Blood vessel segmentation method for ultra-wide-angle fundus images assisted by blood vessel type information
Through the method assisted by vascular type information, the U-Net model was used to segment the thick blood vessels and thin blood vessels in the ultra-wide fundus image, and the results were combined to solve the problem of low segmentation accuracy caused by sample imbalance and low contrast, achieving higher segmentation accuracy and information acquisition.
Patent Information
- Application Number
- CN202210521294.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-05-13
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2042-05-13
AI Technical Summary
The uneven sample of thick blood vessels and thin blood vessels in ultra-wide fundus images and the low contrast between thin blood vessels and background leads to the problem of low vascular segmentation accuracy.
By introducing blood vessel type information, two U-Net models were used to segment the coarse blood vessels and thin blood vessels respectively, and the results were combined to increase the weight of blood vessel pixels in the loss function, image augmentation and contrast enhancement processing were used, combined with morphological methods and Model transfer learning to increase receptive field to improve segmentation accuracy.
It significantly improves the segmentation accuracy of the leucorrhea in ultra-wide-angle fundus images, especially shows good application value on the leucorrhea, obtains higher accuracy of vascular information acquisition and segmentation, and improves the Dice coefficient and accuracy.
Smart Images

Figure CN114897832B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of ultra-wide-angle fundus imaging, and in particular to an ultra-wide-angle fundus image blood vessel segmentation method assisted by blood vessel type information. Background Art
[0002] Retinal vascular structure information helps ophthalmologists detect and diagnose a variety of fundus diseases, such as diabetic retinopathy, hypertensive retinopathy, etc. As recorded in the literature [1] Li Tao, Bo Wang, Hu Chunyu, et al. Applications of deep learning in fundus images: A review. Medical Image Analysis, 2021, 69: 101971. Manual segmentation of retinal blood vessels is inefficient and costly, while the automatic segmentation of fundus blood vessels using AI technology can save time and effort and its accuracy will not be affected by subjective factors. It has attracted the attention of many researchers in recent years.
[0003] Reference [2] Chinese patent "Vessel segmentation method based on self-attention mechanism and convolutional neural network hybrid model" (CN113902757A) uses a convolutional neural network based on a continuous residual structure, channel attention, and spatial attention mechanism to fully extract the low-level features of fundus vascular images, and then uses the self-attention mechanism to associate the high-level features extracted by the convolutional neural network to help the hybrid model identify blood vessels.
[0004] Reference [3] Ren Zihui, Cai Manli, Miao Xiaobo, et al. Retinal vessel segmentation algorithm based on full-scale skip connection. Science Technology and Engineering, 2022, 22(07): 2776-2783. Variant convolution is used to replace the ordinary convolution in the U-Net3+ model to capture the characteristics of thin blood vessels. At the same time, a scale feature fusion module with a channel attention mechanism is introduced to improve the feature fusion capability of the U-Net3+ model.
[0005] Reference [4] Chinese patent "A retinal vessel segmentation method based on fusion of multiple low-level features" (CN113781514A) extracts several low-level feature images from fundus images, such as grayscale, color, intensity, edge, morphology, texture and multi-scale linear features. The low-level features can reflect the detailed characteristics of the blood vessels. These low-level feature images are then input into the retinal vessel segmentation model together with the color fundus image for vessel segmentation training. Due to the imbalance of samples of thick and thin blood vessels in fundus images and the low contrast between thin blood vessels and background, the semantic segmentation method based on deep learning technology still has difficulties in segmenting tiny blood vessels consisting of only a few pixels, resulting in low overall vessel segmentation accuracy.
