A robust blood vessel segmentation method for watermarked fundus images

Abstract

The application relates to the field of artificial intelligence and discloses a robust blood vessel segmentation method for watermarked fundus images, which comprises the following steps: acquiring a clean fundus image sample, corresponding blood vessel labels and watermark protection parameters; processing the clean fundus image through a watermark-related disturbance generator to generate a corresponding watermark protection view; inputting the clean fundus image and the watermark protection view into a blood vessel segmentation backbone network sharing parameters respectively to output a clean prediction image and a watermark protection prediction image; training the network by relying on the clean fundus image and the blood vessel labels to obtain first-stage pre-training parameters; taking the parameters as initial values, combining a clean image segmentation loss, a watermark protection view segmentation loss and a structure stability loss before and after protection to update network parameters, completing robust fine-tuning and obtaining a robust blood vessel segmentation model. The application can improve the analyzability of the fundus image after watermark protection or watermark-related disturbance in a downstream blood vessel segmentation task.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology and relates to medical image processing, digital watermark security protection, deep learning robust training and computer-aided diagnosis in artificial intelligence integrated systems. In particular, it relates to a robust blood vessel segmentation method for watermark protection of fundus images. Background Technology

[0002] With the deep penetration of artificial intelligence technology into ophthalmological diagnosis and treatment, fundus images have become important imaging data in the diagnosis and treatment of diabetic retinopathy, glaucoma, hypertensive retinopathy, and cardiovascular and cerebrovascular risk assessment. The width, density, branching angle, tortuosity, central line continuity, and terminal branching morphology of retinal vessels can reflect the microcirculation status. Accurate vessel segmentation is a fundamental step in vessel width measurement, arteriovenous ratio estimation, lesion localization, image registration, and disease grading.

[0003] In telemedicine, inter-institutional consultations, cloud storage, and multi-center research, fundus images need to be frequently transmitted, copied, and reused, containing patient privacy, diagnostic evidence, and medical liability information. Digital watermarking technology can embed authentication information, source information, or integrity verification information into the image itself, providing technical support for the authenticity, traceability, copyright protection, and tamper-proof authentication of medical images. Existing medical image watermarking methods typically focus on evaluating indicators such as invisibility, robustness, embedding capacity, recovery rate, bit error rate, security, peak signal-to-noise ratio, and structural similarity.

[0004] However, in intelligent medical workflows, watermarked fundus images not only need to be recognized by the human eye or a watermark decoder, but also continue to be input into vessel segmentation, optic disc-cup segmentation, lesion detection, and disease grading models. Even if the watermarked image has minimal visual distortion, frequency domain coefficient adjustments, texture perturbations, compression block effects, residual noise, and geometric resampling can still alter the local gradients, vessel edges, and background statistical features that are sensitive to deep networks. Retinal vessels are characterized by sparse foreground, elongated shape, low contrast, and topological sensitivity; even weak perturbations can lead to peripheral vessel rupture, increased background false positives, boundary drift, or discontinuity of the centerline.

[0005] Existing robust vessel segmentation methods typically improve the model's adaptability to general imaging differences through ordinary data augmentation, domain generalization, adversarial training, or semi-supervised structural stabilization training. However, their perturbation designs do not necessarily originate from watermark embedding, compressed transmission, or security attack processes, nor do they explicitly utilize the constraint that the vascular anatomy of the same fundus image remains unchanged before and after watermark protection. Simply applying traditional augmentation training to protected fundus images makes it difficult to guarantee that the watermarked images can still be stably analyzed by downstream artificial intelligence models.

[0006] Therefore, there is an urgent need for a specialized training method for watermark-protected fundus images, expanding the evaluation of watermark protection in medical images from "whether the watermark is recoverable and whether the image is visually undamaged" to "whether the protected image can be reliably analyzed by downstream models." This method should utilize the pairwise relationship between clean and protected images, introducing watermark-related perturbations and predictive structural stability constraints during the training phase. This will enable the segmentation model to learn vascular structural representations that are stable against watermark embedding, compressed transmission, noise attacks, and spatial degradation. Summary of the Invention

[0007] The main objective of this invention is to provide a robust vascular segmentation method for watermark-protected fundus images, in order to solve the technical problems in the prior art where, although watermark protection of medical images can maintain high visual quality or watermark recovery rate, downstream fundus vascular segmentation models are prone to fine vessel breakage, background false detection, and topological discontinuity under frequency domain, compression, noise, and spatial degradation conditions.

[0008] Based on a first key aspect of the invention, a robust vessel segmentation method for watermark-protected fundus images is provided, the method comprising the following steps performed by a computer system:

[0009] Obtain clean fundus image samples and their vascular annotations, and obtain watermark protection parameters for generating watermark-protected views, wherein the vascular annotations are used to characterize retinal vascular regions in the same fundus image;

[0010] The clean fundus image is processed by a watermark-related perturbation generator to generate a watermark-protected view corresponding to the clean fundus image. The watermark-protected view shares the same blood vessel annotation as the clean fundus image.

[0011] The clean fundus image and the watermark protection view are used as paired inputs and fed into a vascular segmentation backbone network with shared parameters to obtain the clean prediction map and the watermark protection prediction map.

[0012] Based on the cleanliness prediction map and the blood vessel annotations, the cleanliness image segmentation loss is calculated. The blood vessel segmentation backbone network is trained using the cleanliness image segmentation loss to obtain the first-stage pre-training parameters. Using the first-stage pre-training parameters as initialization parameters, the cleanliness image segmentation loss, the watermark protection view segmentation loss, and the structural stability loss before and after protection are calculated based on the cleanliness prediction map, the watermark protection view segmentation loss, and the structural stability loss before and after protection, respectively. Robust fine-tuning is performed based on the cleanliness image segmentation loss, the watermark protection view segmentation loss, and the structural stability loss before and after protection to update the parameters of the blood vessel segmentation backbone network, thereby obtaining a robust blood vessel segmentation model.

[0013] In the above scheme, clean fundus images and corresponding watermark-protected views are used as paired inputs to share parameters for the vascular segmentation network. The network is first pre-trained with clean images, and then robust fine-tuned by combining clean segmentation loss, watermark-protected view segmentation loss and structural stability loss. This can improve the stability and accuracy of vascular segmentation of fundus images after watermark protection, compression, noise and other disturbances while maintaining the vascular segmentation performance of clean images.

[0014] Optionally, in the above method, the watermark-related perturbation generator performs at least three of the following five types of perturbations: Discrete Wavelet Transform (DWT), Discrete Cosine Transform (DCT), JPEG, noise, and spatial perturbation.

[0015] The blood vessel segmentation backbone network includes at least: an initial feature extraction unit for extracting local texture and blood vessel edge features from fundus images; a multi-scale encoding unit for hierarchical representation of main blood vessels, fine blood vessels, and background structures; a decoding and fusion unit for restoring spatial resolution and outputting pixel-level blood vessel probability maps; and a prediction output unit for obtaining a binary blood vessel mask based on a threshold or probability mapping.

[0016] The structural stability loss before and after protection is used to constrain the vascular prediction structure of the same fundus image to remain stable before and after watermark protection.

