Ovarian cancer metastasis assessment local feature and global feature extraction method and device
By combining fully supervised learning and a multi-path CNN encoder with a sliding window transformer and a convolutional neural network, we have achieved accurate assessment of ovarian cancer metastases, solving the problem of difficult early diagnosis and improving treatment outcomes.
Patent Information
- Application Number
- CN202511372771.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-24
- Publication Date
- 2025-12-23
AI Technical Summary
Current technology makes it difficult to accurately assess ovarian cancer metastases, leading to difficulties in early diagnosis, poor treatment outcomes, and the discovery of metastases often occurring later than the primary lesion.
The model is trained using a fully supervised learning approach, and multi-phase information is fused using a multi-path CNN encoder and a sliding window transformer. Combined with convolutional neural networks to extract local features, the model achieves accurate assessment of ovarian cancer metastases.
It improves the accuracy of ovarian cancer metastasis assessment, enabling early detection of lesions, effective prevention of metastasis, and timely treatment.
Smart Images
Figure CN121190780A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of medical image processing, and in particular to a local feature and global feature extraction method for ovarian cancer metastasis evaluation and a local feature and global feature extraction device for ovarian cancer metastasis evaluation. BACKGROUND
[0002] Ovarian cancer is a malignant tumor of the ovary, which is a malignant tumor growing on the ovary. 90-95% of it is primary cancer of the ovary, and the other 5-10% is cancer metastasis to the ovary from other parts. Since there are no symptoms in the early stage of ovarian cancer, even if there are symptoms, they are not specific, and the screening is also limited, so early diagnosis is difficult. 60-70% of patients are in the advanced stage when they seek medical treatment, and the treatment effect of advanced cases is not good. Therefore, although the incidence of ovarian cancer is lower than that of cervical cancer and endometrial cancer, it ranks third among gynecological malignant tumors, but the mortality rate is higher than the sum of cervical cancer and endometrial cancer, ranking first among gynecological cancers, which is the biggest threat to women's health.
[0003] Ovarian cancer metastasis refers to the spread of cancer cells from the original lesion to other sites such as the peritoneum, abdominal wall, fallopian tube, pelvis, uterus, bone, liver, lung, etc. during the development of ovarian cancer. It can also occur at the beginning of the lesion, but more often it occurs after the lesion has developed to a certain extent. Since ovarian cancer usually does not have obvious symptoms, the discovery of ovarian cancer metastasis is often later than that of the primary lesion. Ovarian cancer metastasis can lead to tumor enlargement, blood dissemination, soft tissue dissemination, etc., which can affect the patient's health and quality of life, therefore, the evaluation of ovarian cancer metastasis can effectively prevent ovarian cancer metastasis, help medical personnel discover the lesion early, so that effective treatment measures can be taken in time. SUMMARY
[0004] To overcome the defects of the prior art, the technical problem to be solved by the present application is to provide a local feature and global feature extraction method for ovarian cancer metastasis evaluation, which can improve the accuracy of ovarian cancer metastasis evaluation, effectively prevent ovarian cancer metastasis, help medical personnel discover the lesion early, and thus take effective treatment measures in time.
[0005] The technical solution of the present application is: the local feature and global feature extraction method for ovarian cancer metastasis evaluation comprises the following steps:
[0006] (1) Using a fully supervised learning method, using doctor-labeled data to train and identify 7 key anatomical regions in the upper abdominal CT image;
[0007] (2) Multi-phase information fusion based on multi-channel CNN encoder;
[0008] (3) the global feature of the multi-phase information is learned by using a sliding window transformer, long-distance dependence modeling of the sliding window transformer and processing of global information of the image are utilized, and integration of multi-phase image information is realized;
[0009] (4) the local feature of the multi-phase information is learned by using a convolutional neural network (CNN), local feature extraction of the CNN is utilized to capture detailed information in each phase, and local texture, shape and edge information are extracted from the multi-phase medical image through convolution operation and nonlinear transformation.
[0010] The application can improve the accuracy of ovarian cancer metastasis evaluation, effectively prevent ovarian cancer metastasis, help medical personnel discover lesions as early as possible, and thus timely effective treatment measures can be taken.
