A pattern recognition method and system for urological stone imagery

Abstract

The present application relates to the field of urinary surgery stone image pattern recognition, and provides a urinary surgery stone image pattern recognition method and system. By acquiring the urinary surgery stone image, the stone image is recognized to determine the suspected stone area and its boundary area; for each pixel point in the boundary area, based on its gray scale feature and neighborhood feature, the probability of the pixel point belonging to the stone is calculated to generate a boundary probability map; based on the boundary probability map, the size, volume and Hounsfield unit value of the stone are determined through a preset multi-scale feature fusion manner. Thus, the problem of missed diagnosis, misdiagnosis and Hounsfield unit value deviation of small stones under low-dose CT images in the prior art can be effectively solved, and the accuracy and robustness of urinary surgery stone image pattern recognition are significantly improved.

Description

Technical Field

[0001] This application relates to the field of pattern recognition in urological stone images, and more specifically, to a method and system for pattern recognition in urological stone images. Background Technology

[0002] In urological clinical practice, the analysis of stone images is a crucial step in the diagnosis and treatment of urinary system stones. Traditional stone imaging diagnosis relies primarily on manual interpretation by physicians, which is not only inefficient but also susceptible to the influence of physician experience and fatigue, leading to missed or misdiagnosed small stones. To improve diagnostic efficiency and accuracy, intelligent pattern recognition systems have been introduced into the auxiliary diagnosis of urological stone images. However, existing intelligent recognition systems face numerous challenges in practical clinical applications. For example, under low-dose CT scans, the signal-to-noise ratio of the images is significantly reduced, and a large amount of particulate noise is present. Because existing systems are primarily trained on clear standard-dose images, they struggle to reliably identify the outlines of small stones in complex noisy backgrounds, leading to missed diagnoses of stones smaller than 3 mm in diameter or misidentification of areas with high local noise as suspected stones, thus increasing the false alarm rate. Furthermore, noise interference can cause deviations in the calculated Heinz unit values, affecting the physician's initial assessment of the stone's composition.

[0003] Furthermore, existing systems also struggle with calcifications near the anatomical structures of the urinary system. For example, stones in the distal ureter and pelvic venous stones may appear similar on imaging, but current systems lack the ability to integrate anatomical knowledge for comprehensive reasoning. On low-dose CT images, these subtle distinguishing features become blurred, often leading the system to misdiagnose venous stones adjacent to the ureter as ureteral stones, potentially causing unnecessary examinations and treatments. Moreover, in enhanced CT images, especially during the excretion phase of CT urography (CTU), urine containing contrast agent will make the renal pelvis, ureter, and bladder appear as high-density bright white. Current intelligent recognition systems primarily identify stones by searching for high-density substances within the urinary system; therefore, they may misinterpret contrast agent as large or multiple stones, rendering the system completely ineffective when processing these common images and losing its diagnostic value.

[0004] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0005] This application discloses a pattern recognition method and system for urological stone images, aiming to solve the technical problems of existing intelligent recognition systems for urological stone images, such as difficulty in identifying small stones in low-dose CT images, susceptibility to noise interference leading to Henle unit value deviation, difficulty in distinguishing calcifications near anatomical structures, and misidentification of contrast agents as stones in enhanced CT images. The technical solution of this application is as follows:

[0006] In a first aspect, this application discloses a pattern recognition method for urological stone images, the method comprising: Acquire images of urological stones, identify the stone images to determine the suspected stone area and its boundary area; For each pixel within the boundary region, the probability that the pixel belongs to a stone is calculated based on its grayscale features and neighborhood features, and a boundary probability map is generated. Based on the boundary probability map, the size, volume, and Heinz unit value of the stone are determined by a preset multi-scale feature fusion method.

[0007] Furthermore, in the above-mentioned pattern recognition method for urological stone images, the suspected stone region includes a core region and a boundary region; identifying the stone image to determine the suspected stone region and its boundary region specifically includes: identifying at least one first stone region from the stone image according to a predetermined range of Henlein unit values; for any first stone region, expanding a first pixel layer outward to form an extension region, and contracting a second pixel layer inward to form a core region; and determining the region between the extension region and the core region as the boundary region of the suspected stone region.

[0008] Furthermore, in the above-mentioned pattern recognition method for urological stone images, the grayscale features include the Henle unit value of a pixel; the neighborhood features include the average Henle unit value of pixels in the neighborhood centered on the pixel, the standard deviation of the Henle unit value of pixels in the neighborhood centered on the pixel, and the grayscale gradient intensity in the neighborhood centered on the pixel. Based on its grayscale features and neighborhood features, the probability that a pixel belongs to a kidney stone is calculated, and a boundary probability map is generated. Specifically, this includes: for each pixel within the boundary region, obtaining the Heinz unit value of that pixel, as well as the Heinz unit mean, Heinz unit standard deviation, and grayscale gradient intensity within the first neighborhood centered on that pixel; inputting the Heinz unit value, Heinz unit mean, Heinz unit standard deviation, and grayscale gradient intensity into a probability function to obtain the pixel feature probability value that the pixel belongs to a kidney stone; and combining all pixels and their corresponding pixel feature probability values ​​to form the boundary probability map.

[0009] Building upon the above, this application further proposes a method for determining the size, volume, and Henle unit value of a stone based on a boundary probability map and a preset multi-scale feature fusion approach. Specifically, this includes: acquiring a low-resolution image of the stone and estimating the first size and first volume of the stone based on the low-resolution image; estimating the second size and second volume of the stone based on the boundary probability map and the spatial contribution of each pixel within the suspected stone region using probability weighting; the pixel feature probability value of pixels within the core region is 1; weighted fusion of the first and second sizes to obtain the size of the stone; and weighted fusion of the first and second volumes to obtain the volume of the stone.

[0010] In some preferred embodiments, the size, volume, and Henle unit value of the stone are determined based on the boundary probability map by a preset multi-scale feature fusion method. The method also includes: locally adaptive window width and window level adjustment of the suspected stone area; selecting boundary pixels with pixel feature probability values ​​higher than a preset threshold based on the boundary probability map; and obtaining core pixels within the core area. Calculate the average value of the boundary pixels and the core pixels in Henle units as the Henle unit value of the stone.

[0011] As a technological improvement, in the above-mentioned pattern recognition method for urological stone images, after identifying the stone image to determine the suspected stone area and its boundary area, the method further includes: Anatomical structure identification is performed on stone images to determine at least one target anatomical structure in the images; the target anatomical structure includes at least one of the kidney, ureter, bladder, and vascular structures.

[0012] To enhance functionality, the above-mentioned pattern recognition method for urological stone images, in calculating the probability that a pixel belongs to a stone, further includes: determining a first spatial probability value for the pixel based on the relative positional relationship between any pixel within the boundary region and the target anatomical structure; pre-assigning the first spatial probability value according to the anatomical structure to which the pixel belongs; adjusting the first spatial probability value based on the physiological characteristic correlation between the pixel and the target anatomical structure to obtain a second spatial probability value; and determining the probability that the pixel belongs to a stone based on the second spatial probability value and the pixel feature probability value.

[0013] Based on the above, this application further proposes adjusting the first spatial probability value according to the physiological characteristic correlation between the pixel and the target anatomical structure, specifically including: in response to the Euclidean distance from the pixel to the nearest blood vessel wall being less than a preset distance threshold, determining that the pixel has a physiological characteristic correlation with the blood vessel structure, and reducing the first spatial probability value; wherein, reducing the first spatial probability value includes: in response to the Euclidean distance from the pixel to the nearest blood vessel wall being less than a preset distance threshold, multiplying the first spatial probability value by a preset attenuation factor β to obtain a second spatial probability value; wherein 0 < β < 1.

