Passive domain-adaptive fundus segmentation method based on evidence representation and marginal screening
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANJING UNIV OF POSTS & TELECOMM
- Filing Date
- 2026-06-10
- Publication Date
- 2026-07-10
Smart Images
Figure CN122368508A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of medical image processing, deep learning, and passive domain adaptation, specifically a passive domain adaptation fundus segmentation method based on evidence representation and marginal screening. Background Technology
[0002] Accurate segmentation of the optic disc and optic cup structures in fundus images is a crucial foundation for glaucoma screening, cup-to-disc ratio calculation, optic nerve head morphology analysis, and follow-up assessment of fundus diseases. With the development of deep learning technology, medical image segmentation methods based on convolutional neural networks have achieved good results in fundus structure segmentation tasks, and can improve the automation level of clinical auxiliary diagnosis to a certain extent.
[0003] However, fundus images exhibit significant cross-domain variability. Different hospitals, acquisition devices, imaging procedures, and individual differences among subjects can lead to substantial variations in color distribution, brightness contrast, sharpness, noise levels, and structural morphology. This results in a distribution shift between the training and target domains, causing segmentation models trained on source domain data to often exhibit significantly lower performance on new target domain data, particularly prone to problems such as inaccurate boundaries, missed classifications, and misclassifications.
[0004] To address the aforementioned issues, domain adaptation methods have been applied to medical image segmentation tasks. Existing methods typically improve target domain performance by transferring knowledge from the source domain. One type of method requires continued access to the source domain images and their annotations during the adaptation phase, reducing inter-domain differences through joint training, feature alignment, or adversarial learning. However, in medical data applications, factors such as privacy protection, data compliance, and sharing restrictions often prevent the original source domain images and annotations from being retrieved again during model deployment or transfer. In such cases, the only usable information is usually the model parameters after training in the source domain and the unannotated target domain images. Therefore, passive domain adaptation becomes a more practical approach.
[0005] In existing passive domain adaptive fundus image segmentation methods, a Teacher-Student or Mean-Teacher self-training framework is often used. This involves a Teacher network generating pseudo-labels on the target domain image, which are then used to supervise the Student network for iterative optimization. This type of method is simple to implement and does not rely on the original source domain data, thus having high application value. However, its performance is highly dependent on the quality of the pseudo-labels. If the Teacher network generates incorrect foreground pseudo-labels in the early predictions of the target domain, noisy supervision will be continuously introduced into the subsequent training process, easily causing error accumulation and noise propagation, thereby affecting the final segmentation result.
[0006] Most existing pseudo-label selection methods rely on threshold judgments based on a single predicted probability, confidence level, or entropy value. For example, pixels with predicted probabilities greater than a fixed threshold are directly used as positive pseudo-labels, or low-confidence pixels are removed using uncertainty indicators. While these methods can reduce the impact of noise to some extent, they still have significant shortcomings. On the one hand, a single probability value can only reflect the model's surface confidence in the foreground, making it difficult to distinguish between different situations such as "sufficient foreground support," "strong background counter-evidence," and "insufficient overall evidence." On the other hand, uncertainty estimation methods based on multiple sampling usually require repeated forward inference, resulting in high computational costs and hindering efficient training and deployment.
[0007] The aforementioned problems are even more pronounced for optic cup and optic disc segmentation tasks in fundus images. The optic disc region is relatively large and has clearer boundaries, while the optic cup region is typically smaller, has more blurred boundaries, and is more susceptible to the effects of vascular texture, local reflections, lesion occlusion, and changes in image quality. If pseudo-labels are selected based solely on a fixed probability threshold, some high-probability but unreliable pixel errors can easily be introduced into the training. If a uniform and simple selection rule is applied to different segmentation channels, large regions or background pixels can easily dominate the optimization process, resulting in insufficient learning of the optic cup channel and affecting cross-domain segmentation accuracy. Summary of the Invention
[0008] This invention provides a passive domain adaptive fundus segmentation method based on evidence representation and marginal screening. It aims to reduce the adverse effects of pseudo-label noise propagation on model performance and improve the accuracy, stability and reliability of optic cup and optic disc segmentation in target domain fundus images under the condition that the source domain training images and their pixel-level annotations are inaccessible. This solves the problems in the background art.
[0009] To achieve the above objectives, the technical solution of the present invention is as follows: A passive domain-adaptive fundus image segmentation method based on dual-channel beta evidence representation and evidence margin screening includes the following steps: S1. Obtain the unlabeled target domain fundus image set and the parameters of the source domain pre-trained segmentation model. Construct the Student deep segmentation network and the Teacher deep segmentation network under the condition that the source domain training images and their labels are inaccessible. S2, input the target domain fundus image into the Teacher depth segmentation network, extract the segmentation features of the target domain fundus image, and input the segmentation features into the evidence head set at the network output end. The evidence head outputs foreground positive evidence and background negative evidence for each pixel in each segmentation channel, and constructs the corresponding Beta distribution parameters. S3, calculate the pixel-level foreground probability based on the mean of the Beta distribution parameters, and construct the evidence margin and evidence strength based on the Beta distribution parameters to filter out reliable positive false labels and background false labels; S4, construct a target loss function based on the reliable positive pseudo-labels and background pseudo-labels, and use the target loss function to update the parameters of the Student deep segmentation network; S5. Based on the evidence margin and evidence strength corresponding to the reliable positive and negative labels and the background negative labels, the preset exponential moving average coefficients when constructing the Student deep segmentation network and the Teacher deep segmentation network are dynamically adjusted, and the Teacher deep segmentation network parameters are updated according to the updated Student deep segmentation network parameters using the dynamically adjusted exponential moving average coefficients. S6 uses the updated Student deep segmentation network to segment the target domain fundus image, and outputs the optic cup segmentation result and optic disc segmentation result of the target domain fundus image.
[0010] Preferably, the specific process steps of S1 include: S11, Obtain a set of unlabeled target-domain color fundus images. and source domain supervised pre-trained segmentation model parameters , A color fundus photograph containing the optic nerve head region. Image sequence number The total number of fundus images in the target domain is defined, and the segmentation objects are divided into two foreground channels: the optic cup and the optic disc. Construct Student and Teacher deep segmentation networks with identical structures, and ensure that their initial parameters satisfy the following conditions: ; Where s represents the Student deep segmentation network and t represents the Teacher deep segmentation network. These represent the initial parameters of the Student deep segmentation network. This represents the initial parameters of the Teacher deep segmentation network; S12, perform weak enhancement transformation on each target domain fundus image. and strong enhancement transformation The system includes a Teacher deep segmentation network that takes a weakly enhanced image as input and a Student deep segmentation network that takes a strongly enhanced image as input. Both the Student and Teacher deep segmentation networks employ the same encoder-decoder deep segmentation structure. The deep segmentation structure includes a backbone feature extraction network, a multi-scale context aggregation module, and a decoding and reconstruction module. The Teacher deep segmentation network is used to generate pseudo-label supervision information for the target domain, and the Student deep segmentation network is used to perform parameter optimization.
