Method and system for detecting anemia based on fundus images

Abstract

Provided is an anemia detection method based on fundus images, comprising: optic disc segmentation; fundus blood vessel segmentation and arteriovenous classification; blood vessel skeleton extraction and blood vessel tree construction; blood vessel parameter extraction, the blood vessel parameters including blood vessel tortuosity, blood vessel diameter and blood vessel original image gray scale features; anemia classification and detection. Also provided is an anemia detection system based on fundus images, which, after combining the extracted blood vessel tortuosity, blood vessel diameter, original image features and features extracted by a deep neural network, improves the accuracy, sensitivity and precision of anemia detection.

Description

Technical Field

[0001] This invention belongs to the field of fundus image detection, specifically relating to a method and system for detecting anemia based on fundus images. Background Technology

[0002] Globally, especially in developing countries, the incidence of anemia is significant. Anemia patients are prone to symptoms such as fatigue, weakness, and difficulty concentrating, impacting their daily lives and work efficiency. Clinical anemia testing is usually invasive, involving the collection of a blood sample and diagnosis based on the measured hemoglobin levels.

[0003] Therefore, it is necessary to study an anemia detection method based on fundus images to solve one or more of the above-mentioned technical problems. Summary of the Invention

[0004] To address at least one of the aforementioned technical problems, this invention provides a method for detecting anemia based on fundus images. This method quantifies fundus vascular-related indicators through steps such as fundus image vessel segmentation, arteriovenous classification and vascular skeleton extraction, and fundus image feature extraction. It also combines deep learning with manually extracted features for feature fusion for anemia detection. Furthermore, the extracted fundus feature parameters can be visualized and fed back to clinicians as supplementary information, demonstrating significant clinical application value.

[0005] Specifically, the anemia detection method based on fundus images is characterized by including the following steps:

[0006] View disc splitting;

[0007] Fundus vessel segmentation and arteriovenous classification;

[0008] Vascular skeleton extraction and vascular tree construction;

[0009] Vascular parameter extraction, wherein the vascular parameters include vascular tortuosity, vascular diameter and grayscale features of the original vascular image;

[0010] Anemia classification and detection;

[0011] The classification and detection of anemia specifically include:

[0012] Classify raw fundus images using GoogLeNet, SKNet, or DenseNet networks;

[0013] Change the output layer of the GoogLeNet, SKNet, or DenseNet network so that it outputs image features as the first feature vector before the linear classifier;

[0014] Extract the class probabilities output by the linear classifier of the GoogLeNet, SKNet, or DenseNet network as the second feature vector;

[0015] The vascular parameters are extracted into a third feature vector;

[0016] The first to third feature vectors are input into SVM and / or Xgboost for classification, and the anemia detection results are output.

[0017] According to another aspect of the invention, the first feature vector is specifically extracted between the linear classifier and the layer preceding the linear classifier.

[0018] According to another aspect of the present invention, the vascular parameters are visualized.

[0019] According to another aspect of the present invention, the spectral segmentation is performed using an adjusted Attention-U-Net, which includes a downsampling part and an upsampling part. The downsampling part includes four sets of sequentially alternately connected (3×3 convolution + ReLU function)×2 modules and pooling modules. The upsampling part includes four sets of sequentially alternately connected upsampling modules and (3×3 convolution + ReLU function)×2 modules. The downsampling part and the upsampling part are connected by the (3×3 convolution + ReLU function)×2 modules. The upsampling part further includes: an attention gate fusion module for fusing the same-level image features of the downsampling part and the image features of the previous layer; and a cascade module for cascading the output of the attention gate fusion module with the output of the upsampling module.

[0020] According to another aspect of the present invention, the fundus vessel segmentation and arteriovenous classification specifically adopt a two-stage arteriovenous classification model for classification. In this model, the grayscale images of the original image, the result of the first-stage arteriovenous classification, and the result of the first-stage fundus vessel segmentation are respectively input into the R, G, and B channels to obtain a synthesized image, which is used as the input of the second-stage arteriovenous classification network. The output of the second-stage arteriovenous classification network is used as the final result.

