Semi-supervised OCTA blood vessel segmentation method and imaging method

Abstract

The invention discloses a semi-supervised OCTA blood vessel segmentation method. The method comprises the steps that an existing OCTA eye image is acquired; selecting a plurality of images to carry out blood vessel labeling, not labeling the remaining images, and carrying out random data enhancement on the images to construct a training data set; constructing a semi-supervised OCTA blood vessel segmentation primary model based on the convolution module and the attention mechanism, and training to obtain a semi-supervised OCTA blood vessel segmentation model; and adopting the obtained semi-supervised OCTA blood vessel segmentation model to carry out blood vessel segmentation of an actual OCTA eye image. The invention further discloses an imaging method comprising the semi-supervised OCTA blood vessel segmentation method. According to the method, blood vessel segmentation and imaging of the OCTA eye image can be achieved, a large number of accurately-labeled OCTA eye images do not need to serve as training data, the reliability is higher, and the accuracy is better.

Description

Technical Field

[0001] The present invention belongs to the field of image processing, and particularly relates to a semi-supervised OCTA blood vessel segmentation method and an imaging method. Background Art

[0002] OCTA is a non-invasive imaging method that can clearly show the morphology and changes of microvessels in the retina. The corresponding OCTA images are important data supports for clinical applications and basic research applications in the field of ophthalmology. Therefore, segmenting and annotating blood vessels in OCTA images is of great significance for clinical applications and basic research applications in the field of ophthalmology.

[0003] At present, the traditional blood vessel segmentation schemes for OCTA images mainly include traditional segmentation methods based on machine learning and segmentation methods based on deep learning, and both of these two types of schemes require a large number of accurate OCTA annotated images. However, OCTA images have high blood vessel resolution, and the labor cost and time cost required to calibrate such high-resolution images are very high. Therefore, the blood vessel annotation data of high-quality and accurate OCTA images is extremely scarce. Therefore, in the case of lacking reliable, accurate and sufficient training data, the reliability and accuracy of existing OCTA blood vessel segmentation schemes are relatively poor. Summary of the Invention

[0004] One of the purposes of the present invention is to provide a semi-supervised OCTA blood vessel segmentation method with high reliability and good accuracy.

[0005] Another purpose of the present invention is to provide an imaging method including the semi-supervised OCTA blood vessel segmentation method.

[0006] The semi-supervised OCTA blood vessel segmentation method provided by the present invention includes the following steps:

[0007] S1. Obtain existing OCTA eye images;

[0008] S2. For the OCTA eye images obtained in step S1, select several images for blood vessel annotation, do not annotate the remaining images, and perform random data augmentation on the images to construct a training data set;

[0009] S3. Based on a convolutional module and an attention mechanism, construct a semi-supervised OCTA blood vessel segmentation primary model;

[0010] S4. Use the training data set obtained in step S2 to train the semi-supervised OCTA blood vessel segmentation primary model constructed in step S3 to obtain a semi-supervised OCTA blood vessel segmentation model;

[0011] S5. Use the semi-supervised OCTA vascular segmentation model obtained in step S4 to perform vascular segmentation on actual OCTA eye images.

[0012] For the OCTA eye images obtained in step S1 described in step S2, select several images for vascular annotation, do not annotate the remaining images, and perform random data augmentation on the images to construct a training dataset, which specifically includes the following steps:

[0013] Among the OCTA eye images obtained in step S1, according to a set ratio, perform vascular annotation on several OCTA eye images and use them as labeled data images; for the remaining OCTA eye images, do not perform vascular annotation and use them as unlabeled data images; all labeled data images and unlabeled data images together constitute the first dataset;

[0014] In the first dataset, randomly select several images and ensure that the selected images include both labeled data images and unlabeled data images; for the selected images, perform data augmentation to obtain a training dataset; the training dataset includes the original labeled data images, the original unlabeled data images, the augmented labeled data images, and the augmented unlabeled data images;

[0015] During training, ensure that the images input into the model in each round of training include the original labeled data images, the original unlabeled data images, the augmented labeled data images, and the augmented unlabeled data images.

