A method for optical coherence tomography choroidal vascular layer angiography

Abstract

The present invention provides a method for optical coherence tomography choroidal vascular layer angiography, which belongs to the field of medical imaging and image processing technology. The present invention is based on domain adaptation and semi-supervised image segmentation. A new integrated domain discriminator is used to identify the domain features of the source domain and the integrated target domain, and adversarial training is performed on the segmentation model. At the same time, a new type of confusion domain sample filter is used to filter pseudo labels to reduce the label noise caused by domain differences, and then the low-quality OCT samples after screening are used to perform semi-supervised training on the segmentation model. The present invention is a non-invasive angiography method that does not rely on any contrast agent assistance, so it can be used for large-scale analysis and screening of choroidal blood vessels. In addition, the present invention can also provide three-dimensional vascular distribution of the choroid, so it can be used to evaluate the sub-layer structure of the choroid and the calculation of three-dimensional biological annotation quantities such as vascular density and vascular index, and provide a more effective tool for the analysis and research of choroid-related diseases.

Description

Technical Field

[0001] The present invention provides a choroidal angiography method, in particular to an optical coherence tomography choroidal angiography method based on domain adaptation and semi-supervised image segmentation, belonging to the technical field of medical imaging and image processing. Background Art

[0002] The choroid is the most densely vascularized tissue in the human eye, and its structure and vascular changes are the basis for the diagnosis of many ophthalmic diseases. Existing choroidal angiography methods can be mainly divided into two categories. One is the fundus fluorescein angiography method that relies on intravenous injection of contrast agents [1]. This method uses indocyanine green (ICGA) as a contrast agent for intravenous injection, which can achieve the effect of visualizing the two-dimensional vascular structure of the choroidal blood vessels. However, this method can only achieve two-dimensional visualization of the choroidal blood vessels, and cannot provide three-dimensional information of the blood vessels for the evaluation of the blood vessels of each sublayer of the choroid and will be greatly interfered with by retinal artifacts. At the same time, due to the need for intravenous injection of contrast agents, this method will also cause damage to the patient's renal function [2], and cannot be used for large-scale analysis of the pathological characteristics of the choroidal disease. In addition to fluorescein angiography, optical coherence tomography (OCT) is also a method that can qualitatively evaluate the choroid. Through optical coherence tomography, the fundus structure can be visualized in three dimensions. However, optical coherence tomography is not a vascular angiography method. Therefore, in addition to blood vessels, other tissue information of the fundus, such as the choroidal stroma, will also be captured simultaneously. While the current optical coherence tomography angiography (OCTA) [3] can effectively capture retinal blood vessels or capillaries, it will show black artifacts in the large blood vessels of the choroid [4], and therefore cannot capture the important choroidal blood vessel information. This also makes OCTA unable to be used for quantitative analysis of choroidal blood vessels in clinical practice.

[0003] Currently, most image segmentation methods for choroidal structure aim to separate the choroid layer from the retinal layer in OCT images, and then visualize the choroid structure in two dimensions by maximum density projection. Early choroidal layer segmentation methods were performed in two steps, first extracting edge features from the OCT image, and then locating the choroid layer based on the edge features to complete the choroidal structure segmentation. Among them, Sui et al. [5] used a deep neural network to learn the affinity matrix between pixels and used it as a graph model for the graph search method to locate the choroid structure. He et al. [6] used l2-l qThe regularization term constrains the choroid edge detection part, and then uses a graph search method to locate the upper and lower boundaries of the choroid. It can be found that the above methods all need to rely on graph search as a subsequent step to locate the choroid layer boundary, so it is impossible to achieve end-to-end training of the neural network. Compared with the above methods, some works model the choroid layer segmentation as a semantic segmentation task, that is, by predicting whether each pixel in the OCT image belongs to the choroid, the choroid structure is located. Among them, Zhang et al. [7] used a semantic segmentation model to extract retinal pixels and choroid pixels in the OCT image respectively, and then eliminated the retinal artifacts of the former to better project the choroid layer. Zhu et al. [8] then used a multi-task learning method to simultaneously segment the choroidal blood vessels and retinal blood vessels. This method also provides a tool for quantitative evaluation of choroidal blood vessels in high-quality OCT sections. However, although the above methods can complete the visualization of the projection surface or section surface of the choroidal blood vessels, these methods all need to be based on high-quality OCT scan data and cannot adapt to the low-quality dense scan images used for clinical reconstruction. This problem can be partially solved by using domain adaptation technology.