[0017] The above scheme employs at least three types of watermark-related perturbation processing on the image and uses a multi-scale feature extraction-based vascular segmentation backbone network. It relies on structural stability loss to constrain prediction consistency, thereby adapting to the image degradation characteristics of watermark protection scenarios, accurately extracting the features of main blood vessels and fine blood vessels, and ensuring the overall stability of the predicted vascular structure.

[0018] As a further preferred embodiment, the watermark protection view is generated by a watermark-related perturbation generator, including:

[0019] A discrete wavelet transform is performed on a clean fundus image or its feature map to obtain a low-frequency sub-band and a high-frequency sub-band. A watermark template or a perturbation template is superimposed in at least one high-frequency or mid-high-frequency sub-band, and a wavelet domain watermark protection view is reconstructed by inverse discrete wavelet transform.

[0020] The fundus image is divided into multiple local image blocks, DCT transformation is performed on the local image blocks, frequency domain perturbation template is superimposed on the mid-frequency coefficients or mid-high frequency coefficients, and DCT domain watermark protection view is obtained through inverse DCT transformation.

[0021] Further, at least one of the following operations is performed on the image: JPEG compression, Gaussian noise superposition, brightness adjustment, cropping, scaling, rotation, or local spatial displacement, to simulate image degradation in watermark embedding, compressed storage, remote transmission, and post-attack processing.

[0022] By superimposing perturbation templates in the discrete wavelet / cosine transform domain and combining JPEG compression, noise, and spatial transformation to simulate real watermark embedding and transmission degradation, fundus views consistent with clinical watermark protection procedures can be generated. This provides degradation samples that fit real-world applications for model training, improving the model's adaptability to real-world scenarios.

[0023] As a further preferred embodiment, the watermark protection view satisfies:

[0024] ;

[0025] in, This indicates a clean fundus image. This indicates a view with watermark protection. This indicates a watermark-related perturbation generator. Indicates watermark protection parameters;

[0026] The cleanliness prediction map and the watermark protection prediction map satisfy the following:

[0027]

[0028]

[0029] in, Represents a vessel segmentation backbone network with shared parameters. This represents a cleanliness prediction diagram. The watermark protection prediction diagram is shown; the structural stability loss before and after protection is one of the mean square error, norm distance, and vascular perception weighted distance.

[0030] The above mathematical formulas clarify the generation mapping relationship between the watermark protection view, the cleanliness prediction map, and the watermark protection prediction map, and limit the selectable types of structural stability loss before and after protection, providing a standardized mathematical basis for model calculation and ensuring that loss constraints can be stably implemented.

[0031] Optionally, the weighted distance for blood vessel sensing satisfies:

[0032] ;

[0033] in, Satisfying any of the following forms:

[0034] ;

[0035] or

[0036] ;

[0037] in, and These represent the image height and width, respectively. Indicates the first Line 1 The blood vessel annotation value at the column pixel, This represents the structural weights obtained from vessel annotation, predicted probability maps, vessel centerline maps, edge maps, or fine vessel weight maps. Foreground weight coefficient;

[0038] The vascular sensing weighted distance enables structural stability constraints to focus on the vascular foreground, terminal branches, and topologically continuous regions.

[0039] The above formula for weighted distance of blood vessel perception is constructed based on blood vessel annotation or structural weight. It increases the constraint weight of key areas such as blood vessel foreground and terminal branches, so that the structural stability constraint is accurately applied to the core blood vessel area, thereby improving the continuity and topological integrity of fine blood vessels.

[0040] As a further preferred embodiment, in the aforementioned method, the robust fine-tuning uses the clean fundus image and the watermark-protected view as paired inputs; the paired inputs and the constraints on structural stability loss before and after protection are only enabled during the training phase, and the inference phase after the model training is completed only retains the single-branch vessel segmentation backbone network, without adding additional watermark decoders, additional alignment branches, or paired input requirements.

[0041] The above approach uses clean fundus images and watermark-protected views as paired inputs for robust fine-tuning. Paired inputs and structural stability constraints are used only during the training phase, while a single-branch network is retained during the inference phase. This improves the model's robustness to watermark perturbations without increasing inference complexity or adding extra modules.

[0042] Optionally, the total loss function for robust fine-tuning includes at least three parts:

[0043] Clean image segmentation loss is used to constrain the pixel-level differences between the clean prediction map and the blood vessel annotation;

[0044] Watermark-protected view segmentation loss is used to constrain pixel-level differences between the watermark-protected prediction map and the same blood vessel annotation.

[0045] Protecting structural stability loss before and after protection is used to constrain structural differences between clean prediction maps and watermark protection prediction maps;

[0046] The total loss function satisfies:

[0047]

[0048] in, Indicates the loss from partitioning. This indicates the structural stability loss before and after protection, and Selected from one of mean squared error, norm 1 distance, and vascular sensing weighted distance; Supervision weights for watermark protection view segmentation. Weights for structural stability;

[0049] By adjusting This allows the robust blood vessel segmentation model to improve the stability of watermark-protected view analysis while avoiding excessive sacrifice of clean image segmentation performance.

[0050] The above constructs a total loss function consisting of clean image segmentation loss, watermark-protected view segmentation loss, and structural stability loss before and after protection. By adjusting the weights to balance the model performance, the segmentation stability of watermark-protected images is improved while avoiding sacrificing the segmentation effect of clean images.

[0051] Based on the second main aspect of the present invention, a robust blood vessel segmentation system for watermark-protected fundus images is provided for implementing the aforementioned method, including a data acquisition module, a watermark-protected view generation module, a pairwise segmentation module, a robust fine-tuning module, and an inference output module.

[0052] The data acquisition module is used to acquire clean fundus images, blood vessel annotations, and watermark protection parameters;

[0053] The watermark protection view generation module is used to perform at least three operations on the clean fundus image based on the watermark protection parameters: wavelet domain watermark perturbation, DCT domain perturbation, JPEG compression perturbation, noise perturbation, and spatial perturbation, to generate a watermark protection view that shares vascular annotations with the clean fundus image.

[0054] The paired segmentation module is used to input the clean fundus image and the watermark protection view into the vascular segmentation backbone network with shared parameters, and output the clean prediction map and the watermark protection prediction map.

[0055] The robust fine-tuning module is used to perform two-stage optimization of the blood vessel segmentation backbone network based on the clean image segmentation loss, the watermark-protected view segmentation loss, and the structural stability loss before and after protection.

[0056] The inference output module is used to receive clean fundus images or watermark-protected fundus images during the deployment phase and output a vascular probability map or a vascular binary mask.

[0057] Based on a third key aspect of the present invention, an electronic device is provided, comprising one or more processors;

[0058] Storage device for storing one or more programs;

[0059] When the one or more programs are executed by the one or more processors, the one or more processors implement the robust vessel segmentation method for watermark-protected fundus images as described above.

[0060] Based on a fourth key aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed, implements the aforementioned robust vessel segmentation method for watermark-protected fundus images.

[0061] Compared with existing technologies, this invention incorporates the downstream analyzability of watermarked medical images into the model training objective. It not only focuses on the visual fidelity of the image and the recoverability of the watermark, but also on the reliability of the protected image as input to the artificial intelligence model, thus better meeting the actual needs of remote ophthalmology and cloud-based diagnosis and treatment scenarios.

