OCTA image blood vessel and FAZ region segmentation method and system

Abstract

The embodiment of the invention provides an OCTA image blood vessel and FAZ region segmentation method and system, and belongs to the technical field of image processing. The method comprises the following steps: collecting OCTA image information, inputting the OCTA image information into a convolution module, and performing recursive convolution for preset times based on the convolution module to obtain a feature image; performing feature segmentation on the feature image based on a feature extraction network to obtain blood vessel features and FAZ region features; performing feature extraction on the blood vessel features and the FAZ region features based on a pre-constructed transformer module, and outputting blood vessel extraction features and FAZ region extraction features; capturing relation features between the FAZ region extraction features and blood vessel extraction features based on a pre-constructed attention module; and outputting a blood vessel segmentation result based on the blood vessel extraction features, and outputting an FAZ region segmentation result based on the FAZ region extraction features and the relation features. According to the scheme, the problem that blood vessel and FAZ region segmentation is difficult in an OCTA image in an existing scheme is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to an OCTA image blood vessel and FAZ region segmentation method and an OCTA image blood vessel and FAZ region segmentation system. Background Art

[0002] Optical coherence tomography angiography (OCTA) is a revolutionary non-invasive retinal imaging technology that has gradually become an important tool in clinical and research fields. OCTA uses dense scanning and high-frequency data capture technology to generate three-dimensional retinal volume data, providing micron-level resolution of the retinal vascular system. This technology can effectively visualize the subtle vascular structure of the retina by mapping data from different retinal layers, thereby deepening the understanding of retinal vascular diseases.

[0003] With the advancement of the latest research, OCTA scans have been proven to accurately reveal pathological changes associated with retinal diseases, such as neovascularization of retinal vein occlusion (RVO), retinal ischemia, and abnormalities in the avascular zone (FAZ). Therefore, OCTA images have begun to be widely used as biomarkers to support early clinical assessment of retinopathy. In addition, OCTA imaging provides an intuitive basis for assessing the health of the retina by showing changes in blood flow in the retinal vascular system.

[0004] Although OCTA provides a powerful tool for diagnosing retinal diseases, there are still technical challenges in simultaneously segmenting the retinal vessels (RV) and FAZ. The vascular network structure in the retina is complex and changeable, coupled with the uneven image quality and the unclear boundary between the non-perfused area and the FAZ area, all of which increase the difficulty of segmentation. Based on this, the existing solutions are difficult to segment the blood vessels and FAZ areas in OCTA images and need to be solved urgently. Summary of the invention

[0005] The purpose of the embodiments of the present invention is to provide a method and system for segmenting blood vessels and FAZ regions in OCTA images, so as to at least solve the problem that it is difficult to segment blood vessels and FAZ regions in OCTA images in existing solutions.

[0006] In order to achieve the above-mentioned purpose, the first aspect of the present invention provides a method for segmenting blood vessels and FAZ areas in OCTA images, the method comprising: collecting OCTA image information, and inputting the OCTA image information into a convolution module to obtain a feature image after performing a preset number of recursive convolutions based on the convolution module; performing feature segmentation on the feature image based on a feature extraction network to obtain blood vessel features and FAZ area features; performing feature extraction on the blood vessel features and the FAZ area features respectively based on a pre-built transformer module, and outputting blood vessel extraction features and FAZ area extraction features; capturing the relationship features between the FAZ area extraction features and the blood vessel extraction features based on a pre-built attention module; outputting a blood vessel segmentation result based on the blood vessel extraction features, and outputting a FAZ area segmentation result based on the FAZ area extraction features and the relationship features.

[0007] Optionally, the convolution module includes 32 convolution channels; and the convolution of the convolution module is 3×3 convolution.

[0008] Optionally, the feature extraction network includes multiple network modules, each network module includes three network layers; wherein the first network layer includes 7x7 longitudinally separable convolutions, each convolution is composed of two 3x3 repeated convolutions; the second network layer includes two 3x3 convolution layers; the third network layer includes two 3x3 convolution layers.