[0004] Domain adaptation technology considers the difference between the distribution of training data and test data, so that the deep neural network trained in the source domain can be better applied to test data with different distributions, thereby improving the effect of the model in practical applications. Among them, traditional domain adaptation methods reasonably measure the difference between the training distribution and the test distribution and reduce this difference during the training process, so that the neural network can better obtain cross-domain features [9]. In the field of image segmentation, most methods use adversarial learning to achieve the above purpose. Among them, Tsai et al.

[10] used domain discriminators in both the output domain and the feature domain to perform adversarial learning to ensure that the model can finally obtain cross-domain feature distribution. Liu et al.

[11] introduced the mean teacher model to pseudo-label the target domain data to further enhance the performance of the model. Spaddotto et al.

[12] chose to use a generative adversarial network and used an additional discriminator to distinguish the prediction results of the model in the source domain from the manual annotations. Tranheden et al.

[13] generated cross-domain samples by data augmentation and trained the model based on the cross-domain samples to enhance the cross-domain performance of the model.

[0005] It can be found that for domain adaptation methods, using unlabeled target domain samples to generate their pseudo labels and using self-supervised models to assist model training is a more reasonable model training strategy. For OCT images, we can easily obtain a large number of low-quality densely scanned target domain images. However, compared with other domain adaptation tasks, due to the difficulty of choroidal vascular labeling, our source domain images are not sufficient. From this point of view, the present invention also hopes to use the semi-supervised learning idea that mainly studies how to use a large number of unlabeled samples to assist model training to better assist domain adaptation tasks. Among them, semi-supervised learning methods based on consistency constraints mostly achieve the purpose of semi-supervised learning by applying perturbations to unlabeled images and ensuring the consistency of prediction results before and after perturbations

[14] . Although this type of method can be better applied to training scenarios with sample distribution differences

[15] and is also adopted by some domain adaptation methods

[11] , this strategy will cause the model prediction to be too smooth, resulting in some models being unable to recognize difficult samples. In addition, there is a type of semi-supervised learning method that uses hard pseudo labels as supervisory signals for unlabeled samples and trains the model together with labeled samples

[16] . For example, Chen et al.

[17] trained two segmentation models with different parameter initializations and regarded their own prediction results as pseudo labels of the other model. Sohn et al.

[18] replaced the supervisory signal of the mean teacher model with a hard cross-entropy loss, and used the prediction results of the teacher model after perturbation as the supervisory signal of the student model before perturbation. Compared with consistency constraints, this type of pseudo-label-based semi-supervised learning strategy can make more full use of the information of unlabeled images, thereby more effectively promoting the effect of the model on difficult samples. However, this type of method cannot adapt to the situation where the distribution of labeled samples and unlabeled samples is different as well as the consistency constraint-based method

[15] . Summary of the Invention

[0006] In view of the many defects existing in the above-mentioned existing methods, the present invention provides an optical coherence tomography choroidal vascular layering angiography method based on domain adaptation and semi-supervised image segmentation.

[0007] The technical solutions provided by the present invention are as follows:

[0008] A method for optical coherence tomography choroidal vascular layer angiography, comprising the following steps:

[0009] S1. Image acquisition and training set preparation. The specific implementation steps are:

[0010] S1a High-quality OCT image acquisition. Use OCT image acquisition equipment to collect OCT images of K key scanning positions of the fundus of N subjects and form the source domain dataset in Indicates the color intensity of a frame in a high-quality OCT scan. To ensure the clarity of OCT images in the choroidal region, the following should be fully considered during the process: (1) Enhanced Deep OCT (EDI-OCT) or Swept Source OCT (SS-OCT) equipment should be used for image acquisition to ensure that the choroid is visible; (2) Professional optometrists should be employed whenever possible to select a reasonable focus area when acquiring key scanning positions frame by frame to ensure the integrity and clarity of the choroidal structure; (3) For scanning key positions, multiple averaging of the same position should be used to reduce speckle noise in the acquired image; and (4) Imaging equipment with fast imaging speed and convenient operation should be selected whenever possible to reduce eye movement interference during multi-position averaging.

[0011] S1b High-quality OCT image annotation. Hire professional ophthalmologists or practitioners to annotate the high-quality OCT key frame tomographic image set. The choroidal middle vessel layer, choroidal great vessel layer, choroidal middle vessel segment, and choroidal great vessel segment are finely labeled pixel by pixel and verified, and the source domain label set is constructed based on this annotation in, Represents X i The choroid sublayer annotation results. Represents X i The choroidal vessel segment annotation results.

