Retinal blood vessel segmentation method based on 3D and 2D combined OCTA image

Abstract

The application discloses an OCTA image retinal blood vessel segmentation method based on 3D and 2D combination, which comprises the following steps: carrying out cutting and cutting block operation on input 3D data, and carrying out pixel value conversion; inputting the pretreated 3D data into a 3D adaptive projection learning network to generate a projection map of the 3D data; inputting the projection map and a 2D image into a 2D blood vessel segmentation network to obtain a prediction map of the retinal blood vessel; generating a blood vessel segment according to the prediction map and a corresponding blood vessel label map, calculating a thickness inconsistency weight of a blood vessel segment pixel and a blood vessel thickness inconsistency loss; reserving network parameters with minimum total loss as final prediction network based on Tversky loss of the prediction map and the blood vessel label map; and segmenting retinal blood vessels of an OCTA image to be processed by using the final prediction network. The application can provide more objective and accurate evaluation means for diagnosis and research of diseases related to the retinal blood vessel, and can quickly and accurately segment the retinal blood vessel.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, and particularly relates to an OCTA image retinal blood vessel segmentation method based on 3D and 2D combination. BACKGROUND

[0002] Optical coherence tomography angiography (OCTA) is an extension of optical coherence tomography (OCT). OCTA provides images with details of retinal and choroidal vascular structures by extracting changes in OCT signals caused by red blood cell movement, overcoming the shortcomings of OCT that cannot provide good blood vessel details. OCTA imaging can distinguish different levels of vascular abnormalities, including retinal superficial, deep, and choroidal circulation abnormalities. The three-dimensional blood flow data generated by OCTA imaging needs to be compressed into multiple two-dimensional images (en face images) according to the physiological structure layer to obtain an ideal evaluation of the lesion. OCTA is a non-invasive imaging method. Compared with traditional fundus fluorescein angiography (FFA) and indocyanine green angiography (ICGA), OCTA does not have the penetration and staining of fluorescein, and can better measure the boundaries and areas of capillaries and neovascularization. This is conducive to doctors accurately extracting retinal blood vessels, so OCTA is gradually commonly used for the diagnosis and evaluation of diseases related to retinal blood vessels.

[0003] OCTA can better provide details of capillaries and neovascularization, and can better capture the tiny blood vessels around the foveal area than color fundus images. At the same time, there are many noises in the OCTA image, the scale of the retinal changes is large, and potential retinal diseases may cause deformation of the retinal blood vessels. For example, high blood sugar in type 1 diabetes can damage the retinal blood vessel wall, leading to retinal hemorrhage and further damage to the retina; wet age-related macular degeneration (AMD) can cause increased retinal blood vessel permeability and increased choroidal neovascularization; retinal blood vessel obstruction can cause glaucoma. Therefore, ophthalmologists need to statistically analyze the morphological properties of retinal blood vessels and quantify them when diagnosing related retinal diseases, such as blood vessel tortuosity, branch angle, and blood vessel diameter. Manual segmentation of retinal blood vessels is too tedious and time-consuming, and too dependent on the skills and professional knowledge of related personnel. In order to solve this problem, many researchers have proposed some retinal blood vessel segmentation methods for OCTA images, which can be mainly divided into traditional methods and deep learning-based methods.

[0004] Traditional OCTA image retinal vessel segmentation methods primarily rely on designing various filters to capture vascular structures and then segment the retinal vessels. For example, Azzopardi et al. proposed a bar selection filter combined with thresholding to segment retinal vessels. Zhang et al. proposed a filter based on a 3D rotating frame with orientation fractions, finally obtaining a predicted image of the vessels through thresholding. In recent years, deep learning has achieved great success in computer vision (CV) tasks such as image segmentation, image classification, object detection, and recognition. Simultaneously, various deep learning-based methods have been applied to medical image processing. Pissas et al. proposed iUnet, which improves OCTA retinal vessel segmentation results by iteratively using the Unet network and providing intermediate predictions multiple times within the network. Ma et al. proposed a novel two-stage vessel segmentation network (OCTA-Net) that improves the segmentation of small vessels by adaptively refining the vessel prediction results in the coarse stage. Li et al. constructed a projection learning module (IPN) that aggregates 3D data into a 2D projection plane, thus mitigating the potential impact of inaccurate layer segmentation on retinal vessel segmentation to some extent.

