Retinal vessel image segmentation method based on parent loss integration

Abstract

The present application relates to a retinal blood vessel image segmentation method based on parent loss integration, comprising: inputting a retinal blood vessel image to be segmented into a PL-TransUNet model to obtain an image segmentation visualization result, wherein the PL-TransUNet model is obtained based on a training set, and the training set comprises retinal blood vessel images and corresponding true values; during training of the PL-TransUNet model based on the training set, comprising: predicting the retinal blood vessel image through an initial PL-TransUNet model, calculating a loss channel according to the prediction result of the retinal blood vessel image and the true value, performing reverse filling and feature enhancement according to the loss channel to obtain a final enhanced feature map, and using the final enhanced feature map to obtain a new prediction result for iteration until convergence, thereby obtaining the PL-TransUNet model. The present application can improve the accuracy of retinal blood vessel image segmentation.

Description

Technical Field

[0001] The present invention relates to the technical field of image segmentation, and in particular to a retinal blood vessel image segmentation method based on parent loss integration. Background Art

[0002] Currently, machine vision-based methods are commonly used to automatically segment retinal blood vessels. U-Net, a deep learning image segmentation method, is a representative example. It has advantages such as simple structure, high efficiency, and multi-scale feature fusion, making it widely used. Although the U-Net encoder-decoder structure has demonstrated excellent performance in some medical image segmentation tasks, it also has some drawbacks. For example, the encoder primarily uses convolutional neural networks, but the local nature of convolution limits its ability to represent global features. This can be improved by integrating methods such as the Transformer into the encoder. For example, Zhang et al. proposed an enhancement network using the SwinTransformer as an encoder. This network enhances features by fusing features from different perceptual levels and improving the utilization of spatial and channel information in the image. Fan et al. used two parallel branches, a CNN and a Transformer, in the encoder to extract local and global information, and employed channel and spatial attention for feature fusion. However, these methods need further improvement in their ability to integrate prior knowledge to guide segmentation. Furthermore, Zhang et al. integrated spatial prior information about the pancreas into a deep CNN to improve organ segmentation accuracy. Wang et al. proposed a progressive dual-prior network that first locates and crops the tumor region and then gradually refines the tumor using weak semantic priors and cross-scale correlation priors. Zhang et al. proposed a Gaussian mixture prior method that embeds prior information into a self-attention mechanism to guide segmentation. These methods all introduce priors from the perspective of estimating target location but ignore loss information.

[0003] In order to improve the prior knowledge integration and global feature enhancement capabilities of the codec architecture neural network, the present invention proposes a retinal vascular image segmentation method based on parent loss integration. Summary of the Invention

[0004] The purpose of the present invention is to provide a retinal vascular image segmentation method based on parent loss integration, which strengthens the automatic integration of prior knowledge and global feature enhancement capabilities of the encoding-decoding neural network to improve the segmentation accuracy of retinal vascular images.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] Retinal vessel image segmentation method based on parent loss integration, including:

[0007] Acquire a retinal blood vessel image to be segmented;

[0008] Inputting the retinal vascular image to be segmented into a PL-TransUNet model to obtain an image segmentation visualization result, wherein the PL-TransUNet model is trained based on a training set, and the training set includes retinal vascular images and corresponding true values;

[0009] The process of training the PL-TransUNet model based on the training set includes: predicting the retinal vascular image through the initial PL-TransUNet model, calculating the loss channel according to the prediction result of the retinal vascular image and the true value, performing backfilling and feature enhancement according to the loss channel, obtaining a final enhanced feature map, and using the final enhanced feature map to obtain a new prediction result for iteration until convergence to obtain the PL-TransUNet model.

[0010] Optionally, the PL-TransUNet model includes an encoder, a decoder, and a BPL module, wherein the encoder includes a CNN local encoder and an RLFE module, and the RLFE module includes several Transformer submodules;

[0011] The BPL module is used to calculate a loss channel based on the prediction result of the retinal vascular image and the true value;

[0012] The encoder is configured to perform feature enhancement on the retinal vascular image based on the loss channel to obtain the final enhanced feature map;

[0013] The decoder is configured to output the new prediction result based on the final enhanced feature map.

[0014] Optionally, calculating the loss channel according to the prediction result and the true value of the retinal vascular image includes: inputting the prediction result and the true value into the BPL module to calculate the loss matrix;

[0015] The maximum value of the loss matrix is ​​selected according to the channel and binarized to obtain the loss channel.

