing system and the processing method can shorten the time for reading medical images and providing recommendations, improve the reading efficiency of clinicians, and shorten the waiting time of patients.

[0025] The above technical effects of the present application are not limited to the effects described above, and the skilled person in the technical field can clearly understand the additional technical effects not mentioned from the following description. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 is a diagram of the structure of the medical image processing system according to the present embodiment.

[0027] Figure 2 is a diagram of the concept of loading a diagnostic assistance algorithm in the medical image processing system according to the present embodiment.

[0028] Figure 3 is a diagram of the concept of an abnormal lesion classification model of the diagnostic assistance algorithm according to the present embodiment.

[0029] Figure 4 is a diagram of the concept of a labeling method of the abnormal lesion classification model according to the present embodiment.

[0030] Figure 5 is a diagram of the concept of a learning method of the abnormal lesion classification model according to the present embodiment.

[0031] Figure 6 is a diagram of the concept of a similarity evaluation method for extracting a representative frame from a video frame processing model of the diagnostic assistance algorithm according to the present embodiment.

[0032] Figure 7This is a diagram illustrating the concept of a representative frame forcibly extracted from the video frame processing model of the diagnostic assistance algorithm described in this embodiment.

[0033] Figure 8 This is a diagram illustrating the concept of extracting representative frames from the video frame processing model of the diagnostic assistance algorithm described in this embodiment.

[0034] Figure 9 This is a diagram illustrating the concept of an abnormal lesion classification model for a diagnostic assistance algorithm described in another embodiment. Detailed Implementation

[0035] Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. However, these embodiments are not limited to the embodiments disclosed below, but can be implemented in many different forms. These embodiments are intended to fully disclose the invention and to fully inform those skilled in the art of its scope. For the sake of clarity, the shapes of elements in the drawings may be exaggerated. In the drawings, elements represented by the same symbols are the same elements.

[0036] <Medical image processing system>

[0037] Figure 1 This is a schematic diagram of the structure of the medical image processing system described in this embodiment. Figure 2 This is a schematic diagram of the medical image processing system described in this embodiment loading a diagnostic auxiliary algorithm.

[0038] like Figure 1 and Figure 2 As shown, the medical image processing system 1000 (hereinafter referred to as the processing system) described in this embodiment receives images of the small intestine from the capsule endoscope 10. The processing system 1000 filters out lesion images, including abnormal lesions, from the images. Furthermore, the processing system 1000 reduces the number of selected lesion images to generate optimal readable images.

[0039] In one example, the processing system 1000 may include a communication unit 100, a storage unit 200, a processing unit 300, and a display unit 400.

[0040] The communication unit 100 receives the captured images from the capsule endoscope 10. At this time, the communication unit 100 can communicate with the capsule endoscope 10 to receive the captured images. However, the communication method of the communication unit 100 can be various and is not limited to human body communication.

[0041] The storage unit 200 stores a diagnostic assistance algorithm 210 for generating readable images based on captured images. The diagnostic assistance algorithm 210 may include an abnormal lesion classification model 211 and a video frame processing model 212. The abnormal lesion classification model 211 and the video frame processing model 212 may be deep learning models that have been pre-learned.

[0042] The processing unit 300 generates a readable image based on the captured image. The processing unit 300 sequentially applies the captured image to the abnormal lesion classification model 211 and the video frame processing model 212. Thus, the processing unit 300 generates a readable image based on the diagnostic assistance algorithm 210.

[0043] The display unit 400 can output the processing procedure for readable images and the final readable images generated. Therefore, clinicians can perform lesion diagnosis based on the readable images output by the display unit 400.

[0044] <Abnormal lesion classification model YOLO v4>

[0045] Figure 3 This is a schematic diagram of the abnormal lesion classification model of the diagnostic assistance algorithm described in this embodiment. Figure 4 This is a schematic diagram of the labeling method for the abnormal lesion classification model described in this embodiment.

[0046] like Figure 3 and Figure 4 As shown, the abnormal lesion classification model 211 described in this embodiment performs binary classification on the captured image. That is, the abnormal lesion classification model 211 classifies the captured image into lesion images and normal images. Furthermore, the abnormal lesion classification model 211 applies a bounding box to the region where the lesion is located in the lesion image and outputs the lesion image.

