A method for automatic segmentation and recognition of ultrasound images based on deep learning

Abstract

The present invention relates to the field of image processing technology, and discloses a method for automatic segmentation and recognition of ultrasound images based on deep learning, including the following steps: S1, data preparation and preprocessing, S2, self-supervised learning pre-training, S3, transfer learning and incremental learning, S4, multi-task joint learning, S5, image quality enhancement and super-resolution reconstruction: S6, multi-modal information fusion, information fusion by combining modal images, S7, model training and optimization, increasing data diversity by using data enhancement technology. By fusing ultrasound images with modal image information, the segmentation accuracy and lesion recognition ability of ultrasound images are improved, and different modal images provide complementary features, showing significant advantages in the recognition of complex lesions, complex locations, complex morphology and early lesions. At the same time, by fusing images of different modalities through a deep learning model, image quality can be improved and diagnostic accuracy can be improved.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a method for automatic segmentation and recognition of ultrasonic images based on deep learning. Background Art

[0002] Ultrasound imaging is a commonly used medical imaging technology, which is widely used in clinical diagnosis, especially in maternal examination, heart disease diagnosis, tumor detection, etc. Compared with other medical imaging technologies, ultrasound imaging has the advantages of no radiation, strong real-time performance, and convenient equipment. However, the quality of ultrasound images is affected by many factors, such as noise, artifacts, low resolution, etc., which makes the automatic segmentation and lesion identification of ultrasound images face many challenges. Especially in tasks such as tumor, organ classification, and lesion localization, due to the large image noise and complex organ morphological changes, traditional image processing methods often find it difficult to achieve high accuracy.

[0003] With the rapid development of deep learning technology, methods based on convolutional neural networks (CNNs) have made significant progress in the field of image segmentation and recognition, especially showing great potential in medical image processing. Deep learning can effectively improve the accuracy of ultrasound image segmentation and recognition by automatically extracting features and learning complex image patterns. However, due to the diversity of ultrasound images, the scarcity of annotated data, and image quality issues, existing deep learning models still face certain limitations in practical applications. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention provides an automatic segmentation and recognition method for ultrasound images based on deep learning, which solves the problem that the prior art generally relies on a single modality image, such as using only ultrasound images for lesion segmentation. Since the quality of ultrasound images is easily affected by noise and artifacts, and the recognition ability of complex morphology and early lesions is limited, it leads to insufficient performance in complex disease identification and early lesion diagnosis, and it is difficult to achieve efficient and accurate recognition of complex, irregular morphology or deep lesions.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: a method for automatic segmentation and recognition of ultrasound images based on deep learning, comprising the following steps:

[0006] S1. Data preparation and preprocessing: First, collect and preprocess the ultrasound image dataset. Since ultrasound images are usually greatly affected by noise, image denoising technology is used to improve image quality, and the images are uniformly resized and standardized.

[0007] S2. Self-supervised learning pre-training: Through pre-training on ultrasound images, the model learns the basic features of the image. In the absence of a large amount of labeled data, the model automatically extracts useful features using self-supervised learning.

[0008] S3. Transfer learning and incremental learning: Using transfer learning, the model pre-trained on the image is transferred to the ultrasound image segmentation task. The training features extracted from the pre-trained model are used to gradually train the model through the incremental learning module, so that as new data arrives, the model can continue to be updated and optimized without retraining.

[0009] S4, multi-task joint learning: by jointly training ultrasound images and then sharing the underlying features, the model performs image segmentation and then performs the lesion target localization task in the same network framework;

[0010] S5. Image quality enhancement and super-resolution reconstruction: Generate high-resolution ultrasound images using the GAN network, enhance the images, and then reconstruct low-resolution ultrasound images into high-resolution images using super-resolution reconstruction technology.

[0011] S6. Multimodal information fusion: Information fusion is performed by combining modal images, aligning images of different modalities through image registration technology, and fusion processing using a deep learning model to generate high-quality and recognizable ultrasound images and identify complex lesions;

[0012] S7. Model training and optimization: By using data augmentation technology to increase data diversity, cross-validation is used to optimize the model's hyperparameters, and loss functions are used to evaluate segmentation parameters. Finally, the model is connected to the hospital system and treatment equipment to help doctors diagnose patients' conditions.

