Method for monitoring retinal development in preterm infants based on vascular morphology and system therefor

Abstract

The application belongs to the technical field of medical image processing, and discloses a premature infant retina development monitoring method and system based on blood vessel morphology, which comprises six steps of image enhancement, blood vessel segmentation, morphological parameter calculation, development stage identification, longitudinal tracking and report generation, a blood vessel network description system is established by quantitatively analyzing blood vessel diameter, branch angle, bending degree and density, automatic development stage identification and longitudinal tracking are realized, and an objective basis is provided for premature infant visual development evaluation.

Description

Technical Field

[0001] This invention relates to the fields of medical image processing and ophthalmic diagnostic technology, specifically to a method and system for monitoring retinal development in premature infants based on vascular morphology. Background Technology

[0002] Retinopathy of prematurity (ROP) is an angiogenic eye disease affecting premature infants, and its pathogenesis is closely related to incomplete retinal vascular development. In normal full-term newborns, retinal vessels are largely fully developed at birth, extending from the optic disc to the vicinity of the ora serrata. However, premature infants, having prematurely transitioned from the hypoxic environment of the uterus to the relatively hyperoxygenous environment outside the body, exhibit large areas of avascularity in the peripheral retina. Subsequent vascular development is influenced by various factors, potentially leading to pathological changes such as delayed vascular development, abnormal vascular course, and even abnormal neovascularization. Therefore, systematic, objective, and quantitative monitoring of retinal vascular development in premature infants is of significant clinical importance, aiding in the early identification of developmental abnormalities and guiding clinical intervention decisions.

[0003] With advancements in perinatal medicine and neonatal intensive care, an increasing number of extremely low birth weight and premature infants are surviving, significantly increasing their risk of retinopathy of prematurity. Clinically, indirect ophthalmoscopy or fundus photography is commonly used to screen for retinopathy of prematurity, but the interpretation of results relies heavily on the ophthalmologist's subjective experience. This qualitative assessment method has limitations such as low inter-observer consistency, difficulty in accurately quantifying vascular morphological changes, and inability to objectively track developmental progress. However, with the development of digital fundus imaging and computer-aided diagnostic technologies, quantitative assessment of retinal vessels using image analysis methods has become possible.

[0004] Chinese patent CN110689526A discloses a method and system for retinal vessel segmentation based on retinal fundus images. This method constructs an overall network model based on the features of retinal fundus images. The overall network model includes N cascaded basic modules using an attention mechanism, each constructed based on the features of the retinal fundus images. Specifically, the scheme uses a simplified U-Net architecture as the basic modules, and cascades multiple basic modules to form the final S-UNet network. In the overall network model, the foreground features of the previous basic module are fed into the next basic module along with the original image, allowing subsequent basic modules to inherit the learning experience of previous modules, thereby accelerating the training process and effectively solving the data imbalance problem. This scheme achieves good segmentation performance on the DRIVE and CHASE_DB1 datasets.

[0005] However, the aforementioned comparative documents have the following technical defects, making them difficult to directly apply to objective monitoring scenarios of retinal development in premature infants:

[0006] First, this scheme only achieves binary segmentation of retinal vessels and does not involve quantitative calculation and analysis of vascular morphological parameters. Although it can obtain the segmentation results of vessels and mark the pixel assignments of vessels and background, it cannot further extract quantitative indicators such as vessel diameter distribution, branching angle, tortuosity, and vessel density. These morphological parameters are of great significance for assessing the developmental status and maturity of vascular networks, and the lack of quantitative analysis capabilities limits the application value of this method in clinical diagnosis.

[0007] Second, this scheme was designed and trained for adult retinal vessel segmentation, failing to fully consider the unique characteristics of premature infant retinal images. Premature infant retinal vessels possess unique features such as small diameter, low contrast between vessels and background, incomplete vascular networks, and large avascular areas surrounding them—characteristics significantly different from adult retinal vessels. Directly applying a segmentation model designed for adults to premature infant images may lead to problems such as missed detection of fine vessels, blurred segmentation boundaries, and increased false positive rates, affecting the accuracy of subsequent analysis.

[0008] Third, this scheme lacks the function of identifying and determining the stage of vascular development. Retinal vascularization in premature infants is a dynamic developmental process, gradually expanding outwards from the optic disc. Vascular networks at different developmental stages exhibit different topological structures and morphological characteristics. This comparative document does not establish a correspondence between vascular network characteristics and developmental stages, making it impossible to determine the current developmental stage of the retina based on the current vascular morphology, and thus failing to provide clinicians with an objective basis for assessing developmental progress.

[0009] Fourth, this method lacks longitudinal tracking and analysis capabilities. Premature infants typically require multiple fundus examinations to monitor retinal development, with intervals usually between 1 and 2 weeks. Fundus images acquired at different time points cannot be directly compared at the pixel level due to differences in imaging angle, magnification, and eye position. This comparison file does not provide image registration and temporal analysis functions, making it impossible to align images of the same infant at different times to a unified coordinate system, and thus difficult to quantify the dynamic growth changes of vascular networks and the formation process of neovascularization.

[0010] In summary, there is an urgent need for a technical solution that can objectively and quantitatively monitor the development of retinal vessels in preterm infants. This solution should have comprehensive calculation capabilities for vascular morphology parameters, optimized processing strategies for preterm infants, automatic identification of developmental stages, and longitudinal tracking and analysis capabilities, thereby providing technical support for the scientific formulation of visual system development assessment and clinical follow-up plans for preterm infants. Summary of the Invention

[0011] To address the aforementioned shortcomings of existing technologies, this invention provides a method and system for monitoring the retinal development of preterm infants based on vascular morphology. This method enables comprehensive quantitative morphological analysis of the retinal vascular network in preterm infants, automatically identifies the stages of vascular development, and monitors the dynamic changes in the vascular network through longitudinal tracking.

[0012] To achieve the above technical objectives, the present invention adopts the following technical solution:

[0013] A method for monitoring retinal development in preterm infants based on vascular morphology includes the following steps: Step S1, image acquisition and enhancement processing: A wide-angle fundus camera is used to acquire retinal images of preterm infants. Adaptive contrast enhancement processing is performed on the retinal images, and local histogram equalization is used to improve the visualization of low-contrast vascular regions, resulting in an enhanced image; Step S2, vascular network segmentation and extraction: The enhanced image is input into a pre-trained vascular segmentation network. The vascular segmentation network adopts an encoder-decoder architecture. The encoder extracts multi-scale vascular features through multi-layer convolution and pooling operations, and the decoder recovers vascular spatial details through upsampling and skip connections, outputting a binary segmentation mask for retinal vessels. Skeletonization processing is performed based on the binary segmentation mask to obtain the vascular network structure; Step S3, vascular morphology parameter calculation: Morphological analysis is performed on the vascular network structure to extract morphological feature parameters, including vascular diameter distribution, etc. Step S4: Developmental Stage Identification. Based on morphological feature parameters, the topological complexity index of the vascular network is calculated. The topological complexity index includes network connectivity, branch level depth, and vascular coverage. The topological complexity index is input into the developmental stage classification model, which outputs developmental stage labels based on the vascularization process characteristics. Step S5: Longitudinal Tracking Analysis. Image registration is performed on retinal images of the same child at different examination time points. Registration transformation parameters are obtained through feature point matching. Based on the registration transformation parameters, the vascular network structure at different stages is aligned to a unified coordinate system, and the longitudinal change characteristics of the vascular network are calculated. Step S6: Developmental Assessment Report Generation. Combining morphological feature parameters, developmental stage labels, and longitudinal change characteristics, an individualized retinal developmental assessment report is generated. The developmental assessment report includes vascularization progress status, network maturity index, and developmental trajectory prediction information.