[0047] The abnormal lesion classification model 211 can apply various object detection models. In one example, the abnormal lesion classification model 211 can apply a YOLO (You Only Look Once) based model. Therefore, based on the model, various labeling methods can be implemented to label the image data applied by the abnormal lesion classification model 211.

[0048] In one example, labels on the image data are used by experts to identify lesions present in the data. Furthermore, experts can add textual annotations related to the lesion.

[0049] In one example, labels on the image data are used by experts to identify lesions present in the images. Furthermore, experts create labels based on the text of the lesion, directly identifying its presence in the image data. In this case, experts can draw quadrilaterals around the lesion in the image data to mark its location.

[0050] In one example, the label of the image data is for an expert to confirm a lesion present in the image data. Also, the expert makes a label for the text regarding the lesion to directly identify the lesion present in the image data. At this time, the expert can draw a polygon to identify the position of the lesion according to the lesion present in the image data.

[0051] Also, hereinafter, in order to help understanding of the invention, an embodiment in which the abnormal lesion classification model 211 is configured as a YOLO model is described. Regarding the label of such a YOLO model, a method in which the expert makes a label for the text and draws a quadrangle around the lesion can be applied.

[0052] In one example, the YOLO model can be configured as a YOLO v4 model. The YOLO v4 model adopts YOLO v3, CSPDarknet53, SPP, PAN, BoF ALC BoS structures. The conventional YOLO model has a disadvantage that it is not easy to detect a small object. However, the YOLO v4 model can increase input resolution to show high performance in detecting a small lesion.

[0053] Thus, the abnormal lesion classification model 211 applying the YOLO model classifies the image including a predetermined lesion when inputting the photographed image, and performs binary classification regardless of the type of lesion, classifying the lesion image and the normal image. At this time, the predetermined lesion can be an image including a hemorrhagic, inflammatory, vascular, and polyp lesion.

[0054] Generally, the YOLO model classifies into 4 classes according to whether the photographed image includes a hemorrhagic, inflammatory, vascular, and polyp lesion when inputting the photographed image. However, when the number of classes to be classified increases, the performance of the deep learning model is inevitably reduced.

[0055] Thus, the abnormal lesion classification model 211 does not classify into 4 classes when inputting the photographed image. The abnormal lesion classification model 211 classifies the image including the lesion regardless of the type of lesion and performs binary classification into two classes, i.e., the lesion image and the normal image. That is, the abnormal lesion classification model 211 classifies the photographed image including at least any one of a hemorrhagic, inflammatory, vascular, and polyp lesion, and reclassifies the classified photographed image into a lesion image and a normal image. At this time, the threshold values of the hemorrhagic, inflammatory, vascular, and polyp lesions are respectively calculated, and then the difference criteria of the lesion image and the normal image are determined based on the 4 threshold values.

[0056] <Learning and evaluation of abnormal lesion classification model>

[0057] Figure 5is a schematic diagram of the learning method of the abnormal lesion classification model described in this embodiment.

[0058] As Figure 5 indicated, according to this embodiment, learning of the abnormal lesion classification model 211 performs model learning, which can be subjected to verification and performance testing.

[0059] In one example, in the learning of the abnormal lesion classification model 211, 5462422 small intestine images were extracted from the clinical cases of 10386 patients. Also, the small intestine images were classified into development data and test data.

[0060] The development data was used in total 162160 times. At this time, the normal images were 109902, and the lesion images were 52258. The lesion images were composed of 10000 hemorrhagic lesion images, 28818 inflammatory lesion images, 9103 vascular lesion images, and 4337 polyp lesion images. The development data was divided into a learning data group and a development verification data group. At this time, 80% of the development data was used for the learning data group, and 20% of the development data was used for the development verification data group.

[0061] In the learning data group, the normal images were 87922, and the lesion images were 41807. The lesion images were composed of 8000 hemorrhagic lesion images, 23055 inflammatory lesion images, 7282 vascular lesion images, and 3470 polyp lesion images. Thus, in the learning of the abnormal lesion classification model, a total of 129729 learning data groups were used.

