Blood vessel segmentation method and imaging method for OCTA images
By using a dual decoder model of image transformation and feature perturbation in OCTA image vascular segmentation, the problems of large errors in vascular segmentation, low accuracy and dependence on labeled data in the prior art are solved, and higher segmentation accuracy and accuracy are achieved , suitable for scenes with few labels.
Patent Information
- Application Number
- CN202210303004.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-24
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2042-03-24
AI Technical Summary
The existing OCTA image vascular segmentation methods have problems such as large errors, poor accuracy of microvascular area segmentation, relying on manual parameter adjustment, full supervision methods require a large amount of labeling data, heavy burden on doctors, complex local supervision learning, and the accuracy needs to be improved.
Using a dual-decoder model based on image transformation and feature perturbation, backpropagation is performed by monitoring partial losses and consistency losses, multi-scale random grid transformation is used to alleviate network overfitting, and the potential information of labelless data is fully utilized. Construct a OCTA image vascular segmentation method with high reliability and good practicality.
It improves the accuracy and accuracy of OCTA image vascular segmentation, reduces dependence on annotated data, reduces the burden on doctors, and improves the segmentation effect of the model. It is suitable for scenes with less labeling of OCTA blood vessel images.
Smart Images

Figure CN114693928B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of image processing, and in particular relates to a blood vessel segmentation method and an imaging method of an OCTA image. Background Art
[0002] With the development of economy and technology and the improvement of people's living standards, people are paying more and more attention to health. With the development of image processing technology and deep learning technology, more and more deep learning solutions and image processing solutions are being applied in the medical field, bringing endless convenience to people.
[0003] The state of retinal blood vessels is of great significance in both clinical and experimental research. At present, with the development of technology, a non-invasive imaging technology has emerged - Optical Coherence Tomography Angiography (OCTA). This technology is based on the existence of flowing blood cells in the fundus blood vessels. It images the same cross section multiple times and calculates the blood flow signal, and then performs three-dimensional reconstruction of the blood flow to restore the image of the fundus blood vessels; therefore, this technology can observe the morphological structure of the tiny capillaries in the fundus. For this reason, automatic detection and segmentation of blood vessels in OCTA images is particularly important.
[0004] There are currently some research methods on retinal OCTA image vascular segmentation, which can be roughly divided into traditional segmentation methods and deep learning-based segmentation methods.
[0005] (1) Traditional segmentation method
[0006] In 2016, Gopinath et al. proposed a method for classifying glaucoma using capillary features in OCTA images; this method practiced the segmentation of capillaries in OCTA, but the segmentation effect was not ideal and may cause large errors. In 2017, Eladawi et al. proposed a method using Markov-Gibbs random field to segment blood vessels in OCTA images. In addition to the prior probability model of OCTA images, this method also integrated the appearance model and the spatial model. In addition to the first-order intensity model, the high-order MGRF (HO-MGRF) model is also used to consider spatial information to overcome the low contrast between blood vessels and other tissues; this method achieves a high accuracy, but the segmentation accuracy of some microvascular areas is not ideal. In 2020, Le et al. proposed a filter based on the Hessian matrix and the Ostu threshold method to detect and enhance small blood vessels; this method is simple but the segmentation effect is not ideal. These filter-based methods usually rely heavily on manual parameter adjustment during implementation.