[0011] Preferably, the specific process steps of S2 include: S21, the weakly enhanced target domain fundus image The input is the Teacher deep segmentation network, which passes through the backbone feature extraction network, the multi-scale context aggregation module, and the decoding and reconstruction module to obtain the target domain feature map with segmentation semantic information. S22, at the network output end, set a positive evidence output branch and a negative evidence output branch for each segmented channel. For any pixel and any segmented channel, the positive evidence output branch provides the original response of the foreground positive evidence as follows: The negative evidence output branch provides background negative evidence and the original response is... ; S23. The original responses of the foreground positive evidence and the background negative evidence are processed by a non-negative activation function. By performing mapping, positive foreground evidence is obtained respectively. and negative evidence from the background ,Right now:
[0012]
[0013]
[0014] Where e represents evidence and z represents the original response. This is a Softplus function or another activation function that guarantees a non-negative output; S24. After obtaining positive foreground evidence and negative background evidence, construct the Beta distribution parameters respectively. and :
[0015]
[0016] in, This parameter represents evidence that the current pixel belongs to the foreground. This parameter represents evidence that the current pixel belongs to the background.
[0017] Preferably, the specific process steps of S3 include: S31, after obtaining the pixel-level Beta distribution parameters, calculate the average predicted probability that the current pixel belongs to the foreground, i.e., the foreground probability. Total strength of evidence They are respectively:
[0018]
[0019] Among them, the prospect probability The strength of evidence reflects the model's average tendency to classify the current pixel as belonging to the foreground. This reflects the total amount of evidence upon which the model made that judgment; S32, let the boundary of constructing evidence be... The evidentiary advantage of foreground assignments relative to background assignments is measured, and the marginality of the evidence is determined by the risk of foreground assignments. Risk of background assignment The difference construction, namely:
[0020]
[0021]
[0022] in, The Digamma function indicates that when the evidence margin reaches a preset margin threshold, it means that the current pixel has a clearer foreground tendency under the comparison of positive and negative evidence and can be used as a reliable pseudo-label candidate. S33, during pseudo-label generation, all pixels are initially screened based on foreground probability to obtain initial foreground pseudo-labels. Indicates the preset probability threshold, then the initial foreground pseudo-label. for:
[0023] in, The indicator function, based on the initial foreground pseudo-labels, retains only pixels that simultaneously satisfy the following conditions: foreground probability not lower than a preset probability threshold, evidence margin not lower than a preset margin threshold, and evidence strength not lower than a preset strength threshold, as reliable positive pseudo-labels. respectively , and Let probability threshold, marginal threshold, and intensity threshold represent the reliable positive / false labels, then they can be represented as:
[0024] in, Indicates a reliable, genuine / counterfeit label; S34, For reliable positive and negative label pixels obtained through joint screening, assign supervisory weights based on their evidence margin and evidence strength. :
[0025] in, This represents a weighted mapping function that monotonically increases with the margin and strength of evidence.
[0026] Preferably, the specific process steps of S4 include: S41, for binary classification pixel labels The cross-entropy loss for binary classification is:
[0027] in, For the true class label of the sample, the foreground probability Modeled using the Beta distribution, it satisfies:
[0028] Calculating the expectation of the binary cross-entropy over the Beta distribution yields the Bayes risk-based segmentation loss under the Beta distribution:
[0029] Where E represents the expected value, expanding the expression, we get:
[0030] The Beta distribution satisfies the following expected property:
[0031]
[0032] Substituting the above properties back, we obtain a closed-form expression for pixel-level Beta loss:
[0033] When the pixel label is foreground, the loss encourages the addition of positive evidence and suppresses negative evidence; when the pixel label is background, the loss encourages the addition of negative evidence and suppresses positive evidence. S42, let the Beta distribution parameters of the Student depth segmentation network output for the current pixel be denoted as... and The pseudo-labels generated by the Teacher deep segmentation network and jointly filtered are denoted as... Then weigh the evidence. Incorporating pixel-level Beta loss, we construct an overall weighted segmentation loss:
[0034] in, This represents the overall weighted loss. To prevent extremely small constants with a denominator of zero; S43, in addition to the weighted segmentation loss, introduces the Kullback-Leibler regularization loss. and the loss constrained by the strength of evidence Together, they constitute the target loss function:
[0035] in, This represents the Kullback-Leibler regularized loss. Indicates the loss constrained by the strength of evidence. and The weighting coefficients are used; the Kullback-Leibler regularization loss is used to constrain the current Beta distribution to approach the prior distribution, and the evidence strength constraint loss is used to suppress... Overconfident predictions caused by excessively large projections; S44. Backpropagation and parameter updates are performed on the Student deep segmentation network using the target loss function, enabling the Student network to segment the optic cup and optic disc on the unlabeled image in the target domain.
[0036] Preferably, the specific process steps of S5 include: S51 introduces a class-balanced reweighting coefficient on the background pseudo-label pixels of the visual cup channel to reduce the impact of the imbalance between the number of foreground and background samples on parameter updates. S52, after completing the current round update of the Student network, the difference in prediction probabilities between the Student and Teacher networks is weighted and statistically analyzed based on the evidence weights corresponding to the reliable positive and negative labels, to obtain the degree of difference:
[0037] in, The evidence weight corresponding to the reliable positive and negative label pixels, This represents the predicted probability of the Student deep segmentation network. This represents the predicted probability of the Teacher deep segmentation network; the dissimilarity is used to reflect the degree of consistency between the teacher and student networks in the high-reliability region. S53, Based on the aforementioned degree of difference, adaptively determine the dynamic exponential moving average coefficient. Its expression is:
[0038] in, and Let represent the upper and lower bounds of the EMA coefficient, respectively, and satisfy . ; S54, after obtaining the dynamic EMA coefficients, update the Teacher depth segmentation network parameters:
[0039] Where r represents the exponential moving average coefficient; S55, after completing the Teacher network update, uses the new Teacher network to regenerate the target domain pseudo-labels in the next training round, thus forming a closed-loop iterative mechanism.
[0040] Preferably, the specific process steps of S6 include: S61, after completing the target domain adaptation training, the target domain fundus image is input into the updated Student depth segmentation network for forward inference to obtain the Beta distribution parameters corresponding to the optic cup channel and optic disc channel, and the foreground probability maps of the optic cup and optic disc are calculated respectively:
[0041]
[0042] in, This represents the probability diagram of the viewing cup channel. This represents the probability diagram of the viewing channel; S62, generate corresponding binary segmentation results based on the probability map of the viewing cup channel and the probability map of the viewing disk channel, so as to... and The threshold for segmentation of the optic cup and optic disc is represented as follows:
[0043]
[0044] in, This represents the result of the sight cup segmentation. This indicates the result of the visual disc segmentation.
