Urinary calculus image recognition and analysis system based on deep learning

Abstract

The invention discloses a urinary calculus image recognition and analysis system based on deep learning, which relates to the technical field of urinary calculus image recognition and comprises a structure communication mapping module, an edge disturbance analysis module, a structure fidelity coding module, a trusted path assignment module and a self-adaptive screening regulation and control module. The edge disturbance analysis module is used for constructing an edge direction difference matrix based on the region connectivity vector, carrying out gradient direction analysis on boundary pixels in the enhanced image and extracting edge frequency disturbance characteristics; and the structure fidelity coding module is used for correspondingly fusing the region connectivity vector and the edge frequency disturbance characteristics according to position indexes, constructing a structure integrity description vector and constructing a structure fidelity kernel function based on the structure integrity description vector so as to generate a structure fidelity score. According to the method, the problem of training misleading caused by incomplete enhanced image structure is solved, structural integrity screening and path optimization of training samples are realized, and the model recognition precision and stability are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of urinary stone image recognition, and in particular to a urinary stone image recognition and analysis system based on deep learning. Background Art

[0002] Urinary stone image recognition involves processing and analyzing medical imaging data (such as CT, ultrasound, or X-ray images) to automatically detect, locate, and identify key features of urinary stones, such as their location, size, morphology, and density, thereby providing decision support for clinical diagnosis and treatment. Traditional urinary stone recognition relies primarily on manual image reading by physicians, which is not only time-consuming and labor-intensive but also prone to missed diagnoses and misdiagnoses due to experience differences and visual fatigue, making it difficult to meet the needs of efficient and precise medical care. Deep learning, an artificial intelligence technology with powerful image recognition and pattern extraction capabilities, has particularly excelled in the field of medical image analysis. By constructing and training multi-layer neural networks, it can automatically learn and extract complex feature representations from large amounts of annotated images, enabling precise detection and classification of urinary stones and significantly improving recognition accuracy and processing efficiency. Urinary stone image recognition and analysis technology based on deep learning not only reduces physician workload and diagnostic subjectivity, but also enables a standardized and intelligent analysis process, making it a key direction for the development of intelligent medical image-assisted diagnosis systems.

[0003] Existing deep learning-based urinary stone image recognition and analysis technologies primarily rely on building multi-layer neural network models to automatically process medical image data. The core process typically includes key steps such as image acquisition and preprocessing, data annotation and enhancement, model building and training, image recognition and analysis, and output and visualization. First, urinary system images are acquired using imaging devices such as CT or ultrasound. These images undergo preprocessing, including grayscale normalization, noise removal, and edge enhancement, to improve image quality and feature representation. A training set is then constructed using expert-annotated image data, and data augmentation techniques are used to increase sample diversity and enhance model robustness. Next, mainstream deep learning architectures such as convolutional neural networks (CNNs), U-Net, or YOLO are used to automatically detect, segment, and classify stone regions in the images, extracting multidimensional feature information such as location, size, and density. During model training, network parameters are continuously adjusted through loss function optimization and backpropagation algorithms to improve recognition accuracy and generalization. Finally, the model analysis results are visualized in the form of image heat maps, coordinate boxes, or stone attribute tables, providing clinicians with intuitive and accurate diagnostic information. Overall, this technology realizes a complete closed-loop process from "image to intelligent analysis" for urinary stones, greatly improving diagnostic efficiency and accuracy.

[0004] The existing technology has the following deficiencies: When training deep learning models on urinary stone imaging data, image enhancement techniques are often used to improve the model's generalization capabilities, including random cropping, rotation, and flipping. However, because urinary stones are small and concentrated in the image target area, when the enhancement strategy adopts a fixed cropping grid or center-biased sampling method, it is easy for the training samples to only contain partial stone edges or local texture fragments. Due to the incomplete target structure, these enhanced samples cannot provide accurate stone feature representations, leading to misleading model learning and incorrectly learning abnormal edge regions or background information as stone features. Because the existing data augmentation process is based only on the principle of uniform sampling in image space and lacks a mechanism to effectively evaluate the integrity of the target area, existing deep learning-based urinary stone image recognition and analysis technologies cannot determine whether the training sample has the feature completeness to constitute a valid target based on the incomplete structural features of the stone fragments sampled during the training image enhancement process, thereby filtering out invalid or misleading samples. This problem will cause feature offset in the model during the edge learning stage, reduce its recognition accuracy of boundary structures, and ultimately cause phenomena such as edge expansion and pseudo-structure recognition during the reasoning process, affecting the accuracy and robustness of the overall recognition system.

[0005] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not form the prior art that is already known to a person of ordinary skill in the art. Summary of the Invention

[0006] The purpose of the present invention is to provide a urinary stone image recognition and analysis system based on deep learning to solve the problems in the above-mentioned background technology.

[0007] To achieve the above objectives, the present invention provides the following technical solutions: a urinary stone image recognition and analysis system based on deep learning, comprising a structural connectivity mapping module, an edge perturbation analysis module, a structural fidelity coding module, a trusted path assignment module, and an adaptive screening and control module; The structural connectivity mapping module performs pixel-level mapping between the enhanced urinary stone image and its original annotated area, extracts the structural connectivity information of the corresponding area in the enhanced image, and generates a regional connectivity vector; The edge perturbation analysis module constructs an edge direction difference matrix based on the regional connectivity vector, performs gradient direction analysis on the boundary pixels in the enhanced image, and extracts edge frequency perturbation features; The structural fidelity encoding module fuses the regional connectivity vector and edge frequency perturbation features according to the position index to construct a structural integrity description vector, and constructs a structural fidelity kernel function based on the structural integrity description vector to generate a structural fidelity score; The trusted path assignment module classifies the enhanced images into trustworthy levels according to the structural fidelity scores and assigns them to the main training path, auxiliary training path or re-evaluation path respectively; The adaptive screening and control module constructs a dynamic control factor based on the distribution of structural fidelity scores and boundary prediction gradient fluctuations in consecutive training rounds, and adjusts the parameters of the structural fidelity kernel function accordingly.

[0008] Preferably, in the structural connectivity mapping module, a pixel-level mapping relationship between the enhanced image and the original annotated area is established by presetting reference reference points in the original image and recording the geometric transformation parameters corresponding to the enhancement processing, and the enhanced image coordinates are restored to the original image reference frame through inverse affine transformation to achieve target area alignment; after the mapping is completed, the structural connectivity mapping module extracts the target area corresponding to the annotated area in the enhanced image, constructs a binary graph structure and introduces the eight-neighborhood connection rule to identify the connected area, performs pixel composition density, contour closure degree, area ratio and central axis distribution calculations on each connected area, obtains the corresponding structural connectivity information, and uses the structural connectivity information to generate a regional connectivity vector including the number of connected blocks, the pixel coverage of the largest connected block, the edge smoothness, the gradient difference between the inner and outer boundaries, the center of gravity offset rate and the main axis distribution of the structural direction.

[0009] Preferably, the edge perturbation parsing module extracts the boundary pixel set of the target area in the enhanced image based on the target pixel connectivity label, boundary pixel identification mark and structural closure index contained in the regional connectivity vector, and constructs a direction perception window with each boundary pixel as the center. The direction perception window defines eight direction sequences according to the image coordinate system and is numbered as the 1st to 8th directions in sequence. At the same time, the relative position offset relationship between the adjacent pixels and the current boundary pixel in each direction is recorded to generate a direction mapping index table; the edge perturbation parsing module uses central difference to calculate the grayscale difference between the current pixel and the adjacent pixels in each direction according to the direction mapping index table and performs normalization processing to construct an edge direction difference matrix. Each row of the edge direction difference matrix corresponds to a boundary pixel, and each column corresponds to a direction. The value in the cell is the standardized grayscale difference response value in the direction, which is used to describe the gradient change behavior of the boundary in the spatial direction.

[0010] Preferably, the edge perturbation analysis module performs gradient direction analysis based on the directional response sequence of each boundary pixel in the edge directional difference matrix, and adopts an angular interpolation mechanism to construct a quasi-continuous angle sequence to improve the sensitivity of directional change modeling. The angular interpolation mechanism fills the intermediate angle points between the original eight directions through cubic spline interpolation, and applies the Laplace directional difference function based on the interpolation result to detect the local gradient mutation position and directional jump frequency; on this basis, the edge frequency perturbation features are extracted, and the edge frequency perturbation features include three indicators: directional gradient mean square error, directional jump number and local gradient peak distribution, and uniformly adopts a spatial convolution response reduction mechanism for regional scale normalization and response calibration to form an edge frequency perturbation feature set with directional sensitivity and perturbation aggregation.

[0011] Preferably, the structural fidelity coding module uses the two-dimensional pixel coordinates of the enhanced image as a unified index benchmark, and obtains the regional connectivity vector and edge frequency perturbation features corresponding to the coordinates for each pixel position, wherein the regional connectivity vector includes the number identifier of the connected region to which the pixel belongs, the pixel density of the connected region, the boundary closure score and the axial structure distribution angle, and the edge frequency perturbation features include the mean square error of the directional gradient, the number of directional jumps and the density of local extreme points; by establishing a spatial position index mapping table, the structural attributes in the regional connectivity vector and the directional response information in the edge frequency perturbation features are fused one-to-one at the same coordinate position, and by introducing normalization adjustment, boundary direction weighting and spatial stability restriction mechanism, a fused multi-dimensional pixel-level feature combination is formed, providing a corresponding fusion data set for the subsequent construction of the structural integrity description vector.