[0007] (2) Segmentation method based on deep learning
[0008] In 2019, Deng et al. from the University of Iowa proposed a method for measuring vascular loss in OCTA; this method cleverly avoids locations where microvascular structures are difficult to segment, and can more accurately measure the vascular loss area in OCTA, but because the vascular network cannot be accurately segmented, parameter information such as capillary density cannot be obtained. Mou et al. proposed a unified curve structure segmentation network based on U-net's channel and spatial attention; however, this method is a fully supervised method, and OCTA is particularly difficult to annotate for vascular networks, requiring experienced doctors to spend a lot of time and energy on manual annotation, which increases the burden on doctors. In 2020, Li et al. proposed an end-to-end image projection network (IPN) structure that can achieve three-dimensional to two-dimensional image segmentation in OCTA images; however, this method only segments large blood vessels on OCTA, but does not segment microvascular networks. Mou et al. proposed a curve structure segmentation network, which contains an attention mechanism in the encoder and decoder to learn the rich hierarchical representation of the curve structure. However, this method is based on a fully supervised segmentation method and relies on large-scale labeled training data. However, the OCTA image has dense and interlaced blood vessels, and the image contrast is not high. Labeling is time-consuming and laborious, and it is difficult to collect a large amount of labeled data. In 2021, Xu et al. proposed a segmentation method of local supervised learning. This method locally annotates part of the image. This method adopts an active learning strategy and selects the patch with the largest amount of information for further annotation. However, this method is relatively complex and the accuracy of vascular segmentation needs to be further improved. Summary of the invention
[0009] One of the purposes of the present invention is to provide a blood vessel segmentation method for OCTA images based on image transformation and feature perturbation, which has high reliability, good practicability and good segmentation effect.
[0010] A second object of the present invention is to provide an imaging method including the blood vessel segmentation method of the OCTA image.
[0011] The blood vessel segmentation method of the OCTA image provided by the present invention comprises the following steps:
[0012] S1. Obtain a retinal OCTA image dataset and divide it into labeled images and unlabeled images;
[0013] S2. Construct an encoder and a dual decoder to form a segmentation model; the dual decoder includes a main decoder and an auxiliary decoder;
[0014] S3. Select a number of labeled images and unlabeled images from the images obtained in step S1;
[0015] S4. Input the labeled image into the encoder and the main decoder in sequence to obtain the blood vessel segmentation result, calculate the supervision part loss, and perform back propagation to update the parameters of the encoder and the main decoder;
[0016] S5. Input the unlabeled image into the encoder, and input the output result of the encoder into the main decoder or the auxiliary decoder to obtain the blood vessel segmentation result, calculate the consistency loss, and perform back propagation to update the parameters of the encoder and the auxiliary decoder;
[0017] S6. Repeat steps S3 to S5 until the model training is completed, and obtain the final vascular segmentation model of the OCTA image;
[0018] S7. Using the final OCTA image blood vessel segmentation model obtained in step S6, perform blood vessel segmentation on the actual OCTA image.
[0019] Step S1 is to obtain a retinal OCTA image dataset, which is divided into labeled images and unlabeled images. Specifically, the retinal OCTA image dataset is obtained, which is divided into labeled images. and unlabeled images in, is the i-th labeled image, is the label of the i-th labeled image, is the i-th unlabeled image, N L is the total number of labeled images, N U is the total number of unlabeled images.
[0020] The encoder and the dual decoder described in step S2 are constructed to form a segmentation model; the dual decoder includes a main decoder and an auxiliary decoder; specifically, the encoder and the dual decoder are constructed to form a segmentation model; the dual decoder includes a main decoder and an auxiliary decoder; the encoder parameter is θ e , the parameters of the main decoder are θ m , the auxiliary decoder parameter is θ a .
[0021] Step S4, in which the labeled image is sequentially input into the encoder and the main decoder to obtain the blood vessel segmentation result, calculate the supervision part loss, and perform back propagation to update the parameters of the encoder and the main decoder, specifically includes the following steps:
[0022] The labeled image is sequentially input into the encoder and the main decoder to obtain the blood vessel segmentation result;
[0023] Supervision of partial loss L s Defined as the pixel-by-pixel cross entropy as the loss function; the cross entropy loss function L CE (X L ,YL θ e ,θ m ) is defined as Where X L is a labeled image, Y L is the label corresponding to the labeled image, θ e is the encoder parameter, θ m are the parameters of the main decoder, h is the height of the image, w is the width of the image, c is the number of categories, y L is the true label, log is the logarithm with base 2, is the decoding operation function of the main decoder, is the encoding operation function of the encoder, x L To mark the sample;
[0024] Back-propagation is performed via the gradient descent algorithm to update the parameters of the encoder and main decoder.