[0045] As can be seen from the above technical solution compared with the prior art, the present invention has the following beneficial effects: This invention achieves effective adaptation to target domain images under conditions where source domain data is inaccessible by employing a Student-Teacher dual-network structure and co-training with weak and strong enhancements. Through dual-channel Beta evidence representation, it extends the traditional single-probability output to joint modeling of foreground positive evidence and background negative evidence, enabling the model output to simultaneously possess probabilistic and evidentiary information. By jointly filtering foreground probability, evidence margin, and evidence strength, it effectively reduces the risk of noisy pseudo-labels entering the training process. The stability of self-training optimization is improved through Bayes risk segmentation loss under Beta distribution, Kullback-Leibler regularization loss, and evidence strength constraint loss. Furthermore, the dynamic EMA update mechanism based on evidence quantity enhances the Teacher network's adaptive tracking ability to changes in Student network parameters. Therefore, this invention can effectively improve the accuracy, robustness, and generalization ability of optic cup and optic disc segmentation in target domain fundus images. Attached Figure Description
[0046] Figure 1 This is an overall structural diagram of the method of the present invention; Figure 2 This is a visualization comparison of the medical image segmentation method of this invention with other medical image segmentation methods; Figure 3 This is a structural block diagram of a computer device according to an embodiment of the present invention. Detailed Implementation
[0047] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.
[0048] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are used to illustrate the present invention, but should not be used to limit the scope of the present invention.
[0049] Example: like Figure 1 As shown in this embodiment, a passive domain adaptive fundus segmentation method based on evidence representation and marginal screening is applied to the passive domain adaptive fundus image segmentation task. It is particularly suitable for automatically segmenting the optic cup and optic disc in the target domain fundus image using only the unlabeled target domain fundus image and the source domain supervised pre-trained model when the source domain training image and its pixel-level annotations are inaccessible.
[0050] The core idea of this method is to no longer directly use the point probabilities output by the Teacher network as pseudo-supervision. Instead, it first performs dual-channel beta evidence representation on the target domain pixels, and then combines evidence margin and evidence strength to screen reliable pseudo-labels, thereby achieving progressive denoising learning oriented towards the target domain.
[0051] This embodiment includes the following steps: S1. Obtain the unlabeled target domain fundus image set and the parameters of the source domain pre-trained segmentation model. Construct the Student deep segmentation network and the Teacher deep segmentation network under the condition that the source domain training images and their annotations are inaccessible. Specifically, the steps in S1 are as follows: S11. Obtain unlabeled fundus images of the target domain and parameters of the pre-trained model of the source domain; First, the Fundus dataset was obtained as the source of fundus image data used in this embodiment. The Fundus dataset is a commonly used cross-domain benchmark dataset in the field of fundus cup and optic disc segmentation. It contains four subdomains composed of different devices, acquisition conditions and institutional sources: Drishti-GS dataset, RIM-ONE-r3 dataset, REFUGE training subset, and REFUGE validation subset, which are used to simulate the inter-domain distribution differences of fundus images in real-world scenarios.
[0052] In this embodiment, the REFUGE dataset is used as the source domain to train the parameters of the pre-trained segmentation model that provides source domain supervision. The Drishti-GS and RIM-ONE-r3 datasets are used to perform passive domain adaptation training under the condition that the source domain training images and their annotations are inaccessible. That is, during the target domain adaptation stage, the original images and pixel-level annotations of Domain3 are no longer accessed; only the model parameters obtained from Domain3-supervised training are retained. Subsequently, unlabeled fundus images from other subdomains are input into the method for adaptation. This setup can simulate a passive domain adaptation scenario in real-world applications where a model, after being trained on a certain data source, is transferred to data from other unknown devices or institutions.
[0053] S12. Construct the Student-Teacher dual network and initialize its parameters; Then, obtain a set of unlabeled target-domain color fundus images. and source domain supervised pre-trained segmentation model parameters ,in A color fundus photograph containing the optic nerve head region. Image sequence number The target domain consists of a total number of fundus images. The target segmentation object includes two foreground channels: the optic cup and the optic disc, corresponding to the smaller internal region of the cup / disc and the outer disc-shaped structure, respectively. Two identical depth segmentation networks, Student and Teacher, are constructed, with their initial parameters set to satisfy the following:
[0054] Where s represents the Student deep segmentation network and t represents the Teacher deep segmentation network. These represent the initial parameters of the Student deep segmentation network. This represents the initial parameters of the Teacher deep segmentation network. By employing a homogeneous dual-network design, the Teacher network serves as the pseudo-label generation branch, while the Student network serves as the parameter optimization branch. Both maintain a consistent structural foundation in the parameter space, facilitating stable updates via exponential moving averages. Furthermore, the Teacher network is not directly updated via backpropagation; instead, the Student network parameters are smoothly tracked through a subsequent exponential moving average mechanism, thereby improving the stability of the pseudo-label generation process.
[0055] S13. Perform weak and strong enhancement on the target domain image; For each target domain fundus image Perform weak enhancement transformation respectively and strong enhancement transformation and strong enhancement transformation The image after strong enhancement is input into the Teacher deep segmentation network. Input the Student deep segmentation network.
[0056] Weak enhancements preferably include size adjustment, normalization, slight flipping, or other conservative transformations that preserve the original fundus structure information as much as possible. Strong enhancements can further include random noise addition, brightness perturbation, color shifting, blurring, random erasure, or other perturbation operations that can increase input diversity. This "two-branch, different enhancement strength" input method is adopted because the Teacher network, responsible for pseudo-label generation, requires relatively stable input to reduce pseudo-label fluctuations; the Student network, responsible for robust learning of the target domain, needs to improve its adaptability to changes in the target domain distribution under stronger data perturbations. This setup forms a self-training mechanism where "the Teacher generates stable pseudo-supervision, and the Student learns robust representations under perturbation conditions."
[0057] S14. Construct a DeepLabV3+-style basic segmentation framework; The Student deep segmentation network and the Teacher deep segmentation network adopt the same encoder-decoder deep segmentation structure. This embodiment uses a DeepLabV3+ network as the basic segmentation framework. Its writing style is consistent with the organization method in the template of "first stating the use of a mature architecture, and then further expanding the encoder, decoder and output layer".
[0058] More specifically, the encoder utilizes a backbone feature extraction network to extract features layer by layer from the input fundus image. Through continuous convolution operations, it progressively reduces the spatial resolution of the feature map and increases the number of channels, thereby obtaining high-level features containing rich semantic information. The multi-scale context aggregation module employs multiple dilated convolution branches with different dilation rates and a global context branch to process the encoded features in parallel and fuse contextual information from different receptive fields, enhancing the network's ability to represent the optic cup region, optic disc region, and their surrounding background at different scales. The decoder concatenates and fuses the high-level semantic features with shallow detail features, and through convolution and upsampling, progressively restores the feature map resolution, thus obtaining pixel-level segmentation feature responses corresponding to the input image. The Teacher network is mainly used to generate pseudo-label supervision information for the target domain, while the Student network is mainly used for parameter optimization.