[0021] According to another aspect of the present invention, the vascular skeleton extraction and vascular tree construction include: firstly, performing dilation operations on the binarized vascular images of arteries and veins respectively, then performing connected component analysis, setting an area threshold, and removing image blocks with areas smaller than the threshold; extracting the vascular skeleton based on the arteriovenous classification results; and constructing a vascular tree using vascular endpoints and bifurcation points as seed points.

[0022] According to another aspect of the present invention, an anemia detection system based on fundus images is also provided, characterized by comprising:

[0023] The first module is used for video disc splitting;

[0024] The second module is used for retinal vessel segmentation and arteriovenous classification.

[0025] The third module is used for vascular skeleton extraction and vascular tree construction;

[0026] The fourth module is used for extracting vascular parameters, which include vascular tortuosity, vascular diameter, and grayscale features of the original vascular image.

[0027] The fifth module is used for the classification and detection of anemia;

[0028] Specifically, the fifth module includes:

[0029] The sixth module is used to classify raw fundus images using GoogLeNet, SKNet, or DenseNet networks;

[0030] The seventh module is used to modify the output layer of the GoogLeNet, SKNet, or DenseNet network so that it outputs image features as the first feature vector before the linear classifier.

[0031] The eighth module is used to extract the class probabilities output by the linear classifier of the GoogLeNet, SKNet, or DenseNet network into a second feature vector;

[0032] The ninth module is used to extract the vascular parameters into a third feature vector;

[0033] The tenth module is used to input the first to third feature vectors into SVM and / or Xgboost for classification and output the anemia detection results.

[0034] According to another aspect of the present invention, the first module is specifically an adjusted Attention-U-Net module, the adjusted Attention-U-Net module including a downsampling unit and an upsampling unit, the downsampling unit including four sets of sequentially alternately connected (3×3 convolution + ReLU function)×2 modules and pooling modules; the upsampling unit including four sets of sequentially alternately connected upsampling modules and (3×3 convolution + ReLU function)×2 modules; the downsampling unit and the upsampling unit are connected through (3×3 convolution + ReLU function)×2 modules, wherein the upsampling unit further includes: an attention gate fusion module for fusing the same-level image features of the downsampling unit and the image features of the previous layer; and a cascade module for cascading the output of the attention gate fusion module and the output of the upsampling module.

[0035] According to another aspect of the present invention, the second module is a two-stage arteriovenous classification module, wherein the two-stage arteriovenous classification module is used to input the grayscale images of the original image, the result of the first-stage arteriovenous classification, and the result of the first-stage fundus vessel segmentation into the R, G, and B channels respectively to obtain a synthesized image as the input of the second-stage arteriovenous classification network, and to use the output of the second-stage arteriovenous classification network as the final result.

[0036] According to another aspect of the present invention, the third module includes:

[0037] The eleventh module is used to perform dilation operations on the binary blood vessel images of arteries and veins respectively, then perform connected component analysis, and then set an area threshold to remove image blocks with an area smaller than the threshold.

[0038] The twelfth module is used to extract the vascular skeleton based on the arteriovenous classification results;

[0039] The thirteenth module is used to construct a vascular tree using vascular endpoints and bifurcation points as seed points.

[0040] This invention can achieve one or more of the following technical effects:

[0041] 1. This invention improves the accuracy, sensitivity, and precision of anemia detection by combining extracted blood vessel curvature, blood vessel diameter, original image features, and features extracted by deep neural networks.

[0042] 2. The adjusted Attention-U-Net improves segmentation performance on fundus images;

[0043] 3. Images synthesized using a two-stage arteriovenous classification model improve vascular continuity, which can enhance subsequent detection results;

[0044] 4. Performing connected component analysis on binarized blood vessel images of arteries and veins can eliminate small scattered points in blood vessel segmentation and improve detection results. Attached Figure Description

[0045] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0046] Figure 1 This is a schematic diagram of an anemia classification and detection module according to a first preferred embodiment of the present invention.

[0047] Figure 2 This is a schematic diagram of an anemia classification and detection module according to a second preferred embodiment of the present invention.

[0048] Figure 3 This is a schematic diagram of an anemia classification and detection module according to a third preferred embodiment of the present invention.

[0049] Figure 4 This is a schematic diagram of the adjusted Attention-U-Net structure according to a preferred embodiment of the present invention.

