OCT blood vessel stratification intelligent segmentation system and method thereof

Abstract

The application discloses an OCT blood vessel layering intelligent segmentation system and method, and relates to the technical field of medical image processing.The OCT blood vessel imaging image is acquired and spliced with an initialized feedback correction map, a double-channel segmentation model is input to obtain a first probability map set; a first certainty map is calculated according to the set, and a first segmentation mask is generated; end point sets are obtained through skeletonization and recognition processing, and correction values are calculated based on the end points to update the feedback correction map; the updated feedback correction map is spliced with the first segmentation mask, and second forward propagation, mean value processing and binary processing are performed to obtain a second segmentation mask; whether the Dice similarity coefficient of the second segmentation mask and the first segmentation mask is greater than a preset threshold value is judged, if yes, the second segmentation mask is output, otherwise, the feedback correction map is updated according to the correction value and iteration is continued until convergence or the maximum iteration number is reached; and the application can improve the accuracy and topological coherence of the OCT blood vessel image segmentation result.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to an OCT blood vessel layering intelligent segmentation system and method. Background Technology

[0002] Optical coherence tomography (OCTA) is a non-invasive, high-resolution retinal vascular imaging technique that can non-invasively visualize the microvascular networks of different retinal layers. By accurately segmenting and quantitatively analyzing the vascular networks of different anatomical layers (such as superficial and deep vascular plexuses) in OCTA images, clinicians can effectively diagnose and monitor various retinal vascular diseases, such as diabetic retinopathy and age-related macular degeneration. Therefore, developing automated and high-precision intelligent segmentation methods for OCT vascular layers is of great significance for improving the efficiency and objectivity of clinical diagnosis.

[0003] Currently, mainstream OCT blood vessel layering intelligent segmentation methods mainly rely on deep learning models or traditional image processing algorithms. These methods typically follow the following paradigm: first, a forward inference or optimization is performed on the OCT image using a pre-trained model or algorithm to directly output the final segmentation result; subsequently, some post-processing operations may be added to optimize the visual quality of the segmentation mask.

[0004] However, such methods share a common limitation: their segmentation process is unidirectional and open-loop, meaning that the model or algorithm lacks the ability to self-evaluate and iteratively optimize during the testing and inference phase. Specifically, when dealing with vascular topological structure fractures caused by image noise, signal attenuation, or pathological changes, existing technical solutions often directly output defective segmentation results or can only perform limited repairs through simple post-processing. They cannot fundamentally guarantee the connectivity and physiological rationality of the vascular network, resulting in insufficient reliability of the segmentation results in terms of topological coherence.

[0005] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an OCT vascular layering intelligent segmentation system and method.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows:

[0008] In a first aspect, this invention discloses an OCT blood vessel layering intelligent segmentation method, comprising the following steps:

[0009] Obtain the OCT angiography image to be segmented;

[0010] The OCT angiography image is stitched together with the feedback correction image initialized to zero in the channel dimension and input into the pre-trained dual-channel segmentation model for multiple forward propagations to obtain the first probability map set.

[0011] The first deterministic graph is obtained by calculating the standard deviation based on the first set of probability graphs;

[0012] The first segmentation mask is obtained by mean normalization and binarization of the first probability map set;

[0013] Based on the first segmentation mask, the endpoint set in the skeleton map is obtained through skeletonization and recognition processing;

[0014] Based on the average uncertainty value of each endpoint in the local region corresponding to the endpoint in the first uncertainty map and the Euclidean distance from the endpoint to the blood vessel pixel outside its connected domain, the correction value of each endpoint is calculated, and the feedback correction map is updated according to the correction values ​​of all endpoints.

[0015] The second segmentation mask is obtained by stitching the OCT angiography image with the updated feedback correction image, performing forward propagation, meanization, and binarization;

[0016] Determine whether the Dice similarity coefficient between the second segmentation mask and the first segmentation mask is greater than a preset similarity threshold: if yes, output the second segmentation mask; otherwise, update the feedback correction map according to the correction value of the endpoint corresponding to the second segmentation mask, and repeat the splicing input until the convergence condition is met or the maximum number of iterations is reached.

[0017] Secondly, this invention discloses an OCT vascular layering intelligent segmentation system, comprising:

[0018] The data acquisition module is used to acquire the OCT angiography images to be segmented;

[0019] The stitching input module is used to stitch the OCT angiography image with the feedback correction image initialized to zero in the channel dimension, and input it into the pre-trained dual-channel segmentation model for multiple forward propagations to obtain the first probability map set.

[0020] The first segmentation module is used to calculate the standard deviation based on the first probability map set to obtain a first deterministic map; and to perform mean and binarization processing based on the first probability map set to obtain a first segmentation mask.

[0021] The feedback correction map update module is used to obtain the endpoint set in the skeleton map through skeletonization and recognition processing based on the first segmentation mask; calculate the correction value of each endpoint based on the average uncertainty value of the corresponding local region in the first uncertainty map of each endpoint in the endpoint set and the Euclidean distance from the endpoint to the blood vessel pixel outside its connected domain; and update the feedback correction map according to the correction values ​​of all endpoints.

[0022] The second segmentation module is used to stitch the OCT angiography image with the updated feedback correction image, perform forward propagation, meanization, and binarization to obtain the second segmentation mask.

[0023] The iterative control module is used to determine whether the Dice similarity coefficient between the second segmentation mask and the first segmentation mask is greater than a preset similarity threshold: if yes, the second segmentation mask is output; otherwise, the feedback correction map is updated according to the correction value of the endpoint corresponding to the second segmentation mask, and the splicing input is repeated until the convergence condition is met or the maximum number of iterations is reached.

[0024] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0025] 1. By introducing a feedback correction mechanism and multiple iterative optimizations, a closed-loop segmentation system was constructed. This system can dynamically evaluate the uncertainty of each segmentation result and perform precise correction for ruptured blood vessels, effectively solving the problem that traditional unidirectional reasoning methods cannot self-optimize, thereby significantly improving the connectivity and physiological rationality of the vascular network;

[0026] 2. Based on the vascular topology and local uncertainty, dynamic updates of the correction value are introduced, which can adaptively adjust the feedback correction intensity according to the local environment of the endpoint and the vascular width, effectively avoiding over-correction or under-correction, improving the detail processing capability of vascular segmentation, and showing better adaptability, especially in complex vascular networks.