[0009] Optionally, the pre-built transformer module performs feature extraction on the vascular features and the FAZ region features respectively, and outputs the vascular extraction features and the FAZ region extraction features, including: based on three generators in the pre-built transformer module, converting the input vascular features and the FAZ region features into corresponding query vectors, key vectors and value vectors respectively; wherein the three generators are generator Q, generator Q, and generator Q. and generator A pairwise function is constructed based on the corresponding query vector and key vector, expressed as:

[0010]

[0011] Among them, F is the input blood vessel feature or the FAZ area feature; Q(F) is the query vector; Q(F) is the key vector; T is the matrix transposition operator; feature combination is performed based on the pairwise function and the value vector to obtain the attention feature; the attention feature is used as the residual item of the input blood vessel feature and the FAZ area feature, and the blood vessel extraction feature and the FAZ area extraction feature are output based on the residual connection.

[0012] Optionally, the rule for performing feature combination based on the paired function and the value vector to obtain attention features is:

[0013]

[0014] in, To note the features; is a value vector; softmax is used for normalization The softmax function of C′ is the number of channels; the residual items of the blood vessel features and the FAZ area features that take the attention features as input, and the rules for outputting blood vessel extraction features and FAZ area extraction features based on the residual connection are:

[0015]

[0016] Among them, F i Extract features for blood vessels or FAZ regions; i∈{l,v} represents the corresponding FAZ regions and blood vessels respectively; W is linear embedding.

[0017] Optionally, the attention module includes a self-attention head and a cross-attention head; the self-attention head is used to capture the intra-class dependency between the FAZ area and the vascular extraction features; the cross-attention head is used to capture the inter-class relationship between the FAZ area and the vascular extraction features.

[0018] Optionally, the pre-built attention module captures the relationship features between the FAZ region extraction features and the blood vessel extraction features, including: constructing a pair function of the self-attention head and a pair function of the cross-attention head for FAZ region extraction features based on a self-attention head and a cross-attention head, respectively, expressed as:

[0019]

[0020] Among them, s is the self-attention head, c is the cross-attention head; Q S (F l ), Q c (F l ) is a query vector for extracting features from the FAZ area; Key vector for extracting features for the FAZ area F is the key vector for extracting blood vessel features; l Extract features for the FAZ area; F v Extract features for blood vessels; perform feature combination based on the pairwise function, the value vector of the FAZ region extracted features and the value vector of the blood vessel extracted features, and obtain the self-attention head attention features and the cross-attention head attention features respectively; expressed as:

[0021]

[0022] in, Extract the value vector of features for the FAZ area; Extract the value vector of the blood vessel features; softmax is used for normalization The softmax function of ; the output features are obtained based on the attention features of each attention head and the attention features of the cross attention head, expressed as:

[0023]

[0024] Among them, F i is the output feature; W i is a linear embedding implemented as a 1x1 convolution; The representation is performed by residual connection with element-wise addition; the relational features are output based on the output features of the self-attention head and the output features of the cross-attention head; it is expressed as:

[0025] F out =[F S ; F C ];

[0026] Among them, [;] indicates the connection on the channel size.

[0027] Optionally, outputting a blood vessel segmentation result based on the blood vessel extraction feature, and outputting a FAZ region segmentation result based on the FAZ region extraction feature and the relationship feature, includes: constructing a total network loss based on a blood vessel segmentation loss function and a FAZ region segmentation loss function, expressed as:

[0028]

[0029] in, is the total network loss; is the blood vessel segmentation loss function; is the FAZ area segmentation loss function; λ is the preset first weight; is the intersection-over-union ratio between the predicted area of ​​the blood vessel and FAZ region and the area of ​​the true label; β is the preset second weight; based on the total loss of the network, the blood vessel segmentation result and the FAZ region segmentation result are output.

[0030] The second aspect of the present invention provides an OCTA image blood vessel and FAZ area segmentation system, the system comprising: an input module, used to collect OCTA image information, and input the OCTA image information into a convolution module, so as to obtain a feature image after performing a preset number of recursive convolutions based on the convolution module; a feature segmentation module, used to perform feature segmentation on the feature image based on a feature extraction network, and obtain blood vessel features and FAZ area features; a feature extraction module, used to perform feature extraction on the blood vessel features and the FAZ area features respectively based on a pre-built transformer module, and output blood vessel extraction features and FAZ area extraction features; a relationship feature module, used to capture the relationship features between the FAZ area extraction features and the blood vessel extraction features based on a pre-built attention module; an output module, used to output a blood vessel segmentation result based on the blood vessel extraction features, and output a FAZ area segmentation result based on the FAZ area extraction features and the relationship features.