[0062] Secondly, this invention constructs a set of watermark-related perturbations based on the watermark protection process, rather than simply performing ordinary random enhancement. DWT, DCT, JPEG, noise, and spatial perturbations correspond to frequency domain watermark embedding, transform domain modulation, lossy transmission, noise attack, and geometric degradation, respectively, which is beneficial for the model to learn robust vascular features that match the actual processing chain of the protected medical image.

[0063] Furthermore, this invention leverages the fact that vascular annotations remain unchanged in the same fundus image under both clean and watermarked conditions to construct paired supervision between clean and protected views. By sharing parameters to constrain structural stability, vascular rupture and background false positives under strong disturbance conditions can be reduced.

[0064] This invention also introduces a vascular perception-weighted structural stability loss, which focuses the constraint on clinically high-value regions such as the vascular foreground, fine vessels, edges, and terminal branches, rather than uniformly constraining the entire image pixels. This effectively enhances the topological continuity of blood vessels and the integrity of fine vessels, significantly improves the clinical analyzability of the segmentation results, and provides reliable support for subsequent downstream tasks such as vascular diameter measurement, arteriovenous classification, and disease grading.

[0065] Finally, this invention employs a two-stage optimization and single-branch deployment strategy, using a lightweight deployment scheme that decouples training and inference. Paired inputs, structural stability constraints, and dynamic perturbation chains are only used during the training phase. The deployment phase does not require additional watermark decoders or paired inputs, nor does it increase the complexity of inference on the clinical or cloud end. Therefore, it is suitable for robust fine-tuning and engineering implementation of existing segmentation models.

[0066] This invention is designed for scenarios such as remote ophthalmic screening, cloud-based medical image archiving, multi-center scientific research collaboration, and secure transmission. Without changing the inference structure of the segmentation model, it improves the analyzability of fundus images protected by watermarks or watermark-related perturbations in downstream vascular segmentation tasks. Attached Figure Description

[0067] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, obtaining other drawings based on these drawings without creative effort still falls within the scope of the present invention.

[0068] Figure 1 The following is a flowchart illustrating the execution of a robust blood vessel segmentation method for watermark-protected fundus images according to one embodiment of the present invention.

[0069] Figure 2 A general framework diagram of the invention is shown in one embodiment. The diagram illustrates the relationship between clean fundus images, a perturbation generator, a shared segmentation backbone network, cleanliness prediction, perturbation prediction, segmentation loss, consistency loss, and the final optimization objective.

[0070] Figure 3 A schematic diagram of a two-stage training process for the method of the present invention in one embodiment is shown. This diagram illustrates the overall process of the first stage of clean pre-training and the second stage of pairwise view robust fine-tuning, serving to illustrate the training mechanism of the present invention.

[0071] Figure 4 A schematic diagram of representative fundus images from different datasets under various perturbation conditions is shown in one embodiment. The figure illustrates typical samples from the DRIVE, CHASE_DB1, and HRF datasets under Clean, DWT-S, JPEG-S, Noise-S, DCT-S, and Spatial-S conditions, illustrating the protected image degradation forms targeted by this invention.

[0072] Figure 5 A comparison of the response plot and prediction results of the method of the present invention and the clean training model in one embodiment is shown. This figure illustrates the correspondence between the model's response distribution, prediction results, and ground truth annotations under different perturbation conditions, illustrating the advantages of the present invention in preserving vascular areas of interest and fine structures.

[0073] Figure 6 A qualitative segmentation comparison diagram under a representative strong perturbation in one embodiment is shown. This diagram illustrates the difference in segmentation results between the method of the present invention and the clean training model on the U-Net and Attention U-Net backbones, and is used to illustrate the improvement effect of the present invention in terms of capillary continuity, background suppression, and terminal branch recovery. Detailed Implementation

[0074] The preferred embodiments of the present invention will be described in detail below to provide a clearer understanding of the purpose, features, and advantages of the invention. It should be understood that the following embodiments are not intended to limit the scope of the invention, but are merely illustrative of the essential spirit of the technical solution of the invention.

[0075] In the following description, certain specific details are set forth for the purpose of illustrating various disclosed embodiments in order to provide a thorough understanding of the various disclosed embodiments. However, those skilled in the art will recognize that embodiments may be practiced without one or more of these specific details. In other instances, well-known techniques associated with the invention may not have been shown or described in detail to avoid unnecessarily obscuring the description of the embodiments.

[0076] Throughout this specification, references to "an embodiment" or "an embodiment" indicate that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment. Therefore, the appearance of "in an embodiment" or "an embodiment" in various places throughout the specification does not necessarily refer to the same embodiment. Furthermore, a particular feature, structure, or characteristic may be combined in any manner in one or more embodiments.

[0077] The following is a description of the specific meanings of technical terms, English abbreviations, and formula parameters that may be used in this invention:

[0078] Clean fundus images: Original fundus images that have not been processed by watermark embedding, perturbation, compression, noise, or any other processing.

[0079] Watermark-protected view: A protected fundus image that corresponds to the clean fundus image and shares the same vascular annotation after being processed by the watermark-related perturbation generator.

[0080] Blood vessel annotation: A pixel-level mask representing the retinal vascular region in a fundus image, used to distinguish the foreground and background of blood vessels.

[0081] Paired input: Robust fine-tuning stage, simultaneously inputting a clean fundus image and a corresponding watermark-protected view combination.

[0082] Robust fine-tuning: The training process of the blood vessel segmentation network is optimized by using clean pre-training parameters as initialization and combining three composite losses.

[0083] Structural stability loss before and after protection: A loss function that constrains the structure of the predicted blood vessel map to remain consistent before and after watermark protection for the same fundus image.

[0084] Vascular perception weighted distance: A method for calculating structural stability loss that assigns higher constraint weights to key areas such as vascular foreground, capillaries, and terminal branches.

[0085] Blood vessel segmentation backbone network: A deep learning network for pixel-level blood vessel segmentation in fundus images, with feature extraction and segmentation output capabilities.

[0086] Capillary continuity: The characteristic of small blood vessels remaining connected and unbroken in the segmentation results.

[0087] Topologically continuous regions: key regions such as vascular branches and terminals that need to maintain an intact topological structure.

[0088] DWT Discrete Wavelet Transform: This refers to the operation of decomposing an image into low-frequency and high-frequency sub-bands through discrete wavelet transform, and then superimposing watermark templates or perturbation templates on the mid-to-high frequency or high-frequency sub-bands.

[0089] Discrete Cosine Transform (DCT): This refers to the operation of performing a discrete cosine transform on a local image block, modulating the coefficients in the intermediate frequency or intermediate-high frequency range, and obtaining a watermark-protected view through the inverse transform.

[0090] JPEG perturbation: refers to the process of simulating lossy compression storage and remote transmission through DCT coefficient quantization and dequantization, which may weaken high-frequency details and produce block effects.