[0027] 3. By introducing connectivity constraints based on vascular topology, potential breakpoints are identified and virtual connections are established to effectively repair vascular breaks across connected regions, enhancing the physiological rationality of the segmentation results. This method can prioritize the repair of vascular break regions, ensuring that the segmentation results can truly reflect the topological structure of retinal vessels. Attached Figure Description

[0028] 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, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is an overall block diagram of the method in Embodiment 1 of the present invention;

[0030] Figure 2 This is a flowchart of the method according to Embodiment 1 of the present invention;

[0031] Figure 3 This is an overall block diagram of the system in Embodiment 2 of the present invention. Detailed Implementation

[0032] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] Application Overview: In the traditional optical coherence tomography (OCT) vascular imaging segmentation process, deep learning models directly output a segmentation mask after a single forward inference, lacking a dynamic verification and iterative optimization mechanism for the prediction results. This one-way processing mode prevents the system from effectively identifying and correcting vascular network topological breaks caused by image noise, signal attenuation, or pathological features, thus impairing the physiological rationality of vascular connectivity domains in the segmentation mask. Specifically, this manifests as non-physiological gaps or isolated branches between adjacent vascular segments. This lack of topological coherence directly affects the accuracy of vascular density calculation and perfusion zone analysis, causing deviations in subsequent quantitative parameter extraction.

[0034] For example, in the superficial vascular plexus segmentation task of patients with diabetic retinopathy, when the preprocessed OCTA image is output as the initial segmentation result by the U-Net architecture, two breaks appear in the capillary network in the superior temporal quadrant due to microaneurysms. Traditional methods directly perform thresholding on the probability map to generate a binary mask, failing to detect the uncertainty distribution characteristics of the broken areas. This results in subsequent morphological closing operations only filling localized micro-holes, failing to reconstruct vascular connections that conform to the anatomical structure. This error is transmitted to the hemodynamic analysis module through vascular skeleton extraction, incorrectly calculating the area of ​​the perfusion-deficient region and generating a diagnostic report that deviates from the true pathological state.

[0035] If the above problems are not addressed, topological errors in the segmentation mask will lead to systematic distortion of quantitative parameters of the vascular network, such as underestimation of vascular length density and overestimation of the number of branch points. In longitudinal disease monitoring scenarios, such errors will mask the spatial progression pattern of microvascular degeneration in patients with diabetic retinopathy, resulting in misdiagnosis of disease staging. More seriously, topological breaks may cause automated diagnostic systems to misidentify key pathological signs such as neovascularization, directly affecting the timeliness and accuracy of treatment decisions.

[0036] Faced with the aforementioned problems, this application first recognizes that a one-way inference mechanism cannot dynamically correct defects in vascular topological fractures, necessitating the establishment of a closed-loop feedback system. Traditional methods rely on a single forward propagation to generate segmentation results, lacking the ability to quantitatively assess and iteratively optimize prediction uncertainties. To address this, this application attempts to reconstruct the segmentation process into a multi-stage iterative mechanism, constructing a dynamic correction system by introducing uncertainty quantification indicators and spatial distance features. Further analysis reveals that vascular fracture regions often exhibit high variance in prediction results across multiple inferences, and the fracture endpoints have spatial proximity relationships with adjacent vessels. Based on this, this application proposes generating a deterministic graph through a set of probabilistic graphs to quantify the model's prediction confidence, and constructing a feedback correction graph by combining endpoint spatial distance features, thus forming a closed-loop optimization system that can dynamically adjust the segmentation results.

[0037] Example 1:

[0038] like Figure 1-2 As shown, the OCT blood vessel layering intelligent segmentation method includes the following steps:

[0039] Obtain the OCT angiography image to be segmented;

[0040] The OCT angiography image and the zero-initialized feedback correction image are concatenated along the channel dimension and input into a pre-trained dual-channel segmentation model for multiple forward propagations to obtain the first probability map set. The feedback correction image is an auxiliary image generated through multiple iterations to correct the segmentation results. Specifically, it can be an initialized zero matrix updated with correction values, dynamically adjusting the model input to optimize subsequent segmentation accuracy. The dual-channel segmentation model is a deep learning model capable of processing both the original image and the feedback correction image simultaneously. It typically uses a U-Net architecture, with the first channel weights initialized using ImageNet pre-training and the second channel weights initialized using a zero-mean Gaussian distribution. The number of forward propagations is usually set to 8-12 to balance computational efficiency and uncertainty estimation accuracy. Its function is to fuse image information and correction information to improve the model's ability to perceive vascular structures.

[0041] The first deterministic map is obtained by calculating the standard deviation based on the first probability map set; this step predicts reliability using a statistical discrete quantization model of the probability map set, and the calculation formula is as follows:

[0042]

[0043] in, This represents the uncertainty value at pixel (i,j). Let be the probability value of the k-th forward propagation. Let be the probability mean of N propagations. Higher uncertainty values ​​(e.g., >0.3) usually correspond to blood vessel boundaries or noise regions.

[0044] The first segmentation mask is obtained by mean averaging and binarization based on the first set of probability graphs; this step generates the initial segmentation result through probability fusion and decision-making, and the mean averaging is calculated using an arithmetic mean.

[0045]

[0046] The binarization threshold is fixed at 0.5, and the generated segmentation mask is a binary matrix in which 1 represents blood vessel pixels and 0 represents background.

[0047] The endpoint set in the skeleton map is obtained through skeletonization and recognition processing based on the first segmentation mask. The skeletonization process to obtain the endpoint set refers to the process of extracting the vessel centerline and locating its end points from the binary segmentation result. Specifically, it can be implemented by combining morphological thinning algorithm with endpoint detection algorithm. Its role is to capture potential breakage locations in the vessel topology to generate correction targets. The correction value calculation refers to the process of generating feedback intensity based on the uncertainty measure and spatial distance information of the local region of the endpoint. Specifically, it can be implemented by a linear combination of weighted average uncertainty value and Euclidean distance. Its role is to quantify the contribution of different endpoints to the segmentation error in order to accurately adjust the feedback correction map.

[0048] Based on the average uncertainty value of the local region corresponding to each endpoint in the first uncertainty map and the Euclidean distance from the endpoint to the blood vessel pixel outside its connected domain, the correction value of each endpoint is calculated, and the feedback correction map is updated according to the correction values ​​of all endpoints.

[0049] The second segmentation mask is obtained by stitching the OCT angiography image with the updated feedback correction image, inputting it forward, averaging, and binarizing it.