[0014] Furthermore, in the above-mentioned pattern recognition method for urological stone images, the probability of a pixel belonging to a stone is determined based on the second spatial probability value and the pixel feature probability value. Specifically, the probability of a stone is: P_final = (1 - α). P_image + α W_space; where P_image is the pixel feature probability value; W_space is the second spatial probability value; α is the fusion coefficient, and 0 ≤ α ≤ 1; the fusion coefficient α is adaptively set according to the quality of the stone image: in response to the proportion of artifact areas in the stone image being higher than the preset area threshold, the fusion coefficient α is set to a value greater than 0.5.

[0015] Secondly, this application also discloses a pattern recognition system for urological stone images. The system includes: an acquisition and recognition module for acquiring urological stone images and recognizing the stone images to determine suspected stone regions and their boundary regions; a probability generation module for calculating the probability that each pixel in the boundary region belongs to a stone based on its grayscale features and neighborhood features, and generating a boundary probability map; and a determination module for determining the size, volume, and Heinz unit value of the stone based on the boundary probability map through a preset multi-scale feature fusion method.

[0016] Beneficial effects This application discloses a pattern recognition method for urological stone images. By acquiring urological stone images and identifying suspected stone areas and their boundaries, it can initially locate the possible range of stone presence. Based on this, for each pixel within the boundary area, the probability that the pixel belongs to a stone is calculated based on its grayscale features (such as Henlein unit value) and neighborhood features (such as the mean, standard deviation, and grayscale gradient intensity of neighboring pixels in Henlein units), generating a boundary probability map. This multi-feature fusion probability calculation method effectively overcomes the problems of low signal-to-noise ratio and large particulate noise interference in low-dose CT images caused by single features, improving the stability of small stone contour recognition and reducing the risk of misidentifying local noise as a stone. Finally, based on the boundary probability map, the size, volume, and Henlein unit value of the stone are determined through a preset multi-scale feature fusion method. This multi-scale fusion strategy combines information from different scales, making the estimation of stone size and volume more accurate. Simultaneously, by calculating the Henlein unit values ​​of high-probability pixels, it effectively reduces the bias of noise in the calculation of Henlein unit values, thereby improving the accuracy of doctors' initial judgment of stone composition. In summary, the method of this application can effectively solve the problems of missed diagnosis, misdiagnosis, and Henle unit value deviation of small stones under low-dose CT imaging in the prior art, and significantly improve the accuracy and robustness of urological stone imaging pattern recognition. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the steps of the pattern recognition method for urological stone images disclosed in an embodiment of the present invention. Figure 2 This is a schematic diagram of the pattern recognition system for urological stone imaging disclosed in an embodiment of the present invention. Detailed Implementation

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which these embodiments belong; the terminology used herein and in the specification of the application is for the purpose of describing particular embodiments only and is not intended to limit these embodiments; the terms "comprising" and "having," and any variations thereof, in the specification of these embodiments and the foregoing drawings, are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification of these embodiments and the foregoing drawings are used to distinguish different objects, not to describe a particular order.

[0020] The implementation details of the technical solution in this embodiment are described in detail below: In urological clinical practice, the analysis of stone images is a crucial step in the diagnosis and treatment of urinary system stones. Traditional stone imaging diagnosis relies primarily on manual interpretation by physicians, which is not only inefficient but also susceptible to the influence of physician experience and fatigue, leading to missed or misdiagnosed small stones. To improve diagnostic efficiency and accuracy, intelligent pattern recognition systems have been introduced into the auxiliary diagnosis of urological stone images. However, existing intelligent recognition systems face numerous challenges in practical clinical applications. For example, under low-dose CT scanning protocols, the signal-to-noise ratio of the images is significantly reduced, and a large amount of particulate noise exists in the images, leading to missed diagnoses of small stones or misidentification of areas with high local noise as suspected stones. Simultaneously, existing systems also have difficulty identifying calcifications near the anatomical structures of the urinary system, and in contrast-enhanced CT images, they may misinterpret contrast agents as stones, causing system failure.

[0021] In response, this application proposes a pattern recognition method for urological stone images, such as... Figure 1 As shown, the method includes: S101, acquire images of urological stones, identify the stone images to determine the suspected stone area and its boundary area; S102, For each pixel within the boundary region, calculate the probability that the pixel belongs to a stone based on its grayscale features and neighborhood features, and generate a boundary probability map. S103, based on the boundary probability map, determines the size, volume and Heinz unit value of the stone through a preset multi-scale feature fusion method.

[0022] This application aims to overcome the limitations of traditional methods and existing intelligent recognition systems in identifying stones in complex imaging environments by introducing boundary probability maps and multi-scale feature fusion mechanisms, thereby improving the accuracy and diagnostic rate of stone identification. In particular, it can provide more reliable auxiliary diagnostic information when processing low-dose CT, calcifications near anatomical structures, and enhanced CT images.

[0023] To better understand the technical solution proposed in this application, some key terms are explained first. Urological stone images generally refer to image data obtained through medical imaging techniques such as computed tomography (CT), X-ray, and ultrasound, used for diagnosing urinary system stones. These images contain information such as the shape, location, and density of the stones. A suspected stone region refers to the area in the image where stones may be present; its identification is the first step in pattern recognition. The boundary region refers to the transition area between the suspected stone region and the surrounding normal tissue; accurate analysis of this region is crucial for accurate stone identification. Gray-scale features refer to the brightness information of a pixel, such as the Hounsfield Unit (HU) value, which reflects the degree of tissue absorption of X-rays and is an important basis for distinguishing stones from surrounding tissue. Neighborhood features refer to the characteristics of the area surrounding a pixel, such as the average Hounsfield Unit value, standard deviation, and gray-scale gradient intensity of pixels in the neighborhood. These features provide local texture and structural information, helping to more accurately determine whether a pixel belongs to a stone. A boundary probability map is an image where the value of each pixel represents the probability that the pixel belongs to a kidney stone, providing refined probabilistic information for subsequent determination of kidney stone parameters. Multi-scale feature fusion refers to a method that combines information from different resolutions or analytical scales to comprehensively judge kidney stone characteristics, aiming to improve the robustness and accuracy of identification.

[0024] The pattern recognition method for urological stone images proposed in this application is based on the refined analysis of stone images and the fusion of multi-scale features.

[0025] First, it is necessary to acquire images of urological stones and identify these images to determine suspected stone areas and their boundaries. In practice, acquiring stone images can be achieved in several ways. For example, raw DICOM (Digital Imaging and Communications in Medicine) format image data can be directly obtained from CT scans, or stored images can be retrieved from a PACS (Picture Archiving and Communication System). Identifying suspected stone areas and their boundaries can be done using threshold-based segmentation methods. For instance, high-density areas can be initially screened as potential stone areas based on a preset range of Henlein unit values ​​(e.g., 100 HU to 1000 HU). Another approach is to utilize pre-trained deep learning models, such as U-Net or Mask R-CNN, to perform semantic segmentation on the images, thereby identifying suspected stone areas. These models, by learning from large amounts of labeled data, can automatically identify the shape and location of stones.

[0026] Secondly, for each pixel within the boundary region, the probability that the pixel belongs to a kidney stone is calculated based on its grayscale features and neighborhood features, generating a boundary probability map. After obtaining the boundary region, each pixel within that region needs to be analyzed in detail. Grayscale features can simply be the Heinz unit value of the pixel itself. Neighborhood features can include the Heinz unit mean and standard deviation of pixels within a 3x3 or 5x5 neighborhood centered on the pixel, as well as the grayscale gradient intensity calculated using the Sobel or Prewitt operator. These features can be input into a preset probability function, such as a classifier based on logistic regression, support vector machine (SVM), or a small neural network, which outputs the probability value of each pixel belonging to a kidney stone. Combining all pixels and their corresponding probability values ​​generates the boundary probability map.