[0006] Compared with traditional fundus imaging, ultra-wide-angle fundus imaging technology has the advantages of large field of view, fast imaging, and no mydriasis, and is gradually gaining widespread attention from researchers. At present, there is still little research on the segmentation of blood vessels in ultra-wide-angle fundus images. Reference [5] Wei Feng, Lie Ju, Lin Wang, et al. Unsupervised Domain Adaptation for Retinal Vessel Segmentation with Adversarial Learning and Transfer Normalization [J]. arXiv preprint arXiv: 2108.01821, 2021. proposed an entropy-based adversarial training strategy for cross-domain retinal vessel segmentation, which showed certain advantages in the migration task between conventional fundus datasets and ultra-wide-angle fundus datasets. Reference [6] Li Ding, Ajay E Kuriyan, Rajeev S Ramchandran, et al. Weakly-supervised vessel detection in ultra-widefield fundus photography via iterative multi-modal registration and learning [J]. IEEE Transactions on Medical Imaging, 2020, 40(10): 2748-2758. When segmenting retinal vessels, the ultra-wide-angle fundus image and ultra-wide-angle fluorescein angiography image collected simultaneously were used, and the multi-modal registration step and the weakly supervised learning step of the U-Net model were iterated to obtain a highly accurate ultra-wide-angle fundus vessel segmentation result. However, fluorescein angiography images are difficult to obtain quickly because the fluorescent agent takes some time to enter the fundus blood vessels, and some patients are allergic to them, making them unsuitable for rapid screening of fundus vascular diseases. Summary of the invention
[0007] In view of the problem that the samples of thick and thin blood vessels in fundus images are unbalanced and the low contrast between thin blood vessels and background leads to low precision of thin blood vessel segmentation, the present invention provides a blood vessel segmentation method for ultra-wide-angle fundus images assisted by blood vessel type information. The method introduces blood vessel type information to extract and segment the thick and thin blood vessels in the ultra-wide-angle fundus image respectively, and then fuses the results of the two to obtain a complete ultra-wide-angle fundus blood vessel segmentation image.
[0008] The technical solution adopted by the present invention is:
[0009] The blood vessel segmentation method of ultra-wide-angle fundus image assisted by blood vessel type information comprises the following steps:
[0010] Step 1: Extract the vascular region of interest;
[0011] Step 2: Perform image augmentation and image contrast enhancement on the ultra-wide-angle fundus image, and divide the image into blocks;
[0012] Step 3: Divide the blood vessels into thick blood vessels and thin blood vessels for blood vessel segmentation;
[0013] Step 4: Use two U-Net models to perform thick blood vessel segmentation and thin blood vessel segmentation respectively, and finally fuse the segmentation results of the two U-Net models to obtain a complete ultra-wide-angle fundus blood vessel segmentation image.
[0014] The step 1 comprises the following steps:
[0015] S1.1: First, in the ultra-wide-angle fundus image dataset, an area 2 to 3 times the size of the optic disc is annotated as a labeling box to obtain a training set;
[0016] S1.2: Then train the target detection network YOLOv4 to locate the optic disc area;
[0017] S1.3: Based on the acquired positioning information of the optic disc area, 12 circular areas were extracted with the optic disc as the center and combined into a region of interest with a size of 1120×1120.
[0018] The step 2 comprises the following steps:
[0019] S2.1: Select the green channel image of the ultra-wide-angle fundus image and perform CLAHE processing to highlight blood vessels and reduce noise;
[0020] S2.2: Increase the amount of data required for model training by image segmentation, where the image block size for training the U-Net model for segmenting thick blood vessels is set to 48×48, and the image block size for training the U-Net model for segmenting thin blood vessels is set to 96×96.
[0021] The step 3 comprises the following steps:
[0022] S3.1: First, extract the vascular skeleton from the ultra-wide-angle fundus vascular label, and calculate the ratio of the vascular skeleton to the original vascular area in a sliding window of 8×8 size, which is used as an indicator to measure the thickness of the blood vessels;
[0023] S3.2: Then, a threshold is used to classify the vascular skeleton into two categories: thick vascular skeleton and thin vascular skeleton;
[0024] S3.3: Then compare the sizes of thick and thin blood vessel pixels in the 8-neighborhood of other blood vessel skeleton pixels. If the number of thick blood vessel pixels is greater than that of thin blood vessel pixels, the blood vessel pixel is a thick blood vessel pixel; otherwise, it is a thin blood vessel pixel. This process is repeated 40 times.
[0025] S3.3: Finally, to address the problem of thick and thin blood vessel plaques in the cross-fragmentation area of thick and thin blood vessels, a morphological method is used to classify the parts of thick blood vessels with a connectivity area less than 50 as thin blood vessels, and the parts of thin blood vessels with a connectivity area less than 50 as thick blood vessels. This is iterated 5 times to obtain the final thick and thin blood vessel classification results.
[0026] In step 4, before the thick blood vessel segmentation and the thin blood vessel segmentation, the model migration method is used to first pre-train the thick blood vessel segmentation model and the thin blood vessel segmentation model using traditional fundus images, and then the model is migrated to the ultra-wide-angle fundus image domain to fine-tune the model parameters and learn the blood vessel characteristics.