[0062] Also, in the development verification data group, the normal images were 21980, and the lesion images were 10451. The lesion images were composed of 2000 hemorrhagic lesion images, 5763 inflammatory lesion images, 1821 vascular lesion images, and 867 polyp lesion images. Thus, a total of 32431 development verification data groups were used to verify the learning of the abnormal lesion classification model.

[0063] After the learning and verification, the abnormal lesion classification model 211 performs performance testing.

[0064] The test data was used in total 5300262 times. At this time, the normal images were 4829022, and the lesion images were 471240. The lesion images were composed of 471240 hemorrhagic lesion images, 150237 inflammatory lesion images, 17242 vascular lesion images, and 27204 polyp lesion images. The test data was composed of images in the same manner as in the actual clinic.

[0065] Then, the performance of the abnormal lesion classification model 211 that has been trained was evaluated, and the inference time for the test was measured, and as a result, it was confirmed that the abnormal lesion classification model 211 using the YOLO model showed high performance in the binary classification of the captured images.

[0066] <Video frame processing model>

[0067] Figure 6 is a schematic diagram of a similarity evaluation method for extracting representative frames from a video frame processing model of the diagnostic assistance algorithm described in this embodiment, Figure 7 is a schematic diagram of forced extraction of representative frames from a video frame processing model of the diagnostic assistance algorithm described in this embodiment, Figure 8 is a schematic diagram of an example of extracting representative frames from a video frame processing model of the diagnostic assistance algorithm described in this embodiment.

[0068] As shown in Figure 6 to Figure 8 , according to this embodiment, the video frame processing model 212 selects representative images, i.e., representative frames, from the lesion images of the abnormal lesion classification model 211. At this time, the selected representative frames can be output from the display unit 400.

[0069] The video frame processing model 212 compares the similarity of multiple frames of the lesion images. The video frame processing model 212 selects representative frames that represent similar frames. At this time, the video frame processing model 212 extracts color (Color) and texture (Texture) features of the frames and compares the similarity of the two frames.

[0070] At this time, the video frame processing model 212 analyzes the similarity of the two frames by the Bhattacharyya Distance.

[0071] In one example, the range of the value of the Bhattacharyya Distance can be 0 to 1. At this time, the closer the value of the Bhattacharyya Distance in the video frame processing model 212 is to 0 or 0, the higher the similarity, and thus the similarity between the frames is determined.

[0072] In one example, as shown in Figure 6 , it can be assumed that the multiple frames are composed of the 1st frame to the 10th frame. Thus, the video frame processing model 212 compares the standard frame and the subsequent frame in order with the 1st frame as the reference. At this time, the value of the Bhattacharyya Distance from the 1st frame to the 7th frame can be 0 or can be close to 0. Also, the Bhattacharyya Distance between the 1st frame and the 8th frame can be 1 or can be close to 1.

[0073] Thus, it is considered that the 1st frame to the 7th frame in the video frame processing model 212 are similar frames. Also, the video frame processing model 212 selects representative frames of the similar frames. Then, the video frame processing model 212 compares the standard frame (the 8th frame) and the subsequent frames in order, based on the judgment that the 8th frame is different from the 1st frame. Also, the video frame processing model 212 selects representative frames until the similar frames are cut off.

[0074] In addition, the video frame processing model 212 can filter the standard frames as the representative frames in the process of selecting the representative frames. That is, the frames that become the comparison objects of the initial similarity can be filtered as the representative frames. In an example, as shown in FIG. 8, the 1st frame and the 8th frame can be filtered as the representative frames, respectively. Figure 6

[0075] In addition, in the Bhattacharyya distance calculation, the video frame processing model 212 performs normalization for the two frames that become the comparison objects. Also, the video frame processing model 212 can calculate the Bhattacharyya distance through histogram analysis.

[0076] In an example, hereinafter, the video frame processing model 212 calculating the Bhattacharyya distance of the 1st frame and the 2nd frame will be described. At this time, the frames can be RGB images of 320 x 320 in size. Thus, the video frame processing model 212 adjusts the size of each image to 1 / 2. Also, the video frame processing model 212 performs an HSV conversion with respect to the images of which the size has been adjusted. That is, the video frame processing model converts the red (Red), green (Green), and blue (Blue) values into hue (Hue), saturation (Saturation), and value (Value) values.