[0027] Finally, based on the boundary probability map, the size, volume, and Henlein unit value of the stone are determined through a preset multi-scale feature fusion method. After obtaining the boundary probability map, its fine-grained probability information can be used to more accurately determine various parameters of the stone. One approach is to first binarize the boundary probability map, set a probability threshold (e.g., 0.5), mark pixels with probabilities higher than the threshold as stones, and then calculate the size and volume of the stone through connected component analysis. Another approach is to overlay the boundary probability map with the original image and combine morphological operations (such as dilation and erosion) to optimize the contour of the stone, thereby calculating a more accurate size and volume. For determining the Henlein unit value, the average Henlein unit values ​​of pixels with high probability values ​​(e.g., probability values ​​greater than 0.7) in the boundary probability map can be used as the Henlein unit value of the stone.

[0028] The pattern recognition method for urological stone images proposed in this application works by using multi-stage, multi-dimensional information processing to achieve accurate identification and parameter quantification of urological stones. First, by acquiring urological stone images and preliminarily identifying suspected stone areas and their boundaries, a foundation is laid for subsequent refined analysis. This step can quickly locate areas in the image where stones may exist, avoiding complex pixel-by-pixel calculations of the entire image, thereby improving processing efficiency.

[0029] Next, for the initially determined boundary region, this application introduces a mechanism to calculate the probability of it belonging to a kidney stone based on pixel grayscale features and neighborhood features. Traditional methods often rely solely on a single grayscale threshold, which is easily affected by noise and partial volumetric effects, leading to blurred boundaries or misjudgments. By comprehensively considering the pixel's own Henlein unit value (grayscale feature) and the statistical information of its surrounding pixels (neighborhood features, such as mean, standard deviation, and gradient strength), the local characteristics of the pixel can be captured more comprehensively. For example, pixels inside a kidney stone usually have a high Henlein unit value with a gradual change, while pixels at the edge of the kidney stone may exhibit a large grayscale gradient. Inputting these features into the probability function generates a fine probability value for each pixel, forming a boundary probability map. This probability map not only provides the possibility of the presence of a kidney stone but also reflects the spatial distribution of this possibility, providing a reliable basis for subsequent accurate measurements.

[0030] Finally, this application utilizes boundary probability maps to determine the size, volume, and Henlein unit value of the stone through a pre-defined multi-scale feature fusion method. This multi-scale fusion strategy aims to combine information from different levels to improve the robustness and accuracy of the measurement. For example, the overall size of the stone can be quickly estimated at low resolution, while on a high-resolution boundary probability map, the fine contour of the stone can be probabilistically weighted, thus calculating its size and volume more accurately. By calculating the average Henlein unit value of pixels in high-probability regions of the boundary probability map, a more representative Henlein unit value of the stone can be obtained, which helps doctors make a preliminary judgment on the stone's composition. The entire process organically combines preliminary identification, fine-grained probability modeling, and multi-scale fusion measurement through interconnected steps, enabling this application to effectively address the challenges of complex imaging environments and provide more accurate and reliable stone diagnostic information.

[0031] The pattern recognition method for urological stone imaging proposed in this application demonstrates significant technological advancements and innovations in several aspects compared to existing technologies. Traditional stone imaging diagnosis primarily relies on manual image interpretation by physicians. This method is not only inefficient but also susceptible to the influence of physician experience and fatigue, leading to missed or misdiagnosed small stones. This application significantly improves diagnostic efficiency and accuracy by introducing an automated pattern recognition method.

[0032] Existing intelligent recognition systems suffer from significantly reduced signal-to-noise ratios (SNR) in low-dose CT scans, with a large amount of granular noise in the images. This leads to missed diagnoses of tiny stones smaller than 3 mm in diameter, or misidentification of areas with high local noise as suspected stones, thus increasing the false alarm rate. Furthermore, noise interference can cause deviations in the calculated Heinz unit values. This application addresses this issue by calculating the probability that each pixel within the boundary region belongs to a stone based on its grayscale and neighborhood features, generating a boundary probability map. This allows for more precise capture of the local features of stones, effectively distinguishing stones from noise, thereby reducing missed diagnoses and false alarm rates. The introduction of neighborhood features enables the system to utilize information from surrounding pixels for comprehensive judgment when facing low SNR images, improving the stability of tiny stone contour recognition.

[0033] Furthermore, existing systems face difficulties in identifying calcifications near the anatomical structures of the urinary system. For example, stones in the distal ureter and pelvic venous stones may appear similar on imaging, and the system lacks the ability to integrate anatomical knowledge for comprehensive reasoning. This application generates a boundary probability map and, based on this map, determines the size, volume, and Henlein unit value of the stone through a pre-defined multi-scale feature fusion method, providing more accurate basic data for subsequent differential diagnosis based on anatomical information. Although this independent solution does not directly incorporate anatomical information, the precise stone parameters it provides offer physicians a more reliable basis for differential diagnosis.

[0034] Furthermore, in enhanced CT images, especially excretion-phase images from CT urography (CTU), urine containing contrast agent appears as a high-density bright white color in the renal pelvis, ureter, and bladder. Existing intelligent recognition systems may misidentify the contrast agent as large or multiple stones. This application, through detailed probability calculation of the boundary region and combined with multi-scale feature fusion, can more effectively identify the true boundary of the stone and reduce the interference of the contrast agent in stone identification. By comprehensively analyzing the grayscale features of pixels and neighborhood features, the system can better distinguish the subtle differences in texture and density between the contrast agent and the real stone, thereby avoiding misjudgment.

[0035] In summary, this application, by introducing boundary probability maps and multi-scale feature fusion mechanisms, demonstrates superior performance compared to existing technologies in handling noise in low-dose CT images, distinguishing calcifications near anatomical structures, and dealing with contrast agent interference in enhanced CT images, providing a more accurate and reliable solution for the auxiliary diagnosis of urological stones.

[0036] Specifically, in the above-mentioned pattern recognition method for urological stone images, the step of identifying the stone image to determine the suspected stone region and its boundary region can be further refined. The suspected stone region includes a core region and a boundary region; identifying the stone image to determine the suspected stone region and its boundary region includes: identifying at least one first stone region from the stone image according to a predetermined range of Henlein unit values; for any first stone region, expanding a first pixel layer outward to form an extension region, and contracting a second pixel layer inward to form a core region; determining the region between the extension region and the core region as the boundary region of the suspected stone region.

[0037] The suspected stone area is divided into a core region and a boundary region to allow for refined processing of different parts of the stone. Specifically, when identifying the stone image, at least one first stone region is first identified from the stone image based on a preset range of Henlein unit values. This preset range of Henlein unit values ​​typically corresponds to the typical density value of the stone in CT images; for example, calcified stones usually have higher Henlein unit values.

[0038] Furthermore, for any identified first stone region, an extensional region can be formed by expanding the first pixel layer outward. This extensional region includes the first stone region and its immediately surrounding pixels. Simultaneously, a core region can be formed by contracting the second pixel layer inward. This core region represents the dense central portion of the stone. Thus, the area between the extensional region and the core region is defined as the boundary region of the suspected stone region. This boundary region typically includes pixels at the interface between the stone and surrounding tissue; its grayscale and neighborhood features may be more complex, requiring more refined probability calculations.