[0031] On the other hand, the present invention provides a computer-readable storage medium having instructions stored thereon, which, when executed on a computer, enables the computer to execute the above-mentioned OCTA image blood vessel and FAZ region segmentation method.

[0032] Through the above technical scheme, the scheme of the present invention uses a convolution module to perform recursive convolution processing on the OCTA image, which can accurately extract useful feature images from complex retinal images. This method can deeply mine image data and provide high-quality input for subsequent feature analysis. The obtained feature image is processed by the feature extraction network to distinguish and extract key retinal vascular features and FAZ area features. This step is the key to diagnosing and treating retinal diseases because it is directly related to disease identification and classification. The application of the pre-built transformer module to further extract vascular features and FAZ area features enhances the model's ability to understand details and can more accurately identify subtle lesions. Using the attention module to explore the intrinsic relationship between the FAZ area extraction features and the vascular extraction features helps to understand the interaction between different retinal tissues, which is particularly important for understanding the overall impact of the disease. Finally, based on the vascular extraction features and the FAZ area extraction features and their relationship features, high-precision vascular and FAZ area segmentation results are output. The scheme of the present invention can automatically obtain objective and quantitative evaluation results, which is also more accurate and reliable, and solves the existing OCTA image vascular FAZ segmentation accuracy and low efficiency problems.

[0033] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The accompanying drawings are used to provide a further understanding of the embodiments of the present invention and constitute a part of the specification. Together with the following specific embodiments, they are used to explain the embodiments of the present invention, but do not constitute a limitation on the embodiments of the present invention. In the accompanying drawings:

[0035] Figure 1 It is a flowchart of the steps of a method for segmenting blood vessels and FAZ regions in OCTA images provided by one embodiment of the present invention;

[0036] Figure 2 is a schematic diagram of a neural network structure provided by one embodiment of the present invention;

[0037] Figure 3 is a schematic diagram of a first segmentation example provided by an embodiment of the present invention;

[0038] Figure 4 is a schematic diagram of a second segmentation example provided by an embodiment of the present invention;

[0039] Figure 5 It is a system structure diagram of an OCTA image blood vessel and FAZ region segmentation system provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0040] The specific implementation of the present invention is described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation described here is only used to illustrate and explain the present invention, and is not used to limit the present invention.

[0041] Figure 1 FIG. 1 is a flow chart of a method for segmenting blood vessels and FAZ regions in OCTA images provided by an embodiment of the present invention. Figure 1 As shown, an embodiment of the present invention provides a method for segmenting blood vessels and FAZ regions in OCTA images, the method comprising:

[0042] Step S10: Acquire OCTA image information, and input the OCTA image information into a convolution module, so as to obtain a feature image after performing a preset number of recursive convolutions based on the convolution module.

[0043] Specifically, the convolution module includes 32 convolution channels; the convolution of the convolution module is 3×3 convolution.

[0044] In an embodiment of the present invention, the scheme of the present invention sets the number of convolution channels of FRNet-base to 32, aiming to enhance the feature processing capability of the network by increasing the number of channels. The use of 3x3 convolution kernels instead of traditional 1x1 convolution kernels is to expand the receptive field of the model, enabling it to capture a wider range of contextual information, thereby improving the understanding of image details. In addition, by introducing the concept of recursive convolution in this module, that is, letting the data circulate R times through the same convolution layer, the processing depth of the network on the input data can be further deepened.

[0045] Based on the solution of the present invention, increasing the number of convolution channels can give the network more parameters to learn complex features in the data, which is particularly important for processing highly complex visual tasks. Secondly, using a 3x3 convolution kernel instead of a 1x1 convolution kernel can effectively expand the receptive field and improve the model's ability to capture local image structures without significantly increasing the computational burden. This is very helpful for the accurate recognition and positioning of objects in images. Finally, the application of recursive convolution makes the feature extraction process more in-depth. By repeatedly using the same convolution layer, the network can learn deeper feature information from repeated data processing. This method improves the efficiency and accuracy of feature extraction, allowing the network to perform higher performance when processing complex images or sequence tasks.

