Coagulation specimen quality screening and classifying system and method based on image recognition

Abstract

The invention relates to the technical field of specimen image detection, and discloses a blood coagulation specimen quality screening and classification system and method based on image recognition, and the method comprises the steps: obtaining the image information of each blood coagulation specimen in a to-be-screened batch, and carrying out the image preprocessing of the image information; extracting feature information of each blood coagulation specimen in the preprocessed image information, and performing primary evaluation on the blood coagulation specimens according to the feature information and configured preset feature information; and trust score evaluation is carried out on the primary evaluation result based on a fused convolutional neural network, whether secondary evaluation is carried out on the blood coagulation specimen is determined according to the relationship between the evaluated trust score and the configured preset trust score, and if the trust score is lower than the preset trust score, secondary evaluation is determined to be carried out on the blood coagulation specimen. According to the method, through multi-level image preprocessing, feature extraction, primary evaluation and trust score evaluation, the accuracy of blood coagulation specimen quality screening is effectively improved in combination with a judgment mechanism fusing the convolutional neural network.

Description

Technical Field

[0001] The present invention relates to the technical field of specimen image detection, and in particular to a coagulation specimen quality screening and classification system and method based on image recognition. Background Art

[0002] In clinical testing, coagulation specimens, as an essential component of blood samples, have a quality that directly impacts the accuracy and reliability of coagulation test results. However, because coagulation specimens are susceptible to multiple factors, such as collection method, storage time, and transportation status, some specimens may exhibit issues such as unclear distribution of the plasma and cell layers, abnormal proportions, color deviation, and turbid sediment, seriously interfering with test accuracy. Therefore, pre-test specimen quality screening is crucial to ensuring experimental effectiveness.

[0003] Currently, laboratories typically rely on manual visual inspection of coagulation specimens, including observation of specimen stratification, color, and visible precipitation. However, this manual screening method is subject to significant subjectivity and reliance on experience, resulting in low screening efficiency and a high risk of missed detections or misjudgments when processing large numbers of samples. Meanwhile, some methods have attempted to utilize traditional image processing algorithms for automated screening, but these methods still face challenges such as unstable image feature extraction and a lack of trust mechanisms, making them difficult to meet the demands of refined quality control.

[0004] Therefore, there is an urgent need to invent a coagulation specimen quality screening technology to solve the problem that the existing coagulation specimen quality screening method relies on manual judgment or traditional image algorithms, has strong subjectivity, low efficiency, lacks stability and reliable evaluation mechanism, and is difficult to achieve efficient and accurate automated screening. Summary of the Invention

[0005] In view of this, the present invention proposes a coagulation specimen quality screening and classification system and method based on image recognition, aiming to solve the problems that existing coagulation specimen quality screening methods rely on manual judgment or traditional image algorithms, are highly subjective, inefficient, lack stability and a reliable evaluation mechanism, and are difficult to achieve efficient and accurate automated screening.

[0006] The present invention proposes a method for screening and classifying coagulation specimen quality based on image recognition, comprising:

[0007] Obtaining image information of each coagulation specimen in the batch to be screened, and performing image preprocessing on the image information;

[0008] Extracting characteristic information of each coagulation specimen from the pre-processed image information, and performing an evaluation on the coagulation specimen based on the characteristic information and the configured preset characteristic information;

[0009] The confidence score of the first evaluation result is evaluated based on the fusion convolutional neural network, and the relationship between the evaluated confidence score and the configured preset confidence score is used to determine whether to perform a second evaluation on the coagulation sample:

[0010] If the confidence score is higher than or equal to the preset confidence score, it is determined that the coagulation specimen will not be reassessed;

[0011] If the confidence score is lower than the preset confidence score, a second evaluation of the coagulation specimen is performed.

[0012] Furthermore, obtaining image information of each coagulation specimen in the batch to be screened includes:

[0013] Obtaining the mean value of the R, G, and B channels in the image information of each coagulation sample, and determining a deviation index in the image information of each coagulation sample based on the mean value of the R, G, and B channels;

[0014] Based on the relationship between the deviation index of the image information and the configured preset deviation threshold, it is determined whether the coagulation sample needs to obtain image information a second time:

[0015] When the deviation index is lower than or equal to the preset deviation threshold, it is determined that the coagulation sample does not require secondary acquisition of image information;

[0016] When the deviation index is higher than the preset deviation threshold, it is determined that the coagulation sample needs to obtain image information a second time.