[0014] This invention also provides a retinal development monitoring system for preterm infants based on vascular morphology, including an image acquisition and enhancement module, a vessel segmentation module, a morphological analysis module, a developmental stage identification module, a longitudinal tracking module, and a report generation module. The image acquisition and enhancement module acquires retinal images of preterm infants and performs adaptive contrast enhancement processing; the vessel segmentation module outputs a binary segmentation mask and vascular network structure for the retinal vessels; the morphological analysis module extracts morphological feature parameters such as vessel diameter distribution, branch angle, curvature, and vessel density; the developmental stage identification module calculates topological complexity indices and outputs developmental stage labels; the longitudinal tracking module performs image registration and calculates the longitudinal variation characteristics of the vascular network; and the report generation module integrates the analysis results to generate an individualized retinal development assessment report.

[0015] Compared with the prior art, the present invention has the following beneficial effects:

[0016] Firstly, by quantitatively calculating morphological parameters such as vessel diameter distribution, branching angle, curvature, and vessel density, an objective descriptive system for the retinal vascular network of preterm infants was established, overcoming the shortcomings of poor inter-observer consistency in traditional qualitative assessment methods.

[0017] Secondly, through vascular network topology complexity analysis and developmental stage classification model, the automatic identification of vascular development stages was achieved, and the correspondence between vascular network morphological characteristics and vascularization process was established, which can help clinicians judge the degree of retinal development maturity.

[0018] Third, through image registration based on feature point matching and thin-plate spline transformation technology, precise alignment of images of the same child at different examination time points was achieved, enabling quantitative monitoring of the dynamic growth changes of the vascular network, including the expansion of the vascular coverage area and the formation of new blood vessels.

[0019] Fourth, by comprehensively analyzing the results from multiple dimensions, an individualized developmental assessment report is generated, which includes network maturity indicators and developmental trajectory predictions, providing an objective basis for the scientific formulation of clinical follow-up plans and the timing of interventions. Attached Figure Description

[0020] Figure 1 This is a flowchart of the method for monitoring retinal development in premature infants based on vascular morphology, as described in this invention.

[0021] Figure 2 This is an architectural diagram of the preterm infant retinal development monitoring system based on vascular morphology, which is part of the present invention. Detailed Implementation

[0022] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0023] This invention provides a method for monitoring retinal development in preterm infants based on vascular morphology, such as... Figure 1 As shown, the method comprises six core steps, from S1 to S6. These steps form a deep data flow coupling and closed-loop collaborative relationship: the output of image enhancement in step S1 serves as a key input for vessel segmentation in step S2; the quality of the segmentation result in step S2 directly affects the accuracy of morphological parameter calculation in step S3; the morphological feature parameters from step S3 simultaneously flow to developmental stage identification in step S4 and report generation in step S6; the developmental stage labels from step S4 provide a stage comparison benchmark for longitudinal tracking in step S5 and are also output to step S6; and the longitudinal change features from step S5 are incorporated into step S6 to complete the final report generation. More importantly, the evaluation report generated in step S6 can provide feedback to optimize the processing parameters of preceding steps: when image quality issues are detected in the report, the contrast enhancement parameters of step S1 are automatically adjusted; when the confidence level of developmental stage identification is low, online learning and updating of the model in step S4 is triggered. This closed-loop collaborative mechanism generates a non-linear gain effect greater than the sum of its parts (1+1>2) between the modules, significantly improving the overall monitoring performance and clinical application value of the system.

[0024] Step S1: Image Acquisition and Enhancement. The technical objective of this step is to acquire high-quality images of the preterm infant's retina and improve the visualization of low-contrast vascular regions through adaptive enhancement processing, laying a solid data foundation for subsequent vascular segmentation. The acquisition and preprocessing of preterm infant retinal images is the starting point of the entire monitoring process; the quality of the images directly affects the accuracy and reliability of all subsequent analysis steps.

[0025] In one embodiment of the present invention, a wide-angle fundus camera is used to acquire retinal images of preterm infants. Preferably, the field of view of the wide-angle fundus camera is not less than 130°, which can cover the complete area from the center of the optic disc to the peripheral retina, which is crucial for assessing the retinal vascularization process in preterm infants. Due to their small pupil diameter, low cooperation, and small eye size, preterm infants cannot easily obtain clear and complete visual field images with traditional fundus cameras with a field of view of 30° to 50°. However, wide-angle or even ultra-wide-angle fundus cameras can capture a wide retinal area, including the peripheral avascular zone, in a single image, facilitating the observation of the advancement of the vascular front and the degree of peripheral vascularization. In one embodiment of the present invention, the image acquisition resolution is set to 1280×1024 pixels or higher, with a spatial resolution of 3.5μm to 5μm per pixel, to ensure that microvessels with a diameter of 5μm or more can be distinguished. During the acquisition process, the influence of random noise can be reduced by multi-frame averaging technology. Preferably, 3 to 5 frames are acquired for each examination, and the image with the best quality is automatically selected by an image quality scoring algorithm for subsequent processing. Image quality can be rated based on a combination of indicators such as sharpness, contrast, and illumination uniformity.

[0026] After acquiring the raw retinal image, adaptive contrast enhancement processing needs to be performed. Premature infant retinal images have several unique imaging characteristics compared to adults: First, blood vessels are generally much smaller, with microvessels possibly only 5μm to 15μm in diameter, occupying only 1 to 3 pixels in the image width; second, the contrast between blood vessels and background retinal tissue is low, especially in peripheral neovascularization where grayscale differences are often insignificant; third, the retinal pigment epithelium is not yet fully developed, resulting in different overall reflective properties compared to adults; fourth, due to differences in fundus illumination and mydriasis, localized brightness unevenness often exists in the image. These characteristics lead to the easy omission of microvessels and unclear segmentation boundaries when directly using the raw image for blood vessel segmentation, making it difficult to obtain ideal segmentation results.

[0027] In one embodiment of the present invention, a contrast-limited adaptive histogram equalization algorithm is used for image enhancement. The core idea of ​​this algorithm is to divide the image into multiple local sub-blocks and perform histogram equalization independently on each sub-block, thereby effectively addressing the problem of uneven illumination in images of premature infants' retinas. Compared with global histogram equalization, this algorithm can better preserve local contrast details and avoid overexposure or underexposure in certain areas due to overall enhancement. Preferably, the sub-block size is set to 8×8, that is, the original image is divided into 64 sub-regions for processing. Histogram equalization within each sub-block can enhance the contrast within that local region, making the originally small grayscale differences between blood vessels and the background more clearly distinguishable.

[0028] The contrast limiting threshold is a key parameter affecting the enhancement effect. In standard adaptive histogram equalization, gray-level mapping in local areas may lead to excessive contrast amplification, thereby amplifying image noise. By setting a contrast limiting threshold, components exceeding the threshold in the histogram can be truncated and redistributed, effectively suppressing the noise amplification effect. A threshold that is too low will result in an insignificant enhancement effect, with blood vessels still difficult to distinguish from the background; a threshold that is too high may amplify interference information such as choroidal vessel shadows and noise. In one embodiment of the present invention, the contrast limiting threshold ranges from 2.0 to 4.0, with a preferred value of 3.0. At this threshold setting, a good balance can be achieved between effectively enhancing blood vessel visibility and reasonably suppressing noise amplification.

[0029] The relationship between the enhanced image and the original image can be represented as:

[0030] ,in: Indicates the original image in coordinates The pixel value at that location, for an 8-bit grayscale image, ranges from 0 to 255, where 0 represents pure black and 255 represents pure white; This represents the pixel value at the corresponding position in the enhanced image, and its value range is also from 0 to 255; The contrast limiting threshold is a dimensionless parameter with a value range of 2.0 to 4.0. This parameter controls the upper limit of the intensity of local contrast enhancement. When set to 3.0, it can increase the contrast of blood vessels in typical fundus images of premature infants by about 40% to 60%. The sub-block size parameter is set to 8×8 pixels. This parameter determines the size of the region where local histogram statistics are performed. Sub-blocks that are too small may introduce block artifacts, while sub-blocks that are too large will weaken local adaptability.