[0013] Preferably, in S1, the preprocessing includes operations such as denoising, image alignment, and grayscale normalization; the image denoising technology includes median filtering and convolutional neural network denoising; the ultrasound image data set includes images of different parts of the human body, images of different imaging modes, and normal and diseased images; the noise includes speckle noise, thermal noise, electrical noise, and reverberation noise; the resizing includes scaling, cropping, and padding; and the normalization processing includes pixel value normalization and standardization of the mean and standard deviation.

[0014] Preferably, in S2, the basic features include edges and textures, the ultrasound image pre-training includes geometric transformation or image block reconstruction tasks, the annotation data includes lesion site annotation, organ structure annotation, lesion type annotation and disease severity annotation, and the useful features include texture features, shape features, boundary features and echo features.

[0015] Preferably, in S3, the images include natural images and medical imaging data sets, the transfer learning includes model transfer, parameter transfer, and cross-domain transfer, the image segmentation includes semantic segmentation, boundary segmentation, and multimodal image segmentation, the training features include bottom-level features, middle-level features, and high-level semantic features, and the new data includes new ultrasound image data, updated annotation data, ultrasound image data under different conditions, and data integrating other modal information.

[0016] Preferably, in S4, the joint training includes segmentation, classification and positioning, and the lesion target includes tumor, organ classification and target positioning.

[0017] Preferably, in S5, the image enhancement includes removing noise and artifacts, and the reconstruction techniques include nearest neighbor interpolation, convex set projection, super-resolution convolutional neural network and deep recursive convolutional network.

[0018] Preferably, in S6, the modal images include CT or MRI, the deep learning model includes FusionNet, the modal images include CT images, MRI images, PET images and SPECT images, the image registration technology includes feature registration, grayscale registration and transformation model registration, the fusion processing includes pixel-level fusion, feature-level fusion and decision-level fusion, and the complex lesions include lesions with complex locations, lesions with complex morphology, lesions with multiple tissue components and early lesions.

[0019] Preferably, in S7, the loss function includes the Dice coefficient and IoU, the data enhancement technology includes geometric transformation, color transformation, cropping and padding, the data diversity includes image content diversity, image acquisition condition diversity and annotation information diversity, and the above-mentioned cross-validation includes leave-one-out cross-validation and stratified cross-validation.

[0020] The present invention provides an automatic segmentation and recognition method for ultrasonic images based on deep learning. It has the following beneficial effects:

[0021] 1. The present invention improves the segmentation accuracy and lesion recognition ability of ultrasound images by fusing ultrasound images with image information of modalities such as CT, MRI, PET and SPECT. Images of different modalities provide complementary features and show significant advantages in the recognition of complex lesions, complex locations, complex morphologies and early lesions. At the same time, images of different modalities are effectively fused through deep learning models, thereby improving image quality and diagnostic accuracy.

[0022] 2. The present invention automatically extracts image features without a large amount of labeled data by combining self-supervised learning with incremental learning. Self-supervised learning helps the model understand basic features through geometric transformation and image block reconstruction tasks, while incremental learning enables the model to be continuously updated as new data is added, thereby improving the adaptability and learning efficiency of the model.

[0023] 3. The present invention combines convolutional neural network denoising technology and generative adversarial network enhancement. The present invention effectively improves the diagnostic capability of low-quality ultrasound images. The denoising technology removes speckle noise and thermal noise, and the adversarial network enhancement improves the image resolution, enhances the recognizability of the lesion area, and improves the segmentation accuracy of the model on images with more noise. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is a flow chart of a method for automatic segmentation and recognition of ultrasound images based on deep learning in the present invention. DETAILED DESCRIPTION

[0025] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0026] Please refer to the attached Figure 1 The embodiment of the present invention provides a method for automatic segmentation and recognition of ultrasound images based on deep learning, comprising the following steps:

[0027] S1. Data preparation and preprocessing: First, collect and preprocess the ultrasound image dataset. Since ultrasound images are usually greatly affected by noise, image denoising technology is used to improve image quality, and the images are uniformly resized and standardized.