[0046] Step S20: performing feature segmentation on the feature image based on a feature extraction network to obtain blood vessel features and FAZ area features.

[0047] Specifically, the feature extraction network includes multiple network modules, each of which includes three network layers; wherein the first network layer includes 7x7 longitudinally separable convolutions, each convolution consists of two 3x3 repeated convolutions; the second network layer includes two 3x3 convolution layers; and the third network layer includes two 3x3 convolution layers.

[0048] In an embodiment of the present invention, the first network layer adopts a 7x7 longitudinal separable convolution. This convolution kernel is larger and can cover a wider input area, providing richer contextual information. Each 7x7 convolution is actually composed of two 3x3 convolutions repeated. This structural design not only expands the receptive field, but also strengthens the network's ability to capture the details of the input data, because multiple filtering operations can enhance the model's learning and abstract capabilities for features. The second and third network layers each include two 3x3 convolution layers, which refine the analysis of features through further local information processing. The 3x3 convolution kernel is a convolution kernel size widely used in deep learning. It can effectively capture important features in the image, such as edges, corners, etc., while maintaining computational efficiency. Using two convolution layers in these two layers can deepen the hierarchy of features and enhance the nonlinear processing capabilities of the network, thereby improving the model's ability to recognize complex patterns.

[0049] Based on the scheme of the present invention, by using larger convolution kernels and repeated convolution operations, the receptive field of the network is significantly expanded, allowing the model to observe more contextual information, thereby improving the accuracy of feature segmentation. The multi-layer network structure makes the conversion from the original image to the feature map more detailed and hierarchical, which helps to extract more accurate blood vessel and FAZ area features. Although the number of network layers increases, the computational burden is controlled through reasonable convolution kernel design (such as the use of separable convolution and small-sized convolution kernels), ensuring efficient data processing. The multi-layer convolution structure enhances the network's adaptability to blood vessel features of different sizes and types, making the model more robust and reliable when processing various blood vessel and FAZ area features.

[0050] Step S30: performing feature extraction on the blood vessel features and the FAZ region features respectively based on the pre-built transformer module, and outputting blood vessel extracted features and FAZ region extracted features.

[0051] Specifically, based on the three generators in the pre-built transformer module, the input vascular features and the FAZ area features are respectively converted into corresponding query vectors, key vectors and value vectors; wherein the three generators are respectively generators Generator and generator A pairwise function is constructed based on the corresponding query vector and key vector, expressed as:

[0052]

[0053] Wherein, F is the input blood vessel feature or the FAZ area feature; is the query vector; is a key vector; T is a matrix transposition operator; feature combination is performed based on the pairwise function and the value vector to obtain an attention feature; the attention feature is used as the residual item of the input vascular feature and the FAZ area feature, and a vascular extraction feature and a FAZ area extraction feature are output based on the residual connection.

[0054] In the embodiment of the present invention, a feature generated from the backbone is input The output is a refined feature Detect blood vessels and FAZ areas separately. Use three generators and The input features F are converted into queries, keys, and values ​​respectively. In this module, the generator Implemented with 3x3 convolution and global average pooling, and outputs the query vector The number of channels is designed to be C' = C / 8. and With The same architecture as expected, with the reshape operation replacing the global average pooling, resulting in keys and values

[0055] Furthermore, the global transformer module can be roughly described as an attention mechanism, which first combines the channel weighted features with the input information, and then combines the spatial weighted features. The rule for performing feature combination based on the pairwise function and the value vector to obtain the attention feature is:

[0056]

[0057] in, To note the features;

[0058] is a value vector;

[0059] Softmax is used for normalization The softmax function;

[0060] C′ is the number of channels;

[0061] The rule for outputting the vessel extraction feature and the FAZ region extraction feature based on the residual connection of the attention feature as the input of the vessel feature and the FAZ region feature is:

[0062]

[0063] Among them, the + operation represents the sum operation of broadcast elements, F i Extract features for blood vessels or FAZ regions; i∈{l,v} represents the corresponding FAZ regions and blood vessels respectively; W is linear embedding.