[0077] Also, the video frame processing model 212 adjusts the number of bins of the color and saturation values in order to perform 2D histogram analysis. At this time, the video frame processing model 212 adjusts the number of bins of the color values from 180 to 60. Also, the video frame processing model 212 adjusts the saturation values from 256 to 32. Then, the video frame processing model 212 calculates the histogram (H-Histogram) of the color values and the histogram (S-Histogram) of the saturation values.

[0078] Also, the video frame processing model 212 can calculate the value of the Bhattacharyya distance based on the histogram of each color value and the histogram of the saturation value of the 1st frame and the 2nd frame.

[0079] ​In addition, the video frame processing model 212 prevents frames from excessively skipping when screening representative frames. When analyzing the similarity of frames by the value of the Bhattacharyya distance, numerous frames existing between representative frames can skip. However, when simply skipping frames by similarity analysis in the medical field, an unanticipated disadvantage can be found. Thus, the video frame processing model 212 can forcibly extract representative frames corresponding to a preset frame interval N.

[0080] In one example, the video frame processing model 212 can screen representative frames not exceeding 10 frames. Referring to FIG. 4, the first frame and the 18th frame can be known. The value of the Bhattacharyya distance of the first frame to the 12th frame can be 0 or can be close to 0. Also, from the 13th frame, the value of the Bhattacharyya distance can be 1 or can be close to 1. Also, the value of the Bhattacharyya distance of the 14th frame and the 13th frame can be 1 or can be close to 1. At this time, when using only the value of the Bhattacharyya distance of the video frame processing model 212, the first frame, the 13th frame, and the 14th frame are screened as representative frames. Figure 8

[0081] However, as described above, the video frame processing model 212 can set a screening interval of representative frames to not exceed 10 frames. Thus, even though it is determined that the similarity is high until the 12th frame, the video frame processing model 212 can screen the first frame, the 11th frame, the 13th frame, and the 14th frame as representative frames. That is, the video frame processing model 212 screens the first frame, which becomes an initial standard. Also, the video frame processing model 212 screens the 11th frame, which exceeds 10 frames from the first frame, as a representative frame, and screens the 13th frame, which has different similarity from the 11th frame, as a representative frame. Also, the video frame processing model 212 can screen the 14th frame, which has different similarity from the 13th frame, as a representative frame.

[0082] Then, the video frame processing model 212 outputs the selected representative frames from the display unit 400. Thus, a clinician can shorten the time required for diagnosing a lesion through representative frames.

[0083] <Performance evaluation results of medical image processing system>

[0084] ​To evaluate the performance of the processing system 1000 described in the present application, performance evaluation was performed equal to or higher than NVIDIA GeForce RTX 2080 and Window 10. The processing system 1000 can shorten the reading time of the clinician, which previously required 30 minutes to 2 hours or more, to within 10 minutes. Also, based on 129600 frames of clinical cases, the generation speed of the readable image is approximately 40 minutes. Also, the sensitivity (Sensitivity) of the processing system 1000 is 93.0%, the specificity (Specificity) is 89.0%, and the accuracy is 90.0%, and in terms of performance, it shows high performance. Also, the video frame number compression rate of the processing system 1000 shows more than 80%.

[0085] In addition, the abnormal lesion classification model described in the present embodiment is configured as YOLO v4. However, the abnormal lesion classification model in another embodiment can include YOLO v8 when configured.

[0086] <Abnormal lesion classification model YOLO v8>

[0087] Figure 9 FIG. 1 is a schematic diagram of an abnormal lesion classification model of a diagnostic assistance algorithm according to another embodiment.

[0088] As shown in FIG. 1, the YOLO model described in the present embodiment can be configured as a YOLO v8 model. Figure 9

[0089] In one example, the YOLOv8 model uses a modified (Modified) CSPDarknet53 backbone. In particular, the YOLO v8 model replaces the CSP layer used for YOLO v5 with a C2f module. As a result, the YOLO v8 model has the advantage that the Partial Pyramid Pooling Fast (SPPF) layer performs pooling on image features on a fixed-size map, thereby speeding up the operation. At this time, each convolution of the YOLO v8 model applies Batch Normalization (BN) and SiLU Activation, and the Head part is divided into Process Objectness, Classification, and Regression Task, and since it is configured, each service can be individually implemented.

