Corneal neovascularization automatic quantitative analysis and grading method, system and terminal based on image processing

Abstract

The invention discloses a corneal neovascularization automatic quantitative analysis and grading method, system and terminal based on image processing. The problems that in the existing corneal neovascularization evaluation process, a complete automatic technical system from image input, blood vessel segmentation to clinical quantitative scoring is lacked, a large amount of manual intervention is needed in the actual process, and high-efficiency and standardized automatic evaluation cannot be achieved are mainly solved. The invention provides a corneal neovascularization automatic quantitative analysis and grading method, system and terminal based on image processing, and provides a set of end-to-end automatic analysis system to realize a complete process of automatic CoNV segmentation, quantitative analysis and clinical structured scoring, so that the objectivity, consistency and efficiency of CoNV evaluation can be effectively improved, and the method and system are suitable for popularization and application. The obvious clinical application value and popularization potential are realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent ophthalmic diagnostic technology, specifically to an automatic quantitative analysis and grading method, system, and terminal for corneal neovascularization based on image processing. Background Technology

[0002] The cornea is an avascular, transparent tissue, and its transparency is crucial for visual function. Factors such as infection, hypoxia, immune diseases, and inflammation can lead to corneal neovascularization. These pathological vessels can significantly affect transparency and cause long-term visual decline, serving as an important indicator of the severity of many corneal diseases. Clinical slit-lamp examination is the most commonly used assessment method, but it has the following significant limitations:

[0003] 1) Highly subjective and lacking unified standards Currently, the clinical assessment of the severity of corneal neovascularization (CoNV) primarily relies on ophthalmologists' visual observation and experience under a slit lamp. Doctors typically make a comprehensive judgment based on multiple subjective indicators, including the number of quadrants in which the vessels are distributed, the maximum radial length of the vessels extending centrally, the vessel thickness, morphological characteristics, and the depth of the vessels that can only be discerned under ideal conditions. However, this assessment method is highly dependent on individual experience and observational perspective; different doctors may have significant differences in their judgment criteria, sensitivity, and interpretation of vessel morphology. For example, for the same image, one doctor might consider the vessels to extend into the mid-peripheral region, while another doctor might consider them to have entered the central region; regarding the extent of quadrant involvement, some doctors tend to include only significant vessels, while others may include small, superficial vessels in their calculations. Such inconsistencies in subjective judgment accumulate into scoring biases during clinical follow-up, making objective comparisons across time points difficult.

[0004] Furthermore, traditional manual record-keeping typically relies on textual descriptions or qualitative grading methods (such as mild, moderate, and severe), lacking pixel-level or structured geometric indicators. This makes it difficult for subsequent assessments to accurately reproduce the original state and hinders the quantitative tracking of disease progression. In scenarios involving multiple follow-ups, treatment efficacy evaluations, and research data collection, the lack of unified standards and quantitative output not only affects reliability but also limits data reusability and large-scale statistical analysis capabilities. Therefore, establishing an automated, objective CoNV assessment system with standardized quantitative output has become a key requirement for addressing clinical challenges.

[0005] 2) Traditional deep learning segmentation methods cannot directly output clinical grading. Existing AI research on CoNV (CoV) primarily focuses on pixel-level segmentation of vascular regions. Typical methods include feature extraction structures based on Convolutional Neural Networks (CNNs), U-Net and its variants (such as Attention U-Net, R2U-Net, etc.), and other medical image segmentation models. These methods can identify vascular pixels in slit-lamp images to a certain extent and generate vascular probability maps or binary segmentation maps. However, their functionality is usually limited to the level of detecting "where the blood vessels are" and cannot directly provide doctors with the structured, high-level quantitative information required for clinical assessment.

[0006] First, these methods only output pixel-level vascular masks, lacking a deep understanding of the overall spatial distribution, orientation patterns, and lesion geometry of blood vessels. The two most crucial indicators in clinical diagnosis—the number of affected quadrants and the maximum radial infiltration depth of blood vessels into the central region—are high-level geometric and positional features, not simply the area or number of pixel blocks. Without geometric modeling and spatial structure analysis, relying solely on pixel-level segmentation cannot directly derive these clinically significant indicators.

[0007] 3) Lack of automated end-to-end analysis systems that can be directly applied to clinical workflows Current research on the automation of corneal neovascularization largely focuses on single tasks, such as pixel-level segmentation of vascular regions or feature extraction of local areas. Existing technologies have not yet established a complete automated workflow from image input, automatic corneal region identification, fine vascular segmentation, geometrically structured modeling, to clinical quantitative grading. While some studies can generate vascular masks, manual intervention is still required to complete quadrant division, determine centripetal infiltration depth, and score severity, failing to achieve a truly end-to-end automated analysis system.

[0008] Furthermore, real-world clinical scenarios require stable, reproducible, and interpretable quantitative results. However, existing methods lack mechanisms to automatically correlate segmented outputs with structured clinical quantitative indicators, making them unsuitable for direct use in diagnostic records, follow-up comparisons, or efficacy assessments. The lack of such end-to-end automated analysis systems hinders the large-scale clinical application and promotion of quantitative analysis of CoNV. Summary of the Invention

[0009] To overcome the shortcomings of the prior art, this invention provides an automatic quantitative analysis and grading method, system and terminal for corneal neovascularization based on image processing. This invention mainly solves the following core technical problems existing in the current corneal neovascularization assessment process: the lack of a complete automated technology system from image input, vessel segmentation to clinical quantitative scoring, and the need for a large amount of manual intervention in the actual process, which makes it impossible to achieve high-efficiency and standardized automatic assessment.

