Endoscopic surgery video processing method, computer device and readable storage medium

Abstract

This application discloses a method for processing laparoscopic surgery videos, a computer device, and a readable storage medium. The method includes: performing self-supervised pre-training on an image classification model using a first training set, which includes multiple first training images; supervised fine-tuning of the image classification model using a second training set, which includes multiple second training images; assigning corresponding category labels to each frame of the target video using the image classification model; dividing the target video into information segments and non-information segments based on the assigned category labels; and removing the non-information segments from the target video. This application can ensure the accuracy of the image classification model in recognizing non-information frames in laparoscopic surgery videos, even with a limited amount of labeled data.

Description

Technical Field

[0001] This application relates to the field of surgical video processing, and more particularly to laparoscopic surgical video processing methods, computer equipment, and readable storage media. Background Technology

[0002] In existing technologies, the editing and cleaning of laparoscopic surgical videos largely rely on manual operation, which is highly subjective and labor-intensive. Therefore, some studies have attempted to use traditional computer vision algorithms (such as edge detection and histogram analysis) to identify non-informational segments (such as black screens or blurred images) in laparoscopic surgical videos. However, their recognition capabilities are limited to specific categories, and their accuracy is low. Among these, the most similar approach to this application is a screening method based on thresholds or simple image statistical features. For example, it uses an average brightness below a certain threshold to determine the presence of black screens, or relies on the Sobel operator, Laplacian gradient, jitter detection, and inter-frame motion estimation methods to detect blurred images in laparoscopic surgical videos. These methods still have limited effectiveness in standardized surgical environments, are difficult to generalize, exhibit poor stability when dealing with various types of "invalid images," and lack end-to-end processing capabilities.

[0003] With the development of deep learning, image classification and segmentation techniques based on models such as Convolutional Neural Networks (CNNs) and Visual Transformers have made significant progress in the field of medical imaging. However, the task of identifying non-information fragments in laparoscopic surgical videos still faces two major challenges: (1) The lack of public datasets and calibration standards makes it difficult to directly supervise the training of image classification models; (2) Manual labeling of high-quality datasets is too costly and inefficient.

[0004] Therefore, even though sufficient laparoscopic surgery video data can be collected, the amount of manually labeled laparoscopic surgery video data is relatively small. Although theoretically any image classification model can be used to identify non-informative segments in a target video, directly supervised training of image classification models often fails to achieve ideal results when the amount of labeled dataset is limited. Summary of the Invention

[0005] The purpose of this application is to provide a method for processing laparoscopic surgical videos, a computer device, and a readable storage medium that can ensure the classification accuracy of an image classification model even when the amount of data in the labeled dataset is limited.

[0006] To achieve the above objectives, this application provides a method for processing laparoscopic surgical videos to remove non-information segments from surgical videos, including: The image classification model is self-supervised pre-trained using a first training set, which includes multiple first training images that are not labeled. The image classification model is a deep learning model. The image classification model is fine-tuned under supervision using a second training set, which includes multiple second training images labeled with corresponding category tags. The image classification model is used to assign the corresponding category label to each frame of the target video; Based on the category labels assigned to each frame of the target video, the target video is divided into information segments and non-information segments. Remove the non-information segments from the target video.

[0007] Optionally, the self-supervised pre-training of the image classification model using the first training set includes: A covered image is obtained by covering a portion of the image region of the first training image; The overlay image is input into the image classification model, and the image classification model is used to obtain the first predicted feature corresponding to the overlay image; The predicted image is obtained based on the first predicted feature; The first training image is compared with the predicted image to obtain the first loss function value; The image classification model is optimized using the value of the first loss function.

[0008] Optionally, comparing the first training image with the predicted image to obtain the first loss function value includes: Based on the image region covered by the overlay image, the corresponding image regions are extracted from the prediction image and the first training image respectively as the first image feature and the second image feature; The first image feature is compared with the corresponding second image feature using the mean squared error loss function to obtain the value of the first loss function.

[0009] Optionally, obtaining the predicted image based on the first predicted feature includes decoding the first predicted feature using a decoder to obtain the predicted image, wherein the decoder is a deep learning model; The first loss function value is also used to optimize the decoder.