[0027] The present invention provides a method for blood vessel segmentation of ultra-wide-angle fundus images assisted by blood vessel type information, and the technical effects are as follows:
[0028] 1) Compared with the U-Net model in the literature [7] Xu Guangzhu, Lin Wenjie, Chen Sha, et al. Fundus vascular segmentation method combining U-Net with adaptive threshold pulse coupled neural network. Computer Applications, 2022, 42(3): 825-832, the difference of the present invention is:
[0029] ①: For the U-Net model for segmenting thick blood vessels, due to the imbalance in the number of thick blood vessel pixels and background pixels, the present invention increases the weight of thick blood vessel pixels in the loss function based on the binary cross entropy loss function, and balances the influence of thick blood vessels and background on the U-Net model;
[0030] ②: For the U-Net model for segmenting fine blood vessels, the present invention replaces the ordinary convolution operation in the U-Net model with a dilation rate of 1 dilated convolution, and increases the training block size from 48×48 to 96×96. This is done to increase the receptive field, obtain more vascular structure information, and assist in the segmentation of fine blood vessels. At the same time, due to the imbalance in the number of fine blood vessel pixels and background pixels, the present invention also increases the weight of fine blood vessel pixels in the loss function.
[0031] 2) Aiming at the problem of unbalanced samples of thick and thin blood vessels in fundus images and low contrast between thin blood vessels and background leading to low segmentation accuracy of thin blood vessels, the present invention provides a method for blood vessel segmentation of ultra-wide-angle fundus images with the assistance of blood vessel type information. This method uses a thick and thin blood vessel classification algorithm to classify blood vessel types into thick and thin blood vessels, and then uses two pre-trained U-Net models to perform thick and thin blood vessel segmentation respectively, and finally fuses the results of the two to obtain an ultra-wide-angle fundus blood vessel segmentation image. Compared with the widely used blood vessel segmentation:
[0032] Document [8]: Ronneberger Olaf, Fischer Philipp, Brox Thomas. U-net: Convolutional networks for biomedical image segmentation. International Conference on Medical image computing and computer-assisted intervention, 2015: 234-241. Recorded U-Net;
[0033] Reference [9]: Chen Liang-Chieh, Papandreou George, Schroff Florian, et al. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. DeepLabV3, and
[0034] Compared with the SegNet model in literature
[10] : Badrinarayanan Vijay, Kendall Alex, Cipolla Roberto. Segnet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE transactions on pattern analysis and machine intelligence, 2017, 39(12): 2481-2495., the vascular segmentation algorithm of the present invention has obvious advantages and can obtain more vascular information, especially on fine blood vessels, and has good application value. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] Figure 1(a) is a conventional fundus image;
[0036] Figure 1(b) is an ultra-wide-angle fundus image.
[0037] Figure 2 This is the flow chart of blood vessel segmentation in ultra-wide-angle fundus images.
[0038] Figure 3 Flowchart for vascular region of interest extraction.
[0039] Figure 4 Schematic diagram of image contrast enhancement and image augmentation.
[0040] Figure 5 Schematic diagram of the thick and thin blood vessel classification process in ultra-wide-angle fundus images.
[0041] Figure 6 Schematic diagram of model transfer learning method.
[0042] Figure 7 This is a diagram of the U-Net network structure used for fundus blood vessel segmentation.
[0043] Figure 8 This is the flow chart of ultra-wide-angle fundus blood vessel segmentation.
[0044] Fig. 9 This is the result of the ultra-wide-angle fundus coarse and fine blood vessel segmentation, where:
[0045] a The unit images contained in the column represent the original images;
[0046] b Columns contain cell maps representing thick blood vessel labels;
[0047] The cell images contained in column c represent the result images of thick blood vessel segmentation;
[0048] d The cell maps contained in the vertical column represent the capillary label maps;
[0049] The unit images contained in column e represent the result images of fine blood vessel segmentation;
[0050] The cell maps contained in column f represent the vascular labeling maps.
[0051] The unit graphs contained in the g column represent the fusion result graph.
[0052] Fig.10 This is a comparison of the results of the ultra-wide-angle fundus image blood vessel segmentation method, where:
[0053] The unit images contained in the h column represent the original image;
[0054] The cell graph contained in column i represents the label graph;
[0055] The unit graph contained in the j column represents the blood vessel segmentation result graph of the U-Net model;
[0056] The unit graph contained in the k column represents the blood vessel segmentation result graph of the DeepLabv3 model;
[0057] The unit graphs contained in the l column represent the blood vessel segmentation result graph of the SegNet model;
[0058] The unit graphs contained in the m columns represent the blood vessel segmentation result graphs of the method of the present invention.
