Retinal image-based blood vessel segmentation method and system

Abstract

The application discloses a blood vessel segmentation method and system based on a retinal image, and belongs to the technical field of image blood vessel segmentation, and comprises the following steps: a first encoder-decoder receives a retinal image requiring blood vessel segmentation, adjusts a feature representation dimension, analyzes an abnormal score of artifact noise detection of a first target image, determines interpolation size adjustment requirement information, performs feature compression and reconstruction to output a blood vessel segmentation image, and adaptively adjusts a learning rate based on image segmentation image quality representation parameters. The application significantly improves the weak blood vessel detection capability through feature representation dimension adjustment and feature compression and reconstruction, and solves the problem of insufficient weak blood vessel detection in the prior art.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image blood vessel segmentation, in particular to a blood vessel segmentation method and system based on a retinal image. BACKGROUND

[0002] Retinal blood vessel segmentation is a core technology in ophthalmic medical image analysis, which captures retinal blood vessel details through fundus photography for the diagnosis of diabetic retinopathy, hypertensive retinopathy and retinopathy of prematurity. Traditional methods rely on manual segmentation or rule-based automation techniques such as matched filtering and edge detection. In recent years, deep learning technology has significantly improved segmentation performance, especially models based on convolutional neural networks such as U-Net, which improve feature transmission and preservation through skip connections and become the mainstream method.

[0003] For example, the invention patent with publication number CN110689526B discloses a retinal blood vessel segmentation method and system based on a retinal fundus image, which includes the following steps: obtaining a retinal fundus image to be detected; constructing a basic module according to the features of the retinal fundus image; using N basic modules in cascade as the final network model, inputting the retinal fundus image to be detected as the input of the overall network model, and obtaining the segmentation result of the retinal blood vessels. By inputting the foreground features of the previous basic module and the original image into the next basic module, the subsequent basic module can inherit the learning experience of the previous basic module, thereby accelerating the training process and effectively solving the problem of data imbalance. Inputting the retinal fundus image to be detected as the input of the overall model S-UNet, the segmentation result of the retinal blood vessels is more accurate.

[0004] For example, the invention patent with publication number CN115222638B discloses a retinal blood vessel image segmentation method and system based on a neural network model, which includes: first, obtaining a plurality of retinal blood vessel images of the same position of the fundus retina in different shooting modes, performing image fusion to obtain a fused image; performing image preprocessing on the fused image to obtain a to-be-segmented image; training a blood vessel image segmentation model and a blood vessel target detection model; inputting the to-be-segmented image into the trained blood vessel image segmentation model and the trained blood vessel target detection model; comparing the output results of the blood vessel image segmentation model and the blood vessel target detection model, when the output results are inconsistent, returning to continue training the model, and when the output results are consistent, outputting the segmentation result.

[0005]

[0006] However, in the process of implementing the technical scheme of the present application, the present application has found that the above-mentioned technology at least has the following technical problems:

[0007] ​1. Weak vascular signal detection is insufficient: In images with high noise or low contrast, traditional methods and some deep learning models are difficult to accurately capture small blood vessel structures.

[0008] 2. Insufficient generalization ability across datasets: Differences in resolution, contrast, and pathological features between datasets such as DRIVE and CHASE_DB1 result in unstable performance of existing models on different datasets.

[0009] 3. Target object and background segmentation is not clear: In complex blood vessel networks, existing methods often produce missegmentation, resulting in unclear image boundary division and affecting diagnostic accuracy SUMMARY

[0010] The first aspect of the embodiment of the application provides a blood vessel segmentation method based on a retinal image, comprising the following steps:

[0011] The first encoder-decoder receives a retinal image requiring blood vessel segmentation, denoted as a first target image, collects image feature parameters of the first target image, analyzes image feature indicators of the first target image, and adjusts the feature representation dimension of the first target image accordingly.

[0012] After adjusting the feature representation dimension, collect the artifact noise data of the first target image, analyze the artifact noise detection abnormal score of the first target image, determine the interpolation size adjustment requirement information, and analyze the effect of feature representation dimension adjustment based on the interpolation size adjustment requirement information.

[0013] Perform feature compression and reconstruction, and make the second encoder-decoder receive the output image of the first encoder-decoder through symmetric skip connection, output a blood vessel segmentation map, and denote it as a second target image.

[0014] Collect the segmentation image quality data of the second target image, analyze the segmentation image quality representation parameters of the second target image, and perform adaptive adjustment of the learning rate based on the segmentation image quality representation parameters of the second target image.

[0015] Further, the feature representation dimension of the first target image is adjusted, and the specific analysis method is as follows:

[0016] According to the image feature parameters of the first target image, analyze the image feature indicators of the first target image.

[0017] Extract the image feature comprehensive threshold stored in the database.

[0018] Subtract the image feature comprehensive threshold from the image feature indicators of the first target image to obtain the image feature deviation factor of the first target image.

[0019] An image feature deviation factor interval corresponding to the representation dimension adjustment set is extracted from the first target image stored in the database, and a representation dimension adjustment set corresponding to the interval in which the image feature deviation factor of the first target image is located is mapped, denoted as a feature representation dimension adjustment set.

[0020] The feature representation dimension adjustment set includes a spatial dimension adjustment value and a channel depth adjustment value.

[0021] The current spatial dimension and the current channel depth are extracted.

[0022] The feature representation dimension adjustment of the first target image is performed according to the current spatial dimension, the current channel depth, and the feature representation dimension adjustment set.

[0023] Further, the image feature indicators of the first target image are analyzed, and the specific analysis method is as follows:

[0024] The image feature parameters include the intersection-over-union ratio, the dice coefficient, and the accuracy.

[0025] The image feature indicators of the first target image are analyzed based on the image feature parameters.

[0026] The image feature indicators of the first target image are the quantification indicators of the influence degree of the intersection-over-union ratio, the dice coefficient, and the accuracy on the image feature state, and the specific analysis process is as follows: the intersection-over-union ratio, the dice coefficient, and the accuracy are respectively processed by ratio with the corresponding reference values, and each ratio processing result is respectively combined with the corresponding feature allocation coefficient and coupled to obtain the image feature indicators of the first target image.

[0027] Further, the interpolation size adjustment requirement information is determined, and the specific analysis method is as follows:

[0028] After the feature representation dimension adjustment, the artifact noise parameters of the first target image are obtained.

[0029] The artifact noise parameters include the signal-to-noise ratio, the structural similarity, the average gradient amplitude, and the gradient direction change rate.

[0030] The image artifact noise detection abnormal score is analyzed based on the image artifact noise parameters.