[0010] The technical solution of this invention is an automated quantitative analysis and grading method for corneal neovascularization based on image processing, comprising the following steps: S1: Acquire and preprocess slit lamp images; S2: Input the preprocessed slit lamp image into the corneal region segmentation network to obtain the corneal region mask; S3: The preprocessed slit-lamp image is overlaid with a corneal region mask and then input into a vascular region segmentation network to obtain the corneal neovascularization prediction results (image, see appendix). Figure 2 (middle column C); S4: Based on the corneal region mask, perform ellipse fitting to obtain the geometric parameters of the cornea, and divide the cornea into multiple quadrants and multiple concentric regions based on the geometric parameters; for each quadrant, calculate the total score for that quadrant based on the predicted corneal neovascularization results within that quadrant and the deepest region reached; specifically: Ellipse fitting is performed based on a corneal region mask to obtain a least-squares fitted ellipse for the corneal region. The ellipse is then divided into four quadrants based on its major and minor axes. Two new concentric ellipses are constructed based on the trisection points of the semi-major and semi-minor axes, dividing the original ellipse into three regions: the peripheral region, the mid-peripheral region, and the central region. For each quadrant, the total score is calculated based on the presence of corneal neovascularization prediction results within that quadrant and the deepest region reached by the predicted neovascularization within that quadrant (i.e., the region containing the closest pixel to the center point of the ellipse within that quadrant). S5: Summarize the scores from all quadrants to generate the final score, which is the quantitative score of the severity of corneal neovascularization, and output the visualization results (which can be a chart).

[0011] Both the corneal region segmentation network and the vascular region segmentation network are encoder-decoder structures based on U-Net.

[0012] The blood vessel region segmentation network is trained using a composite loss function consisting of a weighted combination of Focal Loss, Masked Tversky Loss, and clDice Loss; specifically... Loss_total = α Focal Loss + β Masked Tversky Loss + γ clDice Loss; Where α=0.2, β=0.6, and γ=0.2 are weighting coefficients.

[0013] The severity quantification scoring rule in S5 is to calculate the total score for each quadrant of S4. The scoring rules for each quadrant of S4 are as follows: If new blood vessels are present in a certain quadrant, the basic score for that quadrant is recorded as 1 point; If the nearest pixel to the center of the ellipse in the quadrant is located in the outer perimeter, middle perimeter, or central perimeter, then add 1, 2, or 3 points respectively. The system calculates and sums the results for each of the four quadrants, ultimately obtaining a quantitative score for severity ranging from 1 to 16.

[0014] In the corneal region segmentation network, low-level texture features and high-level semantic features are extracted through convolution operations. The ReLU activation function is used to improve non-linear expressive power. The encoder achieves spatial downsampling through max pooling to obtain multi-scale contextual information. The decoder recovers the spatial dimension by upsampling layer by layer and uses skip connections to directly pass the feature mapping of the corresponding layer of the encoder to the decoder, so that the network has both local details and global structural information.

[0015] In the vascular region segmentation network, the intermediate layers employ higher-density feature channels (the number of feature channels for the five scales of the U-shaped network is set to 64, 128, 256, 512, and 1024 respectively), and the convolutional receptive field is expanded (two 3×3 dilated convolutions are used in the deepest layer of the network, with dilation rates of 2 and 4, respectively, whose equivalent receptive fields are approximately 5×5 and 9×9 pixels) to capture the extension patterns of blood vessels. Simultaneously, during inference, the network overlays a corneal region mask onto the preprocessed slit-lamp image, focusing the network on the internal corneal region. The network output is a vascular prediction probability map. Through thresholding and post-processing (e.g., binarizing the vascular prediction probability map with a 0.5 threshold, and combining a first-order morphological closing operation of the 3×3 structuring element and connected component area filtering to remove isolated noise points and compensate for minor breaks), the corneal neovascularization prediction results can be further obtained for subsequent geometric analysis and clinical scoring.

[0016] An automated quantitative analysis and grading system for corneal neovascularization based on image processing, comprising: The image preprocessing module is used to acquire and preprocess slit lamp images; A corneal region segmentation module, including a corneal region segmentation network, is used to segment corneal regions from a preprocessed image to obtain a corneal region mask; The vascular region segmentation module includes a vascular region segmentation network, which is used to overlay the preprocessed slit lamp image with the corneal region mask to obtain a corneal region image, and input it into the vascular region segmentation network to obtain corneal neovascularization prediction results. The geometric modeling and hierarchical module is used to perform the following operations: Based on the corneal region mask, ellipse fitting is performed and quadrants and concentric regions are divided; The total score for this quadrant is calculated based on the predicted corneal neovascularization within the quadrant and the deepest region reached. A severity score is generated by combining the total scores from all quadrants. The results output module is used to output the scoring and visualization analysis results.

[0017] A terminal device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the above-described image processing-based automatic quantitative analysis and grading method for corneal neovascularization.

[0018] The beneficial effects of this invention are as follows: This invention provides an automatic quantitative analysis and grading method, system and terminal for corneal neovascularization based on image processing. It proposes an end-to-end automated analysis system to realize a complete process of automatic CoNV segmentation, quantitative analysis and clinical structured scoring, which can effectively improve the objectivity, consistency and efficiency of CoNV assessment, and has significant clinical application value and promotion potential. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the system flow of an embodiment of the present invention, illustrating the overall workflow of a method for automatic quantitative analysis and grading of corneal neovascularization based on image processing, including data acquisition, annotation, data augmentation, model training and inference, and final grading output. Figure 1 As shown, the present invention first acquires raw images from the historical case database, namely corneal neovascularization slit-lamp images, through a data collection unit. Figure 1 A), and the initial annotation was completed independently by two ophthalmologists ( Figure 1 B), and then the final decision is made by a senior ophthalmologist. Figure 1 C), while quality control is performed by another physician ( Figure 1 D), to ensure the accuracy of the true labels of the corneal region and corneal neovascularization, this annotation step yields 200 grayscale images of the true labels of the corneal region and corneal neovascularization. Figure 1E). Subsequently, the system performs various data enhancement processes on the original image, including photometric enhancement (such as random gamma correction, color jitter, random Gaussian blur, etc.) and geometric augmentation (such as random horizontal flipping, random rotation, random small-angle rotation, and random scaling, etc.). Figure 1 F) is used to improve the robustness of the model under different imaging environments. The data-augmented images are used to train the automatic corneal neovascularization grading model. This invention employs a U-Net-based deep learning structure during the model training phase (F). Figure 1 G), using a five-fold cross-validation method, 200 annotated blood vessel images were trained and tested. Figure 1 H), to fully utilize limited data and improve the model's generalization ability. During the inference phase, the model can automatically complete corneal region identification, blood vessel segmentation, ellipse fitting, structured region partitioning, and hierarchical calculation. A typical example is shown in the lower part of the figure ( Figure 1 The input raw slit-lamp image, after being processed by the model, outputs the following from left to right: an augmented corneal region image, a grayscale image of the corneal region and corneal neovascularization with accurate labels, a predicted grayscale image of the corneal region and corneal neovascularization, and a comprehensive visualization interface including ellipse fitting, region segmentation, and the final scoring results. This invention, through this end-to-end automated analysis system, achieves automated detection, quantitative analysis, and clinical-grade grading assessment of corneal neovascularization.

