Automatic detection method and system for optic disc neovascularization based on bimodal optical scanning

Abstract

The present application relates to the technical field of medical image processing, in particular to a method and system for automatic detection of optic disc neovascularization based on bimodal optical scanning, comprising the following steps: step S1, scanning around the optic disc of a single eye and outputting N continuous bimodal optical scanning images containing the boundary of the optic disc, and establishing a data set containing the boundary label of the optic disc according to the N frequency domain optical coherence tomography images; step S2, training, verifying and testing the data set using a convolution network to obtain the detection result of the boundary of the optic disc; step S3, screening out the candidate area of optic disc neovascularization on each frequency domain optical coherence tomography image and optical coherence tomography angiography image; step S4, screening out the optic disc neovascularization area in each candidate area of optic disc neovascularization; and step S5, calculating the volume of the optic disc neovascularization of a single eye.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image processing, in particular to a disc neovascularization automatic detection method and system based on dual-mode optical scanning. BACKGROUND

[0002] Disc neovascularization (NVD) is a new blood vessel that appears in the disc area, commonly seen in proliferative diabetic retinopathy (PDR), and is one of the main causes of severe vision loss in patients. According to reports, PDR patients with NVD have a high risk of vision loss. Hemorrhage caused by NVD accounts for 85% of PDR hemorrhage. Half of the NVD patients without treatment will be blind within 3 years after the initial diagnosis. Therefore, early diagnosis and timely treatment of NVD are crucial to reduce the risk of vision loss. Frequency domain optical coherence tomography (SD-OCT) and optical coherence tomography angiography (OCTA) are two non-invasive imaging techniques that can display the structure of the disc from a three-dimensional perspective, which helps clinicians diagnose and evaluate the efficacy of NVD. The existing OCTA equipment can provide SD-OCT and OCTA disc images at the same time, which provides a data basis for combining multi-modal images for NVD detection. Automatic and accurate NVD detection can help clinicians quickly assess and analyze NVD, which is of great significance.

[0003] The existing algorithms in the prior art are based on single-mode images for NVD detection, such as fundus images or OCTA images. In the method of detecting NVD based on fundus images, the fundus images cannot obtain the blood flow information of NVD, so the detection algorithm is complex and the accuracy is low; in the method of detecting NVD based on OCTA images, it can obtain the blood flow information of NVD, but the morphological features of NVD are not obvious enough, so the detection of NVD often does not achieve ideal results. Moreover, most of these methods manually design the disc features and NVD features, and the generalization and robustness of the detection results need to be improved, especially for images with small NVD regions or containing noise, which cannot obtain good detection results. SUMMARY

[0004] The present application also aims to provide a disc neovascularization automatic detection method and system based on dual-mode optical scanning, which solves the above technical problems.

[0005] The disc neovascularization automatic detection method based on dual-mode optical scanning comprises the following steps:

[0006] Step S1, scanning and outputting N continuous bimodal optical scanning images containing the optic disc boundary of a single eye, the bimodal optical scanning images including frequency domain optical coherence tomography images and optical coherence tomography angiography images, establishing a dataset containing optic disc boundary annotations according to N frequency domain optical coherence tomography images, N being a natural number greater than 1;

[0007] Step S2, training, verifying and testing the dataset using a convolutional network to obtain optic disc boundary detection results;

[0008] Step S3, screening out optic disc neovascularization candidate regions on each of the frequency domain optical coherence tomography images and the optical coherence tomography angiography images according to the optic disc boundary detection results;

[0009] Step S4, screening out optic disc neovascularization regions in each of the optic disc neovascularization candidate regions;

[0010] Step S5, calculating the volume of the optic disc neovascularization of the single eye according to all the optic disc neovascularization regions.

[0011] Preferably, step S2 comprises:

[0012] Step S21, a feature extraction module of the convolutional network extracts optic disc boundary features of the frequency domain optical coherence tomography images using multiple cascaded down-sampling modules;

[0013] Step S22, an attention module of the convolutional network refines the optic disc boundary features;

[0014] Step S23, a feature fusion module of the convolutional network fuses the optic disc boundary features of different scales, and outputs optic disc boundary segmentation results after restoring the feature maps of the optic disc boundary features to the original image size;

[0015] Step S24, refining the optic disc boundary segmentation results to obtain initial optic disc boundary detection results;

[0016] Step S25, using a graph search algorithm to optimize the optic disc boundary to obtain the optic disc boundary detection results.