[0005] OCTA images contain significant noise, and retinal vascular changes are large-scale. Underlying retinal diseases can cause retinal vascular deformation. The projection of OCTA 3D data into a 2D image depends on the results of layer segmentation, and underlying retinal diseases may disrupt the layer structure of the retina, affecting the accuracy of layer segmentation. Summary of the Invention

[0006] The technical problem solved by this invention is to overcome the shortcomings of segmentation on the projection map of OCTA, which is easily affected by potential retinal diseases and may lead to inaccurate segmentation results. This invention provides a retinal vessel segmentation method for OCTA images that combines 3D and 2D based on deep learning. This method has the characteristics of high accuracy and fast speed.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A method for retinal vessel segmentation based on OCTA images combining 3D and 2D methods includes the following steps:

[0009] S1. Perform cropping and slicing operations on the input 3D data, and convert the pixel values ​​of the input 3D data and 2D image; at the same time, obtain the blood vessel label map of the 2D image;

[0010] S2. Input the preprocessed 3D data into the 3D adaptive projection learning network to generate a projection map of the 3D data;

[0011] S3, input the projection map of the 3D data and the corresponding 2D image into the 2D blood vessel segmentation network to obtain a predicted map of retinal blood vessels;

[0012] S4, generate a blood vessel segment according to the predicted map and a blood vessel label map corresponding to the predicted map, calculate the thickness of the blood vessel segment, calculate the thickness inconsistency weight of the pixels in the blood vessel segment according to the calculated thickness, and calculate the blood vessel thickness inconsistency loss;

[0013] S5, calculate the Tversky loss based on the predicted map and the blood vessel label map, iteratively optimize the 2D blood vessel segmentation network according to the calculated total loss, and test and verify the segmentation result using the optimized network, and reserve the network parameters with the minimum total loss as the final prediction network;

[0014] S6, use the final prediction network to segment the retinal blood vessels of the OCTA image to be processed.

[0015] Further, in step S1, the size of the input 3D data is 304px×304px×640px or 400px×400px×640px.

[0016] Further, in step S1, the input 3D data is cropped and cut, and the pixel value of the input 3D data and the 2D image is converted, which specifically includes:

[0017] S11, in the vertical direction, the size of the 640px dimension is cropped, and only the pixels with the vertical direction coordinates located in the [161, 480] interval are reserved; in the horizontal direction, the entire 3D data is divided into 16 small blocks, and the size of each block obtained finally is: 19px×19px×320px or 25px×25px×320px;

[0018] S12, convert the pixel value range of the input 3D data and the 2D image from [0, 255] to [0, 1].

[0019] Further, in step S2, the 3D adaptive projection learning network includes three cascaded learning networks, each of which includes an input layer, an attention weighting layer and a projection layer connected in turn, and the first two learning networks further include a dimension reduction layer, and step S2 specifically includes:

[0020] S21, in each learning network, a series of 3D data blocks after cutting is input into the input layer in turn;

[0021] S22, in each learning network, the input 3D data block is unidirectionally pooled through the attention weighting layer, the attention weight of the 3D data block is calculated and weighted;

[0022] S23, in the first two learning networks, the weighted data is reduced in the vertical direction by the dimension reduction layer, and the reduced data is input into the next layer learning network;

[0023] S24, in each learning network, the weighted data is projected in the vertical direction by the projection layer to generate a series of projection blocks, and the series of projection blocks are spliced in the horizontal direction to generate a projection map.