[0091] Vascular perception structural stability loss: refers to the loss function that assigns higher weights to vascular foreground, vascular centerline, vascular edge, capillaries, or high uncertainty regions when calculating the difference between clean prediction maps and protected prediction maps.

[0092] RobustMean: refers to the average segmentation metric under multiple watermark-related strong perturbation conditions, which can be used to measure the overall analyzability of protected fundus images in downstream segmentation tasks.

[0093] clDice: A segmentation evaluation index for the topological continuity of tubular structures, used to measure the degree of connectivity preservation between the vascular centerline and slender structures.

[0094] like Figure 1 As shown, in a feasible embodiment, a robust vessel segmentation method for watermark-protected fundus images includes steps S110 to S140 executed by a computer system. The implementation of this invention requires a computer system equipped with a processor, storage device, and a deep learning training and medical image processing framework. This system can complete the entire process of fundus image and parameter acquisition, watermark-protected view generation, paired input network training, robust fine-tuning, and vessel segmentation inference output. This computer system can be deployed in hospital imaging workstations, ophthalmic screening servers, cloud training platforms, telemedicine platforms, or mobile assisted diagnostic devices.

[0095] Step S110, Data Acquisition: Acquire clean fundus image samples and their vascular annotations, and acquire watermark protection parameters for generating watermark protection views, wherein the vascular annotations are used to characterize retinal vascular regions in the same fundus image.

[0096] In practice, a clean fundus image is provided. and blood vessel labeling ,in This refers to fundus images that have not undergone watermark embedding or perturbation protection processing. This represents the actual blood vessel mask corresponding to the fundus image. If the fundus image size is... The number of channels is A clean fundus image can then be represented as:

[0097] ;

[0098] Blood vessel labeling can be represented as:

[0099] ;

[0100] in, Indicates the image height. Indicates the image width. This indicates the number of image channels. For color fundus images, The value is typically 3; for grayscale images or feature inputs. It can be 1 or any other number of channels. In the blood vessel labeling y, Indicates the first Line 1 The column pixels belong to the blood vessel region. This indicates that the pixel belongs to the background area.

[0101] Watermark protection parameters are also obtained during the training phase. Watermark protection parameters may include wavelet subband embedding strength, DCT coefficient perturbation strength, JPEG quality factor, noise standard deviation, spatial displacement amplitude, cropping ratio, scaling ratio, brightness variation range, contrast variation range, and random seed. These parameters can be derived from actual watermark protection processes or preset according to the hospital's image transmission, storage, and security assessment needs.

[0102] Step S120, perturbation view generation. The clean fundus image is processed by a watermark-related perturbation generator to generate a watermark-protected view corresponding to the clean fundus image. The watermark-protected view shares the same vascular annotation as the clean fundus image.

[0103] In practice, a watermark-related perturbation generator is used. Convert the clean fundus image x into a perturbed image. This perturbation image is used to simulate fundus images after watermark protection, compressed transmission, noise attacks, or spatial degradation. The relationship is expressed as:

[0104] ;

[0105] in, This indicates a perturbed image. This indicates a watermark-related perturbation generator. This indicates the disturbance parameter or watermark protection parameter.

[0106] The set of disturbances can be represented as:

[0107] ;

[0108] in, This indicates wavelet domain watermark perturbation. Represents the perturbation in the discrete cosine transform domain. This indicates JPEG compression perturbation. Indicates noise disturbance. This indicates spatial disturbance.

[0109] like Figure 4 As shown, the perturbation generator can generate corresponding clean fundus images, DWT-S, JPEG-S, Noise-S, DCT-S, and Spatial-S images on different fundus datasets such as DRIVE, CHASE_DB1, and HRF. Figure 4 In the dataset, A1-A6 correspond to clean fundus images, DWT-S, JPEG-S, Noise-S, DCT-S, and Spatial-S images, respectively, while B1-B3 correspond to DRIVE, CHASE_DB1, and HRF fundus datasets, respectively.

[0110] Among them, DRIVE, CHASE_DB1, and HRF are the three major standard fundus datasets in this field. Their specific meanings are as follows:

[0111] DRIVE: short for Digital Retinal Images for Vessel Extraction, is the most classic open standard dataset in the field of fundus vessel segmentation, used to evaluate the accuracy and robustness of retinal vessel extraction algorithms.

[0112] CHASE_DB1: The full name is Child Heart and Health Study in England Database 1. It is a commonly used dataset for evaluating fundus vessel segmentation, which contains clinically acquired retinal images and vessel annotations.

[0113] HRF stands for High Resolution Fundus, a high-resolution fundus image dataset suitable for validating the segmentation ability of algorithms on details such as fine blood vessels and terminal branches.

[0114] DWT-S, JPEG-S, Noise-S, DCT-S, and Spatial-S are standard intensity watermarked or degraded perturbation images. -S=Standard indicates the standard intensity perturbation set by this invention. All are watermark-protected fundus images after watermark-related processing.

[0115] DWT-S: Fundus image processed with standard intensity discrete wavelet transform (DWT) frequency domain watermarking / perturbation.

[0116] DCT-S: Fundus image processed with standard intensity discrete cosine transform (DCT) frequency domain watermarking / perturbation.

[0117] JPEG-S: Fundus images degraded by standard strength JPEG compression, then stored / transmitted compressed after simulating watermarking.

[0118] Noise-S: Fundus images superimposed with standard intensity Gaussian noise to simulate watermark embedding noise degradation.

[0119] Spatial-S: Fundus images processed by standard intensity spatial domain transformation (cropping, scaling, rotation, brightness / displacement adjustment).

[0120] Depend on Figure 4 It can be seen that different perturbations have different effects on fundus images: DWT-S and DCT-S mainly simulate frequency domain watermark embedding or transform domain modulation, JPEG-S mainly simulates detail loss caused by compressed transmission, Noise-S mainly simulates noise attack or acquisition interference, and Spatial-S mainly simulates spatial deformation, resampling or local structural perturbation.

[0121] Step S130, Paired Segmentation. The clean fundus image and the watermark protection view are used as paired inputs and fed into a shared parameter vascular segmentation backbone network to obtain the clean prediction map and the watermark protection prediction map.

[0122] In practice, clean fundus images will be used. and perturbation images Blood vessel segmentation backbone network with shared parameters input separately A cleanliness prediction map was obtained. And perturbation prediction map The relationship is expressed as:

[0123] ;

[0124] ;

[0125] in, The parameter is The vascular segmentation backbone network, This represents a cleanliness prediction diagram. This represents the perturbation prediction map. The vessel segmentation backbone network can be U-Net, Attention U-Net, MixerCSeg, or other medical image segmentation networks capable of outputting pixel-level vessel probability maps.

[0126] Shared parameter design ensures that clean and perturbed images are processed and compared within the same model space, avoiding prediction discrepancies caused by different network parameters. Because clean images... and perturbation images For the same fundus anatomical structure, the actual blood vessel labeling remains consistent; therefore, both share the same blood vessel labeling. .