[0028] S2. Self-supervised learning pre-training: Through pre-training on ultrasound images, the model learns the basic features of the image. In the absence of a large amount of labeled data, the model automatically extracts useful features using self-supervised learning.

[0029] S3. Transfer learning and incremental learning: Using transfer learning, the model pre-trained on the image is transferred to the ultrasound image segmentation task. The training features extracted from the pre-trained model are used to gradually train the model through the incremental learning module, so that as new data arrives, the model can continue to be updated and optimized without retraining.

[0030] S4, multi-task joint learning: by jointly training ultrasound images and then sharing the underlying features, the model performs image segmentation and then performs the lesion target localization task in the same network framework;

[0031] S5. Image quality enhancement and super-resolution reconstruction: Generate high-resolution ultrasound images using the GAN network, enhance the images, and then reconstruct low-resolution ultrasound images into high-resolution images using super-resolution reconstruction technology.

[0032] S6. Multimodal information fusion: Information fusion is performed by combining modal images, aligning images of different modalities through image registration technology, and fusion processing using a deep learning model to generate high-quality and recognizable ultrasound images and identify complex lesions;

[0033] S7. Model training and optimization: By using data augmentation technology to increase data diversity, cross-validation is used to optimize the model's hyperparameters, and loss functions are used to evaluate segmentation parameters. Finally, the model is connected to the hospital system and treatment equipment to help doctors diagnose patients' conditions.

[0034] In S1, preprocessing includes denoising, image alignment, grayscale normalization and other operations; image denoising techniques include median filtering and convolutional neural network denoising; ultrasound image datasets include images of different parts of the human body, images of different imaging modes, and normal and diseased images; noise includes speckle noise, thermal noise, electrical noise and reverberation noise; resizing includes scaling, cropping and padding; and standardization includes pixel value normalization and standardization of mean and standard deviation.

[0035] Specifically, since ultrasound images usually have speckle noise, thermal noise, electrical noise, and reverberation noise, median filtering and convolutional neural network (CNN) denoising techniques are used to effectively remove these noises. Median filtering is suitable for removing salt and pepper noise, while convolutional neural network denoising automatically identifies and removes multiple noises from images through deep learning while maintaining image details. The ultrasound image dataset includes images of different parts of the human body, different imaging modes (such as B-ultrasound, color ultrasound, etc.), as well as images of normal and diseased areas, ensuring the diversity of the dataset and the generalization ability of the model;

[0036] The preprocessing process also includes image alignment and resizing. Image alignment ensures the spatial consistency of images from different devices or angles, eliminating the interference of device differences on model training. Resizing unifies images to a fixed size through scaling, cropping, and padding to meet the requirements of deep learning model input. Scaling is used to adjust the resolution, cropping is used to extract key information, and padding is used to avoid information loss.

[0037] Through standardization, the brightness differences between images are removed through pixel value normalization and mean standard deviation standardization to ensure the consistency of image features. Pixel value normalization adjusts the image pixel values ​​to a fixed range, and mean standard deviation standardization converts the image to a standard normal distribution to ensure high efficiency during model training. Through preprocessing operations, the image quality is significantly improved, providing stable and high-quality data input for subsequent deep learning model training.

[0038] In S2, basic features include edges and textures, ultrasound image pre-training includes geometric transformation or image block reconstruction tasks, annotated data includes lesion site annotation, organ structure annotation, lesion type annotation and disease severity annotation, and useful features include texture features, shape features, boundary features and echo features.

[0039] Specifically, ultrasound images are pre-trained through self-supervised learning to help the model learn the basic features of the image without a large amount of labeled data. In the geometric transformation task, the model learns the spatial structural features of the image by rotating, translating, scaling and other operations on the image; in the image block reconstruction task, the model cuts the image into multiple small blocks and learns the local and global features of the image by reconstructing these small blocks. These pre-training tasks enable the model to effectively extract basic features such as edges and textures of ultrasound images, laying the foundation for subsequent segmentation and recognition tasks. These labeled data provide supervision information for the model, helping the model understand the key areas and lesion features in ultrasound images during the learning process. The model learns how to identify important areas in ultrasound images and their corresponding lesion types, structures and severity through these labeled data.