[0012] Preferably, after fusing the regional connectivity vector and the edge frequency perturbation feature, the structural fidelity encoding module constructs a structural integrity description vector in a unified data dimension format, wherein the structural integrity description vector includes a structural dimension sub-vector composed of the pixel density of the connected region, the boundary closure score and the structural main axis angle, and a perturbation dimension sub-vector composed of the mean square error of the directional gradient, the number of jumps and the extreme point density; the structural integrity description vector is encoded in a floating point format, and a centralized coordinate alignment, gradient scale normalization and boundary continuity weighted adjustment mechanism are introduced for all vectors to ensure that all description vectors have a unified scale and stable directional feature expression in the same image space, and are stored in the structural integrity description vector pool according to the pixel index, providing a continuous feature input basis for the structural fidelity kernel function.

[0013] Preferably, the structural fidelity coding module constructs a structural fidelity kernel function based on the structural integrity description vector, wherein the structural fidelity kernel function is composed of a combination of a Gaussian radial basis function and a direction cosine similarity function, the former being used to measure the difference in structural scale and density between the description vector and the standard structural sample, and the latter being used to measure the consistency in edge direction features between the description vector and the standard structural sample; a reference vector set is constructed by selecting manually verified high-fidelity structural samples, and the structural fidelity kernel function is used to perform response calculations on each structural integrity description vector to generate a structural fidelity response surface, and a structural fidelity score is output based on the numerical gradient and change trend of the response distribution, and the score value represents the numerical performance of each pixel in the current image in the structural stability dimension in the form of a continuous floating-point interval.

[0014] Preferably, the trusted path assignment module divides the enhanced image into trust levels based on the structural fidelity score generated by the structural fidelity encoding module. The structural fidelity score is a continuous floating-point number. The trusted path assignment module constructs a dynamic score three-partition model by statistically analyzing the global mean, standard deviation, skewness coefficient and score distribution slope of all scores, and adopts a multi-parameter joint quantile analysis method to divide the enhanced image samples into three levels: high confidence zone, medium confidence zone and low confidence zone, and binds the trust zone label and the original score to each image sample as a path assignment reference coefficient; the high confidence zone image samples are assigned to the main training path, the medium confidence zone image samples are assigned to the auxiliary training path, the low confidence zone image samples are marked as samples to be re-evaluated and assigned to the re-evaluation path, and independent index pools and data buffer mechanisms are configured for the three types of paths respectively.

[0015] Preferably, the adaptive screening and control module constructs a dynamic control factor based on the structural fidelity score distribution and boundary prediction gradient fluctuation in continuous training rounds, and dynamically updates the key parameters in the structural fidelity kernel function according to the dynamic control factor. The structural fidelity score is generated by the structural fidelity kernel function for the structural integrity description vector. The structural fidelity score is aggregated and statistically analyzed at the image level in each training round to form a score statistical matrix containing the score mean, score variance, distribution slope, number of extreme values ​​and score density interval. While obtaining the score evolution vector, the adaptive screening and control module extracts the boundary offset amplitude, gradient response jump frequency and local error density distribution based on the pixel-level difference comparison between the boundary prediction map of the model and the original labeled area after each round of training to form a predicted gradient fluctuation spectrum. The score evolution vector and the predicted gradient fluctuation spectrum are then fused to form a dual-channel dynamic control factor, which is used to adjust the scale factor of the Gaussian radial basis function and the tolerance angle parameter of the direction cosine similarity function in the structural fidelity kernel function respectively.

[0016] In the above technical solution, the technical effects and advantages provided by the present invention are: 1. The present invention forms a complete structural quality-driven training sample screening and control mechanism by constructing a structural connectivity mapping module, an edge perturbation analysis module, a structural fidelity coding module, a trusted path assignment module and an adaptive screening and control module. The system first extracts pixel-level structural connectivity information from the stone area after image enhancement, and further analyzes the edge direction response to construct the edge frequency perturbation feature. The two types of features are then fused one by one under the pixel position index to generate a high-dimensional structural integrity description vector, and a structural fidelity kernel function is constructed based on the description vector to output a structural fidelity score. The scoring results are used to guide the trust level division and path assignment of training samples, so that high-integrity samples participate in the main training path, and structurally unstable or edge-abnormal samples enter the auxiliary or re-evaluation path, which comprehensively improves the support strength of the training data for model structure learning.

[0017] 2. The present invention also introduces an adaptive screening and control module, which constructs a dynamic control factor based on the changes in score distribution and fluctuations in boundary prediction behavior in consecutive training rounds, and continuously optimizes the parameters of the structural fidelity kernel function, thereby ensuring that the structural scoring standard can adaptively evolve with the model learning state, avoiding problems such as score failure or discrimination drift. The overall system breaks through the limitation of the lack of structural integrity discrimination mechanism in existing enhanced image training through the technical chain of "structure extraction-perturbation modeling-fidelity scoring-path scheduling-kernel function self-adjustment", realizes the accurate identification and elimination of invalid or misleading samples, effectively alleviates the feature offset problem caused by boundary fragment learning in model training, and improves the model's recognition accuracy of urinary stone boundary structure and the stability of the inference stage. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction to the drawings required for use in the embodiments will be given below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0019] Figure 1 This is a module diagram of the urinary stone image recognition and analysis system based on deep learning of the present invention. DETAILED DESCRIPTION

[0020] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these example embodiments are provided so that the description of this disclosure will be thorough and complete and will fully convey the concepts of the example embodiments to those skilled in the art.

[0021] The present invention provides Figure 1The deep learning-based urinary stone image recognition and analysis system shown includes a structural connectivity mapping module, an edge perturbation analysis module, a structural fidelity encoding module, a trusted path dispatching module, and an adaptive screening and control module; The structural connectivity mapping module performs pixel-level mapping between the enhanced urinary stone image and its original annotated area, extracts the structural connectivity information of the corresponding area in the enhanced image, and generates a regional connectivity vector; In this embodiment, the structural connectivity mapping module first establishes a pixel-level mapping relationship between the enhanced urinary stone image and its original annotated area to ensure spatial consistency between the target structure in the enhanced image and the original annotated information. In order to achieve pixel-level accuracy, reference reference points are preset in the original image before enhancement, and affine parameters are used to record geometric transformation information including rotation angle, cropping boundary and scaling ratio. After completing the enhancement operation, the structural connectivity mapping module restores the coordinate space of the enhanced image to the original image reference frame through an inverse affine transformation, and performs isotropic mapping on the annotated stone area mask in the original image to obtain an annotated area that is completely aligned with the enhanced image. This pixel mapping process ensures that each pixel point in the enhanced image can correspond to the stone area in the original annotation, thereby ensuring that the subsequently extracted connectivity structure information has verifiable anatomical consistency and avoiding the problem of erroneous feature extraction due to coordinate drift.

[0022] The original annotated region refers to the set of pixels generated by accurately annotating the target region in the original urinary stone image through manual labeling or semi-automatic tools. This region is usually expressed in the form of a mask image, and its data structure is a binary image with the same size as the original image, where pixels representing the urinary stone target are assigned a value of 1, and other background areas are assigned a value of 0. This can accurately describe the morphology, position, and boundary contour of the stone in the image. The annotated region is usually outlined by professional radiologists based on clinical imaging judgment and has high-precision spatial semantic expression. The original image reference frame refers to the spatial coordinate system established during the acquisition process of the original image. Each pixel in this coordinate system has a fixed row and column index and corresponds to the scanning orientation and scale information in the real world. It is the reference coordinate space for all image enhancement transformation operations. During the image enhancement process, operations such as cropping, rotation, and affine stretching applied to the original image will record the geometric relationship between them and the reference frame. Through reverse mapping, the pixels in the enhanced image can be accurately mapped to the coordinate positions of the original annotated area in the original reference frame, thereby achieving pixel-level spatial alignment between the enhanced image and the annotated area, providing an accurate basis for subsequent connectivity structure extraction and feature consistency analysis.

[0023] Based on the established pixel-level mapping relationship, the structural connectivity mapping module further performs connectivity analysis on the mapped stone area to accurately depict its spatial structural state in the enhanced image. In this process, the corresponding target area in the enhanced image is first cut out according to the labeled mask, a binary graph structure representation is constructed, and a pixel intensity threshold is set to distinguish the stone area from the background area. Subsequently, the eight-neighborhood connection rule is introduced to perform adjacency judgment pixel by pixel at the image structure level to identify all connected areas with strong closure and stable internal connections. For each connected area, the structural connectivity mapping module calculates its pixel composition density, contour closure degree, regional area ratio, and central axis distribution, and establishes a preliminary measurement of its spatial integrity and morphological extensibility. The above processing not only accurately retains the physical boundary characteristics of the target area in the enhanced image, but also provides a stable structural foundation for subsequent vectorization processing.