[0050] Determine whether the Dice similarity coefficient between the second segmentation mask and the first segmentation mask is greater than a preset similarity threshold: if yes, output the second segmentation mask; otherwise, update the feedback correction map according to the correction value of the endpoint corresponding to the second segmentation mask, and repeat the splicing input until the convergence condition is met or the maximum number of iterations is reached. The Dice similarity coefficient is an index that measures the degree of overlap between the segmentation results of two iterations. Specifically, it can be achieved by calculating the ratio of the intersection and union of the two masks. Its function is to determine whether the segmentation result tends to be stable in order to determine the iteration termination condition.

[0051] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0052] First, acquire the OCT angiography image to be segmented, with a resolution of 512×512 pixels. Initialize a feedback correction image with the same size as the OCT image, setting all initial values ​​to 0. Then, concatenate the OCT image and the feedback correction image along the channel dimension to obtain a 512×512×2 dual-channel image.

[0053] Using a pre-trained U-Net as a dual-channel segmentation model, the stitched image undergoes 10 Monte Carlo Dropout forward propagations to obtain 10 probability maps, forming the first probability map set. The pixel-level standard deviation of these 10 probability maps is calculated to obtain the first deterministic map. Simultaneously, the mean of the 10 probability maps is taken, and then binarized using a threshold of 0.5 to obtain the first segmentation mask.

[0054] The Zhang-Suen thinning algorithm is applied to the first segmentation mask for skeletonization, and endpoints are identified through 8-neighborhood connectivity analysis to obtain the endpoint set in the skeleton graph. For each endpoint in the endpoint set, a 7×7 local window is defined on the first deterministic graph, and the average uncertainty value of the pixels within the window is calculated. Simultaneously, the Euclidean distance from the endpoint to the nearest blood vessel pixel outside its connected domain is calculated.

[0055] Based on the two features mentioned above, a correction value is calculated for each endpoint. When the Euclidean distance is less than a preset threshold (e.g., 10 pixels), the correction value is positive, indicating a tendency towards connection breaks; when the Euclidean distance is greater than the threshold and uncertainty is high, the correction value is negative, indicating a tendency to remove isolated points. The feedback correction map is updated based on the correction values ​​of all endpoints.

[0056] The original OCT image is stitched together with the updated feedback correction image again, and then input into the U-Net model for forward propagation, mean normalization, and binarization to obtain the second segmentation mask. The Dice coefficient between the second segmentation mask and the first segmentation mask is calculated. If it is greater than 0.99, the iteration stops and the second segmentation mask is output; otherwise, the above process is repeated based on the second segmentation mask until the convergence condition is met or the maximum number of iterations (e.g., 20 times) is reached.

[0057] The above-described scheme effectively addresses the problem of vascular network topological breaks in traditional OCT vascular layering intelligent segmentation methods. By introducing a feedback correction mechanism and multiple iterative optimizations, this method can dynamically assess the uncertainty of segmentation results and specifically correct broken vascular regions. This closed-loop optimization strategy significantly improves the topological coherence of the segmentation mask, making the connectivity of the vascular network more consistent with physiological structures. Compared to traditional unidirectional reasoning methods, the scheme in this application can more accurately reconstruct complete vascular networks, reducing erroneous breaks caused by image noise or pathological changes, thus providing a more reliable basis for subsequent vascular density calculation and perfusion zone analysis. This improvement is of great significance for improving the diagnostic accuracy of fundus diseases such as diabetic retinopathy, and can help clinicians more accurately assess patient disease progression and treatment outcomes.

[0058] In some of the solutions described above in this application, when updating the feedback correction map, it is necessary to determine the correction direction and intensity. However, existing methods lack precise basis for determining whether the endpoints need correction and how to quantify the correction value, resulting in unstable correction effects. This may not be able to effectively repair vascular rupture or cause artifacts due to overcorrection.

[0059] This application further proposes a method for updating the feedback correction map by calculating the minimum Euclidean distance at the endpoints and the average uncertainty value of the neighborhood.

[0060] The minimum Euclidean distance is determined by calculating the minimum Euclidean distance between the endpoint and a blood vessel pixel that is not part of its own connected component. This distance reflects the proximity of the endpoint to its adjacent connected components. The neighborhood window size can be set to 5×5 or 7×7, with the specific value adjusted according to the image resolution. For example, a 7×7 window is used in a 512×512 image. The average uncertainty value is obtained by calculating the arithmetic mean of all pixels within the neighborhood window on the first uncertainty map. This value reflects the reliability of the local region segmentation result. When updating the correction value, positive, negative, or zero correction values ​​are generated by combining the minimum Euclidean distance and the average uncertainty value. For example, a positive correction is applied when the minimum distance is less than 2 pixels and the average uncertainty value is higher than 0.7.

[0061] Specifically, when neighboring blood vessel pixels exist outside the connected component to which the endpoint belongs, the minimum Euclidean distance quantifies the proximity of the endpoint to the external blood vessel. The average uncertainty value within the neighborhood window assesses the instability of the segmentation result for that region. By combining the two, a positive correction value is activated when the endpoint is close to the external blood vessel and the local uncertainty is high, prompting the model to expand the current connected component in subsequent iterations; a negative correction value suppresses erroneous expansion when the endpoint is far from the external blood vessel but the local uncertainty is high. For example, when the minimum distance to the endpoint is 1 pixel and the average uncertainty is 0.8, a positive correction value of +0.6 is applied; when the minimum distance is 4 pixels and the average uncertainty is 0.85, a negative correction value of -0.4 is applied. This process precisely controls the correction direction and intensity through quantitative indicators, avoiding the subjectivity of manual parameter tuning and improving the topological coherence of the segmentation results.

[0062] Through the above technical solution, this application can effectively utilize local information of endpoints to update the feedback correction map. By calculating the minimum Euclidean distance, the spatial relationship between endpoints and other vascular structures can be evaluated. By calculating the average uncertainty value, the segmentation reliability of the region surrounding the endpoints can be evaluated. Combining these two indicators to calculate the correction value can more accurately guide the segmentation optimization process in subsequent iterations, improving the connectivity and topological rationality of the vascular network. This feedback mechanism based on local features makes the segmentation process adaptive, enabling targeted optimization for the characteristics of different regions, thereby improving the overall accuracy and reliability of segmentation.

[0063] In some of the solutions described above in this application, a mechanism for calculating the correction value based on the minimum Euclidean distance of the endpoints and the average uncertainty value is proposed. However, when determining the sign and magnitude of the correction value, there is a problem that it cannot dynamically adapt to different fracture scenarios, which may lead to the correction process incorrectly amplifying noise interference or inhibiting real vascular connections.