[0031] To eliminate the grayscale staircase effect that may occur at the boundaries of adjacent sub-blocks, bilinear interpolation is used to smoothly transition the enhancement results of adjacent sub-blocks. Consider a pixel located at the intersection of four adjacent sub-blocks; its final enhanced pixel value is obtained by weighted interpolation of the corresponding enhancement values ​​of the surrounding four sub-blocks. Let the pixel value at the sub-block boundary be... Its calculation method is expressed as follows:

[0032] ,in: , , , These represent the enhanced pixel values ​​at corresponding positions of four adjacent sub-blocks, in gray levels; This represents the horizontal interpolation coefficient, ranging from 0 to 1, determined by the normalized distance of the pixel in the horizontal direction relative to the boundary of the left sub-block. When the pixel is located at the left sub-block boundary, It is located at the right sub-block boundary; This represents the vertical interpolation coefficient, ranging from 0 to 1, determined by the normalized distance of the pixel relative to the upper sub-block boundary in the vertical direction. This bilinear interpolation process yields a smooth, artifact-free enhanced image, ensuring the continuity of blood vessel edges is unaffected by block boundaries.

[0033] In one embodiment of the present invention, green channel extraction preprocessing can be performed before CLAHE enhancement. Among the red, green, and blue color channels of a color fundus image, the contrast between blood vessels and the background is highest in the green channel image because hemoglobin in retinal vessels absorbs green light most strongly. Therefore, for a color fundus image, first extracting its green channel component and then performing CLAHE enhancement on the green channel grayscale image can further improve the visualization of blood vessels.

[0034] Furthermore, in one embodiment of the present invention, vascular enhancement filtering preprocessing can also be performed. Using a multi-scale Frangi filter or Hessian matrix eigenvalue analysis method, tubular structures are enhanced at different scales, which can further highlight the difference between blood vessels and the background, while suppressing non-tubular noise interference. This filtering step can be performed in series with CLAHE enhancement to form an image preprocessing pipeline.

[0035] Step S2: Vascular Network Segmentation and Extraction. The technical objective of this step is to accurately extract the complete network structure of retinal vessels from the enhanced image, obtaining a binary segmentation mask and a skeletonized network map of the vessels, providing accurate input data for subsequent morphological analysis. Vessel segmentation is the core component of the entire monitoring system, and the segmentation quality directly determines the reliability of morphological parameter calculation and developmental stage identification.

[0036] In one embodiment of the present invention, the blood vessel segmentation network adopts an encoder-decoder architecture, preferably the U-Net network structure or a variant optimized for fundus images of premature infants. U-Net is a classic fully convolutional neural network, whose symmetrical encoder-decoder structure and skip connection mechanism are particularly suitable for medical image segmentation tasks. The encoder part extracts multi-scale semantic features through layer-by-layer downsampling, the decoder part restores spatial resolution through layer-by-layer upsampling, and the skip connections pass the high-resolution feature map of the encoder to the decoder, enabling the network to simultaneously utilize deep semantic information and shallow spatial detail information, accurately locating blood vessel boundaries while maintaining high-resolution details.

[0037] The encoder section is responsible for extracting multi-scale vascular features, and its structure includes four downsampling stages. Each downsampling stage consists of the following components: two 3×3 convolutional layers for feature extraction, each convolutional layer followed by a batch normalization layer to accelerate training convergence and improve generalization ability, a ReLU activation function following the batch normalization layer to introduce non-linear transformation capability, and finally a 2×2 max pooling layer for downsampling to reduce the feature map size. The number of channels increases with the depth of the downsampling layer to enhance the expressive power of deep features, at 64, 128, 256, and 512 respectively.

[0038] Specifically, the first downsampling stage receives the enhanced image as input, passes it through two 3×3 convolutions, and outputs a 64-channel feature map. Then, 2×2 max pooling reduces the feature map size to half the original image size. The second downsampling stage takes the pooling output from the first stage as input, passes it through two 3×3 convolutions, and outputs a 128-channel feature map, which is then pooled to one-quarter the original image size. The third downsampling stage outputs a 256-channel feature map, one-eighth the size of the original image. The fourth downsampling stage outputs a 512-channel feature map, one-sixteenth the size of the original image. At the bottom layer of the encoder, the bottleneck layer, two 3×3 convolutional layers are used to further extract highly abstract deep semantic features, with 1024 channels. This layer captures the global contextual information of the image.

[0039] The decoder is responsible for recovering the spatial details of blood vessels, and its structure also includes four upsampling stages. Each upsampling stage consists of the following components: a 2×2 transposed convolution is used to double the size of the feature map for upsampling; skip connections concatenate and fuse the feature map of the same level as the encoder with the upsampling result in the channel dimension; after fusion, it is further integrated with two 3×3 convolutional layers to integrate multi-scale information. The number of channels decreases with the progression of the upsampling layers, gradually decreasing from 512 to 256, 128, and 64. The transposed convolution enables learnable upsampling, which has a stronger feature reconstruction capability than fixed methods such as bilinear interpolation. The skip connection mechanism is one of the core innovations of U-Net. It directly passes the high-resolution feature maps of each stage of the encoder to the corresponding stage of the decoder, enabling the decoder to recover the details of blood vessel edges using the fine spatial information of shallow layers, avoiding the problem of losing positional accuracy of deep features during multiple upsampling processes.

[0040] Finally, the decoder output is passed through a 1×1 convolutional layer to map the 64-channel features into a single channel, and then the output value is compressed to the range of 0 to 1 using a sigmoid activation function to obtain a probability map of the same size as the input image. In the probability map, each pixel value represents the probability that the location belongs to a blood vessel; the closer the value is to 1, the higher the probability that the pixel is a blood vessel.

[0041] Binarize the probability map to obtain the segmentation mask. Its calculation formula is expressed as:

[0042] ,in: For binary segmentation masks in coordinates The value at this location can be either 0 or 1. 0 indicates that the pixel belongs to the background, i.e., the non-vascular area, while 1 indicates that the pixel belongs to the foreground, i.e., the vascular area. This represents the probability value of the blood vessel at the corresponding location on the probability map, ranging from 0 to 1. The threshold value is used for binarization, ranging from 0.3 to 0.7, with a preferred value of 0.5. This threshold determines the balance between sensitivity and specificity in vessel segmentation: a lower threshold results in more potential vessel pixels, increasing sensitivity but potentially introducing more false positives; a higher threshold results in more conservative segmentation, increasing specificity but potentially missing some real vessels, especially microvessels. In practical applications, this threshold can be adjusted appropriately based on image characteristics and clinical needs.

[0043] The blood vessel segmentation network is trained using a fundus image dataset specifically designed for premature infants. In one embodiment of the invention, the training dataset contains retinal images of at least 200 premature infants and their corresponding ground truth blood vessel masks meticulously annotated by ophthalmologists. A composite loss function is used as the optimization objective during training, comprising a weighted combination of Dice loss and binary cross-entropy loss. Dice loss directly optimizes the overlap of segmented regions, exhibiting good robustness even in cases of severe imbalance between the number of blood vessel pixels and background pixels; cross-entropy loss performs classification optimization at the pixel level, refining blood vessel boundaries. Preferably, the weight ratio of the two losses is set to 1:1. The initial learning rate is set to 0.001, using the Adam optimizer, with 100 training epochs. Every 20 epochs, the learning rate is decayed to 0.5 times its original value to promote convergence. Data augmentation techniques, including random rotation, flipping, scaling, and brightness perturbation, are employed during training to improve the model's generalization ability.

[0044] After obtaining the binary segmentation mask, skeletonization is required to acquire the vascular network structure. Skeletonization, also known as thinning, aims to shrink the binary vascular region to a centerline of a single pixel width while maintaining the topological connectivity of the vascular network. In one embodiment of this invention, skeletonization employs a morphological thinning algorithm, which iteratively removes pixels from the edges of the vascular region, retaining only the center pixels that do not affect connectivity, until all vascular segments are shrunk to a single pixel width. Skeletonization generates a vascular skeleton map. This skeleton diagram accurately represents the topology of the vascular network and serves as the foundational data for subsequent morphological parameter calculations and longitudinal registration.

