A pixel-level screening strategy-based retinal blood vessel segmentation system

Abstract

The application discloses a kind of retinal blood vessel segmentation systems based on pixel-level screening strategy, comprising: fundus image acquisition unit, for obtaining the fundus image to be segmented;Image pre-processing unit is used to standardize the fundus image to be segmented Preprocessing, obtain standardization image sample;Binary classification label output unit is used to standardization image sample is input into the advanced segmentation model of secondary supervision training, the blood vessel class probability of each pixel unit is output, and the blood vessel class probability threshold is converted into the binary classification label of background;Wherein, the training of the advanced segmentation model includes: loss masking operation is carried out to low-confidence pixel unit in second pseudo-label sample;The application reduces the dependence on a large number of labeled data, avoids the negative influence of error propagation, finally, the advanced segmentation model generated shows higher accuracy in blood vessel segmentation task.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical image sample segmentation, in particular to a retinal blood vessel segmentation system and method based on a pixel-level screening strategy. BACKGROUND

[0002] With the deep integration of artificial intelligence and medical imaging technology, retinal blood vessel segmentation plays an increasingly important role in the early diagnosis of diseases such as diabetic retinopathy and hypertension. Traditional fully supervised segmentation methods rely on a large amount of high-quality labeled data, but due to the high cost of manual labeling, time-consuming and difficulty in ensuring labeling consistency, this method has great limitations. To solve this problem, semi-supervised learning combines a small amount of labeled data with a large amount of unlabeled data, reducing the dependence on manual labeling and effectively reducing the labeling cost.

[0003] The patent document with patent publication number CN115115659B discloses an automatic fundus blood vessel segmentation method based on low-cost noise data, which trains a retinal blood vessel segmentation system through supervised and semi-supervised methods. However, existing semi-supervised methods still face the following challenges:

[0004] False label error propagation: The pseudo-label generated from unlabeled data may contain a large number of pixels with prediction errors, which will gradually accumulate in iterative training, ultimately leading to a decline in model performance.

[0005] Low-trust pixel unit interference: Even if high-quality image samples are retained in the pseudo-label, local errors can still affect the training effect. Some correctly predicted pixels may be included in the discarded pseudo-label samples, limiting the effectiveness of the training data and causing interference from low-trust pixel units during the training process.

[0006] Therefore, how to solve the problems of false label error propagation and low-trust pixel interference has become a research hotspot in the field of retinal blood vessel segmentation. To this end, the present application provides a retinal blood vessel segmentation system based on a pixel-level screening strategy. SUMMARY

[0007] In view of the deficiencies of the prior art, the present application provides a retinal blood vessel segmentation system based on a pixel-level screening strategy, which introduces high and low trust pixel units in the second pseudo-label and trains an advanced segmentation model based on high-trust pixel units to solve the technical problems raised in the background art.

[0008] To achieve the above purpose, the present application is implemented by the following technical solutions:

[0009] A retinal blood vessel segmentation system based on a pixel-level screening strategy, comprising:

[0010] An eye fundus image acquisition unit is configured to acquire an eye fundus image to be segmented;

[0011] An image preprocessing unit is configured to perform standardization preprocessing on the eye fundus image to be segmented to obtain a standardized image sample;

[0012] A binary classification label output unit is configured to input the standardized image sample into an advanced segmentation model trained by secondary supervision, output a blood vessel category probability of each pixel unit, and convert the blood vessel category probability threshold into a binary classification label of blood vessels or background.

[0013] The training of the advanced segmentation model comprises: performing binary classification supervised training on the pixel units of the labeled image samples or the high-confidence pixel units of the second pseudo-label; and further comprises: performing a loss masking operation on the low-confidence pixel units in the second pseudo-label samples.

[0014] In some specific embodiments, the secondary supervision training step of the advanced segmentation model comprises:

[0015] S1, loading N eye fundus image samples to be processed;

[0016] S2, labeling K labeled image samples in the N eye fundus image samples and predicting to generate M first pseudo-label samples;

[0017] Each pixel unit of the first pseudo-label sample is marked with a blood vessel category probability.

[0018] S3, screening and marking the pixel units of the first pseudo-label samples until M second pseudo-label samples are obtained; wherein the pixel units of the second pseudo-label samples are marked as high-confidence pixel units or low-confidence pixel units;

[0019] S4, merging the M second pseudo-label samples and the K labeled image samples to obtain J high-quality image samples; wherein J = M + K;

[0020] S5, inputting the J high-quality image samples into the basic segmentation model for secondary supervision training to generate an advanced segmentation model through iteration.