[0011] The ovarian cancer metastasis evaluation local feature and global feature extraction device is also provided, which comprises:
[0012] The training and recognition module is configured to train and recognize 7 key anatomical regions in the upper abdominal CT image by using the data labeled by doctors through a full-supervised learning method;
[0013] The fusion module is configured to fuse the multi-phase information based on a multi-channel CNN encoder; the global feature extraction module is configured to learn the global feature of the multi-phase information by using a sliding window transformer, long-distance dependence modeling of the sliding window transformer and processing of global information of the image are utilized, and integration of multi-phase image information is realized;
[0014] The local feature extraction module is configured to learn the local feature of the multi-phase information by using a convolutional neural network (CNN), local feature extraction of the CNN is utilized to capture detailed information in each phase, and local texture, shape and edge information are extracted from the multi-phase medical image through convolution operation and nonlinear transformation. BRIEF DESCRIPTION OF DRAWINGS
[0015] Figure 1 It is a flowchart of the ovarian cancer metastasis evaluation local feature and global feature extraction method according to the application.
[0016] Figure 2 The overall framework of the double-branch fusion network of the ovarian cancer metastasis evaluation local feature and global feature extraction method according to the application is shown.
[0017] Figure 3 The cross-phase local attention fusion mechanism of the ovarian cancer metastasis evaluation local feature and global feature extraction method according to the application is shown. DETAILED DESCRIPTION
[0018] The application designs a convolution-Transformer dual-branch fusion network to perform multi-phase feature fusion and global-local multi-scale feature fusion of the ovarian cancer metastasis in the plain scan phase, the arterial phase, the portal phase and the delay phase, so as to complete comprehensive and accurate metastasis segmentation.
[0019] As Figure 1 , Figure 2 , Figure 3 illustrates, the ovarian cancer metastasis evaluation local feature and global feature extraction method comprises the following steps:
[0020] (1) By means of full supervision learning, the data labeled by doctors is used to train and identify 7 key anatomical regions in the upper abdominal CT image;
[0021] Such a training method can make the model have higher sensitivity and recognition accuracy in these specific parts.
[0022] (2) Multi-phase information fusion based on multi-channel CNN encoder;
[0023] There are differences between multi-phase images, and it is difficult to obtain good performance by directly splicing multi-phase inputs for the convolutional network, so multi-phase information fusion is very necessary. Therefore, the application designs a multi-channel convolutional encoder for multi-phase image feature extraction. As Figure 3 illustrated, the input of each phase will pass through a complete convolutional channel for single-phase feature extraction. At the same time, the cross-phase local attention fusion mechanism as Figure 3 illustrated is used for local feature fusion between phases after each convolutional block. This method enhances and supplements the features at this position by calculating the similarity between the spatial features of the to-be-segmented phase and the n 2 neighbors of other phase features, so as to integrate local information.
[0024] (3) Apply sliding window transformer Swin Transformer to learn global features of multi-phase information, use the long-distance dependence modeling and image global information processing of Swin Transformer to realize the integration of multi-phase image information;
[0025] This strategy aims to make full use of the long-distance dependence modeling ability and the image global information processing ability of Swin Transformer, and realize the effective integration of multi-phase image information. At the same time, this method can effectively integrate the features of scattered metastases.
[0026] (4) Apply a convolutional neural network (CNN) to learn local features from multi-phase information. CNN's local feature extraction is used to capture detailed information within each phase. Through convolution operations and nonlinear transformations, local texture, shape, and edge information are extracted from multi-phase medical images.
[0027] This strategy takes advantage of CNN's powerful local feature extraction capabilities to capture detailed information within each phase. Through deep convolution operations and nonlinear transformations, CNN can extract rich local texture, shape, and edge information from multi-phase medical images, which is crucial for modeling complex structures in medical images.
[0028] The application can improve the accuracy of ovarian cancer metastasis evaluation, effectively prevent ovarian cancer metastasis, and help medical personnel detect lesions early, so that effective treatment measures can be taken in time.