[0012] S1c: Low-quality OCT image acquisition. Use OCT image acquisition equipment to perform fast and dense scanning of the fundus of N subjects with a frame number of M, and form the target domain dataset. in Represents the color intensity of a frame in a low-quality dense OCT scan. This process should fully consider: (1) image acquisition using enhanced deep OCT (EDI-OCT) or swept-source OCT (SS-OCT) equipment to ensure choroidal visibility; (2) ensuring that the acquisition interval is extremely small to ensure the accuracy of vascular reconstruction, that is, M should be much larger than K; and imaging equipment with fast imaging speed and convenient operation should be selected as much as possible to minimize image deviation caused by eye movement interference during dense scanning.

[0013] S1d. Preprocessing of the captured images. Based on the characteristics of the acquired data, perform certain preprocessing on the images. Specific preprocessing methods include, but are not limited to, data screening, data enhancement, and data standardization.

[0014] S2. Deep Learning Model Construction. Build a deep learning model for OCT choroidal large vessel segmentation. The specific implementation steps are as follows:

[0015] S2a. Segmentation model construction. You can choose to build any existing deep learning model structure suitable for image segmentation tasks or create a new deep learning model structure. The constructed model is denoted as f·, where Θ is the model parameter group. According to the structural characteristics, the models that can be constructed include but are not limited to: (1) CNN-based segmentation models, such as U-Net and UNet++; (2) Transformer-based segmentation models, such as SwinUNetR and UNetR; (3) Choroidal segmentation models based on multi-task learning, such as C-UNet and C-AttUNet.

[0016] S2b. Construction of the teacher model. Based on the model selected in the segmentation module, a model identical to it is selected as the teacher model to generate pseudo labels for the unlabeled samples required for semi-supervised learning. The constructed teacher model is denoted as f t (·, where Θ t is the parameter group of the teacher model.

[0017] S2c. Construction of integrated domain discrimination model. The network depth of the domain discriminator is determined according to the quality of the low-quality OCT images in S1c to better The domain described by the pixels in the image is discriminated and adversarial learning is performed. The constructed domain discrimination model is d(·), where Θ d is the parameter group of the domain discrimination model;

[0018] S2d. Learning module setup. Based on the number and quality of images collected in S1b and S1c, select the functional modules required for training. These modules include, but are not limited to, the following: parameter initialization module (often using initialization algorithms such as MSRA), model parameter update module (often using algorithms such as stochastic gradient descent), discriminator parameter update module (often using update algorithms such as Adam), teacher update module (often using exponential moving average), and learning rate update module (often using algorithms such as cosine periodic decay).

[0019] S2e. Hyperparameter Setup. Based on the deep learning model and its associated learning modules constructed in S2a, S2b, and S2c, and the training process in S3, set the hyperparameters required for model training. Configurable hyperparameters include, but are not limited to, batch size, learning rate, number of epochs, and loss function weights.

[0020] S3. Using the high-quality OCT image dataset obtained in S1 Its corresponding label set And unlabeled low-quality OCT image dataset The deep learning model is constructed in training S2. The specific implementation steps are as follows:

[0021] S3a. Parameter initialization. Use the parameter initialization method set in S2e to initialize the parameters Θ of the segmentation model in S2a.

[0022] S3b. Source domain data batch packaging. For the high-quality OCT images obtained in S1 and its tag set According to the batch size N set in S2 s Divided into several batches, each batch of data is represented as Contains N s High-quality OCT images and their annotations.

[0023] S3c. Target domain data batch packaging. For the low-quality OCT images obtained in S1c According to the batch size N set in S2 s Divided into several batches, each batch of data is represented as Contains N t Low-quality OCT images.

[0024] S3d. Segmentation model forward propagation. Using the image set obtained in S3b The image collection obtained in S3c As supervised training data set and semi-supervised data training set, respectively, they are input into the segmentation model constructed in S2. If the input image is The prediction result can be expressed as Similarly, if the input image is The prediction result can be expressed as

[0025]

[0026] S3e. Forward propagation of the teacher model. Using the image set obtained in S3c As a semi-supervised training dataset, input the segmentation model constructed in S2. If the input image is Then the prediction result can be expressed as Unlabeled data can then be obtained Pseudo labels

[0027] S3f. Forward propagation of the domain discrimination model. The source domain sample prediction results obtained in S3d are And the target domain teacher prediction results obtained in S3e It is passed into the domain discrimination model constructed in S2 to obtain the domain scores of the source domain and the target domain, that is, And based on and Get the discriminative sample mask M i .

