A chronic kidney disease information processing system and method based on a morphological characteristic parameter of uric acid crystals

Abstract

The present application provides a chronic kidney disease information processing system and method based on the morphological characteristic parameters of uric acid crystals, which comprises: a crystal image acquisition module, which acquires multiple uric acid crystal images formed by inducing crystallization of ex vivo serum samples; an image feature extraction module, which extracts the morphology and / or depth features of each image using image processing and / or deep learning, and outputs a feature vector; a multi-instance learning aggregation module, which takes all feature vectors of the same subject as a sample package, and obtains a subject-level crystal morphology representation by using a gated attention weighted aggregation; and a phenotype score generation module, which generates a continuous kidney state-related phenotype score based on the subject-level features. The system converts serum microenvironment information into crystal morphology quantitative indicators, enriches the phenotype information dimension, improves the consistency, repeatability and phenotype resolution capability of the phenotype information processing results, and provides objective and repeatable phenotype reference information for chronic kidney disease research.

Description

Technical Field

[0001] This invention relates to the field of medical image processing, and in particular to an information processing system and method for chronic kidney disease based on uric acid crystal morphology characteristic parameters. Background Technology

[0002] Current assessments of chronic kidney disease-related status primarily rely on detecting the levels of several specific molecular markers (such as serum creatinine, blood urea nitrogen, uric acid, and urinary protein) and staging the disease based on reference ranges. While this method can reflect changes in some metabolite levels, the information obtained is limited to a finite number of markers, making it difficult to fully reflect the complex internal environment state composed of multiple components such as proteins, metabolites, electrolytes, and inflammatory factors in serum. This quantitative detection, limited to a single or a few molecular indicators, easily leads to biased and discrete assessment results, failing to capture the overall characteristics of individual differences in the internal environment.

[0003] In recent years, although some studies have introduced artificial intelligence algorithms such as BP neural networks, XGBoost, and random forests, their inputs still mainly consist of existing clinical data such as electronic medical records, symptoms, lifestyle habits, medical history, blood tests, and urine tests. Essentially, these are risk prediction models based on structured data, still limited by the information dimensions of traditional clinical data. They cannot extract novel and complementary phenotypic signals from the overall serum microenvironment, resulting in insufficient accuracy and completeness of assessment information. Existing AI models primarily rely on existing structured clinical data and lack crystalline morphology phenotypic signals derived from the isolated serum microenvironment. Summary of the Invention

[0004] This invention provides a chronic kidney disease information processing system and method based on uric acid crystal morphology characteristic parameters to solve the problems existing in related technologies. The technical solution is as follows: In a first aspect, embodiments of the present invention provide a chronic kidney disease information processing system based on uric acid crystal morphology characteristic parameters, comprising: The crystal image acquisition module is used to acquire multiple uric acid crystal images corresponding to the subject's ex vivo serum sample. The multiple uric acid crystal images are acquired under uniform imaging conditions from the same sample of the same subject after in vitro induced crystallization of uric acid crystals. The image feature extraction module is used to extract morphological and / or depth image features from each uric acid crystal image using image processing algorithms and / or deep learning models, and output the corresponding image feature vector. The multi-instance learning aggregation module is used to treat all image feature vectors corresponding to the same subject as a sample bag, and to calculate the attention weight of each image feature vector in the sample bag using a gated attention mechanism. Based on the attention weight, the image feature vectors are weighted and aggregated to obtain the subject-level uric acid crystal morphology features. The phenotypic score generation module is used to generate continuous kidney state-related phenotypic scores based on subject-level uric acid crystal morphology features.

[0005] In one implementation, the kidney state-related phenotypic score includes an abnormal phenotypic index; the system further includes: The reference sample storage module stores a normal sample reference set and an abnormal sample reference set. The phenotypic scoring generation module generates an abnormal phenotypic index based on the similarity or distance between the subject-level uric acid crystal morphology features and each feature vector in the normal sample reference set and / or abnormal sample reference set.

[0006] In one implementation, the kidney status-related phenotypic score includes a progression trend score; the system further includes: The stage reference dataset storage module is used to store stage-related reference datasets divided according to preset stage rules. The stage-related reference datasets include at least the crystal image feature set of the first stage samples and the crystal image feature set of the second stage samples. The phenotypic scoring generation module calculates a first metric between the subject-level uric acid crystal morphology features and the crystal image feature set of the first-stage sample, and a second metric between the subject-level uric acid crystal morphology features and the crystal image feature set of the second-stage sample, and generates a progression trend score based on the comparison results of the first and second metric.

[0007] In one implementation, it further includes: The clinical indicator acquisition and coding module is used to acquire the clinical feature vector of the subject, and to preprocess and encode the clinical feature vector to obtain clinical coded features; wherein, the clinical feature vector includes one or more of serum creatinine, estimated glomerular filtration rate, urine protein, hemoglobin, blood urea nitrogen, and uric acid; The multimodal fusion module is used to fuse subject-level uric acid crystal morphology features with clinical coding features to generate multimodal fusion features; The phenotypic scoring generation module maps multimodal fusion features to fusion morphological scores and outputs them.

[0008] In one implementation, the preprocessing performed on the clinical feature vector by the clinical indicator acquisition and encoding module includes one or more of the following: missing value imputation, normalization, standardization, outlier handling, or logarithmic transformation.

[0009] In one implementation, the multimodal fusion module employs a gated fusion mechanism, including: An image projection layer is used to map subject-level uric acid crystal morphology features into image hidden features; The clinical projection layer is used to map clinical coded features to clinical hidden features; A gating network is used to receive a concatenated vector of image hidden features and clinical hidden features, and output a gating factor. The weighted fusion unit is used to perform weighted summation of image hidden features and clinical hidden features according to the gating factor to obtain multimodal fusion features.

[0010] In one embodiment, the crystal image acquisition module further includes: The crystal segmentation unit is used to segment the crystal foreground region of the uric acid crystal image to obtain one or more of the following: crystal foreground mask, crystal outline, crystal foreground image, crystal region image, or standardized single crystal cropped image. The segmented crystal region image, crystal foreground image, or standardized single crystal cropped image is used as input to the image feature extraction module.

[0011] In one embodiment, the morphological features include explicit morphological parameters, which include one or more of the following: crystal area, perimeter, aspect ratio, roundness, convex hull area, concavity ratio, gray mean, gray standard deviation, entropy, contrast, and fractal dimension.