[0016] Based on the convolutional module and the attention mechanism described in step S3, construct a semi-supervised OCTA vascular segmentation primary model, which specifically includes the following steps:

[0017] The constructed semi-supervised OCTA vascular segmentation primary model includes an encoding module, a latent vector module, and a decoding module;

[0018] Construct an encoding module based on the convolutional module; the encoding module is used to extract high-dimensional features from the input image, gradually compress the spatial dimension and extract semantic information, and provide higher-level semantic information for subsequent processing modules;

[0019] Construct a latent vector module based on serpentine convolution and channel attention mechanism; the latent vector module captures local structural details in the image through serpentine convolution, and at the same time combines the channel attention mechanism to enhance the expression ability of key features, so as to generate a more accurate latent representation for supporting image feature learning;

[0020] Construct a decoding module based on the transposed convolution module; the decoding module is used to map the obtained features back to the original space, gradually restore the spatial resolution of the image through transposed convolution, and finally reconstruct the image segmentation result to ensure the accuracy of the output and retain image details.

[0021] The construction of the encoding module based on the convolutional module specifically includes the following steps:

[0022] The encoding module includes a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a third pooling layer, a fourth convolutional layer, a fourth pooling layer, and a fifth convolutional layer;

[0023] The first convolutional layer, the first pooling layer, the second convolutional layer, the second pooling layer, the third convolutional layer, the third pooling layer, the fourth convolutional layer, the fourth pooling layer, and the fifth convolutional layer are connected in series in sequence;

[0024] The first convolutional layer includes a basic convolutional sub-module; the second convolutional layer, the third convolutional layer, the fourth convolutional layer, and the fifth convolutional layer each include two basic convolutional sub-modules connected in series in sequence; the basic convolutional sub-module includes a 3×3 convolutional layer, a batch normalization layer, and a RELU activation function layer connected in series in sequence;

[0025] The first pooling layer, the second pooling layer, the third pooling layer, and the fourth pooling layer are all 2×2 pooling layers.

[0026] The construction of the latent vector module based on the serpentine convolution and channel attention mechanism specifically includes the following steps:

[0027] The latent vector module includes a serpentine convolution module and a channel attention module; the input data of the latent vector module is divided into two paths: the first path is input into the serpentine convolution module for processing, and the second path is input into the channel attention module for processing; the output of the serpentine convolution module and the output of the channel attention module are added bit by bit to obtain the output of the latent vector module;

[0028] The serpentine convolution module includes a coordinate mapping sub-module and a deformed convolution feature calculation sub-module connected in series in sequence; the coordinate mapping sub-module is implemented based on stacking and decompression. The coordinate mapping sub-module offsets each position of the convolution kernel along the x-axis and y-axis by dynamically adjusting the coordinate mapping of the convolution kernel to adapt to different image structures; the coordinate mapping sub-module ensures that the offset of the convolution kernel can capture the local geometry by gradually selecting positions, thereby improving the information perception ability of the convolution kernel; the deformed convolution feature calculation sub-module is implemented based on bilinear interpolation. The deformed convolution feature calculation sub-module calculates the features of the offset convolution kernel position by the bilinear interpolation method to ensure that the convolution kernel can extract features;

[0029] The channel attention module dynamically adjusts the weights of channels by calculating the similarity between channels, thereby achieving differential responses to different channels.

[0030] The decoding module constructed based on the transposed convolution module specifically includes the following steps:

[0031] The decoding module includes a main decoder and an auxiliary decoder;

[0032] If the input of the encoding module is the enhanced labeled data image or the enhanced unlabeled data image, the output of the latent vector module is used as the input of the main decoder; if the input of the encoding module is the original labeled data image or the original unlabeled data image, the output of the latent vector module is used as the input of the auxiliary decoder;

[0033] The main decoder and the auxiliary decoder have the same structure, both including a first upsampling sub-module, a second upsampling sub-module, a third upsampling sub-module, a fourth upsampling sub-module, and an inference sub-module connected in series in sequence;

[0034] The first upsampling sub-module, the second upsampling sub-module, the third upsampling sub-module, and the fourth upsampling sub-module have the same structure, all including a 2×2 transposed convolution layer, a first 3×3 convolution layer, a first batch normalization layer, a first RELU activation function layer, a second 3×3 convolution layer, a second batch normalization layer, and a second RELU activation function layer connected in series in sequence;

[0035] The inference sub-module includes a 1×1 convolution layer.