[0025] Step S5, in which the unlabeled image is input into the encoder, and the output result of the encoder is input into the main decoder or the auxiliary decoder, to obtain the blood vessel segmentation result, calculate the consistency loss, and perform back propagation to update the parameters of the encoder and the auxiliary decoder, specifically includes the following steps:
[0026] A. Input the unlabeled image into the encoder and main decoder in turn to obtain the unlabeled main decoder blood vessel segmentation result
[0027] B. Randomly transform the unlabeled image And the transformed unlabeled image Input to the encoder to obtain unlabeled random transformation encoding features For random transformation The operation function of
[0028] C. Randomly transform the unlabeled encoding features Input into the auxiliary decoder to obtain the vessel segmentation result of the unlabeled random transformation auxiliary decoder
[0029] D. Obtained unlabeled random transformation auxiliary decoder blood vessel segmentation results Perform inverse transformation to obtain the unlabeled random transformation auxiliary decoder inverse transformation blood vessel segmentation result
[0030] E. Inverse transformation of blood vessel segmentation results based on the unlabeled random transformation auxiliary decoder Blood vessel segmentation results with unlabeled main decoder Compute consistency loss based on image transformation;
[0031] F. Randomly transform the unlabeled encoding features Perform inverse transformation to obtain unlabeled random transformation inverse transformation encoding features And input the auxiliary decoder to obtain the unlabeled random transform inverse transform auxiliary decoder blood vessel segmentation result
[0032] G. Blood vessel segmentation results of auxiliary decoder based on unlabeled random transformation and inverse transformation Blood vessel segmentation results with unlabeled main decoder Calculate the consistency loss based on feature perturbations;
[0033] H. Perform back propagation through the gradient descent algorithm to update the parameters of the encoder and auxiliary decoder.
[0034] Random transformation as described in step B Specifically, random grid transformation and random rotation transformation are used for random transformation; the random grid transformation uses 2×2 and 3×3 grids, and the range of random rotation is [-180°, 180°].
[0035] The calculation of the consistency loss based on image transformation described in step E is specifically to calculate the consistency loss based on image transformation L using the following formula: d (X U θ e ,θ m ,θ a ):
[0036]
[0037] Where X U is an unlabeled image; θ e is the encoder parameter; θ m is the parameter of the main decoder; θ a is the parameter of the auxiliary decoder; mse() is the mean square error function; x U is the unlabeled image currently calculated; For random transformation The operation function of For random transformation The inverse transformation operation function; is the encoding operation function of the encoder; Decoding operation function of the main decoder; It is the decoding operation function of the auxiliary decoder.
[0038] The calculation of the consistency loss based on the feature perturbation in step G is specifically based on the following formula: f (X U θ e ,θm ,θ a ):
[0039]
[0040] Where X U is an unlabeled image; θ e is the encoder parameter; θ m is the parameter of the main decoder; θ a is the parameter of the auxiliary decoder; mse() is the mean square error function; x U is the unlabeled image currently calculated; For random transformation The operation function of For random transformation The inverse transformation operation function; is the encoding operation function of the encoder; Decoding operation function of the main decoder; It is the decoding operation function of the auxiliary decoder.
[0041] The final vascular segmentation model of the OCTA image described in step S6 is specifically obtained by averaging the segmentation results of the main decoder and the auxiliary decoder to obtain the final segmentation result.
[0042] The present invention also discloses an imaging method including the blood vessel segmentation method of the OCTA image, which specifically includes the following steps:
[0043] (1) Acquire OCTA images;
[0044] (2) using the above-mentioned OCTA image blood vessel segmentation method to segment the blood vessels in the OCTA image obtained in step (1);
[0045] (3) based on the segmentation result of step (2), marking and re-imaging the blood vessel segmentation result on the OCTA image acquired in step (1);
[0046] (4) Output the OCTA image with vascular markers to complete the final imaging process.