[0012] Secondly, embodiments of the present invention provide a method for processing information on chronic kidney disease based on uric acid crystal morphology characteristic parameters, applied to the aforementioned chronic kidney disease information processing system based on uric acid crystal morphology characteristic parameters; the method includes: Multiple uric acid crystal images corresponding to the subject's ex vivo serum sample were acquired. These multiple uric acid crystal images were obtained under uniform imaging conditions from the same sample of the same subject after in vitro induced crystallization of uric acid crystals. Image processing algorithms and / or deep learning models are used to extract morphological and / or depth image features from each uric acid crystal image, and the corresponding image feature vector is output. All image feature vectors corresponding to the same subject are taken as a sample bag. The attention weight of each image feature vector in the sample bag is calculated by a gated attention mechanism. Based on the attention weight, each image feature vector is weighted and aggregated to obtain the subject-level uric acid crystal morphology features. A continuous kidney state-related phenotypic score is generated based on subject-level uric acid crystal morphology features.

[0013] In one embodiment, the method for obtaining uric acid crystals includes: Peripheral venous blood was collected from the subjects using a serum separation tube, centrifuged at 3500 rpm for 5 minutes, and the supernatant serum was separated to obtain a serum sample, which was then transferred to an EP tube. Add 0.1 g of uric acid to 3 mL of 0.667 mol / L NaOH solution and mix thoroughly until completely dissolved to form a supersaturated uric acid solution. Add 60 μL of acetate-sodium acetate buffer and 8 μL of serum sample to the reaction plate in sequence, then add 10 μL of supersaturated uric acid solution and mix quickly. The reaction plate was placed in a 37°C incubator and incubated for 1 hour to induce the formation of uric acid crystals.

[0014] The advantages or beneficial effects of the above technical solutions include at least the following: This invention induces the formation of uric acid crystals in serum samples in vitro, transforming the synergistic effects of multiple components in serum, such as proteins, metabolites, and inflammatory factors, into visualized crystal microscopic images. This allows for the extraction of crystal morphology features, overcoming the limitations of traditional methods that rely solely on a few molecular markers or existing clinical data. This invention can more comprehensively capture the differences in the overall serum microenvironment.

[0015] Furthermore, this invention employs a multi-instance learning model with a gated attention mechanism to weighted aggregate multiple crystal images or single crystal instances from the same subject, obtaining stable and reliable subject-level crystal morphology features and generating a continuous phenotypic score related to kidney status. This score originates from the crystal morphology phenotype induced by ex vivo serum samples, providing quantitative information from the overall serum microenvironment that differs from traditional clinical indicators such as serum creatinine, eGFR, and urinary protein. It possesses potential complementarity in information dimensions and can serve as phenotypic reference information beyond existing clinical indicators, thereby enriching the phenotypic descriptive dimensions of kidney-related states. It also helps improve the consistency and reproducibility of information processing results through standardized image analysis and model aggregation processes, reducing subjective human intervention in the image analysis stage.

[0016] The above overview is for illustrative purposes only and is not intended to be limiting in any way. In addition to the illustrative aspects, embodiments, and features described above, further aspects, embodiments, and features of the invention will become readily apparent from the accompanying drawings and the following detailed description. Attached Figure Description

[0017] In the accompanying drawings, unless otherwise specified, the same reference numerals throughout the various drawings denote the same or similar parts or elements. These drawings are not necessarily drawn to scale. It should be understood that these drawings depict only some embodiments disclosed in the invention and should not be construed as limiting the scope of the invention.

[0018] Figure 1 This is a schematic diagram of the modules of the chronic kidney disease information processing system based on uric acid crystal morphology characteristic parameters of the present invention; Figure 2 ROC curve for verifying the abnormal phenotypic index of this invention; Figure 3 This is a confusion matrix diagram for verifying the anomalous phenotypic index of the present invention; Figure 4 ROC curve for verifying the progress trend score of this invention; Figure 5 Confusion matrix diagram for verifying the progress trend score of this invention; Figure 6 This is the ROC curve for the morphology scoring verification of this invention; Figure 7 This is the confusion matrix diagram of the morphology scoring verification of the present invention; Figure 8 This is a schematic flowchart of an information processing method according to an embodiment of the present invention. Detailed Implementation

[0019] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0020] Example 1 This embodiment provides a chronic kidney disease information processing system based on uric acid crystal morphology characteristics, referencing... Figure 1 As shown, the system includes a crystal image acquisition module, an image feature extraction module, a multi-instance learning aggregation module, and a phenotypic score generation module.

[0021] In this embodiment, the crystal image acquisition module is used to acquire multiple uric acid crystal images corresponding to the subject's ex vivo serum sample.

[0022] It should be noted that the multiple images of uric acid crystals were acquired under uniform imaging conditions from the same sample of the same subject after in vitro induced crystallization of uric acid crystals. Specifically, the formation process of uric acid crystals is as follows: Step 1, serum acquisition Peripheral venous blood was collected from the subjects using vacuum blood collection tubes containing serum separating gel. The collection tubes were placed in a centrifuge and centrifuged at 3500 rpm for 5 minutes. After centrifugation, the blood separated into three layers: the bottom layer consisted of red blood cells, the middle layer consisted of separating gel and white blood cells / platelets, and the top layer consisted of pale yellow serum. The upper serum layer was carefully aspirated with a pipette and transferred to a sterile EP tube. After labeling, the tube was stored at -80°C for later use.

[0023] It should be noted that, without changing the core idea of ​​"inducing crystal carrier formation from ex vivo body fluid samples and assessing chronic kidney disease-related phenotypes based on crystal morphology characteristics," the test samples are not limited to serum. That is, in addition to serum, plasma, processed whole blood supernatant, urine, dialysate, or other body fluid samples that can reflect differences in chronic kidney disease-related states can also be used. After appropriate pretreatment, these samples can all be used to induce crystal formation and further obtain crystal images, morphological parameters, and model analysis results. These implementation methods based on other body fluid samples can also serve as alternative implementation methods of this invention.

[0024] The second step is the preparation of the uric acid solution. Weigh 0.1 g of uric acid powder and add it to 3 mL of 0.667 mol / L NaOH solution. Vortex or shake thoroughly to mix until the uric acid is completely dissolved, forming a clear supersaturated uric acid solution.

[0025] It should be noted that the uric acid quality, NaOH concentration, and volume can be optimized and adjusted according to the actual crystallization effect (e.g., increasing or decreasing the amount of uric acid or changing the concentration of the alkaline solution).

[0026] The third step is the serum-induced uric acid crystal reaction system. Add 60 μL of acetate-sodium acetate buffer (pH pre-adjusted to 4.5), 8 μL of serum sample, and finally 10 μL of supersaturated uric acid solution to each well of the reaction plate. Immediately mix by quickly pipetting or gently shaking the plate. Preferably, a 96-well plate can be used in this embodiment.