[0059] S2. Input the target domain fundus image into the Teacher depth segmentation network to extract segmentation features from the target domain fundus image. Input these segmentation features into the evidence header set at the network output. The evidence header outputs foreground positive evidence and background negative evidence for each pixel in each segmentation channel, constructing the corresponding Beta distribution. Specifically, steps S2 are as follows: S21. Extract target domain segmentation features; The weakly enhanced target domain fundus image The input is fed into the Teacher deep segmentation network, which, after passing through the backbone feature extraction network, multi-scale context aggregation module, and decoding and recovery module, yields a target domain feature map with segmentation semantic information. This feature map simultaneously contains spatial location information and semantic distinction information of the visual cup region, visual disc region, and background region, serving as the foundation for subsequent evidence modeling, probability estimation, and pseudo-label screening.
[0060] S22. Set up a dual-channel evidence header and output the original evidence response; At the network output, instead of directly using traditional Softmax or Sigmoid to output a single probability, a positive evidence output branch and a negative evidence output branch are set for each foreground channel. For any pixel and any segmented channel, the positive evidence output branch provides the original foreground positive evidence response. The negative evidence output branch provides the background negative evidence and the original response. Taking the visual cup channel as an example, Used to characterize the raw response that "supports the pixel belonging to the foreground of the viewing cup". The raw response used to characterize "supports that the pixel belongs to the background of the viewing cup"; similarly, the same logic can be applied to the viewing disc channel. and Through this two-branch design, the model no longer outputs just a foreground score that is difficult to interpret. Instead, it explicitly distinguishes between positive evidence supporting the foreground and negative evidence supporting the background, thus providing a richer and more interpretable basis for subsequent judgments on the reliability of pseudo-labels.
[0061] S23. Obtain positive or negative evidence through a non-negative activation function; In order for the network output to represent the amount of non-negative evidence, the original response of the foreground positive evidence is... and background negative evidence original response Input non-negative activation functions respectively Mapping is performed. In this embodiment, the Softplus function is preferably used as the non-negative activation function, i.e.:
[0062]
[0063]
[0064] The non-negative activation function can be the Softplus function or other activation functions that guarantee non-negativity of the output. After this mapping, the model output no longer merely represents classification tendency, but can be further interpreted as the amount of evidence supporting a certain category judgment.
[0065] S24. Construct pixel-level Beta distribution parameters; After obtaining positive foreground evidence and negative background evidence, the Beta distribution parameters are constructed respectively. and ,Right now
[0066]
[0067] in, This parameter represents evidence that the current pixel belongs to the foreground. This represents the evidence that the current pixel belongs to the background. Incrementing by 1 as described above ensures... This satisfies the parameter requirements of the Beta distribution.
[0068] In this way, the prediction result for each pixel and each segmented channel is no longer a single point probability, but a combination of... A jointly determined Beta distribution. The same dual-channel evidence modeling process is performed on both the optic cup and optic disc channels to obtain the Beta distribution parameter fields corresponding to each of the two foreground channels.
[0069] S3. Calculate the pixel-level foreground probability based on the Beta distribution parameters, and construct the evidence margin and evidence strength using the Beta distribution parameters to screen reliable positive and false labels. Specifically, the steps in S3 are as follows: S31. Calculate the probability of the prospect and the strength of evidence; After obtaining the pixel-level Beta distribution parameters, the average predicted probability that the current pixel belongs to the foreground is first calculated based on the parameters. Total strength of evidence They are respectively:
[0070]
[0071] Among them, the prospect probability The strength of evidence reflects the model's average tendency to classify the current pixel as belonging to the foreground. This reflects the total amount of evidence upon which the model made that judgment. It addresses the question of whether something resembles the foreground or the background more. The problem addressed is whether there is sufficient evidence to support this judgment. If a pixel has a high foreground probability but low evidence strength, it means that although the model tends to classify it as foreground, this judgment is not robust and cannot be directly regarded as a high-quality pseudo-label.
[0072] S32. Constructing the Margin of Evidence When filtering pseudo-labels solely based on foreground probabilities, some noisy pixels with high probabilities but unclear positive-negative evidence relationships are easily retained. Therefore, this embodiment further constructs an evidence margin *m* to measure the evidentiary advantage of foreground assignments relative to background assignments. First, the risk of foreground assignment is defined. Risks associated with background assignment for:
[0073]
[0074] in For the Digamma function, further, the evidence margin m is defined as:
[0075] Among them, when A larger value indicates that positive evidence is more dominant than negative evidence, and the current pixel is more suitable as a foreground positive / negative label; conversely, a smaller value indicates that... When the value is small or even negative, it indicates that the background evidence is stronger or the relationship between positive and negative evidence is unclear, and it is not appropriate to directly use it as reliable prospect surveillance.
[0076] S33. Generate initial foreground pseudo-labels; During pseudo-label generation, all pixels are first preliminarily filtered based on foreground probability to obtain initial foreground pseudo-labels. If... Indicates the preset probability threshold, then the initial foreground pseudo-label. It can be written as: in, This indicates an indicator function. When the foreground probability of a pixel is higher than a threshold... If the pixel is in the foreground, it is initially considered a candidate pixel; otherwise, it is considered a non-foreground pixel. This step is equivalent to performing a coarse screening across the entire image to reduce the computational scope of the subsequent fine screening.
[0077] S34. Combined prospect probability, evidence margin, and evidence strength to screen reliable positive and negative labels; Based on the initial foreground pseudo-labels, only pixels that simultaneously satisfy the following conditions are retained as reliable positive pseudo-labels: foreground probability not lower than a preset probability threshold, evidence margin not lower than a preset margin threshold, and evidence strength not lower than a preset strength threshold. If we consider... , and Representing probability threshold, marginal threshold, and intensity threshold, reliable positive / false labels are obtained. It can be represented as:
[0078] Specifically, the foreground probability constraint ensures that the current pixel has sufficient foreground tendency, the evidence margin constraint ensures that positive evidence is dominant over negative evidence, and the evidence strength constraint ensures that the judgment is based on sufficient evidence. Through the joint constraint of these three factors, the probability of noisy foreground pseudo-labels entering the training process can be significantly reduced.
[0079] S35. Assign supervisory weights and construct background pseudo-labels; For reliable positive and negative label pixels obtained through joint screening, supervisory weights are then assigned based on their evidence margin and evidence strength. The supervisory weight is essentially a function of the marginality and strength of evidence, and can be denoted as:
[0080]
[0081] Where M is the set of reliable positive and negative label pixels selected. The sigmoid function is represented by k, which is a scaling factor used to control the steepness of the sigmoid function. In this way, pseudo-labels with more substantial evidence and more obvious margins will have a greater weight in training, while pseudo-labels that just barely exceed the threshold will be given a smaller weight, thus further improving the "pseudo-label screening" to "reliability-weighted application".