[0040] Through pre-training, the model can extract a series of useful image features, including texture features, shape features, boundary features and echo features. Texture features help the model identify detailed structures in the image, shape features help identify the morphological information of organs or lesion areas in the image, boundary features can help the model accurately define the boundaries of the lesion area, and echo features provide sound wave reflection information in ultrasound images, further improving the accuracy of segmentation and recognition. Through these features, the model can process ultrasound images more effectively and provide more accurate feature representation for subsequent image segmentation and recognition tasks.

[0041] In S3, images include natural images and medical imaging datasets, transfer learning includes model transfer, parameter transfer, and cross-domain transfer, image segmentation includes semantic segmentation, boundary segmentation, and multimodal image segmentation, training features include low-level features, mid-level features, and high-level semantic features, and new data include new ultrasound image data, updated annotation data, ultrasound image data under different conditions, and data that integrates other modal information.

[0042] Specifically, the efficiency and accuracy of ultrasound image segmentation can be improved through pre-training models on natural image and medical image datasets. Transfer learning includes model transfer, parameter transfer and cross-domain transfer. Model transfer refers to applying models that have been trained on other tasks to ultrasound image segmentation, borrowing existing knowledge. Parameter transfer applies some parameters of existing models to new tasks to reduce training time. Cross-domain transfer allows the transfer of model knowledge from different fields to help the model adapt to ultrasound image segmentation tasks.

[0043] Image segmentation methods include semantic segmentation, boundary segmentation, and multimodal image segmentation. Semantic segmentation classifies each pixel to identify tissues and lesion areas; boundary segmentation helps identify the precise boundaries of lesion areas; multimodal image segmentation enhances the recognition of complex lesion areas by fusing CT, MRI, and ultrasound image information. During the training process, the model completes the segmentation task through low-level features (such as texture and edges), middle-level features (such as shape and structure), and high-level semantic features (such as lesion type and tissue structure) to ensure high-precision segmentation and recognition. New ultrasound image data, updated annotation data, images under different conditions, and information fused from other modalities will be continuously introduced to optimize the training of the model and improve its adaptability to different clinical environments and conditions.

[0044] In S4, joint training includes segmentation, classification and localization, and lesion targets include tumors, organ classification and target localization.

[0045] Specifically, by adopting a joint training strategy and combining image segmentation, classification and positioning tasks, the comprehensive ability of the model is improved. During the joint training process, the model simultaneously learns how to perform image segmentation, classification and positioning, making full use of shared underlying features and enhancing information complementarity between different tasks. The image segmentation task ensures that the model can identify key areas in the image by accurately delineating the boundaries of the lesion area; the classification task classifies the lesion area, such as distinguishing different types of tumors or normal tissues; the positioning task is responsible for accurately locating the lesion location in the image, such as the specific location of the tumor or organ;

[0046] The lesion targets include the classification and location of tumors, the classification of organs, and the location of other targets. Through joint training, the model can not only accurately segment the lesion area, but also complete the target classification and location tasks in the same framework, thereby improving the overall recognition effect. This multi-task learning strategy enables the model to use the labeled data more efficiently, improve the segmentation accuracy and diagnostic accuracy, and reduce the training time.

[0047] In S5, image enhancement includes noise and artifact removal, and reconstruction techniques include nearest neighbor interpolation, convex set projection, super-resolution convolutional neural network and deep recursive convolutional network.

[0048] Specifically, image enhancement technology is used to remove noise and artifacts, thereby improving the quality of ultrasound images and ensuring the accuracy of subsequent segmentation and recognition tasks. Noise and artifacts in images often affect the accurate recognition of lesion areas, so effective denoising and artifact removal methods are needed. By using convolutional neural network (CNN) denoising technology, the model can automatically identify and remove various types of noise, maintain image details and improve its quality;

[0049] In terms of image reconstruction, through nearest neighbor interpolation, convex set projection, super-resolution convolutional neural network (SRCNN) and deep recursive convolutional network (DRCN), nearest neighbor interpolation is a simple and effective image reconstruction method that can interpolate images according to the values ​​of neighboring pixels and restore the details in the image. Convex set projection technology uses mathematical optimization methods to restore the structural information in the image and eliminate the influence of noise. Super-resolution convolutional neural network uses deep learning models to perform super-resolution reconstruction on low-resolution images, thereby generating higher-resolution images and enhancing detail information. Deep recursive convolutional network gradually optimizes images through multiple recursions, restores high-frequency details in the image, and further improves the clarity and quality of the image. Through reconstruction technology, it can effectively improve the resolution of ultrasound images, remove noise and artifacts, and provide clearer and more accurate image input for subsequent lesion segmentation and recognition tasks.