[0021] In some specific embodiments, labeling K labeled image samples in the N eye fundus image samples and predicting to generate M first pseudo-label samples comprises:

[0022] S2-1, selecting K eye fundus image samples from the N eye fundus image samples;

[0023] S2-2, performing binary classification labeling on the K selected eye fundus image samples, defining the K labeled image samples as K labeled image samples, and defining the eye fundus image samples not selected for labeling as unlabeled image samples;

[0024] wherein each pixel unit of the labeled image sample is labeled with a binary true label, and the binary true label represents that the class of the pixel unit in the fundus image sample is actually labeled as retinal blood vessel or retinal background;

[0025] S2-3, inputting the K labeled image samples into the U-Net network for initial supervised training, and iteratively generating a basic segmentation model; wherein the output target of the basic segmentation model is defined as the blood vessel class probability of the pixel unit;

[0026] S2-4, inputting the unlabeled image sample into the basic segmentation model for forward propagation inference, and predicting to generate M first pseudo-label samples.

[0027] In some specific embodiments, K selected fundus image samples are labeled with binary labels, defined as K labeled image samples; and the fundus image samples not selected for labeling are defined as unlabeled image samples, including:

[0028] S2-2-1, standardizing the N fundus image samples to obtain N standardized image samples;

[0029] S2-2-2, data augmentation is performed on the N standardized image samples to obtain M unlabeled image samples; M>N;

[0030] S2-2-3, selecting K unlabeled image samples from the M unlabeled image samples;

[0031] S2-2-4, performing binary labeling on the pixel units of the K unlabeled image samples to obtain K labeled image samples.

[0032] In some specific embodiments, the pixel units of the first pseudo-label sample are screened and marked until M second pseudo-label samples are obtained, including:

[0033] S3-1, performing median filtering on the first pseudo-label sample to generate a noise-free image sample;

[0034] S3-2, performing edge detection on the noise-free image sample to mark the pixel units as contour pixel units or non-contour pixel units;

[0035] S3-3, performing dynamic threshold judgment on the non-contour pixel units of the noise-free image sample to mark the non-contour pixel units as high-confidence pixel units or low-confidence pixel units;

[0036] S3-4, traversing all pixel units of the M first pseudo-label samples, repeatedly performing median filtering, edge detection and dynamic threshold judgment, until the M first pseudo-label samples are screened and marked as M second pseudo-label samples.

[0037] In some specific embodiments, the median filtering is performed on the first pseudo-label sample to generate a noise-free image sample, comprising:

[0038] S3-1-1, anchoring a to-be-labeled pixel unit of the first pseudo-label sample;

[0039] S3-1-2, constructing a square neighborhood window with the to-be-labeled pixel unit as the center of the window;

[0040] Wherein, the neighborhood window covers a plurality of adjacent pixel units in the first pseudo-label sample;

[0041] S3-1-3, obtaining pixel values of a plurality of adjacent pixel units in the neighborhood window, and constructing a pixel value sequence based on the pixel value sequence;

[0042] S3-1-4, selecting a pixel median value in the pixel value sequence, and replacing the pixel value of the to-be-labeled pixel unit with the pixel median value;

[0043] S3-1-5, traversing all pixel units of the first pseudo-label sample until the pixel values of all pixel units are replaced by the pixel median values in their neighborhood windows to generate the noise-free image sample.

[0044] In some specific embodiments, edge detection is performed on the noise-free image sample to label the pixel units as contour pixel units or non-contour pixel units, comprising:

[0045] S3-2-1, obtaining the gray value of each pixel unit in the noise-free image sample;

[0046] S3-2-2, labeling adjacent pixel units in the noise-free image sample and calculating the gray difference value of the adjacent pixel units;

[0047] S3-2-3, if the gray difference value is not 0, then it is determined that the adjacent pixel units are all contour pixel units; otherwise, it is determined as non-contour pixel units;

[0048] S3-2-4, traversing all pixel units of the noise-free image sample to label the pixel units as contour pixel units or non-contour pixel units.

[0049] In some specific embodiments, dynamic threshold judgment is performed on the non-contour pixel units of the noise-free image sample to label the non-contour pixel units as high-confidence pixel units or low-confidence pixel units, comprising:

[0050] S3-3-1, selecting a non-contour pixel unit in the noise-free image sample and labeling it as a current pixel unit;

[0051] S3-3-2, obtaining the current pixel unit to the current iteration number;

[0052] S3-3-3、According to the current iteration number, the dynamic probability threshold of the current pixel unit is calculated;

[0053] S3-3-4, the blood vessel category probability of the current pixel unit is obtained;

[0054] S3-3-5, if the blood vessel category probability is greater than the dynamic probability threshold, the current pixel unit is labeled as a high-confidence pixel unit, otherwise it is labeled as a low-confidence pixel unit.