[0027] It should be noted that an acetate-sodium acetate buffer solution with a pH of 4.5 provides a stable, weakly acidic environment. The solubility of uric acid decreases sharply under acidic conditions, promoting its crystallization. Serum contains various components such as proteins, lipids, ions, and inflammatory factors, which may participate in crystal nucleation, growth regulation, or surface modification, resulting in crystal morphology (size, shape, degree of aggregation) that reflects the characteristics of the overall serum microenvironment.

[0028] It should be noted that, provided that crystal formation is induced and analyzable image features are generated, the composition and parameters of the reaction system can be adjusted. For example, the buffer solution can be an acetate-sodium acetate buffer system, a phosphate buffer system, a Tris buffer system, or other suitable buffer systems; the pH can be adjusted within an acidic, neutral, or weakly alkaline range according to the crystal formation requirements; the serum volume, buffer volume, and the concentration and ratio of uric acid or other crystallization substrates can be optimized according to the crystal formation effect. The reaction temperature can be room temperature, 37°C, or other suitable temperatures for crystal formation; the incubation time can be extended or shortened according to the crystal formation rate and morphological stability. The reaction vessel can also be a 96-well plate, a 384-well plate, a glass slide, a microfluidic chip, or other carriers suitable for standardized crystallization and imaging, which can also be used as alternative embodiments of the present invention.

[0029] Step 4: Crystal formation After adding the samples, cover the 96-well plate and place it in a 37°C incubator for 1 hour. 37°C is the standardized reaction temperature, which is beneficial for obtaining stable and reproducible crystal morphologies. After 1 hour of incubation, the uric acid crystals reach an observable size, such as 5-50 μm in length, while preventing excessive incubation that could lead to overgrowth or clumping of the crystals, affecting single-crystal segmentation and analysis.

[0030] This reaction system triggers supersaturated crystallization by abruptly altering the solubility environment of uric acid (from alkaline, room temperature to acidic, 37°C). Various biomolecules in serum can act as nucleation sites or growth regulators, and the morphological characteristics of the final uric acid crystals (such as needle-like, rod-like, rhomboid, and degree of aggregation) are closely related to the complex internal environment of serum.

[0031] It should be noted that, without changing the core technical concept of "inducing the formation of crystal carriers with observable morphological differences from samples," the crystal system is not limited to uric acid crystals. Uric acid, sodium urate, calcium oxalate, calcium phosphate, cholesterol, salt crystals, or other crystallization systems that can be stably formed in vitro and are affected by the sample microenvironment can be used. As long as the crystal system can form observable and quantifiable crystal morphological differences with the participation of subject samples, and can be used to generate kidney status-related phenotypic scores, abnormal phenotypic indices, progression trend scores, or other intermediate phenotypic indicators, it can be used as an alternative implementation of this invention.

[0032] Subsequently, an inverted microscope (such as the Nikon Ti2) equipped with a high-resolution color CMOS camera was used to image uric acid crystals in bright-field mode. Image resolution was no less than 1024×1024 pixels, and magnification was controlled within the range of 2x to 10x to ensure clear coverage of crystal structure edges and texture features. Consistent exposure was maintained during acquisition, and the shooting time, well location, and sample number were recorded to ensure data traceability.

[0033] It should be noted that, provided that crystal morphology information can be obtained, image acquisition equipment and imaging methods can be substituted. In addition to ordinary bright-field microscopy, phase-contrast microscopy, polarized light microscopy, fluorescence microscopy, confocal microscopy, high-content imaging systems, automated microscopy scanning systems, or mobile phone microscopy imaging devices can also be used. The image acquisition channels can be selected as bright-field, dark-field, fluorescence, polarized, or multi-channel combined imaging according to the optical characteristics of the crystal. Acquisition parameters such as magnification, exposure time, gain, focal length, number of fields of view, and image resolution can all be adjusted according to the specific application scenario.

[0034] In this embodiment, the crystal image acquisition module also includes a crystal segmentation unit. The crystal segmentation unit is a key intermediate link between crystal image acquisition and image feature extraction. Its core task is to accurately separate the crystal region from the background region from the original microscopic image, thereby obtaining a standardized single crystal image for use by the subsequent image feature extraction module.

[0035] To achieve crystal segmentation, the first step is to construct a labeled training dataset, specifically: Professionals manually annotate the collected serum-induced uric acid crystal microscopic images pixel by pixel. The annotations include the crystal foreground area, crystal outline or crystal boundary, generating a segmentation label map corresponding to the original image, in which the crystal area and background area are distinguished by different pixel values.

[0036] In a preferred embodiment, approximately 200 manually labeled or manually corrected crystal images constitute the initial training set. Based on the initial training set, the Cellpose instance segmentation model is used for training, fine-tuning, or parameter optimization to achieve automatic identification and segmentation of crystal regions. The Cellpose instance segmentation model can output a foreground mask or instance segmentation result corresponding to a single crystal based on the boundary, contour, grayscale distribution, and morphological features in the crystal image, thereby obtaining a standardized crystal region image that can be used for subsequent feature extraction.

[0037] When using the Cellpose instance segmentation model for crystal segmentation, the input image's channel parameters are typically set to single-channel grayscale input to reduce computational redundancy and preserve crystal boundary, contour, and texture information. During model training, the Cellpose framework is used to train, fine-tune, or optimize the crystal instance segmentation model based on manually labeled or corrected crystal image datasets. In one specific embodiment, the learning rate is set to 0.0001, the batch size to 8, and the training cycle to at least 500 epochs to obtain an instance segmentation model suitable for uric acid crystal images. During training, data augmentation techniques such as rotation, scaling, flipping, brightness perturbation, or contrast adjustment can be combined to improve the model's robustness in segmenting crystal images of different sizes, shapes, and imaging backgrounds.

[0038] In addition, deep learning segmentation models with encoder-decoder structures such as U-Net and DeepLab V3+, or traditional image processing algorithms, such as combining Otsu thresholding, morphological operations and edge detection, can be used to achieve automatic identification and segmentation of crystal regions.

[0039] It should be noted that the instance segmentation principle of the Cellpose instance segmentation model, the encoder-decoder structures of U-Net and DeepLab V3+, instance segmentation models of Mask R-CNN and StarDist, or traditional image processing algorithms such as Otsu thresholding, morphological operations, and edge detection are already publicly available in existing technologies and will not be described again here. Besides the Cellpose instance segmentation model, the aforementioned deep learning segmentation models or traditional image processing algorithms can also be used to achieve automatic identification and segmentation of crystal regions.

[0040] After training, fine-tuning, or parameter optimization, the Cellpose instance segmentation model outputs a crystal instance mask, contour information, or instance segmentation result for each original crystal image during the inference phase. Subsequently, the crystal contour is extracted based on the mask corresponding to each crystal instance, and a standardized single-crystal image is obtained by cropping from the original image based on the contour's bounding rectangle. Alternatively, the crystal mask can be used as a transparency channel or foreground region constraint to generate a standardized single-crystal image containing crystal foreground information. Further, post-processing steps such as removing small-area noise, filling holes, smoothing boundaries, applying area thresholds, or separating adjacent crystals can yield relatively clean crystal masks and single-crystal images.