[0028] S3g. Representative sample screening. Score the source domain generated in S3f and target domain integrated scoring Pass the confusion domain sample filter to generate a discriminative sample mask M i .

[0029] S3h. Segmentation model loss calculation. Define the loss function As the loss function used in model training:

[0030]

[0031] in, is the supervised segmentation loss, which can be expressed as:

[0032]

[0033] is the semi-supervised segmentation loss, which can be expressed as:

[0034]

[0035] is the domain adaptive segmentation loss, which can be expressed as:

[0036]

[0037] S3i. Update the parameters of the segmentation model. Use the parameter update algorithm selected in S2d according to The calculation results update the segmentation model parameters Θ.

[0038] S3j. Teacher model parameter update. Update the teacher model parameters Θ using the exponential sliding average of the segmentation model parameters t .

[0039] S3k. Calculation of loss function for integrated domain discrimination model. Definition As the loss function for the domain discrimination model.

[0040]

[0041] S3l. Update the parameters of the integrated domain discrimination model. Using the parameter update algorithm selected in S2d, according to The calculation results update the domain discrimination model parameter group Θ d

[0042] S3m. Repeat steps S3c to S3l until the batch is lost. The model converges to a stable minimum. The model parameter state at this point is the final trained deep learning model, which is used for choroidal small and medium-layer vascular angiography in low-quality OCT images.

[0043] S4. Model testing. Test the segmentation performance of the deep learning model trained in S3. The specific implementation steps are:

[0044] S4a. Test data preparation. The test data requires dense low-quality OCT image data collected according to module S1c. To evaluate the segmentation capability of the model, the test data also needs to generate a label set of some of the collected low-quality OCT images according to module S2b.

[0045] S4b. Low-quality OCT image segmentation. The test images are fed frame by frame into the trained deep learning model to obtain the segmentation results of the choroidal large vessels.

[0046] S4c. Choroidal angiography. For a given subject, the segmentation results of the medium and large choroidal vessels in all frames of a low-quality OCT dense scan are arranged in order and displayed as a whole or in layers using maximum intensity projection (maximum intensity projection is performed based on the medium or large vessel layer only).

[0047] Segmentation effect evaluation: For the low-quality OCT image segmentation results predicted in S4b, the predicted results of the annotated frames are compared with the manually annotated ones using the Dice coefficient for objective quantitative evaluation.

[0048] S4e. Evaluation of angiographic effects. The reconstruction effects of the choroidal angiography obtained in S4c were qualitatively compared with those of the clinical gold standard ICGA and other angiographic methods such as OCT choroidal projection and OCTA choroidal projection.

[0049] The overall principle diagram of the present invention is shown in the attached Figure 1 The supervised branch uses high-quality source domain OCT images and their corresponding manual annotations to supervise the segmentation model training; the semi-supervised branch generates corresponding annotations of unlabeled low-quality OCT images as pseudo-labels through the mean teacher model, and relies on a new type of confusion domain sample filter to filter the pseudo-labels to reduce the label noise caused by domain differences, and then uses the filtered low-quality OCT samples to perform semi-supervised training on the segmentation model; the domain adaptation branch uses a new type of integrated domain discriminator to identify the domain features of the source domain and the integrated target domain, and performs adversarial training on the segmentation model so that it can learn cross-domain features.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] (1) As a type of angiography, the present invention provides a non-invasive choroidal sublayer angiography method that does not rely on any contrast agent and can therefore be used for large-scale analysis and screening of choroidal vessels. Furthermore, the present invention can also provide three-dimensional vascular distribution of the choroidal sublayer, thereby being used to evaluate the choroidal sublayer structure and calculate three-dimensional biological markers such as vascular density and vascular index, thereby providing a more effective tool for the analysis and research of choroid-related diseases.

[0052] (2) As a domain adaptive segmentation method, the present invention proposes a new integrated domain discriminator, which simulates the generation mode of low-quality and high-quality images in medical images to prevent the interference caused by image quality drift during its training, so that it can be trained better and more stably, thereby assisting the adversarial learning process of the segmentation model.

[0053] (3) As a semi-supervised segmentation method, the present invention better considers the distribution difference between labeled images and unlabeled images, and proposes a new type of confusion domain sample filter to select representative samples with low pseudo-label noise and high segmentation difficulty, thereby reducing the label noise caused by domain differences in semi-supervised training and improving the accuracy of the segmentation model. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 It is a schematic diagram of the principle of the present invention.