[0050] Figure 5 This is a schematic diagram of a vascular tree constructed according to a preferred embodiment of the present invention.

[0051] Figure 6 This is a preferred embodiment of the blood vessel tortuosity calculation model according to the present invention.

[0052] Figure 7 This is a schematic diagram of blood vessel diameter measurement according to a preferred embodiment of the present invention.

[0053] Figure 8 This is a schematic diagram of the U-Net, 5L, 32W network structure according to a preferred embodiment of the present invention. Detailed Implementation

[0054] The preferred embodiments of the present invention will now be described with reference to the accompanying drawings. These specific embodiments are intended to illustrate the present invention in detail, but should not be construed as limiting the present invention. Various modifications and variations can be made without departing from the spirit and scope of the present invention, and all of these should be included within the protection scope of the present invention.

[0055] Example 1

[0056] According to a preferred embodiment of the present invention, see Figure 1-4 This paper provides a method for detecting anemia based on fundus images, characterized by the following steps:

[0057] View disc splitting;

[0058] Fundus vessel segmentation and arteriovenous classification;

[0059] Vascular skeleton extraction and vascular tree construction;

[0060] Vascular parameter extraction, wherein the vascular parameters include vascular tortuosity, vascular diameter and grayscale features of the original vascular image;

[0061] Anemia classification and detection.

[0062] Preferably, the anemia classification and detection specifically includes:

[0063] Classify raw fundus images using GoogLeNet, SKNet, or DenseNet networks;

[0064] Change the output layer of the GoogLeNet, SKNet, or DenseNet network so that it outputs image features as the first feature vector before the linear classifier;

[0065] Extract the class probabilities output by the linear classifier of the GoogLeNet, SKNet, or DenseNet network as the second feature vector;

[0066] The vascular parameters are extracted into a third feature vector;

[0067] The first to third feature vectors are input into SVM and / or Xgboost for classification, and the anemia detection results are output.

[0068] Understandably, GoogLeNet, SKNet, or DenseNet networks are deep neural networks in the existing technology.

[0069] According to another preferred embodiment of the present invention, see Figure 1-3 The first feature vector is specifically extracted between the linear classifier and the layer preceding the linear classifier.

[0070] According to another preferred embodiment of the present invention, the vascular parameters are visualized.

[0071] According to another preferred embodiment of the present invention, see Figure 4 The spectral segmentation is performed using an adjusted Attention-U-Net, which includes a downsampling part and an upsampling part. The downsampling part comprises four sets of alternately connected (3×3 convolution + ReLU function)×2 modules and pooling modules. The upsampling part comprises four sets of alternately connected upsampling modules and (3×3 convolution + ReLU function)×2 modules. The downsampling and upsampling parts are connected by the (3×3 convolution + ReLU function)×2 modules. The upsampling part further includes: an attention gate fusion module for fusing the same-level image features from the downsampling part and the image features from the previous layer; and a cascade module for cascading the output of the attention gate fusion module with the output of the upsampling module. It can be understood that the same-level image features refer to... Figure 4 The dotted-dash line connecting the downsampling portion (left) and the upsampling portion (right) indicates that they have the same level. Figure 4 (Displayed at the same height), the image features of the same level in this downsampled portion are input to the corresponding attention gate. The image features of the upper layer specifically refer to... Figure 4 The image features connected by the central attention gate.

[0072] Favorably, see Figure 4The main structure of the Attention-U-Net is divided into three parts: (1) downsampling, corresponding to the left half of the figure; (2) upsampling, corresponding to the right half of the figure; and (3) skip connections, i.e., the horizontal arrow part. Downsampling corresponds to the feature compression process, i.e., the encoding process, in which the image feature size is continuously reduced. The skip connections in the middle are used to combine deep and shallow features. In the network, each neuron is followed by an activation function, which provides non-linear connections. This activation function is, for example, the ReLU function.

[0073] The size of the input image features is continuously reduced under the action of the pooling layer until the output of the deepest hidden layer, at which point the feature size of the image becomes 1 / 16 of the original size. Then, through upsampling, the original size is gradually restored. At each layer, while restoring the size, shallow features obtained by fusing the same-level features from the downsampling process through the attention gate (AG) and the features from the previous layer are received. Finally, the size is the same as the original input size, achieving an end-to-end segmentation result.