[0031] The image artifact noise detection abnormality score is a quantitative index of the influence degree of the signal-to-noise ratio, the structural similarity, the average gradient amplitude and the gradient direction change rate on the image artifact noise abnormality, and the specific analysis process is as follows: the reference values corresponding to the signal-to-noise ratio and the structural similarity are respectively processed by ratio with the signal-to-noise ratio and the structural similarity, the gradient direction change rate is processed by ratio with the corresponding reference value, when the average gradient amplitude is greater than the defined value, the average gradient amplitude is processed by ratio with the reference value corresponding to the average gradient amplitude, when the average gradient amplitude is less than or equal to the defined value, the reference value corresponding to the average gradient amplitude is processed by ratio with the average gradient amplitude, and each ratio processing result is coupled by combining with the corresponding feature allocation coefficient, so as to obtain the image artifact noise detection abnormality score of the first target image.

[0032] The abnormality score threshold value in the database is extracted.

[0033] If the image artifact noise detection abnormality score is less than the abnormality score threshold value, the interpolation size adjustment requirement information determination result of the first target image is recorded as no interpolation size adjustment requirement.

[0034] If the image artifact noise detection abnormality score is greater than or equal to the abnormality score threshold value, the interpolation size adjustment requirement information determination result of the first target image is recorded as requiring interpolation size adjustment.

[0035] Further, the feature representation dimension adjustment effect is analyzed based on the interpolation size adjustment requirement information determination result, and the specific analysis method is as follows:

[0036] If the interpolation size adjustment requirement information determination result is no interpolation size adjustment requirement, the first image feature correction factor is extracted according to the image artifact noise detection abnormality score, and the image feature correction value is obtained according to the image feature index of the first target image and the first image feature correction factor.

[0037] If the interpolation size adjustment requirement information determination result is requiring interpolation size adjustment, interpolation size adjustment is performed, and the image artifact noise detection abnormality score is reacquired.

[0038] The image feature index is reacquired.

[0039] The image feature correction value is analyzed based on the reacquired image artifact noise detection abnormality score and the reacquired image feature index.

[0040] The image feature comprehensive threshold value in the database is extracted.

[0041] The image feature correction value is processed by difference with the image feature comprehensive threshold value to obtain the image feature correction deviation factor.

[0042] The ideal interval of the image feature correction deviation factor in the database is extracted.

[0043] If the image feature correction deviation factor is within the image feature correction deviation factor ideal interval, the feature representation dimension adjustment effect analysis result is recorded as effective adjustment, and prompt information is generated synchronously.

[0044] If the image feature correction deviation factor is outside the image feature correction deviation factor ideal interval, the feature representation dimension adjustment effect analysis result is recorded as invalid adjustment, and a warning message is generated synchronously.

[0045] Further, interpolation size adjustment is performed, and the specific analysis method is as follows:

[0046] Subtract the image artifact noise detection anomaly score threshold from the image artifact noise detection anomaly score to obtain the image artifact noise detection anomaly deviation score.

[0047] Extract the interpolation size adjustment value corresponding to each image artifact noise detection anomaly deviation score interval stored in the database, and map the interpolation size adjustment value corresponding to the interval in which the image artifact noise detection anomaly deviation score is located, and record it as the interpolation size adjustment value of the first target image.

[0048] Extract the current image interpolation size value.

[0049] Adjust the interpolation size of the first target image according to the current image interpolation size value and the interpolation size adjustment value of the first target image.

[0050] Further, feature compression and reconstruction are performed, and the specific analysis steps are as follows:

[0051] A1, global average pooling and global maximum pooling are independently applied along the height and width dimensions, respectively, to obtain the average and maximum values in the channel dimension.

[0052] A2, dynamically control the number of compressed channels based on trainable parameters, and restore the compressed features to the original channel dimension through a learnable upsampling operation.

[0053] A3, fuse global average pooling and global maximum pooling to generate enhanced features.

[0054] A4, pass the enhanced features to the reconstruction module to generate enhanced feature maps and feature reconstruction.

[0055] Further, the segmentation image quality representation parameter of the second target image is analyzed as follows:

[0056] The segmentation image quality data of the second target image includes boundary curvature, recall rate, and texture contrast.

[0057] Based on the segmentation image quality data of the second target image, the segmentation image quality representation parameter of the second target image is analyzed.

[0058] The segmentation image quality characteristic parameter of the second target image is a combined influence degree quantification index of the boundary curvature, the recall rate and the texture contrast, and the specific analysis process is as follows: the boundary curvature, the recall rate and the texture contrast are respectively processed by ratio with the corresponding reference value, and each ratio processing result is coupled by combining the corresponding feature allocation coefficient, so as to obtain the segmentation image quality characteristic parameter of the second target image.

[0059] Further, the adaptive adjustment of the learning rate is performed, and the specific analysis method is as follows:

[0060] According to the image feature deviation factor of the first target image, a segmentation image quality correction coefficient is extracted.

[0061] According to the segmentation image quality characteristic parameter of the second target image and the segmentation image quality correction coefficient, a segmentation image quality correction parameter of the second target image is analyzed.

[0062] A preset segmentation image quality correction threshold in the database is extracted.

[0063] The segmentation image quality correction parameter of the second target image is subtracted by the segmentation image quality correction threshold to obtain a segmentation image quality correction deviation parameter of the second target image.

[0064] According to the segmentation image quality correction deviation parameter of the second target image, a learning rate adjustment value is extracted;

[0065] An initial learning rate preset in the database is obtained, and the adaptive adjustment of the learning rate is performed based on the initial learning rate and the learning rate adjustment value.

[0066] Further, the feature representation dimension adjustment module is used for the first encoder-decoder to receive a retinal image requiring blood vessel segmentation, denoted as a first target image, to collect image feature parameters of the first target image, to analyze image feature indicators of the first target image, and to adjust the feature representation dimension of the first target image accordingly.

[0067] The adjustment effect analysis module is used to collect artifact noise data of the first target image after the feature representation dimension adjustment, to analyze artifact noise detection abnormal scores of the first target image, to determine interpolation size adjustment requirement information, and to analyze the feature representation dimension adjustment effect based on the interpolation size adjustment requirement information determination result.

[0068] The feature compression and reconstruction is used for feature reconstruction, and the second encoder-decoder receives the output image of the first encoder-decoder through symmetric skip connection, outputs a blood vessel segmentation map, and is denoted as a second target image.

[0069] The learning rate adaptive adjustment module is configured to collect segmentation image quality data of the second target image, analyze segmentation image quality characteristic parameters of the second target image, and perform adaptive adjustment of the learning rate based on the segmentation image quality characteristic parameters of the second target image.