[0010] Optionally, the supervised fine-tuning of the image classification model using the second training set includes: The second training image is input into the image classification model, and the second predicted feature corresponding to the second training image is obtained using the image classification model; The second predicted feature is compared with the category label corresponding to the second training image to obtain the second loss function value; The image classification model is optimized using the value of the second loss function.

[0011] Optionally, dividing the target video into information segments and non-information segments according to the category labels assigned to each frame of the target video includes: Based on the category labels corresponding to each frame of the target video, determine the start and end timestamps of the non-information segments; Based on the start timestamp and the end timestamp, the non-information segments are correspondingly divided in the target video.

[0012] Optionally, the image classification model is the ViT model.

[0013] Optionally, the category labels include informational and non-informational categories; The category labels in the second training set are the non-information categories.

[0014] Optionally, assigning category labels to each frame of the target video using the image classification model includes: Each frame of the target video is input into the image classification model, and the image classification model outputs the third predicted feature corresponding to each frame of the target video. The category label of the corresponding frame in the target video is determined based on each of the third prediction features.

[0015] To achieve the above objectives, this application provides a computer device, including a memory and a processor; The memory is connected to the processor. The memory is used to store computer programs, and the processor is used to call the computer programs so that the computer device can execute the above-described laparoscopic surgery video processing method.

[0016] To achieve the above objectives, this application provides a computer-readable storage medium storing a computer program adapted to be loaded by a processor and to execute the laparoscopic surgery video processing method of the claims above.

[0017] This application employs a large number of unlabeled laparoscopic surgery video frames for self-supervised pre-training of the image classification model. This enables the model to extract structural information and contextual features from the laparoscopic surgery video frames, thereby learning general features representing laparoscopic video scenes. Simultaneously, this application utilizes a second training set for supervised fine-tuning of the image classification model, improving its classification accuracy and generalization ability. Furthermore, supervised fine-tuning does not require a large dataset; therefore, the second training set can contain only a small number of training images, eliminating the need for extensive data labeling. Thus, this application can ensure the accuracy of the image classification model in recognizing non-informative frames in laparoscopic surgery videos, even with a limited amount of labeled data. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the laparoscopic surgery video processing method according to an embodiment of this application.

[0019] Figures 2 to 7 This is a partial flowchart of the laparoscopic surgery video processing method according to an embodiment of this application.

[0020] Figure 8 This is a table showing the comparison results between this application and the prior art.

[0021] Figure 9 This is a schematic diagram of the structure of a computer device according to an embodiment of this application. Detailed Implementation

[0022] To explain in detail the technical content, structural features, objectives and effects of this application, the following description is provided in conjunction with the embodiments and accompanying drawings.

[0023] Laparoscopic surgical videos, typically acquired directly using recording equipment, contain numerous informational and non-informational segments. Informational segments are those containing valid information, such as clearly showing surgical instruments operating on the patient and clearly visible anatomical structures. Non-informational segments, on the other hand, are those that do not contain valid information.

[0024] In some embodiments, non-information fragments include the following types: Black screen clip: This clip is entirely black, and no surgical content can be seen. This is usually caused by the recording equipment not being turned on.

[0025] Blurry footage: In this footage, only the instruments performing the dissection can be vaguely seen, but the anatomical structures cannot be clearly seen. The reason may be that the lens of the filming equipment is not focused.

[0026] Light spot interference segment: There is at least one light spot in the image of this segment, which makes it impossible to clearly show the image information of the anatomical site. The reason may be that the light source of the operation shines on the lens of the shooting equipment.

[0027] Shaky footage: The footage in this clip is shaky, making it difficult to clearly see the anatomical area. This may be due to the lens not being securely mounted during filming, causing it to slip during the surgery.

[0028] Obscured segment: In this segment, the image of the anatomical area is blocked by other instruments, making it impossible to see the complete anatomical structure.

[0029] External view clip: This clip shows part of the patient's body outside the body, but the complete anatomical structure cannot be captured, possibly because the camera was removed from the patient's body after the surgery.

[0030] No-operation segment: The instruments in this segment are not being used in any way and cannot effectively reflect the surgical process.

[0031] No instruments in the footage: This footage does not show any instruments, therefore it does not belong to the process of surgical treatment of the affected area using instruments.