[0016] Optionally, the loss matrix can be calculated as:

[0017] L m =-α(1-p l ) γ log(p l )

[0018] Among them, L m is the loss matrix, α is the balance factor, p l is the prediction result, and γ is the focusing parameter.

[0019] Optionally, the method to obtain the loss channel is:

[0020]

[0021] C index =argmax c (L m )

[0022] Among them, Lc is the loss channel, max c (·) means taking the maximum value by channel, Cindex means the channel number of the maximum value, L m is the loss matrix, and (x,y) is the coordinate of the channel element.

[0023] Optionally, performing reverse filling and feature enhancement according to the loss channel to obtain a final enhanced feature map includes:

[0024] Reversely filling the loss channel into the idle channel to obtain the filled idle channel;

[0025] Taking the retinal blood vessel image and the padded idle channel as image input, inputting the image input and the loss channel into the CNN local encoder respectively, and calculating image embedding and loss embedding;

[0026] The image embedding and the loss embedding are input into the RLFE module for global feature enhancement to obtain the final enhanced feature map.

[0027] Optionally, inputting the image embedding and the loss embedding into the RLFE module for global feature enhancement includes:

[0028] Embedding the image into the N-1 layer Transformer submodule in the RLFE module for global encoding, embedding the globally encoded image and the loss into the feature enhancement submodule to obtain an enhanced feature map;

[0029] The enhanced feature map is input into the N-layer Transformer submodule for global encoding to obtain the final enhanced feature map.

[0030] Optionally, the method for obtaining the enhanced feature map is:

[0031] F enhanc =F embed +F L

[0032] F L =L embed ·(mask F mask L )

[0033]

[0034] Among them, F enhanc is the enhanced feature map, F embed is the image embedding, F L is the enhancement component based on the loss channel, mask F Embed mask for image, mask L is the parent loss embedding mask, (x, y) is the channel element coordinate, L embed is the loss embedding.

[0035] The beneficial effects of the present invention are as follows: the method of the present invention reverse fills the parent loss (BPL) module, converts the loss information of the segmentation model into a loss channel and reverse fills it into the idle channel of the next generation training input, and participates in the next generation training as prior knowledge to guide segmentation. Then, the parent loss is continued to be used for reverse loss feature enhancement (RLFE), first using Transformer to globally encode the image embedding, and then based on the attention mechanism, the loss embedding and the processed image embedding are feature enhanced and fused, which improves the segmentation accuracy from the two aspects of prior integration and feature enhancement, and has better segmentation effect in retinal vascular image segmentation, especially low-contrast microvascular segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0037] Figure 1 Comparison of the training process of PL-TransUNet according to an embodiment of the present invention and the traditional method;

[0038] Figure 2 This is a flow chart of a retinal vascular image segmentation method based on parent loss integration according to an embodiment of the present invention;

[0039] Figure 3 A flowchart of a BPL module for calculating loss channels according to an embodiment of the present invention;

[0040] Figure 4 This is a structural diagram of an RLFE module according to an embodiment of the present invention;

[0041] Figure 5 1 is an architecture diagram of PL-TransUNet according to an embodiment of the present invention;

[0042] Figure 6This is a comparison chart of the visualization results of segmentation results of different methods on different data sets in an embodiment of the present invention. DETAILED DESCRIPTION

[0043] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0044] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0045] Inspired by the reflection effect in biology, this embodiment proposes a retinal vascular image segmentation method based on TransUNet, which is based on backfilling parent loss (BPL) and reverse loss feature enhancements (RLFE), and is named PL-TransUNet. Figure 1 Figure 4 shows a comparison between PL-TransUNet and traditional methods. Top: At the end of each round of training, the traditional model calculates the loss and then directly performs backpropagation. Bottom: PL-TransUNet calculates the loss and further converts it into a loss matrix, which is then used for prior integration and feature enhancement in the next round of model training.

[0046] This embodiment provides a retinal vessel image segmentation method based on parent loss integration, including:

[0047] Acquire a retinal blood vessel image to be segmented;

[0048] Input the retinal vascular image to be segmented into the PL-TransUNet model to obtain the image segmentation visualization result, wherein the PL-TransUNet model is trained based on the training set, which includes the retinal vascular image and the corresponding true value;

[0049] The process of training the PL-TransUNet model based on the training set includes: predicting the retinal vascular image through the initial PL-TransUNet model, calculating the loss channel according to the predicted result and the true value of the retinal vascular image, performing reverse filling and feature enhancement according to the loss channel, obtaining the final enhanced feature map, and using the final enhanced feature map to obtain a new prediction result for iteration until convergence to obtain the PL-TransUNet model.