[0050] In S6, modal images include CT or MRI, deep learning models include FusionNet, modal images include CT images, MRI images, PET images and SPECT images, image registration technology includes feature registration, grayscale registration and transformation model registration, fusion processing includes pixel-level fusion, feature-level fusion and decision-level fusion, and complex lesions include lesions with complex locations, lesions with complex morphologies, lesions with multiple tissue components and early lesions.

[0051] Specifically, by adopting multimodal image fusion technology and combining image information of different modalities such as CT, MRI, PET and SPECT, the segmentation and recognition accuracy of ultrasound images can be improved. Different modality images provide complementary information and can fully reveal the characteristics of lesions. Using the FusionNet deep learning model for image fusion can effectively integrate the important features of different modality images.

[0052] Image registration technology includes feature registration, grayscale registration and transformation model registration. Feature registration extracts key points or edge information to ensure that the corresponding areas of different modal images are consistent. Grayscale registration optimizes brightness differences. Transformation model registration aligns images in space through geometric transformation to eliminate imaging angle differences.

[0053] Fusion processing includes pixel-level fusion, feature-level fusion and decision-level fusion. Pixel-level fusion directly merges the pixel information of the image, feature-level fusion enhances the lesion representation capability by synthesizing feature maps, and decision-level fusion combines the segmentation results of each modality image to provide more accurate predictions. Through multimodal image fusion, the present invention can effectively identify complex lesions, such as those with complex locations, complex morphologies, multiple tissue components or early lesions, thereby significantly improving the accuracy and robustness of diagnosis.

[0054] In S7, the loss functions include Dice coefficient and IoU, the data enhancement techniques include geometric transformation, color transformation, cropping and padding, the data diversity includes image content diversity, image acquisition condition diversity and annotation information diversity, and the above cross-validation includes leave-one-out cross-validation and stratified cross-validation.

[0055] Specifically, the loss functions include the Dice coefficient and IoU (intersection over union), which are often used to measure the accuracy of image segmentation and can effectively evaluate the performance of the model in the segmentation task. The Dice coefficient calculates the overlap between the predicted result and the true label, while IoU evaluates the ratio of the intersection and union of the predicted area and the true area, which helps to balance the precision and recall rate;

[0056] In terms of data enhancement, the present invention uses geometric transformation, color transformation, and cropping and padding techniques to increase the diversity of training data. Geometric transformation includes rotation, scaling, and flipping. Color transformation adjusts the brightness and contrast of the image. Cropping and padding are used to ensure that the image has a uniform size and avoid information loss. These data enhancement methods can effectively improve the robustness and generalization ability of the model and avoid overfitting.

[0057] Data diversity includes image content diversity, image acquisition condition diversity, and annotation information diversity. Image content diversity ensures that the model performs well on different types of images (such as different lesion types); image acquisition condition diversity includes images from different devices, different angles, or different scanning modes, ensuring that the model can adapt to various actual environments; annotation information diversity further enhances the model's learning ability by providing diverse annotation data (such as annotations of lesion locations and types);

[0058] Finally, the cross-validation methods include leave-one-out cross-validation and stratified cross-validation. These two methods can ensure the generalization ability of the model on different data sets. Leave-one-out cross-validation ensures that each sample can be evaluated as a test set by setting aside one sample for testing at each validation. Stratified cross-validation avoids the problem of class imbalance by ensuring that the proportion of sample categories in each fold is the same as that of the entire data set, so that the performance of the model is fully evaluated and optimized.