[0074] According to another preferred embodiment of the present invention, the fundus vessel segmentation and arteriovenous classification specifically employ a two-stage arteriovenous classification model. Specifically, the grayscale images of the original image, the result of the first-stage arteriovenous classification, and the result of the first-stage fundus vessel segmentation are input into the R, G, and B channels, respectively, to obtain a synthesized image, which serves as the input to the second-stage arteriovenous classification network. The output of the second-stage arteriovenous classification network is then used as the final result. Advantageously, the image synthesized using the two-stage arteriovenous classification model improves vessel continuity, thereby enhancing subsequent detection results.

[0075] Specifically, fundus vessel segmentation, first-stage arteriovenous classification, and second-stage arteriovenous classification can utilize existing U-Net network structures. For example, see [link to relevant documentation]. Figure 8 The U-Net network has 5 hidden layers, with an input image size of 512×512 and an output image size of 512×512. The initial feature map has 32 channels. With each additional layer of depth, the feature size is halved while the number of channels doubles. Pooling layers halve the feature size, while deconvolutional layers gradually restore the feature size. Skip connections combine features from shallower layers of the network.

[0076] According to another preferred embodiment of the present invention, the vascular skeleton extraction and vascular tree construction include: performing dilation operations on the binarized vascular images of arteries and veins respectively, then performing connected component analysis, then setting an area threshold and removing image blocks with an area smaller than the threshold; extracting the vascular skeleton based on the arteriovenous classification results; and constructing a vascular tree using vascular endpoints and bifurcation points as seed points.

[0077] Preferably, the extraction of the vascular skeleton involves refining the results obtained from vascular segmentation. For a segment of blood vessel, edge pixels are progressively removed until only the central line of the vessel remains. Since parameters need to be extracted separately for arteries and veins, the vascular skeleton extraction operation is performed separately for each artery and vein channel. Before extracting the vascular skeleton based on the arterial and vein classification results, further processing is required to improve vascular continuity and remove scattered points.

[0078] For small scattered points in blood vessel segmentation, such as arterial scattered points, including small blood vessel segments that remain isolated after blood vessel dilation, these segments are very short, and studying their morphological characteristics is not very meaningful and may even lead to erroneous results. Therefore, these small blood vessel segments need to be removed. The specific method is as follows: perform dilation operations on the binarized blood vessel images of arteries and veins respectively, then perform connected component analysis, and then set an area threshold to remove image blocks with an area smaller than the threshold.

[0079] Specifically, for small, narrow gaps in blood vessels, such as discontinuous arteries or veins, morphological dilation can be used to fill them. Dilation (dilation operation) allows the vessel to grow into the background area while maintaining its center position, effectively thickening the vessel. This process can merge vessels with small gaps at both ends. Because dilation does not change the center position of the vessel, it does not affect the selection of the center point for skeleton extraction.

[0080] Regarding the construction of vascular trees, when studying vascular morphology, it is necessary to select appropriate seed points. The vascular tree is traversed starting from these seed points, and vascular information is measured during the traversal. Seed points can be selected manually, i.e., the starting point and detection range are manually selected using interactive software, or they can be automatically selected by a computer according to certain rules. To extract morphological statistical information of blood vessels, vascular endpoints and bifurcation points can be used as automatically selected seed points. The vascular tree construction result is as follows: Figure 5 As shown.

[0081] According to another preferred embodiment of the present invention, the distance factor (DF) approximates the tortuosity of a blood vessel by calculating the ratio between the straight-line distance between the two endpoints of a blood vessel and the pixel-by-pixel distance integral. Figure 6 The diagram shows a DF (Directional Function) representation of a blood vessel segment. The curve represents the vascular skeleton, and DF is the ratio of the length of the curve to the length of the straight dashed line. The mathematical definition of DF is shown in equation (1):

[0082]

[0083] Where P1(x1,y1) is the starting point of the blood vessel segment, P n (x n ,y n() represents the endpoint of the blood vessel segment. n represents the pixel count traversed by this segment of the blood vessel. When the blood vessel is strictly straight, the DF value is zero; the larger the DF value, the more tortuous the blood vessel can be considered. Since the length of blood vessels varies, DF cannot reflect the local features of very long blood vessels. Therefore, in actual calculations, a sliding window is used to calculate the DF features of different parts of a blood vessel segment. That is, starting from the seed point, a fixed length of 2m is taken each time, and the DF features of point P are calculated in the directions P1 and P2 respectively. n Given m pixel coordinates in the direction, calculate the DF of this segment of the blood vessel as the DF feature of point P.