[0017] Furthermore, when image information is preprocessed, the following steps are included:

[0018] Remove noise from image information based on median filtering, and enhance the local contrast of image information based on the denoised image information;

[0019] The brightness and contrast of the image information are normalized based on histogram equalization.

[0020] Furthermore, when extracting characteristic information of each coagulation specimen in the pre-processed image information, it includes:

[0021] The pre-processed image information is segmented based on threshold segmentation, and the effective analysis area of ​​the coagulation specimen is extracted;

[0022] Based on the extracted effective analysis area, color feature information and texture feature information of the effective analysis area are obtained, wherein the color feature information includes the channel mean, standard deviation and color deviation index of the image pixels in the effective analysis area calculated in the RGB color space; the texture feature information includes the grayscale change pattern of the local structure of the area extracted based on the local binary pattern;

[0023] Based on the color feature information, the color balance features and potential color cast features of the image in the effective analysis area are obtained;

[0024] Based on the texture feature information, the surface structural features and precipitation particle distribution features of the coagulation specimen area in the effective analysis area are obtained;

[0025] Based on the color balance characteristics, potential color cast characteristics, surface structure characteristics and precipitation particle distribution characteristics of the image in the effective analysis area, the shape characteristics of the plasma layer, the shape characteristics of the cell layer, and the distribution characteristics and proportion characteristics between the plasma layer and the cell layer in the effective analysis area are obtained.

[0026] Furthermore, when the coagulation specimen is evaluated based on the characteristic information and the configured preset characteristic information, the following steps are included:

[0027] Obtaining a shape feature of a preset plasma layer, a shape feature of a preset cell layer, a preset distribution feature and a preset ratio feature between the plasma layer and the cell layer in the preset feature information;

[0028] Based on the Mahalanobis distance, the degree of overlap between the shape features of the plasma layer, the shape features of the cell layer, the distribution features and the ratio features between the plasma layer and the cell layer in the feature information and the preset shape features of the plasma layer, the preset shape features of the cell layer, the preset distribution features and the preset ratio features between the plasma layer and the cell layer is obtained:

[0029]

[0030] Among them, D M is the overlap between the feature information and the preset feature information, Fsample is the feature information, where Fsample=[f1,f2,f3,f4] T , f1 is the shape feature of the plasma layer, f2 is the shape feature of the cell layer, f3 is the distribution feature between the plasma layer and the cell layer, and f4 is the ratio feature between the plasma layer and the cell layer; Fref = [f1 ref ,f2 ref ,f3 ref ,f4 ref ] T , f1 ref is the shape feature of the preset plasma layer, f2 ref To preset the shape characteristics of the cell layer, f3 ref is the preset distribution feature between the plasma layer and the cell layer, f4 ref is the preset ratio feature between the plasma layer and the cell layer, Σ is the covariance matrix between the feature information and the preset feature information;

[0031] According to the relationship between the degree of coincidence and the preset degree of coincidence, the primary score of the coagulation specimen is determined, and based on the primary score, whether the coagulation specimen is qualified is determined:

[0032] When the primary score is lower than the preset qualified score, the coagulation specimen is determined to be unqualified;

[0033] When the primary score is higher than or equal to the preset qualified analysis, the coagulation specimen is determined to be qualified.

[0034] Furthermore, when determining a primary score of a coagulation specimen based on the relationship between the degree of coincidence and a preset degree of coincidence, the following steps are included:

[0035] Determine a primary score of the coagulation sample based on the relationship between the coincidence degree and the configured first preset coincidence degree and second preset coincidence degree:

[0036] When the coincidence degree is lower than the first preset coincidence degree, the primary score of the coagulation sample is determined to be L3;

[0037] When the coincidence degree is higher than or equal to the first preset coincidence degree and the coincidence degree is lower than the second preset coincidence degree, the primary score of the coagulation sample is determined to be L2;

[0038] When the coincidence degree is higher than or equal to the second preset coincidence degree, the primary score of the coagulation sample is determined to be L1;

[0039] The first preset overlap is smaller than the second preset overlap, and L1<L2<L3.

[0040] Furthermore, when conducting a trust score assessment on an assessment result based on a fused convolutional neural network, the following steps are included:

[0041] A fusion convolutional neural network model is constructed based on network training. The characteristic information of the coagulation specimen with an unqualified evaluation result and the preprocessed image are substituted into the fusion convolutional neural network model.