[0064] This application further proposes a calculation process for the correction value, including: if the minimum Euclidean distance is less than a preset distance threshold, the correction value is a positive correction value, and the calculation formula is as follows:

[0065]

[0066] If the minimum Euclidean distance is greater than or equal to a preset distance threshold, and the average uncertainty value is greater than a preset uncertainty threshold, then the correction value is a negative correction value, and its calculation formula is as follows:

[0067]

[0068] Otherwise, the correction value is zero;

[0069] in, Let p be the average uncertainty value. This is the minimum Euclidean distance. The preset uncertainty normalization factor, The preset distance threshold, This is a gain coefficient that can be dynamically adjusted based on the number of iterations k. It is a function for maximizing the value.

[0070] The calculation of positive correction values ​​uses a dual constraint of distance threshold and uncertainty value to ensure that connections are enhanced only when uncertainty is low and the connected components are nearby. The calculation of negative correction values ​​uses an uncertainty threshold to filter out high-noise regions and suppress erroneous connections. The gain coefficient decays exponentially with the number of iterations. For example, the initial gain coefficient is 0.5 and the decay rate is set to 0.1, so that early iterations focus on rapid correction and later iterations focus on fine-tuning.

[0071] Specifically, when the minimum Euclidean distance between an endpoint and another connected component is less than a preset threshold (e.g., 5 pixels), the magnitude of the positive correction value is determined by both local uncertainty and the remaining distance ratio. For example, if an endpoint It is 3 pixels. It is 5 pixels. It is 0.2. If the value is 1, then the positive correction value is 0.5 × (1 - 0.2) × (1 - 3 / 5) = 0.5 × 0.8 × 0.4 = 0.16. The feedback correction map superimposes positive values ​​in this region to promote vascular connectivity. When the distance exceeds a threshold and the average uncertainty is higher than the threshold (e.g., 0.7), the negative correction value is directly proportional to the uncertainty. If the gain is 0.8, then the negative correction value is -0.5 × 0.8 = -0.4, suppressing artifacts in high-noise regions. By dynamically adjusting the gain coefficient, for example, after 3 iterations, γ = 0.37, the correction strength is gradually reduced to avoid overshoot. This mechanism effectively distinguishes between fracture repair and noise suppression scenarios, improving the topological coherence of the segmentation mask.

[0072] Through the above technical solution, this application achieves accurate calculation of endpoint correction values ​​during OCT vascular layering intelligent segmentation. By introducing the concept of positive and negative correction values, combined with minimum Euclidean distance and average uncertainty value, potential errors in the segmentation results can be more accurately evaluated and corrected. This method can effectively improve the accuracy and robustness of segmentation, especially showing advantages when processing complex vascular structures and low-quality images. Furthermore, through dynamically adjusted gain coefficients, this method can adapt to the needs of different iteration stages, further optimizing the segmentation effect.

[0073] In some of the schemes described above in this application, the gain coefficient remains a fixed value during the iteration process. This may lead to overshoot due to excessive gain or slow convergence due to insufficient gain in multiple iterations, affecting the stability of the segmentation results and the iteration efficiency.

[0074] This application further proposes a dynamic adjustment process for the gain coefficient, including the initial gain coefficient, the decay rate, and the current iteration number. A mathematical relationship between the three is established through an exponential function, so that the gain coefficient decays exponentially with the increase of the iteration number.

[0075] The initial gain coefficient is used to set the initial correction intensity, the attenuation rate controls the rate of gain decrease, and the exponential function ensures that the gain monotonically decreases with the number of iterations. The initial gain coefficient and attenuation rate need to be pre-calibrated according to the image resolution and vessel morphology. For example, in an OCTA image with a resolution of 10 μm / pixel, the initial gain coefficient can be set to 0.5, and the attenuation rate to 0.1. The exponential function converts the number of iterations into an exponential attenuation factor; when the number of iterations is 5, the gain coefficient decreases to 60.7% of its initial value.

[0076] Specifically, a higher gain coefficient is used in the first iteration to achieve rapid correction. As the number of iterations increases, the gain coefficient is gradually reduced to avoid oscillations caused by overcorrection. For example, with an initial gain coefficient of 0.5 and a decay rate of 0.1, the gain coefficients for the first five iterations are 0.500, 0.452, 0.407, 0.368, and 0.332, respectively, forming a gradually decreasing correction strength. This dynamic adjustment process, through the mathematical properties of the exponential function, ensures effective correction of significant errors in the initial stage while automatically reducing the correction amplitude as iterations progress, achieving a smooth transition from coarse to fine adjustment. This adjustment method effectively balances convergence speed and stability, avoiding undercorrection or overshoot caused by a fixed gain.

[0077] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0078] The dynamic adjustment process of the gain coefficient includes:

[0079]

[0080] in, The initial gain coefficient, The preset attenuation rate is greater than zero. This represents the current iteration number.

[0081] Specifically, in practical applications, the initial gain coefficient It can be set to 0.1, the preset attenuation rate. It can be set to 0.05. The gain coefficient decreases exponentially as the number of iterations k increases. For example, when k=0, =0.1; when k=10, ≈0.061; when k=20, ≈0.037. Therefore, by dynamically adjusting the gain coefficient, a larger correction amplitude can be maintained in the early stage of the iteration, while the correction amplitude can be gradually reduced in the later stage of the iteration to achieve more stable convergence.

[0082] Through the above technical solution, this application can adaptively adjust the correction intensity during the iteration process, effectively balancing the convergence speed and stability of the segmentation results. In the early stage of iteration, a larger gain coefficient can quickly correct obvious segmentation errors; while in the later stage of iteration, a smaller gain coefficient can avoid oscillations caused by overcorrection. This dynamic adjustment mechanism improves the robustness of the segmentation algorithm, enabling it to better adapt to OCT vascular images of varying complexity, thereby obtaining more accurate and reliable segmentation results.

[0083] In some of the schemes described above in this application, the feedback correction map may introduce high-frequency noise due to abrupt changes in the correction value of local regions during the iterative update process, which may cause undesirable oscillations or divergence of the segmentation mask in subsequent iterations, affecting the convergence stability of the algorithm.

[0084] This application further proposes to perform a weighted summation between the updated feedback correction graph and the feedback correction graph of the current iteration, and use the summation result as the final feedback correction graph for the next iteration.

[0085] The weight coefficients of the weighted summation are dynamically adjusted according to the image resolution. For example, in low-resolution images, the current iterative feedback correction map is given a higher weight to suppress noise, while in high-resolution images, the weight of the updated correction map is increased to preserve details.