[0032] It is evident that manually annotating segments of laparoscopic surgery videos requires constant monitoring of the integrity and clarity of anatomical images, as well as ensuring the relationship between surgical instruments and anatomical sites accurately reflects the surgical process. This necessitates highly specialized knowledge and significant time investment, resulting in high costs for manual annotation and a limited availability of datasets suitable for supervised training of image classification models. In response, this application specifically discloses the following embodiments: Example 1 For this, please refer to Figure 1 This application discloses a method for processing laparoscopic surgical videos to remove non-information segments from surgical videos, including: S1. Self-supervised pre-training is performed on the image classification model using the first training set. The first training set includes multiple first training images, which are unlabeled. The image classification model is a deep learning model. Self-supervised pre-training is a type of unsupervised learning, which trains the image classification model without training labels. Common self-supervised pre-training methods include autoencoders.

[0033] S2, supervised fine-tuning of the image classification model is performed using a second training set. The second training set includes multiple second training images, each labeled with a corresponding class. It can be understood that supervised fine-tuning only involves making minor adjustments to the parameters of the image classification model using a small amount of data; the second training set does not need to include a large number of second training images.

[0034] S3 uses an image classification model to assign corresponding category labels to each frame of the target video. The category labels assigned by the image classification model include informational and non-informational categories, while the category labels in each of the second training sets include informational and / or non-informational categories.

[0035] S4. Based on the category labels defined in each frame of the target video, the target video is divided into information segments and non-information segments.

[0036] S5 removes non-information segments from the target video.

[0037] It is understandable that steps S1 to S2 are for obtaining an image classification model that can assign category labels to each frame of the target video, and steps S1 to S2 are not executed before each execution of steps S3 to S5.

[0038] This application employs a large number of unlabeled laparoscopic surgery video frames for self-supervised pre-training of the image classification model. This enables the model to extract structural information and contextual features from the laparoscopic surgery video frames, thereby learning general features representing laparoscopic video scenes. Simultaneously, this application utilizes a second training set for supervised fine-tuning of the image classification model, improving its classification accuracy and generalization ability. Furthermore, supervised fine-tuning does not require a large dataset, so the second training set can contain only a small number of training images, eliminating the need for extensive data labeling. Therefore, this application can maintain the accuracy of the image classification model in recognizing non-informative frames in laparoscopic surgery videos even with a limited amount of labeled data. Moreover, the laparoscopic surgery video processing method of this application can be applied to other fields such as gastroscopy and thoracoscopy simply by changing the image content of the first and second training sets, demonstrating good transferability.

[0039] Please see Figure 2 Specifically, step S1 includes: S11, covering a portion of the image region of the first training image to obtain the covered image.

[0040] S12, input the overlay image into the image classification model, and use the image classification model to obtain the first predicted feature corresponding to the overlay image.

[0041] S13, Obtain the predicted image based on the first predicted feature. The predicted image is used to predict image information of the covered image region in the covered image.

[0042] S14, compare the first training image with the predicted image to obtain the first loss function value.

[0043] S15, optimize the image classification model using the first loss function value.

[0044] Since the first predicted feature includes feature information of the overlay image, information about the image region covered by the overlay image can be decoded from the first predicted feature. During the training of the image classification model, the first loss function value of the image classification model is obtained by comparing the difference between the first training image and the predicted image. That is, the final optimization goal of the image classification model is to make the predicted image reconstructed based on the first predicted feature as close as possible to the first training image. Therefore, in this process, the model needs to learn how to extract more accurate features from the undisturbed image regions of the overlay image.

[0045] For example, if the first training image is a tumor and the covered area is the left half of the tumor, then the image classification model needs to infer the specific appearance of the left half of the tumor based on the right half of the tumor. Based on the right half of the tumor in the covered image and the left half of the tumor inferred by the image classification model, a first predictive feature is formed. The better the predicted image obtained after decoding the first predictive feature is to be as close as possible to the first training image, the better. In this way, during the self-supervised training process, the image classification model can learn to infer the content of the covered image area based on the uncovered part of the covered image, so that the image classification model has a certain contextual feature extraction capability.

[0046] Please see Figure 3 Optionally, step S13 includes: S131, the first predicted feature is decoded using a decoder to obtain the predicted image. The decoder is a deep learning model.

[0047] Accordingly, in step S15, the first loss function value is also used to optimize the decoder.