[0041] To ensure the accuracy and reliability of the segmentation results, some segmentation results can be manually reviewed, focusing on whether the crystal boundary segmentation is accurate and whether there are any omissions, missegments, adjacent crystal adhesion, boundary offsets, or background misidentifications. For the segmentation errors found, iterative improvements can be made by manually correcting annotations, supplementing training samples, adjusting model parameters, or optimizing post-processing rules, thereby improving the accuracy and robustness of crystal region segmentation.

[0042] Based on the crystal mask or standardized single crystal image, the crystal's morphological features such as area, perimeter, aspect ratio, and roundness can be directly calculated. The standardized crystal image obtained after segmentation can also be used as input for the image feature extraction module, so that subsequent depth image feature extraction focuses on the crystal itself rather than background interference, providing high-quality instance-level representation for multi-instance learning aggregation.

[0043] It should be noted that, without changing the purpose of crystal foreground region segmentation, in addition to the Cellpose instance segmentation model, other deep learning segmentation models or traditional image processing algorithms can also be used to achieve crystal region segmentation.

[0044] The image feature extraction module in this embodiment takes a standardized single uric acid crystal image as input after processing by the crystal segmentation unit, and outputs a fixed-length image feature vector corresponding to each image. Its core function is to extract explicit morphological parameters using image processing algorithms and / or extract implicit image features using deep neural networks, converting the original pixel information into a numerical representation that can be processed by a computer.

[0045] It should be noted that the crystal image features in this embodiment may include explicit morphological parameters and / or depth image features. Explicit morphological parameters may include one or more of the following: crystal area, perimeter, aspect ratio, roundness, convex hull area, concavity ratio, mean gray level, standard deviation of gray level, entropy, contrast, and fractal dimension. Implicit depth image features can be automatically extracted from the crystal image by a deep neural network and are used to characterize morphological information such as crystal edges, texture, local contrast, overall morphological category, arrangement, size uniformity, aspect ratio tendency, or surface roughness pattern. The aforementioned explicit morphological parameters and / or depth image features can both be used to achieve quantitative characterization of the morphological phenotype of uric acid crystals and can serve as alternative or combined implementation methods of the present invention.

[0046] The image feature extraction module can employ various mainstream deep neural network architectures, including convolutional neural networks (such as ResNet, EfficientNet, DenseNet, and ConvNeXt) and visual Transformers (such as ViT and SwinTransformer). In this embodiment, the image feature extraction module uses the ConvNeXt-Base model, which demonstrates superior performance in image classification and feature transfer tasks. Specifically, the classification layer at the end of the ConvNeXt-Base model is removed, retaining only its feature extraction portion (from the input layer to the global average pooling layer or after), outputting a 1024-dimensional floating-point vector as the image feature vector for each input normalized crystal image. This process can be performed sequentially on multiple original crystal field-of-view images, multiple crystal region images, or multiple single crystal instances from the same subject. After segmentation, the original crystal images of the same subject can generate a corresponding number of normalized crystal region images or single crystal images, and further generate a corresponding number of image feature vectors. The dimensionality of each image feature vector remains consistent. For example, when ConvNeXt-Base is used as the backbone network for feature extraction, each crystal image instance can output a 1024-dimensional floating-point vector. The image feature vectors output by the image feature extraction module can characterize the crystal morphology phenotypic information in the original image, and have uniform dimensions and stable values, which can be directly used as input instances for the multi-instance learning aggregation module.

[0047] In this embodiment, the multi-instance learning aggregation module is used to treat all image feature vectors corresponding to the same subject as a sample bag, calculate the attention weight of each image feature vector within the sample bag using a gated attention mechanism, and perform weighted aggregation of each image feature vector based on the attention weight to obtain the subject-level uric acid crystal morphology features. Specifically: Each image feature vector is input into two parallel nonlinear transformation branches: the first branch obtains an attention candidate representation through a linear transformation and a tanh activation function, and the second branch obtains a gated representation through a linear transformation and a sigmoid activation function. Subsequently, the outputs of the two branches are multiplied element-wise to obtain the gated attention representation, which is then mapped to the attention score corresponding to each image instance through a linear transformation. The attention scores of all image instances within the same sample bag are softmax normalized to obtain the attention weights for each image instance. Based on these attention weights, the feature vectors of each image are weighted and summed to obtain the subject-level uric acid crystal morphology features.

[0048] It should be noted that in this weighted aggregation process, image instances with high characterization value of crystal morphology (e.g., images with typical crystal morphology, clear contrast, and no impurities) will automatically receive higher attention weights, while low-value or interfering instances will receive lower weights, thereby achieving adaptive and robust aggregation.

[0049] It should be noted that, given a single subject or a single sample as the sample package, the aggregation method for multiple crystalloid images can be replaced. Besides gated attention pooling, mean pooling, max pooling, Top-K pooling, attention pooling, Transformer aggregation, graph neural network aggregation, or weighted voting can also be used to integrate information from multiple crystalloid images from the same subject into a subject-level feature representation. Any method that can summarize the information from multiple crystalloid images into sample-level phenotypic output results, sample-level phenotypic representations, or sample-level intermediate phenotypic parameters can be used as an alternative implementation of this invention.

[0050] Subject-level crystal morphology phenotypic characterization integrates comprehensive morphological information from multiple valid crystal images or multiple single crystal instances from the subject, which can be used as input to the subsequent phenotypic scoring generation module to generate kidney status-related phenotypic scores.

[0051] In this embodiment, the phenotypic scoring generation module receives subject-level uric acid crystal morphology features from the multi-instance learning aggregation module and generates continuous scores such as abnormal phenotypic index, progression trend score, and fusion morphology score according to a preset task type. Specifically: The first type of generated continuous score is an abnormal phenotypic index. The phenotypic score generation module receives subject-level uric acid crystal morphology features from the multi-instance learning aggregation module. Simultaneously, it reads the normal sample reference set and the abnormal sample reference set from the reference sample storage module.

[0052] It should be noted that the normal sample reference set refers to a collection of multiple subject-level uric acid crystal morphology feature vectors obtained from serum samples of a group of subjects who were pre-identified as healthy or did not exhibit any target kidney-related abnormalities. These vectors were processed using the same induction crystallization, image acquisition, feature extraction, and multi-instance learning aggregation workflow as described in this method. The normal sample reference set reflects the distribution of crystal morphology phenotypes in the absence of specific kidney abnormalities.