[0082] After obtaining reliable positive and negative pseudo-labels, their complementary regions can be used to construct background pseudo-labels, which, together with the foreground pseudo-labels, serve as training supervision signals for the target domain. For the two segmentation channels, the visual cup and visual disk, this process can simultaneously enhance the learning of the foreground region and the suppression of the background region, providing a stable supervision basis for subsequent loss construction.
[0083] S4. Construct a target loss function based on the reliable positive and negative labels and the background negative labels, and use the target loss function to update the parameters of the Student deep segmentation network. Specifically, the steps in S4 are as follows: S41. Constructing pixel-level Bayes risk loss under a Beta distribution For binary classification pixel labels The standard binary classification cross-entropy loss is written as:
[0084] in For the true class label of the sample, in traditional methods, The deterministic probability of the network output is directly taken; however, in this invention, the foreground probability is... Modeled using the Beta distribution, it is no longer a fixed value, but satisfies:
[0085] Therefore, if the standard BCE is still used, it cannot reflect the impact of the evidence distribution itself on the prediction reliability. To ensure that the loss function is consistent with the evidence modeling method, this invention calculates the expectation of the aforementioned binary cross-entropy over the Beta distribution, obtaining the Bayes risk-form segmentation loss under the Beta distribution:
[0086] Where E represents the expected value, expanding the expression yields:
[0087] The Beta distribution satisfies the following expected property:
[0088]
[0089] Substituting the above properties back, we can obtain a closed-form expression for pixel-level Beta loss:
[0090] This loss form is essentially an evidence extension of the traditional BCE: when the pixel label is foreground, the loss encourages the addition of positive evidence and suppresses negative evidence; when the pixel label is background, the loss encourages the addition of negative evidence and suppresses positive evidence.
[0091] S42. Construct a weighted segmentation loss; Let the Beta distribution parameters of the Student deep segmentation network output for the current pixel be denoted as... and The pseudo-labels generated by the Teacher deep segmentation network and jointly filtered are denoted as... The corresponding pixel-level segmentation loss It can be written as:
[0092] This formula shows that during the training process in the target domain, the Student network does not simply fit a point value probability label, but rather approximates the category judgment corresponding to high-quality pseudo-labels at the evidence distribution level. Considering the different reliability of different pseudo-label pixels, this invention further increases the evidence weights. Incorporating pixel-level Beta loss, we construct an overall weighted segmentation loss:
[0093] in, To prevent extremely small constants with a denominator of zero, this weighting method allows high-confidence, high-evidence, and marginal pseudo-labels to have a greater impact on parameter updates, while less reliable pseudo-labels are automatically weakened, thereby reducing the interference of erroneous pseudo-supervision on the training of the Student network.
[0094] S43. Introduce KL regularization loss and evidence strength constraint loss; In addition to the weighted segmentation loss, this invention also introduces the Kullback-Leibler regularization loss. and the loss constrained by the strength of evidence Together, they constitute the target loss function:
[0095] in, and The weighting coefficients are for the weighted segmentation loss and the Kullback-Leibler regularization loss. This represents the weighted segmentation loss described earlier. This represents the Kullback-Leibler regularization loss, used to constrain the Beta posterior distribution of the Student deep segmentation network output to move closer to the prior distribution, in order to prevent the model from forming an unreasonable sharp distribution in the early stages of training.
[0096] For any pixel and any segmentation channel, let the posterior distribution of the Student depth segmentation network output be... The prior distribution is taken as uniform prior. The corresponding Kullback-Leibler divergence can be expressed as:
[0097] Where C is the number of categories, If the index is a category, then the corresponding pixel-level evidence strength constraint loss is... Represented as:
[0098] in, To preset the maximum threshold for the strength of evidence, The evidence strength of the current pixel u is represented by the average of this strength across all training pixels and all segmentation channels, yielding the overall evidence strength constraint loss.
[0099] in, This represents the evidence strength constraint loss, used to suppress overconfident predictions caused by an excessive amount of positive and negative evidence. For all segmented pixels; the strength of evidence for the current pixel. No more than No additional penalty is imposed at the beginning; constraints are only applied to the excess portion when the total amount of evidence exceeds a threshold. The purpose of this loss is to prevent the model from blindly increasing the amount of evidence when there are still errors in the false labels, thereby reducing the adverse effects of overconfident predictions on subsequent training.
[0100] S44. Update Student network parameters; Using the aforementioned total loss function, backpropagation and parameter updates are performed on the Student deep segmentation network, enabling it to gradually develop a more accurate segmentation capability for the view cup and view disk in unlabeled images of the target domain. As training epochs progress, the Student network's ability to represent the structural boundaries of the target domain gradually increases, while its sensitivity to pseudo-label noise gradually decreases.
[0101] S5, dynamically adjust the exponential moving average coefficient based on the amount of evidence corresponding to the pseudo-labels, and update the Teacher depth segmentation network parameters using the dynamically adjusted exponential moving average coefficient and the updated Student depth segmentation network parameters. Specifically, steps S5 are as follows: S51. Introduce class-balanced reweighting for pseudo-labels on the background of the viewing cup channel. In target domain adaptation training, the view cup region is typically small, while the background region accounts for a significantly larger proportion. If supervision is applied to all background pixels equally, the training process can easily become biased towards the background category. To address this, this invention introduces a class-balanced reweighting coefficient on the background pseudo-label pixels of the view cup channel to mitigate the impact of the imbalance between foreground and background samples on parameter updates, thereby making it more beneficial for the Student network to learn the small target structure of the view cup.
[0102] S52. Calculate the evidence-weighted dissimilarity of teacher-student networks. After completing the current round update of the Student network, instead of using a fixed EMA coefficient to update the Teacher network, we first perform a weighted statistical analysis on the difference in prediction probabilities between the Student and Teacher networks based on the evidence weights corresponding to reliable positive and negative labels, to obtain the degree of difference:
[0103] in The evidence weight corresponding to the reliable positive and negative label pixels, This represents the predicted probability of the Student deep segmentation network. This represents the predicted probability of the Teacher deep segmentation network. This dissimilarity is used to reflect the degree of consistency between the teacher and student networks in high-reliability regions; when... When the value is larger, it indicates that there are more significant new changes in the Student relative to the Teacher; when When the value is small, it indicates that the Teacher is able to track the Student's current state quite well.
[0104] S53. Determine the dynamic EMA coefficient based on the degree of difference; Based on the degree of difference Adaptive determination of dynamic exponential moving average coefficients Its expression is:
[0105] in and Let represent the upper and lower bounds of the EMA coefficient, respectively, and satisfy:
[0106] Therefore, when When it is large, This will decrease accordingly, allowing the Teacher network to absorb new parameters from the Student network more quickly; when When smaller, This will increase the size of the network, allowing the Teacher network to maintain a smoother parameter evolution, thus balancing rapid adaptation and training stability.