[0070] The feature representation dimension adjustment module is configured to receive a retinal image requiring blood vessel segmentation, denoted as a first target image, by the first encoder-decoder, collect image feature parameters of the first target image, analyze image feature indicators of the first target image, and thereby adjust the feature representation dimension of the first target image.

[0071] The adjustment effect analysis module is configured to collect artifact noise data of the first target image after the feature representation dimension adjustment, analyze artifact noise detection abnormality scores of the first target image, determine interpolation size adjustment requirement information, and perform feature representation dimension adjustment effect analysis based on the interpolation size adjustment requirement information determination result.

[0072] The feature compression and reconstruction is configured to perform feature reconstruction, make the second encoder-decoder receive the output image of the first encoder-decoder by symmetric skip connection, output a blood vessel segmentation map, and denote the output image as a second target image.

[0073] The learning rate adaptive adjustment module is configured to collect segmentation image quality data of the second target image, analyze segmentation image quality characteristic parameters of the second target image, and perform adaptive adjustment of the learning rate based on the segmentation image quality characteristic parameters of the second target image.

[0074] The one or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages:

[0075] 1. The feature representation dimension adjustment improves the sensitivity to key features, the adjusted dimension can retain and strengthen details, the segmentation result is more detailed and accurate, and the problem of insufficient detection of weak blood vessel signals in the prior art is solved.

[0076] 2. The feature representation dimension adjustment improves the sensitivity to key features, the adjusted dimension can retain and strengthen details, the segmentation result is more detailed and accurate, and the problem of insufficient detection of weak blood vessel signals in the prior art is solved. The double-encoder-decoder architecture method provided by the present application significantly enhances the adaptability of the model to complex scenes by introducing a multi-path feature extraction and fusion mechanism, effectively solving the problem of insufficient detection of weak blood vessel signals in the prior art.

[0077] 3. By feature compression and reconstruction, the feature channel is dynamically adjusted, thereby significantly improving the detection ability of weak blood vessels, effectively solving the problem of insufficient detection of weak blood vessels in the prior art. By symmetric jump connection, the multi-scale features are fused by element addition, thereby preserving the details of complex blood vessel structure, solving the problem of discontinuous segmentation results in the prior art, and by the X-MARs network structure, the multi-path design enhances feature extraction and detail recovery, thereby improving the segmentation accuracy and continuity.

[0078] 4. By image artifact noise detection abnormal score, the interpolation size of the image is adjusted, thereby accurately identifying the real target area, reducing the missegmentation or wrong segmentation phenomenon caused by artifact noise, thereby improving the accuracy of the segmentation result. BRIEF DESCRIPTION OF DRAWINGS

[0079] Figure 1 A blood vessel segmentation method based on a retinal image is provided for the embodiments of the present application.

[0080] Figure 2 A structural schematic diagram of a blood vessel segmentation system based on a retinal image is provided for the embodiments of the present application.

[0081] Figure 3 A feature compression and reconstruction flowchart is provided for the embodiments of the present application.

[0082] Figure 4 An X-MARs network structure diagram is provided for the embodiments of the present application.

[0083] Figure 5 A mind map of the embodiments of the present application. DETAILED DESCRIPTION

[0084] The technical solutions in the embodiments of the present application will be described clearly and completely below 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, not all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application.

[0085] Referring to Figure 5 , a mind map of the embodiments of the present application.

[0086] First, the input image is processed according to the first decoder, then the feature representation dimension is adjusted, the artifact noise detection abnormal score is analyzed, then the feature compression and reconstruction are performed, and finally the learning rate is adaptively adjusted. The second encoder-decoder outputs a blood vessel segmentation map.

[0087] Referring to Figure 4As shown, the X-MARs network structure involved in the embodiment of the application.

[0088] The first encoder stride convolution down-sampling and then the first decoder transpose convolution up-sampling and then the second encoder-decoder further refine the features to generate the target image, wherein the down-sampling and up-sampling paths are optimized by the jump connection.

[0089] Referring to Figure 1 As shown, the first aspect of the embodiment of the application provides a blood vessel segmentation method based on a retinal image, including the following steps:

[0090] The first encoder-decoder receives the retinal image requiring blood vessel segmentation, denoted as the first target image, collects the image feature parameters of the first target image, analyzes the image feature indicators of the first target image, and thus adjusts the feature representation dimension of the first target image.

[0091] In this embodiment, the feature representation dimension of the first target image is adjusted, and the specific analysis method is as follows:

[0092] According to the image feature parameters of the first target image, the image feature indicators of the first target image are analyzed.

[0093] The image feature comprehensive threshold stored in the database is extracted.

[0094] The image feature indicators of the first target image are subtracted by the image feature comprehensive threshold to obtain the image features of the first target image.

[0095] The representation dimension adjustment set corresponding to each image feature deviation factor interval of the first target image stored in the database is extracted, and the representation dimension adjustment set corresponding to the interval where the image feature deviation factor of the first target image is located is mapped, denoted as the feature representation dimension adjustment set.

[0096] If the deviation factor is less than zero, it means that the image quality state is poor at this time, and the spatial dimension needs to be increased to increase the calculation amount and enhance the image quality.

[0097] If the deviation factor is greater than or equal to zero, it means that the image quality state is good at this time, and the spatial dimension can be appropriately reduced to increase the channel depth to increase the calculation efficiency.

[0098] The feature representation dimension adjustment set includes the spatial dimension adjustment value and the channel depth adjustment value.

[0099] It should be noted that the image feature deviation factor represents the deviation of the image feature indicators and the image feature comprehensive threshold, and the image feature deviation factor dynamically selects or adjusts the spatial dimension and the channel depth of the feature space by quantifying the distribution difference of the features in different dimensions, thereby optimizing the capture ability of the model for key information.

[0100] extract the current spatial dimension and the current channel depth.

[0101] adjust the feature representation dimension of the first target image according to the current spatial dimension, the current channel depth and the feature representation dimension adjustment set, specifically, the numerical result of adding the current spatial dimension parameter, the current channel depth and the feature representation dimension adjustment set as the feature representation dimension of the first target image.

[0102] The image feature parameters include intersection over union, dice coefficient and accuracy.

[0103] It should be noted that the intersection over union and the accuracy can be obtained by the YOLO algorithm, and the dice coefficient can be obtained by the VNet (Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation, full convolutional neural network for volumetric medical image segmentation) model.

[0104] The intersection over union refers to the degree of overlap between the predicted region and the real region.

[0105] The dice coefficient refers to the similarity between the predicted region and the real region, and is more sensitive to small regions.

[0106] The accuracy refers to the proportion of correctly classified pixels among all pixels.