[0047] The OCTA image blood vessel segmentation method and imaging method provided by the present invention introduce consistency based on feature perturbation without modifying the network structure, make full use of the potential information of unlabeled data through double consistency regularization, and adopt multi-scale random grid transformation to alleviate network overfitting; the method of the present invention improves the segmentation accuracy of the model while using only limited labeled data and a large amount of unlabeled data, making the present invention more suitable for actual scenes with few OCTA blood vessel images, and has high reliability, good practicability and good segmentation effect. BRIEF DESCRIPTION OF THE DRAWINGS
[0048] Figure 1 Schematic diagram of the process flow of the segmentation method of the present invention.
[0049] Figure 2 It is a schematic diagram of the effect of random transformation of the present invention.
[0050] Figure 3 It is a visualization diagram of the segmentation result of the present invention.
[0051] Figure 4 The figure is a schematic diagram of the method flow of the imaging method of the present invention. DETAILED DESCRIPTION
[0052] like Figure 1 The figure shows a flow chart of the segmentation method of the present invention: the blood vessel segmentation method of the OCTA image provided by the present invention comprises the following steps:
[0053] S1. Obtain a retinal OCTA image dataset and divide it into labeled images and unlabeled images; specifically, obtain a retinal OCTA image dataset and divide it into labeled images and unlabeled images in, is the i-th labeled image, is the label of the i-th labeled image, is the i-th unlabeled image, N L is the total number of labeled images, N U is the total number of unlabeled images;
[0054] S2. Construct an encoder and a dual decoder to form a segmentation model; the dual decoder includes a main decoder and an auxiliary decoder; specifically, construct an encoder and a dual decoder to form a segmentation model; the dual decoder includes a main decoder and an auxiliary decoder; the parameter of the encoder is θ e , the parameters of the main decoder are θ m , the auxiliary decoder parameter is θ a ;
[0055] S3. Select a number of labeled images and unlabeled images from the images obtained in step S1;
[0056] S4. Input the labeled image into the encoder and the main decoder in sequence to obtain the blood vessel segmentation result, calculate the supervision part loss, and perform back propagation to update the parameters of the encoder and the main decoder; Input the labeled image into the encoder and the main decoder in sequence to obtain the blood vessel segmentation result;
[0057] Supervision of partial loss L s Defined as the pixel-by-pixel cross entropy as the loss function; the cross entropy loss function LCE (X L ,Y L θ e ,θ m ) is defined as Where X L is a labeled image, Y L is the label corresponding to the labeled image, θ e is the encoder parameter, θ m are the parameters of the main decoder, h is the height of the image, w is the width of the image, c is the number of categories, y L is the true label, log is the logarithm with base 2, is the decoding operation function of the main decoder, is the encoding operation function of the encoder, x L To mark samples;
[0058] Back-propagation is performed through the gradient descent algorithm to update the parameters of the encoder and main decoder;
[0059] S5. Input the unlabeled image into the encoder, and input the output result of the encoder into the main decoder or the auxiliary decoder to obtain the blood vessel segmentation result, calculate the consistency loss, and perform back propagation to update the parameters of the encoder and the auxiliary decoder; specifically, the steps include:
[0060] A. Input the unlabeled image into the encoder and main decoder in turn to obtain the unlabeled main decoder blood vessel segmentation result
[0061] B. Randomly transform the unlabeled image And the transformed unlabeled image Input to the encoder to obtain unlabeled random transformation encoding features For random transformation Operation function of; random transformation Specifically, random grid transformation and random rotation transformation are used for random transformation; the random grid transformation uses 2×2 and 3×3 grids, and the range of random rotation is [-180°, 180°]. The specific effects are as follows Figure 2 As shown;
[0062] C. Randomly transform the unlabeled encoding features Input into the auxiliary decoder to obtain the vessel segmentation result of the unlabeled random transformation auxiliary decoder
[0063] D. Obtained unlabeled random transformation auxiliary decoder blood vessel segmentation results Perform inverse transformation to obtain the unlabeled random transformation auxiliary decoder inverse transformation blood vessel segmentation result