[0048] In other words, this application first uses an image classification model to encode the overlay image to obtain a first predicted feature; then uses a decoder to decode the first predicted feature to obtain a predicted image; compares the predicted image with the corresponding first training image to obtain a first loss function value; and then uses the first loss function value to optimize the image classification model and decoder. The image classification model and decoder essentially constitute an autoencoder structure.

[0049] Please see Figure 4 Optionally, step S14 includes: S141, based on the image region covered by the overlay image, extract the corresponding image regions from the prediction image and the first training image respectively as the first image feature and the second image feature.

[0050] S142, using the mean square error loss function, the first image feature is compared with the corresponding second image feature to obtain the first loss function value.

[0051] For example, if the first training image is a tumor and the covered area is the left half of the tumor, then the image classification model needs to infer the specific appearance of the left half of the tumor based on the right half of the tumor, and compare the image of the left half of the tumor in the predicted image (first image feature) with the image of the left half of the tumor in the first training image (second image feature) to obtain the first loss function value.

[0052] It's important to explain that the most difficult part to predict in an image is the covered region within the overlay image. Therefore, when calculating the first loss function value, we can compare only the first and second image features. This allows the first loss function value to more directly reflect whether the image classification model can infer the specific content of the covered region based on the contextual information of the covered region. In other words, compared to directly using the predicted image and the first training image to obtain the first loss function value, the first loss function value obtained by comparing the first and second image features more directly reflects whether the image classification model has the ability to extract structural information and contextual features from the image. This makes the optimization using the first loss function value more targeted. Furthermore, since the comparison involves only a portion of the predicted image and the first training image, rather than the entirety, it helps reduce computational costs.

[0053] More specifically, step S1 also includes: S16. Confirm whether the first convergence condition has been met. If not, replace the first training image and repeat steps S11 to S15.

[0054] Optionally, in step S11, before covering a portion of the image region of the first training image, image enhancement is performed on the first training image to prevent the image classification model from overfitting.

[0055] Please see Figure 5 Specifically, step S2 includes: S21, input the second training image into the image classification model, and use the image classification model to obtain the second predicted feature corresponding to the second training image.

[0056] S22, compare the second predicted feature with the category label corresponding to the second training image to obtain the second loss function value.

[0057] S23, optimize the image classification model using the second loss function value.

[0058] More specifically, the loss function used to calculate the second loss function value is the cross-entropy loss function.

[0059] In a specific example, the category labels can be divided into only informational and non-informational categories. In this case, the cross-entropy loss function can be the binary cross-entropy loss function.

[0060] More specifically, step S2 also includes: S24. Confirm whether the second convergence condition has been met. If not, replace the second training image and repeat steps S21 to S23.

[0061] Optionally, in step S21, before inputting the second training image into the image classification model, image enhancement of the second training image is also performed to prevent the image classification model from overfitting.

[0062] Specifically, the category labels in the second training set are all non-informational. By using only non-informational images to fine-tune the image classification model, the model can be made more sensitive to non-informational images, thereby improving the classification accuracy and generalization ability of the image classification model.

[0063] Please see Figure 6 Specifically, step S3 includes: S31, input each frame of the target video into the image classification model, and the image classification model outputs the third predicted feature corresponding to each frame of the target video.

[0064] S32, determine the category label of the corresponding frame in the target video based on each third prediction feature.

[0065] Please see Figure 7 Specifically, step S4 includes: S41, determine the start and end timestamps of non-information segments based on the category labels corresponding to each frame of the target video.

[0066] S42, based on the start and end timestamps, divide the target video into non-information segments. It can be understood that after dividing into non-information segments, the remaining segments of the target video are the information segments.

[0067] For example, if the target video is 8 seconds long, and the category labels for frames from 0-3 seconds and 5-8 seconds are both non-informational, while the category labels for frames from 3-5 seconds are both informational, then one set of start and end timestamps would be 0 seconds and 3 seconds, and another set would be 5 seconds and 8 seconds. Correspondingly, the segments from 0-3 seconds and 5-8 seconds in the target video would be non-informational segments, while the segment from 3-5 seconds would be informational segments.

[0068] Alternatively, one could determine the start and end timestamps of the information segments, divide the target video into information segments, and then the remaining segments of the target video would be non-information segments.