[0127] Step S140, Robust Fine-tuning. Based on the cleanliness prediction map and the vessel annotations, calculate the cleanliness image segmentation loss. Use this loss to train the vessel segmentation backbone network, obtaining the first-stage pre-training parameters. Using these parameters as initialization parameters, calculate the cleanliness image segmentation loss, watermark protection view segmentation loss, and structural stability loss before and after protection based on the cleanliness prediction map, the watermark protection view segmentation loss, and the structural stability loss before and after protection, respectively. Perform robust fine-tuning based on these losses to update the parameters of the vessel segmentation backbone network, obtaining a robust vessel segmentation model.

[0128] In practice, the model is first pre-trained on clean images to obtain initial parameters. Subsequently, during the robust fine-tuning phase, with As initialization parameters, pairs of samples are constructed in each training batch. The algorithm is jointly optimized using clean image segmentation loss, perturbed image segmentation loss, and pre- and post-preservation structural stability constraints. After training, only the robust single-branch model is retained. Used for reasoning.

[0129] like Figure 2 As shown, the overall optimization objective of this invention can be expressed as:

[0130] ;

[0131] in, Represents the total loss function. This represents the segmentation loss on the clean image. Represents the segmentation loss on the perturbed image λ1 and λ2 represent the structural stability constraint between the clean prediction map and the perturbation prediction map, respectively, and represent the perturbation image segmentation loss weight and the structural stability constraint weight.

[0132] In one feasible implementation, the watermark-related perturbation generator consists of wavelet domain perturbation, DCT domain perturbation, JPEG compression perturbation, noise perturbation, and spatial perturbation. The generator can randomly select one perturbation in each training batch, or it can randomly combine multiple perturbations to form a composite perturbation chain to simulate the various degradations that fundus images may encounter during watermark embedding, secure transmission, compressed storage, and post-attack processing.

[0133] For DWT wavelet domain perturbation, the system performs one or more levels of discrete wavelet transform on clean fundus images to obtain low-frequency subband LL and high-frequency detail subbands LH, HL, and HH. The low-frequency subband mainly carries the overall brightness of the fundus image, the background of the optic disc and macular, the orientation of the main blood vessels, and large-scale anatomical structures; the high-frequency subband mainly carries the texture, edge, and fine vascular details in the horizontal, vertical, and diagonal directions.

[0134] DWT decomposition can be represented as:

[0135] ;

[0136] To simulate the impact of watermark embedding on frequency domain coefficients while avoiding excessive damage to low-frequency diagnostic information, a watermark template or perturbation template can be superimposed on at least one high-frequency or mid-high-frequency sub-band, with the modification amplitude of the low-frequency sub-band constrained to below a preset threshold. The high-frequency sub-band perturbation can be expressed as:

[0137] ;

[0138] in, Indicates the first The original wavelet subband coefficients, This represents the modified sub-band coefficients. This indicates a watermark template or a perturbation template. This represents the wavelet domain embedding intensity.

[0139] For DCT domain perturbation, the system divides the fundus image into multiple local image blocks, performs a discrete cosine transform on each image block to obtain frequency domain coefficients, and superimposes a frequency domain perturbation template onto the mid-frequency or mid-high-frequency coefficients. The DCT domain perturbation can be expressed as:

[0140] ;

[0141] in, Indicates position The original DCT coefficients at the location, This represents the DCT coefficients after perturbation. This represents a frequency domain perturbation template. This indicates the intensity of the DCT domain perturbation.

[0142] JPEG compression perturbation is used to simulate lossy compression degradation of fundus images during storage, transmission, or sharing. Its main effect is to attenuate high-frequency details and produce blockiness, easily causing blurring of fine blood vessel edges. The JPEG quantization process can be represented as:

[0143] ;

[0144] in, Represents the position in the quantization matrix Quantization step size at the point, This represents the DCT coefficients after quantization and dequantization.

[0145] Gaussian noise perturbation is used to simulate acquisition noise, transmission noise, or malicious noise attacks, and its expression is:

[0146] ;

[0147] in, This represents Gaussian noise, where N represents a Gaussian distribution. This represents the standard deviation of noise.

[0148] Spatial perturbations are used to simulate geometric changes such as resampling, local displacement, clipping, scaling, and rotation. Their expression is:

[0149] ;

[0150] in, and It represents the amount of local spatial displacement.

[0151] Figure 4 The effects of different perturbations on the appearance of fundus images are demonstrated. It can be seen that while DWT-S, JPEG-S, Noise-S, DCT-S, and Spatial-S do not necessarily completely destroy the visible fundus structures, they do alter local texture, frequency distribution, and the edges or spatial relationships of fine blood vessels. These changes can easily lead to instability in depth segmentation model predictions. Therefore, using these perturbations to construct training samples helps the model learn a more stable vascular representation for the watermark protection process.

[0152] In one feasible embodiment, such as Figure 3 As shown, the present invention preferably employs a two-stage training method, including a clean pre-training stage and a robust fine-tuning stage.

[0153] Phase 1 is the clean pre-training phase. In this phase, only clean fundus images (x) and blood vessel annotations (y) are used to train the blood vessel segmentation backbone network, allowing the model to first learn the main blood vessels, fine blood vessels, background texture, and boundary structures in normal fundus images. The optimization objective of Phase 1 can be expressed as:

[0154] ;

[0155] in, This represents the model parameters obtained after cleanroom pre-training.

[0156] Phase two is the robust fine-tuning phase. This phase focuses on... As initialization parameters, the clean image x and the perturbed image x' are simultaneously input into a vessel segmentation backbone network with shared parameters to obtain a clean prediction map. And perturbation prediction map The optimization objective of Phase Two can be expressed as follows: (The optimization is performed using a combination of clean image segmentation loss, perturbed image segmentation loss, and structural stability constraints before and after protection.)

[0157] ;

[0158] in, This represents the final model parameters obtained after robust fine-tuning. The subscript represents the expectation on the training samples and the perturbation generator. This indicates that the expectation is calculated over the joint distribution. This represents the total loss function.

[0159] like Figure 3 As shown, the first stage ensures the model has stable basic blood vessel segmentation capabilities, while the second stage further utilizes perturbed images and structural stability constraints to enhance the model's robustness in watermark protection scenarios. This training strategy avoids strong perturbations to the model from the early stages of training, which is beneficial for learning blood vessel structural features first, and then learning stable representations of the protected images.

[0160] Clean image With perturbation image These are two observational forms of the same fundus image before and after protection; the actual vascular anatomy of both should remain consistent. Therefore, during the training phase, not only is a clean prediction image required... And perturbation prediction map Each closely approximates the annotation of real blood vessels. Furthermore, both are required to maintain structural stability.

[0161] The stability constraint of the basic structure can be expressed as mean square error or norm distance. The mean square error structural stability loss can be expressed as:

[0162] ;

[0163] The structural stability loss of the first norm can be expressed as:

[0164] ;

[0165] in, and These represent the image height and width, respectively. The clean prediction map is shown in the first... Line 1 Predicted value at column pixel, The perturbation prediction map is shown in the first... Line 1 Predicted value at column pixel.