[0050] Further, the PL-TransUNet model comprises an encoder, a decoder and a BPL module, wherein the encoder comprises a CNN local encoder and an RLFE module, and the RLFE module comprises a plurality of Transformer sub-modules. The BPL module is configured to calculate a loss channel according to a prediction result of the retinal blood vessel image and a true value; the encoder is configured to perform feature enhancement on the retinal blood vessel image based on the loss channel to obtain a final enhanced feature map; and the decoder is configured to output a new prediction result based on the final enhanced feature map.

[0051] As shown in Figure 5 , in the embodiment, the PL-TransUNet model is constructed based on the proposed BPL and RLFE modules and TransUNet. The PL-TransUNet model comprises an encoder and a decoder. The encoder comprises two inputs, i.e., an image input (4-channel input) and a loss channel. The encoder mainly comprises a CNN local encoder and an RLFE module. The CNN adopts ResNet-50, and the RLFE module comprises 12 Transformer sub-modules. The decoder has the same structure as the decoder of TransUNet and mainly comprises a 1x1 convolution, an up-sampling layer and a convolution layer. After the decoder outputs a prediction result, the BPL module is used for reverse parent loss filling.

[0052] The overall flow of the method of the embodiment is shown in Figure 2 . First, a prediction result is obtained through forward propagation of the PL-TransUNet. On one hand, a loss is calculated according to the prediction result and a true value GT and is back propagated. On the other hand, the prediction result is input into the BPL module to be converted into a loss channel, which is back filled into an idle channel. Next, the image and the filled idle channel (collectively referred to as a 4-channel image input) are input into the CNN local encoder to calculate image embedding, and the loss channel is also input into the CNN to obtain loss embedding. Then, the two kinds of embedding are input into the RLFE module for global feature enhancement. Finally, the enhanced feature map is input into the decoder to obtain a prediction result. This process is repeated until the algorithm converges.

[0053] Specifically, in this embodiment, during the first training of PL-TransUNet, the two inputs are a 4-channel input (i.e., the original image and an idle channel C0 filled with all zeros) and a loss channel filled with all zeros. These two inputs first undergo independent forward propagation, and after being encoded by ResNet-50, image embedding and loss embedding are obtained respectively; then the two embeddings are input to the RLFE module for feature enhancement and fusion, and then enter the decoder to output the prediction result. In addition to participating in backpropagation, this output also enters the BPL module to calculate the loss channel. In subsequent training, the loss channel of the previous training is used to fill C0 and serve as another input. The subsequent process is the same as the first training until the convergence condition is reached and the algorithm stops.

[0054] Furthermore, calculating the loss channel according to the prediction result and the true value of the retinal vascular image includes: inputting the prediction result and the true value into the BPL module to calculate the loss matrix;

[0055] Select the maximum value of the loss matrix according to the channel and perform binarization to obtain the loss channel.

[0056] Specifically, this embodiment proposes a reverse padding parent loss method to carry GT information during forward propagation. Inspired by biological reflex mechanisms, the traditional training process can be viewed as a simple reflex, where the model simply makes predictions and adjustments based on the input data. The BPL method, on the other hand, is similar to a complex reflex, where the model can leverage additional GT information to adjust its own parameters.

[0057] like Figure 3 As shown, the BPL module takes GT and prediction results as input and first calculates the loss matrix L m :

[0058] L m =-α(1-p l ) γ log(p l )

[0059] Among them, L m is the loss matrix, the balance factor α∈[0,1], the balance factor is α for positive samples and 1-α for negative samples, p l To predict the results, γ is the focusing parameter (γ ≥ 0), which controls the number of easy or hard samples.

[0060] Next, for L m Take the maximum value by channel and perform binarization to obtain the loss channel L carrying GT information c :

[0061]

[0062] Cindex =argmax c (L m )

[0063] Among them, Lc is the loss channel, max c (·) means taking the maximum value by channel, Cindex means the channel number of taking the maximum value, when C index = 1, the maximum loss value of the position is the value of the first channel (foreground), thereby retaining the loss information (L c Set to 1); otherwise, L c Set to 0, no loss information is retained, (x, y) is the coordinate of the channel element.