[0024] Further, the 2D blood vessel segmentation network is composed of a coarse segmentation network and a fine segmentation network in cascade, and the coarse segmentation network and the fine segmentation network each contain three encoders and three decoders, and step S3 specifically comprises:

[0025] S31, the projection map generated by the first learning network and the corresponding 2D image are input into the first encoder of the coarse segmentation network;

[0026] S32, the output of the first encoder and the projection map generated by the second learning network are input into the second encoder of the coarse segmentation network;

[0027] S33, the output of the second encoder and the projection map generated by the third learning network are input into the third encoder of the coarse segmentation network;

[0028] S34, the encoding results of the encoders are decoded by the three decoders of the coarse segmentation network to generate a coarse prediction result;

[0029] S35, the output of the first encoder of the coarse segmentation network and the coarse prediction result are taken as the input of the first encoder of the fine segmentation network, and the input of the remaining encoders of the fine segmentation network is the output of the previous layer encoder;

[0030] S36, the encoding results of the encoders are decoded by the three decoders of the fine segmentation network to generate a prediction map of retinal blood vessels.

[0031] Further, in step S4, the blood vessel segments are generated according to the prediction map and the corresponding blood vessel label map, specifically comprising:

[0032] S41, the center line of the blood vessels in the blood vessel label image is generated, and the center line is divided into a series of blood vessel skeleton segments of a predetermined length, and the blood vessel skeleton segments in the whole blood vessel label image constitute a blood vessel skeleton segment set Wherein, N is the total number of blood vessel skeleton segments, S j is a blood vessel skeleton segment;

[0033] S42, in the blood vessel skeleton segment set I S , each blood vessel skeleton segment S jThe maximum inscribed circle completely covered by the blood vessel pixel in the blood vessel label image is calculated as the center of each pixel on the image, with a radius of r; for S j Each pixel on the image is searched with a search radius R = r + 1, and finally a blood vessel skeleton I S The corresponding blood vessel segment set where V j is the blood vessel segment.

[0034] Further, in step S4, the thickness of the blood vessel segment is calculated, and the thickness inconsistency weight of the blood vessel segment pixel is calculated according to the calculated thickness, which specifically includes:

[0035] S43, in the blood vessel segment set I v , the relative thickness T j of each blood vessel segment is calculated based on the following formula:

[0036]

[0037] wherein, is the total number of pixels of the current blood vessel segment, is the length of the blood vessel skeleton segment corresponding to the current blood vessel segment;

[0038] S44, in the blood vessel segment set I v , the thickness inconsistency weight of the blood vessel segment pixel is calculated based on the following formula

[0039]

[0040] wherein, T max is the maximum relative thickness of the blood vessel segment in the blood vessel label image, is the relative thickness of the blood vessel segment in which the current pixel is located in the blood vessel label image.

[0041] Further, in step S4, the blood vessel thickness inconsistency loss L ti is calculated based on the following formula:

[0042]

[0043] wherein, N p is the number of pixels, p i ∈[0,1] and g i ∈{0,1} respectively represent the predicted value of the pixel and the pixel value in the corresponding blood vessel label image, is the sample balance weight, which is defined as follows:

[0044]

[0045] Further, in step S5, the Tversky loss LTversky (α,β):

[0046]

[0047] Wherein, alpha, beta are self-defined hyperparameters, when alpha = beta = 0.5, the Tversky coefficient at this time is equal to the dice coefficient;

[0048] The total loss calculated is:

[0049] L total =L Tversky (α, beta) + lambda L ti

[0050] Wherein, lambda > 0 is a self-defined weight.

[0051] The beneficial effects of the present application are:

[0052] 1. The present application uses deep learning technology to realize the automatic segmentation of retinal blood vessels, avoids the tediousness of manual segmentation of retinal blood vessels and the interference of human factors, and improves the objectivity of retinal blood vessel segmentation results.

[0053] 2. The present application can provide more objective and accurate evaluation means for the diagnosis and research of diseases related to retinal blood vessels, and can quickly and accurately segment out the retinal blood vessels. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 The step flowchart of the embodiment of the present application.