[0086] Specifically, by linearly superimposing the feedback correction map of the current iteration with the updated correction map, local overshoot or undershoot caused by abrupt changes in a single correction value can be effectively suppressed. During the weighted summation process, the feedback correction map of the current iteration retains the low-frequency components of historical correction information, while the updated correction map introduces the high-frequency correction components of the current iteration. The superposition of the two forms a smoothly transitioning correction field.

[0087] As a preferred embodiment, the solution of this application is implemented as follows: After obtaining the correction values ​​of all endpoints and updating the feedback correction map, smoothing is performed. The updated feedback correction map is weighted and summed with the feedback correction map of the current iteration, and the sum is used as the final feedback correction map for the next iteration. Specifically, an exponential weighted average method can be used for smoothing. For example, let the current iteration number be k, and the updated feedback correction map be... The feedback correction graph for the current iteration is as follows: The feedback correction graph after smoothing It can be represented as:

[0088]

[0089] in, This is a smoothing coefficient, with a value ranging from 0 to 1. The smaller the value, the stronger the smoothing effect, which helps to suppress drastic fluctuations in the feedback correction graph; A larger value indicates a higher proportion of updated feedback information, which is beneficial for quickly responding to new correction information. In practical applications, it can be dynamically adjusted according to specific image features and iteration patterns. The value is determined to balance stability and response speed.

[0090] Through the above technical solution, this application can effectively smooth the update process of the feedback correction map, reducing noise and instability introduced by a single iteration. This smoothing mechanism can suppress abrupt changes and outliers in the feedback correction map while preserving effective correction information, thereby improving the stability and convergence of the entire iterative process. Furthermore, by reasonably selecting the smoothing coefficient, a high update speed can be maintained in the early stage of iteration, while the smoothing effect is gradually enhanced in the later stage of iteration, achieving adaptive adjustment of the feedback correction process. Thus, the solution of this application can improve the robustness and generalization ability of the algorithm while ensuring segmentation accuracy, especially when processing OCT vascular images with high noise or complex vascular structures, exhibiting better segmentation performance.

[0091] This application further proposes a scheme to add local connectivity verification when the Dice similarity coefficient meets the standard. Specifically, it includes: extracting the skeleton graphs corresponding to the current and previous iteration segmentation masks, identifying the endpoint sets of all connected components in the two skeleton graphs; calculating the minimum Euclidean distance from each endpoint in the current endpoint set to the endpoint set in the previous iteration, taking the maximum value of all minimum distances as the maximum displacement value, and counting the number of endpoints in the two endpoint sets; if the maximum displacement value is less than or equal to the preset local displacement threshold and the number of endpoint sets is equal, then convergence is determined, otherwise the iteration continues.

[0092] The maximum displacement value is calculated by traversing each endpoint, and the endpoint positions of adjacent iterations are quickly matched by the nearest neighbor search algorithm to ensure computational efficiency. The endpoint count needs to exclude isolated skeleton branches with a length less than a preset threshold to avoid noise interference. The local displacement threshold is dynamically adjusted according to the image resolution. For example, in an OCTA image where each pixel corresponds to 5 micrometers, it is set to 3 pixels, which corresponds to an actual displacement of 15 micrometers.

[0093] Specifically, when the Dice similarity coefficient exceeds a preset threshold for the first time, the system automatically triggers a local connectivity verification process. First, the current segmentation mask is skeletonized, extracting the topological skeleton of all connected components and identifying endpoint coordinates. Simultaneously, the skeleton graph data generated in the previous iteration is loaded, and the same endpoint extraction operation is performed. Then, for each coordinate point in the current endpoint set, its Euclidean distance to all points in the historical endpoint set is calculated, and the minimum distance value is recorded. After traversal, the system selects the maximum displacement value among all endpoints and compares it with a preset threshold. Simultaneously, the system counts the total number of endpoints in the two sets; if the difference exceeds zero, it determines that a blood vessel has been severed or a new branch has been added. Only when the maximum displacement value is below the threshold and the number of endpoints is consistent does the system determine that the segmentation result has truly converged and output the final mask; otherwise, feedback correction and iterative optimization continue. This mechanism effectively avoids misjudgments caused by high global similarity but abrupt local structural changes, ensuring that the output result simultaneously meets the requirements of overall consistency and local connectivity.

[0094] Through the above technical solution, this application can effectively determine whether the segmentation result has reached a stable state, avoiding unnecessary iterative calculations. By setting a reasonable similarity threshold, the convergence speed and computational efficiency of the algorithm can be improved while ensuring segmentation accuracy. In addition, this solution provides a clear termination condition for the iteration process, preventing it from falling into infinite loops and enhancing the robustness and reliability of the algorithm.

[0095] In some of the above-mentioned schemes in this application, the correction is performed only based on the uncertainty value of the endpoints themselves and the Euclidean distance when updating the feedback correction map, but the potential connection requirements between endpoints in the case of rupture of blood vessel segments between different connected domains are not considered. This makes it difficult to effectively repair the vascular topological rupture caused by noise or signal attenuation during the iteration process, affecting the physiological rationality of the segmentation results.

[0096] This application further proposes introducing connectivity constraints based on vascular topology during the update of the feedback correction graph. Specifically, this includes: extracting the topological skeleton of all vascular connected domains in the first segmentation mask; establishing a graph structure based on the topological skeleton, where skeleton intersections and endpoints are graph nodes and skeleton branches are edges; calculating the branch order of each endpoint in the endpoint set on its respective skeleton branch, where the branch order is defined as the shortest number of hops from the endpoint to the main skeleton of its connected domain; identifying all endpoint pairs in the endpoint set that are located in different connected domains but are spatially adjacent, and if the Euclidean distance between a pair of endpoints is less than a preset connection threshold and the difference between their branch orders is less than the order tolerance, then they are determined to be a potential breakpoint pair, and a virtual connection is established for them; when calculating the correction value of any endpoint, if the endpoint has one or more virtual connections, then its minimum Euclidean distance is corrected to the minimum Euclidean distance among all virtual connection endpoints.

[0097] The topological skeleton is extracted using a morphological thinning algorithm. The endpoints and intersections of the skeleton branches are identified by traversing the neighborhood connection states of the skeleton pixels. The branch order is calculated using a breadth-first search algorithm, which traverses the branches layer by layer starting from the main skeleton node and records the number of hops. The establishment of virtual connections must simultaneously satisfy the conditions of spatial proximity and branch order similarity. The connection threshold is set to 3-5 pixels according to the image resolution, and the order tolerance is set to 1-2 hops. The minimum Euclidean distance is corrected by comparing the distances between all virtual connection endpoints and taking the minimum value as the effective distance parameter of the current endpoint.