[0055] In some specific embodiments, J high-quality image samples are input into the basic segmentation model for secondary supervision training, and an advanced segmentation model is generated through iteration, comprising:

[0056] S5-1, mark the current iteration pixel unit of the high-quality image sample;

[0057] S5-2, according to the unit category inherited by the current iteration pixel unit, perform binary classification supervision training or loss shielding operation;

[0058] S5-3, if the unit category is a pixel unit of a labeled image sample or a high-confidence pixel unit of a second pseudo label, perform binary classification supervision training based on the pixel unit of the labeled image sample or the high-confidence pixel unit of the second pseudo label.

[0059] S5-4, if the sample category is a low-confidence pixel unit of a second pseudo label, perform loss shielding operation on the low-confidence pixel unit in the second pseudo label sample.

[0060] In some specific embodiments, the shielding operation of loss calculation on the low-confidence pixel unit in the second pseudo label sample comprises:

[0061] S5-4-1, anchor the low-confidence pixel unit;

[0062] S5-4-2, forward propagate the low-confidence pixel unit to predict the blood vessel category probability of the low-confidence pixel unit;

[0063] S5-4-3, skip the cross-entropy loss calculation of the low-confidence pixel unit, and directly enter the secondary supervision training of the next iteration pixel unit.

[0064] The application provides a retinal blood vessel segmentation system based on a pixel-level screening strategy, which has the following beneficial effects:

[0065] The application effectively expands the scale of training samples by generating second pseudo-label samples, improves the training effect of the model on limited labeled fundus images by generating high-quality pseudo-labels from unlabeled images, reduces the dependence on a large amount of labeled data, and still improves the accuracy of retinal blood vessel segmentation in the case of data scarcity. Further, when processing low-confidence pixel units in the second pseudo-label sample, a loss masking operation is used to skip the loss calculation of these low-quality pixel units, thereby avoiding the interference of low-confidence data on the training process. The training process is more focused on the optimization of high-quality high-confidence pixel units, avoiding the negative impact of error propagation. Finally, the generated advanced segmentation model exhibits higher accuracy in the blood vessel segmentation task. BRIEF DESCRIPTION OF DRAWINGS

[0066] Figure 1 A structural block diagram of a retinal blood vessel segmentation system based on a pixel-level screening strategy according to the application is shown in FIG. 1.

[0067] Figure 2 A retinal blood vessel segmentation process according to the application is shown in FIG. 2.

[0068] Figure 3 A training diagram of the advanced segmentation model according to the application is shown in FIG. 3.

[0069] Figure 4 A labeling process diagram of the second pseudo-label sample according to the application is shown in FIG. 4.

[0070] Figure 5 A shielding operation diagram of the loss calculation of the low-confidence pixel unit according to the application is shown in FIG. 5.

[0071] Figure 6 A contrast diagram of the fundus image in an embodiment according to the application is shown in FIG. 6. DETAILED DESCRIPTION

[0072] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only a part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the application.

[0073] Please refer to Figures 1 to 6 The application provides a retinal blood vessel segmentation system based on a pixel-level screening strategy, which comprises:

[0074] A fundus image acquisition unit is configured to acquire a fundus image to be segmented.

[0075] An image preprocessing unit is configured to perform standardization preprocessing on the fundus image to be segmented to obtain a standardized image sample;

[0076] A binary classification label output unit is configured to input the standardized image sample into the advanced segmentation model trained by secondary supervision, output a blood vessel category probability of each pixel unit, and convert the blood vessel category probability threshold into a binary classification label of blood vessels or background.

[0077] The training of the advanced segmentation model comprises: performing binary classification supervised training on the pixel units of the labeled image sample or the high-confidence pixel units of the second pseudo-label; and further comprises: performing a loss masking operation on the low-confidence pixel units in the second pseudo-label sample.

[0078] In this embodiment, the secondary supervision training step of the advanced segmentation model comprises:

[0079] S1, loading N fundus image samples to be processed;

[0080] S2, labeling K labeled image samples in the N fundus image samples and predicting to generate M first pseudo-label samples;

[0081] Each pixel unit of the first pseudo-label sample is marked with a blood vessel category probability.

[0082] S3, screening and marking the pixel units of the first pseudo-label sample until M second pseudo-label samples are obtained; wherein the pixel units of the second pseudo-label sample are marked as high-confidence pixel units or low-confidence pixel units.

[0083] S4, merging the M second pseudo-label samples and the K labeled image samples to obtain J high-quality image samples; wherein J = M + K.

[0084] S5, inputting the J high-quality image samples into the basic segmentation model for secondary supervision training to generate an advanced segmentation model through iteration.