[0036] The training described in step S4 specifically includes the following steps:

[0037] If the input of the encoding module is the enhanced labeled data image, the segmentation loss L is additionally calculated using the following formula Seg :

[0038] L Seg = θ × Dice(f m (A(x)), A(y)) + (1 - θ) × D E (T(f m (A(x))), T(A(y)))

[0039] In the formula, θ is the training epoch parameter, and k is the training epoch, τ is the temperature coefficient; Dice() is the Dice loss function; A() represents image enhancement; x is the original image; y is the corresponding label; f m () is the processing function of the main decoder; D E () is the Euclidean distance calculation function; T() is the triangular feature extraction function;

[0040] For all input data images, the topological skeleton loss L is calculated using the following formula TS :

[0041] L TS = Dice(f m (A(x)), A(f a (x)))+ Dice(S(f m (A(x))), S(A(f a (x))))

[0042] where f a () is the processing function of the auxiliary decoder; S() is the topological skeleton extraction function;

[0043] The following formula is used as the total loss function L:

[0044] L = L S + ω × L U

[0045] where ω is the ramp-up coefficient; L S is the total loss function of the first stage; L U is the total loss function of the second stage; for the labeled data processing stage, the calculation formula of the total loss function L S of the first stage is L S = L Seg + L TS ; for the unlabeled data processing stage, the calculation formula of the total loss function L U of the second stage is L U = L TS .

[0046] Using the semi-supervised OCTA vascular segmentation model obtained in step S4 in step S5, perform actual vascular segmentation on OCTA eye images, specifically including the following steps:

[0047] Obtain the actual OCTA eye image and input it into the semi-supervised OCTA vascular segmentation model obtained in step S4;

[0048] Take the output of the auxiliary decoder as the vascular segmentation result of the actual OCTA eye image.

[0049] The present invention also provides an imaging method including the above-mentioned semi-supervised OCTA vascular segmentation method, further including the following steps:

[0050] S6. Mark and perform secondary imaging on the vascular segmentation result obtained in step S5 on the actual OCTA eye image to obtain an OCTA eye image with the vascular segmentation result.

[0051] The semi-supervised OCTA vascular segmentation method and imaging method provided by the present invention perform partial annotation on OCTA eye images, combine a convolutional module and an attention mechanism, and adopt a semi-supervised learning method to train and apply a semi-supervised OCTA vascular segmentation model. Therefore, the present invention can not only achieve vascular segmentation and imaging of OCTA eye images, but also does not require a large number of accurately annotated OCTA eye images as training data, with higher reliability and better accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a schematic flowchart of the method for the segmentation method of the present invention.

[0053] Figure 2 It is a schematic visualization comparison diagram of the segmentation results of the segmentation method of the present invention and the existing scheme on the first data set.

[0054] Figure 3 It is a schematic visualization comparison diagram of the segmentation results of the segmentation method of the present invention and the existing scheme on the second data set.

[0055] Figure 4 It is a schematic flowchart of the imaging method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0056] As Figure 1 shown is a schematic flowchart of the method for the segmentation method of the present invention: The semi-supervised OCTA vascular segmentation method provided by the present invention includes the following steps:

[0057] S1. Obtain existing OCTA eye images;

[0058] S2. For the OCTA eye images obtained in step S1, select several images for vascular annotation, do not annotate the remaining images, and perform random data augmentation on the images to construct a training data set; specifically, it includes the following steps:

[0059] Among the OCTA eye images obtained in step S1, according to a set ratio, several OCTA eye images are subjected to vascular annotation and used as labeled data images; for the remaining OCTA eye images, no vascular annotation is performed and they are used as unlabeled data images; all labeled data images and unlabeled data images together form a first data set;