[0053] The abnormal sample reference set refers to a collection of multiple subject-level uric acid crystal morphology feature vectors obtained from serum samples of a group of subjects pre-identified as having a certain abnormal kidney condition (such as different stages of chronic kidney disease or specific pathological changes), after undergoing the same identical processing procedure. The abnormal sample reference set represents the crystal morphology phenotypic features associated with a specific abnormal condition.

[0054] For the morphological features of uric acid crystals at the subject level, the similarity or distance between them and each feature vector in the normal sample reference set and / or abnormal sample reference set is calculated. The measurement methods include: Euclidean distance involves treating the uric acid crystal morphology feature vector at the subject level as a point in a high-dimensional space. The Euclidean distance from this point to each feature vector in the normal reference set is calculated, and the average distance (or the distance to the center vector of the set) is taken. Similarly, the average Euclidean distance to the abnormal reference set is calculated. The ratio of the two average distances, or the normalized result, reflects the degree to which the crystal morphology of the tested sample deviates from normal. For example, the smaller the distance to the normal reference set and the larger the distance to the abnormal reference set, the closer the abnormal phenotypic index is to 0; conversely, it is closer to 1.

[0055] Cosine similarity is calculated by taking the average of the cosine similarity between the uric acid crystal morphology feature vector at the subject level and each feature vector in the normal sample reference set (or the similarity with the central vector of the set). A cosine similarity closer to 1 indicates a consistent direction, meaning similar crystal morphology patterns. Similarly, the average cosine similarity with the abnormal sample reference set is calculated. By comparing the two similarities, for example, mapping the difference between normal and abnormal similarities to the 0-1 interval, an abnormal phenotypic index is obtained.

[0056] Mahalanobis distance is calculated by plotting the covariance matrix and its inverse matrix against a reference set (e.g., a normal sample reference set), and then calculating the Mahalanobis distance from the subject-level uric acid crystal morphology feature vector to the center of this set. This distance has eliminated the influence of inter-dimensional correlation and dimensional differences. The Mahalanobis distances from the subject-level uric acid crystal morphology feature vectors to the normal sample reference set and from the subject-level uric acid crystal morphology feature vectors to the abnormal sample reference set are compared, and the abnormal phenotypic index is obtained through distance ratios or logistic mapping.

[0057] In practical use, the most suitable measurement method can be pre-selected or adaptively selected based on the distribution characteristics of the reference samples. It should be noted that regardless of the measurement method used, the final output abnormal phenotypic index is a continuous value within the range of 0 to 1. This index is a quantitative indicator used to quantify the degree of deviation between the subject-level uric acid crystal morphology characteristics and the normal / abnormal sample reference set. It can serve as reference data for subsequent statistical analysis, trend monitoring, or comparison with other clinical information. To verify the effectiveness of the abnormal phenotypic index generated by the system in this embodiment, a reference group validation set including healthy samples and chronic kidney disease samples was constructed. The system in this embodiment performed a series of processing steps on the serum induced uric acid crystal images of each subject in the validation set, ultimately obtaining the corresponding abnormal phenotypic index. By comparing the index distributions of the healthy sample group and the chronic kidney disease sample group, a receiver operating characteristic (ROC) curve was plotted. The horizontal axis of this curve represents 100% - specificity (%), and the vertical axis represents sensitivity (%). The ROC curve was used to evaluate the system's discriminative ability. Figure 2 As shown, when the ConvNeXt series model is used as the backbone network for image feature extraction, the area under the ROC curve (AUC) reaches 0.975, indicating that the abnormal phenotypic index has a good ability to distinguish the phenotypic differences between the healthy reference group and the chronic kidney disease reference group.

[0058] like Figure 3 As shown, Figure 3 The confusion matrix is ​​specifically a 2×2 table, where columns represent the true reference group and rows represent the biased grouping of the system output. According to... Figure 3The results show that the proportion of samples that were actually in the healthy reference group but were biased towards being in the healthy reference group was 0.919; the proportion of samples that were actually in the healthy reference group but were biased towards being in the chronic kidney disease reference group was 0.081; the proportion of samples that were actually in the chronic kidney disease reference group but were biased towards being in the healthy reference group was 0.089; and the proportion of samples that were actually in the chronic kidney disease reference group and were biased towards being in the chronic kidney disease reference group was 0.911. These results suggest that the abnormal phenotypic index can reflect the differences in crystal morphology phenotype between the healthy reference group and the chronic kidney disease reference group, and can serve as a continuous quantitative reference for the degree of deviation between the tested sample and different reference sets.

[0059] The second type of generated continuous score is the progression trend score, used to quantify the tendency of subject-level uric acid crystal morphology features in a preset stage. In this embodiment, the system includes a stage reference dataset storage module, which divides the stage-related reference dataset according to preset staging rules (e.g., early and mid-to-late stage of chronic kidney disease), including at least the crystal image feature set of the first stage samples and the crystal image feature set of the second stage samples. The construction method of these two feature sets is the same as that of the normal / abnormal sample reference set: multiple subjects clinically confirmed to be in the first stage are selected, and their respective subject-level feature vectors are obtained through serum-induced crystallization, image acquisition, feature extraction, and multi-instance learning aggregation, which together constitute the first stage reference set; similarly, multiple subjects in the second stage are selected to construct the second stage reference set.

[0060] The phenotypic scoring generation module first receives the subject-level uric acid crystal morphology feature vector from the multi-instance learning aggregation module. Then, it calculates the similarity or distance between the subject-level uric acid crystal morphology feature vector and the first-stage reference set, denoted as the first metric; simultaneously, it calculates the similarity or distance between the subject-level uric acid crystal morphology feature vector and the second-stage reference set, denoted as the second metric.

[0061] It should be noted that the similarity or distance can be measured by Euclidean distance, cosine similarity, or Mahalanobis distance, and the specific method can be selected based on the distribution characteristics of the reference set.

[0062] The phenotypic scoring module generates a progress trend score based on the comparison between the first and second metrics. Specifically, the comparison method varies depending on the type of metric used. If distance is used as a metric, a smaller distance indicates closer proximity to that stage. The first distance is compared with the second distance, and a progress trend score between 0 and 1 is obtained through normalization. A score closer to 0 indicates the sample is closer to the first stage, and a score closer to 1 indicates it is closer to the second stage. Specifically, the first distance refers to the distance between the subject-level uric acid crystal morphology feature vector and the first-stage reference set; the second distance refers to the distance between the subject-level uric acid crystal morphology feature vector and the second-stage reference set.

[0063] If similarity is used as a metric, a higher similarity indicates closer proximity to the target stage. The second similarity is then compared to the first similarity, and normalized to a progression trend score between 0 and 1. A score closer to 0 indicates closer proximity to the first stage, and a score closer to 1 indicates closer proximity to the second stage. Specifically, the second similarity refers to the similarity between the subject-level uric acid crystal morphology feature vector and the second-stage reference set; the first similarity refers to the similarity between the subject-level uric acid crystal morphology feature vector and the first-stage reference set.