[0064] Furthermore, reverse filling and feature enhancement are performed according to the loss channel to obtain the final enhanced feature map including:

[0065] Reversely fill the loss channel into the idle channel to obtain the filled idle channel;

[0066] The retinal vascular image and the padded idle channel are used as image inputs, and the image input and loss channel are input into the CNN local encoder respectively to calculate the image embedding and loss embedding;

[0067] The image embedding and loss embedding are input into the RLFE module for global feature enhancement to obtain the final enhanced feature map.

[0068] Specifically, in this embodiment, the image and the padded idle channels (collectively referred to as the 4-channel image input) are fed into the CNN local encoder to calculate the image embedding. The loss channel is also fed into the CNN to obtain the loss embedding. These two embeddings are then fed into the RLFE module for global feature enhancement.

[0069] Furthermore, the image embedding and loss embedding are input into the RLFE module for global feature enhancement, including:

[0070] The image is embedded into the N-1 layer Transformer submodule in the input RLFE module for global encoding. The globally encoded image embedding and loss are embedded into the input feature enhancement submodule to obtain the enhanced feature map.

[0071] The enhanced feature map is input into the N-layer Transformer submodule for global encoding to obtain the final enhanced feature map.

[0072] Specifically, in this embodiment, the loss channel contains rich GT information. In order to fully utilize this information to guide feature extraction, a reverse loss feature enhancement (RLFE) method is proposed. Figure 4 As shown, the RLFE module contains two inputs: image embedding F embed and loss embedding Lembed First, F embed Input to N-1 layer Transformer submodule for global encoding, and then with L embed Input to the feature enhancement submodule to calculate the enhanced feature map F enhanc ::

[0073] F enhanc =F embed +F L

[0074] F L =L embed ·(mask F mask L )

[0075]

[0076] Among them, F enhanc is the enhanced feature map, F embed is the image embedding, F L For the enhancement component based on the loss channel, special attention is paid to the area where the loss is not 0 but the feature is 0, so that the features of the area are enhanced. F Embed mask for image, mask L Embedding mask for the parent loss.

[0077] like Figure 1 As shown in the figure, compared with traditional segmentation models, PL-TransUNet further converts the loss into a loss channel after calculating it. On the one hand, it is used as a priori reverse filling to the idle channel of the next generation training input. On the other hand, it converts the loss channel into a loss embedding to further enhance the image features of the encoder. The main effects are as follows:

[0078] (1) A reverse padding parent loss module is designed. This module calculates an additional loss channel after forward propagation, and the loss channel is reversed to fill the idle channel of the next generation training input. This method enables the next generation training to obtain additional ground truth (GT) information and use the parent loss prior to guide segmentation.

[0079] (2) A reverse loss feature enhancement method based on parent loss is proposed. This module is integrated after the CNN local encoder of the model. First, a multi-layer Transformer module extracts global features from the image embedding. Then, the extracted image features are enhanced based on the attention mechanism using the loss embedding. Finally, another layer of Transformer module is used to perform a global feature fusion. This method effectively utilizes the ground-truth information of the loss channel and further enhances the image features globally.

[0080] This example is compared with other SOTA methods such as U-Net and TransUnet on three retinal vessel segmentation datasets: DRIVE, CHASE_DB1, and LES-AV. The results show that PL-TransUNet has a higher accuracy of approximately 98.94%, 98.48%, and 98.30%, respectively.

[0081] The experiments were implemented using PyTorch 1.13 and run on a computer with a 3.5GHz Intel i5-13600KF processor, 32GB of 3600MHz DDR4 RAM, and a 12GB NVIDIA RTX4070Ti processor, running Windows 11. The experiments were conducted on three public retinal vascular datasets: DRIVE, CHASE_DB1, and LES-AV. First, the segmentation accuracy was compared with state-of-the-art methods, including AM-Unet, DA-TransUNe, and RRWNet. Second, ablation experiments were conducted to verify the effectiveness of BPL and RLFE. Finally, the segmentation results were visualized and discussed. The experimental parameters were set as follows: SGD was used as the optimizer with weight decay of 0.0001, momentum of 0.9, an initial learning rate of 0.01, and exponentially decayed learning rate updates. Max_epoch = 200, batch size = 1, α = 0.5, and γ = 2.