[0060] In the first data set, randomly extract several images and ensure that the extracted images include both labeled data images and unlabeled data images; for the extracted images, perform data augmentation to obtain a training data set; the training data set includes the original labeled data images, the original unlabeled data images, the augmented labeled data images, and the augmented unlabeled data images;

[0061] During training, ensure that the images input into the model in each round of training include the original labeled data images, the original unlabeled data images, the enhanced labeled data images, and the enhanced unlabeled data images;

[0062] S3. Based on the convolutional module and the attention mechanism, construct a semi-supervised OCTA vascular segmentation primary model; specifically, it includes the following steps:

[0063] The constructed semi-supervised OCTA vascular segmentation primary model includes an encoding module, a latent vector module, and a decoding module;

[0064] Construct an encoding module based on the convolutional module; the encoding module is used to extract high-dimensional features from the input image, gradually compress the spatial dimension, and extract semantic information, providing higher-level semantic information for the subsequent processing module;

[0065] The encoding module includes a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a third pooling layer, a fourth convolutional layer, a fourth pooling layer, and a fifth convolutional layer;

[0066] The first convolutional layer, the first pooling layer, the second convolutional layer, the second pooling layer, the third convolutional layer, the third pooling layer, the fourth convolutional layer, the fourth pooling layer, and the fifth convolutional layer are connected in series in sequence;

[0067] The first convolutional layer includes a basic convolutional sub-module; the second convolutional layer, the third convolutional layer, the fourth convolutional layer, and the fifth convolutional layer each include two sequentially connected basic convolutional sub-modules; the basic convolutional sub-module includes a 3×3 convolutional layer, a batch normalization layer, and a RELU activation function layer connected in series in sequence;

[0068] The first pooling layer, the second pooling layer, the third pooling layer, and the fourth pooling layer are all 2×2 pooling layers;

[0069] Construct a latent vector module based on serpentine convolution and channel attention mechanism; the latent vector module flexibly captures local structural details in the image through serpentine convolution, and at the same time combines the channel attention mechanism to enhance the expression ability of key features, thereby generating a more accurate latent representation for supporting higher-quality image feature learning;

[0070] The latent vector module includes a serpentine convolution module and a channel attention module; the input data of the latent vector module is divided into two paths: the first path is input into the serpentine convolution module for processing, and the second path is input into the channel attention module for processing; the output of the serpentine convolution module and the output of the channel attention module are added bit by bit to obtain the output of the latent vector module;

[0071] The serpentine convolution module includes a coordinate mapping sub-module and a deformable convolution feature calculation sub-module connected in series in sequence; the coordinate mapping sub-module is implemented based on stacking and decompression. The coordinate mapping sub-module dynamically adjusts the coordinate mapping of the convolution kernel, offsets each position of the convolution kernel along the x-axis and y-axis to adapt to different image structures; the coordinate mapping sub-module ensures that the offset of the convolution kernel can effectively capture complex local geometries, especially slender tubular structures, by gradually selecting positions, thereby enhancing the convolution kernel's perception ability for key information; the deformable convolution feature calculation sub-module is implemented based on bilinear interpolation. The deformable convolution feature calculation sub-module calculates features for the offset convolution kernel positions through bilinear interpolation, ensuring that even if the offset causes position changes, the convolution kernel can still stably extract features; this sub-module not only ensures that the perception area does not shift significantly during the deformation process but also provides more accurate and delicate feature extraction capabilities on local structures, especially suitable for slender blood vessels or tubular structures, ensuring the correct capture of their geometries;

[0072] The channel attention module dynamically adjusts the weights of channels by calculating the similarity between channels, thereby achieving differential responses to different channels;

[0073] The decoding module is constructed based on the transposed convolution module; the decoding module is used to map the obtained features back to the original space and gradually restore the spatial resolution of the image through transposed convolution, ultimately reconstructing a high-quality image segmentation result to ensure the accuracy of the output and the retention of details;

[0074] The decoding module includes a main decoder and an auxiliary decoder;