[0064] In this embodiment, the progress trend score is used to quantitatively reflect the tendency of the crystal morphology of the sample under test in a preset stage: the closer to 0, the more it resembles the phenotype of the first stage (such as the early stage), and the closer to 1, the more it resembles the phenotype of the second stage (such as the middle and late stages). This score is only used as a quantitative indicator of the morphological change trend.

[0065] Finally, the progression trend score was validated. During validation, subjects known to be in stage one were grouped together, and subjects known to be in stage two were grouped separately. A progression trend score was calculated for each subject. Since a progression trend score closer to 1 indicates a stronger bias towards stage two, and closer to 0 indicates a stronger bias towards stage one, samples could be propensally assigned to either the stage one or stage two reference group based on different thresholds. ROC curves were then plotted based on the propensity grouping results at different thresholds to evaluate the ability of the progression trend score to distinguish between crystal morphology differences between stage one and stage two samples.

[0066] like Figure 4 As shown, when ConvNeXt is used as the backbone network for image feature extraction, the area under the ROC curve (AUC) reaches 0.865. These results suggest that the progression trend score has a certain ability to distinguish the differences in crystal morphology phenotype between the first-stage and second-stage samples, indicating that the morphological phenotype in serum-induced uric acid crystal images can undergo quantifiable changes with stage progression.

[0067] like Figure 5 As shown, Figure 5A confusion matrix is ​​used to visually demonstrate the tendency of the progress trend score to classify samples in the first and second stages. In one specific embodiment, the first-stage samples may correspond to samples in the early stage, and the second-stage samples may correspond to samples in the progress stage. Figure 5 The data shows that in the early stage samples, a higher proportion tended to be classified as early stage, while a lower proportion tended to be classified as progressive stage; conversely, in the progressive stage samples, a higher proportion tended to be classified as progressive stage, while a lower proportion tended to be classified as early stage. (Combined with...) Figure 4 The ROC curve (AUC=0.865) and Figure 5 The confusion matrix indicates that the progression trend score based on serum-induced uric acid crystal images can reflect the differences in crystal morphology phenotype between the first-stage and second-stage samples.

[0068] The third type of continuous score is the fusion morphology score, which is generated through the system's clinical indicator acquisition and coding module, multimodal fusion module, and phenotypic score generation module. Specifically: This embodiment of the system includes a clinical indicator acquisition and encoding module, used to acquire the subject's clinical feature vector. This vector contains a series of routine indicators related to kidney status, such as serum creatinine, estimated glomerular filtration rate, urine protein, hemoglobin, blood urea nitrogen, uric acid, etc., each of which is numerical data. Because the original clinical data may contain missing values, different dimensions, and non-normal distributions, the clinical indicator acquisition and encoding module first performs preprocessing.

[0069] Preprocessing includes imputing missing values ​​(e.g., median imputation), performing logarithmic transformation on skewed indicators (e.g., log1p transformation), and normalizing or standardizing all indicators (e.g., Z-score or Min-Max scaling) to make the indicators comparable in numerical range. The preprocessing results in a well-formed clinical feature vector.

[0070] The clinical indicator acquisition and encoding module transforms clinical feature vectors into higher-dimensional or more abstract clinical encoded features through encoding mapping. The encoding mapping can use a multilayer perceptron or a fully connected network to map the original dimensional clinical data into a fixed-length feature vector (e.g., 64-dimensional or 128-dimensional), thereby learning the nonlinear interaction information between clinical indicators and providing a richer representation for subsequent multimodal fusion.

[0071] The multimodal fusion module receives two inputs: subject-level uric acid crystal morphology features from the multi-instance learning aggregation module, and clinical coding features from the clinical indicator acquisition and coding module. Since these two features originate from image modality and structured data modality respectively, and have different dimensions and semantics, effective integration is required. The multimodal fusion module can employ methods such as splicing fusion, weighted fusion, attention fusion, cross-attention fusion, post-fusion voting, or gated fusion. In this embodiment, a gated fusion mechanism is used to achieve adaptive fusion. The specific process is as follows: First, the subject-level uric acid crystal morphology features are input into the image projection layer of the multimodal fusion module, such as a fully connected layer with an activation function, to obtain image hidden features. Simultaneously, clinical coding features are input into the clinical projection layer of the multimodal fusion module, i.e., another fully connected layer with an activation function, to obtain clinical hidden features. Then, the image hidden features and clinical hidden features are concatenated into a long vector and fed into the gating network. The gating network typically consists of linear transformations, layer normalization, ReLU activation, Dropout, and a final sigmoid function, outputting a gating factor, which is a scalar between 0 and 1. Finally, the image hidden features and clinical hidden features are weighted and summed according to the gating factor to obtain the multimodal fusion feature, whose expression is: Multimodal fusion feature = gating factor × image hiding feature + (1 - gating factor) × clinical hiding feature.

[0072] In this way, the multimodal fusion module can automatically learn whether it relies more on image information or clinical information in the current sample, achieving dynamic adjustment. The fused multimodal features integrate dual information from crystal morphology and clinical indicators.

[0073] It should be noted that when conducting phenotypic analysis related to renal disease subtypes, phenotypic stratification studies, or other combined phenotypic modeling, in addition to crystalloid image information, clinical laboratory indicators, demographic information, medical history information, urinalysis indicators, pathological information, or other omics data can also be integrated. Multimodal fusion methods are not limited to gated fusion; direct splicing, weighted fusion, attention fusion, cross-attention fusion, dual-tower network fusion, or post-fusion voting can also be used. Clinical variables are not limited to serum creatinine, eGFR, urinary protein, hemoglobin, urea nitrogen, uric acid, albumin, and electrolytes; relevant indicators can be added or removed according to the specific research objectives.

[0074] The phenotypic scoring generation module takes multimodal fusion features as input, passes them through an output layer (such as one or more fully connected layers followed by a linear neuron), and maps them to a fused morphological score. The fused morphological score is a continuous numerical value, typically ranging from 0 to 1, used to characterize the tendency of the tested sample to fall within a predefined subtype reference group.