[0024] Constructing a binary image structure means cutting out the target area corresponding to the pixel level of the original annotated area in the enhanced urinary stone image, and performing grayscale threshold processing on the cut result so that pixels with grayscale values ​​above the set threshold are assigned a value of 1, and pixels below the threshold are assigned a value of 0, thereby forming a binary image structure containing only the target area. This structural image can clearly present the target boundary outline, remove background interference, and provide basic input for connectivity analysis. On this basis, the eight-neighborhood connection rule is introduced, that is, in a two-dimensional pixel grid, for any pixel, it is judged whether its value is consistent with the adjacent pixels in the upper, lower, left, right, and four diagonal directions. If the adjacent pixel value is also 1, it is considered to belong to the same connected area. This rule is more robust than the four-neighborhood rule and can more comprehensively identify micro-structure connections. For each connected region, the pixel density is further calculated—the ratio of target pixels to total pixels within the region. Contour closure is determined by comparing the end-to-end connectivity of the region's boundary curves and the continuity of the boundary gradient. The region area ratio reflects the dominance of the target body by comparing the area of ​​the largest connected region to the total area of ​​all connected regions. The central axis distribution measures the symmetry and spatial extensibility of the target structure based on the center of mass position and principal direction vector of the connected regions. Together, these metrics form a multidimensional description of the target region's structural state, providing a comprehensive and quantifiable foundation for determining structural integrity.

[0025] After completing the extraction of the connected structure of the target area, the structural connectivity mapping module converts these connected areas into a quantifiable structural description form, namely the regional connectivity vector. In this implementation, the regional connectivity vector is realized through a multi-dimensional feature encoding method, which includes the number of connected blocks, the pixel coverage of the largest connected block, the degree of edge smoothness, the gradient difference between the inner and outer boundaries, the center of gravity offset rate and the main axis distribution of the structural direction. Each parameter is obtained through statistical analysis of the pixel arrangement pattern, spatial density distribution and edge change trend in the target area. The structural connectivity mapping module aligns the scales of all parameters through a unified normalization strategy, and forms a multi-dimensional feature vector structure through a series combination, so that the regional connectivity vector can not only accurately reflect whether the target is a complete structure in the enhanced image, but also has a high degree of distinguishability, supporting subsequent feature evaluation and training data screening.

[0026] To further enhance the stability of regional connectivity vectors in discriminating target structural validity, the Structural Connectivity Mapping module introduces a robustness validation mechanism to compare and analyze the response of each vector under different image enhancement strategies. Specifically, after applying various geometric enhancement techniques to the same original image data, multiple sets of enhanced images are generated, and regional connectivity vectors are extracted for each set of images. The module calculates cosine similarity and Manhattan distance for the connectivity vectors under different enhancement conditions and sets a tolerance interval to assess their fluctuation. When the connectivity vector of an enhanced image deviates from the mean vector by more than a predetermined threshold, it is marked as an anomalous sample with low structural stability. This vector is subsequently assigned a lower confidence weight when used for training sample screening or fidelity score calculation. In this way, the Structural Connectivity Mapping module not only generates regional connectivity vectors with geometric semantic expression capabilities, but also achieves dual optimization in the quantitative and robustness of enhanced image structural connectivity assessment, ensuring that the samples ultimately used for deep learning training possess high-quality structural expression capabilities.

[0027] The edge perturbation analysis module constructs an edge direction difference matrix based on the regional connectivity vector, performs gradient direction analysis on the boundary pixels in the enhanced image, and extracts edge frequency perturbation features; In this embodiment, the edge perturbation analysis module first performs boundary extraction and direction mapping preprocessing on the target boundary area in the enhanced image based on the regional connectivity vector. The regional connectivity vector records the spatial extension relationship, boundary closure characteristics and connectivity strength information between adjacent pixels of the target area in each enhanced image. The edge perturbation analysis module uses the pixel position marked as the boundary point in the vector as a reference to locate all target edge point sets in the enhanced image, and constructs a direction perception window to capture its edge response at different angles. Each direction perception window is centered on the boundary pixel and covers a fixed neighborhood around the current pixel according to the eight-way pixel connection mode, and assigns a direction number consistent with the image coordinate system to each direction. Based on this, the edge perturbation analysis module establishes a pixel-level direction correspondence matrix and generates a direction mapping index table for each boundary pixel to provide clear direction guidance for subsequent differential calculations.

[0028] The edge perturbation resolution module performs boundary extraction and direction mapping preprocessing based on the region connectivity vector. Its core is to accurately identify the edges of each target region in the enhanced image using the target pixel connectivity labels, boundary pixel identifiers, and structural closure metrics contained in the region connectivity vector. First, the boundary pixel identifier field in the region connectivity vector annotates all target edge pixels with significant gradient transitions, enabling the edge perturbation resolution module to accurately extract the boundary pixel set in the enhanced image. After boundary extraction, the edge perturbation resolution module constructs a direction-aware window centered on each boundary pixel. This window uses a fixed size of 3×3 or 5×5 units and is strictly oriented according to the image coordinate system. Directions are numbered 1 to 8 clockwise from the upper left, corresponding to the eight directions: upper left, upper, upper right, right, lower right, lower, lower left, and left. Each window covers the pixels adjacent to the current boundary pixel in all eight directions, and its grayscale or edge response value is used as the initial analysis variable. To ensure directional consistency in the analysis, the edge perturbation analysis module also generates a directional mapping index table, recording the relative position offset vectors between each boundary pixel and its adjacent pixels in each direction, and encoding them into standardized index labels. Through this mapping relationship, the edge perturbation analysis module achieves a unified conversion between the image coordinate system and boundary direction information, laying a complete data foundation and directional indexing mechanism for the subsequent precise construction of the directional difference matrix. This process does not rely on manual annotation or the introduction of external directional templates, but is completely dynamically generated based on the image's own structural data, ensuring high adaptability and analytical stability.

[0029] Based on the established direction mapping index relationship, the edge perturbation analysis module constructs an edge direction difference matrix in the neighborhood of each boundary pixel. Each row of the matrix corresponds to a boundary pixel in the image, each column corresponds to a direction angle, and the value in the cell represents the grayscale difference response of the pixel in the corresponding direction. In order to ensure that the differential value can accurately reflect the detailed characteristics of the edge structure, the edge perturbation analysis module calculates the grayscale difference between the current pixel and the adjacent pixel in the corresponding direction through the central difference method. The numerical result is normalized and filled into the edge direction difference matrix. This matrix fully retains the grayscale change characteristics of the boundary in all directions, has significant directionality and local response characteristics, and provides basic data support for the subsequent frequency domain analysis of edge perturbation behavior.

[0030] Based on the acquired directional mapping index relationship, the edge perturbation analysis module models the directional differential response of each boundary pixel. Its core is to comprehensively capture the directional grayscale changes of the local boundary by constructing an edge directional difference matrix. Specifically, the edge perturbation analysis module first traverses the set of all pixels marked as boundary points in the enhanced image and determines the precise coordinate positions of its eight adjacent pixels in each direction based on the directional mapping index table. In each direction, the module uses the central difference calculation formula, that is, taking the current boundary pixel as the reference point, taking the grayscale value of the adjacent pixel in that direction minus the grayscale value of the reference point, and dividing it by the pixel spacing between the two to obtain the standardized grayscale difference. To enhance the comparability and statistical consistency of the differential values, the edge perturbation analysis module performs maximum and minimum value normalization on all directional differences, so that the grayscale difference response values ​​in the matrix are compressed to between 0 and 1, ensuring that no offset or abnormal response is caused by grayscale scale differences in subsequent analysis. The constructed edge direction difference matrix exists in the form of a two-dimensional tensor. The row dimension of the matrix corresponds to the boundary pixel number, and the column dimension strictly corresponds to the eight directional angles. Each cell in the matrix represents the grayscale mutation intensity of the pixel in a specific direction. This matrix not only fully reflects the grayscale change trend of each boundary point in different directions, but also has clear spatial gradient structure characteristics, providing highly structured raw data support for further identification of disturbance types, frequency distribution, and edge structure stability. This pixel-level directional difference modeling method breaks through the traditional edge feature extraction method based solely on convolution kernel response, and has stronger directional decoupling capabilities and edge dynamic analysis capabilities.

[0031] After constructing the edge direction difference matrix, the edge perturbation analysis module further performs gradient direction analysis on the direction difference sequence of each row in the matrix. Specifically, the edge perturbation analysis module extracts the direction response sequence of each row, regards it as the local angular gradient signal sequence of the boundary, and identifies its fluctuation amplitude, direction jump position and local gradient steep change point through the angular change rate model. In order to improve the sensitivity of capturing angle changes, the edge perturbation analysis module introduces an angular interpolation mechanism, constructs interpolation points between adjacent directions to simulate angular continuous response, and uses the Laplace directional difference function to measure the mutation trend between each angle. Through this analysis process, the edge perturbation analysis module not only obtains the change trend of each boundary pixel in different directions, but also captures its high-frequency structural behaviors such as gradient peak concentration area, response extreme value area and change inflection point, providing a clear directional basis for the extraction of perturbation features.

