Optic nerve sheath ultrasonic image analysis system based on deep learning

Abstract

The invention relates to the technical field of medical image processing, and particularly discloses an optic nerve sheath ultrasonic image analysis system based on deep learning, which comprises the following steps: acquiring an eye dynamic video stream through ultrasonic equipment, and automatically screening key frames to construct an initial image set; performing multi-scale registration on the initial image and a standard anatomical template to generate a reference image; reversely deriving a displacement vector field based on an image space corresponding relation; performing smooth optimization and geometric reconstruction on the displacement field under the elastic mechanical constraint of the biological tissue to generate a corrected image; reinforcing boundary features in the gradient domain and completing accurate segmentation through edge connection and region growth; and finally, performing morphological verification and diameter measurement on the segmentation result, and outputting a structured analysis report.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and more specifically to a deep learning-based ultrasound image analysis system for the optic nerve sheath. Background Technology

[0002] Ultrasound measurement of the optic nerve sheath diameter is an important non-invasive method for assessing intracranial pressure and has significant clinical value in neurocritical care and emergency medicine. However, the clinical application of this technology has long faced two major challenges: First, ultrasound examination is highly operator-dependent; the minute relative movements between the probe and the surface of the eyeball can produce imperceptible shearing deformations within the image plane. These deformations do not cause image blurring or structural loss, but they distort the local geometric features of the optic nerve sheath. Second, existing deep learning-based automatic segmentation methods typically treat deformation correction and segmentation as two independent tasks, leading to a discrepancy between the image quality metrics optimized during correction and the semantic features required for segmentation. This inherent contradiction means that traditional methods may destroy boundary features crucial for segmentation during image correction or fail to effectively distinguish between real anatomical structures and deformation artifacts during segmentation, ultimately affecting the accuracy and reliability of diameter measurement.

[0003] This invention addresses the following technical problem: the conflicting optimization objectives of deformation correction networks and segmentation networks at the feature level. Specifically, deformation correction networks use image similarity metrics as their optimization objective, tending to generate visually smooth and natural images. However, this process may erase weak boundary features crucial for segmentation tasks. Segmentation networks, on the other hand, focus on distinguishing different semantic categories and are insufficiently sensitive to boundary features that may be weakened after correction. This conflict at the feature level results in the system's inability to maintain image naturalness while ensuring segmentation boundary accuracy when dealing with common in-plane shearing deformations encountered in clinical practice. This is particularly evident in the transition region between the optic nerve sheath and surrounding tissues, where jagged edges or local breaks in the boundary are prone to occur, severely impacting the accuracy of diameter measurements. Summary of the Invention

[0004] The purpose of this invention is to provide a deep learning-based ultrasound image analysis system for the optic nerve sheath to solve the problems mentioned above.

[0005] The objective of this invention can be achieved through the following technical solutions: A deep learning-based ultrasound image analysis system for the optic nerve sheath includes: The dynamic sequence acquisition and quality screening module is used to acquire real-time dynamic ultrasound video streams of the eye through ultrasound equipment, and automatically screen out key frames containing complete optic nerve sheath anatomical structures based on image structural integrity assessment to form an initial image set. The anatomical structure template matching module is used to perform multi-scale registration between the initial image set and the pre-constructed standard anatomical template of the optic nerve sheath to generate reference images for structural alignment. The deformation field inverse derivation module is used to inversely derive the displacement vector field characterizing the local geometric deformation of the image based on the pixel spatial correspondence between the initial image set and the reference image. The physical constraint geometric correction module is used to smooth and optimize the displacement vector field under the elastic mechanical constraints of biological tissue, and to perform geometric reconstruction of the initial image based on the optimized displacement field to generate a corrected image that conforms to anatomical reality. The boundary enhancement segmentation module is used to enhance the boundary features of the optic nerve sheath in the gradient domain of the corrected image. It extracts continuous boundary features through a multi-scale edge perception algorithm and combines it with a region growing algorithm to complete the accurate segmentation of the optic nerve sheath. The result verification and output module is used to verify the morphological integrity of the segmentation results, calculate the optic nerve sheath diameter measurement based on the verified segmentation mask, and output the final analysis report.

[0006] As a further aspect of the present invention: the specific process of constructing the initial image set includes: The structural sharpness index and boundary continuity index of each frame image are calculated simultaneously. The structural sharpness index is obtained by calculating the weighted sum of gradient amplitudes in the region of interest of the optic nerve sheath, and the boundary continuity index is obtained by detecting the degree of breakage and curvature consistency of the edge of the optic nerve sheath. The filtering threshold is adaptively adjusted based on the quality change trend of adjacent frames in the video sequence. When a significant drop in quality is detected in consecutive frames, the threshold requirement is automatically increased to ensure that the selected key frames remain uniform in temporal distribution and have stable quality. First, frames that obviously do not meet the requirements are eliminated based on rapid pre-screening. Then, the anatomical integrity of the candidate frames is verified by fine-grained verification. The integrity is confirmed by checking whether the spatial relationship between the optic nerve sheath and the surrounding tissues conforms to the anatomical features. Finally, an initial image set of consistent quality is constructed.

[0007] As a further aspect of the present invention: the reference image for generating structural alignment specifically includes: Feature maps with different receptive fields are extracted from the initial image and the standard anatomical template respectively. Local structural similarity measures are calculated at each scale level to establish a coarse-to-fine feature correspondence. An initial deformation field is generated based on the feature correspondence. At the same time, the confidence of feature matching is calculated as the attention weight. The smoothing constraint strength of the deformation field is dynamically adjusted using the attention weight. The flexibility of the deformation field is maintained in the reliable feature matching region, and the smoothing constraint is enhanced in the uncertain matching region. The deformation fields generated at different scales are weighted and fused, with the coarse-scale deformation field providing global structural alignment and the fine-scale deformation field optimizing local detail registration, ultimately generating a reference image that accurately reflects the spatial correspondence of the anatomical structure of the optic nerve sheath.