[0039] This application's scheme first identifies the possible location of the stone, i.e., the first stone region, by utilizing the typical Henlein unit value characteristics of stones in images. Based on this, the stone region is further subdivided into a core region and a boundary region through the expansion and contraction of pixel layers. The core region represents the denser, more clearly defined part of the stone, while the boundary region encompasses pixels where the stone edge transitions to the surrounding tissue. These pixels often have features intermediate between those of stones and non-stones, representing a challenging and crucial aspect of pattern recognition. This division allows subsequent probability calculations to be optimized for the characteristics of different regions, particularly providing a more refined probability assessment for pixels in the boundary region, thereby improving the accuracy of stone identification.

[0040] The above-mentioned technical solution enables a more refined and structured division of suspected stone areas in urological stone images. By clearly distinguishing between the core and boundary regions, targeted feature analysis and probability calculations can be performed on different parts of the stone. Especially in the boundary region, the transition features between the stone and surrounding tissues can be captured more accurately, thus laying the foundation for the precise determination of stone size, volume, and Henle unit values, effectively improving the accuracy and robustness of stone identification.

[0041] Specifically, in some of the above embodiments, the calculation method for the probability that a pixel belongs to a stone has been further refined and explained. The grayscale feature includes the Heinz unit value of the pixel; the neighborhood feature includes the average Heinz unit value of the pixels in the neighborhood centered on the pixel, the standard deviation of the Heinz unit value of the pixels in the neighborhood centered on the pixel, and the grayscale gradient intensity in the neighborhood centered on the pixel. The step of calculating the probability that a pixel belongs to a kidney stone based on its grayscale features and neighborhood features, and generating a boundary probability map, includes: for each pixel in the boundary region, obtaining the Heinz unit value of the pixel, and the Heinz unit average, Heinz unit standard deviation, and grayscale gradient intensity in the first neighborhood centered on the pixel; inputting the Heinz unit value, Heinz unit average, Heinz unit standard deviation, and grayscale gradient intensity into a probability function to obtain the pixel feature probability value that the pixel belongs to a kidney stone; and combining all pixels and their corresponding pixel feature probability values ​​to form the boundary probability map.

[0042] The grayscale feature refers to the grayscale information of the pixel itself, specifically its Henlein unit value. The Henlein unit value is a standard for measuring tissue density in CT images. For urological stones, the Henlein unit value is usually much higher than that of the surrounding soft tissue, making it an important basis for stone identification. The neighborhood feature refers to the statistical characteristics of pixels within a certain area (i.e., the first neighborhood) centered on the pixel, including the Henlein unit mean, Henlein unit standard deviation, and grayscale gradient intensity. The Henlein unit mean reflects the overall density level of the local area, the Henlein unit standard deviation reflects the uniformity or heterogeneity of the local density, and the grayscale gradient intensity characterizes the drastic change in grayscale in the local area, often used to detect edge information. The probability function can be a pre-trained classification model, such as a support vector machine (SVM), logistic regression model, or neural network model, which outputs a probability value between 0 and 1 based on multiple input feature values, representing the probability that the pixel belongs to a stone. The pixel feature probability value is the output of this probability function, which integrates the pixel's own density information and the texture and edge information of its local environment. The boundary probability map is an image composed of all pixels within the boundary area and their corresponding pixel feature probability values, where the brightness or color of each pixel can represent the probability of it belonging to a stone.

[0043] This application's scheme, by defining grayscale and neighborhood features in detail and using a probability function to comprehensively analyze these features, can more accurately assess the probability that each pixel within the boundary region belongs to a stone. Specifically, the Henlein unit value of a pixel provides direct density evidence, while the mean and standard deviation of the Henlein units in the neighborhood provide contextual information for the local region, helping to distinguish stones from structures with similar Henlein unit values, such as calcified blood vessels or bones. Furthermore, the grayscale gradient intensity can effectively capture the edge information of the stone, further enhancing the accuracy of boundary recognition. Inputting these multi-dimensional features into the probability function for fusion calculation makes the judgment of the probability that a pixel belongs to a stone more comprehensive and robust. Thus, the generated boundary probability map can more accurately reflect the true boundary of the stone, laying the foundation for subsequent accurate measurement of stone size, volume, and Henlein unit value.

[0044] The above technical solution overcomes the limitations of relying solely on a single feature for stone boundary identification, significantly improving the accuracy and reliability of urological stone image pattern recognition. By comprehensively utilizing the Heinz unit value of a pixel, the average Heinz unit value within its neighborhood, the standard deviation of the Heinz unit value, and the gray-level gradient intensity, and inputting these into a probability function for calculation, it is possible to more precisely distinguish between stone and non-stone areas, especially when stone boundaries are blurred or artifacts are present, generating a more accurate boundary probability map. This not only helps improve the accuracy of stone boundary identification but also provides more reliable data support for subsequent quantitative analysis of stone parameters, thereby enhancing the accuracy of diagnosis and treatment planning.

[0045] This application further proposes a pattern recognition method for urological stone images, wherein, based on the boundary probability map, the size, volume, and Henle unit value of the stone are determined through a preset multi-scale feature fusion method, specifically including: A low-resolution image of the stone image is acquired, and the first size and first volume of the stone are estimated based on the low-resolution image. Based on the boundary probability map, the spatial contribution of each pixel in the suspected stone region is weighted according to the pixel feature probability value, and the second size and second volume of the stone are estimated. The pixel feature probability value of the pixel in the core region is 1. The first size and the second size are weighted and fused to obtain the size of the stone. The first volume and the second volume are weighted and fused to obtain the volume of the stone.

[0046] Specifically, when acquiring a low-resolution image of the stone, this can be achieved by performing operations such as downsampling, average pooling, or max pooling on the original high-resolution stone image. The low-resolution image provides information on the overall outline and approximate extent of the stone, which helps in preliminary size and volume estimation at a macroscopic level, i.e., estimating the stone's initial size and volume. This preliminary estimation has good robustness and can effectively avoid interference from local noise or detail artifacts on the overall measurement.

[0047] Specifically, based on the boundary probability map, the second size and second volume of the stone are estimated by weighting the spatial contribution of each pixel within the suspected stone region according to its pixel feature probability value. Here, the pixel feature probability value reflects the confidence level that each pixel belongs to a stone. For pixels in the core region, the probability of them belonging to a stone is extremely high, so their pixel feature probability value is set to 1, indicating that they have a complete spatial contribution to the stone entity. For pixels in the boundary region, their pixel feature probability value is between 0 and 1, indicating the degree to which they belong to a stone. By weighting these probability values, the second size and second volume of the stone can be calculated more precisely, thereby capturing the accurate boundaries and morphology of the stone at the microscopic level.

[0048] In practical applications, the first size and the second size are weighted and fused to obtain the stone's size; the first volume and the second volume are weighted and fused to obtain the stone's volume. This weighted fusion mechanism aims to combine the advantages of estimation results at different scales. The first size and the first volume provide a global, robust estimation, while the second size and the second volume provide a local, fine-grained estimation. By using appropriate weighting coefficients, the two estimation results can be balanced, thereby obtaining more accurate and stable measurements of the stone's size and volume. For example, the weighting coefficients can be adjusted based on image quality, stone size, or preset empirical values ​​to optimize the fusion effect.