[0042] Feature extraction is performed on the characteristic information of unqualified coagulation specimens and preprocessed images, and multi-scale image features are extracted based on the main network branch and auxiliary network branch of the fusion convolutional neural network model;

[0043] Based on the Sigmoid function and multi-scale image features, the confidence score of the unqualified coagulation specimen is determined:

[0044]

[0045] Among them, T is the confidence score of unqualified coagulation specimen, x i is the image feature extracted at the i-th scale, w i is the weight of the i-th feature, b is the bias term, and exp is the exponential function.

[0046] Furthermore, when building a fusion convolutional neural network model based on network training, it includes:

[0047] Perform parallel convolution operations on image information based on multiple groups of convolution kernels of different scales to extract the multi-scale color, texture and structural features of the image;

[0048] Aggregate multi-scale feature maps based on feature fusion mechanism, which includes feature concatenation and channel attention weighting;

[0049] The fused features are input into the regression output layer, and the confidence score is output based on the Sigmoid activation function;

[0050] Using historical annotated data as supervision labels, supervised training is performed based on minimizing the loss function between the network output trust score and the true credible label, where the loss function is a binary cross entropy loss function.

[0051] Furthermore, when using historically labeled data as supervisory labels and performing supervised training based on minimizing the loss function between the network output trust score and the true trustworthy label, the following steps are involved:

[0052] The historical annotated data is used as the supervision label input to construct a training sample set consisting of multi-scale image features;

[0053] Perform forward propagation on the image samples based on the fused convolutional neural network model to obtain the corresponding trust score output;

[0054] The trust score is normalized to the interval [0,1] based on the Sigmoid function as the credibility probability value predicted by the model;

[0055] Based on the training mechanism that aims to minimize the loss function between the network output trust score and the true credible labels in the historical annotated data, the fused convolutional neural network is supervised trained, and the model parameters are iteratively updated based on the backpropagation and gradient optimization algorithms.

[0056] Compared with existing technologies, the present invention offers the following advantages: Through standardized collection and preprocessing of coagulation specimen image information, a uniformly high-quality input data source is constructed, providing a stable foundation for subsequent automated analysis and improving overall anti-interference capabilities and adaptability, making it particularly suitable for batch testing scenarios. Secondly, by extracting multi-dimensional features such as color, texture, and morphology from the image and combining them with the Mahalanobis distance for a quantitative assessment, the quality of each coagulation specimen can be scientifically and objectively judged using a unified metric. Furthermore, by constructing a mapping relationship between the degree of overlap between a preset feature model and sample features, this approach achieves a shift from "empirical judgment" to "data-driven" quality assessment, significantly improving the consistency and traceability of quality assessment and reducing the risk of human misjudgment. Finally, a fused convolutional neural network is introduced to analyze the trust score of the primary assessment results, providing self-judgment capabilities. By training the mapping relationship between image features and labels through deep learning, the system can automatically perceive the credibility of the assessment results and decide whether to conduct a secondary review based on the trust score. This mechanism not only enhances the robustness of processing border samples or complex samples, but also achieves the optimal allocation of resource utilization. That is, while ensuring screening accuracy, it avoids unnecessary repeated processing of high-confidence samples, significantly improving the intelligence level and operational efficiency of the entire process.

[0057] On the other hand, the present application also provides a coagulation specimen quality screening and classification system based on image recognition, comprising:

[0058] an image acquisition module configured to acquire image information of each coagulation specimen in the batch to be screened and perform image preprocessing on the image information;

[0059] An evaluation module electrically connected to the image acquisition module, the evaluation module being configured to extract characteristic information of each coagulation specimen from the pre-processed image information, and to perform an evaluation of the coagulation specimen based on the characteristic information and configured preset characteristic information;

[0060] The analysis module is electrically connected to the evaluation module. The analysis module is configured to perform a trust score evaluation on the primary evaluation result based on a fused convolutional neural network, and determine whether to control the evaluation module to perform a secondary evaluation on the coagulation sample based on the relationship between the evaluated trust score and the configured preset trust score: if the trust score is higher than or equal to the preset trust score, the analysis module determines not to control the evaluation module to perform a secondary evaluation on the coagulation sample; if the trust score is lower than the preset trust score, the analysis module determines to control the evaluation module to perform a secondary evaluation on the coagulation sample.