[0008] As a further aspect of the present invention: the step of inversely deriving the displacement vector field characterizing the local geometric deformation of the image based on the pixel spatial correspondence between the initial image set and the reference image specifically includes: By establishing the spatiotemporal correspondence between key anatomical landmarks between the initial image and the reference image, the motion trajectory of each pixel between the image pairs is calculated, and an initial displacement vector field is constructed based on the continuity constraint of adjacent pixel trajectories. Based on the displacement vector of the key anatomical region, the displacement vector is propagated to the surrounding area according to the similarity of local texture features of the image. In areas with gentle changes in texture features, the propagation range is based on the uniformity of tissue texture, while in areas with drastic changes in texture features, the propagation range is based on the complexity of edge structure. The displacement vector field based on trajectory backtracking and the displacement vector field based on adaptive propagation are weighted and fused. The trajectory backtracking result is given priority in the edge region of the anatomical structure, while the propagation result is given priority in the homogeneous tissue region. Finally, a displacement vector field with spatial continuity and anatomical structure accuracy is generated.

[0009] As a further aspect of the present invention: the smoothing optimization of the displacement vector field under the elastic mechanical constraints of biological tissue specifically includes: An anisotropic smooth constraint matrix was constructed based on the biomechanical properties of the periorbital tissues, wherein the smooth constraint strength along the direction of the optic nerve sheath is less than the constraint strength in the vertical direction, so as to preserve normal anatomical deformation. The strain tensor of each pixel is calculated based on the displacement vector field, and the area that needs to be smoothed is determined according to the strain energy density distribution. While keeping the displacement of key anatomical points constant, the displacement vector field is gradually optimized by minimizing the weighted sum of strain energy and smoothness constraints to obtain a smooth displacement field that conforms to the mechanical properties of biological tissues.

[0010] As a further aspect of the present invention: the generation of the corrected image that conforms to anatomical realism specifically includes: Based on the anatomical features of the optic nerve sheath and surrounding tissues, the image is divided into different reconstruction regions, and different reconstruction parameters are applied to each region. An edge-preserving resampling algorithm is used in the optic nerve sheath boundary region, while a conventional resampling algorithm is used in the homogeneous tissue region. Initial geometric reconstruction is first completed at a lower resolution, and then the resolution is gradually increased to optimize details, generating a corrected image that maintains both the authenticity of the anatomical structure and the preservation of image details.

[0011] As a further aspect of the present invention: the enhancement of optic nerve sheath boundary features within the gradient domain of the corrected image specifically includes: By calculating the gradient magnitude response at different scales, the expression of boundary features from coarse to fine is enhanced, capturing the overall contour at the coarse scale and preserving local details at the fine scale. An anisotropic filter kernel is designed based on the orientation characteristics of the optic nerve sheath boundary, and smoothing is performed along the boundary tangent direction while enhancing the gradient response in the normal direction. The enhancement intensity is dynamically adjusted based on the statistical characteristics of local gradients. The enhancement intensity is increased in the boundary region where the gradient response is weak, while the original characteristics are maintained in the region where the gradient response is strong.

[0012] As a further aspect of the present invention: the step of extracting continuous boundary features using a multi-scale edge perception algorithm and combining it with a region growing algorithm to achieve accurate segmentation of the optic nerve sheath specifically includes: The edge feature maps extracted at different scales are weighted and fused, where coarse-scale edges provide guidance on topological structure and fine-scale edges supplement detailed information. Based on the curvature characteristics and spatial proximity of edge points, the broken edge segments are connected into a complete boundary profile; Using complete boundary contours as constraints, region growth is performed starting from seed points inside the optic nerve sheath. During the growth process, a boundary distance penalty term is used to ensure that the segmentation results are consistent with the enhanced boundary features.

[0013] As a further aspect of the present invention: the output of the final analysis report specifically includes: By calculating the rate of change of boundary curvature and the topological features of connected components of the segmentation mask, the smoothness and continuity of the optic nerve sheath contour are detected, and adaptive repair is initiated for regions with breaks or abnormal mutations. Within the validated optic nerve sheath region, diameter values ​​in multiple directions are calculated at equal angular intervals along its principal axis. Abnormal measurements are eliminated through statistical distribution analysis, and the mode diameter is taken as the final measurement result. Based on the clarity of the segmentation boundary, the consistency of the diameter measurement, and the morphological integrity score, a comprehensive confidence index is generated, and all validation data, measurement results, and confidence assessments are integrated to output a structured analysis report.

[0014] The beneficial effects of this invention are: (1) Traditional methods directly segment and measure the original ultrasound images, ignoring the in-plane shear deformation generated during probe operation, leading to systematic errors in the measurement results. This invention effectively eliminates image deformation caused by differences in operator technique by constructing a standard anatomical template, inversely deriving the displacement vector field, and performing geometric correction under the constraints of biological tissue elasticity. This ensures that images acquired by different operators at different times can meet a unified measurement standard. This correction method reduces the average error in optic nerve sheath diameter measurement and improves the reliability of measurement results in clinical follow-up and efficacy evaluation.

[0015] (2) A complete automated analysis process has been established, achieving standardization of the entire process from image acquisition to report generation, thus improving clinical work efficiency. Traditional optic nerve sheath measurement relies on manual selection and measurement by physicians, which takes 5 to 10 minutes, and the results are greatly affected by the operator's experience. This invention shortens the analysis process through dynamic sequence automatic screening, multi-scale registration, boundary enhancement segmentation, and intelligent verification output, while ensuring the quality of results through a confidence assessment mechanism. Attached Figure Description

[0016] The invention will now be further described with reference to the accompanying drawings.