[0069] In a specific example, if this application is used to output editing suggestions for non-information segments to editing software, then it is only necessary to output the aforementioned start timestamp and end timestamp to the editing software. Users can then crop the target video based on the start timestamp and end timestamp, thereby improving the user's editing efficiency.

[0070] Specifically, the image classification model is the ViT (Vision Transformer) model. ViT possesses superior global perception capabilities and, after training, exhibits better image classification performance.

[0071] Specifically, non-information categories include black screen, blurred image, light spot interference, shaking, occlusion, external image, no operation, and no instrument. Correspondingly, non-information segments include black screen segments, blurred image segments, light spot interference segments, shaking segments, occlusion segments, external image segments, no operation segments, and no instrument segments.

[0072] The following section compares this application with some existing technologies. For ease of comparison, the pre-training dataset used in this experiment is different, while the supervised fine-tuning dataset is the same (both are the same laparoscopic surgery video dataset containing a small amount of labeled data). Finally, the tests were performed on the same test set, and the results are as follows: Figure 8 The experimental tables are shown below. The meanings of the table header names are as follows: Model: The image classification model used in this application is ViT. ResNet, LLM1 Zero-shot, LLM1 Few-shot, LLM2 Zero-shot, LLM2 Few-shot, LLM3 Zero-shot, and LLM3 Few-shot are all existing models used for segmenting non-information segments. LLM1 to LLM3 are existing methods for removing non-information segments from laparoscopic surgery videos using large language models (LLM). Few-shot means that the large model is trained using a few-shot method, and Zero-shot means that the large model is trained using a zero-shot method.

[0073] Pre-trained Dataset: The pre-trained dataset, where Endo700K is an unlabeled dataset of laparoscopic surgery videos, meaning the pre-training for the Endo700K dataset is unsupervised. Accordingly, this application uses Endo700K to indicate that pre-training is performed using steps S11 to S17 above. The ImageNet dataset is a labeled general dataset, and the pre-training method for the ImageNet dataset here is supervised pre-training. N / A indicates no pre-training. It can be understood that since LLM1 to LLM3 are trained using few-shot or zero-shot methods, no pre-training is required.

[0074] Image Augmentation: Whether data augmentation is used during training. Yes means yes, No means no.

[0075] Accuracy, Precision, Recall, and F1 are all metrics used to measure model performance. F1 is the harmonic mean of precision and recall; higher values ​​for these four metrics indicate better model performance.

[0076] Please see Figure 8 As can be seen from the figure, the training method of this application is superior to the training methods of the prior art. Comparing the various experiments for this application in the figure, it can be seen that compared with other prior art (data in rows 3 to 11 of the figure), training using the training method of this application (data in row 1 of the figure) can basically achieve the best score.

[0077] Furthermore, based on the prior art, an ablation experiment was conducted on the laparoscopic surgery video processing method of this application to obtain the data in rows 1 to 4 of the figure, where rows 1 and 2 correspond to the laparoscopic surgery video processing method of this application, and rows 3 and 4 correspond to the laparoscopic surgery video processing method of the prior art. The comparison shows that without pre-training ViT, only data augmentation is performed during supervised fine-tuning of ViT, resulting in a score ranking fourth (4th row of data in the figure). However, if ViT is pre-trained using a general dataset and then fine-tuned under supervised conditions, the score ranks second (3rd row of data in the figure). If ViT is pre-trained under self-supervised conditions only on unlabeled laparoscopic surgery videos, without data augmentation during supervised fine-tuning, the score ranks third (2nd row of data in the figure), but only slightly behind second place (e.g., a difference of only 0.001 in accuracy). Finally, if ViT is pre-trained under self-supervised conditions only on unlabeled laparoscopic surgery videos and data augmentation is performed during supervised fine-tuning, the score ranks first (1st row of data in the figure), and significantly ahead of second place (e.g., a difference of 0.13 in accuracy). Therefore, the laparoscopic surgery video processing method of this application not only reduces the workload of dataset labeling but also ensures recognition accuracy.

[0078] Example 2 Figure 9 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Figure 9 As shown, the computer device may include a processor 201 and a memory 202. The memory 202 is connected to the processor 201 and is used to store computer programs. The processor 201 is used to invoke the computer programs to cause the computer device to execute the aforementioned laparoscopic surgery video processing method. Furthermore, the computer device may also include at least one communication bus 203. The communication bus 203 is used to enable communication between components. The memory 202 may be a high-speed RAM or a non-volatile memory, such as at least one disk storage device.