[0075] During the training phase, a subtype-related reference dataset is first constructed according to predefined subtype grouping rules (such as pathological or clinical classification). This dataset is divided based on these rules, for example, according to pathological, clinical, or research stratification criteria. Each subtype contains several subject samples clearly belonging to that subtype. For each known subtype reference sample, subject-level crystal morphology features are obtained through serum-induced crystallization, image acquisition, segmentation, deep feature extraction, and gated attention aggregation. Simultaneously, clinical indicators are preprocessed and encoded into clinical coded features. These features, along with the subtype labels, constitute the reference dataset. Based on this dataset, the output module is trained using supervised learning or metric learning. One approach is to train a classifier (e.g., a fully connected layer with softmax) that can predict subtype categories based on the input fusion features. During training, the model learns feature mappings related to subtype discrimination. Another approach is to use similarity comparison learning. For example, a contrastive loss function is used to make the fusion features of samples of the same subtype close together in space, while the fusion features of samples of different subtypes are far apart. Simultaneously, a learnable center vector is reserved for each subtype, or the feature mean of a reference sample is used directly as the subtype representative. After training, the output module gains the ability to map fusion features to continuous subtype-related scores.

[0076] During the application phase, the system extracts the multimodal fusion features of the subjects and inputs them into the pre-trained output module. The phenotypic scoring generation module generates one or more continuous fusion morphological scores based on the degree of matching between the multimodal fusion features and each reference subtype. This score only reflects the degree to which the tested sample tends to each reference subtype in terms of overall morphology and clinical indicators, and can be used for propensity observation, similarity comparison, or trend monitoring. It is a non-diagnostic quantitative reference information.

[0077] Figure 6 , Figure 7 This document presents the validation results of a subtype-related fusion scoring model based on the fusion of serum-induced uric acid crystalloid images and clinical indicators. The validation task involved dividing the subject samples into an IgA nephropathy-related reference group and a non-IgA nephropathy-related reference group. The system outputs a fusion morphology score through multimodal fusion, namely the fusion of crystalloid image features and clinical indicator codes, and performs subtype-related propensity analysis based on this score. Figure 6 The ROC curve shows an area under the ROC curve (AUC) of 0.849, indicating that the fusion morphology score has a certain ability to distinguish phenotypic differences between the IgA nephropathy-related reference group and the non-IgA nephropathy-related reference group. Figure 7For the confusion matrix, the diagonal values ​​show that: in the true IgA nephropathy-related reference group samples, 0.828 were biased towards being classified as IgA nephropathy-related reference group; in the true non-IgA nephropathy-related reference group samples, 0.870 were biased towards being classified as non-IgA nephropathy-related reference group. These results suggest that the fusion morphology score can reflect the subtype-related phenotypic differences between the IgA nephropathy-related and non-IgA nephropathy-related reference groups and can serve as a quantitative reference for subtype propensity analysis.

[0078] Example 2 This embodiment provides a method for processing information on chronic kidney disease based on uric acid crystal morphology parameters. This method is applied to the chronic kidney disease information processing system based on uric acid crystal morphology parameters, as described in Embodiment 1. Figure 8 As shown, the method includes the following steps: Step S1: Preparation of uric acid crystals. Methods for obtaining uric acid crystals include: Peripheral venous blood was collected from the subjects using a serum separation tube and centrifuged at 3500 rpm for 5 minutes. The supernatant serum was separated to obtain a serum sample, which was then transferred to an EP tube. 0.1 g of uric acid was added to 3 mL of 0.667 mol / L NaOH solution and mixed thoroughly until completely dissolved to form a supersaturated uric acid solution. 60 μL of acetate-sodium acetate buffer and 8 μL of serum sample were added sequentially to the reaction plate, followed by 10 μL of supersaturated uric acid solution, and the mixture was quickly mixed. The reaction plate was then incubated at 37°C for 1 hour to induce the formation of uric acid crystals.

[0079] Step S2: Acquire multiple uric acid crystal images corresponding to the subject's ex vivo serum sample. The multiple uric acid crystal images are acquired under uniform imaging conditions and from the same sample of the same subject after in vitro induced crystallization of uric acid crystals.

[0080] Step S3: Use image processing algorithms and / or deep learning models to extract morphological and / or depth image features from each uric acid crystal image, and output the corresponding image feature vector.

[0081] Step S4: Take all image feature vectors corresponding to the same subject as a sample bag, use a gated attention mechanism to calculate the attention weight of each image feature vector in the sample bag, and perform weighted aggregation of each image feature vector based on the attention weight to obtain the subject-level uric acid crystal morphology features.

[0082] Step S5: Generate a continuous kidney state-related phenotypic score based on subject-level uric acid crystal morphology features.

[0083] According to the method steps of this embodiment, the following beneficial effects can be achieved: 1. Fully utilize the overall serum microenvironment information: By inducing the formation of uric acid crystals in serum in vitro, the synergistic effect of multiple molecules (proteins, metabolites, inflammatory factors, etc.) in serum is transformed into a visualized crystal morphology, breaking through the limitations of traditional methods that rely on only a few biochemical indicators, and more comprehensively capturing individual differences in the internal environment.

[0084] 2. Automated and standardized image segmentation and feature extraction: Based on deep learning segmentation models (such as Cellpose), automatic identification and segmentation of crystal regions are achieved. Combined with deep neural networks (such as ConvNeXt), morphology-related depth image features are automatically extracted, reducing human subjective intervention in the image analysis stage and ensuring the consistency and repeatability of feature extraction.

[0085] 3. Adaptive multi-image aggregation to improve the stability of subject-level representations: The multi-instance learning aggregation module based on gating attention mechanism is adopted to treat multiple crystal images of the same subject as a sample package, automatically assigning high weight to high-information images and low weight to low-quality images, and weighted aggregation to obtain more robust subject-level crystal morphology features, overcoming the defects of simple averaging or max pooling being susceptible to noise interference.

[0086] 4. Generate continuous kidney status-related phenotypic scores and provide refined quantitative output: Based on the comparison or multimodal fusion of subject-level characteristics with a preset reference set (normal / abnormal, early / late, different subtypes), output continuous scores (abnormal phenotypic index, progression trend score, fusion morphology score), which can more delicately reflect the degree of deviation or tendency between the sample and the reference state compared with discrete staging or binary classification results.

[0087] 5. Multimodal fusion enriches phenotypic information dimensions: Crystal morphology features and routine clinical indicators (serum creatinine, eGFR, urinary protein, etc.) are adaptively integrated through a gating fusion mechanism, dynamically adjusting the contribution ratio of image and clinical information, generating a fused morphology score, providing additional information dimensions that complement single modalities, and enriching phenotypic information dimensions.

[0088] 6. Supplementary information from sources different from traditional clinical indicators: The generated crystal morphology characteristics and scores are derived from the overall serum microenvironment, which is different from the information dimensions of existing blood / urine test indicators. It has potential complementarity and can serve as an additional reference for existing assessment methods, enriching the phenotypic reference information dimensions of chronic kidney disease-related states.