[0059] Fig.11 This is a quantitative comparison chart of the optic disc area extraction results of the method proposed in the present invention and other methods.
[0060] Fig.12 This is a localization map of the optic disc area. DETAILED DESCRIPTION
[0061] The present invention is directed to the application of rapid fundus vascular health status screening based on ultra-wide-angle fundus images. In order to solve the problem of unbalanced samples of thick and thin blood vessels in ultra-wide-angle fundus images and low contrast between thin blood vessels and background leading to low segmentation accuracy of thin blood vessels, the thick and thin blood vessels in ultra-wide-angle fundus images are processed separately by introducing vascular type information to reduce the mutual influence. The main process is: first, based on the acquired optic disc position information, the region of interest covering the standard 7-field of view is extracted with the optic disc as the center; then the blood vessels are divided into thick and thin blood vessels, and the thick blood vessel segmentation model and the thin blood vessel segmentation model are pre-trained using traditional fundus image datasets, and then the models are migrated to the ultra-wide-angle fundus image domain for vascular segmentation training; finally, the vascular segmentation results of the two models are fused to obtain the vascular segmentation image of the ultra-wide-angle fundus image. Figure 2 As shown, the specific steps include:
[0062] Step 1: In order to detect a smaller optic disc area more robustly, the present invention first marks an area 2-3 times the size of the optic disc as a labeling box in the ultra-wide-angle fundus image to obtain a training set. Then, the target detection network YOLOv4 is trained to achieve accurate positioning of the optic disc area, such as Fig.12 The YOLOv4 model uses the CSPDarknet53 module as the backbone network to extract the deep semantic features of the optic disc in ultra-wide-angle fundus images, then uses spatial pyramid pooling and path aggregation network to fuse the multi-scale features of the optic disc, and finally uses the multi-scale output module to obtain the optic disc position information.
[0063] Next, in order to fully exploit the advantage of the wide shooting range of the ultra-wide-angle fundus image, while taking into account the subsequent migration problem of the vascular segmentation model based on the traditional fundus image, the present invention extracts 12 circular areas with the optic disc as the center based on the acquired optic disc position information, and combines them into a region of interest with a size of 1120×1120. The combination process can be regarded as a smoothing filtering process to filter out part of the image noise. The combined image is a large area with the optic disc as the center and the range far exceeds the standard 7-field fundus image, such as Figure 3 shown.
[0064] Step 2: To address the problem of insufficient data and poor quality of ultra-wide-angle fundus images, the present invention uses image augmentation and image contrast enhancement. The present invention selects the green channel image of the ultra-wide-angle fundus image and performs CLAHE processing to highlight blood vessels and reduce noise. At the same time, the present invention increases the amount of data required for model training by dividing the image into blocks, such as Figure 4 As shown in the figure, the image block size of the U-Net model used to train the segmentation of thick blood vessels is set to 48×48, while the image block size of the U-Net model used to train the segmentation of thin blood vessels is set to 96×96.
[0065] Step 3: To address the problem of unbalanced samples of thick and thin blood vessels in fundus images, as well as the low contrast between thin blood vessels and the background leading to low segmentation accuracy of thin blood vessels, the present invention divides blood vessels into thick and thin blood vessels to assist in the subsequent blood vessel segmentation. Figure 5 As shown, the details are as follows:
[0066] First, the sternum skeleton is extracted from the ultra-wide-angle fundus vascular labels, and the ratio of the vascular skeleton to the original vascular area is calculated in an 8×8 sliding window, which is used as an indicator to measure the thickness of the blood vessels.
[0067] Then, the threshold is used to classify the vascular skeleton into two categories: thick vascular skeleton and thin vascular skeleton;
[0068] Next, the sizes of thick and thin blood vessel pixels in the 8-neighborhood of other blood vessel skeleton pixels are compared. If the number of thick blood vessel pixels is greater than that of thin blood vessels, the blood vessel pixel is a thick blood vessel pixel, otherwise it is a thin blood vessel pixel. This process is repeated 40 times.
[0069] Finally, to address the problem of thick and thin blood vessel plaques in the cross-fragmentation area of thick and thin blood vessels, a morphological method was used to classify the parts of thick blood vessels with a connectivity domain less than 50 as thin blood vessels, and the parts of thin blood vessels with a connectivity domain less than 50 as thick blood vessels. This was iterated 5 times to obtain the final thick and thin blood vessel classification results.