[0029] Preferably, the step (1) comprises the following sub-steps:
[0030] (1.1) Extract features in the image through a series of convolutional layers;
[0031] (1.2) Generate candidate regions through a region proposal network (RPN), and identify whether the candidate regions contain metastases through a region of interest (ROI) pooling layer and a fully connected layer.
[0032] Preferably, in step (2), the input of each phase is processed through a complete convolutional path for feature extraction of a single phase, and a cross-phase local attention fusion mechanism is used after each convolutional block to integrate local features between phases.
[0033] Preferably, in step (2), the cross-phase local attention fusion mechanism enhances and supplements the features at this position by calculating the similarity between the spatial features of the phase to be segmented and the n2 neighbors of other phase features, thereby integrating local information.
[0034] Preferably, in step (3), the Swin Transformer integrates information about lesions that are far apart and have similar morphological structures.
[0035] Preferably, in step (4), for metastases with surrounding tissue infiltration, blurred boundaries, and small volumes, image features are learned at multiple levels, from basic edge and texture information to higher-level shape and pattern features, thereby modeling and identifying small and unclear boundary metastases. In the case of complex background and similar tissue structure interference, the accuracy of metastasis segmentation can be significantly improved.
[0036] Those skilled in the art can understand that all or part of the steps in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a program, and the program can be stored in a computer-readable storage medium. When the program is executed, it includes the steps of the above-mentioned embodiment method, and the storage medium can be ROM / RAM, a magnetic disc, an optical disc, a memory card, etc. Therefore, corresponding to the method of the present application, the present application also simultaneously includes an ovarian cancer metastasis evaluation local feature and global feature extraction device, which is usually represented in the form of a functional module corresponding to each step of the method. The device includes:
[0037] The training recognition module is configured to train and recognize 7 key anatomical regions in the upper abdominal CT image by using the data labeled by doctors through a fully supervised learning method;
[0038] The fusion module is configured to perform multi-phase information fusion based on a multi-channel CNN encoder; the global feature extraction module is configured to apply Swin Transformer to learn global features of multi-phase information, utilize long-distance dependency modeling of Swin Transformer and processing of global information of an image, and realize integration of multi-phase image information;
[0039] The local feature extraction module is configured to apply convolutional neural network CNN to learn local features of multi-phase information, utilize local feature extraction of CNN to capture detailed information within each phase, and extract local texture, shape and edge information from multi-phase medical images through convolution operation and nonlinear transformation.
[0040] Preferably, the training recognition module performs:
[0041] (1.1) extracting features in the image through a series of convolutional layers;
[0042] (1.2) generating candidate regions through a region proposal network RPN, and identifying whether the candidate regions contain metastases through a region of interest ROI pooling layer and a fully connected layer.
[0043] Preferably, in the fusion module, the input of each phase is subjected to a complete convolutional channel for feature extraction of a single phase, and a cross-phase local attention fusion mechanism is used after each convolutional block to perform local feature fusion between phases. The cross-phase local attention fusion mechanism enhances and supplements the features at the position by calculating the similarity between the spatial features of the phase to be segmented and the n2 neighbors of the features of other phases, so as to integrate local information.
[0044] Preferably, in the global feature extraction module, the Swin Transformer integrates the information of lesions with long distances and similar morphological structures; in the local feature extraction module, for metastatic lesions with surrounding tissue infiltration, blurred boundaries and small volumes, image features are learned hierarchically from basic edge and texture information to higher level shape and pattern features, thereby modeling and identifying small and unclear boundary metastatic lesions.
[0045] The above is only a preferred embodiment of the present application, and does not limit the present application in any form. Any simple modification, equivalent change and modification of the above embodiment according to the technical essence of the present application are still within the protection scope of the technical solution of the present application.