[0049] This application's solution effectively addresses the accuracy and robustness limitations that may arise from single-scale measurements by introducing a multi-scale feature fusion mechanism. Specifically, the estimation of low-resolution images provides macroscopic information about the stones, effectively filtering out local noise and detail interference, and providing a stable benchmark for the overall size and volume of the stones. Simultaneously, the pixel feature probability-weighted estimation based on boundary probability maps fully utilizes the confidence information of each pixel in the high-resolution image belonging to a stone. Especially in the stone boundary region, probability weighting allows for a more refined depiction of the true shape and boundaries of the stones, resulting in more accurate local measurement results. It is precisely by combining macroscopic robust estimation with microscopic refined measurement, and organically integrating the two through weighted fusion, that the final stone size and volume measurement results possess both global stability and local accuracy.

[0050] Through the above-described technical solution, this application can significantly improve the accuracy and robustness of urological stone size and volume measurement. The multi-scale feature fusion method effectively overcomes the limitations of single-scale measurement in the face of complex imaging conditions or irregular stone morphologies, making the measurement results closer to the true physical parameters of the stone. This improvement is of great significance for clinical diagnosis, treatment planning, and efficacy evaluation, helping physicians to more accurately assess stone burden and thus optimize patient management.

[0051] In some preferred embodiments, a specific example is illustrated below. Suppose a CT image of a urological stone is acquired. First, the CT image is downsampled to generate a lower-resolution image. Based on this low-resolution image, the approximate outline of the stone can be quickly estimated, and its first size (e.g., maximum diameter) and first volume can be calculated. For example, a preliminary stone region can be obtained through simple thresholding or connected component analysis, and the first size and first volume can be calculated accordingly.

[0052] Meanwhile, according to the above method, a boundary probability map of the CT image has been generated, where pixels in the core region have a pixel feature probability value of 1, and pixels in the boundary region have pixel feature probability values ​​between 0 and 1. Next, for each pixel within the suspected stone region, its spatial contribution is weighted according to its pixel feature probability value. For example, if a pixel's pixel feature probability value is 0.8, its contribution to the stone volume is considered to be 0.8 units of pixel volume. By summing all weighted pixel contributions, the second size and second volume of the stone can be estimated.

[0053] Finally, the first size estimated from the low-resolution image and the second size estimated using the boundary probability map are weighted and fused. For example, a fusion coefficient λ can be used, and the final size = λ. First dimension + (1-λ) The second dimension. Similarly, the first and second volumes are weighted and fused to obtain the final stone volume. In this way, even when there are artifacts in the image or the stone boundaries are unclear, more reliable and accurate stone size and volume measurements can be obtained.

[0054] This application further proposes a more accurate and robust method for determining the Henle unit value. Based on the boundary probability map, the size, volume, and Henle unit value of the stone are determined through a preset multi-scale feature fusion method. The method also includes: locally adaptively adjusting the window width and window level of the suspected stone region; selecting boundary pixels with pixel feature probability values ​​higher than a preset threshold within the boundary region based on the boundary probability map; obtaining core pixels within the core region; and calculating the average Henle unit value of the boundary pixels and core pixels as the Henle unit value of the stone.

[0055] Specifically, locally adaptive window width and level adjustment for suspected stone areas refers to dynamically adjusting the display window width and level of the image based on the local grayscale distribution characteristics of the suspected stone area. The aim is to optimize the contrast of the suspected stone area, making the boundary between the stone and surrounding tissue clearer, facilitating subsequent pixel selection and analysis, while reducing detail loss or information blurring caused by improper globally fixed window width and level settings.

[0056] The selection of boundary pixels with feature probability values ​​higher than a preset threshold based on the boundary probability map can be understood as using the probability information provided by the boundary probability map that each pixel belongs to a stone to filter out those boundary pixels that are highly likely to belong to a stone. The preset threshold is used to distinguish between high-confidence stone boundary pixels and low-confidence non-stone pixels. Simultaneously, obtaining core pixels within the core area means including pixels in the suspected stone area that are identified as the core part of the stone. Pixels within the core area usually have a high certainty of being a stone, and their pixel feature probability value is set to 1 in the above embodiment, representing a high degree of stone ownership. The purpose is to ensure that the set of pixels used to calculate the Heinz unit value has high confidence, thereby improving the accuracy of the Heinz unit value measurement.

[0057] In practical applications, calculating the average Henlein unit value of boundary and core pixels as the Henlein unit value of the stone involves averaging the Henlein unit values ​​of the selected high-confidence boundary pixels and pixels within the core region. For example, an arithmetic mean, weighted average, or other statistical methods can be used. The aim is to obtain a Henlein unit value that better represents the overall density of the stone by integrating the Henlein unit information from these high-confidence pixels, effectively avoiding interference from blurred boundary areas or low-confidence pixels in the Henlein unit value calculation.

[0058] This application's solution optimizes the display of kidney stone images by locally adaptively adjusting the window width and level in the suspected stone area, making the stone's boundary features more prominent. Based on this, it precisely selects high-confidence boundary pixels using quantified information from the boundary probability map, combining them with high-certainty core pixels within the core area. This selective pixel set eliminates low-confidence pixels that may be affected by partial volumetric effects or artifacts, allowing the subsequently calculated Henle unit average to more accurately reflect the true density of the stone. This method avoids the errors that may arise from simply averaging the entire suspected stone area, and its advantages are particularly significant when stone boundaries are unclear or image quality is poor.

[0059] Through the above technical solutions, this application can effectively improve the accuracy and robustness of Heinz unit value measurement for urological stones. Locally adaptive window width and level adjustment helps to better identify stone details against complex imaging backgrounds; while the pixel selection mechanism based on boundary probability maps and core regions ensures that the pixel set used for Heinz unit value calculation has high confidence, thereby significantly reducing measurement errors caused by boundary blurring, partial volume effects, or image artifacts. Therefore, the obtained Heinz unit values ​​for stones can more accurately reflect the chemical composition and hardness of the stones, providing clinicians with more reliable diagnostic evidence and support for treatment selection.

[0060] In some preferred embodiments, a specific example is given below. Suppose an image of a urological stone is acquired, and after initial identification, a suspected stone region is determined. To accurately measure the Heinz unit value of this stone, the suspected stone region is first locally adaptively adjusted in terms of window width and level. For example, based on the gray-level histogram distribution of pixels within the region, an optimal window width and level are dynamically set to enhance the contrast between the stone and the surrounding soft tissue. Subsequently, the system identifies all pixels within the boundary region with a feature probability value higher than 0.7 based on the previously generated boundary probability map; these are considered high-confidence boundary pixels. Simultaneously, all pixels within the already identified core region are also considered. Finally, the Heinz unit values ​​of these high-confidence boundary pixels and core pixels are arithmetically averaged to obtain an accurate Heinz unit value for the stone, for example, 1200 HU. This method ensures that the Heinz unit value is calculated based on the pixels that best represent the stone itself, thereby improving the reliability of the measurement.

[0061] This application further proposes that, after identifying the stone image to determine the suspected stone area and its boundary area, the method further includes: identifying the anatomical structure of the stone image to determine at least one target anatomical structure in the image; the target anatomical structure includes at least one of the kidney, ureter, bladder, and vascular structures.

[0062] Specifically, the anatomical structure recognition refers to the automatic or semi-automatic identification and segmentation of key anatomical structures related to the urinary system from urological stone images using image processing and analysis techniques. This can be achieved using various existing technologies, such as deep learning-based image segmentation models, which, through training on large amounts of labeled data, can accurately identify and delineate structures such as the kidneys, ureters, bladder, and blood vessels. Alternatively, a medical image registration method can be used to register standard anatomical atlases with the patient's stone images, thereby determining the location and extent of each anatomical structure in the image. The target anatomical structure is defined as at least one of the kidneys, ureters, bladder, and blood vessels. These structures are the most common or closely related anatomical sites for urinary system stones. For example, the kidneys are the main organs for stone formation, the ureters and bladder are the channels for stone expulsion, and blood vessels (especially the renal vessels and abdominal aorta) may resemble stone image features due to their high-density calcification; therefore, identifying these structures helps distinguish between true and false stones. By identifying these specific target anatomical structures, accurate anatomical context can be provided for subsequent stone identification.