[0107] S54. Update Teacher network parameters; After obtaining the dynamic EMA coefficients, update the Teacher depth segmentation network parameters according to the following formula:
[0108] Through this dynamic update method, the Teacher network no longer tracks the Student network at a fixed speed, but instead adaptively adjusts the update intensity based on the current training state of the target domain and the reliability of the pseudo-labels. This reduces pseudo-label oscillations and facilitates the generation of more stable and accurate target domain pseudo-labels in the later stages of training.
[0109] S55. Forming a closed-loop iteration. After updating the Teacher network, the new Teacher network is used to regenerate the target domain pseudo-labels in the next training round, thus forming a closed-loop iteration mechanism. As the iteration progresses, the quality of the pseudo-labels generated by the Teacher gradually improves, and the target domain segmentation ability of the Student network also increases simultaneously. The two promote each other, ultimately improving the model's convergence performance in the target domain.
[0110] S6. The updated Student depth segmentation network is used to segment the target domain fundus image, outputting the optic cup and optic disc segmentation results. The segmentation performance is evaluated using the Dice coefficient and the average symmetric surface distance (ASSD). Specifically, the steps in S6 are as follows: S61. Calculate the probability diagrams of the viewing cup and viewing disk channels; After completing the target domain adaptation training, the target domain fundus image is input into the updated Student depth segmentation network for forward inference, obtaining the Beta distribution parameters corresponding to the optic cup channel and optic disc channel, respectively. and .
[0111] Furthermore, the foreground probability maps for the visual cup and visual disk can be calculated separately:
[0112]
[0113] in, This represents the probability diagram of the viewing cup channel. This represents the probability map of the viewport channels. The probability map reflects the predicted result of each pixel belonging to the corresponding foreground category and is the basis for generating the final segmentation result.
[0114] S62. Generate binary segmentation results; Generate corresponding binary segmentation results based on the probability maps of the viewing cup and viewing disc channels. If... and Let represent the segmentation threshold for the optic cup and optic disc. Then, the corresponding segmentation result can be expressed as:
[0115]
[0116] in, This represents the result of the sight cup segmentation. This represents the optic disc segmentation result. Through the above reasoning process, the optic cup segmentation result and optic disc segmentation result of the target domain fundus image can be output, providing a reliable basis for subsequent optic cup-optic disc ratio calculation, glaucoma-assisted screening, and clinical structural analysis.
[0117] S63. The Dice coefficient is used to evaluate the degree of overlap of the segmented regions; To evaluate the segmentation performance of the method of this invention on fundus images in the target domain, the Dice coefficient is used to quantitatively assess the overlap between the predicted segmented region and the ground truth labeled region. Let the predicted segmentation result be M, and the ground truth labeled result be G, then the Dice coefficient can be expressed as:
[0118] in, This represents the number of pixels where the predicted segmentation result intersects with the ground truth annotation result. This indicates the number of foreground pixels in the predicted segmentation result. This represents the number of foreground pixels in the ground truth annotation result. The Dice coefficient ranges from 0 to 1; a larger value indicates a higher degree of overlap between the predicted segmentation region and the ground truth annotation region, resulting in better segmentation performance.
[0119] In this embodiment, the Dice coefficients corresponding to the visual cup segmentation results and the visual disk segmentation results are calculated respectively to evaluate the segmentation accuracy of the method in the sense of region overlap.
[0120] S65. The accuracy of the segmentation boundary is evaluated using the average symmetric surface distance (ASSD). In addition to the Dice coefficient, this embodiment also uses the Average Symmetric Surface Distance (ASSD) as an evaluation metric to further measure the geometric proximity between the predicted segmentation boundary and the ground truth labeled boundary. Let the set of boundary points of the predicted segmentation result be... The true set of labeled boundary points is ,point Point set The shortest distance is denoted as ,point Point set The shortest distance is denoted as Then ASSD can be represented as:
[0121] in, This indicates the number of boundary points in the predicted segmentation result. The number of boundary points in the predicted segmentation result is represented by ASSD. ASSD measures the average symmetric distance between the predicted boundary and the true boundary. The smaller the value, the closer the predicted boundary is to the true boundary, indicating that the model has higher accuracy in boundary localization.
[0122] In this embodiment, the corresponding ASSD is calculated for the optic cup segmentation result and the optic disc segmentation result, respectively, to evaluate the segmentation quality of the method in the sense of boundary localization.
[0123] S66. The segmentation results are comprehensively evaluated based on the evaluation indicators; After segmenting the target domain fundus image, the Dice coefficient and ASSD are used as region overlap and boundary error indicators, respectively, to comprehensively evaluate the optic cup and optic disc segmentation results. The Dice coefficient characterizes the consistency between the predicted and ground truth regions, while ASSD characterizes the average distance between the predicted and ground truth boundaries. By simultaneously employing these two evaluation indicators, the segmentation performance of the method of this invention can be comprehensively evaluated from both region overlap and boundary localization perspectives.
[0124] like Figure 2 A comparison of the results of different methods in optic cup / optic disc segmentation in fundus images; The left side shows an example of the Drishti-GS dataset with relatively small domain differences, while the right side shows an example of the RIM-ONE-r3 dataset with relatively large domain differences.
[0125] Each row, from top to bottom, consists of: the original color fundus image; the true segmentation annotation (optic disc and optic cup boundary); the segmentation result of the existing method CBMT (MICC11 2023, International Conference on Medical Image Computation and Computer-Aided Intervention); and the segmentation result of the method of this invention.
[0126] As can be seen from the comparison, when the domain difference is small (Drishti-GS), both methods can obtain relatively complete segmentation boundaries; while when the domain difference is large (RIM-ONE-r3), CBMT shows obvious shape deviation and edge loss, while the method of the present invention can still maintain the optic disc / optic cup contour that is highly consistent with the real result, verifying the robustness and accuracy of the present method in cross-domain segmentation tasks.
[0127] Table 1 presents a quantitative comparison of the method of this invention with several existing medical image segmentation domain adaptation methods on the optic cup and optic disc segmentation task in the target domain fundus images. Experiments were conducted on the Drishti-GS dataset and the RIM-ONE-r3 dataset, both of which were used as target domain data to verify the segmentation performance of different methods on the optic disc and optic cup structures in fundus images under cross-domain conditions.
[0128] Table 1. Comparison of the performance indicators of the method of the present invention with other medical image segmentation methods.
[0129] In Table 1, the "Method" column represents the different segmentation or domain adaptation methods being compared; the "SF" column indicates whether the method belongs to the passive domain adaptation method, where "√" indicates that the method does not need to access the source domain training images and their annotations during the target domain adaptation stage, and "×" indicates that the method usually still relies on source domain data or source domain-related supervision information during the adaptation process. Evaluation metrics include the Dice coefficient and ASSD.