[0084] Preferably, vessel diameter is an important morphological feature of blood vessels. By segmenting blood vessels and classifying arteries and veins, the boundary information of the fundus vessels is obtained, and then the diameter is measured using binarized images.

[0085] like Figure 7 The diagram shows a schematic of blood vessel diameter measurement. The approximately parallel curves represent the result of blood vessel segmentation, and the dashed line in the middle represents the blood vessel centerline extracted from the blood vessel skeleton. To describe the blood vessel diameter information of the entire blood vessel tree, the blood vessel diameter corresponding to each blood vessel center point is measured. The specific implementation method is as follows: Read the blood vessel skeleton coordinate file, where each coordinate corresponds to a blood vessel center point, and blood vessels belonging to the same segment are placed in the same group. Take a point P from it sequentially, and take the shortest line segment from point P to the two boundaries of this blood vessel as the blood vessel diameter. In the design, with an initial angle theta0 = 0 and an initial length L0, gradually increase the value of L until the line segment with point P as the origin, angle theta, and length L exactly passes through both boundaries. At this point, the length of the line segment from point P to the two boundary points is obtained, denoted as D0. Theta becomes theta0 + k*div (k = 1, 2, 3...n), and div is the angle step value. Repeat the above operation. Until theta = theta + 180°, we obtain D0, D1, D2, D3...D n Finally, select D0, D1, ..., D... n The minimum value in the range is the diameter of the blood vessel corresponding to point P. The diameters of all blood vessel segments are measured sequentially using the method described above.

[0086] According to another preferred embodiment of the present invention, the detection results of the method of the present invention will be described in detail below with reference to specific examples.

[0087] The anemia dataset contains 1184 fundus images. Anemia labels were assigned by professional clinicians according to grading standards to create a balanced dataset. Of these 1184 fundus images, 592 individuals did not have anemia, and 592 individuals had varying degrees of anemia, including 392 with mild anemia, 185 with moderate anemia, and 15 with severe anemia.

[0088] K-fold cross-validation was used to evaluate the classifier's performance. K-fold cross-validation requires dividing the original dataset into K equal parts; for example, when K=5, it is called five-fold cross-validation. The steps of K-fold cross-validation are: (1) Divide the original data into K non-overlapping parts; (2) Use one part as the validation set and the other K-1 parts as the training set; (3) Train the model and evaluate its results on the validation set; (4) Repeat steps 2 and 3 K times, selecting different validation sets each time; (5) Use the average accuracy as the final accuracy. In addition, to better evaluate the model's generalization ability, a portion of the data is randomly selected from the entire dataset beforehand as an independent test set; this portion of data does not participate in the training and validation process.

[0089] The anemia dataset was divided into 6 parts. First, we randomly selected 50 cases each from the normal samples and the anemia samples as independent test sets. The remaining samples were used for 5-fold cross-validation. Based on the sample size, the samples were randomly divided into 5 groups. The number of normal and anemia samples in each group is shown in Table 1.

[0090] Table 1. Distribution of original data for five-fold cross-validation

[0091]

[0092] After processing by the method of this invention, the results are as follows:

[0093]

[0094] The results show that by combining the extracted blood vessel tortuosity, blood vessel diameter, original image features and features extracted by deep neural networks, the accuracy, sensitivity, precision and specificity of the validation set and the test set are high.

[0095] According to another preferred embodiment of the present invention, an anemia detection system based on fundus images is also provided, characterized in that it includes:

[0096] The first module is used for video disc splitting;

[0097] The second module is used for retinal vessel segmentation and arteriovenous classification.

[0098] The third module is used for vascular skeleton extraction and vascular tree construction;

[0099] The fourth module is used for extracting vascular parameters, which include vascular tortuosity, vascular diameter, and grayscale features of the original vascular image.