[0020] Figure 2 The illustrations show typical examples of automatic corneal region segmentation, automatic corneal neovascularization segmentation, and automatic severity grading according to an embodiment of the present invention, demonstrating the applicability of the method in various clinical scenarios. Figure 2 As shown, the left side is the original slit-lamp image, and the four columns on the right, from left to right, show the complete processing flow of the CoNV-AutoGrader automatic analysis system of this invention, including: (column a) the corneal region image after data augmentation; (column b) the grayscale image of the corneal region and corneal neovascularization with real labels obtained by clinical experts; (column c) the grayscale image of the corneal region and corneal neovascularization obtained by the model prediction of this invention; and (column d) the automated severity rating map generated based on the ellipse fitting, quadrant division and three concentric zone calculation rules of the corneal region image. Figure 2The example cases cover several representative clinical scenarios, including: (Row A) Simple cases: fewer corneal neovascularizations, limited distribution, and weak image interference. This invention can accurately identify the corneal contour and capture small corneal neovascular structures. (Row B) Moderately complex cases: a moderate number of corneal neovascularizations, distributed radially or circumferentially, with local interwoven structures. This invention can still output topologically continuous and clearly defined corneal neovascularization prediction results in these cases. (Row C) Highly complex cases: dense corneal neovascularizations, complex branching, and disordered directions, accompanied by complex imaging factors such as corneal opacity or high light reflection. This invention can still effectively separate dense corneal neovascularizations, maintain their connectivity, and preserve their overall structural features, thereby obtaining clearly defined corneal neovascularization prediction results. (Row D) Cases of uneven or lateral corneal neovascularization distribution: corneal neovascularizations are mainly concentrated in one quadrant or exhibit obvious asymmetry. This invention can still accurately obtain corneal neovascularization prediction results through a vascular region segmentation network, and accurately calculate the centripetal infiltration depth of each quadrant under irregular distribution using a geometrically structured grading module, and output corresponding scores. (Line E) Cases with irregular corneal boundaries: Abnormal edge morphology caused by pathological factors such as scarring, postoperative changes, or pterygium. This invention can still accurately obtain accurate corneal region masks through a corneal region segmentation network. Furthermore, it can construct stable structured region divisions under irregular boundaries using an ellipse fitting method. (Line F) Post-corneal transplant cases: Including images with corneal sutures, corneal edema, or corneal graft opacity. This invention relies on a region of interest (ROI) constraint mechanism to effectively exclude interference from non-corneal structures such as sutures, and accurately identify corneal neovascularization growth after corneal transplantation. (Line G) Cases with other ocular lesions: Even under complex pathological conditions such as corneal perforation and corneal ulceration, this invention, relying on the ROI restriction mechanism, can still extract the effective corneal area and identify corneal neovascularization, generating a stable automatic score. Overall, Figure 2 This fully demonstrates that the method of the present invention can output consistent and reliable corneal region and corneal neovascularization segmentation results in a variety of complex clinical scenarios, and generate interpretable structured scores, which can be widely applied to corneal neovascularization assessment in real clinical environments.

[0021] Figure 3 The image shows the corneal region segmentation performance results of an embodiment of the present invention, demonstrating the stability and high accuracy of the proposed CoNV-AutoGrader model's corneal region segmentation network in five-fold cross-validation. Figure 3Table 1 shows that this invention achieved excellent quantitative indicators in corneal region segmentation: in the five-fold validation, the Dice coefficients of each fold remained between 0.968 and 0.976, with an average of 0.971 ± 0.003; the Area Under the Curve (AUC) was close to 0.99 in all five folds, with an average of 0.992 ± 0.003; other indicators such as F1, accuracy, sensitivity, and specificity also remained at a high level, fully demonstrating that the corneal region segmentation network of this invention has excellent stability and generalization ability under various clinical imaging conditions. Furthermore, Figure 3 Table 2 further presents a horizontal comparison of the CoNV-AutoGrader model's corneal region segmentation network with several classic baseline models (including Attention U-Net, R2 U-Net, and U-Net) on the corneal region segmentation task. As can be seen from the figure, the corneal region segmentation network proposed in this invention achieves or surpasses the performance of existing classic baseline models in several key performance indicators such as Dice coefficient, accuracy, and area under the curve on the corneal region segmentation task. Overall, Figure 3 This demonstrates that the corneal region segmentation network of the present invention can maintain stable and accurate corneal region segmentation capabilities even under complex clinical conditions such as reflection, highlights, corneal opacity, and post-corneal transplantation changes, providing a reliable foundation for subsequent ellipse fitting, structured region division, and quantitative analysis of neovascularization.

