Semi-supervised segmentation model based on cross pseudo-supervised network and microaneurysm segmentation method

Abstract

The invention relates to the technical field of image processing, and particularly discloses a semi-supervised segmentation model based on a cross pseudo-supervised network and a micro-aneurysm segmentation method, and the model comprises the cross pseudo-supervised network: employing two segmentation networks based on a PFFNet architecture, enabling input data to pass through the two segmentation networks, generating output features, forming two pseudo labels, and obtaining a semi-supervised segmentation model of the cross pseudo-supervised network; the two pseudo labels supervise each other; loss function: adopting a Combo loss function in the training process of the labeled image, and introducing a cross entropy loss function in the cross supervision training of the pseudo label; and the CBAM attention modules are introduced into the uppermost layer and the lowermost layer of feature branches of each stage of the PFFNet architecture so as to solve the problems that in traditional microaneurysm image segmentation, the number of marked MA images in the actual environment is limited, and a large number of unmarked images are not effectively utilized.

Description

Technical Field

[0001] The present application relates to the field of image processing technology, and specifically discloses a semi-supervised segmentation model and a microaneurysm segmentation method based on a cross pseudo-supervised network. Background Art

[0002] Microaneurysms (MAs) are one of the early symptoms of diabetic retinopathy (DR). Accurately and efficiently detecting the location and size of MAs is of great significance for the diagnosis and treatment of DR. Fundus Fluorescein Angiography (FFA) is one of the important diagnostic examinations for fundus diseases and provides a valuable data resource for the diagnosis of DR. Microaneurysms appear as small, circular bright spots in FFA images. Manually labeling microaneurysms is time-consuming, labor-intensive, and has a high error rate. Therefore, using image processing technology to improve the efficiency and accuracy of fundus microaneurysm segmentation has become a key step in auxiliary medical treatment.

[0003] Currently, microaneurysm (MA) image segmentation relies primarily on the deep learning-based model PFFNet. PFFNet employs a Global Pyramid Scene Parsing (GPSP) module and a Spatial Scale-Aware Pyramid (SSAP) module. These modules are embedded in a U-shaped architecture to capture and integrate rich global and multi-scale contextual information, addressing the segmentation challenges posed by the small size of MA and the subtle contrast with the background.

[0004] The above-mentioned segmentation method mainly relies on MA images with pixel-level annotations. However, in practical applications, obtaining annotated MA lesion images usually requires manual annotation. This process is not only inefficient, but also difficult to effectively guarantee the accuracy of the annotations. In the segmentation method based on fully supervised learning, in order to build a high-performance MA segmentation model, a large number of annotated images are required as training data sets. However, in actual environments, annotated images are often scarce. Therefore, how to use a large number of unlabeled images to improve the effect of MA segmentation has become an important research topic that needs to be solved urgently. In view of this, the present invention provides a semi-supervised segmentation model and a microaneurysm segmentation method based on a cross-pseudo-supervised network to solve the above problems. Summary of the Invention

[0005] The purpose of the present invention is to solve the problem in traditional microaneurysm image segmentation that there are limited number of labeled MA images in actual environments, while a large number of unlabeled images are not effectively utilized.

[0006] In order to achieve the above objectives, the basic solution of the present invention provides a semi-supervised segmentation model based on a cross-pseudo-supervised network, including:

[0007] Cross-pseudo-supervisory network: Two segmentation networks based on the PFFNet architecture are used. After the input data passes through the two segmentation networks, the output features are generated respectively and two pseudo labels are formed. The two pseudo labels supervise each other.

[0008] Loss function: In the training process of annotated images, the Combo loss function is used, and in the cross-supervised training of pseudo-labels, the cross-entropy loss function is introduced;

[0009] CBAM attention module: CBAM attention modules are introduced in the top and bottom feature branches of each stage of the PFFNet architecture.

[0010] Furthermore, in the training of the cross pseudo-supervisory network, labeled MA image sets and unlabeled MA image sets are used for learning and training.

[0011] Furthermore, the cross-pseudo-supervisory network includes two parallel segmentation networks, one of which introduces a multi-scale feature fusion module based on PFFNet; the other segmentation network embeds an attention module based on the PFFNet network.

[0012] Furthermore, the Combo loss function is as follows:

[0013]

[0014] in, Refers to the prediction result, y i refers to the ground truth, N refers to the total number of MAs in the fundus image, a refers to the weight of the focal loss, and They are Dice loss and Focal loss respectively.

[0015] Furthermore, the cross entropy loss function is expressed as follows:

[0016]

[0017] Where, and They are the cross-supervised loss function used for labeled images and the cross-supervised loss function used for unlabeled images, respectively.