[0085] In this embodiment, through secondary supervision training, the segmentation performance can be optimized when processing the high-confidence pixel units of the labeled samples and the second pseudo-label samples. During the training process, the model focuses on learning the accurately labeled pixel units, thereby improving the accuracy of blood vessel segmentation. For the low-confidence pixel units in the second pseudo-label sample, a loss masking operation is adopted. In this operation, although the low-confidence pixel units generate prediction results, the calculation of their cross-entropy loss is skipped, avoiding the negative impact of low-quality pixel units on model training, effectively preventing the interference of low-confidence data on the training process, ensuring the accuracy and stability of the training, and finally generating an advanced segmentation model that can more accurately perform retinal blood vessel segmentation on fundus images, thereby improving the image segmentation level of retinal blood vessels.

[0086] Exemplarily, the step S2 specifically comprises:

[0087] S2-1, selecting K fundus image samples from the N fundus image samples;

[0088] S2-2, performing binary classification labeling on the K selected fundus image samples, defining the K fundus image samples as labeled image samples; and defining the fundus image samples not selected for labeling as unlabeled image samples;

[0089] wherein each pixel unit of the labeled image samples is labeled with a binary true label, and the binary true label represents that the class of the pixel unit in the fundus image sample is actually labeled as retinal blood vessels or retinal background;

[0090] S2-3, inputting the K labeled image samples into a U-Net network for initial supervised training, and generating a basic segmentation model through iteration; wherein the output target of the basic segmentation model is defined as the blood vessel class probability of the pixel unit;

[0091] Specifically, the U-Net network is a classical convolutional neural network architecture, which is widely used in medical image segmentation tasks. U-Net can effectively process image segmentation tasks, especially when processing medical images (such as fundus images), it can accurately segment the blood vessel region.

[0092] S2-4, inputting the unlabeled image samples into the basic segmentation model for forward propagation inference to predict M first pseudo-label samples; specifically, each unlabeled image sample is assigned a sample index, and the first pseudo-label sample shares the same sample index with the unlabeled image sample, so as to form a mapping between the unlabeled image sample and the first pseudo-label sample.

[0093] In this embodiment, the pixel units of the K fundus image samples are first explicitly labeled with binary classification, and a basic segmentation model is generated through initial supervised training by the U-Net network. The difference is that the basic segmentation model selects the blood vessel class probability of each pixel unit as the output; then, the unlabeled image samples are input into the basic segmentation model as the input, and finally the first pseudo-label samples labeled with the blood vessel class probability are output; overall, a large number of pixel units covered by a small number of labeled image samples can supervise the training of the basic segmentation model, thereby improving the quality of the first pseudo-label in the semi-supervised process, and ultimately improving the accuracy of the blood vessel segmentation model.

[0094] Specifically, the step S2-2 further comprises:

[0095] S2-2-1, performing standardization preprocessing on the N fundus image samples to obtain N standardized image samples;

[0096] Specifically, the standardized preprocessing includes:

[0097] Uniformly normalize the pixel gray value of each image to adjust the pixel value to the range of [0, 1] to eliminate the brightness difference between images. Adopt histogram equalization or adaptive contrast enhancement method to enhance the contrast of the image, so that the contrast between blood vessels and background is more obvious, and the divisibility of blood vessels in the image is enhanced.

[0098] S2-2-2, data augmentation is performed on the N standardized image samples to obtain M unlabeled image samples; M>N;

[0099] Specifically, the data augmentation includes:

[0100] Rotation: randomly rotate the image, the rotation angle is between -30° and 30°, and the adaptability of the model to different angle images is enhanced.

[0101] Flip: randomly flip the image horizontally and vertically to simulate the blood vessel morphology under different shooting directions.

[0102] Scaling: randomly scale the image, the scaling ratio is within a certain range, which helps the model to identify blood vessels under different scales.

[0103] Translation: randomly translate the image to make the blood vessel position in the image more diverse and improve the generalization ability of the model.

[0104] Brightness and contrast adjustment: randomly adjust the brightness and contrast of the image to simulate images under different lighting environments.

[0105] S2-2-3, select K unlabeled image samples from the M unlabeled image samples;

[0106] Specifically, the unlabeled image sample should be a sub-block with uniform size (such as 96x96 pixels). Such a size-uniform image sub-block helps to reduce the complexity of data processing and improve the stability of training.

[0107] S2-2-4, perform binary classification labeling on the pixel units of the K unlabeled image samples to obtain K labeled image samples.

[0108] The embodiment improves the diversity of training data through standardized preprocessing and data enhancement. Standardized preprocessing unifies the gray value of image pixels, eliminates brightness differences, and improves the separability of blood vessels and background through contrast enhancement. Through data enhancement such as rotation, flipping, scaling, etc., more diverse unlabeled image samples are generated, enhancing the model's adaptability to blood vessels under different conditions. Finally, K unlabeled samples are selected and processed for binary classification labeling to obtain labeled images for supervised training of the U-Net network and sample generation in the inference stage.