[0017] Preferably, step S25 specifically comprises:

[0018] Step S251, determining the starting point and the ending point of the optic disc boundary according to the initial optic disc boundary detection results;

[0019] Step S252, setting pixel values of optic disc boundary points and background pixel values on the frequency domain optical coherence tomography image respectively, and taking the frequency domain optical coherence tomography image after setting as a mask image;

[0020] Step S253, finding the shortest path from the start point to the end point on the mask image using the graph search algorithm.

[0021] Preferably, step S3 comprises:

[0022] Step S30, filtering and denoising the frequency domain optical coherence tomography image;

[0023] Step S31, determining a region of interest above the optic disc boundary according to the optic disc boundary detection result, and removing a part below the optic disc boundary;

[0024] Step S32, setting a threshold to threshold segment the region of interest, and screening out connected regions greater than the threshold;

[0025] Step S33, screening the region of interest using area constraint of the region of interest and average gray value constraint of all pixels in the region of interest, to obtain a new blood vessel candidate region of the optic disc.

[0026] Preferably, step S33 specifically comprises:

[0027] Step S331, calculating pixel area of each connected region, and removing the connected region with pixel area lower than the area constraint;

[0028] Step S332, calculating average gray value of each connected region, and removing the connected region with average gray value lower than the average gray value constraint;

[0029] Step S333, selecting the remaining connected region as the new blood vessel candidate region of the optic disc.

[0030] Preferably, step S4 comprises:

[0031] Step S41, determining a blood vessel screening region with same shape and size as the new blood vessel candidate region of the optic disc at the same position of the optical coherence tomography angiography image according to the new blood vessel candidate region of the optic disc;

[0032] Step S42, judging whether each blood vessel screening region has blood flow signal, if yes, determining the blood vessel screening region as the new blood vessel region of the optic disc; otherwise, removing the blood vessel screening region;

[0033] Preferably, step S5 comprises:

[0034] Step S51, scaling each of the optical coherence tomography angiography images to determine the actual physical area P represented by each pixel point;

[0035] Step S52, calculating the volume V of the optic disc neovascularization of the single eye through N consecutive optical coherence tomography angiography images, the volume V being obtained by the following formula:

[0036]

[0037] Wherein A n is the pixel area of the optic disc neovascularization region on the nth optical coherence tomography angiography image, and U is the physical scanning interval of two consecutive optical coherence tomography angiography images.

[0038] The automatic detection system of optic disc neovascularization based on dual-mode optical scanning is applied to the automatic detection method of optic disc neovascularization and comprises:

[0039] An optical coherence tomography unit is configured to acquire the dual-mode optical scanning image containing the boundary of the optic disc;

[0040] The convolution network is connected to the optical coherence tomography unit

[0041] A blood vessel candidate region screening module is connected to the convolution network;

[0042] A blood vessel region screening module is connected to the blood vessel candidate region screening module;

[0043] A blood vessel volume calculation module is connected to the blood vessel region screening module.

[0044] Preferably, the convolution network comprises:

[0045] The feature extraction module comprises a double-layer convolution module connected to the output end of the optical coherence tomography unit, and a plurality of cascaded down-sampling modules;

[0046] The attention module is connected to the feature extraction module;

[0047] A plurality of cascaded feature fusion modules are connected to the attention module, and the first feature fusion module is connected to the double-layer convolution module through a 1x1 convolution kernel, and each of the remaining cascaded feature fusion modules is connected to a down-sampling module through a 1x1 convolution kernel.

[0048] Preferably, each of the dual-mode optical scanning images comprises:

[0049] an infrared fundus photograph image;

[0050] an optical coherence tomography image in the frequency domain;

[0051] an optical coherence tomography angiography image;

[0052] a composite image of the optical coherence tomography image in the frequency domain and the optical coherence tomography angiography image.