[0055] Figure 2 The network model diagram in the embodiment of the present application.

[0056] Figure 3 The structure diagram of the adaptive projection learning module in the embodiment of the present application.

[0057] Figure 4 The flowchart of the blood vessel thickness inconsistency weight calculation in the embodiment of the present application. DETAILED DESCRIPTION

[0058] The present application will be described in more detail below with reference to the drawings in the embodiments of the present application, which are only exemplary and do not limit the protection scope of the present application. Referring to the drawings, Figure 1 The present application provides an OCTA image retinal blood vessel segmentation method based on the combination of 3D and 2D, and the specific implementation process will be described below with reference to the drawings.

[0059] Step one, crop and cut the input 3D data, and convert the pixel values of the input 3D data and 2D image; at the same time, obtain the blood vessel label map of the 2D image.

[0060] In this embodiment, the image data used comes from the OCTA-500 dataset, which includes two subsets of different field of view types. The first subset is called OCTA_6M, which is composed of 300 OCTA images of normal people or patients of 6mm x 6mm FOV type, with a volume size of 400 pixels (px) x 400px x 640px. Its label image size is 400px x 400px. The second subset is called OCTA_3M, which is composed of 200 OCTA images of normal people of 3mm x 3mm FOV type, with a volume size of 304px x 304px x 640px. Its label image size is 304px x 304px. In addition, OCTA-500 provides six different projections. Among the six projection maps, the maximum projection between the inner limiting membrane layer and the outer plexiform layer based on OCTA (ILM OPL OCTA) is widely used for retinal blood vessel segmentation.

[0061] The cropping and cutting operations and the pixel value conversion specifically include:

[0062] In the vertical direction, the dimension of 640px is cropped, and only the pixels with vertical direction coordinates in the [161, 480] interval are retained; in the horizontal direction, the entire 3D data is divided into 16 small blocks, and finally each block has a size of: 19px x 19px x 320px or 25px x 25px x 320px;

[0063] The pixel value range of the input 3D data and 2D image is converted from [0, 255] to [0, 1].

[0064] Step two, input the preprocessed 3D data into the 3D adaptive projection learning network to generate the projection map of the 3D data.

[0065] Referring to the accompanying drawings Figure 2 In this embodiment, the 3D adaptive projection learning network used includes three cascaded learning networks, each of which includes an input layer, an attention weighting layer and a projection layer connected in turn, and the first two learning networks also include a dimension reduction layer.

[0066] Step two specifically includes:

[0067] 1) In each learning network, a series of 3D data blocks after cutting are input into the input layer in turn;

[0068] 2) In each learning network, the input 3D data block is unidirectionally pooled through the attention weighting layer, the attention weight of the 3D data block is calculated and weighted;

[0069] 3) In the first two learning networks, the weighted data is reduced in the vertical direction by the dimension reduction layer, and the reduced data is input into the next layer learning network;

[0070] 4) In each learning network, the weighted data is projected in the vertical direction by the projection layer to generate a series of projection blocks, and the series of projection blocks are spliced in the horizontal direction to generate a projection map.

[0071] In the process of 3D to 2D, it is very important to retain as much feature information as possible. The features between different dimensions have correlation, and modeling the correlation between different dimensions can retain more feature information. Therefore, in this embodiment, four weighted branches are used to model the feature dependency between any three dimensions. Specifically: in any weighted branch, we first perform one-way pooling on the input 3D data to reduce its dimension in a certain feature dimension, and retain the maximum pooling result and the average pooling result of the dimension with the most texture features and overall information. Then, the convolution result is multiplied by the input data to obtain the weighted feature map. In order to reduce the loss of information during dimension reduction, this embodiment uses one-way pooling and one-way convolution. The weighted feature is subjected to one-way pooling and one-way convolution to obtain the projection map and the 3D data with reduced vertical dimension.

[0072] Step three, input the projection map of the 3D data and the corresponding 2D image into the 2D blood vessel segmentation network to obtain the prediction map of the retinal blood vessels.