[0098] Specifically, during the iterative optimization process, when endpoints belong to different connected domains but have potential breakpoints, virtual connections are established to incorporate the distance between originally isolated endpoints into the correction value calculation. For example, when two endpoints belong to the terminal branches of adjacent connected domains and the distance is less than the connection threshold, if the difference in their branch order is within the tolerance range, they are identified as a breakpoint pair. In this case, when calculating the correction value of any breakpoint, its minimum Euclidean distance is replaced by the minimum distance between it and all virtual connected endpoints of that breakpoint, thereby guiding the feedback correction map to prioritize strengthening the correction intensity of that region in subsequent iterations. This mechanism effectively solves the problem of vascular rupture caused by neglecting cross-connected domain connection requirements in traditional methods, and improves the physiological coherence of the segmentation results through topological constraints.

[0099] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0100] In the process of updating the feedback correction graph based on the endpoint set, connectivity constraints based on the vascular topology are introduced. First, the topological skeleton of all vascular connected domains in the first segmentation mask is extracted. A graph structure is built based on the topological skeleton, with skeleton intersections and endpoints as graph nodes and skeleton branches as edges.

[0101] Furthermore, calculate the branch order of each endpoint in the endpoint set on its corresponding skeleton branch. The branch order is defined as the shortest number of hops from that endpoint to the main skeleton of its connected domain. For example, for an endpoint, if it requires two branch nodes to reach the main skeleton, then its branch order is 2.

[0102] Therefore, all endpoint pairs located in different connected domains but spatially adjacent are identified in the endpoint set. The condition for determining whether an endpoint pair is a potential breakpoint pair is: the Euclidean distance between the endpoints is less than a preset connection threshold, and the difference in their branch orders is less than an order tolerance. As a preferred implementation, the connection threshold can be set to 10 pixels, and the order tolerance can be set to 2. If the above conditions are met, the endpoint pair is determined to be a potential breakpoint pair, and a virtual connection is established for it.

[0103] Specifically, when calculating the correction value for any endpoint, if that endpoint has one or more virtual connections, its minimum Euclidean distance is corrected to the minimum Euclidean distance among all virtual connection endpoints. This correction mechanism guides the feedback correction map to preferentially connect broken vessel segments during iteration.

[0104] Through the above technical solution, this application introduces connectivity constraints based on vascular topology, effectively improving the connectivity and physiological rationality of vascular segmentation results. By establishing a graph structure and calculating the branch order, the method can identify potential vascular breakpoints and prioritize the repair of these breakpoints through a virtual connection mechanism. This topology-based constraint avoids erroneous connections that may be caused by simple Euclidean distance, improving the accuracy of the segmentation results. Simultaneously, by prioritizing the repair of breakpoints, the method can better maintain the overall connectivity of the vascular network, thus providing segmentation results that are more consistent with physiological structures. This is of great significance for subsequent clinical analysis and diagnosis.

[0105] In some of the schemes described above in this application, when the Dice similarity coefficient is greater than the similarity threshold, the overall shape of the segmentation mask may have reached a stable state, but the blood vessel endpoints in local areas may still have slight displacements or changes in number, which may lead to potential inconsistencies in the topological coherence of the segmentation results.

[0106] This application further proposes a step of adding local connectivity verification when the Dice similarity coefficient meets the convergence condition: extract the skeleton graphs corresponding to the current and previous iteration segmentation masks, and identify the endpoint sets of all connected components in the two skeleton graphs respectively; calculate the minimum Euclidean distance from each endpoint in the current endpoint set to the endpoint set of the previous iteration, take the maximum value of all minimum distances as the maximum displacement value, and count the number of endpoints in the two endpoint sets; if the maximum displacement value is less than or equal to the preset local displacement threshold and the number of endpoint sets is equal, then convergence is determined, otherwise the iteration continues.

[0107] The skeleton image extraction is achieved through a morphological skeletonization algorithm, and the endpoint set identification is based on the neighborhood connectivity of the pixels at the ends of the skeleton branches. The maximum displacement value is calculated using Euclidean distance, and the local displacement threshold is dynamically set according to the image resolution, for example, 2-3 pixel units. The endpoint count is completed through connected component labeling and endpoint counting to ensure that the number of endpoints remains stable for the segmentation results of the same blood vessel segment during the iteration process.

[0108] Specifically, after generating a segmentation mask in each iteration, the vascular topology skeleton of the current and previous results is extracted through skeletonization, and the endpoints of all skeleton branches are identified. For each endpoint in the current endpoint set, all endpoints in the previous endpoint set are traversed, their spatial distances are calculated, and the minimum value is recorded. Finally, the maximum distance among all endpoints is selected as the displacement evaluation index. Simultaneously, the total number of endpoints in the two sets is counted. If the difference in the number indicates the presence of newly added or disappeared vascular branches, the topology is considered unstable. By combining the maximum displacement value and the number of endpoints for dual verification, false convergence caused by local small displacements or branch breaks is avoided. For example, when the maximum displacement value is 1 pixel and the number of endpoints is consistent, convergence is determined; if the maximum displacement value is 4 pixels or the number of endpoints decreases, iterative optimization continues. Thus, the local connectivity and global topological stability of the segmentation result are simultaneously verified, improving the reliability of convergence judgment.

[0109] By employing the aforementioned technical solution, this application effectively avoids the problem of false convergence caused by displacement or changes in the number of local vessel endpoints despite meeting the Dice coefficient standard, ensuring the topological stability of the segmentation results upon iteration termination. This solution, through a dual verification mechanism, maintains global shape consistency while ensuring the local coherence of vessel branch terminals, thereby improving the physiological rationality and clinical usability of the vascular network segmentation results.

[0110] This application further proposes adjusting the correction value by incorporating the local width of the blood vessel at the endpoint location during the process of updating the feedback correction map based on the endpoint set, specifically including:

[0111] For each endpoint in the endpoint set, its local vessel width is calculated based on the first segmentation mask. The local vessel width is defined as the half-width value of the skeleton branch where the endpoint is located, obtained through distance transformation or skeleton distance mapping.

[0112] When calculating the correction value for the endpoint, the correction value is adjusted based on the local blood vessel width at that endpoint:

[0113] When the width of the local blood vessel is less than the preset width threshold, the magnitude of the correction value is increased, and the increase is inversely proportional to the width of the local blood vessel.

[0114] When the width of the local blood vessel is greater than or equal to the preset width threshold, the magnitude of the correction value is reduced, and the reduction magnitude is proportional to the width of the local blood vessel.