[0166] To avoid large areas of background pixels dominating structural stability terms, in one possible embodiment, the present invention further introduces a blood vessel sensing weight map ω. This weight map can be constructed based on artificial blood vessel annotations, blood vessel centerlines, edge regions, fine blood vessel regions, uncertain regions, or pseudo-labels. One optional form is:

[0167] ;

[0168] or:

[0169] ;

[0170] in, Indicates the first Line 1 Weights at column pixels, This represents the foreground weight coefficient. Indicates the blood vessel labeling value. This represents the structural weights obtained from the vessel centerline, edge map, or capillary weight map.

[0171] The loss of vascular sensing structural stability can be expressed as:

[0172] ;

[0173] By focusing structural stability constraints on the vessel foreground, terminal branches, and topology-sensitive regions, the model tends to maintain the continuity of capillaries and the topological stability of vessels when faced with watermark perturbations, JPEG compression, noise interference, and spatial degradation, rather than simply learning the consistency of the background region.

[0174] In a feasible embodiment, the segmentation supervision loss consists of a binary cross-entropy loss and a Dice loss. The binary cross-entropy loss provides stable pixel-level supervision, while the Dice loss alleviates the class imbalance problem caused by sparse foreground retinal vessels and a large background area. The segmentation supervision loss can be expressed as:

[0175] ;

[0176] in, This represents the binary cross-entropy loss. This indicates Dice's loss. The weights represent the binary cross-entropy loss. denoted by Dice loss weight, p is the probability value of blood vessels predicted pixel by pixel by the segmentation network (blood vessel segmentation backbone network) for the input image;

[0177] Dice loss can be expressed as:

[0178] ;

[0179] Here, ε is a smoothing term to prevent the denominator from being zero.

[0180] The total loss during the robust fine-tuning phase consists of three parts: the first term maintains segmentation capability on clean fundus images, the second term enhances segmentation capability on perturbed images, and the third term constrains and protects the stability of the predicted structure before and after the perturbation. The total loss function can be expressed as:

[0181] ;

[0182] in, Indicates the supervision weights for perturbed image segmentation. This represents the structural stability loss weight.

[0183] When using the vascular sensing structural stability loss, the total loss function can also be expressed as:

[0184] ;

[0185] Structural stability weights The selection should be based on both the clean image segmentation performance and the robustness performance of the perturbed images on the validation set. If If the value is too small, the model may degenerate into a simple perturbation enhancement; if... If the structural stability weights are too large, the model may overly force predictions of severely degraded images to be completely consistent with clean images, thereby suppressing its ability to adapt to the true degraded appearance. Therefore, this invention preferably employs moderate structural stability weights to achieve a balance between cleanliness performance and robustness after protection.

[0186] To verify the difference between the solution of the present invention and ordinary data augmentation, the following possible embodiments are provided.

[0187] Typical data augmentation training usually only optimizes the segmentation loss of the augmented image, that is, it requires that the augmented image T(x) can be correctly segmented. Its training objective can be expressed as:

[0188] ;

[0189] This approach does not explicitly constrain the predictive relationship before and after enhancement. For general image tasks, ordinary enhancement may be sufficient; however, for watermark-protected fundus images, the core assumption is that the same fundus image should have the same vascular anatomy before and after protection.

[0190] This invention further introduces a prediction difference constraint:

[0191] ;

[0192] This directly constrains the structural stability between cleanliness prediction and perturbation prediction. The robust training objective after incorporating structural stability constraints can be expressed as:

[0193] ;

[0194] in, The measure of difference between two prediction maps can be the mean squared error, norm distance, Dice consistency measure, or vascular perception weighted distance.

[0195] Therefore, this invention differs from ordinary randomized augmentation training and from semi-supervised consistency training without watermark protection. Its perturbation sources, training objectives, and deployment requirements all revolve around the downstream analyzability of medical images after watermark protection. While the watermarking algorithm is responsible for embedding, extracting, and verifying security information, this invention is responsible for training the downstream vascular segmentation model, enabling it to stably output clinically meaningful vascular structures even after the image has been protected, compressed, transmitted, or attacked.

[0196] In one embodiment, the present invention provides a robust vessel segmentation system for watermark-protected fundus images. The system includes a data acquisition module, a watermark-related perturbation generation module, a pairwise segmentation module, a robust fine-tuning module, and an inference output module.

[0197] The data acquisition module is used to acquire clean fundus images and vascular annotations from fundus cameras, medical image storage systems, telemedicine platforms, cloud data warehouses, or local training datasets. It can also read watermark protection parameters, perturbation intensity ranges, and training configurations.

[0198] The watermark-related perturbation generation module generates perturbed images based on watermark protection parameters. This module can call a real watermark embedder or simulate the watermark protection process using perturbation operators such as DWT, DCT, JPEG, Noise, and Spatial. The perturbed images output by this module, together with clean fundus images, form paired samples.

[0199] The pairwise segmentation module is used to input clean fundus images and perturbed images into a shared-parameter vascular segmentation backbone network, and output a clean prediction map. And perturbation prediction map .

[0200] The robust fine-tuning module is used to calculate the segmentation loss of clean images, the segmentation loss of perturbed images, and the structural stability loss before and after protection, and updates the model parameters based on the total loss function.

[0201] The inference output module is used to receive clean fundus images or watermarked fundus images during the deployment phase and output a vascular probability map or a vascular binary mask.

[0202] To illustrate the technical effects of this invention, in one possible embodiment, the system can be validated on publicly available fundus vessel segmentation datasets such as DRIVE, CHASE_DB1, and HRF, employing backbone networks such as U-Net, Attention U-Net, and MixerCSeg. Test conditions include clean fundus images, DWT-S, JPEG-S, Noise-S, DCT-S, and Spatial-S, where strong perturbation conditions are used to simulate watermark-related frequency domain perturbations, lossy transmission, noise attacks, and spatial degradation.

[0203] On the DRIVE, CHASE_DB1, and HRF datasets, the method of this invention demonstrates a higher Best Dice metric than the model trained using only clean images under various perturbation conditions. The improvement is more significant under strong perturbation conditions such as DWT-S, JPEG-S, Noise-S, and Spatial-S, indicating that this method can improve the downstream segmentation stability of protected fundus images.

[0204] In terms of evaluation metrics, in addition to the Dice score under Clean conditions, task-level metrics such as Best_Dice, Best_IoU, clDice, HD95, ASSD, and RobustMean can also be used. RobustMean represents the average of Best_Dice under multiple strong perturbation conditions, and can be expressed as:

[0205] ;

[0206] in, This represents a set of strong perturbations. Indicates the disturbance condition Best_Dice.

[0207] When the set of strong disturbances includes five types of disturbances: DWT-S, JPEG-S, Noise-S, DCT-S, and Spatial-S, it can be represented as:

[0208] ;

[0209] In one experimental example, this invention improved RobustMean under strong perturbation conditions on multiple datasets and backbone combinations, while the Dice of clean images showed minimal change. For high-resolution HRF datasets, which contain many fine blood vessels and high-frequency structures, watermarking, compression, and noise have a more significant impact on peripheral blood vessels. This invention can significantly improve vascular connectivity and topological stability under strong perturbation conditions.