[0107] It should be noted that the intersection over union and the dice coefficient have a strong mathematical correlation, and the larger the intersection over union, the larger the dice coefficient. The increase of the intersection over union and the dice coefficient means that the number of correctly predicted pixels is higher, and the number of incorrectly predicted pixels and the number of incorrectly predicted background pixels is lower. This means that the model correctly segments most of the target region and has less false positives and false negatives. Therefore, the number of correctly predicted background pixels must be high, and thus the accuracy is high.

[0108] Extract the reference intersection over union, the reference dice coefficient and the reference accuracy stored in the database.

[0109] Extract the intersection over union feature allocation coefficient, the dice coefficient feature allocation coefficient and the accuracy feature allocation coefficient preset in the database.

[0110] It should be noted that the value range of the intersection-over-union feature distribution coefficient, the dice coefficient feature distribution coefficient and the accuracy rate feature distribution coefficient is 0 to 1, and the sum of the intersection-over-union feature distribution coefficient, the dice coefficient feature distribution coefficient and the accuracy rate feature distribution coefficient is 1. When used, the pre-set values can be directly extracted from the database. For example, a one-to-one mapping set is constructed by the intersection-over-union, the dice coefficient and the accuracy rate and the corresponding intersection-over-union feature distribution coefficient, the dice coefficient feature distribution coefficient and the accuracy rate feature distribution coefficient. When used, the intersection-over-union, the dice coefficient and the accuracy rate obtained in real time are respectively input into the corresponding mapping set, and the intersection-over-union feature distribution coefficient, the dice coefficient feature distribution coefficient and the accuracy rate feature distribution coefficient are extracted.

[0111] The image feature index of the first target image is analyzed based on the image feature parameters.

[0112] The image feature index of the first target image is a quantitative index of the influence degree of the intersection-over-union, the dice coefficient and the accuracy rate on the image feature state. The specific analysis process is as follows: the intersection-over-union, the dice coefficient and the accuracy rate are respectively subjected to ratio processing with the corresponding reference values, and the ratio processing results are respectively combined with the corresponding feature distribution coefficients and subjected to coupling processing, so as to obtain the image feature index of the first target image.

[0113] In specific embodiments, the image feature index of the first target image is specifically represented as:

[0114]

[0115] Wherein, F g represents the image feature index of the first target image, I represents the intersection-over-union, I0 represents the reference intersection-over-union, a i represents the intersection-over-union feature distribution coefficient, b d represents the dice coefficient feature distribution coefficient, g μ represents the accuracy rate feature distribution coefficient, D represents the dice coefficient, D0 represents the reference dice coefficient, m represents the accuracy rate, and m0 represents the reference accuracy rate.

[0116] After adjusting the feature representation dimension, the artifact noise data of the first target image is collected, the artifact noise detection abnormal score of the first target image is analyzed, the interpolation size adjustment requirement information is determined, and the feature representation dimension adjustment effect is analyzed based on the interpolation size adjustment requirement information determination result.

[0117] After adjusting the feature representation dimension, the artifact noise parameter of the first target image is obtained.

[0118] The artifact noise parameter includes signal-to-noise ratio, structural similarity, average gradient amplitude and gradient direction change rate.

[0119] It should be noted that the signal-to-noise ratio is a measure of the ratio of effective information to noise in the signal, reflecting the image definition, the structural similarity is a measure of the similarity of two images in brightness, contrast and structure, the average gradient amplitude is a measure of the intensity of the image edge and texture, the average gradient amplitude is a measure of the local variation intensity index of the image or signal, and the gradient direction change rate is a measure of the spatial variation frequency of the gradient direction, and the larger the value, the more inconsistent the direction.

[0120] It should be noted that the signal-to-noise ratio, the structural similarity, the average gradient amplitude and the gradient direction change rate are interrelated in the image artifact noise detection abnormal score analysis process. The increase of the signal-to-noise ratio means the reduction of the noise, and the image is closer to the true signal, so as to retain more structural information, resulting in the increase of the structural similarity. The increase of the structural similarity indicates that the edge is closer to the true structure, which will lead to the increase of the edge gradient intensity, i.e. the average gradient amplitude, and the image artifact noise detection abnormal score will be reduced. However, when the average gradient amplitude is greater than the threshold value, the image will be sharpened, resulting in the increase of the image artifact noise detection abnormal score. The increase of the edge gradient intensity means that the edge gradient direction is more consistent, resulting in the decrease of the gradient direction change rate and the image artifact noise detection abnormal score.

[0121] It should be noted that the signal-to-noise ratio, the structural similarity, the average gradient amplitude and the gradient direction change rate are interrelated in the image artifact noise detection abnormal score analysis process. The increase of the signal-to-noise ratio means the reduction of the noise, and the image is closer to the true signal, so as to retain more structural information, resulting in the increase of the structural similarity. The increase of the structural similarity indicates that the edge is closer to the true structure, which will lead to the increase of the edge gradient intensity, i.e. the average gradient amplitude, and the image artifact noise detection abnormal score will be reduced. However, when the average gradient amplitude is greater than the threshold value, the image will be sharpened, resulting in the increase of the image artifact noise detection abnormal score. The increase of the edge gradient intensity means that the edge gradient direction is more consistent, resulting in the decrease of the gradient direction change rate and the image artifact noise detection abnormal score.

[0122] The average gradient amplitude threshold value stored in the database is extracted.

[0123] The image artifact noise detection abnormal score is analyzed based on the image artifact noise parameters.

[0124]

[0125] Wherein, F Prepresents image artifact noise detection abnormal score, X represents signal-to-noise ratio, X0 represents signal-to-noise ratio reference value, G represents structural similarity, G0 represents structural similarity reference value, T represents average gradient amplitude, T0 represents average gradient amplitude reference value, Q represents structural similarity, Q0 represents structural similarity reference value, p represents signal-to-noise ratio feature assignment coefficient, s represents structural similarity feature assignment coefficient, represents average gradient amplitude feature assignment coefficient, and w represents gradient direction change rate feature assignment coefficient.

[0126] The abnormal score threshold value preset in the database is extracted.

[0127] If the image artifact noise detection abnormal score is less than the abnormal score threshold value, the interpolation size adjustment requirement information determination result of the first target image is recorded as no interpolation size adjustment being required.

[0128] It needs to be explained that if the image artifact noise detection abnormal score is less than the abnormal score threshold value, it means that the feature representation dimension adjustment effect is good, the image quality state is good, and the artifact noise does not have a great interference on image segmentation. At this time, the image does not need to be interpolated for size adjustment, so the interpolation size adjustment requirement information determination result of the first target image is recorded as no interpolation size adjustment being required, which can not only ensure the normal operation of image segmentation, but also avoid unnecessary operations to reduce system complexity and energy consumption.