[0018] Furthermore, the expressions of the cross-supervised loss function used for the labeled image and the cross-supervised loss function used for the unlabeled image are as follows:

[0019]

[0020] Where W and H are the width and height of the input image, respectively, and p 1i and p 2i are the predicted values of the two networks, y 1i and y 2i is the true labeled value, l ce is the cross entropy loss.

[0021] Furthermore, the CBAM attention module includes:

[0022] Channel attention module: processes the input feature map, generates channel dimension weight information, and weights the input feature map;

[0023] Spatial attention module: processes the feature map after the channel attention module, generates weight information of the spatial dimension, and combines it with the weighted result of the channel attention module to obtain a doubly weighted output feature map.

[0024] Furthermore, the channel attention module includes the following steps:

[0025] Perform global average pooling and global maximum pooling operations on the input feature map respectively;

[0026] The feature maps obtained by the two pooling operations are mapped to a low-dimensional space and then mapped back to the original space to generate a channel attention map;

[0027] Add the channel attention maps and normalize them to generate channel attention weights;

[0028] Multiply the generated channel attention weights with the original input feature map channel by channel to generate the final output feature map;

[0029] The spatial attention module includes the following steps:

[0030] The output feature map of the channel attention module is used as input, and the maximum pooling operation and the average pooling operation are performed on the input feature map;

[0031] Perform convolution operation on the merged feature map to generate a single-channel spatial attention weight map;

[0032] The generated spatial attention weight map is multiplied pixel by pixel with the input feature map to generate the final output of the entire CBAM attention module.

[0033] Based on the same inventive concept, the present invention discloses a semi-supervised microaneurysm segmentation method based on a cross pseudo-supervised network, including using the above-mentioned semi-supervised segmentation model based on a cross pseudo-supervised network to segment microaneurysm images.

[0034] Furthermore, the steps for segmenting microaneurysm images using the aforementioned semi-supervised segmentation model based on the cross-pseudo-supervised network are as follows:

[0035] Step S1, obtaining a lesion image;

[0036] Step S2, image preprocessing, using operations such as image flipping, random cropping, and image transformation to perform data enhancement on the lesion image;

[0037] In step S3, the preprocessed image is segmented using a semi-supervised segmentation model that integrates the CBAM attention module and the cross pseudo-supervised network.

[0038] The principle and effect of this solution are:

[0039] Based on the fact that there are only a limited number of annotated MA images in the real environment, while a large number of unannotated images are not effectively utilized, the present invention not only performs data augmentation operations based on data preprocessing and then generates a well-calibrated segmentation confidence map through a multi-scale feature fusion module, but also selectively embeds a CBAM attention module in the feature extraction network to enhance the ability to extract lesion feature information. In addition, by constructing two segmentation networks for collaborative training and adopting a supervised loss function, the model's learning effect on complex samples is improved. Compared with supervised segmentation models and methods that only use annotated images, the present invention achieves different degrees of improvement in segmentation effect for annotated images of different scales. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0041] Figure 1 The figure shows the architecture of the semi-supervised segmentation model based on the cross pseudo-supervised network proposed in the embodiment of the present application;

[0042] Figure 2 The cross-pseudo-supervisory network training flow chart in the semi-supervised segmentation model based on the cross-pseudo-supervisory network proposed in an embodiment of the present application is shown;

[0043] Figure 3 The overall structure of the CBAM attention module in the semi-supervised segmentation model based on the cross pseudo-supervised network proposed in the embodiment of the present application is shown;

[0044] Figure 4The structure diagram of the channel attention module in the CBAM in the semi-supervised segmentation model based on the cross pseudo-supervised network proposed in an embodiment of the present application is shown;

[0045] Figure 5 The figure shows the structure of the spatial attention module in the semi-supervised segmentation model based on the cross pseudo-supervised network proposed in the embodiment of the present application;

[0046] Figure 6 A schematic diagram illustrating image preprocessing in the semi-supervised microaneurysm segmentation method based on a cross-pseudo-supervised network proposed in an embodiment of the present application is shown, wherein (a) is the original image, (b) is the vertically flipped image, (c) is the image magnified 20% outward, and (d) is the horizontally flipped image;

[0047] Figure 7 The figure shows the original image containing MA and the corresponding label example in the IDRiD dataset, where (a) refers to the original image containing MA and (b) refers to its corresponding MA label;

[0048] Figure 8 The figure shows the original image containing MA and the corresponding label example in the DDR dataset, where (a) refers to the original image containing MA and (b) refers to its corresponding MA label;