[0045] In one embodiment of the present invention, morphological opening is performed before skeletonization to remove isolated noise points and small false positive regions, using circular structural elements with a diameter of 3 to 5 pixels. After skeletonization, spur trimming is performed to remove short branches less than 5 to 10 pixels in length. These short branches are usually pseudo-forks caused by uneven segmentation edges; removal of these branches results in a smoother and more regular vascular skeleton. Furthermore, for broken vascular segments, endpoint extension and gap bridging techniques can be used for repair to ensure the connectivity and integrity of the vascular network.

[0046] Step S3: Calculation of vascular morphology parameters. The technical purpose of this step is to perform a comprehensive quantitative morphological analysis of the vascular network structure and extract multidimensional characteristic parameters that can characterize the development of retinal vessels in preterm infants. Vascular morphology analysis is one of the core innovations of this invention. By systematically calculating four categories of parameters—vascular diameter distribution, branching angle, tortuosity, and vascular density—a quantitative description system for the vascular network is established, providing an objective basis for subsequent developmental stage identification and maturity assessment.

[0047] Blood vessel diameter is an important indicator of the maturity of vascular development. During normal development, retinal vessels branch off from the optic disc, with the main trunk vessels being thicker and the terminal capillaries being thinner, forming a tree-like hierarchical structure. By statistically analyzing the distribution ratio of vessels within different diameter ranges, the degree of hierarchical differentiation of the vascular network can be assessed.

[0048] In one embodiment of the present invention, the blood vessel diameter is obtained by calculating the width of an orthogonal section along the direction of the vascular skeleton. The specific implementation steps are as follows: First, at fixed intervals on the vascular skeleton... Select sampling points, The value ranges from 5 to 10 pixels, with 8 pixels being the preferred value. The selection of the sampling interval needs to strike a balance between computational efficiency and sampling density: too small an interval will result in a large amount of redundant computation, while too large an interval may miss details of diameter changes. For each sampling point... First, calculate the tangent direction vector of the vascular skeleton at that point. The tangent direction is approximated by the direction of the line connecting adjacent skeleton points. Then, the normal unit vector perpendicular to the tangent is calculated. Extract grayscale profiles from the binary segmentation mask or the original grayscale image along the normal direction.

[0049] Let the sampling point be the center in the normal direction, and with Extracting grayscale profiles from pixels at half length ,in The distance parameter is along the normal direction, and its value range is... , In this embodiment of the invention, half the maximum expected diameter of the blood vessel plus an appropriate margin is added. Setting the resolution to 50 pixels allows for the coverage of large blood vessels with a diameter of 100 pixels, or approximately 350 μm or more. Gaussian function fitting is applied to the grayscale profile to determine the vessel boundary locations; the fitting formula is expressed as:

[0050] ,in: Indicates the first The vertical distance at each sampling point is The grayscale value of the location ranges from 0 to 255 for an 8-bit image; The amplitude parameter of the Gaussian peak is expressed in gray levels and reflects the difference in gray level contrast between blood vessels and the background. The larger the value, the more obvious the distinction between blood vessels and the background. This represents the offset of the Gaussian peak center position relative to the skeleton sampling point, in pixels, and is used to correct for slight deviations that may exist in the skeleton centerline positioning. This represents the standard deviation of the Gaussian distribution, expressed in pixels. This parameter is directly proportional to the width of the blood vessel. This represents the background grayscale baseline, expressed in grayscale levels, and represents the average grayscale value of the non-vascular regions at both ends of the profile.

[0051] After determining the above four fitting parameters through least squares method or nonlinear iterative optimization, the first... Blood vessel diameter at each sampling point The calculation is as follows: ,in: For the first The physical dimensions of the blood vessel diameter at each sampling point are expressed in μm. The standard deviation parameter is obtained from Gaussian fitting, in pixels; Image spatial resolution, i.e., the actual physical length corresponding to each pixel, is expressed in μm / pixel. This value is obtained by calibrating the optical parameters and working distance of the fundus imaging system. In this embodiment of the invention, a typical value is 3.5 μm / pixel to 5 μm / pixel; coefficient This is the conversion factor between the full width at half maximum (FWHM) and the standard deviation of the Gaussian function, with a value of approximately 2.355. This conversion relationship is derived based on the mathematical properties of the Gaussian distribution.

[0052] retinal vessel diameter in premature infants The typical range of values ​​for λ is 5 μm to 200 μm. Based on diameter, blood vessels can be divided into three categories: microvessels or capillaries with diameters ranging from approximately 5 μm to 20 μm, which have the highest sensitivity requirements for image enhancement and segmentation algorithms; medium-sized vessels with diameters ranging from approximately 20 μm to 80 μm, which are the main components of the retinal vascular network; and trunk vessels with diameters ranging from approximately 80 μm to 200 μm, which are mainly distributed in the area near the optic disc.

[0053] By statistically analyzing the blood vessel diameter measurements at all sampling points, the blood vessel diameter distribution can be obtained. In one embodiment of the present invention, the diameter values ​​are discretized into several intervals, and the proportion of the blood vessel segment length to the total blood vessel length within each interval is calculated to form a blood vessel diameter distribution histogram. This histogram can intuitively reflect the hierarchical differentiation characteristics of the vascular network: a mature vascular network should exhibit a multi-peak distribution, corresponding to the main trunk, intermediate, and micro vessels; an early-stage vascular network may have only a few branches, resulting in fewer peaks in the histogram.

[0054] The branching angle of blood vessels is an important characteristic of the geometric properties of a vascular network, reflecting the spatial orientation of vessels at bifurcation points. Normal retinal vessels branch according to certain optimal hydrodynamic laws, and the branching angles typically fall within a specific range. Abnormal proliferation or developmental disorders may cause the branching angles to deviate from the normal range.

[0055] In one embodiment of the present invention, the calculation of the branch angle first requires detecting bifurcation points on the vascular skeleton. A bifurcation point is defined as a pixel on the skeleton with a neighborhood connectivity factor greater than 2, that is, the point connects to three or more vascular segment branches. Under the definition of 8-neighborhood connectivity, the bifurcation point is usually the center location of a three-way bifurcation or a higher-order bifurcation.

[0056] For each detected bifurcation point Identify the branches of each blood vessel segment connected to it, and then extend a certain length along the direction of each branch. Then calculate the direction vector of this branch. Extension length. The selection of the bifurcation point needs to avoid interference from the direction estimation caused by the bend of the blood vessel near the bifurcation point, while also not extending too far beyond the local area of ​​the bifurcation. In one embodiment of the present invention, The value ranges from 10 to 30 pixels, with a preferred value of 20 pixels. Let the first pixel connected to the bifurcation point be... The direction vector of the blood vessel branch is The vector extends from the bifurcation point coordinates to the direction along that branch. The coordinates of the skeleton point after the pixel. Then the... Article and No. Branching angle between blood vessel branches The calculation is as follows: ,in: For the first Article and Section The angle between branches of a blood vessel, measured in degrees, ranges from 0° to 180°. 0° indicates that the two branches are in the same direction, i.e., they turn back, while 180° indicates that the two branches are in completely opposite directions, i.e., they pass through in a straight line. and These are the direction vectors of the two blood vessel branches; This is an inverse cosine function, returning the value in radians; multiplied by... The conversion factor converts radians to degrees.

[0057] For a typical three-way bifurcation point, the angles between the three pairs of branches can be calculated. In normal retinal vessels, the bifurcation of the parent vessel into two daughter vessels typically results in branch angles between 60° and 120°, with an average of approximately 75° to 85°. An angle that is too small indicates that the two daughter vessels are too close in direction, while an angle that is too large indicates that the bifurcation is too flat. The branch angles at all bifurcation points are statistically analyzed, and the average branch angle is calculated. and branch angle standard deviation As a morphological characteristic parameter of vascular networks.