[0053] The present application has the beneficial effects that, by adopting the above technical scheme, the present application obtains a more smooth optic disc boundary detection result, improves the optic disc neovascularization detection precision, and eliminates noise interference. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 a step schematic diagram of the optic disc neovascularization automatic detection method in the embodiment of the present application;

[0055] Figure 2 a step schematic diagram of step S2 in the embodiment of the present application;

[0056] Figure 3 a step schematic diagram of step S25 in the embodiment of the present application;

[0057] Figure 4 a step schematic diagram of step S3 in the embodiment of the present application;

[0058] Figure 5 a step schematic diagram of step S33 in the embodiment of the present application;

[0059] Figure 6 a step schematic diagram of step S4 in the embodiment of the present application;

[0060] Figure 7 a step schematic diagram of step S5 in the embodiment of the present application;

[0061] Figure 8 a structure schematic diagram of the optic disc neovascularization automatic detection system in the embodiment of the present application;

[0062] Figure 9 a convolution network structure schematic diagram in the embodiment of the present application;

[0063] Figure 10 a feature fusion module structure schematic diagram in the embodiment of the present application;

[0064] Figure 11 a double-layer convolution module structure schematic diagram in the embodiment of the present application;

[0065] Figure 12 a down-sampling module structure schematic diagram in the embodiment of the present application;

[0066] Figure 13 Figure 1 is a schematic diagram of an up-sampling module structure in an embodiment of the present application. DETAILED DESCRIPTION

[0067] The technical solutions in the embodiments of the present application will be clearly and completely described with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0068] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0069] The present application will be further described below with reference to the drawings and specific embodiments, but is not limited to the present application.

[0070] An optic disc neovascularization automatic detection method based on dual-mode optical scanning, as shown in FIG. 1, comprises the following steps: Figure 1

[0071] Step S1, scanning around the optic disc of a single eye and outputting N continuous dual-mode optical scanning images containing the optic disc boundary, the dual-mode optical scanning images comprising frequency domain optical coherence tomography images and optical coherence tomography angiography images, establishing a data set containing optic disc boundary labels according to the N frequency domain optical coherence tomography images, N being a natural number greater than 1;

[0072] Step S2, training, verifying and testing the data set using a convolution network 2 to obtain an optic disc boundary detection result;

[0073] Step S3, screening out optic disc neovascularization candidate regions on each frequency domain optical coherence tomography image and optical coherence tomography angiography image according to the optic disc boundary detection result;

[0074] Step S4, screening out optic disc neovascularization regions in each optic disc neovascularization candidate region;

[0075] Step S5, calculating the volume of the optic disc neovascularization of the single eye according to all the optic disc neovascularization regions.

[0076] Specifically, in actual clinical application, dual-mode optical scanning is performed around the optic disc of a single eye with a size of 6mmx6mm, and 512 continuous frequency domain optical coherence tomography images and corresponding 512 optical coherence tomography angiography images can be obtained for each single eye of a patient, that is, the value of N in the present embodiment is 512.

[0077] ​Further, the finally established dataset contains 2289 frequency domain optical coherence tomography images and 2289 optical coherence tomography angiography images; the dataset is divided into a training set (70%, 1602 images) for training the convolutional network 2, a verification set (10%, 229 images) for selecting the optic disc boundary, and a test set (20%, 458 images) for testing the segmentation accuracy of the optic disc boundary.

[0078] In a preferred embodiment, as shown in Figure 2 Step S2 includes:

[0079] Step S21, the feature extraction module 7 of the convolutional network 2 extracts the optic disc boundary features of the frequency domain optical coherence tomography image by using a plurality of cascaded down-sampling modules;

[0080] Step S22, the attention module 8 of the convolutional network 2 performs a refinement process on the optic disc boundary features;

[0081] Step S23, the feature fusion module 9 of the convolutional network 2 fuses the optic disc boundary features of different scales, and outputs the optic disc boundary segmentation result after restoring the feature map of the optic disc boundary features to the original image size;

[0082] Step S24, the optic disc boundary segmentation result is refined to obtain an initial optic disc boundary detection result;

[0083] Step S25, the graph search algorithm is used to optimize the optic disc boundary to obtain the optic disc boundary detection result.

[0084] Further, in this embodiment, the binary cross-entropy loss function is used as the target function for training, and a multi-loss function is used to supervise the training of the network.

[0085] In order to speed up the training speed and save the GPU memory, the input image resolution is set to 256x256 pixels. In terms of network segmentation performance evaluation, the Dice coefficient is used to evaluate the optic disc boundary segmentation result. After 60 epochs of training, the epoch with the highest Dice value on the verification set is selected as the final model, and the Dice value on the verification set is 0.856 and the Dice value on the test set is 0.844.