[0063] This application's solution effectively addresses the potential misidentification problem that can occur when relying solely on image grayscale and neighborhood features for stone identification by introducing anatomical structure recognition. Specifically, once a suspected stone region is initially identified, identifying target anatomical structures such as the kidneys, ureters, bladder, and blood vessels in the image provides crucial spatial localization and physiological background information for the suspected stone region. For example, if a high-density area is initially identified as a suspected stone, but anatomical structure recognition reveals that the area is clearly located within the renal pelvis or calyces of the kidney, the probability of it being a stone is increased. Conversely, if the high-density area is identified as being located on a blood vessel wall, the probability of it being a stone is significantly reduced, thus avoiding misidentification. This judgment mechanism, which incorporates anatomical context, enables the system to more intelligently distinguish between genuine stones and non-stone structures with similar image features, thereby improving the specificity and accuracy of stone identification.

[0064] In some preferred embodiments, a specific example is given below. Suppose a urological CT image of a stone is acquired. First, based on a preset range of Henlein unit values, the system identifies multiple high-density regions in the image as primary stone regions. It further determines the core and boundary regions of these regions, forming a probability map of suspected stone regions and their boundaries. Based on this, to further improve the accuracy of identification, the system performs anatomical structure identification on the CT image. Specifically, a pre-trained deep learning model can be used to perform semantic segmentation of the image, accurately identifying and delineating the patient's kidneys, ureters, bladder, and major vascular structures (such as the renal artery, renal vein, and abdominal aorta). For example, if a suspected stone region is initially identified with a high Henlein unit value, but anatomical structure identification reveals that the region is clearly located in the renal pelvis region within the kidney, the system will combine this anatomical information to increase the confidence that the region is a stone. Conversely, if another high-density region is initially identified with image features similar to a stone, but anatomical structure identification shows that the region is located on the wall of the renal artery with obvious vascular calcification features. In this situation, the system reduces the probability of a stone being located in that area based on its anatomical position within the vascular structure, and may even exclude it from the stone candidate area. In this way, anatomical structure recognition provides crucial contextual information for the final stone identification, making the stone identification results more accurate and effectively avoiding misidentification of non-stone structures such as vascular calcifications as stones, thereby improving the specificity and accuracy of the diagnosis.

[0065] This application further proposes a pattern recognition method for urological stone images, wherein calculating the probability that a pixel belongs to a stone further includes: determining a first spatial probability value of the pixel based on the relative positional relationship between any pixel in the boundary region and the target anatomical structure; the first spatial probability value is preset based on the anatomical structure to which the pixel belongs. Based on the correlation between the pixel and the physiological characteristics of the target anatomical structure, the first spatial probability value is adjusted to obtain a second spatial probability value; based on the second spatial probability value and the pixel feature probability value, the probability that the pixel belongs to a stone is determined.

[0066] Specifically, when calculating the probability that a pixel belongs to a stone, a first spatial probability value is first determined based on the relative position of any pixel within the boundary region to the target anatomical structure. This first spatial probability value is preset based on the anatomical structure to which the pixel belongs. For example, if a pixel is located in the kidney or ureter region, its first spatial probability value might be preset to a relatively high value, as these are common anatomical sites for stones; conversely, if the pixel is located in other non-common stone-producing areas, its first spatial probability value might be preset to a lower value. Further, the first spatial probability value is adjusted based on the physiological characteristic correlation between the pixel and the target anatomical structure to obtain a second spatial probability value. This physiological characteristic correlation refers to whether the pixel's position within the anatomical structure conforms to the physiological distribution patterns of stones; for example, stones typically do not appear inside blood vessels. Finally, the second spatial probability value is combined with the pixel feature probability value to determine the final probability that the pixel belongs to a stone.

[0067] This application's solution introduces spatial probability values ​​and combines the relative positional relationship between pixels and target anatomical structures with physiological feature correlations to conduct a multi-dimensional assessment of the probability that a pixel belongs to a stone. Specifically, determining the first spatial probability value allows the system to utilize macroscopic anatomical information to assign different initial stone probabilities to pixels in different regions, thus introducing spatial context information in the initial stage. Subsequently, by analyzing the physiological feature correlation between pixels and target anatomical structures, such as determining whether a pixel is located inside a blood vessel or adjacent to the vessel wall, the first spatial probability value is further refined. This adjustment mechanism can effectively exclude areas that appear high-density on imaging but are physiologically unlikely to be stones (such as vascular calcification), thereby avoiding misidentification. Finally, the second spatial probability value, corrected by spatial information, is fused with the pixel feature probability value obtained based on pixel grayscale and neighborhood features. This ensures that stone identification not only relies on local image features but also fully considers its position and physiological rationality within the entire anatomical structure, thereby improving the accuracy and robustness of identification.

[0068] Through the above technical solution, this application can effectively solve the problem of insufficient accuracy in stone identification by traditional methods in complex anatomical structures or in the presence of artifacts. By introducing spatial probability values ​​and adjusting them in conjunction with physiological feature correlations, it can more accurately distinguish between stone and non-stone structures, especially showing a significant advantage in distinguishing interference features similar to stone imaging characteristics, such as vascular calcification, thereby reducing the misdiagnosis rate and missed diagnosis rate and improving the reliability of urological stone imaging pattern recognition.

[0069] In some preferred embodiments, it is assumed that pattern recognition is performed on a urological stone image. First, the system identifies target anatomical structures such as the kidneys, ureters, and blood vessels in the image. When a pixel is detected within the boundary region of the kidney area, its first spatial probability value is preset to 0.8 because the pixel is located inside the kidney. Subsequently, the system further analyzes the physiological characteristic correlation between the pixel and the blood vessel structure. If it is found that the Euclidean distance from the pixel to the nearest renal artery wall is less than a preset distance threshold, it indicates that it may be related to vascular calcification. Based on the physiological characteristic correlation, the first spatial probability value is adjusted, for example, by multiplying it by a decay factor of 0.5, thereby obtaining a second spatial probability value of 0.4. At the same time, the pixel feature probability value calculated based on its Henlein unit value, the average Henlein unit value of its neighborhood, the standard deviation of the Henlein unit value, and the gray-level gradient intensity is 0.9. Finally, the second spatial probability value of 0.4 and the pixel feature probability value of 0.9 are fused (e.g., by weighted averaging) to obtain the final probability that the pixel belongs to a stone. In this way, even if the pixel has a high Henlein value, its probability of being identified as a stone will be reduced due to its close association with blood vessels, thus effectively avoiding misidentification of vascular calcification as a stone.

[0070] This application further proposes a more refined method for adjusting the first spatial probability value by considering the physiological characteristics of the correlation between pixels and vascular structures to optimize the calculation of the spatial probability value.

[0071] Adjusting the first spatial probability value based on the physiological characteristic correlation between the pixel and the target anatomical structure includes: in response to the Euclidean distance from the pixel to the nearest blood vessel wall being less than a preset distance threshold, determining that the pixel has a physiological characteristic correlation with the blood vessel structure, and reducing the first spatial probability value; wherein, reducing the first spatial probability value includes: in response to the Euclidean distance from the pixel to the nearest blood vessel wall being less than a preset distance threshold, multiplying the first spatial probability value by a preset attenuation factor β to obtain a second spatial probability value; wherein 0 < β < 1.