[0109] In an exemplary embodiment, the step S3 specifically comprises:

[0110] S3-1, performing median filtering on the first pseudo-label sample to generate a noise-free image sample;

[0111] S3-2, performing edge detection on the noise-free image sample to mark the pixel units as contour pixel units or non-contour pixel units;

[0112] S3-3, performing dynamic threshold judgment on the non-contour pixel units of the noise-free image sample to mark the non-contour pixel units as high-confidence pixel units or low-confidence pixel units;

[0113] S3-4, traversing all pixel units of the M first pseudo-label samples, repeatedly performing median filtering, edge detection and dynamic threshold judgment until the M first pseudo-label samples are filtered and marked as M second pseudo-label samples.

[0114] Specifically, the step S3-1 further comprises:

[0115] S3-1-1, anchoring the pixel unit to be marked of the first pseudo-label sample;

[0116] S3-1-2, constructing a square neighborhood window with the pixel unit to be marked as the window center;

[0117] Wherein, the neighborhood window covers several adjacent pixel units in the first pseudo-label sample;

[0118] S3-1-3, obtaining the pixel values of the several adjacent pixel units in the neighborhood window, and constructing a pixel value sequence based on the pixel value sorting;

[0119] S3-1-4, selecting the pixel median value in the pixel value sequence, and replacing the pixel value of the pixel unit to be marked with the pixel median value;

[0120] S3-1-5, traversing all pixel units of the first pseudo-label sample until the pixel values of all pixel units are replaced by the pixel median values in their neighborhood windows to generate the noise-free image sample.

[0121] The embodiment generates a noise-free image sample by performing median filtering on the first pseudo-label sample. Specifically, first, each pixel unit to be labeled is operated on, and a square neighborhood window is constructed centered on the pixel unit to be labeled, covering multiple adjacent pixel units. Then, the pixel values in the neighborhood window are obtained and sorted to generate a sequence of pixel values, and the median value in the sequence is selected and replaced as the value of the pixel unit to be labeled. By traversing the entire first pseudo-label sample, it is ensured that the value of each pixel unit is replaced by the median value in the neighborhood window, thereby eliminating noise and generating a noise-free image sample. This embodiment effectively removes noise in the image, improves the quality of the image, and thus improves the accuracy of the training of the blood vessel segmentation model.

[0122] Specifically, the step S3-2 further includes:

[0123] S3-2-1, obtaining the gray value of each pixel unit in the noise-free image sample;

[0124] Wherein, the gray value is 0 or 255, representing the background or foreground (blood vessels) respectively.

[0125] S3-2-2, marking adjacent pixel units in the noise-free image sample and calculating the gray difference value of the adjacent pixel units;

[0126] S3-2-3, if the gray difference value is not 0, it is determined that the adjacent pixel units are all contour pixel units; otherwise, it is determined as non-contour pixel units;

[0127] S3-2-4, traversing all pixel units of the noise-free image sample, marking the pixel units as contour pixel units or non-contour pixel units.

[0128] The embodiment effectively distinguishes the contour pixel units and non-contour pixel units in the image by performing edge detection on the noise-free image sample. First, the gray value of each pixel unit is obtained, and 0 represents the background and 255 represents the blood vessel foreground. The boundary is identified by the difference in gray value. If the gray difference value of adjacent pixels is not 0, it is marked as a contour pixel unit, otherwise it is marked as a non-contour pixel unit. This process ensures that the boundary between the blood vessels and the background is effectively captured. By traversing all pixel units and judging whether they are contour pixel units according to the gray difference, the embodiment can clearly identify the boundary region in the image, providing accurate edge information for blood vessel segmentation.

[0129] Specifically, the step S3-3 further includes:

[0130] S3-3-1, selecting a non-contour pixel unit in the noise-free image sample and marking it as the current pixel unit;

[0131] S3-3-2, obtain a current pixel unit to a current iteration number;

[0132] S3-3-3, calculate a dynamic probability threshold of the current pixel unit according to the current iteration number;

[0133] The calculation formula of the dynamic probability threshold is:

[0134]

[0135] Wherein, T represents the dynamic probability threshold, I represents the iteration number, and z represents the number of image samples of the fundus image;

[0136] Specifically, when the iteration number I is 1 or 2, the dynamic probability threshold is set to the reference value 0.5; and when the iteration number is greater than 2 and the number of fundus image samples is less than 20, the dynamic probability threshold of the current pixel unit is calculated according to the above calculation formula. Therefore, in this embodiment, it is required that the number of fundus image samples in each training process is maintained below 20. The introduction of the dynamic probability threshold is because in the training process, the proportion of background pixels is usually higher than that of blood vessel pixels, resulting in that the response of the model to the background is gradually enhanced with the increase of the iteration number, so that the pixels originally belonging to the blood vessels are incorrectly predicted as the background. With the progress of iteration, the prediction probability of part of the blood vessel pixels gradually decreases, and even lower than the predetermined reference threshold. In order to solve this problem, through experimental design, the calculation method of the above dynamic probability threshold is proposed, which can effectively inhibit the decrease of the prediction probability of the blood vessel pixels caused by the increase of the iteration number, so as to improve the accuracy of the blood vessel segmentation.