[0086] In a preferred embodiment, as shown in Figure 3 Step S25 specifically includes:

[0087] Step S251, according to the initial optic disc boundary detection result, the starting point and the ending point of the optic disc boundary are determined;

[0088] Step S252, respectively set the pixel value of the optic disc boundary point and the background pixel value on the frequency domain optical coherence tomography image, and take the set frequency domain optical coherence tomography image as a mask image;

[0089] Step S253, find the shortest path from the start point to the end point on the mask image using a graph search algorithm.

[0090] Specifically, in the embodiment, the pixel value of the optic disc boundary point is set to 1, and the background pixel value is set to 1000, and the pixel value represents the cost of the path passing through this point; the optimization target is to minimize the sum of the pixel values of all pixel points on the path, that is, the lowest cost. A more continuous optic disc boundary can be obtained after optimization by the graph search, which is the final optic disc boundary detection result.

[0091] In a preferred embodiment, as shown in Figure 4 Step S3 includes:

[0092] Step S30, filter and denoise the frequency domain optical coherence tomography image;

[0093] Step S31, according to the optic disc boundary detection result, determine the part above the optic disc boundary as the region of interest, and remove the part below the optic disc boundary;

[0094] Step S32, set a threshold to threshold segment the region of interest, and filter out the connected regions greater than the threshold, specifically, set the threshold to 200;

[0095] Step S33, use the area constraint of the region of interest and the average gray value constraint of all pixels in the region of interest to filter the region of interest, and obtain the optic disc neovascularization candidate region.

[0096] Specifically, remove the part below the optic disc boundary, that is, set the pixel value of the region below the boundary to 0.

[0097] In a preferred embodiment, as shown in Figure 5 Step S33 specifically includes:

[0098] Step S331, calculate the pixel area of each connected region, and remove the connected regions with pixel area lower than the area constraint, specifically, in the embodiment, remove the connected regions with pixel area less than 30;

[0099] Step S332, calculate the average gray value of each connected region, and remove the connected regions with average gray value lower than the average gray value constraint, specifically, in the embodiment, remove the connected regions with average gray value less than 190;

[0100] Step S333, select the remaining connected regions as the optic disc neovascularization candidate region.

[0101] In a preferred embodiment, as shown in Figure 6 Step S4 includes:

[0102] Step S41, according to the optic disc neovascularization candidate region, determine the same size and shape of the blood vessel screening region as the optic disc neovascularization candidate region in the same position of the optical coherence tomography angiography image;

[0103] Step S42, determine whether each blood vessel screening region has blood flow signal, if yes, determine the blood vessel screening region as the optic disc neovascularization region; otherwise, remove the blood vessel screening region;

[0104] In a preferred embodiment, as shown in Figure 7 Step S5 includes:

[0105] Step S51, scale each optical coherence tomography angiography image to determine the actual physical area P represented by each pixel point;

[0106] Step S52, calculate the volume V of the optic disc neovascularization of a single eye through N consecutive optical coherence tomography angiography images, the volume V is obtained by the following formula:

[0107]

[0108] Where A n is the pixel area of the optic disc neovascularization region on the nth optical coherence tomography angiography image, and U is the physical scanning interval of two consecutive optical coherence tomography angiography images.

[0109] In this example, for a picture with a resolution of 256x256, U is 11.742um and P is 167.265um 2 .

[0110] The optic disc neovascularization automatic detection system based on dual-mode optical scanning is applied to the optic disc neovascularization automatic detection method in any one of the above embodiments, as shown in Figure 8 includes:

[0111] The optical coherence tomography unit 1 is used to obtain a dual-mode optical scanning image containing the optic disc boundary, specifically, the optical coherence tomography unit 1 in the present embodiment adopts the German Heidelberg OCTA Spectralis HRA+OCT (Heidelberg Engineering, Germany);

[0112] The convolution network 2 is connected to the optical coherence tomography unit 1;

[0113] The blood vessel candidate region screening module 3 is connected to the convolution network 2;

[0114] Blood vessel region screening module 4 is connected to blood vessel candidate region screening module 3;

[0115] The blood vessel volume calculation module 5 is connected to the blood vessel region filtering module 4.