[0017] Figure 1 This is a flowchart of the system of the present invention; Figure 2 This is a flowchart illustrating the specific implementation of the dynamic sequence acquisition and quality screening module in this invention. Figure 3 This is a flowchart illustrating the specific implementation of the anatomical structure template matching module in this invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 As shown, this invention is a deep learning-based ultrasound image analysis system for the optic nerve sheath, comprising: The dynamic sequence acquisition and quality screening module is used to acquire real-time dynamic ultrasound video streams of the eye through ultrasound equipment, and automatically screen out key frames containing complete optic nerve sheath anatomical structures based on image structural integrity assessment to form an initial image set. The anatomical structure template matching module is used to perform multi-scale registration between the initial image set and the pre-constructed standard anatomical template of the optic nerve sheath to generate reference images for structural alignment. The deformation field inverse derivation module is used to inversely derive the displacement vector field characterizing the local geometric deformation of the image based on the pixel spatial correspondence between the initial image set and the reference image. The physical constraint geometric correction module is used to smooth and optimize the displacement vector field under the elastic mechanical constraints of biological tissue, and to perform geometric reconstruction of the initial image based on the optimized displacement field to generate a corrected image that conforms to anatomical reality. The boundary enhancement segmentation module is used to enhance the boundary features of the optic nerve sheath in the gradient domain of the corrected image. It extracts continuous boundary features through a multi-scale edge perception algorithm and combines it with a region growing algorithm to complete the accurate segmentation of the optic nerve sheath. Please see Figure 2 As shown, the result verification and output module is used to verify the morphological integrity of the segmentation results, calculate the optic nerve sheath diameter measurement value based on the verified segmentation mask, and output the final analysis report.

[0020] In the dynamic sequence acquisition and quality screening module, firstly, a high-frequency linear ultrasound probe is used to acquire real-time dynamic ultrasound video streams of the patient's eye. During acquisition, the operator must keep the probe perpendicular to the surface of the eyeball and slowly scan along the long axis of the optic nerve to ensure complete capture of the longitudinal profile of the optic nerve sheath. Each acquisition lasts 3–5 seconds, acquiring continuous dynamic image sequences at a rate of 30 frames per second to form the raw ultrasound video data. After acquisition, the system automatically decodes the video stream, decomposing it into independent image frame sequences, preparing for subsequent quality assessment and keyframe screening.

[0021] During the image quality assessment phase, the system simultaneously calculates two core quality metrics for each frame: structural sharpness and boundary continuity. The structural sharpness metric is calculated as follows: First, the region of interest (ROI) of the optic nerve sheath is determined in the image. This region is typically located in the lower center of the image and presents as a low-echo, band-like structure. Then, the gradient magnitude of each pixel within this region is calculated. A weighted sum of the gradient magnitudes of all pixels is then used to obtain the sharpness evaluation value, with pixels closer to the center of the optic nerve sheath assigned higher weights and pixels at the edges assigned lower weights. The boundary continuity metric is calculated by detecting the integrity of the optic nerve sheath's edges: First, an edge detection algorithm is used to extract the upper and lower boundaries of the optic nerve sheath. Then, the breakage and curvature changes of the boundary lines are analyzed, the number of discontinuities on the boundary lines and the curvature differences between adjacent boundary segments are calculated, ultimately yielding a quantitative metric reflecting boundary continuity.

[0022] Based on the dynamic characteristics of video sequences, the system employs an adaptive threshold adjustment mechanism for keyframe selection. This mechanism dynamically adjusts the selection criteria by analyzing the quality change trends between adjacent frames: when a declining trend in the quality index of multiple consecutive frames is detected, the system automatically increases the selection threshold to ensure that only the highest quality frames are retained; when the video sequence quality is stable, the baseline selection threshold is maintained. This dynamic adjustment strategy effectively addresses the unavoidable image quality fluctuations during ultrasound scanning, ensuring that the selected keyframes are evenly distributed over time, avoiding excessive concentration of keyframes in a particular time period, and thus guaranteeing that subsequent analysis covers the entire scanning process.

[0023] The keyframe selection process employs a two-stage verification mechanism. The first stage involves rapid pre-screening, quickly eliminating obviously unsuitable frames based on weighted scores of structural clarity and boundary continuity indicators. These typically include frames with incomplete optic nerve sheath structures, severely broken boundaries, or blurred images. The second stage performs refined anatomical integrity verification, conducting more rigorous analysis on the pre-screened candidate frames. Integrity is confirmed by examining the spatial relationship between the optic nerve sheath and surrounding tissues (including the posterior wall of the eyeball and orbital fat) to ensure it conforms to standard anatomical features. Specifically, this includes verifying whether the optic nerve sheath exhibits a typical "V" or "U" shape, whether the sheath maintains the correct angle with the posterior wall of the eyeball, and whether the echo characteristics of surrounding tissues conform to anatomical patterns. Finally, all verified frames are included in the initial image set. This set ensures both the image quality of individual frames and the temporal representativeness of the scanning process, providing reliable input data for subsequent deformation correction and segmentation analysis.

[0024] Please see Figure 3 As shown, in the anatomical structure template matching module, after obtaining the initial image set after quality screening, these images need to be registered with a pre-constructed standard anatomical template of the optic nerve sheath to generate a reference image for structural alignment. The standard anatomical template is a representative structure obtained through statistical analysis of normal optic nerve sheath ultrasound images. These images are all from healthy volunteers and have been rigorously screened and confirmed by professional physicians. During the template construction process, all images are first normalized to uniformly adjust to the same spatial resolution and grayscale range. Then, key anatomical feature points of the optic nerve sheath, including sheath boundary points, curvature feature points, and branching points, are determined through expert annotation. Finally, statistical shape analysis methods are used to calculate the average position and variation range of these feature points, forming a standard template containing typical morphological features and spatial location information of the optic nerve sheath.

[0025] First, multi-scale feature extraction and similarity calculation are performed. This process unfolds at three different scale levels: the first scale uses 1 / 4 resolution of the original image to capture the overall outline and main branch structures of the optic nerve sheath; the second scale uses 1 / 2 resolution to focus on the local curvature features of the sheath; and the third scale uses the original resolution to preserve subtle changes in the boundaries and texture features. At each scale level, features are extracted from the initial image and a standard anatomical template, respectively. Feature extraction uses a sliding window approach, with window sizes set to 32×32 pixels, 16×16 pixels, and 8×8 pixels depending on the scale. Two types of features are calculated within each window: gradient orientation histogram and gray-level co-occurrence matrix (GLCM) features. The gradient orientation histogram evenly divides the gradient direction from 0 to 180 degrees into nine intervals, and the gradient magnitude within each interval is summed. The GLCM features calculate contrast, correlation, and energy features in four directions (0 degrees, 45 degrees, 90 degrees, and 135 degrees). Then, similarity is calculated between the feature maps of the two images. The cosine similarity measure between feature vectors is used to evaluate the degree of similarity by calculating the cosine value of the angle between the two feature vectors, with the value ranging from -1 to 1.