[0100] The fifth module is used for the classification and detection of anemia;

[0101] Specifically, the fifth module includes:

[0102] The sixth module is used to classify raw fundus images using GoogLeNet, SKNet, or DenseNet networks;

[0103] The seventh module is used to modify the output layer of the GoogLeNet, SKNet, or DenseNet network so that it outputs image features as the first feature vector before the linear classifier.

[0104] The eighth module is used to extract the class probabilities output by the linear classifier of the GoogLeNet, SKNet, or DenseNet network into a second feature vector;

[0105] The ninth module is used to extract the vascular parameters into a third feature vector;

[0106] The tenth module is used to input the first to third feature vectors into SVM and / or Xgboost for classification and output the anemia detection results.

[0107] According to another preferred embodiment of the present invention, the first module is specifically an adjusted Attention-U-Net module. The adjusted Attention-U-Net module includes a downsampling unit and an upsampling unit. The downsampling unit includes four sets of sequentially alternately connected (3×3 convolution + ReLU function)×2 modules and a pooling module. The upsampling unit includes four sets of sequentially alternately connected upsampling modules and (3×3 convolution + ReLU function)×2 modules. The downsampling unit and the upsampling unit are connected through the (3×3 convolution + ReLU function)×2 modules. The upsampling unit further includes: an attention gate fusion module for fusing the same-level image features of the downsampling unit and the image features of the previous layer; and a cascade module for cascading the output of the attention gate fusion module and the output of the upsampling module.

[0108] According to another preferred embodiment of the present invention, the second module is a two-stage arteriovenous classification module, wherein the two-stage arteriovenous classification module is used to input the grayscale images of the original image, the result of the first-stage arteriovenous classification, and the result of the first-stage fundus vessel segmentation into the R, G, and B channels respectively to obtain a synthesized image as the input of the second-stage arteriovenous classification network, and to use the output of the second-stage arteriovenous classification network as the final result.

[0109] According to another preferred embodiment of the present invention, the third module includes:

[0110] The eleventh module is used to perform connected component analysis on the binary vascular images of arteries and veins respectively, and then set an area threshold to remove image blocks with an area smaller than the threshold.

[0111] The twelfth module is used to extract the vascular skeleton based on the arteriovenous classification results;

[0112] The thirteenth module is used to construct a vascular tree using vascular endpoints and bifurcation points as seed points.

[0113] This invention can achieve one or more of the following technical effects:

[0114] 1. This invention improves the accuracy, sensitivity, and precision of anemia detection by combining extracted blood vessel curvature, blood vessel diameter, original image features, and features extracted by deep neural networks.

[0115] 2. The adjusted Attention-U-Net improves segmentation performance on fundus images;

[0116] 3. Images synthesized using a two-stage arteriovenous classification model improve vascular continuity, which can enhance subsequent detection results;

[0117] 4. Performing connected component analysis on binarized blood vessel images of arteries and veins can eliminate small scattered points in blood vessel segmentation and improve detection results.

[0118] Those skilled in the art should understand that this invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to this invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for detecting anemia based on fundus images, characterized in that... Includes the following steps: View disc splitting; Fundus vessel segmentation and arteriovenous classification; Vascular skeleton extraction and vascular tree construction; Vascular parameter extraction, wherein the vascular parameters include vascular tortuosity, vascular diameter and grayscale features of the original vascular image; Anemia classification and detection; The classification and detection of anemia specifically include: Classify raw fundus images using GoogLeNet, SKNet, or DenseNet networks; Change the output layer of the GoogLeNet, SKNet, or DenseNet network so that it outputs image features as the first feature vector before the linear classifier; Extract the class probabilities output by the linear classifier of the GoogLeNet, SKNet, or DenseNet network as the second feature vector; The vascular parameters are extracted into a third feature vector; The first to third feature vectors are input into SVM and / or Xgboost for classification, and the anemia detection results are output. The tortuosity of a blood vessel is approximated by calculating the ratio between the straight-line distance between the two endpoints of a blood vessel and the pixel-by-pixel distance integral. The viewport segmentation is performed using an adjusted Attention-U-Net, which includes a downsampling part and an upsampling part. The downsampling part includes four sets of alternately connected (3×3 convolution + ReLU function)×2 modules and pooling modules. The upsampling part includes four sets of alternately connected upsampling modules and (3×3 convolution + ReLU function)×2 modules. The downsampling part and the upsampling part are connected by the (3×3 convolution + ReLU function)×2 modules. The upsampling part further includes: an attention gate fusion module for fusing image features from the same level of the downsampling part and image features from the previous layer; and a cascade module for cascading the output of the attention gate fusion module with the output of the upsampling module. The fundus vessel segmentation and arteriovenous classification specifically adopt a two-stage arteriovenous classification model. In this model, the grayscale images of the original image, the result of the first-stage arteriovenous classification, and the result of the first-stage fundus vessel segmentation are input into the R, G, and B channels, respectively, to obtain a synthesized image, which is then used as the input to the second-stage arteriovenous classification network. The output of the second-stage arteriovenous classification network is then used as the final result.