Claims
1. A method for evaluating local and global features of ovarian cancer metastases, characterized by: It includes the following steps: (1) Through a fully supervised learning approach, using data annotated by doctors, train and identify 7 key anatomical regions in upper abdominal CT images; (2) Multi-phase information fusion based on multi-path convolutional neural network (CNN) encoder; (3) Apply the sliding window transformer to learn the global features of multi-phase information, and use the long-distance dependency modeling of the sliding window transformer and the processing of global image information to realize the integration of multi-phase image information. (4) Apply convolutional neural network (CNN) to learn local features of multi-phase information, use the local feature extraction of CNN to capture detailed information within each phase, and extract local texture, shape and edge information from multi-phase medical images through convolution operation and nonlinear transformation.
2. The method for evaluating local and global features of ovarian cancer metastases according to claim 1, characterized in that: Step (1) includes the following sub-steps: (1.1) Features in the image are extracted through a series of convolutional layers; (1.2) Candidate regions are generated through the Region Proposal Network (RPN), and whether these candidate regions contain metastases are identified through the pooling layer and fully connected layer of the Region of Interest (ROI).
3. The method for evaluating local and global features of ovarian cancer metastases according to claim 2, characterized in that: In step (2), the input of each phase passes through a complete convolutional path for feature extraction of a single phase, and a cross-phase local attention fusion mechanism is used after each convolutional block to perform local feature fusion between phases.
4. The method for evaluating local and global features of ovarian cancer metastases according to claim 3, characterized in that: In step (2), the cross-phase local attention fusion mechanism calculates the spatial features of the phase to be segmented and the features of other phases by multiplying the n values. 2 The similarity between neighbors is used to enhance and supplement the features of the location, thereby integrating local information.
5. The method for evaluating local and global features of ovarian cancer metastases according to claim 4, characterized in that: In step (3), the sliding window transformer integrates information on lesions that are far apart and have similar morphological structures.
6. The method for evaluating local and global features of ovarian cancer metastases according to claim 5, characterized in that: In step (4), for metastatic lesions that infiltrate surrounding tissues, have blurred boundaries, and are small in size, image features are learned hierarchically, from basic edge and texture information to higher-level shape and pattern features, thereby modeling and identifying small and indistinct metastatic lesions.
7. A device for evaluating local and global features of ovarian cancer metastases, characterized in that: It includes: The training and recognition module is configured to train and recognize seven key anatomical regions in upper abdominal CT images using doctor-annotated data through a fully supervised learning approach. The fusion module is configured to fuse multi-phase information based on a multi-path CNN encoder; the global feature extraction module is configured to learn global features of multi-phase information by applying a sliding window transformer, and to integrate multi-phase image information by utilizing long-distance dependency modeling of the sliding window transformer and processing of global image information. The local feature extraction module is configured to apply a convolutional neural network (CNN) to learn local features from multi-phase information. It uses the local feature extraction of CNN to capture detailed information within each phase and extracts local texture, shape, and edge information from multi-phase medical images through convolution operations and nonlinear transformations.
8. The device for extracting local and global features for evaluating ovarian cancer metastases according to claim 7, characterized in that: The training and recognition module performs the following: (1.1) Features in the image are extracted through a series of convolutional layers; (1.2) Candidate regions are generated through the Region Proposal Network (RPN), and whether these candidate regions contain metastases are identified through the pooling layer and fully connected layer of the Region of Interest (ROI).
9. The device for evaluating local and global features of ovarian cancer metastases according to claim 8, characterized in that: In the fusion module, the input of each phase passes through a complete convolutional path for feature extraction of a single phase. After each convolutional block, a cross-phase local attention fusion mechanism is used to fuse local features between phases. The cross-phase local attention fusion mechanism calculates the n-value of the spatial features of the phase to be segmented with the features of other phases. 2 The similarity between neighbors is used to enhance and supplement the features of the location, thereby integrating local information.
10. The device for evaluating local and global features of ovarian cancer metastases according to claim 9, characterized in that: In the global feature extraction module, the sliding window transformer integrates information on lesions that are far apart and have similar morphological structures; in the local feature extraction module, for metastatic lesions that infiltrate surrounding tissues, have blurred boundaries, and are small in size, image features are learned hierarchically, from basic edge and texture information to higher-level shape and pattern features, thereby modeling and identifying small and unclear metastatic lesions.