[0090] ​As described above, according to the present application, the medical image processing system and the processing method have the following technical effects: shortening the time of reading medical images and providing suggestions, improving the reading efficiency of clinicians, and shortening the waiting time of patients.

[0091] As described above, as shown in the drawings, an embodiment of the present application should not be interpreted as limiting the technical idea of the present application. The scope of protection of the present application is only limited in the content recited in the claims, and the technical idea of the present application can be modified and changed in various forms by those skilled in the art. Therefore, such modifications and changes are obvious to those skilled in the art and should belong to the scope of protection of the present application.

Claims

1. A medical image processing system, characterized in that: include: The communication unit receives images captured by the capsule endoscope; and The processing unit uses the lesion model learned through a lesion classification method to generate a readable image based on the captured image. The processing unit applies the captured image to a pre-learned abnormal lesion classification model, classifying the image into images that include preset lesions. Then, the classified images are further classified into lesion images and normal images. Representative images representing the lesions are selected from the classified lesion images. The selected representative image is generated into the readable image.

2. The medical image processing system according to claim 1, characterized in that: The abnormal lesion classification model includes the YOLO series models.

3. The medical image processing system according to claim 1, characterized in that: The abnormal lesion classification model may include at least one of YOLO v4 and YOLO v8.

4. The medical image processing system according to claim 3, characterized in that: In the learning data labels of the abnormal lesion classification model, experts can identify the lesions present in the learning data and mark them using a preset labeling method on the text of the lesions and the lesions present in the learning data.

5. The medical image processing system according to claim 4, characterized in that: The lesion is marked by a quadrilateral or polygon surrounding it.

6. The medical image processing system according to claim 1, characterized in that: The images captured include at least one of the following: hemorrhagic, inflammatory, vascular, and polypoid lesions. The abnormal lesion classification model, during the reclassification process, is independent of the type of lesion and can binary classify the captured image into lesion images and normal images.

7. The medical image processing system according to claim 6, characterized in that: The abnormal lesion classification model applies bounding boxes to the lesions included in the lesion image.

8. The medical image processing system according to claim 6, characterized in that: The abnormal lesion classification model, for lesions including at least one of the hemorrhagic, inflammatory, vascular, and polypoid lesions, calculates thresholds statistically for each. When the captured images are provided, the lesion images and the normal images are classified into two categories based on the threshold values ​​of each lesion.

9. The medical image processing system according to claim 1, characterized in that: In the selection of representative images, a pre-learned video frame processing model is applied. The video frame processing model extracts color and texture features from several of the lesion images and compares the similarity between the images.

10. The medical image processing system according to claim 9, characterized in that: When analyzing the similarity, the similarity is based on the Bach distance analysis.

11. The medical image processing system according to claim 9, characterized in that: In the similarity analysis, HSV transformation is performed on several lesion images that are compared with each other. Adjust the number of orientation boxes for color and chromaticity values, and calculate the histograms for color and chromaticity values.

12. The medical image processing system according to claim 1, characterized in that: In the selection of representative images, the standard image is compared with subsequent images in turn. When a subsequent image with a different similarity than the standard image appears, the subsequent image with the different similarity is used as the standard image, and the standard image and subsequent images are compared sequentially. The standard image is selected as the representative image.

13. The medical image processing system according to claim 12, characterized in that: When filtering the representative images, the number of images between a pair of representative images can not exceed n.

14. The medical image processing system according to claim 1, characterized in that: When selecting the representative images, the standard image is compared with subsequent images in sequence. The subsequent images n images starting from the standard image are selected as standard images, and the standard images are selected as representative images.

15. A medical image processing method, characterized in that: include: Receive images captured by a capsule endoscope; The captured images are applied to a pre-learned abnormal lesion classification model to classify images including preset lesions. Then, the classified images are further classified into lesion images and normal images. Representative images representing the lesions are selected from the classified lesion images; and The selected representative image is generated into a readable image.

Images