Ovarian cancer ultrasonic image segmentation method based on feature fusion mechanism
Through a network model based on feature fusion mechanism, the ovarian ultrasound images are automatically processed, which solves the accuracy and efficiency of manual interpretation of ultrasound images in the prior art, and achieves efficient and accurate image segmentation and diagnosis, which improves medical level.
Patent Information
- Application Number
- CN202310865286.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-07-14
- Publication Date
- 2025-01-14
AI Technical Summary
The prior art relies on manual interpretation of ultrasound images in the diagnosis of ovarian diseases, resulting in the diagnostic accuracy being affected by doctors' experience, time-consuming and labor-intensive and error-prone, and lacks efficient and automated diagnostic methods.
Design an ultrasonic image segmentation method of ovarian cancer based on feature fusion mechanism. Through a network model composed of feature extraction module, feature fusion module, auxiliary decoder module and decoder module, multi-round training is performed using multi-scale features and labeled data, and automatically Perform image segmentation and diagnosis.
Automatic ovarian ultrasound image segmentation has been achieved, which improves the accuracy and efficiency of diagnosis, reduces the burden of reading films by medical workers, and improves the unity of medical standards.
Smart Images
Figure CN119313602A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of image processing, and in particular to a medical image segmentation method and a network model based on a feature fusion mechanism. Background Art
[0002] Medical ultrasound imaging technology is widely used in medical examinations and clinical diagnosis due to its advantages of low cost, good real-time performance and convenience. Imaging doctors evaluate the abnormal parts in ultrasound images based on their clinical experience and make the final diagnosis.
[0003] The diversity and multiplicity of ovarian diseases have become a special field of research in modern medicine. How to diagnose ovarian diseases efficiently is an important medical issue. At present, the examination of ovarian diseases is carried out with the help of ultrasound examination equipment, and doctors make judgments on ovarian diseases based on clinical experience. Therefore, whether the doctor's judgment of ultrasound images is accurate becomes a key factor affecting clinical decision-making. The accuracy of the diagnosis is also affected by the doctor's ability to interpret it. Interpreting medical images and making diagnostic or treatment recommendations is extremely dependent on specially trained medical experts. Currently, the interpretation of medical images is labor-intensive, time-consuming, expensive and error-prone.
[0004] Thanks to the rapid progress of computer performance today, deep learning has become a hot topic for many scholars with its powerful self-learning ability, and has even surpassed humans in some tasks such as image recognition, facial recognition, and natural language processing. If we can rely on the powerful performance of deep learning to design a computer-aided diagnosis system that can automatically diagnose based on ultrasound images, we can reduce the burden on doctors and reduce the subjective differences in prenatal examinations, so that the diagnosis results become more objective and accurate, and improve the overall level of prenatal examinations. Therefore, it is necessary to apply artificial intelligence methods to ovarian ultrasound image data to provide doctors with a more reliable diagnostic basis. Summary of the invention
[0005] The purpose of this application is to provide an ovarian cancer ultrasound image segmentation method, which can improve the accuracy of image segmentation by fully utilizing the rich information contained in manually annotated data and the multi-scale feature fusion method to train the segmentation model.
[0006] The present invention designs an ovarian cancer ultrasound image segmentation method based on a feature fusion mechanism, which consists of four parts: a feature extraction module, a feature fusion module, an auxiliary decoder module, and a decoder module. First, the acquired ovarian ultrasound image and the segmentation mask annotation image corresponding to the original ovarian ultrasound image and the lesion area are input into the feature extraction module to obtain the first feature map of the ovarian ultrasound image; the first feature map is input into the feature fusion module to obtain the fused second feature map; and the first feature map is input into the auxiliary decoder module to obtain the loss value of the first segmentation result; the second feature map and the loss value of the first segmentation result are input into the decoder module to obtain the second segmentation result; based on the second segmentation result and the sample annotation image, the feature extraction module and the decoder module are trained for multiple rounds to obtain a trained ultrasound image segmentation model.
[0007] The first feature map extracted by the feature extraction module includes multi-scale feature maps of different depths and scales, which will be input into the auxiliary decoding module and the feature fusion module. The feature fusion module fuses feature maps of different depths from deep to shallow, and fuses feature maps of different sizes from small to large. The obtained second feature map includes the features after the fusion of multiple multi-scale feature maps. The auxiliary decoder module facilitates the first feature map to segment the ovarian cancer ultrasound image, and calculates and obtains the first loss value based on the annotated image.
[0008] In the decoder module, the image segmentation result of the ovarian ultrasound data set is obtained through the second feature map, and the second loss value is calculated; the total loss value is obtained according to the first loss value, the second loss value, the first weight corresponding to the first loss value, and the second weight corresponding to the second loss value; based on the total loss value, the feature extraction module and the decoder module are trained for multiple rounds to obtain the trained ultrasound image segmentation model.
[0009] Compared with the prior art, the system disclosed in the present invention is of great significance: (1) Saving human resources: Compared with the traditional manual film reading diagnosis method, this method can help medical workers to automatically segment ovarian ultrasound images, extract the lesion area, and reduce the number of film readings for medical workers; (2) Efficient and accurate: Using deep learning technology, the network model can quickly and accurately segment the lesion area in the ultrasound image. (3) Improving the overall medical level: Due to differences in development levels, there are certain differences in the medical diagnosis level between different cities and different hospitals. By introducing this method, it will be able to indicate the overall medical level to a certain extent. BRIEF DESCRIPTION OF THE DRAWINGS
[0010] Figure 1 It is a system block diagram of an ovarian cancer ultrasound image segmentation method based on a feature fusion mechanism of the present invention.