[0115] As a preferred embodiment, the solution of this application is implemented as follows: In the process of calculating the endpoint correction value, the segmentation mask is first processed by the distance transformation method to extract the half-width value of the skeleton branch where each endpoint is located as the local blood vessel width. For endpoints with a width less than a preset threshold, the correction value is amplified by an inverse proportional coefficient. For example, when the local blood vessel width is 3 pixels and the preset threshold is 5 pixels, the correction value is amplified to 1.67 times the original value. For endpoints with a width exceeding the preset threshold, the correction value is reduced by a linear ratio. For example, when the local blood vessel width is 8 pixels, the correction value is reduced to 0.8 times the original value. This adjustment process introduces vascular morphological features, making the correction intensity of narrow blood vessel areas significantly higher than that of wide blood vessel areas.

[0116] Through the above technical solution, this application achieves adaptive correction of local morphological features of blood vessels, generating stronger correction signals in narrow and ruptured vascular regions to promote connectivity restoration, while avoiding artifacts caused by overcorrection in wide vascular regions. This mechanism effectively solves the correction bias caused by neglecting morphological differences in complex vascular networks by traditional methods, and improves the segmentation consistency of vascular structures at different scales.

[0117] In some of the solutions described above in this application, the segmentation mask may contain isolated pixels or hole regions caused by noise or image acquisition errors. These defects can affect the accuracy of subsequent vascular connectivity analysis results.

[0118] This application further proposes morphological post-processing of the first segmentation mask: applying morphological opening operations to remove noisy pixels and applying morphological closing operations to fill holes, wherein the opening operation uses a first structuring element and the closing operation uses a second structuring element, and the structuring elements are all rectangles with sizes that adapt to the image resolution, to obtain an optimized segmentation mask.

[0119] The morphological opening operation uses a first structuring element to erode the segmentation mask, followed by a dilation operation to eliminate isolated noise points or small protrusions. The morphological closing operation uses a second structuring element to dilate the segmentation mask, followed by an erosion operation to fill holes or fracture areas. The sizes of the first and second structuring elements are dynamically adjusted according to the image resolution. Specifically, the physical size is converted to pixel size by calculating the number of pixels per millimeter in the image. For example, when the image resolution is 10 micrometers / pixel, the structuring element size is set to 3×3 pixels to match the width range of blood vessels. The rectangular shape of the structuring element maintains the straight-line characteristics of the blood vessel edges, avoiding edge blurring caused by circular structuring elements.

[0120] Specifically, after the segmentation mask is generated, an opening operation is first applied, using erosion to remove isolated noise points with areas smaller than the structuring element, followed by dilation to restore the morphology of the main blood vessel region. Then, a closing operation is applied, using dilation to connect adjacent blood vessel segments and fill holes, followed by erosion to restore the original blood vessel size. Adaptive adjustment of the structuring element ensures that the processing effect matches the image resolution; for example, smaller structuring elements are used for high-resolution images to preserve details, while larger structuring elements are used for low-resolution images to enhance denoising capabilities. Through these operations, noise pixels in the segmentation mask are effectively suppressed, blood vessel connectivity is enhanced, and distortion of the blood vessel structure caused by excessive smoothing is avoided.

[0121] Through the above technical solutions, this application effectively eliminates isolated pixels caused by image noise or model misjudgment in the segmentation results, while filling in the voids inside blood vessels caused by signal loss. This improves the visual coherence of the segmentation mask while maintaining the integrity of the blood vessel's main morphology. The adaptive adjustment mechanism for the size of structural elements in morphological operations avoids over-erosion or under-filling problems caused by manually preset parameters, ensuring a balance between preserving blood vessel details and suppressing noise in images of different resolutions, and providing segmentation results with complete topological structure for subsequent quantitative analysis.

[0122] Example 2:

[0123] like Figure 3 As shown, the OCT vascular layering intelligent segmentation system includes:

[0124] The data acquisition module is used to acquire the OCT angiography images to be segmented;

[0125] The stitching input module is used to stitch the OCT angiography image with the feedback correction image initialized to zero in the channel dimension, and input it into the pre-trained dual-channel segmentation model for multiple forward propagations to obtain the first probability map set.

[0126] The first segmentation module is used to calculate the standard deviation based on the first probability map set to obtain a first deterministic map; and to perform mean and binarization processing based on the first probability map set to obtain a first segmentation mask.

[0127] The feedback correction map update module is used to obtain the endpoint set in the skeleton map through skeletonization and recognition processing based on the first segmentation mask; calculate the correction value of each endpoint based on the average uncertainty value of the corresponding local region in the first uncertainty map of each endpoint in the endpoint set and the Euclidean distance from the endpoint to the blood vessel pixel outside its connected domain; and update the feedback correction map according to the correction values ​​of all endpoints.

[0128] The second segmentation module is used to stitch the OCT angiography image with the updated feedback correction image, perform forward propagation, meanization, and binarization to obtain the second segmentation mask.

[0129] The iterative control module is used to determine whether the Dice similarity coefficient between the second segmentation mask and the first segmentation mask is greater than a preset similarity threshold: if yes, the second segmentation mask is output; otherwise, the feedback correction map is updated according to the correction value of the endpoint corresponding to the second segmentation mask, and the splicing input is repeated until the convergence condition is met or the maximum number of iterations is reached.

[0130] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

[0131] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0132] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. An OCT-based intelligent segmentation method for blood vessel layering, characterized in that, Includes the following steps: Obtain the OCT angiography image to be segmented; The OCT angiography image is stitched together with the feedback correction image initialized to zero in the channel dimension and input into the pre-trained dual-channel segmentation model for multiple forward propagations to obtain the first probability map set. The first deterministic graph is obtained by calculating the standard deviation based on the first set of probability graphs; The first segmentation mask is obtained by mean normalization and binarization of the first probability map set; Based on the first segmentation mask, the endpoint set in the skeleton map is obtained through skeletonization and recognition processing; Based on the average uncertainty value of each endpoint in the local region corresponding to the endpoint in the first uncertainty map and the Euclidean distance from the endpoint to the blood vessel pixel outside its connected domain, the correction value of each endpoint is calculated, and the feedback correction map is updated according to the correction values ​​of all endpoints. The second segmentation mask is obtained by stitching the OCT angiography image with the updated feedback correction image, performing forward propagation, meanization, and binarization; Determine whether the Dice similarity coefficient between the second segmentation mask and the first segmentation mask is greater than a preset similarity threshold: if yes, output the second segmentation mask; otherwise, update the feedback correction map according to the correction value of the endpoint corresponding to the second segmentation mask, and repeat the splicing input until the convergence condition is met or the maximum number of iterations is reached.