[0026] After obtaining the feature correspondences at each scale, the initial deformation field is generated and the matching confidence is calculated. Based on the correspondences of feature points, a thin-plate spline interpolation algorithm is used to generate the initial deformation field. This algorithm first establishes a mapping relationship between corresponding feature points, and then solves for a smooth deformation field by minimizing the bending energy function. The bending energy function consists of two parts: feature point matching error and the smoothness constraint of the deformation field. Simultaneously, the system calculates the confidence of each feature point match. The confidence calculation comprehensively considers three factors: the similarity score of the feature match, the structural consistency score of the 5×5 pixel region surrounding the feature point, and the spatial consistency score of the matching results of the three adjacent feature points. The weights of these three factors are set to 0.6, 0.2, and 0.2, respectively. These confidence values ​​are converted into attention weights in the range of 0 to 1 using a sigmoid function, which are used to guide the subsequent deformation field optimization process.

[0027] Next, multi-scale deformation field fusion optimization is performed. The deformation fields generated at the three scale levels are weighted and fused, with the weight of each scale determined based on the average matching confidence at that scale. Specifically, the weight of the first scale is 0.4 times the base weight, the second scale is 0.35 times, and the third scale is 0.25 times, then normalized by multiplying by the average confidence of each scale. During the fusion process, a coarse-to-fine iterative optimization strategy is adopted, with a total of three iterations. The first iteration uses only the first-scale deformation field for initial alignment; the second iteration incorporates the second-scale deformation field for local adjustments; and the third iteration incorporates the third-scale deformation field for detailed optimization. Each iteration re-evaluates the registration effect, guiding subsequent optimization by calculating the similarity between the registered image and the template on five key anatomical features: overall curvature of the optic nerve sheath, boundary smoothness, branch angle, width change rate, and local curvature features.

[0028] After deformation field fusion, a final spatial transformation is performed to generate a reference image. The initial image is resampled using the fused deformation field, employing a cubic spline interpolation algorithm. This algorithm calculates the value of each target pixel using a 4×4 neighborhood of pixels, achieving smooth interpolation through a linear combination of 16 basis functions. During resampling, a special protection mechanism is used for the optic nerve sheath boundary region: first, the region within three pixels on either side of the boundary is detected, and interpolation parameters with reduced smoothness are used in these regions to preserve more edge details; while for the internal tissue regions, standard parameters are used for interpolation. The final generated reference image maintains a high degree of spatial structural consistency with the standard anatomical template, while preserving the texture details and grayscale features of the original image.

[0029] The parameter settings for the entire registration process are based on the statistical analysis results of clinical data, including the weights of each scale, the strength parameters of smoothing constraints, and the termination conditions for iterative optimization. These parameters fully consider the imaging characteristics of optic nerve sheath ultrasound images and the needs of clinical applications. Registration accuracy is evaluated using two indicators: feature point localization error and boundary overlap, requiring an average error of less than 2 pixels and a boundary overlap greater than 85%. Through this systematic registration method, even with clinical ultrasound images of varying quality, structurally accurate and detailed reference images can be generated.

[0030] In the deformation field inverse derivation module, after obtaining the reference image, the deformation field needs to be inversely derived to establish the spatial correspondence between the initial image and the reference image. First, feature trajectory backtracking analysis is performed, establishing a spatial mapping by selecting corresponding key anatomical landmarks in the initial and reference images. These key anatomical landmarks include boundary inflection points of the optic nerve sheath, curvature extrema, and connection points with other tissues; typically, 16 representative feature points are selected. For each feature point, a 5×5 pixel neighborhood window is taken around it, and the feature descriptor within this window is calculated, including the gradient direction histogram and local binary pattern features. By comparing the similarity of the feature descriptors in the two images, the correspondence between feature points is established. Based on these correspondences, the thin-plate spline interpolation method is used to calculate the motion trajectory of each pixel. This method constructs a smooth deformation field between known feature points and ensures that the displacement at the feature points is consistent with the observed values. During the calculation process, a continuity constraint on the trajectories of adjacent pixels is introduced, requiring that the change in the motion direction of adjacent pixels be gradual and the difference in motion amplitude be within a certain range, thereby constructing the initial displacement vector field.

[0031] Next, an adaptive propagation mechanism is implemented to effectively transmit displacement information of key anatomical regions to the entire image area. Using the displacement vectors of the optic nerve sheath boundary and key internal regions as reference points, the propagation range and intensity are determined based on the statistical characteristics of local image texture features. In areas with gentle texture changes, such as the homogeneous tissue region within the optic nerve sheath, a propagation strategy based on tissue texture uniformity is adopted. First, the texture uniformity index within a 7×7 pixel region centered on the reference point is calculated; this index is obtained by calculating the coefficient of variation of grayscale values ​​within the region. The higher the texture uniformity, the larger the propagation range is set, up to a maximum of 31×31 pixels. In areas with drastic texture changes, such as the boundary between the optic nerve sheath boundary and surrounding tissue, a propagation strategy based on edge structural complexity is adopted. Structural complexity is assessed by calculating the local variance and rate of curvature change in the edge direction; the higher the complexity, the smaller the propagation range is set, with a minimum of 5×5 pixels. During propagation, the amplitude of the displacement vector attenuates according to the propagation distance, and the attenuation coefficient is proportional to the texture similarity.