[0137] S3-3-4, obtain a blood vessel class probability of the current pixel unit;

[0138] S3-3-5, if the blood vessel class probability is greater than the dynamic probability threshold, the current pixel unit is marked as a high-confidence pixel unit, otherwise it is marked as a low-confidence pixel unit.

[0139] The embodiment solves the problem of increasing proportion of background pixels in training through the dynamic probability threshold, avoids the misjudgment of blood vessel pixels, and enhances the accuracy of the model in different iteration stages. Through this embodiment, the high-confidence pixel unit and the low-confidence pixel unit are accurately distinguished, and the accuracy of the blood vessel segmentation is improved.

[0140] Exemplarily, the step S5 specifically comprises:

[0141] S5-1, mark a current iteration pixel unit of a high-quality image sample;

[0142] S5-2, perform binary classification supervised training or loss shielding operation according to the unit class inherited by the current iteration pixel unit;

[0143] Specifically, each pixel unit in the high-quality image sample has been explicitly labeled with its cell class before the training starts. In this embodiment, the pixel units of the high-quality image sample inherit their cell class before merging, i.e., the class of each pixel (such as blood vessels or background) will remain unchanged after merging into the high-quality sample.

[0144] That is, in the high-quality image sample, the class of the pixel unit has been labeled through previous steps (such as probability-based screening or pseudo-label generation). These class information will be directly inherited as the attribute of the pixel unit, without the need for additional processing operations. Therefore, further supervised training or screening operations can be directly performed according to these labeled class information during the training process.

[0145] S5-3, if the cell class is a pixel unit of a labeled image sample or a high-confidence pixel unit of a second pseudo-label, performing binary classification supervised training on the pixel unit of the labeled image sample or the high-confidence pixel unit of the second pseudo-label;

[0146] S5-4, if the sample class is a low-confidence pixel unit of a second pseudo-label, performing loss masking operation on the low-confidence pixel unit in the second pseudo-label sample.

[0147] This embodiment inputs J high-quality image samples into the basic segmentation model, performs secondary supervised training, and generates an advanced segmentation model through iteration. During the training process, each pixel unit in the high-quality image sample inherits the class information before merging, ensuring that additional labeling is not required when merging the high-quality sample, but the class information generated in the previous step is directly used.

[0148] During training, for pixel units of labeled image samples and high-confidence pixel units of second pseudo-labels, standard binary classification supervised training is performed. For low-confidence pixel units of second pseudo-labels, loss masking operation is performed to skip the loss calculation of these pixel units, thereby avoiding the influence of low-confidence pixel units on model training.

[0149] In this embodiment, the step S5-3 further comprises:

[0150] S5-3-1, anchoring the current pixel unit in the currently recognized pixel unit of the labeled image sample or the high-confidence pixel unit of the second pseudo-label;

[0151] S5-3-2, performing forward propagation on the current pixel unit to predict the blood vessel class probability of the current pixel unit;

[0152] S5-3-3, calculating the cross-entropy loss of the blood vessel class probability and the true label of the current pixel unit;

[0153] S5-3-4, calculate the gradient of the cross-entropy loss with respect to the model parameters;

[0154] S5-3-5, update the model parameters based on the gradient to generate updated parameters;

[0155] S5-3-6, anchor the next iteration pixel unit in the current identified labeled image sample;

[0156] S5-3-7, perform forward propagation on the next iteration pixel unit according to the updated parameters to predict and generate the blood vessel class probability of the next iteration pixel unit.

[0157] S5-3-8, repeat the calculation of the cross-entropy loss and the gradient in the next iteration until the updated parameters make the basic segmentation model converge into an advanced segmentation model.

[0158] The embodiment continuously optimizes and adjusts the parameters of the basic segmentation model through standard binary classification supervised training to improve the segmentation accuracy of the blood vessel image.

[0159] In the embodiment, the step S5-4 further comprises:

[0160] S5-4-1, anchor the low-confidence pixel unit;

[0161] S5-4-2, forward propagate the low-confidence pixel unit to predict and generate the blood vessel class probability of the low-confidence pixel unit;

[0162] S5-4-3, skip the cross-entropy loss calculation of the low-confidence pixel unit and directly enter the secondary supervised training of the next iteration pixel unit.