[0032] After constructing the edge directional difference matrix, the edge perturbation analysis module further performs high-precision gradient directional analysis on the directional difference response sequence of each edge pixel. The key objective is to identify angular abrupt changes in the grayscale variation at the edge. To this end, the module interprets each row in the directional difference matrix as the gradient change trajectory of that pixel in eight directions, forming a directionally ordered sequence of angular gradient signals. This sequence reflects the amplitude and pattern of grayscale transitions along the edge in different directions and is a direct manifestation of local edge perturbation behavior. To improve the resolution and directional sensitivity of the analysis, the module introduces an angular interpolation mechanism. Interpolation points are added to the original eight discrete directions, transitioning the angular distribution from a discrete state to a quasi-continuous state, thereby providing a more detailed description of the angular gradient trend. Cubic spline interpolation is used to maintain smoothness and physical consistency of the response curve. After interpolation, the Laplace directional difference function is applied to the entire angular sequence to quantify the degree of gradient abrupt changes between angles. This function performs a second-order difference operation on the interpolated sequence to identify sign flips and amplitude jumps in sudden gradients between adjacent directions, thereby marking the local maximum gradient peak, the location of the response extreme value, and the inflection point of the change. These indicators are used to measure the degree of directional imbalance of boundary perturbations, local boundary instability, and the location of structural continuity breaks. Compared with traditional static edge detection methods, this angular gradient analysis process achieves microscopic modeling of boundary perturbation behavior and can effectively distinguish between natural edge morphological changes and abnormal structural jumps introduced by enhancement operations, providing a clear direction and prominent change analysis basis for subsequent perturbation feature extraction.

[0033] After modeling the gradient directional behavior of each boundary pixel, the edge perturbation analysis module extracts edge frequency perturbation features based on the angular gradient analysis results. This feature reflects the degree of instability of the boundary response in different directions and the intensity of structural perturbations. It is composed of three components: the mean square error of directional gradients, which measures the central tendency of the boundary's amplitude variation in various directions; the number of directional jumps, which describes the degree of discreteness of the edge structure's directionality; and the distribution of local gradient peaks, which identifies regions of frequent edge perturbations. The edge perturbation analysis module uniformly employs a spatial convolution response reduction mechanism to extract these metrics, ensuring comparability of edge perturbation behavior across different regions. Finally, the edge perturbation analysis module outputs a set of edge frequency perturbation features for each boundary pixel and retains the original directional differential path for subsequent reference in evaluating the structural stability of training samples. This provides a directional basis for identifying structural anomalies in image enhancement samples.

[0034] After modeling the angular gradient behavior, the edge perturbation analysis module further extracts edge frequency perturbation features based on statistical analysis of gradient directional variations. This is used to measure the perturbation intensity and stability distribution of the boundary structure after image enhancement across spatial angles. First, to capture the concentration of boundary response variations in multiple directions, the edge perturbation analysis module calculates the mean square error of directional gradients. Specifically, based on the directional gradient sequence corresponding to each boundary pixel, it first calculates the mean, then calculates the squared differences between the response values ​​in each direction and the mean, and averages them. This metric reflects whether the local boundary exhibits significant anomalous responses in certain directions. Second, to characterize the discontinuous jump behavior of the edge structure, the edge perturbation analysis module counts the number of directional jumps—that is, the number of sign flips and response abrupt changes between consecutive directions in the directional gradient sequence—to quantify the directional complexity of the structural perturbation. Third, the edge perturbation analysis module locates the distribution of local gradient peaks, identifies the extreme response intensity points in the angular response sequence, and maps their distribution density in image space, thereby determining the high-frequency regions of boundary perturbations. To ensure the comparability of feature distribution between different image regions, the above three indicators uniformly introduce a spatial convolution response reduction mechanism, extract local statistical information by setting a fixed receptive field window, and perform scale normalization and regional response calibration to eliminate the impact of local structural differences in the image. After the feature extraction is completed, the edge perturbation analysis module combines the three types of perturbation features into an edge frequency perturbation feature set as the core input in the structural stability assessment, and simultaneously retains the original directional differential path so that subsequent modules can further distinguish and verify based on the complete directional response sequence. This feature set provides an intermediate representation with the dual information expression capabilities of directional sensitivity and perturbation aggregation, and builds a high-dimensional, quantifiable boundary behavior foundation for enhanced sample screening and structural anomaly judgment.

[0035] The structural fidelity encoding module fuses the regional connectivity vector and edge frequency perturbation features according to the position index to construct a structural integrity description vector, and constructs a structural fidelity kernel function based on the structural integrity description vector to generate a structural fidelity score; In this embodiment, the structural fidelity coding module uses the standard pixel coordinate system of the enhanced image as the index reference to perform corresponding fusion processing on the regional connectivity vector and the edge frequency perturbation feature. The regional connectivity vector is a vectorized encoding of the connectivity attributes of each region determined to be a valid structure in the enhanced image, including multi-dimensional structural indicators such as the label number of the connected region, pixel distribution density, boundary closure score and extended axial ratio. The edge frequency perturbation feature is the boundary direction variation parameter extracted by the edge perturbation analysis module in the enhanced image, including quantitative performance such as the mean square error of the directional gradient, the number of directional jumps and the local extreme value density value. The structural fidelity coding module maps these two sets of features one by one based on the two-dimensional coordinates of each pixel, and fuses and pairs the structural dimension of the regional connectivity vector with the perturbation dimension of the edge frequency perturbation feature through the spatial index mapping table to ensure that the structural features and edge behavior of each analyzed region can have complete information semantic expression at the same pixel level.

[0036] During this process, the structural fidelity encoding module uses a two-dimensional position indexing system based on a standard image coordinate space as the key positioning basis for the fusion operation. Specifically, the module first establishes a unique mapping index for the row and column coordinates of each pixel in the enhanced image. It then extracts a subset of structural features associated with that location from the region connectivity vector, including the identification number of the connected domain to which the pixel belongs, local pixel density statistics within the connected domain, the closedness fit score of adjacent boundary points, and the angular distribution parameters between the main axis of the region and the image coordinate axes. These features are then stored in an index mapping table. Simultaneously, the edge frequency perturbation feature accesses the boundary perturbation data table at the same coordinates to obtain the gradient response data for the pixel in all eight angular directions. From this data, the mean square error of the directional gradient is calculated to quantify the strength of edge fluctuations. The number of directional jumps is counted to identify edge discontinuities, and the density of local extreme points is calculated by comparing the gradient amplitudes between adjacent directional response values. These perturbation features are combined with structural features through positional indices to form a multi-channel fusion structure, in which each set of fused data corresponds strictly to a specific image pixel, ensuring that feature mismatch or spatial drift is avoided during the fusion process. In addition, the structural fidelity coding module also introduces a set of specific fusion rules to control the matching strength of structural features and perturbation features at different scales, including normalized amplitude adjustment, directional sensitivity weighting and position stability dynamic adjustment mechanism, to ensure that the fused feature set has high consistency and anatomical expression ability, laying the foundation for the subsequent construction of structural integrity description vector.

[0037] Based on the above fusion results, the structural fidelity coding module combines each set of paired structural features and perturbation features to construct a structural integrity description vector. The structural integrity description vector is a multidimensional real number vector in a unified format. Its dimensionality is determined by the needs of structural expression. It usually includes a three-dimensional connectivity feature subvector and a three-dimensional edge perturbation subvector. During the construction process, a position stability indicator is introduced as an adjustment factor for position dependence. During the construction process, the structural fidelity coding module uses position alignment normalization and gradient scaling to ensure that the description vector has scale uniformity and direction sensitivity within the same image. To improve the expressive efficiency of the description vector, the module also introduces a boundary continuity score to adjust the weight of the perturbation feature part and stores all fused vectors in a pixel-indexed description vector pool. Each structural integrity description vector represents the joint performance of a pixel position in the enhanced image in the two dimensions of spatial structure and edge stability, providing sufficient numerical support for the construction of the kernel function in the next stage.

[0038] In this implementation, the structural fidelity coding module, based on each pixel-level feature set that has been fused in the enhanced image, converts it into a structural integrity description vector in a unified format and uses a clear vector construction logic for normalization. Specifically, the module first sets the dimensional composition of the vector based on the structural expression requirements, divides the connectivity feature into three sub-vectors: pixel density index within the connected domain, boundary closure index, and axial distribution parameter. Each sub-vector is numerically normalized and uniformly mapped to the structural feature channel. At the same time, the edge perturbation feature is decomposed into three types of perturbation behavior: the mean square error sequence of directional gradients, the angular jump count index, and the extreme value density map, and filled into the perturbation feature channel according to a unified data dimensionality standard. To ensure the scale consistency and directional comparability of the description vectors in different images, the structural fidelity coding module introduces a position alignment normalization mechanism, constructs a centralized coordinate system with reference to the image origin, and adjusts the scaling factor of each perturbation dimension based on edge direction sensitivity. Furthermore, to enhance the guiding effect of edge behavior on the final score, the structural fidelity encoding module constructs a boundary continuity score based on the connectivity span and pixel integrity of each boundary segment. This score is then used to weight the perturbation feature channels, thereby enhancing features in areas of high continuity. Each constructed structural integrity description vector is ultimately stored in a pool of description vectors indexed by image pixels. Each vector strictly corresponds to a unique location in the image and contains complete structural expression semantics and perturbation dynamics, providing a precise numerical foundation for the subsequent construction of the structural fidelity kernel function.