[0058] Torque, also known as tortuosity, is a dimensionless index describing the degree of vascular curvature. It is of significant clinical importance in assessing retinal vascular developmental abnormalities and hemodynamic status. Normal blood vessels exhibit moderate physiological curvature to adapt to the curvature of the eyeball and blood flow pressure, while in pathological conditions, blood vessels may exhibit excessive tortuosity or abnormal course.

[0059] In one embodiment of the present invention, a multi-scale analysis method is used to calculate the tortuosity of blood vessels to comprehensively reflect the tortuosity characteristics at different spatial scales. On the vascular skeleton, blood vessel segments are extracted with different window lengths, and tortuosity indices at each scale are calculated separately, and then weighted and fused into the final tortuosity.

[0060] Set at scale Below, starting from a point on the skeleton, a section is cut along the direction of the blood vessel, with a length of... The vascular segment of the pixel, the starting point of which is The endpoint is The actual arc length of the vascular segment along the skeleton is... This is obtained by accumulating the Euclidean distances between adjacent pixels on the skeleton. The straight-line distance between the start and end points, i.e., the chord length, is... Curvature at this scale Defined as the ratio of arc length to chord length minus 1: ,in: For scale The curvature index is dimensionless and ranges from 0 to positive infinity. When the vessel segment is perfectly straight, the arc length equals the chord length. The more tortuous the blood vessel, the greater the arc length relative to the chord length. The higher the value. It represents the cumulative arc length of the blood vessel segment along the skeleton, expressed in pixels. It is calculated by accumulating the distance between adjacent points along the skeleton from the starting point to the ending point. This represents the Euclidean straight-line distance between the starting and ending points, in pixels. Subtracting 1 makes the curvature of the line segment exactly 0, which facilitates interpretation and comparison.

[0061] In one embodiment of the present invention, three feature scales are used for multi-scale analysis. Small scale Set to 20 pixels, it is sensitive to localized minute bends and high-frequency fluctuations, enabling it to capture the subtle tortuosity of blood vessels; large-scale... Set to 100 pixels to reflect the overall curvature of blood vessels and smooth local fluctuations; mesoscale. Set to 50 pixels, somewhere in between, to balance local and overall features. Final curvature metric. Take a weighted average of the curvature across the three scales: ,in: The final multi-scale curvature index is dimensionless; , , The weighting coefficients for the three scales satisfy the normalization constraint. In this embodiment of the invention, the weights are set to... , , The mesoscale weight is maximized to achieve a balance between local details and global features. This weighting ratio has been experimentally validated using a dataset of fundus images from premature infants, demonstrating its effectiveness in distinguishing the vascular tortuosity features of normal development from those of abnormal lesions. (The tortuosity of retinal vessels in normal premature infants is shown in the image.) The typical value range is 0.05 to 0.3, and the tortuosity may exceed 0.5 or even higher when abnormal angiogenesis occurs.

[0062] Vascular density reflects the distribution and coverage of the vascular network in the retinal region and is the most direct quantitative indicator for assessing the vascularization process. In premature infants, retinal vessels gradually grow from the optic disc towards the periphery, and vascular density increases with development.

[0063] In one embodiment of the present invention, blood vessel density Defined as the ratio of the area of ​​a blood vessel pixel to the area of ​​the effective retinal region, its calculation formula is expressed as: ,in: This is a vascular density index, expressed as a percentage, with values ​​ranging from 0% to 100%. This represents the set of pixels representing the effective area of ​​the retina, which excludes invalid black areas at the image edges and the circular area of ​​the optic disc. The optic disc area is excluded because it does not contain retinal vessels and may be overexposed. For binary segmentation masks in coordinates The value at that location is 1 for blood vessel pixels and 0 for background pixels; This represents the total number of pixels in the effective area of ​​the retina. The vascular density of the retina in a normal full-term newborn is about 8% to 15%. In premature infants, because the vascular network is not fully developed, especially the peripheral areas which contain avascular areas, the overall vascular density is usually lower than that of full-term infants, typically ranging from 5% to 12%.

[0064] In one embodiment of the present invention, local vascular density is calculated by dividing the retina into regions based on its distance from the center of the optic disc. The effective retinal area is divided into multiple concentric ring regions centered on the center of the optic disc, for example, into four regions: 0 to 1 disc diameter, 1 to 2 disc diameter, 2 to 3 disc diameter, and greater than 3 disc diameter. The vascular density of each region is calculated separately. This radial vascular density distribution can intuitively reflect the degree of vascular development from the optic disc to the periphery: during normal development, the vascular density in each region should gradually become more uniform; in the early stages of development or in abnormal conditions, the density in the peripheral regions is significantly lower than that in the central region.

[0065] Step S4: Developmental Stage Identification. The technical objective of this step is to automatically identify the developmental stage of the retina in premature infants based on vascular morphological characteristic parameters, establish the correspondence between quantitative characteristics of the vascular network and the vascularization process, and provide clinicians with objective developmental assessment data. Developmental stage identification is a key step in this invention that combines quantitative morphological analysis with clinical application.

[0066] In one embodiment of the present invention, the developmental stage identification first calculates the topological complexity index of the vascular network, and then outputs the developmental stage label through the developmental stage classification model.

[0067] Vascular networks can be modeled and analyzed from a graph theory perspective. The skeletonized vascular network can be abstracted as an undirected graph. , where the set of nodes It includes vessel endpoints and bifurcation points. An endpoint is a skeletal point connected to only one vessel segment, while a bifurcation point is a skeletal point connected to multiple vessel segments; edge set. It contains blood vessel segments connecting adjacent nodes, with each edge corresponding to a continuous blood vessel segment. Let the total number of nodes be... The total number of edges is .

[0068] Network connectivity Defined as the ratio of the number of edges to the number of nodes, it characterizes the connection density of a vascular network. ,in: It is a network connectivity metric, dimensionless, reflecting half the average number of edges connected to each node; This represents the number of edges in the vascular network, i.e., the total number of vascular segments, expressed in units of individual segments. This represents the number of nodes in the vascular network, i.e., the total number of endpoints and branching points, expressed in units of nodes. For a pure tree structure, the number of edges equals the number of nodes minus 1. Approaching 1; as the number of loops and matching branches in the network increases, It will enlarge. A normally developing retinal vascular network. Typical values ​​range from 1.0 to 1.5; abnormal proliferation may result in more loops leading to... Increase.

[0069] Branch level depth This is obtained by performing a breadth-first search (BFS) traversal starting from the center of the optic disc. The blood vessel node closest to the center is designated as the root node. Starting from the root node, a BFS algorithm is used to traverse the entire blood vessel network, recording the number of edges traversed from the root node to each node, i.e., the graph theory path length. The branch depth is the maximum value of all node path lengths. ,in: This is a dimensionless index of branching level depth, representing the number of branching levels traversed from the root node at the center of the optic disc to the farthest terminal node of the blood vessel. The root node is usually selected as the vascular node closest to the center of the optic disc; Let be any node in the vascular network diagram; From the root node to the node The shortest graph theory distance is the distance traversed with the fewest edges. A mature retinal vascular network has completed multi-level branching. Typically between 6 and 10; in early development, the vascular network has fewer branching levels. It may only be 3 to 5.

[0070] Vascular coverage Defined as the ratio of the spatial distribution range of the vascular network to the total area of ​​the retina. In one embodiment of the present invention, the convex hull area formed by the coordinates of all nodes in the vascular network is used to characterize the vascular distribution range: ,in: This is an indicator of vascular coverage, expressed as a percentage, with values ​​ranging from 0% to 100%. The area of ​​the convex hull formed by the coordinates of all nodes in the vascular network is expressed in square pixels. The convex hull is the smallest convex polygon that contains all nodes. This represents the effective area of ​​the retina, expressed in square pixels. In full-term newborns, retinal vessels extend to the vicinity of the ora serrata, with a coverage rate approaching 100%. In premature infants, the peripheral retina is not yet vascularized, resulting in a coverage rate of less than 100%, which is related to gestational age and corrected gestational age.