[0059] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for automatic segmentation and recognition of ultrasound images based on deep learning, characterized in that: The following steps are involved: S1. Data preparation and preprocessing: First, collect and preprocess the ultrasound image dataset. Since ultrasound images are usually greatly affected by noise, image denoising technology is used to improve image quality, and the images are uniformly resized and standardized. S2. Self-supervised learning pre-training: Through pre-training on ultrasound images, the model learns the basic features of the image. In the absence of a large amount of labeled data, the model automatically extracts useful features using self-supervised learning. S3. Transfer learning and incremental learning: Using transfer learning, the model pre-trained on the image is transferred to the ultrasound image segmentation task. The training features extracted from the pre-trained model are used to gradually train the model through the incremental learning module, so that as new data arrives, the model can continue to be updated and optimized without retraining. S4, multi-task joint learning: by jointly training ultrasound images and then sharing the underlying features, the model performs image segmentation and then performs the lesion target localization task in the same network framework; S5. Image quality enhancement and super-resolution reconstruction: Generate high-resolution ultrasound images using the GAN network, enhance the images, and then reconstruct low-resolution ultrasound images into high-resolution images using super-resolution reconstruction technology. S6. Multimodal information fusion: Information fusion is performed by combining modal images, aligning images of different modalities through image registration technology, and fusion processing using a deep learning model to generate high-quality and recognizable ultrasound images and identify complex lesions; S7. Model training and optimization: By using data augmentation technology to increase data diversity, cross-validation is used to optimize the model's hyperparameters, and loss functions are used to evaluate segmentation parameters. Finally, the model is connected to the hospital system and treatment equipment to help doctors diagnose patients' conditions.

2. The method for automatic segmentation and recognition of ultrasonic images based on deep learning according to claim 1, characterized in that: In S1, the preprocessing includes operations such as denoising, image alignment, and grayscale normalization. The image denoising technology includes median filtering and convolutional neural network denoising. The ultrasound image data set includes images of different parts of the human body, images of different imaging modes, and normal and diseased images. The noise includes speckle noise, thermal noise, electrical noise, and reverberation noise. The resizing includes scaling, cropping, and padding. The normalization processing includes pixel value normalization and standardization of the mean and standard deviation.

3. The method for automatic segmentation and recognition of ultrasonic images based on deep learning according to claim 1, characterized in that: In S2, the basic features include edges and textures, the ultrasound image pre-training includes geometric transformation or image block reconstruction tasks, the annotation data includes lesion site annotation, organ structure annotation, lesion type annotation and disease severity annotation, and the useful features include texture features, shape features, boundary features and echo features.

4. The method for automatic segmentation and recognition of ultrasonic images based on deep learning according to claim 1, characterized in that: In S3, the images include natural images and medical imaging data sets, the transfer learning includes model transfer, parameter transfer, and cross-domain transfer, the image segmentation includes semantic segmentation, boundary segmentation, and multimodal image segmentation, the training features include bottom-level features, middle-level features, and high-level semantic features, and the new data includes new ultrasound image data, updated annotation data, ultrasound image data under different conditions, and data integrating other modal information.

5. The method for automatic segmentation and recognition of ultrasonic images based on deep learning according to claim 1, characterized in that: In S4, the joint training includes segmentation, classification and positioning, and the lesion targets include tumors, organ classification and target positioning.

6. The method for automatic segmentation and recognition of ultrasonic images based on deep learning according to claim 1, characterized in that: In S5, the image enhancement includes removing noise and artifacts, and the reconstruction technology includes nearest neighbor interpolation, convex set projection, super-resolution convolutional neural network and deep recursive convolutional network.

7. The method for automatic segmentation and recognition of ultrasonic images based on deep learning according to claim 1, characterized in that: In S6, the modality image includes CT or MRI, the deep learning model includes FusionNet, the modality images include CT images, MRI images, PET images and SPECT images, the image registration technology includes feature registration, grayscale registration and transformation model registration, the fusion processing includes pixel-level fusion, feature-level fusion and decision-level fusion, and the complex lesions include lesions with complex locations, lesions with complex morphology, lesions with multiple tissue components and early lesions.

8. The method for automatic segmentation and recognition of ultrasonic images based on deep learning according to claim 1, characterized in that: In S7, the loss function includes the Dice coefficient and IoU, the data enhancement technology includes geometric transformation, color transformation, cropping and padding, the data diversity includes image content diversity, image acquisition condition diversity and annotation information diversity, and the above cross-validation includes leave-one-out cross-validation and stratified cross-validation.

Images