[0129] If the image artifact noise detection abnormal score is greater than or equal to the abnormal score threshold value, the interpolation size adjustment requirement information determination result of the first target image is recorded as requiring interpolation size adjustment.

[0130] It needs to be explained that if the image artifact noise detection abnormal score is greater than or equal to the abnormal score threshold value, it means that the feature representation dimension adjustment effect is poor, the image quality state is poor, and the artifact noise has a great interference on image segmentation. At this time, the image needs to be interpolated for size adjustment, so the interpolation size adjustment requirement information determination result of the first target image is recorded as requiring interpolation size adjustment, which can ensure the normal operation of image segmentation and avoid interference on subsequent operations of image segmentation.

[0131] Based on the interpolation size adjustment requirement information determination result, the feature representation dimension adjustment effect is analyzed, and the specific analysis method is as follows:

[0132] If the interpolation size adjustment requirement information determination result is no interpolation size adjustment being required, a first image feature correction factor is extracted according to the image artifact noise detection abnormal score, an image feature correction value is obtained according to the image feature index of the first target image and the first image feature correction factor, and the specific analysis process is as follows: the image feature index of the first target image is multiplied by the first image feature correction factor, and the numerical result is recorded as the image feature correction value.

[0133] It should be noted that the first image feature correction factor is extracted according to the image artifact noise detection abnormal score. The specific extraction method is: extracting the feature correction factor corresponding to each image artifact noise detection abnormal score interval stored in the database, and mapping the feature correction factor corresponding to the interval in which the image artifact noise detection abnormal score is located, which is recorded as the first image feature correction factor.

[0134] It should be noted that the larger the image artifact noise detection abnormal score, the more the running state deviates from the ideal state, and the higher the interpolation size adjustment degree. In this case, the image feature is corrected, the image artifact noise detection abnormal score is reduced, the error segmentation area caused by noise and artifact is reduced, the workload of subsequent manual correction or additional processing of the segmentation result is reduced, the efficiency of the entire image segmentation process is improved, time and labor costs are saved, and the segmentation model can better adapt to various complex image conditions.

[0135] If the interpolation size adjustment requirement information determination result is to require interpolation size adjustment, interpolation size adjustment is performed.

[0136] The image feature index is reacquired.

[0137] The image artifact noise detection abnormal score is reacquired, and the image feature correction value is analyzed based on the reacquired image artifact noise detection abnormal score and the reacquired image feature index. The specific analysis process is as follows: multiply the reacquired image artifact noise detection abnormal score by the reacquired image feature index, and record the numerical result as the image feature correction value.

[0138] The image feature comprehensive threshold in the database is extracted.

[0139] The image feature correction value is subjected to difference processing with the image feature comprehensive threshold to obtain an image feature correction deviation factor.

[0140] The image feature correction deviation factor ideal interval preset in the database is extracted.

[0141] If the image feature correction deviation factor is within the image feature correction deviation factor ideal interval, the feature representation dimension adjustment effect analysis result is recorded as effective adjustment, and prompt information is generated synchronously.

[0142] It should be noted that the image feature correction deviation factor being within the image feature correction deviation factor ideal interval means that the interpolation adjustment effect is good, the image quality state is good, and the artifact noise interference on the image has been corrected, so the feature representation dimension adjustment effect analysis result is recorded as effective adjustment, and prompt information is generated.

[0143] It needs to be added that in specific embodiments, the prompt information can be: "feature representation dimension adjustment has been successfully completed, the current image quality state is good, and the subsequent operation can continue".

[0144] If the image feature correction deviation factor exceeds the ideal interval of the image feature correction deviation factor, the feature representation dimension adjustment effect analysis result is recorded as invalid adjustment, and a warning information is generated synchronously.

[0145] It needs to be pointed out that the image feature correction deviation factor is in the image feature correction deviation factor exceeding the ideal interval, which indicates that the interpolation adjustment effect is poor, the image quality state is poor, and the artifact noise interference has not been corrected. Therefore, the feature representation dimension adjustment effect analysis result is recorded as invalid adjustment, and a warning information is generated.

[0146] It needs to be added that in specific embodiments, the warning information can be: "warning! Feature representation dimension adjustment error".

[0147] Subtract the abnormal score threshold from the image artifact noise detection abnormal score to obtain the image artifact noise detection abnormal deviation score.

[0148] Extract the interpolation size adjustment value corresponding to each image artifact noise detection abnormal deviation score interval stored in the database, and map the interpolation size adjustment value corresponding to the interval in which the image artifact noise detection abnormal deviation score is located, which is recorded as the interpolation size adjustment value of the first target image.

[0149] It needs to be added that the larger the image interpolation size adjustment value, the more the running state deviates from the ideal state, and the higher the interpolation size adjustment degree. In this case, image feature correction is performed to reduce the interpolation size adjustment value and reduce the missegmentation or missegmentation caused by artifact noise, thereby improving the accuracy of the segmentation result and improving the efficiency of the entire image segmentation process.

[0150] Extract the current image interpolation size value.

[0151] According to the current image interpolation size value and the interpolation size adjustment value of the first target image, the interpolation size adjustment of the first target image is performed. The current image interpolation size value plus the interpolation size adjustment value of the first target image is taken as the interpolation size of the first target image.

[0152] Referring to Figure 3 The feature compression and reconstruction process involved in the embodiments of the present application performs global average pooling and global maximum pooling operations, restores the compressed features to the original channel dimension through upsampling operation, and fuses the output to generate enhanced feature maps and feature reconstruction.

[0153] Feature compression and reconstruction are performed, the second encoder-decoder receives the first encoder-decoder output image through the symmetric skip connection, outputs a blood vessel segmentation image, and is recorded as a second target image.

[0154] A1, global average pooling and global maximum pooling are independently applied along the height and width dimensions, thereby obtaining the average value and the maximum value in the channel dimension.

[0155] It should be noted that global average pooling refers to extracting average information in the channel dimension along the height and width dimensions, and global maximum pooling refers to extracting significant information in the channel dimension along the height and width dimensions.

[0156] In specific embodiments, the average value in the channel dimension is specifically represented as:

[0157]

[0158] wherein GAP(X) is the average value in the channel dimension, c is the number of channels, H is the channel height, W is the channel width, and X (c,i,j) is the value of the input feature map at the i-th row and the j-th column of the c-th channel, i is the index of the feature map in the height, j is the index of the feature map in the width, i = 1, 2,..., H, and j = 1, 2,..., W.