[0071] Combining the above three indicators, a topological complexity feature vector is constructed. .

[0072] In one embodiment of the present invention, the developmental stage classification model employs a Bayesian classifier. The topological complexity index is used... The morphological feature parameters calculated in step S3 The features are concatenated to form a 9-dimensional input feature vector. ,in This represents the average blood vessel diameter. This represents the standard deviation of the diameter.

[0073] Based on the physiological process of retinal vascularization in premature infants, developmental stages are divided into four categories: Stage I, the initial stage of vascularization, represents the stage where blood vessels begin to grow outward from the optic disc and have limited coverage; Stage II, the rapid development stage, represents the stage where the vascular network is rapidly expanding peripherally and the branching hierarchy is increasing; Stage III, the stable maturation stage, represents the stage where blood vessels have basically covered the entire retina and the network structure is becoming stable; and Stage IV, the abnormal proliferation stage, represents the stage where pathological changes such as abnormal neovascularization occur. Let the developmental stage categories be... Bayesian classifiers calculate posterior probabilities based on conditional and prior probabilities:

[0074] ,in: Given a feature vector When the sample belongs to the category The posterior probability of the class is between 0 and 1, and the sum of the posterior probabilities of the four classes is 1. For category The likelihood function is assumed to have a multivariate Gaussian distribution, and its mean vector and covariance matrix parameters are estimated from a preterm infant retinal development annotation dataset. For category The prior probability is determined based on the actual frequency of occurrence of each developmental stage in the premature infant population; The marginal probability distribution of the feature vector is used as a normalization constant to ensure that the sum of the posterior probabilities of the four categories is 1.

[0075] Final developmental stage tags Choose the category with the highest posterior probability: In one embodiment of the present invention, the classification model also outputs the probability distribution of each category as a confidence reference. When the difference between the maximum posterior probability and the second-largest posterior probability is small, it indicates that the current sample is located in the boundary region between two stages, and clinicians can refer to the probability distribution and combine it with other clinical information to make a comprehensive judgment.

[0076] Step S5: Longitudinal tracking analysis. The technical purpose of this step is to accurately register retinal images of the same preterm infant at different examination time points, quantify the dynamic growth changes of the vascular network over time, and provide longitudinal data support for developmental trend assessment. Longitudinal tracking is a key innovation of this invention compared to existing single-time-point analysis methods.

[0077] Retinal screening for premature infants typically requires multiple follow-up examinations, with intervals set according to developmental status and risk level, usually 1 to 2 weeks. Because image acquisition conditions vary at different examination time points, including imaging angle, magnification, eye posture, and degree of pupil dilation, directly comparing vascular network images from different periods cannot accurately reflect true developmental changes. Therefore, image registration technology is needed to align images from different periods to a unified spatial coordinate system.

[0078] In one embodiment of the present invention, image registration employs a feature point-based method. Compared to pixel-grayscale-based registration methods, the feature point method exhibits better robustness to changes in illumination and imaging conditions, making it more suitable for registration scenarios involving fundus images of premature infants.

[0079] Feature points are selected from key points on the vascular skeleton that possess stable geometric properties, including vascular bifurcation points and vascular endpoints. Bifurcation points, as nodes where three or more vessels converge, have unique topological locations within the vascular network; endpoints, as the terminating locations of vascular fronts, reflect the boundaries of vascular expansion. These two types of points have corresponding relationships in images from different time periods and can be used to establish registration transformations.

[0080] Let the first inspection be the baseline time point. The feature point set is ,Include The first feature point. The second follow-up examination is the time point. The feature point set is ,Include The number of feature points is [number missing]. Due to the continuous development of the vascular network, the number of feature points examined subsequently [increases / decreases]. Typically greater than or equal to the number of feature points checked in the benchmark. .

[0081] Feature point matching employs the Iterative Nearest Point (ICP) algorithm. First, an initial coarse matching is performed based on the local descriptors of the feature points. These local descriptors can utilize geometric features such as the angle combination and branch length ratio of the branches surrounding the vessel bifurcation point. Point pairs with similar descriptors are considered candidate matching pairs. Then, the matching relationship and transformation parameters are iteratively optimized using the ICP algorithm. Each iteration includes two steps: matching update and transformation update. Matching update finds the nearest target point for each source point under the current transformation; transformation update minimizes the registration error after transformation based on the current matched point pair. Iteration continues until convergence or the maximum number of iterations is reached. The registration error is defined as the sum of the squares of the distances between matched point pairs: ,in: Registration error, expressed in square pixels; The final number of matching pairs; Let be the spatial transformation function to be solved; For time points The Coordinates of one feature point; The coordinates of the feature points at the corresponding reference time point; To match the index mapping function.

[0082] After obtaining stable matching point pairs, registration parameters are calculated based on the thin-plate spline TPS transform. The thin-plate spline transform is a flexible, non-rigid transformation model capable of handling local deformations and suitable for non-uniform geometric changes that may occur during the development of the preterm infant's eyeball. The mathematical form of the TPS transform includes affine components and nonlinear deformation components: ,in: Input coordinates The transformed target coordinates, in pixels; It is a 3×3 affine transformation matrix, containing linear transformation components such as translation, rotation, scaling, and shearing; The homogeneous coordinate representation of the input coordinate point; The number of control points must equal the number of matching point pairs; For the first The deformation weight coefficients of each control point are obtained by solving a system of linear equations. For the first The coordinates of each control point are the coordinates of the matching reference point; Let be the radial basis functions of the thin plate spline, where This is the Euclidean distance to the control point.

[0083] Transform the time points using TPS. The vascular network was aligned to the baseline time point. After establishing the coordinate system, precise longitudinal comparisons can be performed to calculate the dynamic changes in the vascular network.

[0084] Expansion of vascular coverage area Defined as the percentage increase in vascular coverage area at the current examination time relative to the baseline time: ,in: The extent of vascular coverage expansion is expressed as a percentage. For time points The registered blood vessel coverage area, in square pixels; As the reference time point The area covered by blood vessels, measured in square pixels. During normal development... A positive and stable growth rate indicates that the vascular network is expanding outwards; if Stagnant growth or abnormally rapid growth may indicate developmental abnormalities.

[0085] Neovascularization detection is achieved by comparing the registered vascular skeleton. Let the vascular skeleton at the baseline time point be... The registered vascular skeleton at the current time point is To tolerate slight registration errors, the baseline skeleton is morphologically expanded to a certain extent. Neovascularization region. Defined as the portion of the current skeleton that exceeds the base skeleton after expansion: ,in: This is the set of pixels representing the detected neovascularization region; Indicated by radius The region after morphological expansion of the baseline skeleton; To determine the registration tolerance radius, in this embodiment of the invention, a radius of 5 to 10 pixels is used. This tolerance is used to absorb residual errors in registration alignment. This represents the set difference operation. Each connected component in the model represents a newly detected neovascular segment, and their number and total length are counted as longitudinal variation features.

[0086] In addition, the morphological evolution trend of the matched blood vessel segment can be calculated, including the rate of change of diameter and the rate of change of tortuosity of the same blood vessel segment at different times, which can be used to assess the development and maturation process of a single blood vessel.

[0087] Step S6: Developmental Assessment Report Generation. The technical objective of this step is to synthesize the analysis results of all preceding steps to generate an individualized retinal developmental assessment report, providing comprehensive data support for clinicians' diagnostic decisions and follow-up planning. Report generation is the final step in translating technical analysis into clinical application.

[0088] In one embodiment of the present invention, the developmental assessment report includes the following core content modules:

[0089] Vascularization Progress: Report the vascular coverage, vascular density, and distribution characteristics of the vascular network in different regions of the retina at the current examination time. For vascular coverage, provide a specific percentage value and compare it with reference values ​​for the same gestational age; for vascular density, provide the overall density and the density of each radial zone. If the current examination is a follow-up, also report the expansion of vascular coverage area and the results of neovascularization detection compared to the previous examination.