[0055] Figure 2 It is a flow chart of a specific embodiment of the present invention.

[0056] Figure 3 FIG. 4 is a schematic diagram of the choroidal sublayer blood vessel segmentation result according to a specific embodiment of the present invention.

[0057] Figure 4 It is a schematic diagram of the choroidal angiography results according to a specific embodiment of the present invention.

[0058] Figure 5 It is a comparison chart of choroidal angiography results in a specific embodiment of the present invention. DETAILED DESCRIPTION

[0059] In order to make the objectives, technical solutions and advantages of the present invention more clear, the specific embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0060] Please refer to Figure 2The present invention provides a choroidal angiography method based on domain adaptation and semi-supervised segmentation. Specifically, the present invention trains a choroidal segmentation model based on high-quality OCT images and applies it to choroidal sublayer vascular segmentation using densely collected low-quality OCT images in clinical practice. The method then performs choroidal angiography using maximum intensity projection. The specific implementation steps include:

[0061] S1. Image acquisition and data annotation. The specific implementation steps are:

[0062] S1a. High-quality OCT image acquisition. Professional optometrists collected 32 OCT slices at equal intervals within a 6 × 6 mm area. Each slice was averaged 30 times at the same location to enhance image quality.

[0063] S1b. High-quality OCT image annotation. 88 OCT images obtained in S1a were selected and annotated by professional ophthalmologists to obtain position masks for the medium and large vessel layers, medium and large vessel segments, and large vessel segments.

[0064] S1c. Low-quality OCT image acquisition. Optometrists scanned 1024 OCT slices at even intervals within a 12 × 12 mm area. Each slice was averaged twice at the same location to increase imaging speed. 768 slices were randomly selected as unlabeled low-quality OCT images for training.

[0065] S1c. Image preprocessing. Considering the uniformity of data flow during model training and the limitation of video memory size, all images in the dataset obtained by the above operation are uniformly cropped to 448×448 pixels.

[0066] S2. Construct a deep learning model for choroidal segmentation. The hardware conditions for the implementation of the present invention are six NVIDIA RTX 3090 graphics cards, each equipped with 24GB of video memory. All the following networks and training modules are built using the Pytorch toolbox. Based on the specific tasks of the embodiment of the present invention and considering the limitations of hardware conditions, the specific implementation steps of the model construction are as follows:

[0067] S2a. Segmentation model construction. Considering both segmentation performance and operational efficiency, a U-Net deep neural network architecture is constructed here as the segmentation model. The encoder gradually reduces the size of the feature map while increasing the number of channels. This configuration allows the encoder to better extract high-level semantic features, including those affected by OCT slices. The decoder gradually increases the size of the feature map while decreasing the number of channels, ultimately restoring the original image size and ensuring accurate segmentation of edge details. Both the encoder and decoder contain five layers. Each encoder layer consists of three operations: two consecutive 3×3 convolutions + batch normalization + leaky ReLU activations, followed by a 2×2 max pooling with a stride of 2. Each decoder layer consists of four operations: a 2×2 max pooling with a stride of 2, a skip connection to a feature map symmetrically located in the encoder, and two consecutive 3×3 convolutions + batch normalization + leaky ReLU activations.

[0068] S2b. Construction of the teacher model. The construction of the teacher model is exactly the same as the construction and details of the segmentation model in S2a.

[0069] S2c. Construction of the integrated domain discriminator. Preferably, considering the amount of data and the difficulty of segmentation, the integrated domain discriminator constructed here consists of five convolutional layers with a kernel of 4×4, where the number of channels of these convolutional layers is: {64, 128, 256, 512, 1}.

[0070] S2d. Learning module setting: Preferably, Kaiming He initialization is set as the parameter initialization module, Adam is set as the parameter update module, and Adam gradient decay is set as the learning rate update module.

[0071] S2c. Hyperparameter settings. Preferably, the batch size is set to 4 and the initial learning rate of the segmentation model is set to 6×10 -4 The learning rate is halved every 10 epochs, and the number of epochs is set to 150. The initial learning rate of the domain discriminator model is 6×10 -5 .

[0072] S3. Using the high-quality OCT image dataset obtained in S1 Its corresponding label set And unlabeled low-quality OCT image dataset Train the deep learning model built in S2. The specific implementation steps are:

[0073] S3a. Parameter initialization. Use Kaiming He initialization algorithm to initialize the parameters Θ of the segmentation model.