[0039] After constructing the structural integrity description vector, the structural fidelity kernel function is constructed based on the description vector. This kernel function is a two-dimensional mapping function specifically used to quantify structural integrity and edge coherence, and has the ability to respond to and amplify the differences in input vectors. The structural fidelity kernel function captures its overall performance in structural compactness and edge perturbation frequency by inputting the structural integrity description vector into the Gaussian radial basis function and direction cosine similarity fusion model. The structural fidelity encoding module constructs the structural fidelity response surface of the position in the current training image based on the function response value between each vector and multiple standard high-fidelity structural sample vectors. The kernel function does not directly output the label classification result, but constructs a nonlinear mapping relationship based on the change trend of the response curvature between the description vector and the reference vector, thereby distinguishing the reliability of the structure of each region in the enhanced image without relying on artificial thresholding.

[0040] In the specific implementation process, after completing the construction of the structural integrity description vector, the structural fidelity coding module further designs a structural fidelity kernel function based on the description vector to achieve a quantitative assessment of the structural reliability of the target area. The structural fidelity kernel function introduces a fusion response modeling mechanism to mathematically couple the Gaussian radial basis function with the direction cosine similarity function to form a two-dimensional kernel mapping framework with local response enhancement and directional alignment sensitivity. Specifically, a set of standard high-fidelity structural sample vectors verified by experts is first selected as a reference set. This set covers common urinary stone structure types and edge state characteristics, and a multi-point response reference network is constructed with each reference vector in the set as the kernel center. The structural fidelity coding module inputs each structural integrity description vector in the current image into the fusion kernel function respectively, compares the response with all reference samples, measures the difference in structural scale and compactness in the Gaussian radial basis part, and evaluates the consistency of the edge perturbation direction in the direction cosine part. The final output of the kernel function is not an absolute numerical label, but a response curvature map of each vector to be evaluated relative to all reference samples. This response map reflects the distribution offset trend of structure and perturbation in high-dimensional space. By analyzing the peak position, response concentration and change gradient of the curvature map, the structural fidelity coding module can extract scoring features that reflect the authenticity of the structure and the reliability of the boundary, thereby avoiding misjudgment of low-integrity areas and significantly improving the structural credibility discrimination ability of training samples. The design of this kernel function does not rely on any fixed threshold, but is completely based on the natural distribution behavior of the description vector in the response space to construct a discrimination basis. It has good versatility and dynamic adaptability, and can adapt to the structural change performance under different image enhancement conditions.

[0041] The structural fidelity encoding module generates a structural fidelity score based on the output of the structural fidelity kernel function. The structural fidelity score quantifies the effectiveness and completeness of each target region in the enhanced image along the structural dimension, expressed as a continuous floating-point score interval. The structural fidelity encoding module scores each set of descriptor vectors in the entire image and constructs a score distribution histogram, calculating the mean, variance, and extreme value trends of each region type. During the scoring process, the structural fidelity encoding module introduces a dynamic score offset factor to absorb fine-tuning changes in the scoring criteria across training rounds and utilizes the synergy of scores between adjacent regions within the image for quadratic fitting to improve the regional stability of the scores. The resulting structural fidelity score not only provides a basis for determining the training credibility of the enhanced samples but also provides a stable input for the sample path assignment mechanism in the subsequent screening phase, ensuring that the deep learning training phase can achieve precise and controllable sample flow and model updates based on the actual performance of structural integrity.

[0042] During the specific implementation of the structural fidelity score generation process, the structural fidelity encoding module applies a nonlinear mapping calculation to each structural integrity description vector using the previously constructed structural fidelity kernel function, obtaining a continuous numerical score output. This score, expressed as a floating-point number, typically varies within a predefined closed interval and quantifies the comprehensive performance of the target region in terms of structural compactness, connectivity integrity, and edge perturbation stability. To improve the accuracy of the score results, the structural fidelity encoding module first constructs the score results of all descriptors in the image into a two-dimensional score map and then calculates a score distribution histogram for the entire image. This histogram is used to statistically analyze the global trend information of the scores, including the mean, variance, skewness, and kurtosis of the scores, to determine the overall structural distribution of the image. Furthermore, the structural fidelity encoding module introduces a dynamic score offset factor, which automatically adjusts the center position of the score in the current round based on the trend of the score mean in successive training rounds, thereby adapting to the convergence trajectory of the training target. Furthermore, to enhance the regional consistency of scores across image space, the structural fidelity encoding module constructs a pixel adjacency graph and performs a weighted quadratic fitting on the scores of all adjacent regions to smooth local score gradients and suppress interference from isolated, outlier scores. Ultimately, this scoring result serves as a quantitative indicator of the structural stability and training credibility of image enhancement samples. This not only serves as a classification reference for subsequent path assignment strategies but also provides feedback signals on the structural reliability dimension for sample flow and parameter optimization in deep learning models, thereby establishing a structure-driven intelligent training closed-loop mechanism.

[0043] The structural fidelity kernel function is a two-dimensional mapping function used to quantify the response of the structural integrity description vector in two dimensions: structural scale and edge disturbance. It is mainly composed of a combination of Gaussian radial basis function and direction cosine similarity function. The specific calculation method is as follows: Let the structural integrity description vector be ,The standard high-fidelity reference vector set is ,in For any vector to be evaluated , first calculate its relationship with each reference vector The structural difference and directional similarity between them, where the structural difference is calculated by the Gaussian radial basis function formula: ; Directional similarity is calculated using the direction cosine similarity function, and the formula is: ;Structural fidelity kernel function is a weighted combination of the above two: ;in is the fusion weight parameter, is the scale factor. The structural fidelity score is The score represents the degree of fidelity in the dimensions of structural scale and directional consistency, as measured by the maximum or weighted average of the response values ​​of all reference vectors. For example, when the description vector of a pixel has both high radial response and high directional consistency with multiple reference vectors, the structural fidelity score at that location will approach 1, reflecting high structural integrity and low boundary perturbations, making it suitable as a sample for the main training path. Through the kernel function calculation method described above, the system can achieve quantitative modeling of the responsiveness of different image structural regions in feature space, providing a mathematical basis and execution path for the generation of structural fidelity scores.

[0044] The trusted path assignment module classifies the enhanced images into trustworthy levels according to the structural fidelity scores and assigns them to the main training path, auxiliary training path or re-evaluation path respectively; In this embodiment, the trusted path assignment module first classifies the enhanced image into trustworthy levels based on the structural fidelity scores generated by the structural fidelity encoding module. Structural fidelity scores are continuous floating-point numbers, and are assigned to each target region in the enhanced image. Therefore, the trusted path assignment module must statistically analyze the distribution patterns of all scores to construct effective classification boundaries. To ensure that the trustworthy levels are meaningful for model training, the trusted path assignment module employs a multi-parameter joint quantile analysis approach, using the global mean, standard deviation, skewness coefficient, and slope of the score distribution as core references to construct a dynamic scoring three-partition model. This model classifies all enhanced image samples into three levels: high confidence, medium confidence, and low confidence. The high confidence zone includes samples with structural fidelity scores above the global mean plus one standard deviation, the medium confidence zone includes samples with scores within one standard deviation of the mean, and the low confidence zone includes all samples with scores below the mean minus one standard deviation. During the partitioning process, the trusted path assignment module binds the interval label of each sample to the sample, while retaining the original score as the path assignment reference coefficient, providing a numerical basis for subsequent sample flow control and path configuration.

[0045] To improve the adaptability of score boundary setting and the accuracy of sample identification during the trusted level classification process, the trusted path assignment module introduces a density function reconstruction mechanism for the score distribution when constructing a dynamic three-partition model. Kernel density estimation is used to probabilistically model the distribution of structural fidelity scores across the entire enhanced image set. A smooth curve is then constructed based on the actual score distribution. This curve is used to identify inflection points and regions of steep gradient changes in the score distribution, assisting in locating the boundaries of possible structural anomaly clusters. Regarding score statistics, the trusted path assignment module not only collects the global mean and standard deviation of the scores but also calculates the skewness coefficient to measure the asymmetry of the distribution. The slope of the score histogram is also evaluated to measure the degree of attenuation of the score density toward both sides, thereby constructing a multidimensional score structure vector. Based on this, the module uses a sliding window scanning mechanism to identify local density outliers in the score ranking sequence as potential candidate reference boundary locations for classification. Finally, a weighted voting mechanism is used to find the optimal boundary location between global statistics and local trends, achieving adaptive classification into three levels: high confidence, medium confidence, and low confidence. After the partitioning is complete, the module assigns a unique trusted label to each image based on its corresponding structural fidelity score's position within the partition. This label is then bound to its original score value to generate a trusted score index table for subsequent training path assignment. This ensures a strict numerical mapping and logical consistency between the score and path assignment. This process ensures that the trusted level classification results not only reflect the differences in image structural quality across spatial dimensions, but also provide mathematical support that is highly coupled to the training path scheduling logic.