[0032] Then, a multi-level fusion optimization is performed, organically combining the displacement vector field based on trajectory backtracking with that based on adaptive propagation. First, the displacement vector fields obtained by the two methods are normalized to eliminate dimensional differences. In the edge regions of anatomical structures, including the boundary of the optic nerve sheath and the edges of important internal structures, the trajectory backtracking result is preferentially used, assigned a weight of 0.7, while the adaptive propagation result is assigned a weight of 0.3. This is because the trajectory backtracking method can provide more accurate displacement estimates in edge regions. In homogeneous tissue regions, the adaptive propagation result is preferentially used, assigned a weight of 0.8, while the trajectory backtracking result is assigned a weight of 0.2. This is because the adaptive propagation method can maintain better spatial continuity in homogeneous regions. In transition regions, a linear weighting method is used, with weights smoothly transitioning according to the distance to the nearest edge. Spatial consistency constraints are also introduced during the fusion process. A 3×3 sliding window is used to check the displacement vector differences between adjacent pixels, and regions with excessively large differences are smoothed.

[0033] Finally, the displacement vector field is refined, including outlier correction and spatial smoothing. First, outliers are detected in the displacement vector field; these outliers are typically displacement vectors that are opposite in direction to their neighboring vectors or have excessively large amplitude differences. By calculating the average difference between each displacement vector and its eight neighboring vectors, those with differences exceeding a threshold of two pixels are marked as outliers. These outliers are corrected using neighborhood median replacement. Then, spatial smoothing is performed using an adaptive Gaussian filtering method, employing a smaller filter kernel (σ=0.5) in edge regions and a larger filter kernel (σ=1.5) in uniform regions. During smoothing, the displacement vectors at key anatomical feature points are kept unchanged to ensure the accuracy of deformation of important structures. After three rounds of iterative optimization, a displacement vector field with spatial continuity and anatomical structural accuracy is finally generated, accurately reflecting the local geometric deformation of the initial image relative to the reference image.

[0034] In the physical constraint geometry correction module, after obtaining the displacement vector field, smoothing optimization based on the elasticity constraints of biological tissue is first required. This process begins with constructing an anisotropic smoothing constraint matrix. According to the biomechanical characteristics of periocular tissues, the optic nerve sheath has a high elastic modulus along the axial direction, while lateral deformation perpendicular to the axial direction is subject to greater constraints. Therefore, when constructing the smoothing constraint matrix, a smaller smoothing constraint coefficient is set along the direction of the optic nerve sheath, typically 0.3; while a larger smoothing constraint coefficient is set in the direction perpendicular to the direction of the optic nerve sheath, typically 0.8. This anisotropic constraint setting effectively preserves the natural deformation of the optic nerve sheath under normal physiological conditions while suppressing non-physical deformations that do not conform to biomechanical laws. The constraint matrix is ​​constructed based on the local orientation field of the optic nerve sheath, and the allocation of constraint coefficients is determined by calculating the principal direction at each pixel.

[0035] Next, strain energy density analysis is performed, calculating the strain state of each pixel based on the displacement vector field. First, the displacement gradient tensor of each point is calculated by analyzing the changes in displacement vectors within a 3×3 neighborhood. Then, the Green-Lagrange strain tensor is calculated based on the displacement gradient, which describes the degree of deformation of the microstructure. Based on the strain tensor, the strain energy density of each point is further calculated, taking into account mechanical parameters such as Young's modulus and Poisson's ratio of the microstructure. Based on the distribution of strain energy density, regions requiring enhanced smoothing are identified. Regions with strain energy densities exceeding twice the average value are typically marked as high-strain areas; these regions often correspond to locations with abnormal deformation or large calculation errors. Additional smoothing constraints are applied to these regions, with the smoothing intensity proportional to the magnitude of the strain energy density.

[0036] While keeping the displacements of key anatomical points constant, iterative optimization of the displacement vector field was implemented. Key anatomical points included boundary feature points, branch points, and curvature extrema of the optic nerve sheath, totaling approximately 20-30 points. The optimization process employed gradient descent, gradually adjusting the displacement vector field by minimizing the weighted sum of strain energy and smoothing constraints. The weight of the strain energy term was set to 0.6, and the weight of the smoothing constraint term was set to 0.4. In each iteration, the total energy of the current displacement field was first calculated, including both strain energy and smoothing constraint energy. Then, the gradient of the energy function with respect to each displacement vector was calculated, and the displacement vectors were updated according to the gradient direction. During the update process, the displacement vectors at the key anatomical points remained unchanged, while the displacements of other points were adjusted based on the gradient information. The entire optimization process underwent 50 iterations to gradually obtain a smooth displacement field that conformed to the mechanical properties of biological tissue.

[0037] Based on the optimized displacement field, image geometric reconstruction begins. First, the image is divided into different reconstruction regions according to the anatomical features of the optic nerve sheath and surrounding tissues. These are mainly divided into four regions: the optic nerve sheath boundary region, the intrasheath core region, the surrounding tissue region, and the background region. Different reconstruction parameters are used for each region: the optic nerve sheath boundary region uses the highest reconstruction precision with a sampling interval of 1 pixel; the intrasheath core region uses medium reconstruction precision with a sampling interval of 2 pixels; the surrounding tissue region uses standard reconstruction precision with a sampling interval of 3 pixels; and the background region uses lower reconstruction precision with a sampling interval of 4 pixels. This partitioned reconstruction strategy improves computational efficiency while maintaining the accuracy of important structures.

[0038] An edge-preserving resampling algorithm is employed in the optic nerve sheath boundary region. This algorithm first detects the boundary's orientation and curvature features, using a higher interpolation order (3rd order) in the boundary tangent direction and a lower interpolation order (1st order) in the boundary normal direction. Specifically, an 8×8 neighborhood window is taken around each sampling point, and the gradient direction distribution within the window is calculated. The anisotropy parameters of the interpolation kernel function are determined based on the dominant gradient direction. In homogeneous tissue regions, a conventional bilinear resampling algorithm is used, employing a 2×2 neighborhood for interpolation calculations. This differentiated resampling strategy effectively preserves the sharpness of the optic nerve sheath boundary while ensuring the quality of reconstruction within the tissue.