[0073] In this embodiment, the 2D blood vessel segmentation network is composed of a coarse segmentation network and a fine segmentation network in cascade, wherein the coarse segmentation network and the fine segmentation network each contain three encoders and three decoders. The input of the coarse segmentation network is the original OCTA projection map and the projection map generated by the 3D adaptive projection path, but the second layer and the third layer of the network perform convolution operation on the projection map generated by the 3D adaptive projection path to ensure the same number of channels as the input features of the previous layer, thereby avoiding semantic gap. The input of the fine segmentation network is the output of the first layer and the last layer of the coarse segmentation network. Both networks output a 2D retinal blood vessel prediction map and are supervised by a loss function. The final output of the 2D blood vessel segmentation path is the output result of the fine network.

[0074] In this embodiment, step three specifically includes:

[0075] 1) input the projection map generated by the first learning network and the corresponding 2D image into the first encoder of the coarse segmentation network;

[0076] 2) input the output of the first encoder and the projection map generated by the second learning network into the second encoder of the coarse segmentation network;

[0077] 3) input the output of the second encoder and the projection map generated by the third learning network into the third encoder of the coarse segmentation network;

[0078] 4) decode the encoding results of the encoders through the three decoders of the coarse segmentation network to generate a coarse prediction result;

[0079] 5) input the output of the first encoder of the coarse segmentation network and the coarse prediction result as the input of the first encoder of the fine segmentation network, and input the output of the last layer of encoder as the input of the remaining encoders of the fine segmentation network;

[0080] 6) decode the encoding results of the encoders through the three decoders of the fine segmentation network to generate a prediction map of retinal blood vessels.

[0081] Step four, generating a blood vessel segment according to the prediction map and the corresponding blood vessel label map, calculating the thickness of the blood vessel segment, calculating the thickness inconsistency weight of the pixel of the blood vessel segment according to the calculated thickness, and calculating the blood vessel thickness inconsistency loss.

[0082] The pixel-level loss is calculated and back-propagated by directly comparing the deviation between the predicted probability of the pixel and the corresponding label. When calculating the pixel-level loss, pixels of the same class are treated equally. However, considering the problem of high imbalance between thick blood vessels and thin blood vessels in retinal blood vessels, the proportion of thick blood vessel pixels is much larger than that of thin blood vessel pixels. In order to guide the network to learn the characteristics of small blood vessels and improve the segmentation result of small blood vessels, the thickness inconsistency loss of blood vessels is introduced in this example. By balancing the importance of thick blood vessels and thin blood vessels, the network is guided to pay equal attention to blood vessels with inconsistent thickness, thereby improving the blood vessel segmentation result.

[0083] wherein the generating a blood vessel segment according to the prediction map and the corresponding blood vessel label map specifically comprises:

[0084] 1) generating a center line of the blood vessels in the blood vessel label image, and dividing the center line into a series of blood vessel skeleton segments of a predetermined length, the blood vessel skeleton segments in the entire blood vessel label image constituting a blood vessel skeleton segment set I wherein N is the total number of blood vessel skeleton segments, S j is a blood vessel skeleton segment;

[0085] 2) in the blood vessel skeleton segment set I S , taking each pixel on each blood vessel skeleton segment S j as the center, calculating the largest inscribed circle completely covered by the blood vessel pixels in the blood vessel label image, and the radius is r; for each pixel on S j , performing pixel-by-pixel search with a search radius R = r + 1, and finally generating a blood vessel skeleton I S corresponding blood vessel segment set wherein V j is a vessel segment. The generation process of the vessel segment is shown in Figure 4 .

[0086] In the vessel segment set I v , the relative thickness T j of each vessel segment is calculated based on the following formula:

[0087]

[0088] wherein, is the total number of pixels of the current vessel segment, is the length of the vessel skeleton segment corresponding to the current vessel segment;

[0089] In order to balance the importance of thick and thin vessels in loss, an adaptive thickness inconsistency balance weight is designed in the present application In the vessel segment set I v , the thickness inconsistency weight of the vessel segment pixel is calculated based on the following formula:

[0090]

[0091] wherein, T max is the maximum relative thickness of the vessel segment in the vessel label map, is the relative thickness of the vessel segment where the current pixel is located in the vessel label map.