[0046] After completing the confidence level classification, the trusted path assignment module further assigns enhanced images to the main training path, auxiliary training path, or re-evaluation path based on their confidence levels, implementing a structure-quality-driven sample allocation mechanism. In practice, the trusted path assignment module assigns image samples in the high-confidence region directly to the main training path, allowing these images with high structural integrity and low boundary perturbations to participate in the primary gradient learning process, enhancing the model's responsiveness to high-fidelity target structures. Image samples in the medium-confidence region are assigned to the auxiliary training path, where they perform parameter updates using edge optimization or background suppression strategies, enhancing the model's robustness in scenarios with moderate structural fluctuations. Image samples in the low-confidence region are temporarily deferred from the main training path and instead are marked as pending re-evaluation samples, transferring them to the re-evaluation path for structural stability re-verification. The trusted path assignment module implements index pools and data buffering mechanisms for each of the three paths, ensuring that enhanced images, driven by structure scores, flow accurately and stably into the corresponding training branches. This prevents highly structurally perturbed samples from misleading the model's main path weight updates and improves the overall distribution of structural confidence in training samples.

[0047] After images are assigned to different trust levels and enter the path allocation phase, the trusted path dispatch module constructs a mapping function between structural fidelity scores and training paths by building path matching rules. This mapping function employs a combination of interval matching and logical conditional filtering to map the trust level labels carried by image samples to training path parameters one-to-one, thus preventing cross-confusion in path allocation. In the primary training path, the trusted path dispatch module implements a high-priority data access mechanism, granting weighted access to image samples with structural fidelity scores above a preset upper threshold to enhance the primary path's ability to focus on learning high-fidelity structures. In the auxiliary training path, the module regulates the access rhythm of intermediate trust samples through a soft margin screening strategy, dynamically adjusting their participation frequency and weight in the auxiliary path based on the score drift rate, thereby providing gentle training guidance for intermediate samples. For the reassessment path, the trusted path dispatch module implements a cyclic scheduling mechanism to periodically trigger a structural stability review process for samples in the low-trust zone. Based on historical score fluctuations, it determines whether to return samples to the auxiliary path or maintain them in the reassessment channel for further screening. In addition, the module introduces an inter-branch isolation index table between the three path branches. Each type of image sample must undergo dual verification of the trust score and path rules before entering the path, ensuring that the path access action is exclusive and directional. This not only improves the accuracy of the dispatched action, but also avoids the interference diffusion problem caused by low-quality structure during training, thereby realizing the collaborative operation of multi-channel training paths driven by structural fidelity.

[0048] While performing path assignment, the trusted path assignment module also monitors the performance of image samples at each level in model feedback during each training cycle in real time, establishing a bidirectional mapping between structural fidelity scores and training error convergence trends. To achieve continuous adaptive adjustment of the training path assignment mechanism, the trusted path assignment module comprehensively analyzes metrics such as the mean structural fidelity scores, local perturbation entropy, and gradient fluctuation amplitude within training rounds of samples in the main and auxiliary training paths. By constructing a dynamic trusted path adjustment factor, it dynamically fine-tunes the boundary threshold based on the model's learning response to samples of varying structural quality. If images of a certain structural level exhibit non-convergence of training loss or abnormal model response in consecutive training rounds, the trusted path assignment module will immediately adjust its path assignment strategy. For example, if samples in the trusted region demonstrate higher-than-expected recognition contribution, these samples can be transferred to the main training path to participate in core training according to the score upward shifting mechanism, thereby increasing the model's adaptability to marginal structural fluctuations with each round. Through the coupled response adjustment between structural fidelity score and training feedback, the trusted path dispatch module realizes the dynamic mapping optimization between enhanced image structure credibility and training strategy, effectively improving the adaptive learning ability and generalization stability of the deep learning model under multi-structure quality input conditions.

[0049] The trusted path assignment module is implemented based on a dynamic three-partition model constructed using structural fidelity scoring. It assigns image samples to different paths through a series of statistical analyses and logical mapping rules. The module first uses the structural fidelity scoring results generated by the structural fidelity encoding module to calculate the score distribution characteristics of all enhanced images, including the global score mean, score standard deviation, skewness coefficient, and the slope of the score histogram. Based on this, it constructs the score partition boundaries for trustworthy classification. To improve the accuracy and adaptability of the classification, the trusted path assignment module uses kernel density estimation to reconstruct the score distribution curve and identify inflection points and abnormal cluster boundaries in the score density changes. After completing the score statistics and distribution modeling, the module classifies image samples into three levels of high confidence, medium confidence, and low confidence based on the score range. Each image sample is assigned a unique trustworthy level label and is written into the trustworthy score index table along with the original score. Based on the trustworthy classification, the trusted path assignment module constructs path matching mapping rules to assign the labeled image samples to the primary training path, auxiliary training path, or re-evaluation path according to their level. Among them, high-confidence samples are directly assigned to the main training path for convergence training of core parameters; medium-confidence samples are assigned to the auxiliary training path to participate in edge structure recognition optimization; low-confidence samples enter the re-evaluation path, and whether to re-assign them will be decided after further verification of structural stability. Each path branch is equipped with an independent sample index pool and buffer mechanism to ensure the data integrity and index consistency of the samples during the path assignment process. In addition, after the path assignment is completed, the trusted path assignment module continues to monitor the training performance of samples of each level in each training cycle, and dynamically adjusts the division boundary based on the response relationship between the structural fidelity score and the training feedback. Through the sample level adjustment mechanism, the medium-confidence samples with excellent performance are promoted to the main path, or the samples with abnormal fluctuations are transferred to the re-evaluation path, thereby realizing dynamic sample path allocation and closed-loop control driven by structural credibility, with a clear implementation process, statistical basis and interaction characteristics between modules.

[0050] The adaptive screening and control module combines the structural fidelity score distribution and boundary prediction gradient fluctuations in consecutive training rounds to construct a dynamic control factor, and adjusts the parameters of the structural fidelity kernel function accordingly to achieve dynamic updating of the enhanced sample screening criteria.

[0051] In this embodiment, the adaptive screening and control module uses continuous training rounds as the time axis to establish a dynamic distribution trajectory of the structural fidelity score as the training evolves. The structural fidelity score is a real-valued output generated by the structural fidelity kernel function for the structural integrity description vector, with clear regional distribution characteristics and statistically quantifiable features. To capture the trend of score changes, the adaptive screening and control module performs image-level aggregation on the score results of all enhanced images in each round of training, and constructs a multidimensional score statistical matrix including the score mean, score variance, distribution slope, number of extreme values, and score density interval. Based on this matrix, the module identifies the structural deviation behavior of the score in a specific training round, such as the decrease in the concentration of high-confidence areas and the expansion of the score extreme value distribution. In addition, to improve the ability to perceive score fluctuations, the module introduces a score momentum update mechanism, which weights and superimposes the changing trends in the score distribution of multiple consecutive rounds through a sliding window and records them as a score evolution vector to quantify the stability of the score during training and lay the foundation for the construction of dynamic control factors.

[0052] On the basis of the above implementation methods, the adaptive screening and control module introduces a set of fine-grained statistical measurement methods for the dynamic distribution characteristics of the structural fidelity score, further improving the accuracy of the analysis of the score evolution trend. The module first uses the structural fidelity score of the enhanced image as the core data source, and establishes a data structure with the image number and the target area pixel index as the dual primary keys to ensure that each scoring sample has a clear spatial attribution. For the scoring data set generated at the end of each round of training, the adaptive screening and control module uses a batch processing method to summarize the image-level scores. By calculating the statistical mean, mean square error, skewness coefficient, kurtosis index and density distribution curve slope of the current round score, a high-dimensional score feature tensor is constructed, and the round label is retained to support horizontal comparison. The module specifically introduces a dynamic quantile sliding window mechanism to construct a 5-quantile score sequence map for the extreme score area to identify the changing trend of the score concentration. The score momentum update mechanism calculates the first-order difference and second-order derivative based on the score change within the sliding window to obtain the slope and acceleration index of the score trend line, thereby extracting potential score stability areas, mutation points and offset directions, and finally encoding these trend characteristics into score evolution vectors. As a form of time series feature expression, the score evolution vector comprehensively describes the statistical drift behavior of the structural fidelity score with the training process, providing a continuous and sensitive scoring basis for the subsequent construction of dynamic control factors that integrate gradient perturbation indicators. This scoring modeling method based on statistical tensors and sliding window mechanisms has good robustness and scalability, and can significantly improve the adaptive screening and control module's ability to respond to score evolution patterns under different training states.