[0116] In a preferred embodiment, such as Figure 9 As shown, convolutional network 2 includes:

[0117] The feature extraction module 7 includes a two-layer convolution module 11 connected to the output end of the optical coherence tomography unit 1, and also includes multiple cascaded downsampling modules 12;

[0118] Attention module 8 connects to feature extraction module 7;

[0119] Multiple cascaded feature fusion modules 9 are connected to attention modules 8, and the first feature fusion module 9 is connected to a two-layer convolution module 11 through a 1×1 convolution kernel. Each of the remaining cascaded feature fusion modules 9 is connected to a downsampling module 12 through a 1×1 convolution kernel.

[0120] Specifically, such as Figure 10 As shown, the feature fusion module 9 includes two input terminals: one connected to the internal upsampling module 14 and then to a convolutional layer 15, and the other directly connected to a convolutional layer 15. After the convolutional layer 15 extracts features, the features are jointly input to the internal feature map addition unit 16. The feature fusion module 9 also includes:

[0121] ReLU unit 17 is connected to feature map addition unit 16;

[0122] 1×1 convolution kernel 13, connected to ReLU unit 17;

[0123] Sigmoid unit 18, connected to 1×1 convolution kernel 13;

[0124] Feature map multiplication unit 19 is connected to Sigmoid unit 18;

[0125] Concat unit 20 connects feature map multiplication unit 19;

[0126] The two-layer convolutional module 11 is connected to the Concat unit 20;

[0127] Furthermore, the feature map multiplication unit 19 is also connected to the input of the feature fusion module 9, which is not connected to the upsampling module 14, and the Concat unit 20 is also connected to the output of the upsampling module 14.

[0128] Furthermore, such as Figure 11 As shown, the dual-layer convolutional module 11 includes two convolutional layers 15; as Figure 12As shown, the down-sampling module 12 includes a max-pooling layer 21 and two convolutional layers 15; as Figure 13 As shown, the up-sampling module 14 includes an up-sampling layer 22 and a convolutional layer 15.

[0129] In a preferred embodiment, each dual-modality optical scan image includes:

[0130] an infrared fundus photograph image;

[0131] a frequency-domain optical coherence tomography image;

[0132] an optical coherence tomography angiography image;

[0133] a composite image of the frequency-domain optical coherence tomography image and the optical coherence tomography angiography image.

[0134] In summary, the technical solution in the present application obtains a more smooth optic disc boundary detection result by combining the optic disc boundary detection based on the convolutional network 2 and the optic disc boundary post-processing algorithm based on graph search, determines the region of interest for the optic disc neovascularization detection through the optic disc boundary detection, improves the optic disc neovascularization detection precision, eliminates noise interference, and combines the morphological features in the frequency-domain optical coherence tomography image and the blood flow signal information in the optical coherence tomography angiography image for the optic disc neovascularization detection, utilizes two kinds of modal information, and improves the optic disc neovascularization detection precision.

[0135] The above only describes preferred embodiments of the present application, and does not limit the implementation manners and protection scope of the present application. It should be understood by those skilled in the art that any equivalent replacement and obvious change made according to the content of the present application should be included in the protection scope of the present application.

Claims

1. A method for automatic detection of neovascularization in the optic disc based on bimodal optical scanning, characterized in that, The method comprises the following steps; Step S1, scanning and outputting N continuous bimodal optical scanning images containing the optic disc boundary of a single eye, the bimodal optical scanning images comprising frequency domain optical coherence tomography images and optical coherence tomography angiography images, establishing a dataset containing optic disc boundary labels according to the N frequency domain optical coherence tomography images, N being a natural number greater than 1; Step S2, training, verifying and testing the dataset using a convolutional network to obtain optic disc boundary detection results; Step S3, screening out optic disc neovascularization candidate regions on each of the frequency domain optical coherence tomography images and the optical coherence tomography angiography images according to the optic disc boundary detection results; Step S4, screening out optic disc neovascularization regions in each of the optic disc neovascularization candidate regions; Step S5, calculating the volume of the optic disc neovascularization of the single eye according to all the optic disc neovascularization regions; The step S4 comprises: Step S41, determining a blood vessel screening region with the same shape and size as the optic disc neovascularization candidate region at the same position of the optical coherence tomography angiography image according to the optic disc neovascularization candidate region; Step S42, judging whether each blood vessel screening region has blood flow signals, if yes, determining the blood vessel screening region as the optic disc neovascularization region; otherwise, removing the blood vessel screening region.