[0061] It is understandable that the image recognition-based coagulation specimen quality screening and classification system and method in the above-mentioned embodiments of the present invention have the same beneficial effects and will not be described in detail. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Various other advantages and benefits will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiment below. The accompanying drawings are for illustration purposes only and are not to be considered as limiting the present invention. The same reference symbols are used throughout the drawings to represent the same components. In the drawings:

[0063] Figure 1 A flowchart of a method for screening and classifying coagulation specimen quality based on image recognition provided by an embodiment of the present invention;

[0064] Figure 2 This is a functional block diagram of the coagulation specimen quality screening and classification system based on image recognition provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0065] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art. It should be noted that, unless there is a conflict, the embodiments of the present disclosure and the features in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in conjunction with the embodiments.

[0066] like Figure 1 As shown in some embodiments of the present application, this embodiment provides a method for screening and classifying coagulation specimen quality based on image recognition, comprising:

[0067] Step S100: Obtain image information of each coagulation specimen in the batch to be screened, and perform image preprocessing on the image information.

[0068] Specifically, when obtaining the image information of each coagulation specimen in the batch to be screened, it includes: obtaining the mean of the three channels R, G, and B in the image information of each coagulation specimen, and determining the deviation index in the image information of each coagulation specimen based on the mean of the three channels R, G, and B; determining whether the coagulation specimen needs to obtain the image information a second time based on the relationship between the deviation index of the image information and the configured preset deviation threshold: when the deviation index is lower than or equal to the preset deviation threshold, it is determined that the coagulation specimen does not need to obtain the image information a second time; when the deviation index is higher than the preset deviation threshold, it is determined that the coagulation specimen needs to obtain the image information a second time.

[0069] Specifically, image preprocessing includes: removing noise from the image information based on median filtering, and enhancing the local contrast of the image information based on the denoised image information; and normalizing the brightness and contrast of the image information based on histogram equalization.

[0070] It can be understood that by introducing an RGB channel mean calculation mechanism during the image acquisition phase to measure the overall image performance across different color channels, statistically averaging the pixel values ​​of the three R, G, and B channels provides a preliminary reflection of the image's overall color balance. Furthermore, a "deviation index" is constructed based on the calculated three-channel means and compared with a preset deviation threshold to determine whether the image exhibits significant color cast or exposure issues. When the deviation index exceeds the threshold, the image is deemed abnormal and requires re-acquisition to ensure input quality for subsequent processing. Secondly, median filtering is a classic spatial domain denoising algorithm with excellent edge preservation and salt-and-pepper noise suppression. In this technical solution, median filtering is used to smooth noise in coagulation specimen images, thereby removing random interference signals caused by changes in the acquisition environment or imaging equipment. After filtering, the image's contour information is preserved while its texture details are sharper, enhancing the accuracy of subsequent feature extraction. Median filtering can significantly improve the robustness of region segmentation and boundary recognition, particularly when processing images containing tiny sediment particles. Finally, the image's brightness and contrast are normalized using histogram equalization. This method redistributes the image's grayscale distribution, resulting in a richer brightness hierarchy, enhanced dark detail, and increased overall contrast. This normalization not only improves image quality but also enhances the image's ability to convey information in subsequent feature analysis. This is particularly effective in samples with uneven exposure or background noise, effectively increasing the recognizability of key features and providing a more discriminative visual input foundation for classification and evaluation.

[0071] As can be seen, by calculating the mean of the three RGB channels in the image and constructing a deviation index that is compared against a preset threshold, the system implements automatic quality control during image acquisition. When an image exhibits significant color cast or exposure anomalies, an image re-acquisition mechanism is triggered immediately, preventing low-quality images from entering the subsequent analysis process. This mechanism effectively reduces errors caused by inconsistent image acquisition conditions, ensuring the accuracy and consistency of coagulation specimen image data from the source, and improving the robustness and reliability of the overall screening system. Secondly, median filtering is used to denoise the image, effectively eliminating random noise introduced by equipment interference or sample particles, and enhancing local image contrast while preserving edge structure. In coagulation specimen images, the visual features of detailed areas such as the cell layer and plasma layer are particularly crucial for screening judgment. This processing method helps highlight these structural differences, improving the discriminative power of subsequent feature extraction algorithms and reducing the risk of false positives and missed detections. Finally, histogram equalization normalizes image brightness and contrast, balancing differences in image appearance under different lighting conditions and achieving more consistent image quality. This processing not only improves the overall appearance of the image, but also significantly enhances the sensitivity of subsequent image analysis algorithms to key areas such as edges and textures, improves the accuracy of feature recognition and classification, and helps maintain stable screening performance in multiple scenarios.

[0072] Step S200: extracting characteristic information of each coagulation specimen from the pre-processed image information, and performing an evaluation on the coagulation specimen based on the characteristic information and configured preset characteristic information.