[0053] While acquiring the score evolution vector, the adaptive screening and control module simultaneously monitors the fluctuations in the predicted behavior of the enhanced image at the structural boundary during each training round to construct a complete control factor structure. Specifically, after each training round, the module extracts the boundary prediction map output by the model and compares it with the original annotated region at the pixel level. It then extracts three boundary fluctuation indicators: boundary offset amplitude, gradient response jump frequency, and local error density distribution. These indicators are normalized at the image scale to form a predicted gradient fluctuation spectrum, which is used to measure the stability of the model's judgment of structural boundaries during the current training phase. Subsequently, the adaptive screening and control module performs a feature-level joint fusion of the score evolution vector and the predicted gradient fluctuation spectrum to construct a dynamic control factor. This control factor is represented by two channels: one channel captures the interval convergence trend of the score dimension, and the other channel records the perturbation strength in the boundary prediction direction. The two channels are linked through a weighted coupling mechanism to form a multidimensional control variable with feedback capability, which serves as the direct input for the subsequent kernel function adjustment process.

[0054] To further enhance the targeted adjustment and feedback loop capabilities of the structural fidelity kernel function parameters, the adaptive screening and control module, based on the construction of the score evolution vector, simultaneously introduces a mechanism to monitor fluctuations in the structural boundary prediction behavior and constructs a dynamic control factor through feature-level joint fusion. In specific implementation, after each training round, the adaptive screening and control module calls the structural boundary prediction map output by the current model and performs pixel-by-pixel registration and alignment of this prediction map with the original manually annotated region based on the pixel coordinate system of the enhanced image. During this alignment, the module first calculates the predicted offset for each boundary pixel—the Euclidean distance between the predicted boundary and the true boundary—and constructs a boundary offset amplitude matrix based on this. Simultaneously, it analyzes the amplitude of the predicted probability gradient map within the boundary neighborhood, extracting the frequency at which the gradient difference between adjacent pixels exceeds a set threshold, thereby quantifying the frequency of predicted gradient response jumps. Furthermore, for locations where the predicted image and the true label are misclassified at the boundary, the module uses a density estimation algorithm to count the number of prediction error points per unit area, forming a local error density distribution map. These three types of boundary fluctuation indicators are normalized and integrated into a predicted gradient fluctuation spectrum to reflect the response stability of the current model in the process of boundary structure recognition. Subsequently, the adaptive screening and control module performs feature-level splicing on the spectrum and the aforementioned score evolution vector to construct a two-dimensional joint feature matrix, and forms a dynamic control factor of the dual-channel structure based on this matrix. Among them, the first channel encodes the interval convergence speed, amplitude change and stability trend of the score dimension on the time axis, and the second channel records the disturbance intensity and directional change in the structural boundary prediction behavior. The two channels establish an associated weight matrix through an adjustable weight coupling model, thereby generating a dynamic control quantity with dual-source information response capability. This control quantity is directly passed to the kernel function adjustment unit as a dynamic input parameter, providing feedback, timing and structural triple guarantees for subsequent parameter tuning, ensuring that the sample screening criteria can dynamically evolve with the model training state, achieving a dual improvement in sample utilization efficiency and structural integrity recognition accuracy.

[0055] The predicted gradient fluctuation spectrum is an important feature expression used in the adaptive screening and control module to quantify the stability of structural boundary prediction behavior. Its specific implementation process includes multiple sequential steps: First, after each round of training, the module calls the structural boundary prediction map output by the current model and aligns it with the original manually annotated region at the pixel level. The prediction offset of each boundary pixel is calculated one by one to reflect the deviation of the model's spatial positioning behavior. Second, the module further analyzes the gradient changes between adjacent pixels in the prediction probability map within the boundary neighborhood, extracting the number of locations where the predicted gradient value changes sharply. This is used to count the jump frequency of the predicted gradient response, thereby evaluating the continuity and stability of the boundary judgment. Third, for misjudgments in the boundary area, the module constructs a local error density distribution based on the number of prediction error points per unit area to reveal the regional error aggregation trend of the model in boundary judgment. The above three boundary fluctuation indicators are all normalized at the image scale and fused based on a unified pixel index structure to generate a predicted gradient fluctuation spectrum that describes the boundary prediction fluctuation pattern. This spectrum has clear spatial positioning properties and multi-dimensional perturbation expression capabilities. It can comprehensively reflect the response characteristics of the model in structural boundary recognition in the current training round, and serve as an important reference for adjusting the parameters of the structural fidelity kernel function. It is used in conjunction with the score evolution vector in the construction process of dynamic regulatory factors, thereby achieving dynamic feedback and fine control of sample screening criteria.

[0056] Based on the aforementioned dynamic control factors, the adaptive screening and control module dynamically updates key parameters in the structural fidelity kernel function to achieve adaptive evolution of the augmentation sample screening criteria. During the kernel function update process, the module dynamically adjusts the scaling factor σ of the Gaussian radial basis function based on the changing trends of the scoring channels in the control factors to adjust the kernel function's response coverage in the scoring space. Furthermore, the module dynamically modifies the tolerance angle parameter θ of the direction cosine similarity function based on the fluctuation amplitude of the boundary perturbation channel to enhance or suppress the influence of structural direction changes on the kernel function response value. To ensure the continuity and convergence of the kernel function adjustment process, the adaptive screening and control module uses an exponential moving average to smooth the influence weights of the control factors and sets upper and lower bounds for the adjustment amplitude. After the update is complete, the module regenerates the kernel function response surface based on the new parameters, thereby updating the augmentation sample fidelity score baseline in real time, achieving self-correction within each round and self-stabilization between rounds of the screening criteria. This dynamic control mechanism ensures that the judgment of structural credibility during model training always evolves in tandem with the actual learning state, significantly improving the rationality of augmentation sample configuration in the training path and the accuracy of structural feature guidance.

[0057] To achieve dynamic evolution and adaptive adjustment of the enhanced sample screening criteria over the training process, the adaptive screening and control module, upon receiving a dynamic control factor constructed from the score evolution vector and the boundary prediction fluctuation spectrum, performs dimensional dynamic adjustment of key parameters of the structural fidelity kernel function. To adjust the structural fidelity kernel function, the module first analyzes the temporal rate of change and current fluctuation direction of the score channel in the control factor. Combining the slope and stability of the control variable across successive training rounds, it applies an exponentially weighted correction to the scaling factor σ in the Gaussian radial basis function. This expands or contracts the kernel's response receptive field in the score space, thereby enhancing or suppressing the influence of high-fidelity structural features in the score as needed. Next, the adaptive screening and control module extracts the mean perturbation intensity and peak transition frequency of the boundary perturbation channel in the control factor. Combined with the angular consistency indicator of the structural boundary, it dynamically adjusts the tolerance angle parameter θ in the directional cosine similarity function to improve the kernel function's sensitivity and directional selectivity to boundary perturbations. To prevent drastic fluctuations in the kernel function parameter adjustment process, the adaptive screening and control module introduces an exponential sliding average filtering mechanism to smooth the weights of the control factors in successive rounds, eliminate local noise interference, and set upper and lower boundary conditions for σ and θ to ensure that parameter changes remain within a stable and controllable range. After completing the parameter update, the module reconstructs the response function surface of the structural fidelity kernel function based on the adjusted σ and θ values ​​and recalculates the structural fidelity score of the enhanced image based on the latest kernel function. Ultimately, the updated scoring baseline not only reflects the dynamic response of the model to structural features during the current training phase but also serves as the core basis for the subsequent trusted path assignment mechanism and training sample configuration strategy. This establishes a self-repairing mechanism for the scoring criteria within the training process, enabling the dynamic and stable evolution of the structural credibility assessment criteria between training rounds, thereby improving the model's adaptive screening capabilities and structural guidance accuracy.

[0058] The dynamic control factor is a set of dual-channel adjustment parameters generated by the fusion of the score evolution vector and the predicted gradient fluctuation spectrum at the feature level. It is used to guide the dynamic adjustment of key parameters in the structural fidelity kernel function. Its calculation process includes time series trend modeling of the score channel and spatial fluctuation quantification of the boundary perturbation channel. The specific calculation method is as follows: Assume that the continuous training round is , No. The set of structural fidelity scores for each image in a round is , then the score evolution vector can be obtained from the score mean , score standard deviation , skewness coefficient and rating momentum (recorded as first-order difference) The predicted gradient fluctuation spectrum is composed of the boundary offset amplitude matrix , gradient jump frequency diagram and error density plot It consists of three parts, which are normalized into 、 、 Finally, the dynamic regulatory factor Expressed as: ,in to is a preset weight coefficient used to control the influence of each dimension feature in dynamic adjustment. For example, if the mean score in the 10th round of model training drops from 0.76 in the previous round to 0.68, ,At the same time, the jump frequency of the prediction graph in the boundary area increases significantly, resulting in Increases, the dynamic control factor will comprehensively reflect the downward trend of the score and the instability of the boundary prediction, thereby driving the structural fidelity kernel function to reduce the Gaussian kernel scale factor , improve the response selectivity of the kernel function to high-fidelity structures, and adjust the cosine function tolerance angle , enhancing the model's sensitivity to directional perturbations. The construction process of this regulatory factor provides a feedback mechanism with clear mathematical expression and strong structural consistency, enabling the structural fidelity scoring system to be optimized in real time with the model training status, thereby improving the adaptability and reliability of enhanced image sample screening.