[0049] Figure 9 A histogram of the ablation experiment of the CBAM-CPS module in the semi-supervised microaneurysm segmentation method based on the cross pseudo-supervised network proposed in an embodiment of the present application is shown;

[0050] Figure 10 A schematic bar chart showing a comparative experiment in a semi-supervised microaneurysm segmentation method based on a cross pseudo-supervised network proposed in an embodiment of the present application is shown;

[0051] Figure 11 A schematic diagram of the segmentation results of a comparative experiment in the semi-supervised microaneurysm segmentation method based on a cross pseudo-supervised network proposed in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0052] In order to further illustrate the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the specific implementation methods, structures, features and effects of the present invention are described in detail below in conjunction with the accompanying drawings and preferred embodiments.

[0053] The semi-supervised segmentation model based on the cross pseudo-supervised network has the following overall architecture: Figure 1As shown in Figure 1, the CBAM-CPS model uses a cross-pseudo-supervisory network as its foundation to build two segmentation networks for collaborative training. The CBAM attention module is selectively embedded in the feature extraction network to enhance the pertinence and effectiveness of feature extraction. Furthermore, the CBAM-CPS model uses generated pseudo-labels for cross-supervision to update the parameters of the segmentation network and further optimize model performance. The architectures of the CBAM-CPS model are as follows:

[0054] Cross-pseudo-supervised network: This is the core architecture of the CBAM-CPS model. It uses two segmentation networks based on the PFFNet architecture. After data is input into these two segmentation networks, two OCRNet networks generate output features and form two pseudo-labels. The mutual supervision of these two pseudo-labels enables collaborative network training.

[0055] In the training of the cross pseudo-supervised network, the dataset is divided into two parts: the labeled MA image set A and the unlabeled MA image set B. The goal of the cross pseudo-supervised network is to use A and B to learn and train a segmentation model. The training process is as follows Figure 2 The cross-pseudo-supervisory network consists of two parallel segmentation networks. The two networks have exactly the same structure but different initialization parameters. Therefore, in the early stages of training, their output results for the same input sample will be different, as shown in the following formula:

[0056] T1=f(X,λ1)

[0057] T2=f(X,λ2)

[0058] Where X is the input sample, and the two networks are initialized with parameters λ1 and λ2, respectively. After processing by the networks, the output features are normalized using a softmax function to generate confidence maps T1 and T2. Based on the confidence maps, corresponding pseudo-labels Y1 and Y2 are further generated. The pseudo-labels generated by one network are used to supervise the training process of the other network.

[0059] The training method for a cross-pseudo-supervised network has the following notable features: it processes the input image through two networks with different initializations, generates their own pseudo-labels, and uses these pseudo-labels for cross-supervision. This mechanism not only fully utilizes the characteristic information of unlabeled data but also, through consistency regularization, encourages the two networks to maintain consistent predictions for the same input image. This consistency constraint helps prevent the network from overfitting to noise or erroneous information, thereby improving the model's robustness and generalization ability. During training, the pseudo-labels are used to supervise and update the network, effectively expanding the size of the training data. This data augmentation strategy not only increases the number of training samples for the network but also improves the training quality through diverse supervisory signals. Compared to traditional supervised learning that relies solely on limited labeled data, this method makes better use of data resources, thereby improving the overall performance of the model. As training progresses, the model's learning of the target gradually stabilizes. Due to the introduction of the cross-pseudo-supervision mechanism, this method ultimately achieves more accurate segmentation results than models trained solely on labeled images. This stability and improved accuracy demonstrate that the cross-pseudo-supervised network effectively exploits the inherent structure of unlabeled data to further optimize the model's learning process.

[0060] Two segmentation networks based on the PFFNet network are trained collaboratively. The first segmentation network is an improvement on PFFNet. By introducing a multi-scale feature fusion module, this network integrates the multi-scale feature information output by PFFNet within a pyramidal horizontal fusion architecture and comprehensively fuses these features to enhance global scale information. Furthermore, a feature enhancement module further optimizes the output, effectively alleviating the semantic ambiguity of MA features. The second segmentation network embeds an attention module on top of PFFNet. PFFNet consists of five stages, with the number of feature branches increasing gradually. The top-level feature branches have the highest resolution and provide rich spatial location information, while the bottom-level feature branches have the lowest resolution but richer semantic information. By adding the CBAM attention module to the top and bottom branches of each stage, the network significantly improves its ability to extract both location and semantic information. The main difference between the two networks lies in the design of the multi-scale feature fusion module and the attention module. During training, the pseudo-labels generated by the two networks are used for cross-supervision, enabling them to learn from each other's strengths and gradually optimize their performance through iterative training.