[0089] 7. Standardized automated image analysis and model aggregation process after image acquisition, which is efficient and reproducible: Based on standardized serum-induced crystallization and unified image acquisition, this invention forms an automated information processing process after image acquisition through automated image segmentation, feature extraction, multi-instance learning aggregation and phenotypic scoring output. This reduces human subjective intervention in the image analysis stage, improves analysis efficiency, result consistency and reproducibility, and is suitable for crystal morphology phenotypic analysis of large-scale samples.

[0090] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0091] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0092] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in the present invention, and these should all be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A chronic kidney disease information processing system based on uric acid crystal morphology characteristic parameters, characterized in that, include: The crystal image acquisition module is used to acquire multiple uric acid crystal images corresponding to the subject's ex vivo serum sample. The multiple uric acid crystal images are acquired under uniform imaging conditions from the same sample of the same subject after in vitro induced crystallization of uric acid crystals. The image feature extraction module is used to extract morphological and / or depth image features from each of the uric acid crystal images using image processing algorithms and / or deep learning models, and output the corresponding image feature vector. The multi-instance learning aggregation module is used to treat all image feature vectors corresponding to the same subject as a sample bag, and to calculate the attention weight of each image feature vector in the sample bag using a gated attention mechanism. Based on the attention weight, the image feature vectors are weighted and aggregated to obtain the subject-level uric acid crystal morphology features. The phenotypic score generation module is used to generate a continuous kidney state-related phenotypic score based on the subject-level uric acid crystal morphology characteristics.

2. The chronic kidney disease information processing system based on uric acid crystal morphology characteristic parameters according to claim 1, characterized in that, The kidney state-related phenotypic score includes an abnormal phenotypic index; the system also includes: The reference sample storage module stores a normal sample reference set and an abnormal sample reference set. The phenotypic scoring generation module generates the abnormal phenotypic index based on the similarity or distance between the subject-level uric acid crystal morphology features and each feature vector in the normal sample reference set and / or the abnormal sample reference set.

3. The chronic kidney disease information processing system based on uric acid crystal morphology characteristic parameters according to claim 1, characterized in that, The kidney status-related phenotypic score includes a progression trend score; The system also includes: A stage reference dataset storage module is used to store stage-related reference datasets divided according to a preset stage division rule. The stage-related reference datasets include at least the crystal image feature set of the first stage samples and the crystal image feature set of the second stage samples. The phenotypic scoring generation module calculates a first metric between the subject-level uric acid crystal morphology features and the crystal image feature set of the first-stage sample, and calculates a second metric between the subject-level uric acid crystal morphology features and the crystal image feature set of the second-stage sample, and generates the progression trend score based on the comparison result of the first metric and the second metric.

4. The chronic kidney disease information processing system based on uric acid crystal morphology characteristic parameters according to claim 1, characterized in that, Also includes: The clinical indicator acquisition and encoding module is used to acquire the clinical feature vector of the subject, and to preprocess and encode the clinical feature vector to obtain clinical coded features; wherein, the clinical feature vector includes one or more of serum creatinine, estimated glomerular filtration rate, urine protein, hemoglobin, blood urea nitrogen, and uric acid; A multimodal fusion module is used to fuse the subject-level uric acid crystal morphology features with the clinical coding features to generate multimodal fusion features; The phenotypic scoring generation module maps the multimodal fusion features into a fusion morphological score and outputs it.

5. The chronic kidney disease information processing system based on uric acid crystal morphology characteristic parameters according to claim 4, characterized in that, The preprocessing performed on the clinical feature vector by the clinical indicator acquisition and encoding module includes one or more of the following: missing value imputation, normalization, standardization, outlier handling, or logarithmic transformation.

6. The chronic kidney disease information processing system based on uric acid crystal morphology characteristic parameters according to claim 4, characterized in that, The multimodal fusion module employs a gated fusion mechanism, including: An image projection layer is used to map the subject-level uric acid crystal morphology features into image hidden features; A clinical projection layer is used to map the clinical encoded features to clinical hidden features; A gating network is used to receive the concatenated vector of the image hidden features and the clinical hidden features, and output a gating factor; The weighted fusion unit is used to perform a weighted summation of the image hidden features and the clinical hidden features according to the gating factor to obtain the multimodal fusion features.

7. The chronic kidney disease information processing system based on uric acid crystal morphology characteristic parameters according to claim 1, characterized in that, The crystal image acquisition module further includes: The crystal segmentation unit is used to segment the crystal foreground region of the uric acid crystal image to obtain one or more of the following: crystal foreground mask, crystal outline, crystal foreground image, crystal region image, or standardized single crystal cropped image. The segmented crystal region image, crystal foreground image, or standardized single crystal cropped image is used as the input of the image feature extraction module.

8. The chronic kidney disease information processing system based on uric acid crystal morphology characteristic parameters according to claim 1, characterized in that, The morphological features include explicit morphological parameters, which include one or more of the following: crystal area, perimeter, aspect ratio, roundness, convex hull area, concavity ratio, gray mean, gray standard deviation, entropy, contrast, and fractal dimension.

9. A method for processing information on chronic kidney disease based on uric acid crystal morphology parameters, characterized in that, The method is applied to the chronic kidney disease information processing system based on uric acid crystal morphology characteristic parameters as described in any one of claims 1 to 8; the method includes: Multiple uric acid crystal images corresponding to the subject's ex vivo serum sample were acquired, wherein the multiple uric acid crystal images were acquired under uniform imaging conditions from the same sample of the same subject after in vitro induced crystallization of uric acid crystals; The morphological features and / or depth image features of each uric acid crystal image are extracted using image processing algorithms and / or deep learning models, and the corresponding image feature vector is output. All image feature vectors corresponding to the same subject are taken as a sample bag. The attention weight of each image feature vector in the sample bag is calculated using a gated attention mechanism. Based on the attention weight, each image feature vector is weighted and aggregated to obtain the subject-level uric acid crystal morphology features. A continuous kidney state-related phenotypic score is generated based on the subject-level uric acid crystal morphology characteristics.

10. The method for processing information on chronic kidney disease based on uric acid crystal morphology parameters according to claim 9, characterized in that, The method for obtaining uric acid crystals includes: Peripheral venous blood was collected from the subjects using a serum separation tube, centrifuged at 3500 rpm for 5 minutes, and the supernatant serum was separated to obtain a serum sample, which was then transferred to an EP tube. Add 0.1 g of uric acid to 3 mL of 0.667 mol / L NaOH solution and mix thoroughly until completely dissolved to form a supersaturated uric acid solution. Add 60 μL of acetate-sodium acetate buffer and 8 μL of serum sample to the reaction plate in sequence, and then add 10 μL of the supersaturated uric acid solution and mix quickly. The reaction plate was placed in a 37°C incubator and incubated for 1 hour to induce the formation of uric acid crystals.

Images