[0074] S3b. Source domain data batch packaging. For the high-quality OCT images obtained in S1 and its annotations According to the batch size N B =4 is divided into 22 batches, each batch of data contains 4 high-quality OCT tomographic images and their corresponding annotations.

[0075] S3b. Target domain data batch packaging. For the low-quality OCT images obtained in S1 According to the batch size N B =4 is divided into 192 batches, each batch of data contains 4 low-quality OCT tomographic images.

[0076] S3d. Segmentation model forward propagation. Using the image set obtained in S3b As supervised training data set and semi-supervised data training set, respectively, they are input into the segmentation model constructed in S2. If the input image is Then the output image can be represented as Similarly, if the input image is Then the output image can be represented as

[0077] S3e. Forward propagation of the teacher model. The image pair set obtained in S3c As a semi-supervised training dataset, input the segmentation model constructed in S2. If the input image is Then the output image can be represented as Unlabeled data can then be obtained Pseudo labels

[0078] S3f. Forward propagation of the domain discrimination model. The source domain sample prediction results obtained in S3d are And the target domain teacher prediction results obtained in S3e It is passed into the domain discrimination model constructed in S2 to obtain the domain scores of the source domain and the target domain, that is, And based on and Get the discriminative sample mask M i .

[0079] S3h. Segmentation model loss calculation. Define the loss function As the loss function used in model training:

[0080]

[0081] in,

[0082]

[0083]

[0084]

[0085] S3e. Segmentation model parameter update: Use the Adam algorithm to update the segmentation model parameter group Θ.

[0086] S3j. Teacher model parameter update. Update the teacher model parameter group Θ using the exponential sliding average of Θ t .

[0087] S3k. Calculation of loss function for integrated domain discrimination model. Definition As the loss function for the domain discrimination model.

[0088]

[0089] S3l. Update the parameters of the integrated domain discrimination model. Using the Adam algorithm, according to The calculation results update the domain discrimination model parameter group Θ d

[0090] S3m. Repeat steps S3c to S3l until the batch is lost. The model parameter state at this time converges to a stable minimum. The final trained choroidal small and medium layer blood vessel segmentation model for low-quality OCT images is obtained.

[0091] S4. Model testing. Test the segmentation performance of the deep learning model trained in S3. The specific implementation steps are:

[0092] S4a. Test data preparation. Optometrists scanned four subjects with 1024 OCT slices at even intervals within a 12 × 12 mm area. Each slice was averaged twice at the same location to increase imaging speed. To evaluate the model's segmentation capabilities, 39 low-quality slices were randomly selected and annotated by professional ophthalmologists.

[0093] S4b. Low-quality OCT image segmentation. The images used for testing are input into the trained segmentation model frame by frame to obtain the segmentation results of the choroidal large blood vessels frame by frame. The sample results are shown in the attached Figure 3 middle.

[0094] S4c. Choroidal angiography. The results of choroidal medium and large vessels segmentation in all frames of the low-quality OCT dense scan of the subject in S4a were selected one by one, and they were arranged in order and displayed as a whole or in layers by maximum intensity projection (maximum intensity projection was performed only based on the medium or large vessel layer). The example results are shown in the attached figure. Figure 4 middle.

[0095] S4d. Segmentation Evaluation. For the low-quality OCT image segmentation results predicted in S4b, we compared the predicted results for the annotated partial frames with the manually annotated results using the Dice coefficient for objective quantitative evaluation. The relevant quantitative evaluation results are shown in Table 1. The optimal result is highlighted in bold. As can be seen from Table 1, the proposed method significantly outperforms current domain adaptive and semi-supervised segmentation methods in terms of sub-layer vessel segmentation.

[0096] S4e. Evaluation of imaging effect. The reconstruction effect of the choroidal angiography of the subjects obtained in S4c was qualitatively compared with the clinical gold standard ICGA and other imaging methods such as OCT choroidal projection and OCTA choroidal projection. Figure 4 As shown, the present invention can clearly observe the distribution of large blood vessels in the choroid, and this method does not require the injection of contrast agents. It is a non-invasive imaging method and can therefore be used for large-scale screening.

[0097] Table 1 Quantitative evaluation results of choroidal sublayer vessel segmentation

[0098]

[0099]

[0100] The above embodiments are only some preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any changes that adopt the design principles of the present invention and any changes made through non-creative work on this basis should fall within the scope of protection of the present invention.