[0061] Loss Function: To improve the CBAM-CPS model's ability to learn small MA features, a combo loss function is used during training with labeled images. Furthermore, a cross-entropy loss function is introduced during cross-supervised training with pseudo-labels. This diverse loss function setting helps enhance the model's learning performance.

[0062] The model is trained on a set of both labeled and unlabeled images, using the aforementioned cross-supervision approach. The labeled images are trained using both supervised and cross-supervised methods. Therefore, the overall loss calculation can be split into two parts: the supervised loss for the labeled images and the pseudo-label supervised loss for cross-training.

[0063] The Combo loss function is used to perform the MA segmentation task. The Combo loss function combines the advantages of Dice loss and Focal loss to alleviate the class imbalance problem in the MA segmentation task and ensure that the network pays more attention to the distribution, area, and boundary of MA when generating the segmentation mask. The Combo loss function is shown in the following formula:

[0064]

[0065] in, Refers to the prediction result, y i refers to the ground truth, N refers to the total number of MAs in the fundus image, a refers to the weight of the focus loss, which is set to 1 in this embodiment, and p refers to the probability that the MA belongs to the positive class. and They are Dice loss and Focal loss respectively.

[0066] In cross-pseudo supervision, pseudo-label Y1 is used to supervise the confidence map T2 output by another network, and pseudo-label Y2 is used to supervise the confidence map T1 output by another network. This includes the supervised loss of labeled images and the supervised loss of unlabeled images. The cross-supervised loss function calculation method used for unlabeled images is shown in the following formula:

[0067]

[0068] Where W and H are the width and height of the input image, respectively, p 1i and p 2i are the predicted values of the two networks, y 1i and y 2i is the true labeled value, l ce is the cross entropy loss.

[0069] The cross-supervision loss function used for labeled images is as follows:

[0070]

[0071] The overall cross-supervision loss is composed of the above two cross losses, and the calculation method is as follows:

[0072]

[0073] Therefore, the loss function of the entire network is as follows:

[0074] L=L Combo +λL cps

[0075] Where λ is a tuning parameter used to balance the supervision loss and the cross loss.

[0076] CBAM Attention Module: In the PFFNet architecture, the network is divided into five stages, with the number of feature branches gradually increasing from the input to the output, while the feature resolution gradually decreases from top to bottom. The top-level feature branches have rich spatial information due to their higher resolution, while the bottom-level feature branches contain strong semantic information. By introducing the CBAM attention module in the top and bottom-level feature branches of each stage, the network can focus on key feature information in the spatial and channel dimensions, thereby enhancing the ability to learn MA features and improving the overall performance of the model.

[0077] In the MA segmentation task, the CBAM attention module can adaptively obtain weight information, highlight important features, and suppress unimportant features, thereby improving segmentation performance. The overall structure of the CBAM attention module is as follows: Figure 3 As shown in Figure 1, the CBAM attention module consists of two parts: the channel attention module and the spatial attention module. Specifically, the input feature map is first processed by the channel attention module to generate channel dimension weight information and weight the input feature map. Subsequently, the feature map processed by the channel attention is further input into the spatial attention module to generate spatial dimension weight information. This is combined with the weighted result of the channel attention module to finally obtain a doubly weighted output feature map. The entire process is shown in the following two equations:

[0078]

[0079]

[0080] Where X represents the input feature map, X' represents the feature map after channel dimension weighting, X" represents the double-weighted output feature map, F and F' represent the weight information of channel dimension and spatial dimension respectively, and M C is the channel attention, M S For spatial attention, the symbol Represents multiplication (weighted).

[0081] The structure of the channel attention module in CBAM is as follows Figure 4 As shown in the figure, the principle is to adaptively adjust the channel weights of the feature map to highlight important feature channels and suppress unimportant channels. First, global average pooling and global maximum pooling operations are performed on the input feature map respectively; second, the feature maps obtained by the above two pooling operations are fed into a shared multi-layer perceptron (MLP). This MLP usually contains two fully connected layers, whose function is to map the pooled feature map to a low-dimensional space and then map it back to the original space to generate a channel attention map. The two feature maps processed by the MLP are added together and normalized using the sigmoid activation function to generate channel attention weights; finally, the generated channel attention weights are multiplied by the original input feature map channel by channel to generate the final output feature map.