[0163] In the embodiment, for the low-confidence pixel unit in the second pseudo-label sample, the loss masking operation effectively avoids the influence of the low-confidence pixel unit on the training. Specifically, although these low-confidence pixel units generate blood vessel class probabilities through forward propagation, their cross-entropy loss calculation is skipped and does not participate in gradient updating. Through this method, the negative impact of the low-confidence pixel unit on parameter updating can be avoided, so as to focus more on the training and optimization of the high-confidence pixel unit; this embodiment ensures that the optimization in the training process focuses on high-quality pixel unit samples, avoids training bias caused by unreliable data, improves the accuracy of the blood vessel segmentation model, and finally generates an advanced segmentation model that can more accurately perform retinal blood vessel segmentation on the fundus image, thereby improving the segmentation accuracy of the retinal blood vessels in the fundus image.

[0164] The above-described embodiments can be implemented in part or in whole through software, hardware, firmware or any combination thereof. When implemented in software, the above-described embodiments can be implemented in the form of a computer program product that includes one or more computer instructions or computer programs. When loaded and executed by a computer, the computer instructions or computer programs can produce the processes or functions described above in accordance with the embodiments of the present application. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable apparatus. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, from a website, a computer, a server or a data center to another website, computer, server or data center through a wired (for example, infrared, wireless, microwave or the like) manner.

[0165] The computer-readable storage medium can be any available medium or a collection of medium accessible by a computer or a data storage device such as a server, a data center or the like that includes one or more of the available medium. The available medium can be a magnetic medium (for example, a floppy disk, a hard disk, a magnetic tape), an optical medium (for example, a DVD), or a semiconductor medium. The semiconductor medium can be a solid state disk.

[0166] In several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented in other ways. For example, the above-described device embodiments are merely illustrative, for example, a plurality of units or components can be combined or integrated into another system, or some features can be omitted or not executed. In addition, the coupling or direct coupling or communication connection between the shown or discussed units can be indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0167] The above description is merely a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application.

Claims

1. A retinal vessel segmentation system based on a pixel-level screening strategy, characterized in that, include: The fundus image acquisition unit is used to acquire fundus images to be segmented; The image preprocessing unit is used to perform standardized preprocessing on the fundus image to be segmented to obtain standardized image samples; The binary classification label output unit is used to input standardized image samples into the advanced segmentation model trained under secondary supervision, output the blood vessel category probability of each pixel unit, and convert the blood vessel category probability into a binary classification label of blood vessel or background. The training of the advanced segmentation model includes: performing binary classification-based supervised training on pixel units of labeled image samples or high-confidence pixel units of the second pseudo-label; and also includes: performing a loss masking operation on low-confidence pixel units in the second pseudo-label samples. The secondary supervised training steps of the advanced segmentation model include: S1. Load the N fundus image samples to be processed; S2. Label K labeled image samples from N fundus image samples, and predict and generate M first pseudo-label samples; In the first pseudo-label sample, each pixel unit is labeled with the probability of blood vessel category; S3. Filter and mark the pixel units of the first pseudo-label sample until M second pseudo-label samples are obtained; Among them, the pixel units of the second pseudo-label sample are marked as high-confidence pixel units or low-confidence pixel units; S4. Merge the M second pseudo-labeled samples with the K labeled image samples to obtain J high-quality image samples; where J = M + K; S5. Input J high-quality image samples into the basic segmentation model for secondary supervised training, and generate an advanced segmentation model through iteration. The step of inputting J high-quality image samples into the basic segmentation model for secondary supervised training, and iteratively generating an advanced segmentation model, includes: S5-1, Mark the current iteration pixel unit of the high-quality image sample; S5-2. Based on the unit category inherited by the current iteration pixel unit, perform binary classification supervised training or loss masking operation; S5-3. If the unit category is a pixel unit of a labeled image sample or a highly reliable pixel unit of the second pseudo-label, then perform binary classification-based supervised training on the pixel unit of the labeled image sample or the highly reliable pixel unit of the second pseudo-label. S5-4. If the sample category is a low-confidence pixel unit of the second pseudo-label, then perform a loss masking operation on the low-confidence pixel units in the second pseudo-label sample.

2. The retinal vessel segmentation system based on a pixel-level screening strategy according to claim 1, characterized in that, From N fundus image samples, label K labeled image samples and predict and generate M first pseudo-label samples, including: S2-1. Select K fundus image samples from N fundus image samples; S2-2. Perform binary classification and labeling on the K selected fundus image samples, defining them as K labeled image samples; and define the fundus image samples that are not selected for labeling as unlabeled image samples; Among them, the pixel units of the labeled image samples are all labeled with binary true labels. The binary true labels indicate that the category of the pixel unit in the fundus image sample is actually labeled as retinal vessels or retinal background. S2-3. Input K labeled image samples into the U-Net network for initial supervised training, and generate a basic segmentation model through iteration; wherein, the output target of the basic segmentation model is defined as the blood vessel category probability of pixel unit; S2-4. Input the unlabeled image samples into the basic segmentation model for forward propagation inference to predict and generate M first pseudo-label samples.