[0090] Network maturity index: In one embodiment of the present invention, a network maturity score is constructed. As a single indicator comprehensively reflecting the maturity of vascular network development, the maturity score is calculated by weighted fusion of multiple morphological and topological features. ,in: The network maturity score is set, with a value ranging from 0 to 1. The closer the value is to 1, the more mature the vascular network development is and the closer it is to the level of a full-term newborn. The total number of features involved in the calculation; For the first The original measurements of each feature; Features The standardized values ​​after normalization are normalized with reference to the characteristic statistical distribution of the normal full-term newborn population, mapping each characteristic to the interval between 0 and 1; For the first The weight coefficients of each feature are determined such that the sum of all weights is 1. The weights are determined based on the correlation analysis between each feature and developmental maturity.

[0091] In one embodiment of the present invention, the main features involved in maturity calculation include vascular density, average branching angle, average tortuosity, vascular coverage, and network connectivity, with corresponding weights set to 0.25, 0.15, 0.15, 0.30, and 0.15, respectively. This weight configuration has been validated with clinical data and can effectively distinguish maturity levels at different developmental stages.

[0092] Developmental Stage Determination: Report the currently identified developmental stage label (early vascularization, rapid development, stable maturation, or abnormal proliferation) and the confidence probability distribution for each stage. A low confidence level suggests to clinicians that the sample may be located in a stage boundary region, requiring comprehensive judgment in conjunction with other clinical information.

[0093] Developmental trajectory prediction: For children who have undergone multiple follow-up examinations, one embodiment of this invention uses a Gaussian process regression model based on longitudinal tracking data to predict future developmental trajectories. Let the observed examination time series be... The corresponding maturity score sequence is The Gaussian process regression model fits a trend curve of maturity over time based on observed data and extrapolates to predict the next inspection time point. Expected maturity The predicted interval and its 95% confidence interval are then compared with a normal developmental reference curve for the same gestational age to assess whether the developmental trend is normal.

[0094] Follow-up recommendations: Based on the current developmental stage, maturity score, and developmental trajectory prediction results, follow-up examination recommendations are automatically generated. For children in the early stages of vascularization or rapid development with low maturity scores, it is recommended to shorten the follow-up interval to 1 week to closely monitor the developmental process; for children in the stable maturation stage with high maturity scores, the follow-up interval can be appropriately extended to 2 weeks; for children whose developmental trends deviate from the normal range or who show abnormal proliferative risk signals, it is recommended to refer them to an ophthalmology specialist as soon as possible for detailed evaluation and possible intervention treatment.

[0095] Closed-loop feedback mechanism: The generation of developmental assessment reports not only summarizes and outputs data but also includes quality feedback and parameter optimization suggestions for preceding processing steps. When image quality issues are detected during report analysis, such as a high false negative rate in vessel segmentation due to insufficient contrast, the system automatically records and suggests adjusting the contrast enhancement parameters in step S1. When the confidence level for developmental stage identification remains consistently low, incremental learning and updating of the classification model in step S4 is triggered, utilizing accumulated labeled data to optimize model parameters. This closed-loop mechanism enables the entire monitoring system to continuously optimize performance in practical applications.

[0096] This invention also provides a vascular morphology-based system for monitoring retinal development in preterm infants, such as... Figure 2 As shown, the system comprises six functional modules: image acquisition and enhancement module 1, blood vessel segmentation module 2, morphological analysis module 3, developmental stage identification module 4, longitudinal tracking module 5, and report generation module 6. Each module corresponds one-to-one with the steps in the method embodiment, realizing the corresponding data processing and analysis functions. The modules are connected through data interfaces to form a complete processing pipeline.

[0097] The image acquisition and enhancement module 1 comprises two sub-units: an image acquisition unit and an image enhancement unit. The image acquisition unit provides a hardware interface with a wide-angle fundus camera, supporting real-time acquisition and transmission of image data. Preferably, it supports medical imaging standard formats such as DICOM, facilitating integration with hospital information systems. The image acquisition unit also includes an automatic image quality assessment function, scoring acquired images based on indicators such as sharpness, contrast, and illumination uniformity, and automatically selecting qualified images for subsequent processing. The image enhancement unit implements adaptive contrast enhancement processing as described in step S1 of the method embodiment, incorporating the CLAHE algorithm and supporting parameter configuration for sub-block size and contrast limit thresholds. Enhancement parameters can be adaptively adjusted according to image characteristics or manually set by the user.

[0098] The vessel segmentation module 2 comprises two sub-units: a deep learning inference unit and a skeleton extraction unit. The deep learning inference unit loads a pre-trained vessel segmentation network model file, supporting GPU-accelerated inference to meet real-time processing requirements. The model employs the U-Net encoder-decoder architecture described in step S2 of the method embodiment, with encoder channels configured as 64, 128, 256, and 512. The decoder recovers spatial details through transposed convolutions and skip connections. The deep learning inference unit receives the enhanced image as input and outputs a vessel probability map. The skeleton extraction unit performs post-processing operations on the probability map, including binarization thresholding, morphological denoising, skeleton thinning, and spur trimming, outputting a binary vessel segmentation mask and a single-pixel-width vessel skeleton map.

[0099] The morphological analysis module 3 comprises four sub-units: a diameter calculation unit, a branch angle calculation unit, a tortuosity calculation unit, and a density calculation unit. The diameter calculation unit calculates the vessel diameter distribution as described in step S3 of the method embodiment, extracts orthogonal grayscale profiles along the skeleton, determines the vessel width through Gaussian fitting, and outputs a diameter distribution histogram and statistical parameters. The branch angle calculation unit detects bifurcation points on the skeleton, calculates the angle between each branch, and outputs statistical parameters of the angle distribution. The tortuosity calculation unit uses a three-scale multi-scale analysis method to calculate the tortuosity of each vessel segment and performs weighted fusion. The density calculation unit calculates the overall vessel density and the density of different regions, and outputs density distribution features. The outputs of all units are summarized to form a complete morphological feature parameter vector.

[0100] The developmental stage identification module 4 comprises two sub-units: a topology analysis unit and a stage classification unit. The topology analysis unit models the vascular skeleton as a graph structure, extracts nodes and edges, and calculates three topology complexity indices: network connectivity, branch level depth, and vascular coverage, as described in step S4 of the method embodiment. The stage classification unit has a built-in pre-trained Bayesian classifier model, receiving the topology complexity indices and morphological feature parameters as input, and outputting developmental stage labels and the posterior probability distribution of each stage. The classifier model supports incremental learning updates, allowing online parameter optimization after accumulating sufficient labeled samples.

[0101] The longitudinal tracking module 5 comprises three sub-units: a feature point extraction unit, a registration calculation unit, and a change analysis unit. The feature point extraction unit identifies bifurcation points and endpoints from the vascular skeleton as registration feature points and calculates local geometric descriptors for initial matching. The registration calculation unit performs ICP feature point matching and thin-plate spline transformation parameter solving as described in step S5 of the method embodiment, aligning images from different time periods to a unified coordinate system. The change analysis unit compares the registered vascular network, calculates the coverage area expansion, detects neovascularization areas, analyzes morphological evolution trends, and outputs longitudinal change feature data.

[0102] The report generation module 6 comprises three sub-units: a data aggregation unit, a maturity assessment unit, and a report output unit. The data aggregation unit collects all analysis results data from the other five modules, performs formatting and validity verification. The maturity assessment unit performs network maturity score calculation and Gaussian process regression developmental trajectory prediction as described in step S6 of the method embodiment. The report output unit generates a standardized developmental assessment report according to a predefined report template, supporting output formats such as PDF and HTML for easy printing, archiving, and electronic distribution. The report content includes modules such as vascularization progress status, network maturity score, developmental stage determination, developmental trajectory prediction, and follow-up recommendations.