[0082] The structure of the spatial attention module is as follows Figure 5 As shown in the figure, the spatial attention module adaptively adjusts the spatial attention weights of the feature map to extract spatial information relevant to the segmentation task. First, the output feature map of the channel attention module is used as input. A maximum pooling operation is performed on the input feature map to capture the salient spatial information in the feature map. Then, an average pooling operation is performed to extract the global spatial information of the feature map. The combined feature map is then convolved with a 7×7 convolution kernel to generate a single-channel spatial attention weight map. Finally, the generated spatial attention weight map is multiplied pixel by pixel with the input feature map to generate the final output of the entire CBAM attention module.

[0083] Based on the same inventive concept, this embodiment also discloses a semi-supervised microaneurysm segmentation method based on a cross pseudo-supervised network, comprising the following steps:

[0084] Step S1, obtaining a lesion image;

[0085] Step S2, image preprocessing, uses image flipping, random cropping, image transformation and other operations to perform data enhancement on the lesion image, as follows:

[0086] Image flipping: Due to the small number of images with MA annotations, the trained model is often prone to overfitting. Therefore, it is necessary to alleviate this problem by expanding the number of images in the data set. This embodiment performs two operations on the original image: vertical flipping and horizontal flipping. Vertical flipping is to flip the image along the horizontal symmetry axis, and the generated new image corresponds to the original image in the vertical direction. The image after vertical flipping remains consistent in the horizontal direction, but the upper and lower positions are opposite, such as Figure 6(b) is the effect of vertical flipping of the image. Horizontal flipping is to flip the image along the vertical symmetry axis. The generated new image corresponds to the original image in the horizontal direction. The image after horizontal flipping remains consistent in the vertical direction, but the left and right positions are opposite, such as Figure 6 (d) is the new image after horizontal flipping.

[0087] Image scaling: Image scaling is to enlarge or reduce an image by linear interpolation to change the scale of the MA features in the image. Its purpose is to simulate the image scale changes caused by different shooting devices or scenes in actual applications. Through image scaling, the trained model can better adapt to various task scenarios, thereby improving the robustness of the model. For example, enlarging or reducing an image by a certain ratio can generate new image samples and enhance the model's recognition ability for MAs of different scales. An example of the scaling effect of an MA image is as follows: Figure 6 The preprocessing operation (c) is shown, which shows the effect after the lesion image is magnified by 20%.

[0088] Image padding: After scaling an image, padding is required to ensure that its width and height are consistent. Here, zero padding is used to maintain the same height and width. This padding ensures that the image maintains a consistent aspect ratio during subsequent processing, while preventing processing errors caused by inconsistent dimensions. The padded image visually retains the core content of the original image while meeting the model input size requirements.

[0089] Image cropping: Due to the high resolution of the images in the dataset, directly using these images as input will increase the hardware burden and reduce the training and inference speed of the model. Therefore, it is necessary to crop the scaled image. This embodiment uses random cropping to crop the image to a fixed size of 512×512 pixels. Random cropping not only reduces the image resolution, but also increases the diversity of the data by randomly selecting the cropped area, preventing the model from overfitting. In addition, random cropping enables the model to learn more local features during training, thereby improving the model's recognition effect and stability on the target.

[0090] In step S3, the preprocessed image is segmented using a semi-supervised segmentation model that integrates the CBAM attention module and the cross pseudo-supervised network.

[0091] In order to verify the effectiveness of the semi-supervised segmentation model (CBAM-CPS) disclosed in this embodiment, the following experiments were conducted:

[0092] Experimental data: IDRiD is a publicly accessible database of fundus retinal images. The image resolution in the dataset is 4288×2848 and takes up about 1M of space. The semantic segmentation task in the dataset includes 81 images. All images have been accurately annotated by ophthalmologists, marking lesion areas such as microaneurysms, hemorrhages, and hard exudates, and corresponding segmentation annotations are provided for each image. Therefore, it is suitable as a dataset for studying MA segmentation algorithms. Since IDRiD does not have a given test or validation set, we performed ten-fold cross-validation on the images in IDRiD to evaluate the performance of the proposed MA segmentation model. Ten-fold cross-validation refers to randomly dividing each dataset into ten partitions of equal size. One of the partitions is used as a test and validation set, and the remaining nine partitions are used as training sets. The cross-validation is repeated ten times, and after obtaining ten performance results, the average of these results is taken as the final estimate. Examples of original images containing MA and corresponding labels in the IDRiD dataset are shown below. Figure 7 As shown in the figure, (a) refers to the original image containing MA, and (b) refers to its corresponding MA label.