[0064] E. Inverse transformation of blood vessel segmentation results based on the unlabeled random transformation auxiliary decoder Blood vessel segmentation results with unlabeled main decoder Calculate the consistency loss based on image transformation; specifically, the consistency loss L based on image transformation is calculated using the following formula: d (X U θ e ,θ m ,θ a ):
[0065]
[0066] Where X U is an unlabeled image; θ e is the encoder parameter; θ m is the parameter of the main decoder; θ a is the parameter of the auxiliary decoder; mse() is the mean square error function; x U is the unlabeled image currently calculated; For random transformation The operation function of For random transformation The inverse transformation operation function; is the encoding operation function of the encoder; Decoding operation function of the main decoder; Decoding operation function of auxiliary decoder;
[0067] F. Randomly transform the unlabeled encoding features Perform inverse transformation to obtain unlabeled random transformation inverse transformation encoding features And input the auxiliary decoder to obtain the unlabeled random transform inverse transform auxiliary decoder blood vessel segmentation result
[0068] G. Blood vessel segmentation results of auxiliary decoder based on unlabeled random transformation and inverse transformation Blood vessel segmentation results with unlabeled main decoder Calculate the consistency loss based on feature perturbation; specifically, the consistency loss L based on feature perturbation is calculated using the following formula: f (X U θ e ,θ m ,θ a ):
[0069]
[0070] Where X U is an unlabeled image; θ eis the encoder parameter; θ m is the parameter of the main decoder; θ a is the parameter of the auxiliary decoder; mse() is the mean square error function; x U is the unlabeled image currently calculated; For random transformation The operation function of For random transformation The inverse transformation operation function; is the encoding operation function of the encoder; Decoding operation function of the main decoder; Decoding operation function of auxiliary decoder;
[0071] H. Perform back propagation through the gradient descent algorithm to update the parameters of the encoder and auxiliary decoder;
[0072] S6. Repeat steps S3 to S5 until the model training is completed, and obtain the final OCTA image blood vessel segmentation model; in specific implementation, the segmentation results of the main decoder and the auxiliary decoder are averaged to obtain the final segmentation result;
[0073] S7. Using the final OCTA image blood vessel segmentation model obtained in step S6, perform blood vessel segmentation on the actual OCTA image.
[0074] In the following, the segmentation method of the present invention is compared with the existing method in combination with the embodiments:
[0075] The method proposed in the present invention is compared with the existing segmentation methods: the fully supervised method uses all the labels of the training set for training, and the other methods use 3.3% of the labeled data. The accuracy of the segmentation results and the Ground Truth, the Dice coefficient and the false discovery rate are used as evaluation criteria; the specific segmentation results are shown in Table 1:
[0076] Table 1 Schematic diagram of OCTA vascular segmentation results
[0077] Segmentation method Accuracy (%) Dice coefficient (%) False discovery rate (%) Fully supervised methods 91.21 75.11 19.97 Semi-supervised methods 90.70 73.48 17.49 Method of the present invention 91.34 75.61 15.99
[0078] It can be seen from Table 1 that the method of the present invention is superior to the two existing different methods in terms of three indicators, including the fully supervised method, and obtains more accurate segmentation results through the dual consistency constraint. Figure 3 The results of the semi-supervised method and the method of the present invention are compared. From left to right, they are the original image, the ground truth, the result of the semi-supervised method and the result of the method of the present invention. It can be seen from the figure that the method of the present invention performs better in the continuity of segmented blood vessels.
[0079] like Figure 4The figure shows a schematic diagram of the method flow of the imaging method of the present invention: the imaging method provided by the present invention including the blood vessel segmentation method of the OCTA image specifically comprises the following steps:
[0080] (1) Acquire OCTA images;
[0081] (2) using the above-mentioned OCTA image blood vessel segmentation method to segment the blood vessels in the OCTA image obtained in step (1);
[0082] (3) based on the segmentation result of step (2), marking and re-imaging the blood vessel segmentation result on the OCTA image acquired in step (1);
[0083] (4) Output the OCTA image with vascular markers to complete the final imaging process.