[0022] like Figure 4 As shown, the proposed CoNV-AutoGrader model for corneal neovascularization segmentation demonstrates significant quantitative advantages in corneal neovascularization segmentation tasks. Figure 4 Table 3 shows that, under the same dataset and annotation conditions, the CoNV-AutoGrader model of this invention, used for corneal neovascularization segmentation, achieved a stable Dice coefficient in five-fold cross-validation for corneal neovascularization segmentation tasks. The coefficient remained within the range of 0.386–0.405, with an average of 0.400 ± 0.008. The average area under the curve reached 0.940 ± 0.005, indicating that the model maintains good discrimination ability even under complex backgrounds, weak contrast, and extremely sparse blood vessels. Other metrics, such as accuracy, F1 score, precision, sensitivity, and specificity, all exhibited consistent high stability, demonstrating the robustness of the blood vessel segmentation network of this invention. Furthermore, Figure 4Table 4 further illustrates the horizontal performance comparison between the CoNV-AutoGrader model's corneal neovascularization segmentation network and various classic baseline models on the corneal neovascularization segmentation task, clearly demonstrating the outstanding advantages of this invention in the vascular segmentation task. Under the same dataset, the same annotation system, and the same training strategy, the Dice coefficient of U-Net is only about 0.202, the Dice coefficient of Attention-UNet is about 0.210, while the Dice coefficient of the CoNV-AutoGrader model's corneal neovascularization segmentation network of this invention can reach 0.400, representing a performance improvement of nearly two times. This difference particularly illustrates the significant advantages of the method of this invention in the recognition of small, sparse, discontinuous, and low-contrast corneal neovascularization structures, which is the key technical value of this invention in solving such medical image problems. Overall, Figure 4 This demonstrates that the CoNV-AutoGrader model of this invention, which is a vascular region segmentation network, not only significantly outperforms existing models in terms of quantitative metrics in the challenging task of automatic segmentation of corneal neovascularization, but also maintains highly consistent and reproducible segmentation performance under complex clinical conditions, providing a solid technical foundation for subsequent calculation of centripetal infiltration depth and clinical severity grading.

[0023] Figure 5 This is a schematic diagram illustrating the severity grading of corneal neovascularization according to an embodiment of the present invention, showing the quantitative scoring results generated by the present invention based on quadrant involvement and the depth of centripetal vascular infiltration. Figure 5 As shown, this invention presents two representative cases of corneal neovascularization. In Case A, neovascularization was observed in all four quadrants (4 points). Based on the centripetal extension of the longest vessel in each quadrant, the upper right, upper left, lower left, and lower right quadrants corresponded to the peripheral zone (2 points), central zone (3 points), peripheral zone (2 points), and peripheral zone (2 points), respectively, resulting in a severity score of 13 points. In Case B, vessels were observed in only three quadrants (3 points). The centripetal extension depth in the upper right, lower left, and lower right quadrants corresponded to the peripheral zone (1 point), central zone (3 points), and central zone (3 points), respectively, resulting in a severity score of 10 points calculated by the system. These examples illustrate that the grading system of this invention can generate intuitive and interpretable quantitative results based on vessel distribution and centripetal invasion depth. Detailed Implementation

[0024] The present invention will be further described below with reference to the accompanying drawings: Step S1: Image acquisition and preprocessing.

[0025] First, anterior segment images acquired using a slit-lamp imaging device are obtained. Exemplary images are typically in PNG or JPG format with a resolution of approximately 2592×1944 pixels. This application does not impose any special restrictions on the acquired anterior segment images. Subsequently, image preprocessing is performed, including image cropping, scaling, and enhancement. To adapt to the input format of the deep learning model, the anterior segment image is cropped or scaled to a standard size of 512×512 pixels and then normalized. After normalization, a series of enhancement operations are performed, including photometric enhancement (such as random gamma correction, color jitter, random Gaussian blur, etc.) and geometric expansion (such as random horizontal flipping, random rotation, random small-angle rotation, and random scaling, etc.) to simulate lighting variations and imaging differences in the actual clinical environment. Through these enhancement operations, preprocessing effectively improves the robustness of the model under conditions of uneven exposure, corneal reflection, imaging noise, and slight displacement. The preprocessed image will serve as input for the subsequent corneal region segmentation module.

[0026] Step S2: Corneal region segmentation.

[0027] The preprocessed image is input into the CoNV-AutoGrader model's corneal region segmentation network, which is based on the U-Net encoder-decoder architecture and aims to automatically identify corneal regions from complex backgrounds. The encoder part of the network obtains multi-scale semantic features through convolution and downsampling, while the decoder part recovers spatial details through upsampling and skip connections. The network performs binary classification prediction on each pixel in the preprocessed image, outputting a corneal region mask. This corneal region mask is crucial for subsequent steps, serving as a constraint mechanism for the vascular region segmentation network. It masks background interference from non-corneal regions, ensuring that vascular segmentation always occurs within the effective corneal region.

[0028] Step S3: Segmentation of the vascular region.

[0029] To accurately identify corneal neovascularization regions, the preprocessed image is overlaid with the corneal mask generated in step S2 to obtain an image containing only the corneal region. This image is then input into the CoNV-AutoGrader model's U-Net-based vascular region segmentation network, which is specifically designed to identify long, sparse, and topologically complex corneal neovascularizations, yielding corneal neovascularization prediction results. Considering that blood vessel widths are typically only a few pixels, to improve the network's sensitivity to small targets, this invention employs a block-based processing strategy during the training phase, dividing the 512×512 image into multiple 256×256 sub-blocks to enhance local detail learning capabilities. Furthermore, the model design utilizes a combined loss function to ensure the coherence of the vascular structure, reducing breaks and missed detections. Finally, the network outputs a probability map of the vascular region and the corresponding binarized corneal neovascularization prediction results. The block processing strategy includes: (1) dividing the 512×512 corneal image into several blocks of 256×256; (2) counting whether each block contains a blood vessel label and distinguishing them into "blood vessel blocks" and "background blocks"; (3) inputting several blocks from the same batch into U-Net for training, so that the network can learn the texture of fine blood vessels in the local high-resolution field of view; (4) using a weighted random sampler to increase the sampling weight of blocks containing blood vessels, thereby achieving oversampling of the fine blood vessel region.

[0030] Step S4: Geometric structuring calculation.

[0031] After obtaining the corneal region mask and corneal neovascularization prediction results, this invention further performs geometric modeling of the corneal morphology to transform pixel-level information into structured spatial features required for clinical scoring. First, edge contour points are extracted from the corneal mask, and the least squares method is used to fit an ellipse to these points, obtaining the corneal center coordinates, major axis length, minor axis length, and ellipse rotation angle. Then, by scaling the fitted ellipse equidistantly at 1 / 3 and 2 / 3 ratios, three regions—peripheral, mid-peripheral, and central—are constructed. These three regions correspond to different grades of centripetal infiltration depth used clinically.