[0064] Step S40: capturing the relationship features between the FAZ region extraction features and the blood vessel extraction features based on a pre-built attention module.

[0065] Specifically, the paired function of the self-attention head and the paired function of the cross-attention head for extracting features from the FAZ region are constructed based on the self-attention head and the cross-attention head, respectively, and are expressed as:

[0066]

[0067] Among them, s is the self-attention head and c is the cross-attention head; A query vector for extracting features for the FAZ area; Key vector for extracting features for the FAZ area F is the key vector for extracting blood vessel features; l Extract features for the FAZ area; Fv Extract features for blood vessels; perform feature combination based on the pairwise function, the value vector of the FAZ region extracted features and the value vector of the blood vessel extracted features, and obtain the self-attention head attention features and the cross-attention head attention features respectively; expressed as:

[0068]

[0069] in, Extract the value vector of features for the FAZ area; Extract the value vector of the blood vessel features; softmax is used for normalization The softmax function of ; the output features are obtained based on the attention features of each attention head and the attention features of the cross attention head, expressed as:

[0070]

[0071] Among them, F i is the output feature; W i is a linear embedding implemented as a 1x1 convolution; The representation is performed by residual connection with element-wise addition; the relational features are output based on the output features of the self-attention head and the output features of the cross-attention head; it is expressed as:

[0072] F out =[F S ; F C ];

[0073] Among them, [;] indicates the connection on the channel size.

[0074] In one possible implementation, in each head, three trainable linear embeddings are used as query, key, and value generators, respectively. This is implemented using 3x3 convolution and reshape operations. Computing query and key values ​​pairwise on self-attention heads and cross-attention heads can be expressed as:

[0075]

[0076] Subscripts s and c denote self-attention and cross-attention heads, respectively. Different from the self-attention head that obtains the query from the input feature F, the cross-attention head generates keywords from the vessel feature F instead of integrating vessel information. The individual attention features of the two heads are calculated as:

[0077]

[0078] Study each header and get the output:

[0079]

[0080] Where W i is a linear embedding implemented as a 1x1 convolution, and the operation is performed via a residual connection with element-wise vector addition. The representation is performed by residual concatenation with element-wise addition.

[0081] The self-attention head computes the response at a position as a weighted sum of features at all positions, thus capturing long-range dependencies well. Given that blood vessels are usually distributed over a wide area, self-attention can exchange information between multiple positions regardless of their positional distance, allowing for modeling the duality of intra-class lesions. The head should be able to distinguish mixtures of more than two lesions and further refine the edges of large patterns for vessel segmentation.

[0082] The cross-attention head queries the global vascular structure from the vascular features, thereby incorporating the interaction between the FAZ and the vessels. Considering that the FAZ region has a strong positional connection with the veins, the cross-attention helps to better eliminate false positives.

[0083] The obtained feature F s , concatenated from the self-attention head and the cross-attention head, to get the final feature output:

[0084] F out =[F S ; F C ]

[0085] where [·;·] denotes connectivity on the channel dimensions.

[0086] Step S50: outputting a blood vessel segmentation result based on the blood vessel extraction feature, and outputting a FAZ region segmentation result based on the FAZ region extraction feature and the relationship feature.

[0087] Specifically, the total network loss is constructed based on the blood vessel segmentation loss function and the FAZ area segmentation loss function, which is expressed as:

[0088]

[0089] in, is the total network loss; is the blood vessel segmentation loss function; is the FAZ area segmentation loss function; λ is the preset first weight; is the intersection-over-union ratio between the predicted area of ​​the blood vessel and FAZ region and the area of ​​the true label; β is the preset second weight; based on the total loss of the network, the blood vessel segmentation result and the FAZ region segmentation result are output.

[0090] Furthermore, is the intersection-over-union ratio between the pre-test and true label areas of the blood vessel and FAZ regions, and β is the weight corresponding to this parameter. It is expressed as:

[0091]

[0092] In the above formula, X represents the predicted area and Y represents the true label area.

[0093] like Figure 2 , is the neural network structure constructed by the scheme of the present invention for OCTA image segmentation. By upgrading the adaptability of the network structure, the accuracy of FAZ area and blood vessel segmentation is guaranteed.