3. The retinal vessel segmentation system based on a pixel-level screening strategy according to claim 2, characterized in that, K selected fundus image samples are binary-classified and labeled, defined as K labeled image samples; and unlabeled fundus image samples are defined as unlabeled image samples, including: S2-2-1. Perform standardized preprocessing on N fundus image samples to obtain N standardized image samples; S2-2-2. Perform data augmentation on N standardized image samples to obtain M unlabeled image samples; M > N; S2-2-3. Select K unlabeled image samples from M unlabeled image samples; S2-2-4. Perform binary classification labeling on the pixel units of the K unlabeled image samples to obtain K labeled image samples.

4. The retinal vessel segmentation system based on a pixel-level screening strategy according to claim 1, characterized in that, The pixel units of the first pseudo-label sample are filtered and labeled until M second pseudo-label samples are obtained, including: S3-1. Perform median filtering on the first pseudo-label sample to generate a noise-free image sample; S3-2. Perform edge detection on noise-free image samples to label pixel units as contour pixel units or non-contour pixel units. S3-3. Perform dynamic threshold determination on non-contour pixel units of noise-free image samples to mark non-contour pixel units as high-confidence pixel units or low-confidence pixel units. S3-4. Traverse all pixel units of the M first pseudo-label samples, and repeatedly perform median filtering, edge detection and dynamic threshold determination until the M first pseudo-label samples are selected and marked as M second pseudo-label samples.

5. The retinal vessel segmentation system based on a pixel-level screening strategy according to claim 4, characterized in that, Median filtering is performed on the first pseudo-label sample to generate noise-free image samples, including: S3-1-1, Anchor the pixel unit to be labeled in the first pseudo-label sample; S3-1-2. Construct a square neighborhood window with the pixel unit to be labeled as the center of the window; The neighborhood window covers several adjacent pixel units in the first pseudo-label sample; S3-1-3. Obtain the pixel values ​​of several adjacent pixel units within the neighborhood window, and construct a pixel value sequence based on the sorted pixel values; S3-1-4. Select the median value of the pixels in the pixel value sequence and replace the pixel value of the pixel unit to be marked with the median value; S3-1-5. Traverse all pixel units of the first pseudo-label sample until the pixel value of all pixel units is replaced by the median value of the pixels in its neighborhood window to generate the noise-free image sample.

6. The retinal vessel segmentation system based on a pixel-level screening strategy according to claim 5, characterized in that, Performing edge detection on noise-free image samples to label pixel units as contour pixel units or non-contour pixel units includes: S3-2-1. Obtain the gray value of each pixel unit in a noise-free image sample; S3-2-2, In a noise-free image sample, mark adjacent pixel units and calculate the gray-level difference between adjacent pixel units; S3-2-3. If the grayscale difference is not 0, then the adjacent pixel units are determined to be contour pixel units; otherwise, they are determined to be non-contour pixel units. S3-2-4. Traverse all pixel units of the noise-free image sample and mark the pixel units as contour pixel units or non-contour pixel units.

7. The retinal vessel segmentation system based on a pixel-level screening strategy according to claim 6, characterized in that, Dynamic thresholding is performed on non-contour pixel units of noise-free image samples to mark them as high-confidence or low-confidence pixel units, including: S3-3-1. Select a non-contour pixel unit in a noise-free image sample and mark it as the current pixel unit; S3-3-2, Obtain the current pixel unit to the current iteration number; S3-3-3. Calculate the dynamic probability threshold of the current pixel unit based on the current iteration number; S3-3-4. Obtain the probability of blood vessel category for the current pixel unit; S3-3-5. If the probability of the blood vessel category is greater than the dynamic probability threshold, then the current pixel unit is marked as a high-confidence pixel unit; otherwise, it is marked as a low-confidence pixel unit.

8. The retinal vessel segmentation system based on a pixel-level screening strategy according to claim 7, characterized in that, A masking operation is performed on the low-confidence pixel units in the second pseudo-label sample, including: S5-4-1, Anchoring low-confidence pixel units; S5-4-2, Forward propagation of low-confidence pixel units to predict the probability of blood vessel category that generates low-confidence pixel units; S5-4-3. Skip the cross-entropy loss calculation for low-confidence pixel units and directly proceed to the secondary supervised training of the next iteration pixel units.

Images