[0032] Using the center point of the fitted ellipse as a reference, the cornea is automatically divided into four quadrants—upper left, lower left, upper right, and lower right—through the horizontal and vertical axes. Finally, for all vascular pixels within each quadrant, the deepest region reached by the predicted corneal neovascularization within that quadrant (i.e., the region containing the pixel closest to the center point of the ellipse) determines whether the vascularization in that quadrant belongs to the peripheral, intermediate, or central region.

[0033] Step S5: Automated clinical grading and terminal output and visualization.

[0034] Based on the quadrant distribution and centripetal infiltration depth obtained in step S4, this invention constructs an automatic grading rule according to the clinically accepted corneal neovascularization severity scoring system. Specifically, if neovascularization exists in a certain quadrant, the basic quadrant score is 1 point; if the longest radial infiltration of the vessel in that quadrant is located in the peripheral, mid-peripheral, or central area, an additional 1, 2, or 3 points are added, respectively. The system calculates and sums the scores for each of the four quadrants (see [link to relevant documentation]). Figure 5 This process yields a quantitative score for the severity of corneal neovascularization, ranging from 1 to 16 points. This score corresponds one-to-one with a clinical manual scoring system and can be used in various clinical scenarios such as disease severity assessment, postoperative follow-up, and treatment effectiveness evaluation.

[0035] Finally, the system outputs the aforementioned calculation results in a structured format, including corneal segmentation maps, vascular segmentation maps, fitted ellipses, quadrant division maps, tri-concentric region division maps, and automatic scoring results. Simultaneously, the system generates multi-layered overlay visualization images that are easy for clinicians to read, allowing them to directly observe the location, extent, and scoring criteria of blood vessels. Furthermore, the system of this invention has excellent engineering deployment capabilities. It can be deployed on hospital servers or Picture Archiving and Communication Systems (PACS) for batch analysis, or on mobile applications, cloud platforms, or portable slit-lamp terminals to meet the needs of various scenarios such as follow-up monitoring, primary care screening, or remote diagnosis and treatment.

[0036] The CoNV-AutoGrader model of this invention employs a deep learning segmentation system, used for corneal region recognition and neovascularization region detection, respectively. Both networks are built based on the U-Net structure, and through an encoder-decoder symmetric architecture and a multi-scale feature fusion mechanism, it achieves high-precision segmentation of complex, subtle, and topologically sensitive target regions in slit-lamp images.

[0037] In the corneal region segmentation network, low-level texture features and high-level semantic features are extracted through convolutional operations. The ReLU activation function is used to improve non-linear expressive power, and max pooling is used to achieve spatial downsampling to obtain multi-scale contextual information. The decoder part restores the spatial dimension by upsampling layer by layer, and uses skip connections to directly pass the feature maps of the corresponding layers of the encoder to the decoder. This allows the network to have both local details and global structural information, thus maintaining the stability of region boundary recognition under challenging conditions such as uneven illumination, corneal reflection, and opacity.

[0038] The vascular region segmentation network adopts a U-Net backbone structure similar to the corneal segmentation network, but with targeted optimizations in feature channel settings and training objectives. Given the extremely narrow, elongated, complex branching structures and blurred boundaries of blood vessels, the network uses higher-density feature channels in the intermediate layers and expands the convolutional receptive field to capture the extension patterns of blood vessels. Simultaneously, during inference, the network applies corneal mask constraints to the input image, focusing the network on the internal corneal region and effectively avoiding false detections caused by background structures (such as eyelash shadows and specular reflections). The network output is a vascular probability map, which can be further processed through thresholding and post-processing to obtain a binary vascular mask for subsequent geometric analysis and clinical scoring.

[0039] In summary, this invention not only achieves division of labor and cooperation in task functions, but also achieves complementary enhancement in information flow design, which helps to improve the accuracy, stability and interpretability of corneal and blood vessel segmentation tasks.

[0040] Composite loss function system: To ensure stable performance of the model in the highly sparse, class-imbalanced, and topology-sensitive slit-lamp vessel segmentation task, this invention novelly designs a composite loss function system consisting of Focal Loss, Masked Tversky Loss, and clDice Loss. This system complements each other at different training stages and in different segmentation dimensions, significantly improving the model's overall representation ability for small targets, continuous structures, and irregular vascular regions.

[0041] Focal Loss Because blood vessel pixels constitute a very small proportion of images with a very large proportion of background pixels, directly applying conventional cross-entropy loss would result in the gradient being dominated by the background, making it difficult for the model to effectively learn blood vessel features. Therefore, this invention applies Focal Loss, which introduces a modulating factor to reduce the loss weight of easily classified background pixels, thereby increasing the model's attention to difficult-to-classify blood vessel pixels. In the early stages of training, Focal Loss can quickly enhance the model's sensitivity to foreground regions; in the later stages of training, it further suppresses background interference, helping to improve blood vessel recall.

[0042] Masked Tversky Loss In vessel segmentation tasks, issues such as blurred boundaries and local breaks are common. If only pixel-level consistency metrics are relied upon, the model is prone to an improper trade-off between false positives (FP) and false negatives (FN). Masked Tversky Loss introduces adjustable parameters to control the impact of FP and FN on the loss, allowing the model to dynamically adjust the penalty intensity based on the actual state of the vessel boundaries. For example, for regions with many branch ends, the FN penalty can be appropriately increased to reduce missed detections; for reflective, bright regions, the FP penalty can be increased to reduce the misidentification of light spots as vessels. By overlaying corneal region masks on the preprocessed image, this loss function only operates within the effective corneal area, improving the targeting of loss optimization.

[0043] clDice Loss A key feature of vascular structure lies in its connectivity. This invention employs clDice Loss, a loss term specifically designed for elongated topologies, to enhance the structural integrity of vascular segmentation. By simultaneously comparing the centerline consistency between predicted and actual vessels, clDice Loss ensures that the network learns not only the pixel regions of the vessels but also their connected topological features. This loss is highly effective in compensating for breaks, bridging vessel interruptions, and maintaining the integrity of vessel orientation, representing a key innovation in improving the stability of clinical quantitative CoNV results.