[0094] like Figure 3 and Figure 4 , are two effect diagrams of embodiments of FAZ region and blood vessel segmentation based on the solution of the present invention. As shown in the figures, the solution of the present invention can achieve accurate segmentation of FAZ region and blood vessels in a single OCTA image at the same time.

[0095] Figure 5 FIG. 1 is a system structure diagram of an OCTA image blood vessel and FAZ region segmentation system provided by an embodiment of the present invention. Figure 5 As shown, an embodiment of the present invention provides an OCTA image blood vessel and FAZ area segmentation system, the system comprising: an input module, used to collect OCTA image information, and input the OCTA image information into a convolution module, so as to obtain a feature image after performing a preset number of recursive convolutions based on the convolution module; a feature segmentation module, used to perform feature segmentation on the feature image based on a feature extraction network, and obtain blood vessel features and FAZ area features; a feature extraction module, used to perform feature extraction on the blood vessel features and the FAZ area features respectively based on a pre-built transformer module, and output blood vessel extraction features and FAZ area extraction features; a relationship feature module, used to capture the relationship features between the FAZ area extraction features and the blood vessel extraction features based on a pre-built attention module; an output module, used to output a blood vessel segmentation result based on the blood vessel extraction features, and output a FAZ area segmentation result based on the FAZ area extraction features and the relationship features.

[0096] The embodiment of the present invention further provides a computer-readable storage medium, on which instructions are stored, which, when executed on a computer, enable the computer to execute the above-mentioned OCTA image blood vessel and FAZ region segmentation method.

[0097] Those skilled in the art will understand that all or part of the steps in the method for implementing the above-mentioned embodiments can be completed by instructing the relevant hardware through a program, and the program is stored in a storage medium, including several instructions for making a single-chip microcomputer, a chip or a processor (processor) perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0098] The optional embodiments of the present invention are described in detail above in conjunction with the accompanying drawings. However, the embodiments of the present invention are not limited to the specific details in the above embodiments. Within the technical concept of the embodiments of the present invention, the technical scheme of the embodiments of the present invention can be subjected to a variety of simple modifications, and these simple modifications all belong to the protection scope of the embodiments of the present invention. It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, the embodiments of the present invention will not further describe various possible combinations.

[0099] In addition, various embodiments of the present invention may be arbitrarily combined, and as long as they do not violate the concept of the embodiments of the present invention, they should also be regarded as the contents disclosed in the embodiments of the present invention.

Claims

1. A method for segmenting blood vessels and FAZ regions in OCTA images, characterized in that: The method comprises: Collecting OCTA image information, and inputting the OCTA image information into a convolution module, so as to obtain a feature image after performing a preset number of recursive convolutions based on the convolution module; Perform feature segmentation on the feature image based on a feature extraction network to obtain blood vessel features and FAZ area features; Based on the pre-built transformer module, feature extraction is performed on the blood vessel feature and the FAZ region feature respectively, and the blood vessel extraction feature and the FAZ region extraction feature are output; Capturing the relationship features between the FAZ region extraction features and the blood vessel extraction features based on a pre-built attention module; A blood vessel segmentation result is output based on the blood vessel extraction feature, and a FAZ region segmentation result is output based on the FAZ region extraction feature and the relationship feature.

2. The method according to claim 1, characterized in that The convolution module includes 32 convolution channels; The convolution of the convolution module is 3×3 convolution.

3. The method according to claim 1, characterized in that The feature extraction network includes multiple network modules, each of which includes three network layers; The first network layer consists of 7x7 vertically separable convolutions, each of which consists of two 3x3 repeated convolutions; The second network layer includes two 3x3 convolutional layers; The third network layer consists of two 3x3 convolutional layers.

4. The method according to claim 1, characterized in that The pre-built transformer module performs feature extraction on the blood vessel feature and the FAZ region feature respectively, and outputs the blood vessel extraction feature and the FAZ region extraction feature, including: Based on the three generators in the pre-built transformer module, the input vascular features and the FAZ area features are respectively converted into corresponding query vectors, key vectors and value vectors; wherein, The three generators are respectively Generator and generator A pairwise function is constructed based on the corresponding query vector and key vector, expressed as: Wherein, F is the input blood vessel feature or the FAZ area feature; is the query vector; is the key vector; T is the matrix transpose operator; Perform feature combination based on the pairwise function and the value vector to obtain an attention feature; The attention feature is used as the residual item of the blood vessel feature and the FAZ area feature which are input, and a blood vessel extraction feature and a FAZ area extraction feature are output based on the residual connection.