2. The optic disc neovascularization automatic detection method according to claim 1, characterized in that, Step S2 comprises: Step S21, a feature extraction module of the convolutional network extracts optic disc boundary features of the frequency domain optical coherence tomography images by using multiple cascaded down-sampling modules; Step S22, an attention module of the convolutional network refines the optic disc boundary features; Step S23, a feature fusion module of the convolutional network fuses the optic disc boundary features of different scales, and outputs an optic disc boundary segmentation result after restoring the feature map of the optic disc boundary features to the original image size; Step S24, refining the optic disc boundary segmentation result to obtain initial optic disc boundary detection results; Step S25, using a graph search algorithm to optimize the optic disc boundary to obtain the optic disc boundary detection results.

3. The optic disc neovascularization automatic detection method according to claim 2, characterized in that, Step S25 specifically comprises: Step S251, determining the starting point and the ending point of the optic disc boundary according to the initial optic disc boundary detection results; Step S252, setting the pixel value of the optic disc boundary point and the background pixel value on the frequency domain optical coherence tomography image respectively, and taking the set frequency domain optical coherence tomography image as a mask image; Step S253, using the graph search algorithm to find the shortest path from the starting point to the ending point on the mask image.

4. The optic disc neovascularization automatic detection method according to claim 3, characterized in that, Step S3 comprises: Step S30, filtering and denoising the frequency domain optical coherence tomography image; Step S31, determining the region of interest above the optic disc boundary according to the optic disc boundary detection results, and removing the region below the optic disc boundary; Step S32, setting a threshold to threshold segment the region of interest, and screening out connected regions greater than the threshold; Step S33, screening the region of interest using the area constraint and the average gray value constraint of all pixels in the region of interest to obtain the optic disc neovascularization candidate region.

5. The optic disc neovascularization automatic detection method according to claim 4, characterized in that, Step S33 specifically includes: Step S331, calculating the pixel area of each connected region, and removing the connected region whose pixel area is lower than the area constraint; Step S332, calculating the average gray value of each connected region, and removing the connected region whose average gray value is lower than the average gray value constraint; Step S333, selecting the remaining connected regions as the optic disc neovascularization candidate region.

6. The optic disc neovascularization automatic detection method according to claim 1, characterized in that, Step S5 includes: Step S51, scaling each of the optical coherence tomography angiography images to determine the actual physical area represented by each pixel point ; Step S52, calculating the volume of the optic disc neovascularization of the single eye through N continuous optical coherence tomography angiography images , the volume is obtained by the following formula: ； wherein is the first is the pixel area of the optic disc neovascularization region on the optical coherence tomography angiography image, is the physical scanning interval of the two consecutive optical coherence tomography angiography images.

7. The system for automatic detection of optic disc neovascularization based on bimodal optical scanning, applied to the method for automatic detection of optic disc neovascularization according to any one of claims 1-6, characterized in that, including: An optical coherence tomography unit for acquiring the dual-mode optical scanning image containing the optic disc boundary; The convolutional network is connected to the optical coherence tomography unit; A blood vessel candidate region screening module connected to the convolutional network; A blood vessel region screening module connected to the blood vessel candidate region screening module; A blood vessel volume calculation module connected to the blood vessel region screening module.

8. The optic disc neovascularization automatic detection system of claim 7, wherein, The convolutional network includes: A feature extraction module including a double-layer convolutional module connected to the output end of the optical coherence tomography unit, and a plurality of cascaded down-sampling modules; An attention module connected to the feature extraction module; A plurality of cascaded feature fusion modules connected to the attention module, and the first feature fusion module is connected to the double-layer convolutional module through a 1×1 convolutional kernel, and each subsequent feature fusion module is connected to a down-sampling module through a 1×1 convolutional kernel.

9. The optic disc neovascularization automatic detection system of claim 8, wherein, Each of the dual-mode optical scanning images includes: An infrared fundus photograph image; A frequency domain optical coherence tomography image; An optical coherence tomography angiography image; A composite image of the frequency domain optical coherence tomography image and the optical coherence tomography angiography image.

Images