[0101] References:

[0102] [1]Slakter JS,Yannuzzi LA,Guyer DR,et al.Indocyanine-greenangiography[J].Current opinion in ophthalmology,1995,6(3):25-32.

[0103] [2]Agrawal R, Ding J, Sen P, et al. Exploring choroidal angioarchitecture in health and disease using choroidal vascularity index[J]. Progress inretinal and eye research, 2020,77:100829.

[0104] [3]De Carlo T E,Romano A,Waheed N K,et al.A review of opticalcoherence tomography angiography(OCTA)[J].International journal of retina andvitreous,2015,1(1):1-15.

[0105] [4]Wang R K,Kirby M,Li C,et al.An explanation for why choroidal bloodvessels appear dark on clinical OCT images[J].Investigative Ophthalmology&Visual Science,2017,58(8):4754-4754.

[0106] [5]Sui X,Zheng Y,Wei B,et al.Choroid segmentation from opticalcoherence tomography with graph-edge weights learned from deep convolutionalneural networks[J].Neurocomputing,2017,237:332-341.

[0107] [6]He F,Chun R K M,Qiu Z,et al.Choroid Segmentation of Retinal OCTImages Based on CNN Classifier and l2-lq Fitter[J].Computational&MathematicalMethods in Medicine,2021.

[0108] [7]Zhang H,Yang J,Zhou K,et al.Automatic segmentation andvisualization of choroid in OCT with knowledge infused deep learning[J].IEEEJournal of Biomedical and Health Informatics,2020,24(12):3408-3420.

[0109] [8]Zhu L,Li J M,Zhu R,et al.Synergistically segmenting choroidallayer and vessel using deep learning for choroid structure analysis[J].Physics in Medicine&Biology,2022,67(8):085001.

[0110] [9]Ganin Y,Lempitsky V.Unsupervised domain adaptation bybackpropagation[C] / / International conference on machine learning.PMLR,2015:1180-1189.

[0111]

[10] Tsai Y H,Hung W C,Schulter S,et al.Learning to adapt structuredoutput space for semantic segmentation[C] / / Proceedings of the IEEE conferenceon computer vision and pattern recognition.2018:7472-7481.

[0112]

[11] Liu P,Kong B,Li Z,et al.CFEA:collaborative feature ensemblingadaptation for domain adaptation in unsupervised optic disc and cupsegmentation[C] / / International Conference on Medical Image Computing andComputer-Assisted Intervention.Springer,Cham,2019:521-529.

[0113]

[12] Spadotto T,Toldo M,Michieli U,et al.Unsupervised domainadaptation with multiple domain discriminators and adaptive self-training[C] / / 2020 25th International Conference on Pattern Recognition(ICPR).IEEE,2021:2845-2852.

[0114]

[13] Tranheden W,Olsson V,Pinto J,et al.Dacs:Domain adaptation viacross-domain mixed sampling[C] / / Proceedings of the IEEE / CVF Winter Conferenceon Applications of Computer Vision.2021:1379-1389.

[0115]

[14] Miyato T,Maeda S,Koyama M,et al.Virtual adversarial training:aregularization method forsupervised and semi-supervised learning[J].IEEEtransactions on pattern analysis and machineintelligence,2018,41(8):1979-1993.

[0116]

[15] Oliver A,Odena A,Raffel C,et al.Realistic evaluation of semi-supervised learningalgortihms[C] / / International conference on learningrepresentations.2018:1-15.

[0117]

[16] Lee D H.Pseudo-label:The simple and efficient semi-supervisedlearning method for deepneural networks[C] / / Workshop on challenges inrepresentation learning,ICML.2013,3(2):896.

[17] Chen X,Yuan Y,Zeng G,etal.Semi-supervised semantic segmentation with cross pseudosupervision[C] / / Proceedings of the IEEE / CVF Conference on Computer Vision andPatternRecognition.2021:2613-2622.