[0075] If the input of the encoding module is the enhanced labeled data image or the enhanced unlabeled data image, the output of the latent vector module is used as the input of the main decoder; if the input of the encoding module is the original labeled data image or the original unlabeled data image, the output of the latent vector module is used as the input of the auxiliary decoder;

[0076] The main decoder and the auxiliary decoder have the same structure, both including a first upsampling sub-module, a second upsampling sub-module, a third upsampling sub-module, a fourth upsampling sub-module, and an inference sub-module connected in series in sequence;

[0077] The first upsampling sub-module, the second upsampling sub-module, the third upsampling sub-module, and the fourth upsampling sub-module have the same structure, all including a 2×2 transposed convolution layer, a first 3×3 convolution layer, a first batch normalization layer, a first RELU activation function layer, a second 3×3 convolution layer, a second batch normalization layer, and a second RELU activation function layer connected in series in sequence;

[0078] The inference sub-module includes a 1×1 convolution layer;

[0079] S4. Use the training data set obtained in step S2 to train the semi-supervised OCTA vascular segmentation primary model constructed in step S3 to obtain a semi-supervised OCTA vascular segmentation model;

[0080] During training, it specifically includes the following steps:

[0081] If the input of the encoding module is the enhanced labeled data image, the segmentation loss L is additionally calculated using the following formula Seg :

[0082] L Seg = θ × Dice(f m (A(x)), A(y)) + (1 - θ) × D E (T(f m (A(x))), T(A(y)))

[0083] In the formula, θ is the training round parameter, which is used to make the model have different preferences in different training stages, and k is the training round, τ is the temperature coefficient; Dice() is the Dice loss function; A() represents image enhancement; x is the original image; y is the corresponding label; f m () is the processing function of the main decoder; D E () is the Euclidean distance calculation function; T() is the triangular feature extraction function;

[0084] For all input data images, the topological skeleton loss L is calculated using the following formula TS :

[0085] L TS = Dice(f m (A(x)), A(f a (x)))+ Dice(S(f m (A(x))), S(A(f a (x))))

[0086] In the formula, f a () is the processing function of the auxiliary decoder; S() is the topological skeleton extraction function;

[0087] The following formula is used as the total loss function L:

[0088] L = L S + ω × L U

[0089] In the formula, ω is the ramp-up coefficient, which is used to make the weight of the unsupervised loss increase continuously with the iteration of the training process, and gradually improve the influence of the unsupervised stage training on the update of network parameters; L S is the total loss function of the first stage; LU is the total loss function for the second stage; for the labeled data processing stage, the total loss function L of the first stage S is calculated as L S = L Seg + L TS ; for the unlabeled data processing stage, the total loss function L of the second stage U is calculated as L U = L TS ;

[0090] S5. Use the semi-supervised OCTA vascular segmentation model obtained in step S4 to perform actual vascular segmentation on OCTA eye images; specifically, it includes the following steps:

[0091] Obtain the actual OCTA eye image and input it into the semi-supervised OCTA vascular segmentation model obtained in step S4;

[0092] Take the output of the auxiliary decoder as the vascular segmentation result of the actual OCTA eye image.

[0093] The segmentation method of the present invention enhances the vascular segmentation accuracy without increasing the requirement for labeled data. It adjusts the position of the convolution kernel through dynamic spatial sampling to adapt to different image structures, especially slender blood vessels or tubular structures, thereby reducing the interference of the background and noise existing in the OCTA image to the network; through double topological consistency, the network has different learning preferences in different stages, thereby enhancing the network's learning of data topological features. The present invention improves the data utilization efficiency and further enhances the segmentation accuracy of the model when only using limited labeled data and a large amount of unlabeled data, making the present invention more suitable for scenarios with limited label quantities.

[0094] The following combines an embodiment to compare and explain the segmentation effects of the segmentation method of the present invention and the existing methods:

[0095] Compare the method of the present invention with the existing segmentation methods on two datasets. Among them, the first dataset contains 30 training images and 9 test images; the second dataset contains 60 training images and 20 test images. On the first dataset, all methods use 3.3% of the labeled data, and on the second dataset, all methods use 5.0% of the labeled data. The accuracy, Dice coefficient, and false discovery rate of the segmentation results are used as evaluation criteria. All experimental results are obtained on the test set.