[0011] Figure 2 It is an application scenario diagram of an embodiment of the present invention.
[0012] Figure 3 It is a schematic block diagram of a method for training an image segmentation model in one embodiment of the present invention. DETAILED DESCRIPTION
[0013] The specific implementation of the present invention is described below in conjunction with the accompanying drawings.
[0014] The method in the embodiment of the present invention can be applied to various scenarios of segmenting an image (ie, image segmentation), which is not limited in the embodiment of the present invention. For example, the method in the embodiment of the present invention can be applied to medical image segmentation.
[0015] Combine the following Figure 2 In the application scenario of the present invention, taking medical image segmentation as an example, the specific application of the embodiment of the present invention is introduced. It should be noted that the following Figure 2 The description is only an example and not a limitation. The method in the embodiment of the present invention is not limited to medical image segmentation, and can also be applied to other scenarios of image segmentation (ie, image segmentation). Figure 2 The application scenario includes an image acquisition device 110 and an image segmentation device 120, wherein the image acquisition device 110 may be a computed tomography (CT) scanner, magnetic resonance imaging (MRI) or other device or equipment for taking medical images, and the image segmentation device 120 may be a computer device, a server (e.g., a cloud server) or other device or equipment capable of performing image segmentation and other processing on an image. The embodiment of the present invention is not limited to medical images, and the medical images in the embodiment of the present invention may include CT images, MRI images or other images applied in the medical field.
[0016] Figure 3 FIG. 1 is a schematic block diagram of a method for training an image segmentation model in an embodiment of the present invention. The training ultrasound image segmentation method can be applied to Figure 2 In the implementation environment provided, the above Figure 2 The image segmentation device 120 shown in FIG. 1 is executed, but the embodiment of the present application is not limited thereto. The following describes an exemplary embodiment of the present application by taking the method executed by a computer device as an example.
[0017] Among them, 210 is to extract the features of the image and form a feature map. This step is completed by the feature extraction module. The feature map is not unique and contains multi-scale features with different depths and sizes. In 220, the feature fusion module fuses the feature maps of different depths from deep to shallow, and fuses the feature maps of different sizes from small to large. The features after the fusion of the multi-scale feature maps are obtained. In 230, the auxiliary decoder module uses the feature map in step 210 to segment the ovarian cancer ultrasound image, and calculates and obtains the first loss value based on the labeled image. Finally, in 240, the decoder module will obtain the image segmentation result of the ovarian ultrasound data set through the feature map in step 210, and calculate the second loss value; according to the first loss value, the second loss value, and the first weight corresponding to the first loss value and the second weight corresponding to the second loss value in step 210, the total loss value is obtained; based on the total loss value, the feature extraction module and the decoder module are trained for multiple rounds to obtain the trained ultrasound image segmentation model.
[0018] The above description is merely an explanation of the present invention, but not a limitation of the present invention. Any simple deformation or modification made without departing from the core of the device described in the present invention shall fall within the protection scope of the present invention.
Claims
1. A method for training an ultrasound image segmentation model, characterized in that: include: Obtaining a plurality of ovarian ultrasound image sets to be processed, each of the ovarian ultrasound image sets to be processed; Inputting the image set into a feature extraction module to obtain a first feature map of the image in the image set; inputting the first feature map into a feature fusion module to obtain a fused second feature map; and inputting the first feature map into an auxiliary decoder module to obtain a loss value of a first segmentation result; The second feature map and the loss value of the first segmentation result are input into the decoder module to obtain a second segmentation result; based on the second segmentation result and the sample annotated image, the feature extraction module and the decoder module are trained for multiple rounds to obtain a trained ultrasound image segmentation model.
2. The ultrasound image segmentation model training method according to claim 1, characterized in that: The ovarian ultrasound image set includes original ovarian ultrasound images and corresponding lesion area segmentation mask annotation images.
3. The ultrasound image segmentation model training method according to claim 1, wherein: The feature extraction module extracts features from the ultrasound image set to obtain a first feature map; the first feature map includes a plurality of multi-scale feature maps of different depths and scales.
4. The ultrasound image segmentation model training method according to claims 1 and 3, wherein: The feature fusion module includes: performing feature fusion on the first feature map, fusing feature maps of different depths from deep to shallow; fusing feature maps of different sizes from small to large; and the second feature map includes the features after the fusion of the multiple multi-scale feature maps.
5. The ultrasound image segmentation model training method according to claims 1 and 3, wherein: The auxiliary decoder module includes: inputting the first feature map into the auxiliary decoder module, obtaining the image segmentation result of the ovarian ultrasound data set, and calculating a first loss value.
6. The ultrasound image segmentation model training method according to claims 4-5, wherein: The decoder module includes: inputting the second feature map into the decoder module to obtain the image segmentation result of the ovarian ultrasound data set, and calculating the second loss value; obtaining the total loss value according to the first loss value, the second loss value, and the first weight corresponding to the first loss value and the second weight corresponding to the second loss value; and performing multiple rounds of training on the feature extraction module and the decoder module based on the total loss value to obtain the trained ultrasound image segmentation model.