[0084] The imaging method provided by the present invention can be applied to the existing OCTA imaging equipment; in specific application, the imaging method of the present invention is integrated into the equipment; when the equipment is working, firstly, the fundus image of the inspector is collected according to the existing technology, and the existing OCTA image is obtained; then, the equipment uses the vascular segmentation method of the OCTA image provided by the present invention to segment the blood vessels in the acquired OCTA image (at this time, the blood vessel segmentation and marking are not completed); then, according to the segmentation result, the blood vessels are highlighted on the OCTA image (for example, marked with a prominent color, etc.) and secondary imaging is performed; secondary imaging can obtain an OCTA image with prominent blood vessel markings; finally, the obtained OCTA image with prominent blood vessel markings is output, thereby completing the final imaging work. At this time, the existing OCTA imaging equipment becomes a multifunctional OCTA imaging equipment with blood vessel segmentation and marking.
Claims
1. A blood vessel segmentation method for OCTA images, comprising the following steps: S1. Obtain a retinal OCTA image dataset and divide it into labeled images and unlabeled images; S2. Construct an encoder and a dual decoder to form a segmentation model; the dual decoder includes a main decoder and an auxiliary decoder; S3. Select a number of labeled images and unlabeled images from the images obtained in step S1; S4. Input the labeled image into the encoder and the main decoder in sequence to obtain the blood vessel segmentation result, calculate the supervision part loss, and perform back propagation to update the parameters of the encoder and the main decoder; S5. Input the unlabeled image into the encoder, and input the output result of the encoder into the main decoder or the auxiliary decoder to obtain the blood vessel segmentation result, calculate the consistency loss, and perform back propagation to update the parameters of the encoder and the auxiliary decoder; S6. Repeat steps S3 to S5 until the model training is completed, and obtain the final vascular segmentation model of the OCTA image; S7. Using the final OCTA image blood vessel segmentation model obtained in step S6, perform blood vessel segmentation on the actual OCTA image.
2. The blood vessel segmentation method of OCTA image according to claim 1, characterized in that Step S1 is to obtain a retinal OCTA image dataset, which is divided into labeled images and unlabeled images. Specifically, the retinal OCTA image dataset is obtained, which is divided into labeled images. and unlabeled images in, is the i-th labeled image, is the label of the i-th labeled image, is the i-th unlabeled image, N L is the total number of labeled images, N U is the total number of unlabeled images.
3. The blood vessel segmentation method of OCTA image according to claim 2, characterized in that The encoder and the dual decoder described in step S2 are constructed to form a segmentation model; the dual decoder includes a main decoder and an auxiliary decoder; Specifically, an encoder and a dual decoder are constructed to form a segmentation model; the dual decoder includes a main decoder and an auxiliary decoder; the encoder parameter is θ e , the parameters of the main decoder are θ m , the auxiliary decoder parameter is θ a .
4. The blood vessel segmentation method of OCTA image according to claim 3, characterized in that Step S4, in which the labeled image is sequentially input into the encoder and the main decoder to obtain the blood vessel segmentation result, calculate the supervision part loss, and perform back propagation to update the parameters of the encoder and the main decoder, specifically includes the following steps: The labeled image is sequentially input into the encoder and the main decoder to obtain the blood vessel segmentation result; Supervision of partial loss L s Defined as the pixel-by-pixel cross entropy as the loss function; the cross entropy loss function L CE (X L ,Y L θ e ,θ m ) is defined as Where X L is a labeled image, Y L is the label corresponding to the labeled image, θ e is the encoder parameter, θ m are the parameters of the main decoder, h is the height of the image, w is the width of the image, c is the number of categories, y L is the true label, log is the logarithm with base 2, is the decoding operation function of the main decoder, is the encoding operation function of the encoder, x L To mark the sample; Back-propagation is performed via the gradient descent algorithm to update the parameters of the encoder and main decoder.