[0070] Step 4: To address the problem of lack of ultra-wide-angle fundus image data, the present invention uses a model migration method, such as Figure 6As shown, the thick vessel segmentation model and the thin vessel segmentation model are first pre-trained using traditional fundus images, and then the model is migrated to the ultra-wide-angle fundus image domain to fine-tune the model parameters and learn the vascular features.
[0071] Pre-training refers to directly inputting traditional fundus images into the coarse vessel segmentation model and the fine vessel segmentation model. The encoder modules of the two models extract vascular features at different levels, which are fused in the decoder module through skip connections, and finally a 3×3 convolution kernel is used to obtain the vascular probability map. Then, the model uses the cross entropy loss function to calculate the error between the predicted result and the vascular label, and then uses stochastic gradient descent to minimize the cost, and then back-propagates to update the weights and biases.
[0072] Fine-tuning the model parameters and learning vascular features refers to: in the ultra-wide-angle fundus image domain, adjusting the parameters of the optimization function Adam function used in model training, reducing its learning rate parameter from the pre-trained 0.001 to 0.00001, and setting the parameters for calculating the gradient average and its square to 0.9 and 0.999, respectively.
[0073] Step 5: The present invention addresses the problem of low accuracy of thin vessel segmentation caused by the imbalance of samples of thick and thin vessels in fundus images and the low contrast between thin vessels and background. The blood vessels are divided into two types: thick and thin vessels. Then two U-Net models are used to perform thick and thin vessel segmentation respectively. Finally, the segmentation results of the two models are fused to obtain a complete ultra-wide-angle fundus vessel segmentation image.
[0074] Two U-Net models are used to perform thick blood vessel segmentation and thin blood vessel segmentation respectively, as follows:
[0075] During the training process of the model, 90% of the image blocks are used to learn the vascular features of the model, and 10% of the data is used to verify the parameters of the model. The 48×48 image blocks are directly input into the thick vessel segmentation model, and the 96×96 image blocks are input into the thin vessel segmentation model to obtain the thick vessel probability image and the thin vessel probability image respectively. Then, the model uses the cross entropy loss function to calculate the error between the predicted result and the vascular label, and then uses stochastic gradient descent to minimize the cost, and then back-propagates to update the weights and biases.
[0076] The segmentation results of the two models are combined to obtain a complete ultra-wide-angle fundus vascular segmentation image, as follows:
[0077] The thick vessel probability image is directly binarized using a hard threshold of 0.35 to obtain a thick vessel binary image, while the thin vessel probability image is binarized using a hard threshold of 0.15 to obtain a thin vessel binary image. Then, the two images are ORed, that is, as long as one of the two images has a pixel at the same position as a vessel pixel, the pixel is determined to be a vessel pixel, and the final ultra-wide-angle fundus vessel segmentation image is obtained.
[0078] In order to solve the problem of unbalanced samples of thick and thin blood vessels in ultra-wide-angle fundus images and low contrast between thin blood vessels and background leading to low segmentation accuracy of thin blood vessels, the present invention divides blood vessels into two types: thick blood vessels and thin blood vessels. Then, two U-Net models are used to perform thick blood vessel segmentation and thin blood vessel segmentation respectively. For the thick blood vessel segmentation model, the weight of thick blood vessel pixels in the loss function is increased. For the thin blood vessel segmentation model, the image block is enlarged to 96×96 during training, and a dilation rate of 1 is used for the dilation convolution. At the same time, the weight of thin blood vessel pixels in the loss function is increased. Finally, the segmentation results of the two are fused to obtain a complete ultra-wide-angle fundus image blood vessel segmentation result, such as Fig. 9 As shown in the figure, using thick and thin blood vessel type information to assist fundus vessel segmentation is more effective than using thick and thin blood vessel type information to retain the information of micro-vessels and reduce the number of misclassified pixels.
[0079] like Fig.10 As shown in the figure, from the blood vessel segmentation results of the ultra-wide-angle fundus image, compared with the U-Net, DeepLabV3 and SegNet models that are widely used in blood vessel segmentation, the ultra-wide-angle fundus blood vessel segmentation algorithm of the present invention performs the best and can segment most of the blood vessels, especially only a very small number of thin blood vessels that are difficult to discern in the original image.