[0130] Among them, the Dice coefficient is used to measure the degree of overlap between the predicted segmented region and the ground truth labeled region. The larger the value, the more accurate the segmentation region. The ASSD represents the average symmetric surface distance, which is used to measure the average distance between the predicted boundary and the ground truth boundary. The smaller the value, the closer the segmentation boundary is to the ground truth label.
[0131] Table 1 shows the Dice and ASSD values for spectral disc segmentation and spectral cup segmentation, respectively. "Source domain only" in Table 1 indicates that the segmentation model is trained using only the source domain labeled data and then directly applied to the target domain image without any target domain adaptation.
[0132] This result reflects the performance of the source domain model in directly generalizing to the target domain data, which is usually affected by the differences in distribution between domains and thus suffers from performance degradation. "Target domain only" means the result obtained by supervised training using the target domain image and its manual annotations. This result can be used as a reference upper limit under supervised training conditions in the target domain, but it is usually not available in actual unlabeled target domain scenarios.
[0133] Table 1 also lists several existing domain adaptation or test-time adaptation methods as comparison methods. Among them, BEAL (MICCAI 2019, International Conference on Computational Medical Images and Computer-Aided Intervention) is a domain adaptation method for fundus image segmentation, which mainly reduces the difference between the source and target domains through strategies such as boundary information and adversarial learning; AdvEnt (CVPR 2019, International Conference on Computer Vision and Pattern Recognition) is a domain adaptation method based on adversarial entropy minimization, which improves the deterministic nature of target domain prediction by constraining the target domain prediction entropy distribution; TENT (ICLR 2021, International Conference on Learning Representations) is a test-time entropy minimization method, which adjusts model parameters by minimizing prediction entropy during target domain testing or adaptation, thereby improving the stability of target domain prediction; DPL (MICCAI 2021, International Conference on Computational Medical Images and Computer-Aided Intervention) is a domain adaptation method based on pseudo-label learning, which constructs pseudo-supervision signals by utilizing the model's prediction results in the target domain; CPR (MICCAI 2023, International Conference on Computational Medical Images and Computer-Aided Intervention) improves target domain feature learning by introducing class relationships, prediction constraints, or regularization mechanisms; CBMT (MICCAI) The 2023 International Conference on Medical Image Computation and Computer-Aided Intervention (WACV 2024, Winter Conference on Applications of Computer Vision) employs a Mean-Teacher framework and incorporates class balancing to alleviate class imbalance during target domain training. PLPB (ISBI 2025, International Symposium on Biomedical Imaging) improves the quality of pseudo-supervision in the target domain through pseudo-label propagation or pseudo-label selection strategies. SBIF (ISBI 2025, International Symposium on Biomedical Imaging) is a recently proposed passive domain-adaptive fundus image segmentation method that enhances cross-domain segmentation performance by improving target domain self-training or feature constraint mechanisms. All of the above methods serve as baselines for verifying the effectiveness of this invention in this embodiment.
[0134] Furthermore, the present invention also discloses a computer device, including a memory and a processor, such as... Figure 3 As shown. The computer device includes a processor, memory, input devices, output devices, and a network interface connected via a system bus. The processor provides computing and control capabilities; the memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs, while the internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The input device is used to receive fundus image data, model training parameters and segmentation detection instructions input by the user; the output device is used to display the optic cup segmentation results, optic disc segmentation results and prediction uncertainty heatmap; the network interface is used to communicate with external medical image archiving and communication systems (PACS) and data servers to realize the uploading of fundus image data and the distribution of segmentation results.
[0135] It is understood that the systems, devices, and storage media provided in the embodiments of the present invention correspond to the methods provided in the embodiments of the present invention, and the explanations, examples, and beneficial effects of the relevant content can be referred to the corresponding parts of the above methods.
[0136] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another.
[0137] For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that a computer can access, or a data storage device such as a server or data center that integrates one or more available media.
[0138] The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid state disks (SSDs)).
[0139] It should be noted that in this document, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations.
[0140] Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0141] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.
[0142] The embodiments of the present invention are given for the purposes of illustration and description. Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.
Claims
1. A passive domain adaptive fundus segmentation method based on evidence representation and marginal screening, characterized in that, Perform the following steps using a computer device: S1. Obtain the unlabeled target domain fundus image set and the parameters of the source domain pre-trained segmentation model. Construct the Student deep segmentation network and the Teacher deep segmentation network under the condition that the source domain training images and their labels are inaccessible. S2, input the target domain fundus image into the Teacher depth segmentation network, extract the segmentation features of the target domain fundus image, and input the segmentation features into the evidence head set at the network output end. The evidence head outputs foreground positive evidence and background negative evidence for each pixel in each segmentation channel, and constructs the corresponding Beta distribution parameters. S3, calculate the pixel-level foreground probability based on the mean of the Beta distribution parameters, and construct the evidence margin and evidence strength based on the Beta distribution parameters to filter out reliable positive false labels and background false labels; S4, construct a target loss function based on the reliable positive pseudo-labels and background pseudo-labels, and use the target loss function to update the parameters of the Student deep segmentation network; S5. Based on the evidence margin and evidence strength corresponding to the reliable positive and negative labels and the background negative labels, the preset exponential moving average coefficients when constructing the Student deep segmentation network and the Teacher deep segmentation network are dynamically adjusted, and the Teacher deep segmentation network parameters are updated according to the updated Student deep segmentation network parameters using the dynamically adjusted exponential moving average coefficients. S6 uses the updated Student deep segmentation network to segment the target domain fundus image, and outputs the optic cup segmentation result and optic disc segmentation result of the target domain fundus image.
2. The passive domain adaptive fundus segmentation method based on evidence representation and marginal screening as described in claim 1, characterized in that: The specific process steps of S1 include: S11, Obtain a set of unlabeled target-domain color fundus images. and source domain supervised pre-trained segmentation model parameters , A color fundus photograph containing the optic nerve head region. Image sequence number The total number of fundus images in the target domain is defined, and the segmentation objects are divided into two foreground channels: the optic cup and the optic disc. Construct Student and Teacher deep segmentation networks with identical structures, and ensure that their initial parameters satisfy the following conditions: ; Where s represents the Student deep segmentation network and t represents the Teacher deep segmentation network. These represent the initial parameters of the Student deep segmentation network. This represents the initial parameters of the Teacher deep segmentation network; S12, perform weak enhancement transformation on each target domain fundus image. and strong enhancement transformation The system includes a Teacher deep segmentation network that takes a weakly enhanced image as input and a Student deep segmentation network that takes a strongly enhanced image as input. Both the Student and Teacher deep segmentation networks employ the same encoder-decoder deep segmentation structure. The deep segmentation structure includes a backbone feature extraction network, a multi-scale context aggregation module, and a decoding and reconstruction module. The Teacher deep segmentation network is used to generate pseudo-label supervision information for the target domain, and the Student deep segmentation network is used to perform parameter optimization.