[0073] Specifically, when extracting the characteristic information of each coagulation specimen in the preprocessed image information, it includes: performing image segmentation on the preprocessed image information based on threshold segmentation, and extracting the effective analysis area of ​​the coagulation specimen; based on the extracted effective analysis area, obtaining the color feature information and texture feature information of the effective analysis area, wherein the color feature information includes the channel mean, standard deviation and color deviation index of the image pixels in the effective analysis area calculated in the RGB color space; the texture feature information includes the grayscale change pattern of the local structure of the area extracted based on the local binary pattern; based on the color feature information, obtaining the color balance feature and potential color cast feature of the image in the effective analysis area; based on the texture feature information, obtaining the surface structure feature and sedimentation particle distribution feature of the coagulation specimen area in the effective analysis area; based on the color balance feature, potential color cast feature, surface structure feature and sedimentation particle distribution feature of the image in the effective analysis area, obtaining the shape feature of the plasma layer, the shape feature of the cell layer, the distribution feature and the proportion feature between the plasma layer and the cell layer in the effective analysis area.

[0074] Specifically, when evaluating a coagulation specimen based on the feature information and the configured preset feature information, the process includes: obtaining the shape feature of the preset plasma layer, the shape feature of the preset cell layer, the preset distribution feature and the preset ratio feature between the plasma layer and the cell layer in the preset feature information; and obtaining the degree of overlap between the shape feature of the plasma layer, the shape feature of the cell layer, the distribution feature and the ratio feature between the plasma layer and the cell layer in the feature information and the shape feature of the preset plasma layer, the shape feature of the preset cell layer, the preset distribution feature and the preset ratio feature between the plasma layer and the cell layer based on the Mahalanobis distance:

[0075]

[0076] Among them, D M is the overlap between the feature information and the preset feature information, Fsample is the feature information, where Fsample=[f1,f2,f3,f4] T , f1 is the shape feature of the plasma layer, f2 is the shape feature of the cell layer, f3 is the distribution feature between the plasma layer and the cell layer, and f4 is the ratio feature between the plasma layer and the cell layer; Fref = [f1 ref ,f2 ref ,f3 ref ,f4 ref ] T , f1 ref is the shape feature of the preset plasma layer, f2 ref To preset the shape characteristics of the cell layer, f3 ref is the preset distribution feature between the plasma layer and the cell layer, f4 ref is the preset ratio feature between the plasma layer and the cell layer, Σ is the covariance matrix between the feature information and the preset feature information; according to the relationship between the overlap degree and the preset overlap degree, the primary score of the coagulation sample is determined, and according to the primary score, whether the coagulation sample is qualified is determined: when the primary score is lower than the configured preset qualified score, the coagulation sample is determined to be unqualified; when the primary score is higher than or equal to the preset qualified analysis, the coagulation sample is determined to be qualified.

[0077] Specifically, when determining the primary score of the coagulation sample based on the relationship between the overlap and the preset overlap, it includes: determining the primary score of the coagulation sample based on the relationship between the overlap and the configured first preset overlap and second preset overlap: when the overlap is lower than the first preset overlap, the primary score of the coagulation sample is determined to be L3; when the overlap is higher than or equal to the first preset overlap, and the overlap is lower than the second preset overlap, the primary score of the coagulation sample is determined to be L2; when the overlap is higher than or equal to the second preset overlap, the primary score of the coagulation sample is determined to be L1; wherein, the first preset overlap is less than the second preset overlap, and L1<L2<L3.