[0118]

[18] Sohn K,Berthelot D,Carlini N,et al.Fixmatch:Simplifying semi-supervised learning withconsistency and confidence[J].Advances in neuralinformation processing systems,2020,33:596-608。

Claims

1. A method for optical coherence tomography choroidal vascular layer angiography, characterized in that: The specific steps include: S1. Image acquisition and data annotation; S1a. Use OCT image acquisition equipment to collect OCT images at K key scanning positions of N subjects and form the source domain dataset in Represents the color intensity of a frame in a high-quality OCT tomography scan; S1b. High-quality OCT key-frame tomographic image set The choroidal middle vessel layer, choroidal large blood layer, choroidal middle vessel segment, and choroidal large vessel segment are finely labeled pixel by pixel and verified, and the source domain label set is constructed based on this annotation in, Represents X i The choroid sublayer annotation results, Represents X i The choroidal vascular segment annotation results; S1c. Use OCT image acquisition equipment to perform fast and dense scanning of N subjects with M frames, and form the target domain dataset where X i Represents the color intensity of a frame in an OCT tomography scan; S1d. Perform certain preprocessing on the image according to the characteristics of the acquired data; S2. Build a deep learning model for choroidal segmentation. The specific steps for model construction are as follows: S2a. Build a segmentation model. Let the constructed model be f(·) and Θ be the parameter set of the model. S2b. Build the teacher model. The construction of the teacher model is exactly the same as the construction and details of the segmentation model in S2a. The constructed teacher model is denoted as f t (·), Θ t is the parameter group of the teacher model; S2c. Build an integrated domain discriminator and denote the constructed domain discrimination model as d(·), Θ d is the parameter group of the domain discrimination model; S2d. Learning module settings: select the parameter update algorithm required for training based on the number and quality of images collected in S1b and S1c; S2e. Set the hyperparameters required for model training based on the deep learning model and its related learning modules constructed in S2a, S2b, and S2c, and the training process in S3. S3. Using the high-quality OCT image dataset obtained in S1 Its corresponding label set And unlabeled low-quality OCT image dataset The deep learning model is constructed in training S2. The specific implementation steps are as follows: S3a. Initialize the parameters Θ of the segmentation model in S2a; S3b. For high-quality OCT images and its tag set Divide into several batches, each batch of data is represented as Contains N s High-quality OCT images and their annotations; S3c. For low-quality OCT images Divide into several batches, each batch of data is represented as Contains N t Low-quality OCT images; S3d. Image collection As supervised training data set and semi-supervised data training set respectively, they are input into the segmentation model constructed in S2. If the input image is The output image is represented as Similarly, if the input image is The output image is represented as S3e. Image pair collection As a semi-supervised training dataset, input the segmentation model constructed in S2. If the input image is The output image is represented as Then get unlabeled data Pseudo labels S3f. Prediction results of source domain samples obtained in S3d And the target domain teacher prediction result P obtained in S3e i tt It is passed into the domain discrimination model constructed in S2 to obtain the domain scores of the source domain and the target domain, that is, And based on and P i tt Get the discriminative sample mask M i ; S3h. Define the loss function As the loss function used in model training: in, S3i. Using the parameter update algorithm selected in S2d, according to The calculation results update the segmentation model parameters Θ; S3j. Update the teacher model parameters Θ using the exponential moving average of the segmentation model parameters t ; S3k.Definition As the loss function of the domain discrimination model: S3l. Use the parameter update algorithm selected in S2d, according to The calculation results update the domain discrimination model parameters Θ d ; S3m. Repeat steps S3c to S3l until the batch is lost. Converge to a stable minimum, the state of the model parameters at this time is the final trained deep learning model; S4. Input the test images frame by frame into the trained deep learning model to obtain the choroidal vascular layer angiography results frame by frame.

2. The optical coherence tomography choroidal vascular laminography method according to claim 1, wherein: Step S1 performs certain preprocessing on the image. Specific preprocessing methods include but are not limited to: data screening, data enhancement, and data standardization.

3. The optical coherence tomography choroidal vascular laminography method according to claim 1, wherein: The learning module settings include but are not limited to: the parameter initialization module selects the MSRA initialization algorithm, the model parameter update module selects the stochastic gradient descent algorithm, the discriminator parameter update module selects the Adam update algorithm, the teacher update module selects the exponential moving average algorithm, and the learning rate update module selects the cosine periodic decay algorithm.

4. The optical coherence tomography choroidal vascular laminography method according to claim 1, wherein: The hyperparameters that can be set include but are not limited to: batch size, learning rate, number of epochs, and loss function weight.

5. The optical coherence tomography choroidal vascular laminography method according to claim 1, wherein: The segmentation model includes but is not limited to: a CNN-based segmentation model, a Transformer-based segmentation model, or a choroidal segmentation model based on multi-task learning.

6. The optical coherence tomography choroidal vascular laminography method according to claim 1, wherein: The S4 choroidal vascular layered angiography results were arranged in order and displayed overall or layered by maximum intensity projection.

Images