5. The method according to claim 4, characterized in that The rule for performing feature combination based on the pairwise function and the value vector to obtain attention features is: in, To note the features; is a value vector; Softmax is used for normalization The softmax function; C′ is the number of channels; The rule for outputting the vessel extraction feature and the FAZ region extraction feature based on the residual connection of the attention feature as the input of the vessel feature and the FAZ region feature is: Among them, F i Extract features for blood vessels or FAZ areas; i∈{l,v} indicates the corresponding FAZ area and blood vessel, respectively; W is a linear embedding.

6. The method according to claim 1, characterized in that The attention module includes a self-attention head and a cross-attention head; The self-attention head is used to capture the intra-class dependency between the FAZ region and the vascular extraction features; The cross-attention head is used to capture the inter-class relationship between the FAZ region and the vessel extracted features.

7. The method according to claim 6, characterized in that The pre-built attention module captures the relationship features between the FAZ region extraction features and the blood vessel extraction features, including: The paired functions of the self-attention head and the cross-attention head for extracting features from the FAZ region are constructed based on the self-attention head and the cross-attention head, respectively, and are expressed as: Among them, s is the self-attention head and c is the cross-attention head; A query vector for extracting features for the FAZ area; Key vector for extracting features for the FAZ area A key vector for extracting features for blood vessels; F l Extract features for the FAZ area; F v Extract features for blood vessels; Based on the pair function, the value vector of the FAZ region extraction feature and the value vector of the blood vessel extraction feature, feature combination is performed to obtain the self-attention head attention feature and the cross-attention head attention feature respectively; expressed as: in, Extract the value vector of features for the FAZ area; A value vector for extracting features of blood vessels; Softmax is used for normalization Softmax function; The output features are obtained based on the attention features of each attention head and the attention features of the cross attention head, respectively, and are expressed as: Among them, F i is the output feature; W i is a linear embedding implemented as a 1x1 convolution; The representation is performed by residual concatenation with element-wise addition; The output features of the self-attention head attention features and the output features of the cross-attention head output relation features; expressed as: F out =[F S ;F C ]; Among them, [;] indicates the connection on the channel size.

8. The method according to claim 1, characterized in that The outputting of the blood vessel segmentation result based on the blood vessel extraction feature, and the outputting of the FAZ region segmentation result based on the FAZ region extraction feature and the relationship feature, include: The total network loss is constructed based on the blood vessel segmentation loss function and the FAZ area segmentation loss function, which is expressed as: in, is the total network loss; is the blood vessel segmentation loss function; is the FAZ area segmentation loss function; λ is the preset first weight; It is the intersection-over-union ratio between the predicted area and the true label area of ​​the blood vessel and FAZ area; β is the preset second weight; Based on the total loss of the network, the blood vessel segmentation results and FAZ area segmentation results are output.

9. An OCTA image blood vessel and FAZ region segmentation system, characterized in that: The system comprises: An input module, used for collecting OCTA image information, and inputting the OCTA image information into a convolution module, so as to obtain a feature image after performing a preset number of recursive convolutions based on the convolution module; A feature segmentation module, used to perform feature segmentation on the feature image based on a feature extraction network to obtain blood vessel features and FAZ area features; A feature extraction module, used to perform feature extraction on the blood vessel feature and the FAZ region feature respectively based on a pre-built transformer module, and output blood vessel extraction features and FAZ region extraction features; A relationship feature module, used for capturing the relationship features between the FAZ region extraction features and the blood vessel extraction features based on a pre-built attention module; An output module is used to output a blood vessel segmentation result based on the blood vessel extraction feature, and to output a FAZ region segmentation result based on the FAZ region extraction feature and the relationship feature.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the OCTA image blood vessel and FAZ region segmentation method according to any one of claims 1 to 8.

Images