[0082] The relevant parameters of the dataset are shown in Table 1. The DRIVE dataset contains 40 retinal vascular images and their corresponding GTs with a resolution of 565×584. It is pre-divided into 20 samples for training and 20 samples for testing. The CHASE_DB1 dataset contains 28 retinal vascular images and their corresponding GTs with a resolution of 990×960. The training subset contains 20 samples, and the rest are used as test subsets. The LES-AV dataset contains 22 retinal vascular images and their corresponding GTs with an image resolution of 1620×1444. Its training set contains 16 images, and the remaining images are used as test sets. In order to alleviate the overfitting problem, 512×512 areas of the original image are randomly cropped as training samples during training.

[0083] Table 1

[0084]

[0085] The segmentation results are compared with the manually segmented ground truth. Sensitivity (SE), specificity (SP), accuracy (AC), Jaccard coefficient (JA), and Dice coefficient (DI, also known as F-measure) are calculated to measure the degree to which each pixel is correctly classified as background or blood vessel. The formulas for these indicators are as follows:

[0086]

[0087]

[0088] Among them, TP represents the number of true positive samples, FP represents the number of false positive samples, TN represents the number of true negative samples, and FN represents the number of false negative samples.

[0089] First, we compared the results with ten methods on the DRIVE dataset, including IterMiU-Net, HRD-Net, IMFF-Net, MILU-Net, AM-Unet, and DA-TransUNet. Table 2 shows the comparison results with existing methods on the DRIVE dataset. Bold values ​​indicate the best performance. PL-TransUNet ranked first in both AC and DI values. AM-Unet, which employs an attention mechanism and enhances the green channel to enhance image contrast, ranked second in AC value. DA-TransUNet, which also improves on the TransUNet structure and integrates spatial and channel attention, ranked third in AC value. HRD-Net and MILU-Net enhance features by fusing branch outputs of different resolutions or integrating multi-scale detail features. Their AC and DI values ​​are very similar, both lower than DA-TransUNet. IterMiU-Net focuses on vascular information from multiple dimensions, enhancing local and global information acquisition capabilities. IMFF-Net adopts a four-layer encoder-decoder framework based on attention pooling feature fusion. Both methods achieved lower AC and DI values.

[0090] Then, the results were compared with nine methods including IterMiU-Net, HRD-Net, IMFF-Net, AM-Unet, and DA-TransUNet on the CHASE_DB1 dataset. Table 3 shows the comparison results with existing methods on the CHASE_DB1 dataset. The bold values ​​indicate the best indicators. PL-TransUNet has the highest AC (98.48%), DI (85.30%), and JA (74.41%). HRD-Net ranks second in DI value. It fuses the outputs of branches with different resolutions through a global aggregation operation and integrates features using feature enhancement cascades. DA-TransUNet ranks third in DI value. It uses a dual attention mechanism in skip connections to capture finer features, but it easily leads to model overfitting. IterMiU-Net and IMFF-Net both have lower DI values.

[0091] Furthermore, the results were compared with eight methods including DA-TransUNet, FSE-Net, and RRWNet on the LES-AV dataset. Table 4 shows the comparison results with existing methods on the LES-AV dataset. The bold values ​​indicate the best indicators. PL-TransUNet also has the highest AC (98.29%), DI (85.86%), and JA (75.24%). TransUNet ranks second in AC value, combining the self-attention mechanism of Transformer and the encoder-decoder structure of U-Net. DA-TransUNet adopts a jump connection with a dual attention mechanism, and its AC value ranks third. FSE-Net ranks fourth in AC value, which processes the features of the distillation layer and the refinement layer through residual convolution and fuses multi-scale features. RRWNet is based on a fully convolutional neural network and has a lower AC value.

[0092] Table 2

[0093]

[0094] Table 3

[0095]

[0096] Table 4

[0097]

[0098]

[0099] Figure 6 The image segmentation visualization results of PL-TransUNet, TransUnet and DA-TransUNet on three datasets are shown. Figure 6 It can be seen that the compared methods can all segment most of the blood vessel images well, but for the blood vessels in the low contrast area, especially the microvessels, the segmentation results are different. Figure 6From the local magnified image, we can see that TransUnet only relies on the local and global features of the image itself for image segmentation. Its segmentation result is the worst, and most of the microvessels are classified as background. The segmentation result of DA-TransUnet is slightly better than that of TransUnet. It integrates the spatial and channel attention mechanism on the basis of TransUnet for feature enhancement. The relatively obvious microvessels are segmented, but the particularly small blood vessels are still not segmented, resulting in more prominent discontinuity of the blood vessels. PL-TransUNet has the best segmentation result, and most of the microvessels are correctly segmented. This is due to its parent loss integration mechanism, which improves the segmentation accuracy from the two aspects of prior integration and feature enhancement. It can be seen that PL-TransUNet has better segmentation effect in retinal vascular image segmentation, especially in the segmentation of low-contrast microvessels.