[0210] Specifically, across eight datasets and backbone combinations, the method of this invention improves RobustMean by an average of approximately 2.93 percentage points under perturbation conditions, while reducing Dice for clean images by approximately 0.62 percentage points. On the HRF dataset U-Net backbone, RobustMean increases from 68.45% to 74.66%, and clDice-RM increases from 65.35% to 72.28%. These results demonstrate that this invention does not simply improve clean image performance, but rather provides a stable gain in the downstream analyzability of protected images.

[0211] The effectiveness of the present invention is further verified by response graph analysis in one of the following embodiments.

[0212] like Figure 5 As shown, under different strong perturbation conditions, comparing the clean training model and the model trained using the method of this invention, significant differences can be observed in their response maps and final prediction maps. Figure 5In the diagram, C1-C7 represent Input (the clean fundus image), Perturbed Input (watermark-protected view), Response Map (Clean-trained) (the feature response map output by the model trained only with clean images), Response Map (CWPC-FT) (the feature response map output by the model after robust fine-tuning with CWPC according to this invention), Prediction (Clean-trained) (the vessel segmentation prediction result of the clean-trained model), Prediction (CWPC-FT) (the vessel segmentation prediction result of the model after robust fine-tuning with this invention), and GT (the ground truth of vessel annotation), respectively. D1-D3 represent DRIVE / Noise-S (test conditions with standard intensity Gaussian noise perturbation applied on the DRIVE fundus vessel segmentation dataset), CHASE_DB1 / JPEG-S (test conditions with standard intensity JPEG compression perturbation applied on the CHASE_DB1 fundus vessel segmentation dataset), and HRF / DWT-S (test conditions with standard intensity discrete wavelet transform (DWT) watermark perturbation applied on the HRF fundus vessel segmentation dataset). Wherein: DRIVE / CHASE_DB1 / HRF are the three major publicly available standard datasets in the field of fundus vessel segmentation; Noise / JPEG / DWT are the core watermark-related perturbation types defined in this invention; -S is an abbreviation for Standard (standard intensity), representing the standard perturbation intensity adopted in this invention. Note: In all embodiments, CWPC is the code name for the method of this invention, and CWPC-FT is the robust fine-tuning training strategy of this invention.

[0213] For clean training models, under perturbation inputs such as Noise-S, JPEG-S and DWT-S, problems such as scattered response regions, incorrect activation of background regions, and insufficient response at the edge of blood vessels are prone to occur, resulting in the prediction results showing broken capillaries, local missing or false positive background.

[0214] In contrast, the model trained using the method of this invention has a response map that is more concentrated in the main blood vessels and small branch areas, and can still maintain a clear vascular response in low-contrast areas and areas with strong perturbation. Figure 5 The magnified local area further illustrates that the present invention can reduce peripheral blood vessel rupture and local prediction loss under perturbation conditions, making the prediction results closer to the real blood vessel annotation.

[0215] therefore, Figure 5 This not only demonstrates that the present invention can improve segmentation metrics, but also shows that it can enhance attention to stable vascular structures at the model response level and reduce the false activation of background textures and noise regions by watermark-related perturbations.

[0216] The following embodiments will qualitatively segment and compare the solutions of the present invention.

[0217] exist Figure 6 In the diagram, E1-E6 represent Image (the input fundus image, including a clean fundus image and a watermark-protected view after watermarking / perturbation), U-Net Clean (the vessel segmentation prediction result trained on a clean fundus image using a basic U-Net network without robust fine-tuning), U-Net CWPT-FT (the vessel segmentation prediction result optimized by the CWPT robust fine-tuning strategy of this invention using a basic U-Net network), Attu-Net Clean (the vessel segmentation prediction result optimized by the CWPT robust fine-tuning strategy of this invention using a basic U-Net network), Attu-Net CWPT-FT (the vessel segmentation prediction result optimized by the CWPT robust fine-tuning strategy of this invention using an attention-based U-Net network), and GT (Ground Truth, i.e., the gold standard annotation (truth mask) for retinal vessel segmentation). F1-F6 represent DRIVE DWT-S (DRIVE DWT-S, DRIVE Noise-S, CHASE_DB1 JPEG-S, CHASE_DB1 Spatial-S, HRF DWT-S, HRF Noise-S), DRIVE Noise-S (fundus test images on the DRIVE dataset with standard intensity Gaussian noise perturbation), CHASE_DB1 JPEG-S (fundus test images on the CHASE_DB1 dataset with standard intensity JPEG compression degradation), CHASE_DB1Spatial-S (fundus test images on the CHASE_DB1 dataset with standard intensity spatial domain transformation (cropping, scaling, rotation, translation, etc.) perturbation), HRF DWT-S (fundus test images on the HRF high-resolution fundus dataset with standard intensity discrete wavelet transform (DWT) frequency domain watermark perturbation), and HRF Noise-S (fundus test images on the HRF high-resolution fundus dataset with standard intensity Gaussian noise perturbation), respectively.

[0218] like Figure 6 As shown, on representative strongly disturbed samples, a qualitative comparison is made between the clean training model and the model trained using the method of this invention. Figure 6The differences between the input fundus image, the predictions from clean U-Net training, the predictions from U-Net using the method of this invention, the predictions from clean Attention U-Net training, the predictions from Attention U-Net using the method of this invention, and the ground truth annotations are shown. It can be seen that under conditions such as DRIVE / DWT-S, DRIVE / Noise-S, CHASE_DB1 / JPEG-S, CHASE_DB1 / Spatial-S, HRF / DWT-S, and HRF / Noise-S, the clean training model is more prone to the following problems:

[0219] One cause is the rupture of capillaries or terminal branches;

[0220] Second, there is an increase in false positives in areas with strong noise or spatial disturbance.

[0221] Third, the local vascular structure deviates from the actual annotation;

[0222] Fourth, small blood vessels are not adequately preserved in HRF high-resolution images.

[0223] The model trained using the method of this invention can output more continuous and stable vascular structures under the same perturbation conditions. Especially in the area marked by the red box, this invention can better preserve the direction of fine vessels, terminal branches, and local topology, reduce background false detections, and make the prediction results closer to the real vascular annotations.

[0224] therefore, Figure 6 The effectiveness of this invention in downstream segmentation tasks of watermark-protected fundus images was verified from a visual perspective, demonstrating that this invention not only improves overall performance but also enhances the continuity and clinical analyzability of actual vascular structures.

[0225] The deployment method of the present invention will be described in the following embodiments.

[0226] During the deployment phase, the trained robust vessel segmentation model receives only one input image. This input can be a clean fundus image without embedded watermarks, or a protected fundus image after watermark embedding, compressed transmission, noise contamination, or spatial processing. The model outputs a vessel probability map, which can be used to generate a binary vessel mask through thresholding, connected component filtering, or post-processing operations.

[0227] Since the paired inputs, structural stability constraints, and dynamic perturbation chains of this invention are only used in the training phase, the deployment phase does not require simultaneous input of clean and perturbed images, nor does it require additional watermark decoders, additional feature alignment modules, or additional backend branches. Therefore, this invention can be embedded into existing fundus image analysis systems with low engineering costs, improving segmentation reliability in remote medical and medical image secure sharing environments.

[0228] The program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0229] The acquisition, storage, and application of user personal information involved in the technical solution of this invention all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0230] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this invention does not impose any limitations on them.