[0080] like Fig.11 As shown, the method of the present invention is better than the direct use of U-Net model, DeepLabV3 model, and SegNet model for ultra-wide-angle fundus image vascular segmentation results. Except for sensitivity, the four quantitative indicators of specificity, precision, accuracy and Dice coefficient are higher than the vascular segmentation results of U-Net, DeepLabv3, and SegNet models, especially the precision and Dice coefficient have obvious advantages.
Claims
1. The blood vessel segmentation method of ultra-wide-angle fundus image assisted by blood vessel type information is characterized by The following steps are involved: Step 1: Extract the vascular region of interest; Step 2: Perform image augmentation and image contrast enhancement on the ultra-wide-angle fundus image, and divide the image into blocks; Step 3: Divide the blood vessels into thick blood vessels and thin blood vessels for blood vessel segmentation; Step 4: Use two U-Net models to perform thick blood vessel segmentation and thin blood vessel segmentation respectively, and finally fuse the segmentation results of the two U-Net models to obtain a complete ultra-wide-angle fundus blood vessel segmentation image; Step 3 The following steps are involved: S3.1: First, extract the vascular skeleton from the ultra-wide-angle fundus vascular label, and calculate the ratio of the vascular skeleton to the original vascular area in the sliding window, which is used as an indicator to measure the thickness of the blood vessel; S3.2: Then, a threshold is used to classify the vascular skeleton into two categories: thick vascular skeleton and thin vascular skeleton; S3.3: Then, the sizes of thick and thin blood vessel pixels in the 8-neighborhood of other blood vessel skeleton pixels are compared. If the number of thick blood vessel pixels is greater than that of thin blood vessel pixels, the blood vessel pixel is a thick blood vessel pixel; otherwise, it is a thin blood vessel pixel, and this process is repeated multiple times. S3.3: Finally, for the problem of thick and thin blood vessel plaques in the cross-fragmentation area of thick and thin blood vessels, a morphological method is used to classify the parts of thick blood vessels with a connected area less than 50 as thin blood vessels, and the parts of thin blood vessels with a connected area less than 50 as thick blood vessels. This is iterated 5 times to obtain the final thick and thin blood vessel classification results; In step 4, during the training of the model, 90% of the image blocks are used to learn the vascular features of the model, and 10% of the data are used to verify the parameters of the model; the image blocks of size 48×48 are directly input into the thick vessel segmentation model, and the image blocks of size 96×96 are input into the thin vessel segmentation model to obtain thick vessel probability images and thin vessel probability images respectively; then, the model uses the cross entropy loss function to calculate the error between the prediction result and the vascular label, and then uses the stochastic gradient descent method to minimize the cost, and then back-propagates to update the weights and biases; For the thick blood vessel segmentation model, the weight of the thick blood vessel pixels in the loss function is increased; for the thin blood vessel segmentation model, the image block is increased to 96×96 during training, and a dilation rate of 1 is used for the hole convolution, and the weight of the thin blood vessel pixels in the loss function is increased.
2. The method for blood vessel segmentation of ultra-wide-angle fundus images assisted by blood vessel type information according to claim 1, characterized in that: The step 1 comprises the following steps: S1.1: First, an area is marked as a labeling box with the optic disc as the center in the ultra-wide-angle fundus image dataset to obtain a training set; S1.2: Then train the target detection network YOLOv4 to locate the optic disc area; S1.3: Based on the acquired positioning information of the optic disc area, a circular area is extracted with the optic disc as the center and combined into a region of interest.
3. The method for blood vessel segmentation of ultra-wide-angle fundus images assisted by blood vessel type information according to claim 1, characterized in that: The step 2 comprises the following steps: S2.1: Select the green channel image of the ultra-wide-angle fundus image and perform CLAHE processing to highlight blood vessels and reduce noise; S2.2: Increase the amount of data required for model training by image segmentation, where the image block size for training the U-Net model for segmenting thick blood vessels is set to 48×48, and the image block size for training the U-Net model for segmenting thin blood vessels is set to 96×96.
4. The method for blood vessel segmentation of ultra-wide-angle fundus images assisted by blood vessel type information according to claim 1, characterized in that: In step 4, before the thick blood vessel segmentation and the thin blood vessel segmentation, the model migration method is used to first pre-train the thick blood vessel segmentation model and the thin blood vessel segmentation model using traditional fundus images, and then the model is migrated to the ultra-wide-angle fundus image domain, the model parameters are fine-tuned and the blood vessel features are learned.