[0103] The various modules of this system are connected via an internal data bus, enabling efficient data transfer and automatic scheduling of processing flows. The system also includes a patient information management module and a historical data storage module, used to manage the basic information of the children and store data from previous examinations, supporting longitudinal tracking analysis and population statistical analysis. The system provides a graphical user interface for easy operation and result viewing by clinicians.

[0104] The embodiments of the present invention are not limited to the specific embodiments described above. Those skilled in the art can make various equivalent changes or substitutions based on the technical solutions of the present invention, and all such changes or substitutions should be included within the protection scope of the present invention.

Claims

1. A method for monitoring retinal development in preterm infants based on vascular morphology, characterized in that, Includes the following steps: Step S1, Image Acquisition and Enhancement Processing: A wide-angle fundus camera is used to acquire retinal images of premature infants. Adaptive contrast enhancement processing is performed on the retinal images, and the visualization effect of low-contrast vascular areas is improved by local histogram equalization to obtain enhanced images. Step S2, Vascular network segmentation and extraction: The enhanced image is input into the pre-trained vascular segmentation network. The vascular segmentation network adopts an encoder-decoder architecture. The encoder extracts multi-scale vascular features through multi-layer convolution and pooling operations. The decoder restores the spatial details of the vascular system through upsampling and skip connections and outputs a binary segmentation mask of the retinal vessels. Based on the binary segmentation mask, skeletonization processing is performed to obtain the vascular network structure. Step S3, Calculation of vascular morphological parameters: Perform morphological analysis on the vascular network structure and extract morphological feature parameters, including vascular diameter distribution, branching angle, tortuosity, and vascular density. The vascular diameter distribution is obtained by calculating the width of the orthogonal section along the vascular skeleton direction. The branching angle is obtained by detecting the bifurcation point of the vascular segment and calculating the angle between adjacent vascular segments. The tortuosity is obtained by calculating the ratio of the arc length to the chord length of the vascular segment. The vascular density is obtained by calculating the proportion of vascular pixels to the effective area of ​​the retina. Step S4, Developmental Stage Identification: Calculate the topological complexity index of the vascular network based on morphological feature parameters. The topological complexity index includes network connectivity, branch level depth, and vascular coverage. Input the topological complexity index into the developmental stage classification model. The developmental stage classification model outputs developmental stage labels based on the vascularization process characteristics. Step S5, longitudinal tracking analysis: Perform image registration on retinal images of the same child at different examination time points, obtain registration transformation parameters through feature point matching, align the vascular network structure at different time periods to a unified coordinate system based on the registration transformation parameters, and calculate the longitudinal change characteristics of the vascular network. The longitudinal change characteristics include the expansion of the vascular coverage area, the number of new blood vessels, and the trend of vascular morphological evolution.

2. The method for monitoring retinal development in preterm infants based on vascular morphology according to claim 1, characterized in that, In step S1, the adaptive contrast enhancement processing uses a limited contrast adaptive histogram equalization algorithm to divide the retinal image into 8×8 sub-blocks, perform histogram equalization independently on each sub-block, set the contrast limit threshold to 2.0 to 4.0, and use bilinear interpolation to eliminate sub-block boundary artifacts.

3. The method for monitoring retinal development in preterm infants based on vascular morphology according to claim 1, characterized in that, In step S2, the encoder of the blood vessel segmentation network contains four downsampling stages, each of which includes two 3×3 convolutional layers, a batch normalization layer, a ReLU activation function, and a 2×2 max pooling layer, with the number of channels being 64, 128, 256, and 512, respectively. The decoder contains four upsampling stages, each of which includes a 2×2 transposed convolution, skip connection feature concatenation, and two 3×3 convolutional layers.

4. The method for monitoring retinal development in preterm infants based on vascular morphology according to claim 1, characterized in that, In step S3, the calculation of the blood vessel diameter distribution includes: calculating a grayscale profile perpendicular to the skeleton direction along the blood vessel skeleton with sampling points spaced 5 to 10 pixels apart as the center, determining the blood vessel boundary position by Gaussian fitting, and the blood vessel diameter range is 5 μm to 200 μm.

5. The method for monitoring retinal development in preterm infants based on vascular morphology according to claim 1, characterized in that, In step S3, the curvature is calculated using a multi-scale analysis method. Blood vessel segments are extracted from the vascular skeleton at different window lengths, including three scales: 20 pixels, 50 pixels, and 100 pixels. The ratio of the arc length of the blood vessel segment to the straight-line distance to the endpoint is calculated at each scale. The weighted average of the curvature at the three scales is taken as the final curvature index, with weights of 0.2, 0.5, and 0.3, respectively.

6. The method for monitoring retinal development in preterm infants based on vascular morphology according to claim 1, characterized in that, In step S4, the calculation of the topological complexity index includes: modeling the vascular network structure as an undirected graph, where the nodes of the graph are the endpoints and branching points of the blood vessels, the edges of the graph are the segments of the blood vessels, the network connectivity is defined as the ratio of the number of edges to the number of nodes, the branch level depth is obtained by performing a breadth-first traversal from the center of the optic disc to obtain the maximum path length, and the blood vessel coverage is defined as the ratio of the convex hull area of ​​the vascular network to the effective area of ​​the retina.

7. The method for monitoring retinal development in preterm infants based on vascular morphology according to claim 1, characterized in that, In step S4, the developmental stage classification model uses a Bayesian classifier, taking the topological complexity index and morphological feature parameters as input feature vectors. Based on the preterm infant retinal development dataset, it obtains the prior probability and conditional probability distribution of each developmental stage. The developmental stages include four categories: early vascularization, rapid development, stable maturation, and abnormal proliferation.

8. The method for monitoring retinal development in preterm infants based on vascular morphology according to claim 1, characterized in that, In step S5, image registration includes: extracting feature points from the vascular network structure at different time points, including vascular bifurcation points and vascular endpoints; using the iterative nearest point algorithm to match feature points; calculating thin-plate spline transformation parameters based on the matched feature point pairs; and aligning the vascular network at subsequent examination time points to the coordinate system of the first examination time point through thin-plate spline transformation.

9. The method for monitoring retinal development in preterm infants based on vascular morphology according to claim 1, characterized in that, It also includes step S6, developmental assessment report generation: by integrating morphological feature parameters, developmental stage labels and longitudinal change features, an individualized retinal developmental assessment report is generated. The developmental assessment report includes vascularization progress, network maturity index and developmental trajectory prediction information. The calculation of the network maturity index includes: constructing a multi-dimensional feature vector based on vascular density, average branch angle, average curvature and topological complexity, inputting the feature vector into a trained regression model, and outputting a maturity score ranging from 0 to 1. The closer the maturity score is to 1, the more mature the vascular network development.

10. A vascular morphology-based preterm infant retinal development monitoring system, used to implement the vascular morphology-based preterm infant retinal development monitoring method of claim 9, characterized in that, It includes an image acquisition and enhancement module, a blood vessel segmentation module, a morphological analysis module, a developmental stage recognition module, a longitudinal tracking module, and a report generation module; The image acquisition and enhancement module is used to acquire images of the retina of premature infants, perform adaptive contrast enhancement processing on the retinal images, and output enhanced images; The vessel segmentation module is used to input enhanced images into a pre-trained vessel segmentation network and output a binary segmentation mask for retinal vessels and the vessel network structure. The morphological analysis module is used to perform morphological analysis on the vascular network structure and extract morphological feature parameters, including vessel diameter distribution, branch angle, tortuosity, and vessel density. The developmental stage identification module is used to calculate the topological complexity index of vascular networks based on morphological feature parameters and output developmental stage labels through a developmental stage classification model. The longitudinal tracking module is used to perform image registration on retinal images of the same child at different examination time points and to calculate the longitudinal variation characteristics of the vascular network. The report generation module is used to generate individualized retinal development assessment reports by integrating morphological feature parameters, developmental stage labels, and longitudinal variation features.

Images