[0092] Further, the vessel thickness inconsistency loss L ti is calculated based on the following formula:

[0093]

[0094] wherein, N p is the number of pixels, p i ∈[0,1] and g i ∈{0,1} represent the predicted value of the pixel and the pixel value in the corresponding vessel label map respectively, is the sample balance weight to alleviate the imbalance problem of positive and negative samples in the OCTA image, which is defined as follows:

[0095]

[0096] Step five, calculate the Tversky loss based on the prediction map and the vessel label map, iteratively optimize the 2D vessel segmentation network according to the calculated total loss, and test and verify the segmentation result with the optimized network, and keep the network parameters with the minimum total loss as the final prediction network.

[0097] The Tversky loss is widely used in medical image segmentation and has achieved excellent results. Therefore, the Tversky loss is introduced to further improve the retinal blood vessel segmentation result.

[0098] In the embodiment, the Tversky loss L is calculated based on the following formula Tversky :

[0099]

[0100] Wherein, alpha, beta are self-defined hyperparameters, when alpha = beta = 0.5, the Tversky coefficient at this time is equal to the dice coefficient;

[0101] The total loss calculated is:

[0102] L total = L Tversky (alpha, beta) + lambda L ti

[0103] Wherein, lambda > 0 is a self-defined weight.

[0104] The above embodiments are only used to help understand the method of the present application and its core idea. It should be pointed out that for those skilled in the art, without departing from the principles of the present application, the present application can be improved and modified, and these improvements and modifications also fall within the scope of protection of the claims of the present application.

Claims

1. A method for retinal blood vessel segmentation based on 3D and 2D combined OCTA images, characterized in that, The method comprises the following steps: S1, cutting and slicing the input 3D data, and converting the pixel values of the input 3D data and 2D image; meanwhile, obtaining the blood vessel label map of the 2D image; S2, inputting the preprocessed 3D data into a 3D adaptive projection learning network to generate a projection map of the 3D data, the 3D adaptive projection learning network comprising three cascaded learning networks, each learning network comprising an input layer, an attention weighting layer and a projection layer connected in sequence, and the first two learning networks further comprising a dimension reduction layer; specifically comprising: S21, in each learning network, a series of 3D data blocks after slicing are input into the input layer in sequence; S22, in each learning network, the input 3D data block is unidirectionally pooled through the attention weighting layer to calculate the attention weight of the 3D data block and weight it; S23, in the first two learning networks, the weighted data is dimensionally reduced in the vertical direction through the dimension reduction layer, and the dimensionally reduced data is input into the next layer learning network; S24, in each learning network, the weighted data is projected in the vertical direction through the projection layer to generate a series of projection blocks, and the series of projection blocks are spliced in the horizontal direction to generate a projection map; S3, inputting the projection map of the 3D data and the corresponding 2D image into a 2D blood vessel segmentation network to obtain a predicted map of the retinal blood vessels, the 2D blood vessel segmentation network being composed of a cascaded coarse segmentation network and a fine segmentation network, and each of the coarse segmentation network and the fine segmentation network comprising three encoders and three decoders; specifically comprising: S31, inputting the projection map generated by the first learning network and the corresponding 2D image into the first encoder of the coarse segmentation network; S32, inputting the output of the first encoder and the projection map generated by the second learning network into the second encoder of the coarse segmentation network; S33, inputting the output of the second encoder and the projection map generated by the third learning network into the third encoder of the coarse segmentation network; S34, decoding the encoding results of the encoders through the three decoders of the coarse segmentation network to generate a coarse prediction result; S35, taking the output of the first encoder of the coarse segmentation network and the coarse prediction result as the input of the first encoder of the fine segmentation network, and taking the output of the previous layer encoder as the input of the remaining encoders of the fine segmentation network; S36, decoding the encoding results of the encoders through the three decoders of the fine segmentation network to generate a predicted map of the retinal blood vessels; S4, generating a blood vessel segment according to the predicted map and the corresponding blood vessel label map, calculating the thickness of the blood vessel segment, calculating the thickness inconsistency weight of the pixels of the blood vessel segment according to the calculated thickness, and calculating the blood vessel thickness inconsistency loss; S5, calculating the Tversky loss based on the predicted map and the blood vessel label map, iteratively optimizing the 2D blood vessel segmentation network according to the calculated total loss, and verifying the segmentation result by using the optimized network, and retaining the network parameters with the minimum total loss as the final prediction network; S6, using the final prediction network to segment the retinal blood vessels of the OCTA image to be processed.