[0044] Weighted combination strategy: This invention achieves comprehensive optimization of pixel accuracy, boundary accuracy, and topological continuity by weighting and combining the three loss functions mentioned above in a specific ratio. The combination strategy can be expressed as: Loss_total = α Focal Loss + β Masked Tversky Loss + γ clDice Loss Here, α, β, and γ are adjustable weights (specifically, α=0.2, β=0.6, γ=0.2 can be selected), dynamically set according to the training stage and data characteristics. Through this combined strategy, the network can balance accurate detection of blood vessel pixels (Focal), accurate edge segmentation (Tversky), and continuity of structural topology (clDice), thereby obtaining highly stable and robust blood vessel segmentation results, providing a reliable foundation for subsequent geometric structuring calculations and clinical grading.

[0045] The system for implementing the methods provided in this application specifically includes the following modules: ① Image preprocessing module ② Corneal region segmentation module ③ Vascular region segmentation module ④ Geometric modeling and grading module, including geometric fitting module and grading calculation module ⑤ Result output module ⑥ May also include terminal communication and storage module ① Image preprocessing module This module is used to receive anterior segment image data from slit-lamp microscopes, portable anterior segment imaging devices, or mobile imaging devices. It supports multi-format and multi-resolution image input, including PNG, JPG, and DICOM, and can synchronize data with hospital image archiving and communication systems / hospital information systems (PACS / HIS) or receive real-time image streaming. The module can be configured with an image quality detection subunit to automatically determine blurriness, overexposure, and the extent of highlight reflection, thus filtering out images that do not meet analysis standards.

[0046] The module performs normalization processing on the input image, including resizing, cropping, normalization, and enhancement, to adapt it to the input requirements of subsequent depth models. It can perform a series of optical and geometric enhancement operations, including brightness adjustment, contrast perturbation, gamma correction, color perturbation, random rotation, scaling, mirroring, and Gaussian blur. This module ensures that the image retains stable feature representation under different lighting conditions, imaging devices, and shooting angles, providing a consistent input basis for subsequent corneal segmentation, vessel detection, and geometric calculations.

[0047] ② Corneal region segmentation module The corneal region segmentation module employs a U-Net-based deep learning network to identify corneal regions from complex backgrounds using a pixel-level binary classification method. This module outputs a binary corneal mask and can further extract the positions of corneal edge contour points to support subsequent ellipse fitting calculations. Through skip connections, multi-scale feature fusion, and data augmentation strategies, this module maintains high robustness even in the presence of eyelash occlusion, reflective points, and corneal opacity. Its output serves as a crucial basis for the system's ROI constraint rules, providing spatial constraints for vessel detection.

[0048] ③Vascular region segmentation module The vascular region segmentation module receives the image generated by the image preprocessing module and the corneal mask generated by the corneal region segmentation module. By superimposing these two images, an input image containing only the corneal region is obtained, which is then fed into a U-Net-based vascular segmentation network for pixel-level vascular prediction. This module employs a block-based training strategy and a composite loss function system (including Focal Loss, Masked Tversky Loss, and clDice Loss) to ensure the coherence and integrity of the vascular structure even under conditions of sparse corneal neovascularization. Finally, it outputs a vascular probability map and binary vascular prediction results, providing basic data for clinical grading.

[0049] ④ The geometric modeling and grading module includes a geometric fitting module and a grading calculation module. The geometric fitting module extracts the contours of the corneal region mask edges and reconstructs the corneal elliptical model using least-squares fitting or robust fitting algorithms, obtaining geometric parameters such as the corneal center point, major axis length, minor axis length, and rotation angle. Based on the fitted ellipse, this module further generates multiple concentric regions (peripheral, intermediate, and central regions) and constructs a four-quadrant structure based on the ellipse center. This module is responsible for converting the segmentation output into a structured geometric space, providing a clear geometric basis for subsequent radial distance calculations and clinical grading.

[0050] The grading calculation module, based on the vessel prediction results output by the vessel region segmentation module and the quadrant and concentric region structures generated by the geometric fitting module, traverses the normalized radial distance of vessel pixels within each quadrant, extracts the maximum radial invasion depth of that quadrant, and assigns a corresponding grade score according to the region (peripheral, mid-peripheral, or central) where that depth is located. Simultaneously, when any vessel is present in a quadrant, that quadrant receives a base score of 1 point. The scores from the four quadrants are accumulated to form the final CoNV severity score (1–16 points). This module can directly output structured indicators consistent with clinical scoring systems, achieving automated clinical grading assessment.

[0051] ⑤ Result Output Module The results output module is responsible for presenting the system analysis results to clinical users in a graphical and structured manner. The module provides overlay views including the original image, corneal mask, vascular prediction results, fitted ellipse contour, quadrant division diagram, three concentric regions, and scores, enabling physicians to intuitively understand the scoring criteria. This module also outputs structured report data (such as JSON, XML, or DICOM SR format) for easy clinical archiving, follow-up comparisons, and research statistics.

[0052] ⑥ It may also include a terminal communication and storage module. The terminal communication and storage module enables system deployment, data interaction, and storage in multi-terminal environments. The module supports communication interfaces with hospital PACS / HIS, ​​cloud servers, mobile applications, and handheld slit lamp devices, and data transmission can be achieved via protocols such as HTTP, HTTPS, DICOM, and WebSocket. Simultaneously, the module provides local or cloud-based result caching, database storage, data encryption, and privacy protection mechanisms, supporting automatic switching between local and cloud-based model inference to ensure stable system operation in environments with good network conditions, limited network access, or complete offline operation.

[0053] Implementation effect Excellent corneal region segmentation performance like Figure 3 As shown, the corneal region segmentation network in this invention achieved high performance metrics of approximately Dice≈0.971 and AUC≈0.992 on the test set. These results demonstrate that the CoNV-AutoGrader model's corneal region segmentation network exhibits extremely high robustness and generalization ability under complex clinical imaging conditions, such as identifying corneal boundaries, handling reflective points, and corneal opacity. Stable and high-precision corneal segmentation is fundamental for subsequent ellipse fitting, region division, and quantitative vascular analysis. This invention ensures reliable extraction of corneal regions through preprocessing strategies, laying a solid foundation for the entire automated grading system.