[0096] The specific test results are shown in Tables 1 to 2:

[0097] Table 1 Comparison schematic table of segmentation results for the first public dataset

[0098] Segmentation method Accuracy rate (%) Dice coefficient False positive rate (%) Semi-supervised method in 2017 0.8734 0.6693 0.3129 Semi-supervised method in 2020 0.9056 0.7395 0.2005 Semi-supervised method in 2022 0.9126 0.7627 0.1852 Semi-supervised method in 2024 0.9132 0.7661 0.1904 Method of the present invention 0.9170 0.7673 0.1518

[0099] Comparison Schematic Table of Segmentation Results of the Second Public Dataset in Table 2

[0100] Segmentation method Accuracy rate (%) Dice coefficient False positive rate (%) Semi-supervised method in 2017 0.9827 0.8576 0.1394 Semi-supervised method in 2020 0.9856 0.8788 0.0944 Semi-supervised method in 2022 0.9862 0.8869 0.1164 Semi-supervised method in 2024 0.9870 0.8964 0.1040 Method of the present invention 0.9877 0.9009 0.0856

[0101] As can be seen from Table 1 and Table 2, the method of the present invention has achieved the optimal results in three different metrics on two different datasets, verifying the feasibility and superiority of the method of the present invention.

[0102] Figure 2 and Figure 3 respectively show the visualization results of the method of the present invention and other semi-supervised methods on two different datasets. From left to right, they are the original image, manual annotation, semi-supervised method in 2017, semi-supervised method in 2020, semi-supervised method in 2022, semi-supervised method in 2024, and the visualization results of the method of the present invention. As can be seen from the figure, the method proposed by the present invention can achieve more accurate segmentation, specifically manifested as fewer abnormal phenomena such as break points and broken line segments in the segmentation results.

[0103] Such as Figure 4 shown is the schematic flow chart of the imaging method of the present invention: This imaging method including the semi-supervised OCTA vascular segmentation method disclosed by the present invention comprises the following steps:

[0104] S1. Obtain existing OCTA eye images;

[0105] S2. For the OCTA eye images obtained in step S1, select several images for vascular annotation, do not annotate the remaining images, and perform random data augmentation on the images, thereby constructing a training dataset;

[0106] S3. Based on the convolutional module and the attention mechanism, construct a primary semi-supervised OCTA vascular segmentation model;

[0107] S4. Use the training dataset obtained in step S2 to train the primary semi-supervised OCTA vascular segmentation model constructed in step S3 to obtain a semi-supervised OCTA vascular segmentation model;

[0108] S5. Use the semi-supervised OCTA vascular segmentation model obtained in step S4 to perform vascular segmentation on actual OCTA eye images;

[0109] S6. Label and perform secondary imaging on the vascular segmentation results obtained in step S5 on actual OCTA eye images to obtain OCTA eye images with vascular segmentation results.

[0110] The imaging method provided by the present invention can be directly applied to existing eye image devices (such as an eye image imaging system) or directly applied to a terminal (such as a computer). When specifically applied, an actual eye image is obtained using an existing solution, and then the acquired data is input into the corresponding machine device or terminal. At this time, the machine device or terminal can, according to the imaging method disclosed by the present invention, obtain the vascular segmentation result of the actual eye image, and display the vascular segmentation result on the original image through different types of representations (such as colors), and then perform secondary imaging and output. At this time, the output image is an OCTA eye image with the vascular segmentation result, and this eye image can reflect the actual eye image and the corresponding vascular segmentation result, thus greatly facilitating the subsequent work of clinical medical staff and laboratory experimenters.