[0072] Specifically, the Euclidean distance refers to the straight-line distance between two points in multidimensional space, and here it specifically refers to the spatial distance between a pixel and its nearest blood vessel wall. The preset distance threshold is a configurable parameter set to define whether there is a sufficiently close physiological correlation between the pixel and the blood vessel structure. For example, it can be obtained based on clinical experience or through training a machine learning model to ensure effective differentiation between stones and blood vessels. The blood vessel structure can be understood as tubular structures such as arteries and veins identified in the image, and its identification can be achieved through specialized blood vessel segmentation algorithms. When the Euclidean distance from the pixel to the nearest blood vessel wall is less than the preset distance threshold, it indicates that the pixel is highly likely to have spatial overlap or close proximity with the blood vessel structure. In this case, to avoid misclassifying the blood vessel structure as a stone, the probability that the pixel belongs to a stone needs to be adjusted.

[0073] One specific way to reduce the first spatial probability value is to multiply it by a preset attenuation factor β. The preset attenuation factor β is a value between 0 and 1 (0 < β < 1), and its function is to reduce the original first spatial probability value. For example, when β is set to 0.5, it means that the original spatial probability value is halved. The specific value of this attenuation factor β can be adjusted according to the actual application scenario, image quality, and tolerance for misjudgment to achieve the best recognition effect. Through this multiplicative attenuation method, the risk of pixels physiologically related to vascular structures being misjudged as stones can be effectively reduced, thereby improving the specificity of stone recognition.

[0074] This application's solution addresses the issue of vascular structures potentially interfering with accurate stone identification in complex anatomical contexts by introducing a judgment on the correlation between pixels and the physiological characteristics of vascular structures. Specifically, when a pixel is initially identified as potentially belonging to a certain anatomical structure (such as the kidney or ureter) and assigned a first spatial probability value, the solution further examines whether the pixel has a close physiological correlation with the vascular structure. By calculating the Euclidean distance from the pixel to the nearest vascular wall and comparing it with a preset distance threshold, pixels that may belong to blood vessels but are adjacent to stone regions can be accurately identified. Once such a correlation is confirmed—that is, the Euclidean distance is less than the preset distance threshold—the probability that the pixel belongs to a stone is effectively reduced by multiplying the first spatial probability value by a preset attenuation factor β. This mechanism avoids misidentifying non-stone structures such as vascular walls or intravascular calcifications as stones, thereby improving the accuracy of spatial probability value calculation.

[0075] As a specific implementation, suppose that in the acquired urological stone image, a pixel is initially identified as being located within the kidney region and assigned a high first spatial probability value, such as 0.8, based on its associated anatomical structure. The system then further detects the relationship between this pixel and the identified vascular structures in the image. If the calculated Euclidean distance from this pixel to the nearest renal artery wall is 1 mm, and the preset distance threshold is set to 2 mm, the system determines that this pixel has a physiological characteristic association with the vascular structure. In this case, the preset attenuation factor β can be set to 0.4. Then, the first spatial probability value of this pixel, 0.8, will be multiplied by 0.4, resulting in an adjusted second spatial probability value of 0.32. In this way, the probability that this pixel belongs to a stone is significantly reduced, effectively avoiding misidentification of the renal artery wall as a stone and improving the overall accuracy of identification.

[0076] In some embodiments described above, while it is proposed to adjust spatial probability values ​​based on the physiological characteristics of the correlation between pixels and target anatomical structures (e.g., vascular structures), and combine this with pixel feature probability values ​​to determine the probability that a pixel belongs to a stone, in practical applications, the quality of urological stone images can be affected by various factors, such as the presence of artifacts. This may cause a simple probability combination method to fail to accurately reflect the true boundary of the stone, thereby affecting the accuracy of stone size, volume, and Henle unit values. If the above problems are not addressed, the identification and quantification results of stones may be biased when the image quality is poor, thus affecting the formulation of clinical diagnosis and treatment plans.

[0077] In response, this application further proposes a method for determining the probability that a pixel belongs to a gallstone based on the aforementioned second spatial probability value and the aforementioned pixel feature probability value, specifically including: the probability of the gallstone is: P_final = (1 - α) P_image + α W_space; where P_image is the pixel feature probability value; W_space is the second spatial probability value; α is the fusion coefficient, and 0 ≤ α ≤ 1; the fusion coefficient α is adaptively set according to the quality of the stone image: in response to the proportion of artifact region in the stone image being higher than a preset region threshold, the fusion coefficient α is set to a value greater than 0.5.

[0078] Specifically, P_image can be understood as the probability that a pixel belongs to a stone, calculated based on its own grayscale features and neighborhood features. It reflects the likelihood of a stone in the local image features of the pixel. W_space can be understood as the spatial probability value obtained after adjusting for the relative positional relationship between the pixel and the target anatomical structure and its physiological feature correlation. It reflects the likelihood of a stone in the anatomical context of the pixel. The fusion coefficient α is used to balance the weights of P_image and W_space in the calculation of the final probability P_final. When α is set to 0, the final probability is completely determined by the pixel feature probability value; when α is set to 1, the final probability is completely determined by the second spatial probability value.

[0079] In practical applications, the adaptive setting of the fusion coefficient α is based on the quality of the stone image. Specifically, when a high proportion of artifact regions is detected in the stone image, for example, if the proportion of artifact regions exceeds a preset region threshold, it is considered that the local pixel features of the image may be disturbed. In this case, the reliability of the pixel feature probability value P_image will decrease. To compensate for this uncertainty, the fusion coefficient α is set to a value greater than 0.5, which means that the second spatial probability value W_space will receive greater weight when calculating the final stone probability P_final. The preset region threshold can be determined based on clinical experience or through training a machine learning model, for example, it can be set to 5%, 10%, or higher.

[0080] This application's solution effectively addresses the problem of inaccurate stone identification when relying solely on local pixel features is problematic in low-quality images by introducing a fusion coefficient α and adaptively setting it according to the image quality. This is because the reliability of the pixel feature probability value P_image decreases when the image artifact region is high, while the second spatial probability value W_space, based on anatomical structure and physiological correlation, is relatively less affected by local artifacts. Therefore, by setting the fusion coefficient α to a value greater than 0.5, W_space dominates the calculation of the final probability P_final, thus enabling a more robust assessment of the probability that a pixel belongs to a stone. This adaptive weight adjustment mechanism allows the system to flexibly adjust its dependence on local features and global anatomical context when facing images of different qualities, thereby improving the accuracy and robustness of stone identification.

[0081] Through the above technical solution, this application can adaptively adjust the fusion weights of pixel feature probability values ​​and spatial probability values ​​based on the actual quality of urological stone images, especially the presence of artifacts. This adaptive adjustment mechanism significantly enhances the robustness of stone identification, especially when image quality is impaired (e.g., artifacts are present), effectively reducing the interference of artifacts on stone boundary judgment, thereby obtaining more accurate stone size, volume, and Henle unit values. Compared to fixed-weight fusion methods, the solution of this application can better adapt to the diversity of clinical images, improve the reliability and accuracy of stone quantification results, and provide a more reliable basis for clinical diagnosis and treatment.