[0093] DDR is a color fundus image dataset for the scientific research of diabetic retinopathy. DDR contains 757 images manually labeled as MA by ophthalmologists. The dataset has various resolutions, ranging from 1088×1920 to 3456×5184 pixels. 383 of them are used as training sets, 149 as validation sets, and the remaining 225 as test sets. As known in the prior art, DDR is the largest lesion segmentation dataset in fundus images. Examples of original images containing MA and corresponding labels in the DDR dataset are as follows: Figure 8 As shown in the figure, (a) refers to the original image containing MA, and (b) refers to its corresponding MA label.

[0094] In addition to the above labeled data sets, another 1,439 unlabeled images are also included. A part of them is added to the labeled image set as the original data. The images need to be preprocessed before inputting the model. In addition to using the preprocessing method in Chapter 3 to process the labeled and unlabeled images, image flipping, random cropping, image transformation and other operations are also used to perform data enhancement on the input lesion images, and then use them as the input of the model.

[0095] Experimental evaluation indicators: In order to compare the experimental results more intuitively, the evaluation indicators used include: union intersection (IOU), recall rate (Recall), precision (Precision) and Dice.

[0096] Experimental environment: Considering that the semi-supervised segmentation method used has high hardware requirements, the training batch size is set to 4 in the experiment.

[0097] This embodiment uses two sets of experiments for analysis. The first set of experiments is an ablation experiment. By constructing multiple variant structures based on a cross-supervised network, the effects of image enhancement, segmentation network design, and attention module on lesion segmentation performance are verified, thereby proving the effectiveness of each module. The second set of experiments is a comparative experiment. By using unlabeled images of different scales, it is verified that the proposed method can effectively utilize the feature information in a large number of unlabeled MA images, thereby improving the model's learning ability for MA. The experimental process and result analysis will be described in detail below.

[0098] Ablation experiments: A series of experiments were conducted to verify the rationality and effectiveness of the design of each module in the CBAM-CPS model, covering data augmentation, multi-scale feature fusion modules, and attention modules. First, the effectiveness of data augmentation was verified by comparing training images with and without data augmentation. Subsequently, ablation experiments were conducted on the CBAM-CPS model, using various variant structures to evaluate the role of the multi-scale feature fusion module and attention module in the overall model. The details are as follows:

[0099] (1) Data augmentation ablation experiment

[0100] To verify the effectiveness of data augmentation in the MA segmentation task, this embodiment uses a set of comparative experiments. The experiments were conducted on the CBAM-CPS model, one group using data augmentation and the other not. In the experiment, the ratio of annotated images to the total number of images was 1:8. According to the results in Table 1, the model using data augmentation outperformed the model without data augmentation in all evaluation indicators. Specifically, after using data augmentation, the model's mIoU increased from 66.95% to 71.32%, mDice increased from 74.48% to 77.55%, mRecall increased from 74.79% to 77.02%, and mPrecision increased from 74.18% to 78.32%. These results show that data augmentation significantly improves the segmentation performance of the model, especially in key indicators such as mIoU and mDice scores.

[0101] Table 1 Results of MA segmentation with and without data augmentation

[0102]

[0103] (2) CBAM-CPS module ablation experiment

[0104] To evaluate the effectiveness of each module in the CBAM-CPS model, this example conducted an ablation experiment. During the dataset processing phase, the experiment implemented a data augmentation strategy on the raw data. This strategy included not only fundus image green channel extraction, contrast-limited adaptive histogram equalization, gamma correction, and non-local means denoising preprocessing, but also image flipping, scaling, padding, and cropping. Furthermore, the ratio of annotated to unannotated images used in the experiment was set at 1:8.

[0105] In the architectural design of the CBAM-CPS model, this embodiment selectively embeds the CBAM attention module and the multi-scale feature fusion module in the segmentation network of the cross-pseudo-supervisory network. The ablation experiment forms multiple variant structures by adding and deleting these modules and verifies their effectiveness. Among them, the CBAM attention module is denoted as C and the multi-scale feature fusion module is denoted as D. The experimental results of different variant structures are recorded in the experiment. The evaluation indicators include mIoU, mDice and mRecall. The results are shown in Table 2. In order to present the segmentation results more intuitively, a bar chart is drawn, as shown in Figure 9 shown.