[0100] This embodiment employs a Backfill Parent Loss (BPL) module to convert the segmentation model's loss information into loss channels and backfill them into the idle channels of the next-generation training input. This information is used as prior knowledge to guide segmentation in the next-generation training. Simultaneously, RLFE utilizes loss embedding to perform attention-based feature enhancement on image embeddings, extracting more effective image features. Its effectiveness is demonstrated through experimental comparisons on three retinal vascular datasets. The proposed BPL and RLFE modules are relatively independent and can be applied to different semantic segmentation models.

[0101] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by persons skilled in the art should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A retinal vascular image segmentation method based on parent loss integration, characterized in that: include: Acquire a retinal blood vessel image to be segmented; Inputting the retinal vascular image to be segmented into a PL-TransUNet model to obtain an image segmentation visualization result, wherein the PL-TransUNet model is trained based on a training set, and the training set includes retinal vascular images and corresponding true values; The process of training the PL-TransUNet model based on the training set includes: predicting the retinal vascular image using the initial PL-TransUNet model, calculating a loss channel according to the prediction result of the retinal vascular image and the true value, performing backfilling and feature enhancement according to the loss channel to obtain a final enhanced feature map, and using the final enhanced feature map to obtain a new prediction result and iterate until convergence to obtain the PL-TransUNet model; The PL-TransUNet model includes an encoder, a decoder and a BPL module, wherein the encoder includes a CNN local encoder and an RLFE module, and the RLFE module includes several Transformer submodules; The BPL module is used to calculate a loss channel based on the prediction result of the retinal vascular image and the true value; The encoder is configured to perform feature enhancement on the retinal vascular image based on the loss channel to obtain the final enhanced feature map; The decoder is configured to output the new prediction result based on the final enhanced feature map.

2. The retinal vascular image segmentation method based on parent loss integration according to claim 1, characterized in that: Calculating a loss channel according to the prediction result and the true value of the retinal vascular image includes: inputting the prediction result and the true value into the BPL module to calculate a loss matrix; The maximum value of the loss matrix is ​​selected according to the channel and binarized to obtain the loss channel.

3. The retinal vascular image segmentation method based on parent loss integration according to claim 2, characterized in that: The method to calculate the loss matrix is: L m =-α(1-p l ) γ log(p l ) Among them, L m is the loss matrix, α is the balance factor, p l is the prediction result, and γ is the focusing parameter.

4. The retinal vascular image segmentation method based on parent loss integration according to claim 2, characterized in that: The method to obtain the loss channel is: Among them, L c is the loss channel, max c (·) means taking the maximum value by channel, C index Indicates the channel number of the maximum value, L m is the loss matrix, and (x,y) is the coordinate of the channel element.

5. The retinal vascular image segmentation method based on parent loss integration according to claim 2, characterized in that: Backfilling and feature enhancement are performed according to the loss channel to obtain the final enhanced feature map, including: Reversely filling the loss channel into the idle channel to obtain the filled idle channel; Taking the retinal blood vessel image and the padded idle channel as image input, inputting the image input and the loss channel into the CNN local encoder respectively, and calculating image embedding and loss embedding; The image embedding and the loss embedding are input into the RLFE module for global feature enhancement to obtain the final enhanced feature map.

6. The retinal vascular image segmentation method based on parent loss integration according to claim 5, characterized in that: Inputting the image embedding and the loss embedding into the RLFE module for global feature enhancement includes: Embedding the image into the N-1 layer Transformer submodule in the RLFE module for global encoding, embedding the globally encoded image and the loss into the feature enhancement submodule to obtain an enhanced feature map; The enhanced feature map is input into the N-layer Transformer submodule for global encoding to obtain the final enhanced feature map.

7. The retinal vascular image segmentation method based on parent loss integration according to claim 6, characterized in that: The method to obtain the enhanced feature map is: F enhanc =F embed +F L F L =L embed ·(mask F ·mask L ) Among them, F enhanc is the enhanced feature map, F embed is the image embedding, F L is the enhancement component based on the loss channel, mask F Embed mask for image, mask L is the parent loss embedding mask, (x, y) is the channel element coordinate, L embed is the loss embedding.

Images