[0039] Finally, multi-resolution geometric reconstruction is performed. Initial geometric reconstruction is first performed at 1 / 4 resolution of the original image, focusing primarily on the accuracy of the overall structure. Then, intermediate layer reconstruction is performed at 1 / 2 resolution to optimize the alignment accuracy of key features. Finally, detail reconstruction is performed at the original resolution to accurately restore the image's texture details. At each resolution level, the reconstruction parameters are adjusted accordingly: larger smoothing parameters are used at lower resolution levels, and smaller smoothing parameters are used at higher resolution levels. A progressive thinning strategy is employed during the transition from low to high resolution, using the low-resolution reconstruction results as the initial values ​​for the high-resolution reconstruction to ensure the stability and continuity of the reconstruction process. Through this multi-resolution reconstruction method, a corrected image that maintains both the realism of the anatomical structure and the preservation of image details is ultimately generated.

[0040] In the boundary enhancement segmentation module, based on the corrected image, boundary feature enhancement processing within the gradient domain is performed first. This process begins with multi-scale gradient magnitude calculation, employing three different scale levels to enhance boundary features. The first scale uses a Gaussian kernel function to smooth the image with a smoothing parameter set to 3, then calculates the gradient magnitude; this scale primarily captures the overall contour features of the optic nerve sheath. The second scale uses a smaller Gaussian smoothing kernel with a smoothing parameter set to 1; the calculated gradient magnitude reflects a moderate level of boundary detail. The third scale directly calculates the gradient magnitude on the original image, preserving the finest boundary information. At each scale, the gradient magnitude is calculated using the Sobel operator, calculating the horizontal and vertical gradient components separately, then obtaining the comprehensive gradient magnitude by taking the square root of the sum of squares. This multi-scale calculation forms a boundary feature representation from coarse to fine, providing comprehensive boundary information for subsequent processing.

[0041] Next, anisotropic filtering is implemented, and a specialized filter kernel is designed based on the orientation characteristics of the optic nerve sheath boundary. First, the main orientation of the boundary at each pixel is determined by analyzing the gradient direction field, and the distribution of gradient directions is statistically analyzed within a 3×3 neighborhood. Based on the boundary orientation information, an anisotropic Gaussian filter kernel is constructed, with a kernel size of 7×7 pixels. A larger standard deviation is set in the boundary tangent direction, typically 2.5, to achieve a smoothing effect along the boundary direction; a smaller standard deviation is set in the boundary normal direction, typically 0.5, to maintain the gradient response in the normal direction. During the filtering process, the orientation parameters of the filter kernel are adjusted for each pixel according to its local boundary orientation to ensure that the filtering operation can adapt to changes in boundary orientation. This anisotropic filtering effectively enhances the continuity of the optic nerve sheath boundary while suppressing noise.

[0042] Based on local gradient statistical characteristics, dynamic enhancement intensity is adjusted. First, the statistical characteristics of the gradient magnitude, including mean, variance, and maximum value, are calculated within a 5×5 neighborhood of each pixel. Based on these statistical characteristics, an adaptive enhancement function is designed: in regions with weak gradient response (i.e., regions where the gradient mean is less than 30% of the overall image gradient mean), a strong enhancement coefficient is used, amplifying the gradient magnitude by a factor of 1.5; in regions with moderate gradient response, a moderate enhancement coefficient is used, amplifying by a factor of 1.2; and in regions with strong gradient response, the original gradient magnitude is kept unchanged. The enhancement process also considers the consistency of gradient direction, appropriately increasing the enhancement intensity in regions with gentle gradient direction changes and maintaining a conservative enhancement strategy in regions with drastic gradient direction changes. This dynamic adjustment mechanism can selectively enhance weak boundaries while avoiding over-enhancing strong boundaries.

[0043] After boundary feature enhancement, multi-scale edge feature fusion processing begins. The edge feature maps extracted from the three scales are weighted and combined, with weights assigned based on the reliability assessment of each scale feature. The first-scale feature primarily provides overall topological information about the optic nerve sheath and is assigned a weight of 0.4; the second-scale feature provides moderate detail information and is assigned a weight of 0.35; and the third-scale feature provides the finest boundary information and is assigned a weight of 0.25. During the fusion process, the feature maps at each scale are first normalized to eliminate scale differences, and then linearly combined according to their weights. To maintain boundary consistency, non-maximum suppression is also performed after fusion, retaining only local maxima points along the gradient direction and refining the width of the boundary lines.

[0044] Based on the fused edge features, edge connection processing is implemented to form a complete boundary contour. First, the curvature features of edge points are detected, and local curvature is estimated by calculating the angle formed by three adjacent edge points. Edge segments with curvature changes less than 15 degrees are marked as smooth regions, and those greater than 45 degrees are marked as high-curvature regions. During the connection process, edge segments with spatial distances within 5 pixels and similar curvature features are prioritized for connection. For edge segments with significant breaks, curve interpolation is used for connection, with the interpolation curve satisfying the requirement of second-order continuity. The continuity of edge intensity is also considered during the connection process, requiring the connected edges to maintain a smooth transition in intensity. Through this processing, discrete edge segments are connected into a complete optic nerve sheath boundary contour.

[0045] Finally, a boundary-constrained region growing algorithm is executed to achieve accurate segmentation. First, seed points are selected within the optic nerve sheath. These seed points are chosen based on anatomical prior knowledge and are typically located in the geometric center of the optic nerve sheath. Region growing begins from the seed points and expands progressively using an 8-neighborhood connection approach. A boundary distance penalty term is introduced during the growing process. When the distance between the boundary of the grown region and the boundary of the enhanced optic nerve sheath exceeds 3 pixels, a stronger penalty is applied to limit further growth. Simultaneously, gray-level similarity constraints are considered during the growing process, requiring that the gray-level difference between newly added pixels and the already grown region does not exceed 15 gray levels. Growth termination conditions include: reaching the enhanced boundary contour, encountering a significant gradient boundary, or the gray-level difference exceeding a threshold. This boundary-constrained growing method ensures that the final segmentation result is consistent with the enhanced boundary features, resulting in an accurate segmented optic nerve sheath region.