[0106] Table 4-2 Ablation experiment results of different variant structures in MA segmentation task

[0107]

[0108]

[0109] from Figure 9 It can be observed that the CBAM attention module and the multi-scale feature fusion module significantly improve segmentation performance. According to the results in Table 2, when using only two U-shaped segmentation networks, the mDice (mean Distance Over Union) is 71.69%. Replacing one of the segmentation networks with a network embedded with the CBAM attention module increases mDice by 3.17 percentage points, demonstrating that this module significantly improves the detection of small-scale objects (MA). This is likely due to the fact that the attention mechanism is embedded in the feature branch with the highest resolution during the feature extraction stage of the first segmentation network, thereby enhancing the ability to focus on small objects. Furthermore, introducing the multi-scale feature fusion module into the cross-pseudo-supervised model based on two U-shaped segmentation networks improves mDice by 3.11 percentage points, and other metrics also show varying degrees of improvement. Specifically, mIoU (mean Intersection Over Union) and mRecall (mean Recall) increase by 3.7 percentage points and 3.96 percentage points, respectively. This demonstrates that the multi-scale feature fusion module effectively filters low-confidence samples, reduces the transfer of erroneous information to the segmentation network, and significantly improves the model's training performance for small-scale MA.

[0110] In summary, the various improved modules introduced in the CBAM-CPS model have a significant effect on improving the MA segmentation performance. The ablation experiment results fully verify the effectiveness of these modules in optimizing the lesion segmentation effect.

[0111] Comparative experiment: In order to verify the rationality of the MA segmentation method based on semi-supervised learning proposed in this embodiment, this section analyzes it through comparative experiments. The semi-supervised learning method introduces unlabeled images as a training set based on the labeled images. In many semi-supervised learning studies, good results can be achieved using only a small amount of labeled data, and the impact of different proportions of unlabeled data on the experimental results is also different. Therefore, based on the CBAM-CPS semi-supervised segmentation method, MA segmentation experiments with different proportions of labeled images were carried out. 81 labeled images were used in the experiment, and their number remained unchanged, while unlabeled images were added as training sets in different proportions. The ratio of the number of labeled images to the total number of images was set to 1 / 2, 1 / 4 and 1 / 8 respectively. The evaluation indicators used in the experiment include mIoU, mDice and mRecall. The experimental results are shown in Table 3. In the first column of Table 3, the left side of the slash in the brackets represents the number of labeled images, and the right side represents the total number of images. In order to present the experimental results more intuitively, a bar chart was drawn, as shown in Figure 10 shown.

[0112] Table 3 Comparative experimental results of the number of annotated images at different scales

[0113]

[0114] from Figure 10It can be seen that the semi-supervised segmentation method proposed in the embodiment achieves varying degrees of improvement in segmentation performance after introducing different proportions of unlabeled images. As the number of unlabeled images increases, various metrics also show an upward trend. According to Table 3, the mDice value of the PFFNet network segmentation is 74.11%. When the ratio of labeled images to the total number of images is 1 / 2, the semi-supervised segmentation method increases the mDice value by 0.37 percentage points, reaching 75.48%. Simultaneously, the mIoU and mRecall metrics also increase by approximately 1 percentage point each. This preliminarily demonstrates that the embodiment method can effectively utilize the feature information of unlabeled images to promote the model's learning of lesion characteristics. When the ratio of labeled images to the total number of images is increased to 1 / 4, the mDice value further improves, increasing by 2.88 percentage points compared to the fully supervised PFFNet network and by 1.51 percentage points compared to the semi-supervised method with a 1 / 2 ratio. All metrics show improvement, indicating that increasing the number of unlabeled images within a certain range can enhance the model's ability to learn lesion characteristics. Specifically, the mIoU and mDice values improved significantly, while the mRecall metric saw a less pronounced improvement. This may be because the number of unlabeled images was still insufficient to fully cover the feature distribution. When the ratio of labeled images to total images was increased to 1 / 8, the mRecall metric increased from 76.72% at 1 / 4 to 77.02%, indicating that significantly increasing the number of unlabeled images significantly improved the model's learning performance.

[0115] Figure 10 The visualization of the indicator results in also draws a similar conclusion: under the premise of fixing the number of labeled images, as the number of unlabeled images increases, the various indicators of the semi-supervised segmentation method perform better and better. Figure 10 Comparing the results of the segmentation from 1 / 2 to 1 / 4 and from 1 / 4 to 1 / 8 scales, we can see that the improvement is more significant when the number of unlabeled images is significantly increased. This may be because the training set is small, and the large number of unlabeled images enriches the distribution information of MA features. This also further proves that the semi-supervised segmentation method proposed in this example can effectively extract feature information from unlabeled images, thereby improving model performance.

[0116] The proposed semi-supervised segmentation method is analyzed by the actual MA segmentation effect. The segmentation results of the comparative experiment are as follows: Figure 11 As shown in the figure, it can be observed that compared with the supervised segmentation method that only uses annotated images, the semi-supervised segmentation method proposed in the embodiment achieves different degrees of improved segmentation effect under different proportions of annotated images. For example, in the MA segmentation results in the fourth row, as the number of unlabeled images increases, the MA segmentation effect gradually improves, which shows that the unlabeled images have a significant impact on the segmentation results in the semi-supervised segmentation method of the embodiment.