2. The OCT vascular layering intelligent segmentation method according to claim 1, characterized in that: The process of updating the feedback correction graph includes: For each pixel corresponding to an endpoint, calculate the Euclidean distance between it and all pixels in the first segmentation mask / second segmentation mask that do not belong to its own connected component, and determine the minimum value among them as the minimum Euclidean distance of the endpoint. Define a neighborhood window of size M×M centered on the endpoint pixel, and calculate the arithmetic mean of the corresponding values ​​of all pixels in the neighborhood window on the first uncertainty map, which is used as the average uncertainty value of the endpoint. The correction values ​​for the endpoints are calculated based on the minimum Euclidean distance and the average uncertainty value, and the feedback correction map is updated.

3. The OCT vascular layering intelligent segmentation method according to claim 2, characterized in that: The calculation process for the correction value includes: If the minimum Euclidean distance is less than a preset distance threshold, the correction value is a positive correction value, and its calculation formula is as follows: If the minimum Euclidean distance is greater than or equal to a preset distance threshold, and the average uncertainty value is greater than a preset uncertainty threshold, then the correction value is a negative correction value, and its calculation formula is as follows: Otherwise, the correction value is zero; in, Let p be the average uncertainty value. This is the minimum Euclidean distance. The preset uncertainty normalization factor, The preset distance threshold, This is a gain coefficient that can be dynamically adjusted based on the number of iterations k. It is a function for maximizing the value.

4. The OCT vascular layering intelligent segmentation method according to claim 3, characterized in that: The dynamic adjustment process of the gain coefficient includes: in, The initial gain coefficient, The preset attenuation rate is greater than zero. This represents the current iteration number.

5. The OCT vascular layering intelligent segmentation method according to claim 1, characterized in that: After obtaining the correction values ​​for all endpoints and updating the feedback correction map, smoothing processing is also required: The updated feedback correction graph is weighted and summed with the feedback correction graph of the current iteration, and the sum is used as the final feedback correction graph for the next iteration.

6. The OCT vascular layering intelligent segmentation method according to claim 1, characterized in that: The specific convergence condition is as follows: Calculate the Dice similarity coefficient between the segmentation mask obtained in the current iteration and the segmentation mask obtained in the previous iteration; Determine whether the Dice similarity coefficient is greater than the preset similarity threshold: if yes, the segmentation result is determined to have converged.

7. The OCT vascular layering intelligent segmentation method according to claim 1, characterized in that: The process of updating the feedback correction graph based on the endpoint set also includes introducing connectivity constraints based on the vascular topology: Extract the topological skeleton of all blood vessel connected domains in the first segmentation mask, and build a graph structure based on the topological skeleton, wherein the skeleton intersections and endpoints are used as graph nodes, and the skeleton branches are used as edges; Calculate the branch order of each endpoint in the endpoint set on its corresponding skeleton branch. The branch order is defined as the shortest number of hops from the endpoint to the main skeleton of its connected domain. Identify all pairs of endpoints in the endpoint set that are located in different connected domains but are spatially close. If the Euclidean distance between a pair of endpoints is less than a preset connection threshold and the difference between their branch orders is less than the order tolerance, then they are determined to be potential breakpoint pairs and a virtual connection is established for them. When calculating the correction value for any endpoint, if the endpoint has one or more virtual connections, its minimum Euclidean distance is corrected to the minimum Euclidean distance among all virtual connection endpoints, thereby guiding the feedback correction map to preferentially connect ruptured blood vessel segments during iteration.

8. The OCT vascular layering intelligent segmentation method according to claim 6, characterized in that: When the Dice similarity coefficient is greater than the similarity threshold, local connectivity verification is also added: Extract the skeleton graphs corresponding to the current and previous iteration segmentation masks, and identify the endpoint sets of all connected components in the two skeleton graphs respectively; Calculate the minimum Euclidean distance from each endpoint in the current endpoint set to the endpoint set of the previous iteration, take the maximum value of all minimum distances as the maximum displacement value, and count the number of endpoints in the two endpoint sets. If the maximum displacement value is less than or equal to the preset local displacement threshold and the number of endpoints is equal, then convergence is determined; otherwise, the iteration continues.

9. The OCT vascular layering intelligent segmentation method according to claim 1, characterized in that: In the process of updating the feedback correction map based on the endpoint set, the local width of the blood vessel at the endpoint location is also introduced to adjust the correction value, specifically including: For each endpoint in the endpoint set, its local vessel width is calculated based on the first segmentation mask. The local vessel width is defined as the half-width value of the skeleton branch where the endpoint is located, obtained through distance transformation or skeleton distance mapping. When calculating the correction value for the endpoint, the correction value is adjusted based on the local blood vessel width at that endpoint: When the width of the local blood vessel is less than the preset width threshold, the magnitude of the correction value is increased, and the increase is inversely proportional to the width of the local blood vessel. When the width of the local blood vessel is greater than or equal to the preset width threshold, the magnitude of the correction value is reduced, and the reduction magnitude is proportional to the width of the local blood vessel.

10. An OCT vascular layering intelligent segmentation system, characterized in that: Using the OCT vascular layering intelligent segmentation method as described in any one of claims 1-9, including: The data acquisition module is used to acquire the OCT angiography images to be segmented; The stitching input module is used to stitch the OCT angiography image with the feedback correction image initialized to zero in the channel dimension, and input it into the pre-trained dual-channel segmentation model for multiple forward propagations to obtain the first probability map set. The first segmentation module is used to calculate the standard deviation based on the first probability map set to obtain a first deterministic map; and to perform mean and binarization processing based on the first probability map set to obtain a first segmentation mask. The feedback correction map update module is used to obtain the endpoint set in the skeleton map through skeletonization and recognition processing based on the first segmentation mask; calculate the correction value of each endpoint based on the average uncertainty value of the corresponding local region in the first uncertainty map of each endpoint in the endpoint set and the Euclidean distance from the endpoint to the blood vessel pixel outside its connected domain; and update the feedback correction map according to the correction values ​​of all endpoints. The second segmentation module is used to stitch the OCT angiography image with the updated feedback correction image, perform forward propagation, meanization, and binarization to obtain the second segmentation mask. The iterative control module is used to determine whether the Dice similarity coefficient between the second segmentation mask and the first segmentation mask is greater than a preset similarity threshold: if yes, the second segmentation mask is output; otherwise, the feedback correction map is updated according to the correction value of the endpoint corresponding to the second segmentation mask, and the splicing input is repeated until the convergence condition is met or the maximum number of iterations is reached.

Images