[0054] Blood vessel segmentation is significantly improved compared to the classic model. like Figure 4 As shown, under the same dataset and annotation conditions, the blood vessel region segmentation network of this invention achieves a segmentation performance of Dice≈0.400, significantly outperforming many classic baseline models. For example: The Dice of the basic UNet is only about 0.202. The Dice of Attention-UNet is approximately 0.210. The blood vessel segmentation performance of this invention is improved by more than 100%, which is mainly attributed to the following technological innovations: Block-based training improves the model's ability to analyze small blood vessels. Composite loss function system enhances vascular topological continuity ROI constraints improve the ability to detect effective regions and avoid background interference, allowing the model to focus on the real structure. By using a labeled mask, the extracorneal region is blacked out in the input image to remove invalid background textures; In the loss function, pixels with a mask value of 255 are set as ignored labels and do not participate in loss calculation and gradient update; During block sampling, the sampling probability of only containing background blocks is reduced, thereby reducing the dominance of background samples in training; Compared to traditional methods that can only produce unstable and severely fragmented segmentation results, this invention can output more coherent vascular masks with more complete topological structures, providing the possibility for reliable clinical quantification.

[0055] The grading results are highly consistent with the expert scores. This invention utilizes a complete process (segmentation → geometric calculation → automatic scoring) to obtain a CoNV severity index, which shows a high degree of consistency with the manual scoring by experienced clinical experts. The consistency evaluation index reaches: ICC (intra-group correlation coefficient) = 0.885 (p < 0.001). The above p < 0.001 indicates that the result is statistically significant.

[0056] This value falls within the high consistency range in the field of medical scale assessment, indicating that the present invention can not only operate stably at the image segmentation level, but also achieve a level close to expert judgment in the final clinical scoring results, demonstrating feasibility and credibility in clinical decision support.

[0057] In other words, the scoring system of this invention not only "can calculate scores," but also "the calculated scores are consistent with expert understanding." This is a key indicator for the successful implementation of the automated quantitative analysis and grading method, system, and terminal for corneal neovascularization, and this invention fully meets this condition.

[0058] In summary, this invention addresses the challenges of automatic segmentation, quantitative analysis, and clinical grading of CoNV in slit-lamp images, proposing a systematic and innovative solution from multiple dimensions, including deep learning architecture design, image geometric modeling, loss function optimization, workflow engineering integration, and terminal deployment. Specific innovations are as follows: ① CoNV-AutoGrader model deep segmentation architecture with dual-network cascade structure (corneal region segmentation network + blood vessel region segmentation network) This invention proposes a CoNV-AutoGrader model depth segmentation architecture with a dual-network cascade structure (corneal region segmentation network + vessel region segmentation network). It consists of two independent but collaboratively cascaded U-Net networks, used for corneal region segmentation and subsequent neovascularization detection, respectively. The corneal region segmentation network generates a high-precision corneal region mask, which limits the effective area for vessel detection, avoiding misclassification of extracorneal regions, reflective points, and conjunctival vessels as corneal neovascularization. This significantly improves the accuracy of vessel localization, thus forming a mandatory ROI constraint mechanism.

[0059] Compared with traditional single-network multi-class segmentation methods, the dual-network cascaded structure of this invention has the following advantages: Each network is optimized for different objects to avoid feature confusion caused by competition among multiple tasks. Using corneal masks as a prerequisite constraint effectively reduces the sensitivity of the vascular region segmentation network to background noise; Combinatorial prediction methods achieve higher stability and generalization ability.

[0060] This architecture enables the blood vessel segmentation network to focus on the texture information inside the corneal region, significantly improving the accuracy and robustness of CoNV detection, which is a fundamental innovation for improving algorithm performance.

[0061] ② Geometric structured calculation module combined with clinical grading system To address the limitation of existing deep learning methods that can only output pixel masks and cannot directly generate clinically meaningful structured indicators, this invention proposes a complete geometric structured computation module directly corresponding to the clinical CoNV grading system. The module's functions include: An ellipse fitting algorithm based on corneal region mask is used to obtain the geometric center, major and minor axes, and rotation angle of the cornea. The fitted ellipse is automatically divided into four quadrants; Based on the method of elliptical concentric scaling, the cornea is further divided into three radial depth regions: the peripheral zone, the intermediate zone, and the central zone. Calculate the maximum radial invasion distance of the blood vessel pixels within each quadrant and map it to a clinical grade (1–3 points). The final product is a clinically applicable severity index of 1–16 points, which accurately reflects the combination of vascular extent and depth.

[0062] This module is the first to achieve an engineering closed loop that automatically transitions from "pixel-level segmentation" to "clinical grading," enabling deep learning models to not only identify lesions but also provide repeatable, interpretable, and quantitative outputs consistent with clinical diagnostic procedures.

[0063] ③ Composite loss function designed specifically for the topological characteristics of blood vessels Corneal neovascularization is characterized by its extremely fine, sparse, dendritic, and highly interconnected structure, making it difficult for traditional loss functions to guarantee its continuity and integrity. This invention introduces a combination of Focal Loss, Masked Tversky Loss, and clDice Loss, and through weight fusion, forms a composite loss function system specifically designed for CoNV scenarios. Focal Loss: Under the condition of severe imbalance between positive and negative samples, by reducing the weight of easily classified background pixels, the model pays more attention to a very small number of foreground blood vessel pixels. Masked Tversky Loss: By finely controlling the penalty ratio between FP and FN, the model has greater flexibility in boundary recognition and improves the stability of blood vessel edge detection; clDice Loss: This loss function is designed to maintain the topological connectivity of slender structures, which can significantly improve the continuity of blood vessel segmentation and reduce the problems of breaks, jumps, and local missed detections.

[0064] This composite loss system not only solves the class imbalance problem in blood vessel segmentation tasks, but also overcomes the scoring instability caused by blood vessel topological fractures. It is a key breakthrough in this invention that is highly technical and innovative.