[0117] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed as a preferred embodiment as above, it is not intended to limit the present invention. Any person skilled in the art can make some changes or modifications to equivalent embodiments using the technical contents disclosed above without departing from the scope of the technical solution of the present invention. However, any indirect modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A semi-supervised segmentation model based on a cross-pseudo-supervised network, characterized by: include: Cross-pseudo-supervisory network: Two segmentation networks based on the PFFNet architecture are used. After the input data passes through the two segmentation networks, the output features are generated respectively and two pseudo labels are formed. The two pseudo labels supervise each other. Loss function: In the training process of annotated images, the Combo loss function is used, and in the cross-supervised training of pseudo-labels, the cross-entropy loss function is introduced; CBAM attention module: CBAM attention modules are introduced in the top and bottom feature branches of each stage of the PFFNet architecture.

2. The semi-supervised segmentation model and microaneurysm segmentation method based on a cross pseudo-supervised network according to claim 1, characterized in that: In the training of the cross pseudo-supervisory network, labeled MA image sets and unlabeled MA image sets are used for learning and training.

3. The semi-supervised segmentation model based on the cross pseudo-supervised network according to claim 2, characterized in that The cross-pseudo-supervised network includes two parallel segmentation networks, one of which introduces a multi-scale feature fusion module based on PFFNet; the other segmentation network embeds an attention module based on the PFFNet network.

4. The semi-supervised segmentation model based on the cross pseudo-supervised network according to claim 1, characterized in that The Combo loss function is as follows: in, Refers to the prediction result, y i refers to the ground truth, N refers to the total number of MAs in the fundus image, a refers to the weight of the focal loss, and They are Dice loss and Focal loss respectively.

5. The semi-supervised segmentation model based on the cross pseudo-supervised network according to claim 1, characterized in that The cross entropy loss function expression is as follows: Where, and They are the cross-supervised loss function used for labeled images and the cross-supervised loss function used for unlabeled images, respectively.

6. The semi-supervised segmentation model based on the cross pseudo-supervised network according to claim 5, characterized in that The expressions of the cross-supervised loss function used for the labeled image and the cross-supervised loss function used for the unlabeled image are as follows: Where W and H are the width and height of the input image, respectively, and p 1i and p 2i are the predicted values of the two networks, y 1i and y 2i is the true labeled value, l ce is the cross entropy loss.

7. The semi-supervised segmentation model based on the cross pseudo-supervised network according to claim 1, characterized in that The CBAM attention module includes: Channel attention module: processes the input feature map, generates channel dimension weight information, and weights the input feature map; Spatial attention module: processes the feature map after the channel attention module, generates weight information of the spatial dimension, and combines it with the weighted result of the channel attention module to obtain a doubly weighted output feature map.

8. The semi-supervised segmentation model based on the cross pseudo-supervised network according to claim 1, characterized in that The channel attention module includes the following steps: Perform global average pooling and global maximum pooling operations on the input feature map respectively; The feature maps obtained by the two pooling operations are mapped to a low-dimensional space and then mapped back to the original space to generate a channel attention map; Add the channel attention maps and normalize them to generate channel attention weights; Multiply the generated channel attention weights with the original input feature map channel by channel to generate the final output feature map; The spatial attention module includes the following steps: The output feature map of the channel attention module is used as input, and the maximum pooling operation and the average pooling operation are performed on the input feature map; Perform convolution operation on the merged feature map to generate a single-channel spatial attention weight map; The generated spatial attention weight map is multiplied pixel by pixel with the input feature map to generate the final output of the entire CBAM attention module.

9. A semi-supervised microaneurysm segmentation method based on a cross-pseudo-supervised network, characterized in that: It includes using the semi-supervised segmentation model based on the cross pseudo-supervised network described in any one of claims 1-8 to segment microaneurysm images.

10. The semi-supervised microaneurysm segmentation method based on a cross-pseudo-supervised network according to claim 9, characterized in that: The steps of segmenting a microaneurysm image using the semi-supervised segmentation model based on a cross pseudo-supervised network according to any one of claims 1 to 8 are as follows: Step S1, obtaining a lesion image; Step S2, image preprocessing, using operations such as image flipping, random cropping, and image transformation to perform data enhancement on the lesion image; In step S3, the preprocessed image is segmented using a semi-supervised segmentation model that integrates the CBAM attention module and the cross pseudo-supervised network.

Images