3. The passive domain adaptive fundus segmentation method based on evidence representation and marginal screening as described in claim 2, characterized in that: The specific process steps of S2 include: S21, the weakly enhanced target domain fundus image The input is the Teacher deep segmentation network, which passes through the backbone feature extraction network, the multi-scale context aggregation module, and the decoding and reconstruction module to obtain the target domain feature map with segmentation semantic information. S22, at the network output end, set a positive evidence output branch and a negative evidence output branch for each segmented channel. For any pixel and any segmented channel, the positive evidence output branch provides the original response of the foreground positive evidence as follows: The negative evidence output branch provides background negative evidence and the original response is... ; S23. The original responses of the foreground positive evidence and the background negative evidence are processed by a non-negative activation function. By performing mapping, positive foreground evidence is obtained respectively. and negative evidence from the background ,Right now: Where e represents evidence and z represents the original response. This is a Softplus function or another activation function that guarantees a non-negative output; S24. After obtaining positive foreground evidence and negative background evidence, construct the Beta distribution parameters respectively. and : in, This parameter represents evidence that the current pixel belongs to the foreground. This parameter represents evidence that the current pixel belongs to the background.
4. The passive domain adaptive fundus segmentation method based on evidence representation and marginal screening as described in claim 3, characterized in that: The specific process steps of S3 include: S31, after obtaining the pixel-level Beta distribution parameters, calculate the average predicted probability that the current pixel belongs to the foreground, i.e., the foreground probability. Total strength of evidence They are respectively: Among them, the prospect probability The strength of evidence reflects the model's average tendency to classify the current pixel as belonging to the foreground. This reflects the total amount of evidence upon which the model made that judgment; S32, let the boundary of constructing evidence be... The evidentiary advantage of foreground assignments relative to background assignments is measured, and the marginality of the evidence is determined by the risk of foreground assignments. Risk of background assignment The difference construction, namely: in, The Digamma function indicates that when the evidence margin reaches a preset margin threshold, it means that the current pixel has a clearer foreground tendency under the comparison of positive and negative evidence and can be used as a reliable pseudo-label candidate. S33, during pseudo-label generation, all pixels are initially screened based on foreground probability to obtain initial foreground pseudo-labels. Indicates the preset probability threshold, then the initial foreground pseudo-label. for: in, The indicator function, based on the initial foreground pseudo-labels, retains only pixels that simultaneously satisfy the following conditions: foreground probability not lower than a preset probability threshold, evidence margin not lower than a preset margin threshold, and evidence strength not lower than a preset strength threshold, as reliable positive pseudo-labels. respectively , and Let probability threshold, marginal threshold, and intensity threshold represent the reliable positive / false labels, then they can be represented as: in, Indicates a reliable, genuine / counterfeit label; S34, For reliable positive and negative label pixels obtained through joint screening, assign supervisory weights based on their evidence margin and evidence strength. : in, This represents a weighted mapping function that monotonically increases with the margin and strength of evidence.
5. The passive domain adaptive fundus segmentation method based on evidence representation and marginal screening as described in claim 4, characterized in that: The specific process steps of S4 include: S41, for binary classification pixel labels The cross-entropy loss for binary classification is: in, For the true class label of the sample, the foreground probability Modeled using the Beta distribution, it satisfies: Calculating the expectation of the binary cross-entropy over the Beta distribution yields the Bayes risk-based segmentation loss under the Beta distribution: Where E represents the expected value, expanding the expression, we get: The Beta distribution satisfies the following expected property: Substituting the above properties back, we obtain a closed-form expression for pixel-level Beta loss: When the pixel label is foreground, the loss encourages the addition of positive evidence and suppresses negative evidence; when the pixel label is background, the loss encourages the addition of negative evidence and suppresses positive evidence. S42, let the Beta distribution parameters of the Student depth segmentation network output for the current pixel be denoted as... and The pseudo-labels generated by the Teacher deep segmentation network and jointly filtered are denoted as... Then weigh the evidence. Incorporating pixel-level Beta loss, we construct an overall weighted segmentation loss: in, This represents the overall weighted loss. To prevent extremely small constants with a denominator of zero; S43, in addition to the weighted segmentation loss, introduces the Kullback-Leibler regularization loss. and the loss constrained by the strength of evidence Together, they constitute the target loss function: in, This represents the Kullback-Leibler regularized loss. Indicates the loss constrained by the strength of evidence. and The weighting coefficients are used; the Kullback-Leibler regularization loss is used to constrain the current Beta distribution to approach the prior distribution, and the evidence strength constraint loss is used to suppress... Overconfident predictions caused by excessively large projections; S44. Backpropagation and parameter updates are performed on the Student deep segmentation network using the target loss function, enabling the Student network to segment the optic cup and optic disc on the unlabeled image in the target domain.
6. The passive domain adaptive fundus segmentation method based on evidence representation and marginal screening as described in claim 5, characterized in that: The specific process steps of S5 include: S51 introduces a class-balanced reweighting coefficient on the background pseudo-label pixels of the visual cup channel to reduce the impact of the imbalance between the number of foreground and background samples on parameter updates. S52, after completing the current round update of the Student network, the difference in prediction probabilities between the Student and Teacher networks is weighted and statistically analyzed based on the evidence weights corresponding to the reliable positive and negative labels, to obtain the degree of difference: in, The evidence weight corresponding to the reliable positive and negative label pixels, This represents the predicted probability of the Student deep segmentation network. This represents the predicted probability of the Teacher deep segmentation network; the dissimilarity is used to reflect the degree of consistency between the teacher and student networks in the high-reliability region. S53, Based on the aforementioned degree of difference, adaptively determine the dynamic exponential moving average coefficient. Its expression is: in, and Let represent the upper and lower bounds of the EMA coefficient, respectively, and satisfy . ; S54, after obtaining the dynamic EMA coefficients, update the Teacher depth segmentation network parameters: Where r represents the exponential moving average coefficient; S55, after completing the Teacher network update, uses the new Teacher network to regenerate the target domain pseudo-labels in the next training round, thus forming a closed-loop iterative mechanism.
7. The passive domain adaptive fundus segmentation method based on evidence representation and marginal screening as described in claim 6, characterized in that: The specific process steps of S6 include: S61, after completing the target domain adaptation training, the target domain fundus image is input into the updated Student depth segmentation network for forward inference to obtain the Beta distribution parameters corresponding to the optic cup channel and optic disc channel, and the foreground probability maps of the optic cup and optic disc are calculated respectively: in, This represents the probability diagram of the viewing cup channel. This represents the probability diagram of the viewing channel; S62, generate corresponding binary segmentation results based on the probability map of the viewing cup channel and the probability map of the viewing disk channel, so as to... and The threshold for segmentation of the optic cup and optic disc is represented by: in, This represents the result of the sight cup segmentation. This indicates the result of the visual disc segmentation.