5. The blood vessel segmentation method of OCTA image according to claim 4, characterized in that Step S5, in which the unlabeled image is input into the encoder, and the output result of the encoder is input into the main decoder or the auxiliary decoder, to obtain the blood vessel segmentation result, calculate the consistency loss, and perform back propagation to update the parameters of the encoder and the auxiliary decoder, specifically includes the following steps: A. Input the unlabeled image into the encoder and main decoder in turn to obtain the unlabeled main decoder blood vessel segmentation result B. Randomly transform the unlabeled image And the transformed unlabeled image Input to the encoder to obtain unlabeled random transformation encoding features For random transformation The operation function of C. Randomly transform the unlabeled encoding features Input into the auxiliary decoder to obtain the vessel segmentation result of the unlabeled random transformation auxiliary decoder D. Obtained unlabeled random transformation auxiliary decoder blood vessel segmentation results Perform inverse transformation to obtain the unlabeled random transformation auxiliary decoder inverse transformation blood vessel segmentation result E. Inverse transformation of blood vessel segmentation results based on the unlabeled random transformation auxiliary decoder Blood vessel segmentation results with unlabeled main decoder Compute consistency loss based on image transformation; F. Randomly transform the unlabeled encoding features Perform inverse transformation to obtain unlabeled random transformation inverse transformation encoding features And input the auxiliary decoder to obtain the unlabeled random transform inverse transform auxiliary decoder blood vessel segmentation result G. Blood vessel segmentation results of auxiliary decoder based on unlabeled random transformation and inverse transformation Blood vessel segmentation results with unlabeled main decoder Calculate the consistency loss based on feature perturbations; H. Perform back propagation through the gradient descent algorithm to update the parameters of the encoder and auxiliary decoder.
6. The blood vessel segmentation method of OCTA image according to claim 5, characterized in that Random transformation as described in step B Specifically, random grid transformation and random rotation transformation are used for random transformation; the random grid transformation uses 2×2 and 3×3 grids, and the range of random rotation is [-180°, 180°].
7. The blood vessel segmentation method of OCTA image according to claim 6, characterized in that The calculation of the consistency loss based on image transformation described in step E is specifically to calculate the consistency loss based on image transformation L using the following formula: d (X U θ e ,θ m ,θ a ): Where X U is an unlabeled image; θ e is the parameter of the encoder; θ m is the parameter of the main decoder; θ a is the parameter of the auxiliary decoder; mse() is the mean square error function; x U is the unlabeled image currently calculated; For random transformation The operation function of For random transformation The inverse transformation operation function; is the encoding operation function of the encoder; Decoding operation function of the main decoder; It is the decoding operation function of the auxiliary decoder.
8. The blood vessel segmentation method of OCTA image according to claim 7, characterized in that The calculation of the consistency loss based on the feature perturbation in step G is specifically based on the following formula: f (X U θ e ,θ m ,θ a ): Where X U is an unlabeled image; θ e is the parameter of the encoder; θ m is the parameter of the main decoder; θ a is the parameter of the auxiliary decoder; mse() is the mean square error function; x U is the unlabeled image currently calculated; For random transformation The operation function of For random transformation The inverse transformation operation function; is the encoding operation function of the encoder; Decoding operation function of the main decoder; It is the decoding operation function of the auxiliary decoder.
9. The blood vessel segmentation method of OCTA image according to claim 8, characterized in that The final vascular segmentation model of the OCTA image described in step S6 is specifically obtained by averaging the segmentation results of the main decoder and the auxiliary decoder to obtain the final segmentation result.
10. An imaging method comprising the blood vessel segmentation method of OCTA images according to any one of claims 1 to 9, characterized in that The specific steps include: (1) Acquire OCTA images; (2) using the OCTA image blood vessel segmentation method according to any one of claims 1 to 9 to segment the blood vessels in the OCTA image acquired in step (1); (3) based on the segmentation result of step (2), marking and re-imaging the blood vessel segmentation result on the OCTA image acquired in step (1); (4) Output the OCTA image with vascular markers to complete the final imaging process.