[0159] In specific embodiments, the maximum value in the channel dimension is specifically represented as:

[0160]

[0161] wherein GMP(X) is the maximum value in the channel dimension, c is the number of channels, and X (c,i,j) is the value of the input feature map at the i-th row and the j-th column of the c-th channel, i is the index of the feature map in the height, i = 1, 2,..., H, H is the channel height, j is the index of the feature map in the width, j = 1, 2,..., W, and W is the channel width.

[0162] A2, dynamically control the number of compressed channels based on trainable parameters, and restore the compressed features to the original channel dimension through a learnable upsampling operation.

[0163] A3, fuse global average pooling and global maximum pooling to generate enhanced features.

[0164] Fusion of global average pooling and global maximum pooling refers to element-level fusion of the outputs of GAP and GMP, addition, and processing through an S-shaped activation function, F fused = σ(GAP(X) + GMP(X)), wherein F fusedGAP(X) represents the average value in the channel dimension, and GMP(X) represents the maximum value in the channel dimension. This fusion, combined with the S-type activation, integrates complementary average and significant information into a unified representation, mapping the output to a probability range suitable for the blood vessel segmentation task.

[0165] A4, the enhanced features are passed to the reconstruction module to generate enhanced feature maps and feature reconstruction.

[0166] GAP and GMP are again applied to the channel dimension of the input X to produce two feature vectors, which are then concatenated, specifically represented as: F concat = Concat(GAP(X), GMP(X)),

[0167] where F concat is the feature concatenation value, GAP(X) is the average value in the channel dimension, and GMP(X) is the maximum value in the channel dimension.

[0168] The concatenated features are passed to the reconstruction module and then through the S-type activation function to generate enhanced feature maps and feature reconstruction while maximizing the preservation of key spatial and channel information.

[0169] Collect segmentation image quality data of the second target image, analyze segmentation image quality characteristic parameters of the second target image, and perform adaptive adjustment of the learning rate based on the segmentation image quality characteristic parameters of the second target image.

[0170] The segmentation image quality data of the second target image includes boundary curvature, recall rate, and texture contrast.

[0171] It should be noted that the boundary curvature, recall rate, and texture contrast are all obtained from the pixels collected by the sensor.

[0172] Boundary curvature is a measure of the local bending degree of the image boundary, reflecting the complexity of the shape. Recall rate is a measure of the model's ability to detect targets, calculated as the proportion of correctly detected target pixels to the total number of real target pixels. Texture contrast reflects the difference in texture intensity in local regions of the image.

[0173] It should be noted that low texture contrast makes it difficult to distinguish the target from the background, resulting in low recall rate. High boundary curvature makes it easy for the model to miss the complex and curved boundary, resulting in low recall rate. The diameter of the small blood vessel branches in the high curvature region is small, resulting in a gentle change in local grayscale (low texture contrast), which is often accompanied by low texture contrast.

[0174] Based on the segmentation image quality data of the second target image, the segmentation image quality characteristic parameters of the second target image are analyzed.

[0175] The segmentation image quality state of the second target image is quantified by a boundary curvature, a recall rate and a texture contrast, and the specific analysis process is as follows: the boundary curvature, the recall rate and the texture contrast are respectively subjected to ratio processing with corresponding reference values, the ratio processing results are respectively coupled with corresponding feature distribution coefficients for processing, and thus the segmentation image quality state of the second target image is obtained.

[0176] It should be noted that the boundary curvature feature distribution coefficient, the recall rate feature distribution coefficient and the texture contrast feature distribution coefficient are all in the range of 0 to 1, and the sum of the boundary curvature feature distribution coefficient, the recall rate feature distribution coefficient and the texture contrast feature distribution coefficient is 1. When used, the pre-set values can be directly extracted from the database. For example, a one-to-one mapping set is constructed by the boundary curvature, the recall rate and the texture contrast and the corresponding boundary curvature feature distribution coefficient, the recall rate feature distribution coefficient and the texture contrast feature distribution coefficient. When used, the real-time acquired boundary curvature, the recall rate and the texture contrast are respectively input into the corresponding mapping set, and thus the boundary curvature feature distribution coefficient, the recall rate feature distribution coefficient and the texture contrast feature distribution coefficient are extracted.

[0177] In specific embodiments, the segmentation image quality state of the second target image is specifically represented as:

[0178]

[0179] Wherein, F e represents the segmentation image quality state of the second target image, K0 represents the reference value corresponding to the boundary curvature, K represents the boundary curvature value, δ represents the boundary curvature segmentation state distribution coefficient, ε represents the recall rate segmentation state distribution coefficient, R represents the recall rate value, R0 represents the reference value corresponding to the recall rate, N represents the texture contrast value, N0 represents the reference value corresponding to the texture contrast, and ∈ represents the texture contrast boundary curvature segmentation state distribution coefficient.

[0180] The segmentation image quality correction coefficient is extracted according to the image feature deviation factor of the first target image, and the specific extraction process is as follows: the image quality correction coefficients corresponding to the image feature deviation factor intervals of each first target image stored in the database are extracted, and the image quality correction coefficient corresponding to the interval in which the image feature deviation factor of the first target image is located is mapped and extracted, which is denoted as the segmentation image quality correction coefficient.

[0181] It should be noted that the larger the image quality correction coefficient, the more the operating state deviates from the ideal state, and the worse the segmentation image quality. In this case, the segmentation image quality is corrected, so that the data distribution of different images is more consistent, which helps to improve the stability and generalization ability of the model

[0182] According to the segmentation image quality characteristic parameter of the second target image and the segmentation image quality correction coefficient, the segmentation image quality correction parameter of the second target image is analyzed, and the specific analysis process is as follows:

[0183] The segmentation image quality characteristic parameter of the second target image is multiplied by the segmentation image quality correction coefficient, and the numerical result is recorded as the segmentation image quality correction parameter of the second target image.

[0184] The preset segmentation image quality correction threshold in the database is extracted.

[0185] The segmentation image quality correction parameter of the second target image is subtracted from the segmentation image quality correction threshold to obtain the segmentation image quality correction deviation parameter of the second target image.

[0186] According to the segmentation image quality correction deviation parameter of the second target image, a learning rate adjustment value is extracted, and the specific extraction process is as follows: the learning rate adjustment value corresponding to the interval of the segmentation image quality correction deviation parameter of each second target image stored in the database is extracted, and the learning rate adjustment value corresponding to the interval in which the segmentation image quality correction deviation parameter of the second target image is located is mapped and extracted, and is recorded as the learning rate adjustment value.