[0046] In the result verification and output module, after obtaining the segmentation results of the optic nerve sheath, morphological integrity verification is performed first. This process begins with the calculation of the boundary curvature change rate. Local curvature is calculated along the boundary line of the segmentation mask in segments of 5 pixels each, quantified by the angle formed by three adjacent boundary points. Simultaneously, the topological features of the connected components are analyzed, including checking whether the boundaries form a single closed contour, whether there are isolated broken regions, and the continuity of the boundary lines. When a curvature change rate exceeding 0.15 radians per pixel or a boundary break is detected, the system initiates an adaptive repair mechanism. The repair process first locates the abnormal region, searching for possible connection paths within a 10-pixel range around the break endpoint. Path selection is based on the principles of boundary direction consistency and grayscale continuity. For regions with abrupt curvature changes, cubic spline interpolation is used for smoothing while preserving the true anatomical structural features. The repaired boundary needs to be re-verified to ensure it meets the anatomical characteristics of the optic nerve sheath.

[0047] After morphological verification, precise measurement of the optic nerve sheath diameter was performed. First, the principal axis direction of the optic nerve sheath region was determined, and the major axis direction was obtained by calculating the second moment of the mask. Along the principal axis, 18 measurement directions were set at 10-degree intervals. In each direction, the distance between the two intersection points of the direction line and the boundary of the optic nerve sheath was calculated as the diameter value for that direction. After obtaining the 18 diameter measurements, statistical distribution analysis was performed to calculate the mean, standard deviation, and distribution pattern of these measurements. Box plots were used to identify outliers, marking measurements exceeding 1.5 times the interquartile range as abnormal. Outlier removal required anatomical rationality judgment to ensure that the retained measurements conformed to the actual morphological characteristics of the optic nerve sheath. Finally, the most frequent value was selected from the remaining measurements as the final measurement result; if multiple values ​​had the same frequency, their average was taken.

[0048] Based on the validation and measurement results, a comprehensive confidence index and a structured analysis report are generated. The confidence index consists of three parts: segmentation boundary sharpness score, diameter measurement consistency score, and morphological integrity score. Boundary sharpness is calculated by measuring the average gradient magnitude of the boundary region, with a value ranging from 0 to 1. Diameter measurement consistency is calculated using the coefficient of variation of 18 measurements; a smaller coefficient of variation results in a higher score. The morphological integrity score is derived from a comprehensive assessment of boundary curvature smoothness and connectivity. The three scores are assigned weights of 0.4, 0.4, and 0.2, respectively, and a weighted average is used to obtain the comprehensive confidence index. The final output structured analysis report includes the following parts: original image information, visualization of segmentation results, diameter measurements and their statistical characteristics, morphological integrity assessment results, comprehensive confidence index, and quality control recommendations. The report uses a standardized format to ensure that all necessary information is clearly presented, facilitating understanding and reference by clinicians.

[0049] A comprehensive quality control system was established throughout the validation and output process. During the validation phase, the system records all detected issues and their corrective actions, generating a detailed quality control log. In the measurement phase, the system saves all intermediate measurement results and statistical calculations, ensuring traceability. During report generation, the system automatically checks the completeness of all necessary data and the accuracy of all indicator calculations, ensuring the professionalism and reliability of the final report. This systematic validation and output process provides accurate and reliable optic nerve sheath analysis results for clinical diagnosis, while confidence level indicators offer crucial reference points for physicians using these results.

[0050] The working principle of this invention is as follows: Real-time dynamic video streams of the eye are acquired using ultrasound equipment. Based on structural clarity and boundary continuity indicators, keyframes containing the complete optic nerve sheath are automatically selected to form an initial image set. The initial images are then registered at multiple scales with a pre-constructed standard anatomical template. A structurally aligned reference image is generated through feature extraction, deformation field generation, and spatial transformation. Based on the spatial correspondence between the initial and reference images, the displacement vector field is inversely derived through feature trajectory backtracking and adaptive propagation mechanisms. The displacement field is smoothed and optimized under the constraints of biological tissue elasticity. A partitioned reconstruction strategy and multi-resolution method are used to generate a corrected image that conforms to anatomical realism. Multi-scale boundary feature enhancement is performed within the gradient domain of the corrected image. Precise segmentation is achieved through edge connectivity and boundary-constrained region growing. Finally, the segmentation results are verified for morphological integrity. Measurement results are obtained through multi-directional diameter measurement and statistical analysis, and a structured analysis report is output based on confidence assessment.

[0051] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A deep learning-based optic nerve sheath ultrasound image analysis system, characterized by, The method comprises the following steps: a dynamic sequence acquisition and quality screening module for acquiring a real-time dynamic ultrasound video stream of an eye by an ultrasound device, and automatically screening out key frames containing complete optic nerve sheath anatomical structures based on image structural integrity evaluation to form an initial image set; an anatomical structure template matching module for multi-scale registration of the initial image set and a pre-constructed standard optic nerve sheath anatomical template to generate a structure-aligned reference image; a deformation field reverse derivation module for reverse derivation of a displacement vector field representing local geometric deformation of the image based on pixel space correspondence between the initial image set and the reference image; a physical constraint geometric correction module for smoothing optimization of the displacement vector field under the constraint of biological tissue elasticity mechanics, and geometric reconstruction of the initial image based on the optimized displacement field to generate a corrected image conforming to anatomical reality; a boundary enhancement segmentation module for strengthening the boundary feature of the optic nerve sheath in the gradient domain of the corrected image, extracting continuous boundary features through a multi-scale edge perception algorithm, and completing accurate segmentation of the optic nerve sheath in combination with a region growing algorithm; a result verification and output module for morphological integrity verification of the segmentation result, calculation of the optic nerve sheath diameter measurement value based on the verified segmentation mask, and output of a final analysis report.