[0059] The above embodiments can be implemented in whole or in part via software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product comprises one or more computer instructions or computer programs. When loaded or executed on a computer, the processes or functions described in the embodiments of this application are fully or partially performed. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium accessible by a computer or a data storage device such as a server or data center that contains a collection of one or more available media. The available medium can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media. The semiconductor media can be a solid-state drive.

[0060] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0061] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0062] In the several embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the embodiments described above are merely illustrative. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. The mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0063] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0064] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0065] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. A urinary stone image recognition and analysis system based on deep learning, characterized by: It includes a structural connectivity mapping module, an edge perturbation analysis module, a structural fidelity encoding module, a trusted path dispatching module, and an adaptive screening and control module; The structural connectivity mapping module performs pixel-level mapping between the enhanced urinary stone image and its original annotated area, extracts the structural connectivity information of the corresponding area in the enhanced image, and generates a regional connectivity vector; The edge perturbation analysis module constructs an edge direction difference matrix based on the regional connectivity vector, performs gradient direction analysis on the boundary pixels in the enhanced image, and extracts edge frequency perturbation features; The structural fidelity encoding module fuses the regional connectivity vector and edge frequency perturbation features according to the position index to construct a structural integrity description vector, and constructs a structural fidelity kernel function based on the structural integrity description vector to generate a structural fidelity score; The trusted path assignment module classifies the enhanced images into trustworthy levels according to the structural fidelity scores and assigns them to the main training path, auxiliary training path or re-evaluation path respectively; The adaptive screening and control module constructs a dynamic control factor based on the distribution of structural fidelity scores and boundary prediction gradient fluctuations in consecutive training rounds, and adjusts the parameters of the structural fidelity kernel function accordingly.

2. The urinary stone image recognition and analysis system based on deep learning according to claim 1, characterized in that: In the structural connectivity mapping module, a pixel-level mapping relationship between the enhanced image and the original annotated area is established by presetting reference reference points in the original image and recording the geometric transformation parameters corresponding to the enhancement processing. The enhanced image coordinates are restored to the original image reference frame through inverse affine transformation to achieve target area alignment. After the mapping is completed, the structural connectivity mapping module extracts the target area corresponding to the annotated area in the enhanced image, constructs a binary graph structure and introduces the eight-neighborhood connection rule to identify the connected area. The pixel composition density, contour closure degree, area ratio and central axis distribution are calculated for each connected area to obtain the corresponding structural connectivity information. The structural connectivity information is used to generate a regional connectivity vector including the number of connected blocks, the pixel coverage of the largest connected block, the edge smoothness, the gradient difference between the inner and outer boundaries, the center of gravity offset rate and the main axis distribution of the structural direction.

3. The urinary stone image recognition and analysis system based on deep learning according to claim 2, characterized in that: The edge perturbation parsing module extracts the boundary pixel set of the target area in the enhanced image based on the target pixel connectivity label, boundary pixel identification mark and structural closure index contained in the regional connectivity vector, and constructs a direction perception window with each boundary pixel as the center. The direction perception window defines eight direction sequences according to the image coordinate system and numbered them from 1 to 8 in sequence. At the same time, the relative position offset relationship between the adjacent pixels and the current boundary pixel in each direction is recorded to generate a direction mapping index table; the edge perturbation parsing module uses central difference to calculate the grayscale difference between the current pixel and the adjacent pixels in each direction according to the direction mapping index table and performs normalization processing to construct an edge direction difference matrix. Each row of the edge direction difference matrix corresponds to a boundary pixel, and each column corresponds to a direction. The value in the cell is the standardized grayscale difference response value in that direction, which is used to describe the gradient change behavior of the boundary in the spatial direction.

4. The urinary stone image recognition and analysis system based on deep learning according to claim 3, characterized in that: The edge perturbation analysis module performs gradient direction analysis based on the directional response sequence of each boundary pixel in the edge direction difference matrix, and uses the angular interpolation mechanism to construct a quasi-continuous angle sequence to improve the sensitivity of directional change modeling. The angular interpolation mechanism fills the intermediate angle points between the original eight directions through cubic spline interpolation, and applies the Laplace directional difference function based on the interpolation result to detect the local gradient mutation position and directional jump frequency; on this basis, the edge frequency perturbation features are extracted. The edge frequency perturbation features include three indicators: the mean square error of the directional gradient, the number of directional jumps, and the local gradient peak distribution. The spatial convolution response reduction mechanism is uniformly used for regional scale normalization and response calibration to form an edge frequency perturbation feature set with directional sensitivity and perturbation aggregation.

5. The urinary stone image recognition and analysis system based on deep learning according to claim 4 is characterized in that: The structural fidelity coding module uses the two-dimensional pixel coordinates of the enhanced image as a unified index benchmark, and obtains the regional connectivity vector and edge frequency perturbation features corresponding to the pixel position coordinates for each pixel position. The regional connectivity vector includes the number of the connected region to which the pixel belongs, the pixel density of the connected region, the boundary closure score and the axial structure distribution angle. The edge frequency perturbation features include the mean square error of the directional gradient, the number of directional jumps and the density of local extreme points. By establishing a spatial position index mapping table, the structural attributes in the regional connectivity vector and the directional response information in the edge frequency perturbation features are fused one-to-one at the same coordinate position. By introducing normalization adjustment, boundary direction weighting and spatial stability restriction mechanisms, a fused multi-dimensional pixel-level feature combination is formed, providing a corresponding fused dataset for the subsequent construction of the structural integrity description vector.

6. The urinary stone image recognition and analysis system based on deep learning according to claim 5, characterized in that: After fusing the regional connectivity vector and the edge frequency perturbation features, the structural fidelity encoding module constructs a structural integrity description vector in a unified data dimension format. The structural integrity description vector includes a structural dimension sub-vector composed of the pixel density of the connected region, the boundary closure score, and the structural main axis angle, and a perturbation dimension sub-vector composed of the mean square error of the directional gradient, the number of jumps, and the density of extreme points. The structural integrity description vector is encoded in a floating-point format, and a centralized coordinate alignment, gradient scale normalization, and boundary continuity weighted adjustment mechanism are introduced for all vectors to ensure that all description vectors have a unified scale and stable directional feature expression in the same image space. They are then stored in the structural integrity description vector pool according to pixel index, providing a continuous feature input basis for the structural fidelity kernel function.

7. The urinary stone image recognition and analysis system based on deep learning according to claim 6, characterized in that: The structural fidelity coding module constructs a structural fidelity kernel function based on the structural integrity description vector, where the structural fidelity kernel function is composed of a combination of Gaussian radial basis function and direction cosine similarity function. The former is used to measure the difference in structural scale and density between the description vector and the standard structure sample, and the latter is used to measure the consistency of the edge direction characteristics between the description vector and the standard structure sample. High-fidelity structural samples that have been manually verified are selected to construct a reference vector set, and the structural fidelity kernel function is used to calculate the response of each structural integrity description vector to generate a structural fidelity response surface. The structural fidelity score is output according to the numerical gradient and change trend of the response distribution. The score value represents the numerical performance of each pixel in the current image in the structural stability dimension in the form of a continuous floating-point interval.

8. The deep learning-based urinary stone image recognition and analysis system according to claim 7, characterized in that: The trusted path dispatch module divides the enhanced images into trust levels based on the structural fidelity score generated by the structural fidelity encoding module. The structural fidelity score is a continuous floating-point number. The trusted path dispatch module constructs a dynamic scoring three-partition model by statistically analyzing the global mean, standard deviation, skewness coefficient and score distribution slope of all scores using a multi-parameter joint quantile analysis method. The enhanced image samples are divided into three levels: high confidence zone, medium confidence zone and low confidence zone. Each image sample is bound to its own trust zone label and original score as a reference coefficient for path dispatch; image samples in the high confidence zone are assigned to the main training path, image samples in the medium confidence zone are assigned to the auxiliary training path, and image samples in the low confidence zone are marked as samples to be re-evaluated and assigned to the re-evaluation path. Independent index pools and data buffering mechanisms are configured for each of the three types of paths.

9. The deep learning-based urinary stone image recognition and analysis system according to claim 8, characterized in that: The adaptive screening and control module constructs a dynamic control factor based on the structural fidelity score distribution and boundary prediction gradient fluctuation in continuous training rounds, and dynamically updates the key parameters in the structural fidelity kernel function according to the dynamic control factor. The structural fidelity score is generated by the structural fidelity kernel function for the structural integrity description vector. The structural fidelity score is aggregated and statistically analyzed at the image level in each training round to form a score statistical matrix containing the score mean, score variance, distribution slope, number of extreme values ​​and score density interval. While obtaining the score evolution vector, the adaptive screening and control module extracts the boundary offset amplitude, gradient response jump frequency and local error density distribution based on the pixel-level difference comparison between the boundary prediction map of the model after each round of training and the original annotated area to form a predicted gradient fluctuation spectrum. The score evolution vector and the predicted gradient fluctuation spectrum are then fused to form a dual-channel dynamic control factor, which is used to adjust the scale factor of the Gaussian radial basis function and the tolerance angle parameter of the direction cosine similarity function in the structural fidelity kernel function respectively.

Images