[0187] It should be noted that when the learning rate adjustment value is larger, the model will jump excessively in the parameter space and cannot converge stably, and even when approaching the optimal solution, it may cross the optimal solution, causing the loss function value to increase instead of decreasing in each iteration, and even diverging to infinity, causing the training process to fail. In this case, learning rate adjustment can enable the model to better learn the features in the data and improve the accuracy and generalization ability of the model, avoiding problems such as overfitting or underfitting caused by excessively large or small learning rates.

[0188] The initial learning rate preset in the database is obtained, and the adaptive adjustment of the learning rate is performed based on the initial learning rate and the learning rate adjustment value, and the specific analysis process is as follows: the initial learning rate is added to the adjustment value as the current learning rate for adjustment.

[0189] By reasonably adjusting the learning rate, the model can use a larger learning rate to quickly converge in the early training stage, and gradually reduce the learning rate in the later training stage, so that the model can more finely adjust the parameters, thereby ensuring the performance of the model while improving the training efficiency, saving training time and computational resources.

[0190] For example, Figure 2As shown, it is a structural schematic diagram of the blood vessel segmentation system based on the retinal image provided by the embodiment of the present application, the blood vessel segmentation system based on the retinal image provided by the embodiment of the present application comprises a feature representation dimension adjustment module, which is used for a first encoder-decoder to receive a retinal image requiring blood vessel segmentation, which is recorded as a first target image, collect image feature parameters of the first target image, analyze image feature indexes of the first target image, and thus adjust the feature representation dimension of the first target image.

[0191] An adjustment effect analysis module is used to collect artifact noise data of the first target image after the feature representation dimension adjustment, analyze artifact noise detection abnormal score of the first target image, determine interpolation size adjustment requirement information, and perform feature representation dimension adjustment effect analysis based on the interpolation size adjustment requirement information determination result.

[0192] Feature compression and reconstruction are used for feature reconstruction, and a second encoder-decoder receives an output image of the first encoder-decoder through symmetric skip connection, outputs a blood vessel segmentation image, and records the blood vessel segmentation image as a second target image.

[0193] A learning rate adaptive adjustment module is used to collect segmentation image quality data of the second target image, analyze segmentation image quality characteristic parameters of the second target image, and perform adaptive adjustment of the learning rate based on the segmentation image quality characteristic parameters of the second target image.

[0194] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can adopt a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can adopt a computer program product implemented on one or more computer usable storage media (including but not limited to disk memory, CD-ROM, optical memory, etc.) containing computer usable program code.

[0195] The present application is described with reference to flowcharts and / or block diagrams of the method, device (system), and computer program product according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be realized by computer program instructions. These computer program instructions can be provided to a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to produce a machine, so that the instructions executed by the computer or other programmable data processing devices produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The device that implements the functions specified in one flow or multiple flows and / or blocks Figure 1 The device that implements the functions specified in one flow or multiple flows and / or blocks

[0196] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the Figure 1 function specified in the flow or flows and / or blocks Figure 1 of the block or blocks.

[0197] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the Figure 1 function specified in the flow or flows and / or blocks Figure 1 of the block or blocks.

[0198] Although preferred embodiments of the application have been described herein, substitutions and alterations are possible in view of the teachings of this application. Accordingly, the appended claims are intended to encompass all such substitutions and alterations. It will be apparent to those skilled in the art that various modifications and variations can be made to the present application without departing from the spirit or scope of the application. Thus, it is intended that the present application cover the modifications and variations of this application provided they come within the scope of the appended claims and their equivalents.

[0199] It will be apparent to those skilled in the art that various modifications and variations can be made to the present application without departing from the spirit or scope of the application. Thus, it is intended that the present application cover the modifications and variations of this application provided they come within the scope of the appended claims and their equivalents.

Claims

1. A method for blood vessel segmentation based on retinal images, characterized in that, The method comprises the following steps: The first encoder-decoder receives a retinal image requiring blood vessel segmentation, denoted as a first target image, collects image feature parameters of the first target image, analyzes image feature indicators of the first target image, and thereby adjusts the feature representation dimension of the first target image; After the feature representation dimension adjustment, the artifact noise data of the first target image is collected, the artifact noise detection abnormal score of the first target image is analyzed, the interpolation size adjustment requirement information is determined, and the feature representation dimension adjustment effect is analyzed based on the determination result of the interpolation size adjustment requirement information; Feature compression and reconstruction are performed, the second encoder-decoder receives the output image of the first encoder-decoder through symmetric skip connection, outputs a blood vessel segmentation image, and is denoted as a second target image; The segmentation image quality data of the second target image is collected, the segmentation image quality characteristic parameters of the second target image are analyzed, and the learning rate is adaptively adjusted based on the segmentation image quality characteristic parameters of the second target image; The feature representation dimension adjustment of the first target image is performed, and the specific analysis method is as follows: According to the image feature parameters of the first target image, the image feature indicators of the first target image are analyzed; An image feature comprehensive threshold stored in a database is extracted; The image feature indicators of the first target image are subtracted from the image feature comprehensive threshold to obtain image feature deviation factors of the first target image; The representation dimension adjustment set corresponding to each image feature deviation factor interval of the first target image stored in the database is extracted, and the representation dimension adjustment set corresponding to the interval in which the image feature deviation factor of the first target image is located is mapped and extracted, denoted as a feature representation dimension adjustment set; The feature representation dimension adjustment set comprises a spatial dimension adjustment value and a channel depth adjustment value; The current spatial dimension and the current channel depth are extracted; The feature representation dimension of the first target image is adjusted according to the current spatial dimension, the current channel depth and the feature representation dimension adjustment set.

2. The blood vessel segmentation method based on retinal images as claimed in claim 1, wherein: The image feature indicators of the first target image are analyzed, and the specific analysis method is as follows: The image feature parameters comprise an intersection over union, a dice coefficient and an accuracy rate; The image feature indicators of the first target image are analyzed based on the image feature parameters; The image feature indicators of the first target image are the quantification indicators of the influence degree of the image feature state of the intersection over union, the dice coefficient and the accuracy rate, and the specific analysis process is that the intersection over union, the dice coefficient and the accuracy rate are respectively subjected to ratio processing with corresponding reference values, each ratio processing result is respectively combined with a corresponding feature allocation coefficient and subjected to coupling processing, and thereby the image feature indicators of the first target image are obtained.