[0231] The technical terms, principles, or means related to the technical solutions of the present invention mentioned in the above embodiments, which are not described in detail above, are all well-known technologies or common practices that are known to those skilled in the art.

[0232] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A robust blood vessel segmentation method for watermark-protected fundus images, characterized in that, The method includes the following steps performed by a computer system: Obtain clean fundus image samples and their vascular annotations, and obtain watermark protection parameters for generating watermark-protected views, wherein the vascular annotations are used to characterize retinal vascular regions in the same fundus image; The clean fundus image is processed by a watermark-related perturbation generator to generate a watermark-protected view corresponding to the clean fundus image. The watermark-protected view shares the same blood vessel annotation as the clean fundus image. The clean fundus image and the watermark protection view are used as paired inputs and fed into a vascular segmentation backbone network with shared parameters to obtain the clean prediction map and the watermark protection prediction map. Based on the clean image prediction map and the blood vessel annotation, the clean image segmentation loss is calculated. The blood vessel segmentation backbone network is trained using the clean image segmentation loss to obtain the first-stage pre-training parameters. Using the first-stage pre-training parameters as initialization parameters, the clean image segmentation loss, the watermark protection view segmentation loss, and the structural stability loss before and after protection are calculated based on the clean image prediction map, the watermark protection view segmentation loss, and the structural stability loss before and after protection, respectively. Robust fine-tuning is performed based on the clean image segmentation loss, the watermark protection view segmentation loss, and the structural stability loss before and after protection to update the parameters of the blood vessel segmentation backbone network, thereby obtaining a robust blood vessel segmentation model. The robust fine-tuning uses the clean fundus image and the watermark-protected view as paired inputs; the paired inputs and the constraints on structural stability loss before and after protection are only enabled during the training phase. After the model training is completed, the inference phase only retains the single-branch vessel segmentation backbone network and does not add additional watermark decoders, additional alignment branches or paired input requirements. The total loss function for robust fine-tuning comprises at least three parts: Clean image segmentation loss is used to constrain the pixel-level differences between the clean prediction map and the blood vessel annotation; Watermark-protected view segmentation loss is used to constrain pixel-level differences between the watermark-protected prediction map and the same blood vessel annotation. The structural stability loss before and after protection is used to constrain the structural differences between the clean prediction map and the watermark protection prediction map.

2. The robust blood vessel segmentation method for watermark-protected fundus images according to claim 1, characterized in that, The watermark-related perturbation generator performs at least three of the following five types of perturbations: Discrete Wavelet Transform (DWT), Discrete Cosine Transform (DCT), JPEG, noise, and spatial perturbations. The vascular segmentation backbone network includes at least: an initial feature extraction unit for extracting local texture and vascular edge features from fundus images; a multi-scale encoding unit for hierarchical representation of main vessels, fine vessels, and background structures; a decoding and fusion unit for restoring spatial resolution and outputting pixel-level vascular probability maps; and a prediction output unit for obtaining a binary vascular mask based on a threshold or probability mapping.

3. The robust blood vessel segmentation method for watermark-protected fundus images according to claim 2, characterized in that, The watermark protection view is generated by a watermark-related perturbation generator, including: A discrete wavelet transform is performed on a clean fundus image or its feature map to obtain a low-frequency sub-band and a high-frequency sub-band. A watermark template or a perturbation template is superimposed in at least one high-frequency or mid-high-frequency sub-band, and a wavelet domain watermark protection view is reconstructed by inverse discrete wavelet transform. The fundus image is divided into multiple local image blocks, DCT transformation is performed on the local image blocks, frequency domain perturbation template is superimposed on the mid-frequency coefficients or mid-high frequency coefficients, and DCT domain watermark protection view is obtained through inverse DCT transformation. Further, at least one of the following operations is performed on the image: JPEG compression, Gaussian noise superposition, brightness adjustment, cropping, scaling, rotation, or local spatial displacement, to simulate image degradation in watermark embedding, compressed storage, remote transmission, and post-attack processing.

4. The robust blood vessel segmentation method for watermark-protected fundus images according to claim 3, characterized in that, The watermark protection view satisfies: ； in, This indicates a clean fundus image. This indicates a view with watermark protection. This indicates the watermark-related perturbation generator. Indicates watermark protection parameters; The cleanliness prediction map and the watermark protection prediction map satisfy the following: in, Represents a vessel segmentation backbone network with shared parameters. This represents a cleanliness prediction diagram. The watermark protection prediction diagram is shown; the structural stability loss before and after protection is one of the mean square error, norm distance, and vascular perception weighted distance.

5. The robust blood vessel segmentation method for watermark-protected fundus images according to claim 4, characterized in that, The weighted distance for vascular sensing satisfies: ； in, Satisfying any of the following forms: ； or ； in, and These represent the image height and width, respectively. Indicates the first Line 1 The blood vessel annotation value at the column pixel, This represents the structural weights obtained from vessel annotation, predicted probability maps, vessel centerline maps, edge maps, or fine vessel weight maps. Foreground weight coefficient, The clean prediction diagram shows the first... Line 1 Predicted values ​​of column pixels, The first watermark protection prediction diagram shows the... Line 1 Predicted values ​​for column pixels; The vascular sensing weighted distance enables structural stability constraints to focus on the vascular foreground, terminal branches, and topologically continuous regions.

6. The robust blood vessel segmentation method for watermark-protected fundus images according to claim 1, characterized in that, The total loss function satisfies: in, Indicates the partition loss. This indicates the structural stability loss before and after protection, and Selected from mean squared error, norm 1 distance, and vascular sensing weighted distance one; Supervision weights for watermark protection view segmentation. Weights for structural stability; By adjusting This allows the robust blood vessel segmentation model to improve the stability of watermark-protected view analysis while avoiding excessive sacrifice of clean image segmentation performance.

7. A robust vessel segmentation system for watermark-protected fundus images to implement the method of claim 1, characterized in that, It includes a data acquisition module, a watermark-protected view generation module, a pairwise segmentation module, a robust fine-tuning module, and an inference output module; The data acquisition module is used to acquire clean fundus images, blood vessel annotations, and watermark protection parameters; The watermark protection view generation module is used to perform at least three operations on the clean fundus image based on the watermark protection parameters: wavelet domain watermark perturbation, DCT domain perturbation, JPEG compression perturbation, noise perturbation, and spatial perturbation, to generate a watermark protection view that shares vascular annotations with the clean fundus image. The paired segmentation module is used to input the clean fundus image and the watermark protection view into the vascular segmentation backbone network with shared parameters, and output the clean prediction map and the watermark protection prediction map. The robust fine-tuning module is used to perform two-stage optimization of the blood vessel segmentation backbone network based on the clean image segmentation loss, the watermark-protected view segmentation loss, and the structural stability loss before and after protection. The inference output module is used to receive clean fundus images or watermark-protected fundus images during the deployment phase and output a vascular probability map or a vascular binary mask.

8. An electronic device, characterized in that, Includes one or more processors; Storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the robust vessel segmentation method for watermark-protected fundus images as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed, the computer program implements the robust vessel segmentation method for watermark-protected fundus images as described in any one of claims 1-6.

Images