Claims

1. A semi-supervised OCTA blood vessel segmentation method, comprising the following steps: S1. Obtain existing OCTA eye images; S2. For the OCTA eye images obtained in step S1, select a number of images for blood vessel annotation, do not annotate the remaining images, and perform random data enhancement on the images, thereby constructing a training data set; S3. Based on the convolution module and attention mechanism, a semi-supervised OCTA vascular segmentation primary model is constructed; S4. Using the training data set obtained in step S2, the semi-supervised OCTA vascular segmentation primary model constructed in step S3 is trained to obtain a semi-supervised OCTA vascular segmentation model; S5. Use the semi-supervised OCTA blood vessel segmentation model obtained in step S4 to perform actual blood vessel segmentation of the OCTA eye image.

2. The semi-supervised OCTA blood vessel segmentation method according to claim 1, characterized in that Step S2, for the OCTA eye images obtained in step S1, selects a number of images for blood vessel annotation, does not annotate the remaining images, and performs random data enhancement on the images, thereby constructing a training data set, specifically comprising the following steps: Among the OCTA eye images acquired in step S1, blood vessels are annotated for some OCTA eye images according to a set ratio and used as labeled data images; blood vessels are not annotated for the remaining OCTA eye images and used as unlabeled data images; all the labeled data images and unlabeled data images together constitute a first data set; In the first data set, a number of images are randomly extracted, and it is ensured that the extracted images include both labeled data images and unlabeled data images; data enhancement is performed on the extracted images to obtain a training data set; the training data set includes the original labeled data images, the original unlabeled data images, the enhanced labeled data images, and the enhanced unlabeled data images; During training, ensure that the images input into the model in each round of training include the original labeled data images, the original unlabeled data images, the enhanced labeled data images, and the enhanced unlabeled data images.

3. The semi-supervised OCTA blood vessel segmentation method according to claim 2, characterized in that The semi-supervised OCTA vascular segmentation primary model is constructed based on the convolution module and the attention mechanism described in step S3, which specifically includes the following steps: The constructed semi-supervised OCTA vascular segmentation primary model includes an encoding module, a latent vector module, and a decoding module; The encoding module is constructed based on the convolution module; the encoding module is used to extract high-dimensional features from the input image, gradually compress the spatial dimensions and extract semantic information, and provide higher-level semantic information for subsequent processing modules; The latent vector module is constructed based on snake convolution and channel attention mechanism. The latent vector module captures the local structural details in the image through snake convolution, and combines the channel attention mechanism to enhance the expression ability of key features, thereby generating a more accurate latent representation to support image feature learning. Construct a decoding module based on the deconvolution module; The decoding module is used to map the obtained features back to the original space, and gradually restore the spatial resolution of the image through deconvolution, and finally reconstruct the image segmentation result to ensure the accuracy of the output and retain the image details.

4. The semi-supervised OCTA blood vessel segmentation method according to claim 3, characterized in that The construction of the encoding module based on the convolution module specifically includes the following steps: The encoding module includes a first convolution layer, a first pooling layer, a second convolution layer, a second pooling layer, a third convolution layer, a third pooling layer, a fourth convolution layer, a fourth pooling layer and a fifth convolution layer; The first convolutional layer, the first pooling layer, the second convolutional layer, the second pooling layer, the third convolutional layer, the third pooling layer, the fourth convolutional layer, the fourth pooling layer and the fifth convolutional layer are connected in series in sequence; The first convolution layer includes a basic convolution submodule; the second convolution layer, the third convolution layer, the fourth convolution layer and the fifth convolution layer each include two basic convolution submodules connected in series; the basic convolution submodule includes a 3×3 convolution layer, a batch normalization layer and a RELU activation function layer connected in series; The first pooling layer, the second pooling layer, the third pooling layer and the fourth pooling layer are all 2×2 pooling layers.