2. The OCTA image-based retinal blood vessel segmentation method according to claim 1, wherein, The size of the input 3D data in step S1 is 304px*304px*640px or 400px*400px*640px.

3. The 3D and 2D combined OCTA image retinal blood vessel segmentation method according to claim 2, characterized in that, In step S1, the input 3D data is cropped and cut, and the pixel value conversion of the input 3D data and 2D image specifically includes: S11, in the vertical direction, the dimension of 640px is cropped, and only the pixels with the vertical direction coordinates in the interval [161, 480] are reserved; in the horizontal direction, the entire 3D data is divided into 16 small blocks, and finally the size of each block is: 19px*19px*320px or 25px*25px*320px; S12, the pixel value range of the input 3D data and 2D image is converted from [0, 255] to [0, 1].

4. The OCTA image-based retinal blood vessel segmentation method according to claim 1, wherein, In step S4, the blood vessel segment is generated according to the prediction map and the corresponding blood vessel label map, specifically including: S41, a center line of the blood vessel in the blood vessel label image is generated, and the center line is divided into a series of blood vessel skeleton segments of a preset length, and the blood vessel skeleton segments in the whole blood vessel label image constitute a blood vessel skeleton segment set wherein N is the total number of blood vessel skeleton segments, is a blood vessel skeleton segment; S42、In the blood vessel skeleton segment set , the maximum inscribed circle completely covered by the blood vessel pixel in the blood vessel label image is calculated with each pixel on each blood vessel skeleton segment as the center, and the radius is r; for each pixel on , a pixel-by-pixel search is performed with a search radius R = r + 1, and finally the blood vessel skeleton corresponding to the blood vessel segment set = , wherein is the blood vessel segment.

5. The OCTA image retinal blood vessel segmentation method based on the combination of 3D and 2D according to claim 4, characterized in that, In step S4, the thickness of the blood vessel segment is calculated, and the thickness inconsistency weight of the pixels of the blood vessel segment is calculated according to the calculated thickness, specifically including: S43、In the set of vessel segments The relative thickness of each vessel segment is calculated based on the following formula : wherein, is the total number of pixels of the current vessel segment, is the length of the vessel skeleton segment corresponding to the current vessel segment; S44、In the set of vessel segments The thickness inconsistency weight of the vessel segment pixel is calculated based on the following formula : wherein, is the maximum relative thickness of the vessel segment in the vessel label map, is the relative thickness of the vessel segment in the vessel label map in which the current pixel is located. 6.The 3D and 2D combined based OCTA image retinal blood vessel segmentation method of claim 5, wherein, In step S4, the blood vessel thickness inconsistency loss is calculated based on the following formula : where, is the number of pixels, and denote the predicted value of a pixel and its corresponding pixel value in the vessel label map, respectively, is the sample balancing weight, defined as follows: 。 7. The OCTA image-based retinal blood vessel segmentation method of claim 6, wherein, In step S5, the Tversky loss is calculated based on the following formula : wherein, is a self-defined hyper-parameter, when The Tversky coefficient at this time is equal to the dice coefficient; the total loss calculated is: wherein, is a custom weight.

Images