[0082] In some preferred embodiments, a specific example is given below. Suppose that during a CT scan of a patient with urological stones, significant motion artifacts or bundle hardening artifacts are present in the acquired stone image due to the patient's breathing movements or metal implants. First, the system evaluates the proportion of artifact regions in the stone image. For example, if the image processing algorithm detects that the artifact region proportion reaches 15%, which is higher than a preset region threshold (e.g., 10%), the fusion coefficient α will be adaptively set to a value greater than 0.5, such as 0.7. This means that when calculating the final probability P_final that a pixel belongs to a stone, the second spatial probability value W_space (reflecting the probability of anatomical context) will account for 70% of the weight, while the pixel feature probability value P_image (the probability of local features that may be affected by artifacts) will account for 30% of the weight. In this way, even if pixel features are blurred or distorted in local areas of the image due to artifacts, the system can rely more on the relative positional relationship and physiological characteristics of pixels with target anatomical structures such as the kidneys and ureters to determine the likelihood of them being stones. This avoids misjudgments or omissions caused by artifacts, ensuring accurate delineation of stone boundaries and the reliability of subsequent quantification results. Conversely, if the image quality is good and the proportion of artifact areas is below a preset threshold, the fusion coefficient α can be set to a value less than or equal to 0.5, such as 0.3. In this case, the pixel feature probability value P_image will receive greater weight to fully utilize the local detail information of high-quality images.

[0083] Secondly, this embodiment proposes a pattern recognition system for urological stone imaging, such as... Figure 2 As shown, it includes: The acquisition and identification module 201 is used to acquire images of urological stones and identify the stone images to determine the suspected stone area and its boundary area. The probability generation module 202 is used to calculate the probability that each pixel in the boundary region belongs to a stone based on its grayscale features and neighborhood features, and generate a boundary probability map. The determination module 203 is used to determine the size, volume and Heinz unit value of the stone based on the boundary probability map by a preset multi-scale feature fusion method.

[0084] This application introduces a modular design, decomposing the pattern recognition process of kidney stone images into functionally independent components. This aims to overcome the limitations of traditional methods and existing intelligent recognition systems in identifying kidney stones in complex imaging environments, improving the accuracy and robustness of the identification, especially when processing low-dose CT, calcifications near anatomical structures, and enhanced CT images, providing more reliable auxiliary diagnostic information. The module described in the second aspect is used to perform the method described in the first aspect, and will not be elaborated further here.

[0085] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A pattern recognition method for urological stone images, characterized in that, include: Acquire images of urological stones, and identify the stone images to determine the suspected stone area and its boundary area; For each pixel within the boundary region, the probability that the pixel belongs to a stone is calculated based on its grayscale features and neighborhood features, and a boundary probability map is generated. Based on the boundary probability map, the size, volume, and Heinz unit value of the stone are determined by a preset multi-scale feature fusion method.

2. The pattern recognition method for urological stone images according to claim 1, characterized in that, The suspected stone area includes a core area and a boundary area; Identifying the stone images to determine suspected stone areas and their boundary areas includes: Based on a predetermined range of Henlein unit values, at least one first stone region is identified from the stone image; For any of the first stone regions, a first pixel layer is extended outward to form an extension region, and a second pixel layer is contracted inward to form a core region; the region between the extension region and the core region is determined as the boundary region of the suspected stone region.

3. The pattern recognition method for urological stone images according to claim 2, characterized in that, The grayscale features include the Heinz unit value of the pixel; the neighborhood features include the average Heinz unit value of the pixels in the neighborhood centered on the pixel, the standard deviation of the Heinz unit value of the pixels in the neighborhood centered on the pixel, and the grayscale gradient intensity in the neighborhood centered on the pixel. The process of calculating the probability that a pixel belongs to a kidney stone based on its grayscale features and neighborhood features, and generating a boundary probability map, includes: For each pixel within the boundary region, obtain the Heinz unit value of that pixel, as well as the average Heinz unit value, standard deviation of Heinz unit, and grayscale gradient intensity of the first neighborhood centered on that pixel. The Heinz unit value, Heinz unit average value, Heinz unit standard deviation, and gray-level gradient intensity are input into a probability function to obtain the pixel feature probability value that the pixel belongs to the stone. The boundary probability map is formed by combining all pixels and their corresponding pixel feature probability values.

4. The pattern recognition method for urological stone images according to claim 3, characterized in that, Based on the boundary probability map, the size, volume, and Henle unit value of the stone are determined through a preset multi-scale feature fusion method, including: A low-resolution image of the stone is acquired, and the first size and first volume of the stone are estimated based on the low-resolution image. Based on the boundary probability map, the second size and second volume of the stone are estimated by weighting the spatial contribution of each pixel in the suspected stone region according to the pixel feature probability value; the pixel feature probability value of the pixel in the core region is 1. The first size and the second size are weighted and fused to obtain the size of the stone; the first volume and the second volume are weighted and fused to obtain the volume of the stone.

5. The pattern recognition method for urological stone images according to claim 4, characterized in that, Based on the boundary probability map, the size, volume, and Heinz unit value of the stone are determined through a preset multi-scale feature fusion method, and the method further includes: The window width and window level are locally adaptively adjusted for the suspected stone area; Based on the boundary probability map, select boundary pixels whose pixel feature probability values ​​within the boundary region are higher than a preset threshold; and obtain the core pixels within the core region. The average value of the boundary pixels and the core pixels in Henle units is calculated as the Henle unit value of the stone.

6. The pattern recognition method for urological stone images according to claim 3, characterized in that, The method further includes identifying the stone image to determine the suspected stone area and its boundary area, and then: Anatomical structure identification is performed on the stone images to determine at least one target anatomical structure in the images; the target anatomical structure includes at least one of the kidney, ureter, bladder, and vascular structures.

7. The pattern recognition method for urological stone images according to claim 6, characterized in that, The calculation of the probability that the pixel belongs to a stone also includes: Based on the relative positional relationship between any pixel within the boundary region and the target anatomical structure, a first spatial probability value for that pixel is determined; the first spatial probability value is preset based on the anatomical structure to which the pixel belongs. Based on the correlation between the pixel and the physiological characteristics of the target anatomical structure, the first spatial probability value is adjusted to obtain the second spatial probability value; Based on the second spatial probability value and the pixel feature probability value, the probability that the pixel belongs to a stone is determined.

8. The pattern recognition method for urological stone images according to claim 7, characterized in that, Based on the correlation between the pixel and the physiological characteristics of the target anatomical structure, the first spatial probability value is adjusted, including: If the Euclidean distance from the pixel to the nearest blood vessel wall is less than a preset distance threshold, it is determined that the pixel is associated with the blood vessel structure by physiological characteristics, and the first spatial probability value is reduced. Reducing the first spatial probability value includes: In response to the Euclidean distance from the pixel to the nearest blood vessel wall being less than a preset distance threshold, the first spatial probability value is multiplied by a preset attenuation factor β to obtain a second spatial probability value; where 0 < β < 1.

9. The pattern recognition method for urological stone images according to claim 8, characterized in that, Based on the second spatial probability value and the pixel feature probability value, the probability that the pixel belongs to a stone is determined, including: The probability of obtaining the stone is: P_final = (1 - α) P_image + α W_space; Where P_image is the pixel feature probability value; W_space is the second space probability value; α is the fusion coefficient, and 0 ≤ α ≤ 1; The fusion coefficient α is adaptively set according to the quality of the stone image: in response to the proportion of artifact regions in the stone image being higher than a preset region threshold, the fusion coefficient α is set to a value greater than 0.

5.

10. A pattern recognition system for urological stone imaging, characterized in that, include: The acquisition and identification module is used to acquire images of urological stones and identify the stone images to determine the suspected stone area and its boundary area; The probability generation module is used to calculate the probability that each pixel in the boundary region belongs to a stone based on its grayscale features and neighborhood features, and generate a boundary probability map. The determination module is used to determine the size, volume, and Heinz unit value of the stone based on the boundary probability map and through a preset multi-scale feature fusion method.

Images