5. The semi-supervised OCTA blood vessel segmentation method according to claim 4, characterized in that The potential vector module based on snake convolution and channel attention mechanism is constructed, which specifically includes the following steps: The latent vector module includes a snake convolution module and a channel attention module. The input data of the latent vector module is divided into two paths: the first path is input to the snake convolution module for processing, and the second path is input to the channel attention module for processing. The output of the snake convolution module and the output of the channel attention module are bitwise added to obtain the output of the latent vector module. The snake convolution module includes a coordinate mapping submodule and a deformation convolution feature calculation submodule connected in series; The coordinate mapping submodule is implemented based on stack decompression. The coordinate mapping submodule dynamically adjusts the coordinate mapping of the convolution kernel to offset the positions of the convolution kernel along the x-axis and y-axis to adapt to different image structures. The coordinate mapping submodule ensures that the offset of the convolution kernel can capture the local geometric shape by gradually selecting the position, thereby improving the convolution kernel's ability to perceive information; the deformation convolution feature calculation submodule is implemented based on bilinear interpolation. The deformation convolution feature calculation submodule calculates the features of the offset convolution kernel position through the bilinear interpolation method to ensure that the convolution kernel can extract features; The channel attention module dynamically adjusts the channel weights by calculating the similarity between channels, thereby achieving differential responses to different channels.

6. The semi-supervised OCTA blood vessel segmentation method according to claim 5, characterized in that The decoding module is constructed based on the deconvolution module, and specifically comprises the following steps: The decoding module includes a main decoder and an auxiliary decoder; If the input of the encoding module is the enhanced labeled data image or the enhanced unlabeled data image, the output of the latent vector module is used as the input of the main decoder; if the input of the encoding module is the original labeled data image or the original unlabeled data image, the output of the latent vector module is used as the input of the auxiliary decoder; The main decoder and the auxiliary decoder have the same structure, and both include a first upsampling submodule, a second upsampling submodule, a third upsampling submodule, a fourth upsampling submodule and an inference submodule connected in series in sequence; The first upsampling submodule, the second upsampling submodule, the third upsampling submodule and the fourth upsampling submodule have the same structure, and all include a 2×2 deconvolution layer, a first 3×3 convolution layer, a first batch of normalization layers, a first RELU activation function layer, a second 3×3 convolution layer, a second batch of normalization layers and a second RELU activation function layer connected in series in sequence; The inference submodule consists of a 1×1 convolutional layer.

7. The semi-supervised OCTA blood vessel segmentation method according to claim 6, characterized in that The training described in step S4 specifically includes the following steps: If the input of the encoding module is an enhanced labeled data image, the segmentation loss L is additionally calculated using the following formula: Seg : L Seg =θ×Dice(f m (A(x)),A(y))+(1-θ)×D E (T(f m (A(x))),T(A(y))) Where θ is the training round parameter, and k is the training round, τ is the temperature coefficient; Dice() is the Dice loss function; A() represents image enhancement; x is the original image; y is the corresponding label; f m () is the processing function of the main decoder; D E () is the Euclidean distance calculation function; T() is the triangular feature extraction function; For all input data images, the topological skeleton loss L is calculated using the following formula: TS : L TS =Dice(f m (A(x)),A(f a (x)))+Dice(S(f m (A(x))),S(A(f a (x)))) Where f a () is the processing function of the auxiliary decoder; S() is the topological skeleton extraction function; The following formula is used as the total loss function L: L=L S +ω×L U Where ω is the ramp-up coefficient; L S is the total loss function of the first stage; L U is the total loss function of the second stage; for the labeled data processing stage, the total loss function of the first stage L S The calculation formula is L S =L Seg +L TS ; For the unlabeled data processing stage, the total loss function L in the second stage U The calculation formula is L U =L TS .

8. The semi-supervised OCTA blood vessel segmentation method according to claim 7, characterized in that The semi-supervised OCTA blood vessel segmentation model obtained in step S4 is used in step S5 to perform actual blood vessel segmentation of the OCTA eye image, which specifically includes the following steps: Acquire the actual OCTA eye image and input it into the semi-supervised OCTA blood vessel segmentation model obtained in step S4; The output of the auxiliary decoder is used as the actual blood vessel segmentation result of the OCTA eye image.

9. An imaging method comprising the semi-supervised OCTA blood vessel segmentation method according to any one of claims 1 to 8, characterized in that The following steps are also included: S6. The blood vessel segmentation result obtained in step S5 is annotated and re-imaged on the actual OCTA eye image to obtain an OCTA eye image with the blood vessel segmentation result.

Images