[0065] ④ Background suppression mechanism based on ROI mask To reduce false detections caused by complex backgrounds in slit-lamp images, this invention proposes a ROI background suppression mechanism based on corneal segmentation results. The invention achieves the following: Set non-corneal regions to zero or low-weight inputs; In the blood vessel segmentation network, shield high-noise areas such as highly reflective areas and eyelash obstruction; Reduce false vascular predictions caused by a large amount of background.

[0066] This mechanism ensures that the input for blood vessel detection is always focused on physiologically relevant regions, significantly improving the robustness, speed, and accuracy of the model while reducing network complexity and false detection probability, making it a crucial factor in engineering deployment.

[0067] This invention constructs a complete end-to-end structured output pipeline to achieve: Original slit lamp image input Corneal segmentation results Blood vessel segmentation results Fitting Ellipses and Region Partitioning Multiquadrant and radial classification Automatically generate numerical severity index Visual overlay plots and structured data output This process seamlessly integrates all stages from "image recognition → segmentation → quantification → clinical interpretation → visualization." Doctors not only receive the scoring results but also intuitively see the basis for the scoring, achieving high transparency and interpretability, which is beneficial for clinical adoption and regulatory requirements.

[0068] The methods provided in this application embodiment are not limited to electronic devices; they can also be implemented by applications or system services.

[0069] The embodiments described with reference to the accompanying drawings are exemplary and intended to explain the invention, and should not be construed as limiting the invention. The embodiments should not be considered as limiting the invention, but any improvements made based on the spirit of the invention should be within the scope of protection of the invention.

Claims

1. An image processing based automatic quantification and grading method for corneal neovascularization, characterized in that: The method comprises the following steps S1: obtaining and preprocessing a slit lamp image; S2: inputting the preprocessed slit lamp image into a corneal region segmentation network to obtain a corneal region mask; S3: superimposing the preprocessed slit lamp image and the corneal region mask to obtain a corneal region image, and inputting the corneal region image into a blood vessel region segmentation network to obtain a corneal neovascularization prediction result; S4: performing ellipse fitting based on the corneal region mask to obtain geometric parameters of the cornea, and dividing the cornea into a plurality of quadrants and a plurality of concentric regions based on the geometric parameters; For each quadrant, the total score of the quadrant is calculated based on the corneal neovascularization prediction result in the quadrant and the deepest region reached; S5: aggregating the total scores of all the quadrants to generate a severity quantification score of the corneal neovascularization and output a visualization result.

2. The automatic quantitative analysis and grading method of corneal neovascularization based on image processing according to claim 1, characterized in that, In step S4, the cornea is divided into a plurality of concentric regions based on the geometric parameters, specifically comprising: obtaining a least square fitting ellipse of the corneal region, and dividing the ellipse into four quadrants based on the major and minor axes of the ellipse, and constructing two new concentric ellipses according to the three equal points of the semi-major axis and the semi-minor axis to divide the original ellipse into three regions, namely, the peripheral region, the mid-peripheral region and the central region.

3. The automatic quantitative analysis and grading method of corneal neovascularization based on image processing according to claim 1, characterized in that, The corneal region segmentation network and the blood vessel region segmentation network are both encoder-decoder structures based on U-Net.

4. The automatic quantitative analysis and grading method of corneal neovascularization based on image processing according to claim 1, characterized in that, The blood vessel region segmentation network adopts a composite loss function composed of Focal Loss, Masked TverskyLoss and clDice Loss during training; specifically Loss_total = a Focal Loss + b Masked Tversky Loss + g clDice Loss; Wherein, α=0.2, β=0.6, γ=0.2 are weight coefficients.

5. The method according to claim 2, wherein, The quadrant total score evaluation rule in S4 is, If there is neovascularization in a quadrant, the quadrant basic score is 1; If the nearest pixel point to the ellipse center point in the quadrant is located in the peripheral region, the mid-peripheral region or the central region, 1, 2 or 3 points are added respectively.

6. The automatic quantitative analysis and grading method of corneal neovascularization based on image processing according to any one of claims 1-5, characterized in that, In the corneal region segmentation network, the encoder extracts low-level texture features and high-level semantic features through convolution operation, uses the ReLU activation function to improve the non-linear expression ability, and realizes spatial down-sampling through max-pooling to obtain multi-scale context information; the decoder part recovers the spatial dimension through layer-by-layer up-sampling, and utilizes the skip connection to directly transmit the feature mapping of the corresponding layer of the encoder to the decoder, so that the network has both local details and global structural information.

7. The automatic quantitative analysis and grading method of corneal neovascularization based on image processing according to any one of claims 1-5, characterized in that, In the blood vessel region segmentation network, the middle layer adopts a high-density feature channel, and expands the convolution receptive field to capture the extension mode of the blood vessels; at the same time, the network superimposes the corneal region mask on the preprocessed slit lamp image during inference, and the output of the network is a blood vessel prediction probability map, which can be further obtained through thresholding and post-processing to obtain a corneal neovascularization prediction result for subsequent geometric analysis and clinical scoring.

8. An image processing based automatic quantification and grading system for corneal neovascularization characterized in that, It comprises: an image preprocessing module for obtaining and preprocessing a slit lamp image; a corneal region segmentation module comprising a corneal region segmentation network for segmenting a corneal region mask from the preprocessed image; The blood vessel region segmentation module comprises a blood vessel region segmentation network, which is configured to superimpose the preprocessed slit lamp image and the corneal region mask to obtain a corneal region image, and input the corneal region image into the blood vessel region segmentation network to obtain a corneal neovascularization prediction result; The geometric modeling and grading module is configured to perform the following operations: performing ellipse fitting and dividing quadrants and concentric regions based on the corneal region mask; calculating a total score of each quadrant based on the corneal neovascularization prediction result in the quadrant and the deepest region reached; generating a severity quantification score by synthesizing the total scores of all the quadrants; The result output module is configured to output the score and a visual analysis result. 9.A terminal device, comprising a memory and a processor, the memory storing a computer program, characterized in that, The processor implements the computer program to implement the method for automatically quantifying and grading corneal neovascularization based on image processing according to any one of claims 1 to 7.

Images