2. The deep learning-based optic nerve sheath ultrasound image analysis system of claim 1, wherein, The specific process of forming the initial image set comprises: synchronously calculating a structural definition index and a boundary continuity index of each frame of image, wherein the structural definition index is obtained by calculating a weighted sum of gradient amplitudes in a region of interest of the optic nerve sheath, and the boundary continuity index is obtained by detecting the fracture degree and curvature consistency of the edge of the optic nerve sheath; adaptingively adjusting the screening threshold according to the quality change trend of adjacent frames in the video sequence, automatically increasing the threshold requirement when a significant decrease in the quality of continuous frames is detected, and ensuring that the key frames selected are uniform in time distribution and stable in quality; firstly excluding obviously non-required frames based on fast pre-screening, then performing fine anatomical structure integrity verification on the candidate frames, confirming the integrity by checking whether the spatial relationship between the optic nerve sheath and the surrounding tissue conforms to the anatomical features, and finally constructing an initial image set with consistent quality. 3.The deep learning-based optic nerve sheath ultrasound image analysis system of claim 1, wherein, The specific process of generating the structure-aligned reference image comprises: extracting feature maps with different receptive fields from the initial image and the standard anatomical template respectively, calculating a local structural similarity measure at each scale level, and establishing a feature correspondence from coarse to fine; generating an initial deformation field according to the feature correspondence, calculating a feature matching confidence as an attention weight, and dynamically adjusting the smoothing constraint strength of the deformation field using the attention weight, keeping the flexibility of the deformation field in the reliable feature matching area and enhancing the smoothing constraint in the uncertain matching area; weighting and fusing the deformation fields generated at different scale levels, wherein the coarse scale deformation field provides global structure alignment, and the fine scale deformation field optimizes local detail registration, and finally generating a reference image that accurately reflects the spatial correspondence of the optic nerve sheath anatomical structure.

4. The deep learning-based optic nerve sheath ultrasound image analysis system of claim 1, wherein, The specific process of reverse derivation of the displacement vector field representing local geometric deformation of the image based on the pixel space correspondence between the initial image set and the reference image comprises: By establishing the spatio-temporal correspondence between the key anatomical landmark points of the initial image and the reference image, the motion trajectory of each pixel point between the image pairs is calculated, and an initial displacement vector field is constructed based on the continuity constraint of adjacent pixel trajectories; Taking the displacement vector of the key anatomical region as a reference, the displacement vector is propagated to the surrounding area according to the similarity of the local texture features of the image, and the propagation range based on the uniformity of the tissue texture is used in the area where the texture feature changes gently, and the propagation range based on the complexity of the edge structure is used in the area where the texture feature changes sharply; The displacement vector field based on trajectory backtracking and the displacement vector field based on adaptive propagation are weighted and fused, the trajectory backtracking result is preferentially used in the edge area of the anatomical structure, and the propagation result is preferentially used in the uniform tissue area, and finally a displacement vector field with spatial continuity and anatomical structure accuracy is generated. 5.The deep learning-based optic nerve sheath ultrasound image analysis system of claim 1, wherein, The displacement vector field is smoothed and optimized under the constraint of the biomechanics of the biological tissue, and specifically includes: According to the biomechanical properties of the ocular tissue, an anisotropic smoothing constraint matrix is constructed, wherein the smoothing constraint strength along the direction of the optic nerve sheath is smaller than the constraint strength in the vertical direction, so as to retain the normal anatomical deformation; Based on the displacement vector field, the strain tensor of each pixel point is calculated, and the area needing to be smoothed is determined according to the strain energy density distribution; Under the premise of keeping the displacement of the key anatomical point unchanged, the displacement vector field is gradually optimized by minimizing the weighted sum of the strain energy and the smoothing constraint, and a smoothed displacement field conforming to the biomechanical properties of the biological tissue is obtained.

6. The deep learning-based optic nerve sheath ultrasound image analysis system of claim 1, wherein, The corrected image conforming to the anatomical reality is generated, and specifically includes: According to the anatomical features of the optic nerve sheath and the surrounding tissue, the image is divided into different reconstruction regions, and different reconstruction parameters are used for each region; In the boundary area of the optic nerve sheath, an edge-preserving resampling algorithm is used, and in the uniform tissue area, a conventional resampling algorithm is used; First, the initial geometric reconstruction is completed at a lower resolution, and then the resolution is gradually improved to optimize the details, so as to generate a corrected image that not only maintains the anatomical structure reality but also retains the image details.

7. The deep learning-based optic nerve sheath ultrasound image analysis system of claim 1, wherein, The boundary feature of the optic nerve sheath is strengthened in the gradient domain of the corrected image, and specifically includes: By calculating the gradient amplitude response at different scales, the boundary feature expression from coarse to fine is enhanced, the overall contour is captured at a coarse scale, and the local details are retained at a fine scale; According to the direction feature of the boundary of the optic nerve sheath, an anisotropic filter kernel is designed, the smoothing processing is performed along the tangent direction of the boundary, and the gradient response in the normal direction is strengthened; Based on the local gradient statistical feature, the enhancement strength is dynamically adjusted, the enhancement strength is improved in the boundary area with weak gradient response, and the original feature is maintained in the area with strong gradient response.

8. The deep learning-based optic nerve sheath ultrasound image analysis system of claim 1, wherein, The continuous boundary feature is extracted by a multi-scale edge perception algorithm, and the accurate segmentation of the optic nerve sheath is completed by combining a region growing algorithm, and specifically includes: The edge feature maps extracted at different scales are weighted and fused, wherein the coarse scale edge provides topological structure guidance, and the fine scale edge supplements detail information; Based on the curvature feature and spatial proximity of the edge points, the broken edge segments are connected into a complete boundary contour. The region growing is performed from a seed point inside the optic nerve sheath with the complete boundary contour as a constraint, and a boundary distance penalty term is used to ensure that the segmentation result is consistent with the enhanced boundary features during the growing process. 9.The deep learning-based optic nerve sheath ultrasound image analysis system of claim 1, wherein, The output final analysis report specifically includes: The smoothness and continuity of the optic nerve sheath contour are detected by calculating the boundary curvature change rate and the connected domain topological features of the segmentation mask, and adaptive repair is started for the region with fracture or abnormal mutation; In the verified optic nerve sheath region, a plurality of directional diameter values are calculated at equal angle intervals along the main axis direction, abnormal measurement values are excluded through statistical distribution analysis, and the mode diameter is taken as the final measurement result; Based on the definition of the segmentation boundary, the consistency degree of the diameter measurement value and the morphological integrity score, a comprehensive confidence index is generated, and all verification data, measurement results and confidence evaluation are integrated into a structured analysis report.

Images