3. The blood vessel segmentation method based on retinal images as described in claim 1, characterized in that: The interpolation size adjustment requirement information is determined, and the specific analysis method is as follows: After the feature representation dimension adjustment, the artifact noise parameters of the first target image are obtained; The artifact noise parameters comprise a signal-to-noise ratio, a structural similarity, an average gradient amplitude and a gradient direction change rate; The image artifact noise detection abnormal score is analyzed based on the image artifact noise parameters; The image artifact noise detection abnormality score is a quantitative index of the influence degree of the signal-to-noise ratio, the structural similarity, the average gradient amplitude and the gradient direction change rate on the image artifact noise abnormality, and the specific analysis process is as follows: the reference values corresponding to the signal-to-noise ratio and the structural similarity are respectively subjected to ratio processing with the signal-to-noise ratio and the structural similarity, the gradient direction change rate is subjected to ratio processing with the corresponding reference value, when the average gradient amplitude is greater than a defined value, the average gradient amplitude is subjected to ratio processing with the reference value corresponding to the average gradient amplitude, when the average gradient amplitude is less than or equal to the defined value, the reference value corresponding to the average gradient amplitude is subjected to ratio processing with the average gradient amplitude, and each ratio processing result is coupled by combining with the corresponding feature assignment coefficient, so as to obtain the image artifact noise detection abnormality score of the first target image; extracting a preset abnormality score threshold value in a database; if the image artifact noise detection abnormality score is less than the abnormality score threshold value, the interpolation size adjustment requirement information judgment result of the first target image is recorded as no interpolation size adjustment being required; if the image artifact noise detection abnormality score is greater than or equal to the abnormality score threshold value, the interpolation size adjustment requirement information judgment result of the first target image is recorded as interpolation size adjustment being required.

4. The blood vessel segmentation method based on retinal images as claimed in claim 1, wherein: The specific analysis method for the feature representation dimension adjustment effect analysis based on the interpolation size adjustment requirement information judgment result is as follows: if the interpolation size adjustment requirement information judgment result is no interpolation size adjustment being required, a first image feature correction factor is extracted according to the image artifact noise detection abnormality score, and an image feature correction value is obtained according to the image feature index of the first target image and the first image feature correction factor; if the interpolation size adjustment requirement information judgment result is interpolation size adjustment being required, interpolation size adjustment is performed, and the image feature index is reacquired; the image artifact noise detection abnormality score is reacquired; the image feature correction value is analyzed based on the reacquired image artifact noise detection abnormality score and the reacquired image feature index; an image feature comprehensive threshold value in the database is extracted; a difference value processing is performed on the image feature correction value and the image feature comprehensive threshold value, so as to obtain an image feature correction deviation factor; a preset image feature correction deviation factor ideal interval in the database is extracted; if the image feature correction deviation factor is within the image feature correction deviation factor ideal interval, the feature representation dimension adjustment effect analysis result is recorded as effective adjustment, and prompt information is synchronously generated; if the image feature correction deviation factor is out of the image feature correction deviation factor ideal interval, the feature representation dimension adjustment effect analysis result is recorded as ineffective adjustment, and warning information is synchronously generated.

5. The method of claim 4, wherein the step of segmenting the blood vessels comprises: The specific analysis method for the interpolation size adjustment is as follows: ​ an image artifact noise detection abnormality deviation score is obtained by subtracting the abnormality score threshold value from the image artifact noise detection abnormality score; an interpolation size adjustment value corresponding to the interval in which the image artifact noise detection abnormality deviation score is located is mapped and extracted, and is recorded as the interpolation size adjustment value of the first target image; the current image interpolation size value is extracted; Adjusting the interpolation size of the first target image according to the current image interpolation size value and the interpolation size adjustment value of the first target image.

6. The blood vessel segmentation method based on retinal images as claimed in claim 1, wherein: The feature compression and reconstruction is performed, and the specific analysis steps are as follows: A1, global average pooling and global maximum pooling are independently applied along the height and width dimensions, thereby obtaining the average value and maximum value in the channel dimension; A2, the number of compressed channels is dynamically controlled based on trainable parameters, and the compressed features are restored to the original channel dimension through a learnable upsampling operation; A3, the global average pooling and the global maximum pooling are fused to generate enhanced features; A4, the enhanced features are transmitted to the reconstruction module to generate enhanced feature maps and feature reconstruction.

7. The blood vessel segmentation method based on retinal images as claimed in claim 1, wherein: The segmentation image quality data of the second target image includes boundary curvature, recall rate, and texture contrast; The segmentation image quality data of the second target image is analyzed to obtain the segmentation image quality characteristic parameter of the second target image; The segmentation image quality characteristic parameter of the second target image is the boundary curvature, the recall rate, and the texture contrast, which are collectively quantified as the influence degree of the segmentation image quality state, and the specific analysis process is as follows: the boundary curvature, the recall rate, and the texture contrast are respectively processed by ratio with the corresponding reference value, and each ratio processing result is coupled by combining the corresponding feature allocation coefficient, thereby obtaining the segmentation image quality characteristic parameter of the second target image. The adaptive adjustment of the learning rate is performed, and the specific analysis method is as follows:

8. The blood vessel segmentation method based on retinal images as claimed in claim 1, wherein: A segmentation image quality correction coefficient is extracted according to the image feature deviation factor of the first target image; A segmentation image quality correction parameter of the second target image is analyzed according to the segmentation image quality characteristic parameter of the second target image and the segmentation image quality correction coefficient; A preset segmentation image quality correction threshold in the database is extracted; The segmentation image quality correction parameter of the second target image is subtracted by the segmentation image quality correction threshold to obtain a segmentation image quality correction deviation parameter of the second target image; A learning rate adjustment value is extracted according to the segmentation image quality correction deviation parameter of the second target image; An initial learning rate preset in the database is obtained, and the adaptive adjustment of the learning rate is performed based on the initial learning rate and the learning rate adjustment value. It includes:

9. System for applying a method for blood vessel segmentation based on retinal images according to any of claims 1 to 8, characterized in that, The feature representation dimension adjustment module is used for the first encoder-decoder to receive the retinal image requiring blood vessel segmentation, denoted as the first target image, to collect the image feature parameters of the first target image, to analyze the image feature indicators of the first target image, and to adjust the feature representation dimension of the first target image; The adjustment effect analysis module is used to collect the artifact noise data of the first target image after the feature representation dimension adjustment, to analyze the artifact noise detection abnormal score of the first target image, to determine the interpolation size adjustment requirement information, and to analyze the feature representation dimension adjustment effect based on the interpolation size adjustment requirement information determination result; The feature compression and reconstruction is performed, and the specific analysis steps are as follows: The feature compression and reconstruction is performed, and the specific analysis steps are as follows: The learning rate self-adaptive adjustment module is configured to collect segmentation image quality data of the second target image, analyze segmentation image quality characteristic parameters of the second target image, and perform self-adaptive adjustment of the learning rate based on the segmentation image quality characteristic parameters of the second target image.

Images