2. The anemia detection method based on fundus images according to claim 1, characterized in that... The first feature vector is specifically extracted between the linear classifier and the layer preceding it.

3. The anemia detection method based on fundus images according to claim 2, characterized in that... Visualize the vascular parameters.

4. The anemia detection method based on fundus images according to claim 3, characterized in that... The vascular skeleton extraction and vascular tree construction include: performing dilation operations on the binarized vascular images of arteries and veins respectively, then performing connected component analysis, and then setting an area threshold to remove image blocks with an area smaller than the threshold; extracting the vascular skeleton based on the arteriovenous classification results; and constructing the vascular tree using vascular endpoints and bifurcation points as seed points.

5. An anemia detection system based on fundus images, characterized in that... include: The first module is used for video disc splitting; The second module is used for retinal vessel segmentation and arteriovenous classification. The third module is used for vascular skeleton extraction and vascular tree construction; The fourth module is used for extracting vascular parameters, which include vascular tortuosity, vascular diameter, and grayscale features of the original vascular image. The fifth module is used for the classification and detection of anemia; Specifically, the fifth module includes: The sixth module is used to classify raw fundus images using GoogLeNet, SKNet, or DenseNet networks; The seventh module is used to modify the output layer of the GoogLeNet, SKNet, or DenseNet network so that it outputs image features as the first feature vector before the linear classifier. The eighth module is used to extract the class probabilities output by the linear classifier of the GoogLeNet, SKNet, or DenseNet network into a second feature vector; The ninth module is used to extract the vascular parameters into a third feature vector; The tenth module is used to input the first to third feature vectors into SVM and / or Xgboost for classification and output the anemia detection results; The tortuosity of a blood vessel is approximated by calculating the ratio between the straight-line distance between the two endpoints of a blood vessel and the pixel-by-pixel distance integral. The first module is specifically the adjusted Attention-U-Net module, which includes a downsampling unit and an upsampling unit. The downsampling unit includes four sets of alternately connected (3×3 convolution + ReLU function)×2 modules and a pooling module. The upsampling unit includes four sets of alternately connected upsampling modules and (3×3 convolution + ReLU function)×2 modules. The downsampling unit and the upsampling unit are connected through the (3×3 convolution + ReLU function)×2 modules. The upsampling unit further includes: an attention gate fusion module for fusing the same-level image features of the downsampling unit and the image features of the previous layer; and a cascade module for cascading the output of the attention gate fusion module with the output of the upsampling module. The second module is a two-stage arteriovenous classification module. The two-stage arteriovenous classification module is used to input the grayscale images of the original image, the result of the first-stage arteriovenous classification, and the result of the first-stage fundus vessel segmentation into the R, G, and B channels respectively to obtain a synthesized image as the input of the second-stage arteriovenous classification network, and the output of the second-stage arteriovenous classification network as the final result.

6. The anemia detection system based on fundus images according to claim 5, characterized in that... The third module includes: The eleventh module is used to perform dilation operations on the binary blood vessel images of arteries and veins respectively, then perform connected component analysis, and then set an area threshold to remove image blocks with an area smaller than the threshold. The twelfth module is used to extract the vascular skeleton based on the arteriovenous classification results; The thirteenth module is used to construct a vascular tree using vascular endpoints and bifurcation points as seed points.

Images