[0079] Example 3 This application also provides a computer-readable storage medium storing a computer program adapted to be loaded by a processor and executed by the above-described laparoscopic surgery video processing method.

[0080] Example 4 This application also provides a computer program product or computer program that includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the electronic device to perform the aforementioned laparoscopic surgery video processing method.

[0081] It should be understood that, in the embodiments of this application, the processor may be a central processing unit (CPU), but it may also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0082] Those skilled in the art will understand that implementing all or part of the processes in the above embodiments can be accomplished by hardware related to computer program instructions. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0083] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0084] The above-disclosed embodiments are merely preferred embodiments of this application and should not be construed as limiting the scope of this application. Therefore, any equivalent variations made in accordance with the scope of this application shall still fall within the scope of this application.

Claims

1. A method for processing videos of endoscopic surgery for removing non-information segments from the videos, characterized in that, include: The image classification model is self-supervised pre-trained using a first training set, which includes multiple first training images that are not labeled. The image classification model is a deep learning model. The image classification model is fine-tuned under supervision using a second training set, which includes multiple second training images labeled with corresponding category tags. The image classification model is used to assign the corresponding category label to each frame of the target video; Based on the category labels assigned to each frame of the target video, the target video is divided into information segments and non-information segments. Remove the non-information segments from the target video.

2. The laparoscopic surgery video processing method of claim 1, wherein, The step of performing self-supervised pre-training of the image classification model using the first training set includes: A covered image is obtained by covering a portion of the image region of the first training image; The overlay image is input into the image classification model, and the image classification model is used to obtain the first predicted feature corresponding to the overlay image; The predicted image is obtained based on the first predicted feature; The first training image is compared with the predicted image to obtain the first loss function value; The image classification model is optimized using the value of the first loss function.

3. The surgical video processing method of claim 2, wherein, The step of comparing the first training image with the predicted image to obtain the first loss function value includes: Based on the image region covered by the overlay image, the corresponding image regions are extracted from the prediction image and the first training image respectively as the first image feature and the second image feature; The first image feature is compared with the corresponding second image feature using the mean squared error loss function to obtain the value of the first loss function.

4. The laparoscopic surgery video processing method as described in claim 2, characterized in that, The step of obtaining the predicted image based on the first predicted feature includes decoding the first predicted feature using a decoder to obtain the predicted image, wherein the decoder is a deep learning model; The first loss function value is also used to optimize the decoder.

5. The laparoscopic surgical video processing method of claim 1, wherein, The supervised fine-tuning of the image classification model using the second training set includes: The second training image is input into the image classification model, and the second predicted feature corresponding to the second training image is obtained using the image classification model; The second predicted feature is compared with the category label corresponding to the second training image to obtain the second loss function value; The image classification model is optimized using the value of the second loss function.

6. The surgical video processing method of claim 1, wherein, The step of dividing the target video into information segments and non-information segments according to the category labels assigned to each frame of the target video includes: Based on the category labels corresponding to each frame of the target video, determine the start and end timestamps of the non-information segments; Based on the start timestamp and the end timestamp, the non-information segments are correspondingly divided in the target video.

7. The surgical video processing method of claim 1, wherein, The image classification model is the ViT model.

8. The laparoscopic surgery video processing method as described in claim 1, characterized in that, The category labels include informational and non-informational categories; The category labels in the second training set are the non-information categories.

9. The laparoscopic surgery video processing method according to any one of claims 1 to 8, characterized in that, The step of assigning category labels to each frame of the target video using the image classification model includes: Each frame of the target video is input into the image classification model, and the image classification model outputs the third predicted feature corresponding to each frame of the target video. The category label of the corresponding frame in the target video is determined based on each of the third prediction features.

10. A computer device, comprising: Including memory and processor; The memory is connected to the processor, the memory is used to store computer programs, and the processor is used to call the computer programs so that the computer device executes the laparoscopic surgery video processing method according to any one of claims 1 to 9.

11. A computer readable storage medium, characterized in that, The computer-readable storage medium stores a computer program adapted to be loaded by a processor and execute the laparoscopic surgery video processing method according to any one of claims 1 to 9.

Images