[0078] As can be understood, based on image preprocessing, a threshold-based image segmentation method is employed to precisely extract the effective analysis region within the coagulation specimen image. By focusing on the critical regions encompassing the plasma and cell layers, non-specimen information, such as external background and tube edges, can be effectively avoided from interfering with subsequent feature extraction and evaluation. This strategy ensures the targeted focus of image analysis and lays the data foundation for subsequent high-precision feature modeling. Secondly, within the effective analysis region, color and texture features are extracted by direction. Color features quantify the overall color distribution and balance of the region using channel mean, standard deviation, and color deviation in the RGB color space, assisting in determining whether color casts or lighting anomalies exist within the image. Texture features, based on local binary patterns (LBP), model grayscale variation patterns, effectively capturing surface structural changes and the distribution of sediment particles within the image. This introduces structural details beyond color information, enhancing the completeness and discriminability of feature representation. Furthermore, by further structuring low-level features such as color balance, potential color casts, surface texture changes, and particle distribution, shape and distribution characteristics of the plasma and cell layers, as well as their proportionality, are constructed. This approach achieves a fusion modeling of image "structure-hierarchy" attributes, not only reproducing the tomographic characteristics of the physical structure of coagulation specimens but also identifying potential anomalies (such as increased precipitation and blurred layer boundaries), thereby improving the ability to discern specimen quality. Furthermore, by constructing the Mahalanobis distance between the specimen feature vector and a preset reference feature vector, a covariance matrix is ​​introduced to model the correlation between multidimensional features and assess the distance of the current specimen from the "standard specimen" in feature space. Compared to methods such as the Euclidean distance, the Mahalanobis distance is more robust when handling highly correlated features and can effectively measure the global consistency of complex image structures, providing a more discriminative quality indicator for screening systems. Finally, based on the relationship between the degree of overlap and two-level preset overlap (thresholds), a three-level scoring mechanism (L1, L2, and L3) is designed to express the degree of similarity in structural features between the specimen and the reference sample, thereby enabling quantitative grading of conformity assessment. This scoring mechanism not only improves the system's flexibility in determining the quality of diverse specimens, but also facilitates subsequent processes (such as trust score assessment and review mechanism) to make automated decisions based on clear results, thereby improving overall screening efficiency and intelligence.

[0079] As can be seen, using threshold segmentation to segment the image effectively identifies and extracts the valid analysis area of ​​the coagulation specimen, avoiding interference from irrelevant factors such as tube background and shadows. This pre-screening strategy ensures that the subsequently extracted color and texture features are more representative of the specimen, enhancing the accuracy and robustness of the overall image analysis. Secondly, image features are extracted from two dimensions: color (channel mean, standard deviation, color deviation) and texture (grayscale pattern based on local binary patterns), ensuring a comprehensive reflection of key indicators such as the surface state and sediment distribution of the coagulation specimen. This multi-dimensional feature extraction approach helps to more meticulously reveal the differences between specimens of varying quality, enhancing the discriminative capabilities of intelligent recognition. Furthermore, by deeply integrating color balance, color cast, surface structure, and sediment particles, the shape, distribution, and proportion features of the plasma and cell layers are further extracted, achieving a transition from low-level image features to high-level semantic representation. This feature fusion modeling facilitates a comprehensive assessment of specimen quality, avoiding misjudgments caused by single-dimensional evaluation. In addition, compared with the traditional Euclidean distance, the Mahalanobis distance fully considers the covariance relationship between features and can more effectively reflect the degree of similarity between the current specimen and the standard sample in the multi-dimensional feature space. By using this distance to calculate the degree of overlap, it is possible to evaluate the consistency of complex image structures and enhance the recognition ability and discrimination stability of abnormal samples. Finally, by setting a multi-level overlap threshold and introducing a graded scoring (L1, L2, L3) mechanism, not only is a refined division of specimen quality achieved, but it also provides clear guidance for subsequent processing procedures such as whether to conduct a secondary review and whether to mark abnormalities. This scoring mechanism is both transparent and practical, which is conducive to efficient operation and supports rapid manual interpretation, thereby improving the intelligence level and practical operability of the overall screening process.

[0080] Specifically, the effective analysis region refers to the relevant area of ​​the coagulation specimen extracted from the image using a threshold-based segmentation method. By removing background and irrelevant parts, only the core area of ​​the specimen is analyzed. This area includes the plasma layer, cell layer, and the structural features between them. By extracting color and texture features, it helps accurately assess specimen quality and distinguishes it from other non-specimen areas in the image, thereby ensuring the accuracy and reliability of the evaluation results.

[0081] Step S300: performing a confidence score evaluation on the primary evaluation result based on the fused convolutional neural network, and determining whether to perform a secondary evaluation on the coagulation sample based on the relationship between the evaluated confidence score and the configured preset confidence score.

[0082] Specifically, when evaluating the trust score of an assessment result based on a fused convolutional neural network, the method includes: constructing a fused convolutional neural network model based on network training, substituting the characteristic information of the unqualified coagulation specimen and the preprocessed image into the fused convolutional neural network model; extracting features from the characteristic information of the unqualified coagulation specimen and the preprocessed image, and extracting multi-scale image features based on the main network branch and auxiliary network branch of the fused convolutional neural network model; and determining the trust score of the unqualified coagulation specimen based on the Sigmoid function and the multi-scale image features:

[0083]

[0084] Among them, T is the confidence score of unqualified coagulation specimen, x i is the image feature extracted at the i-th scale, w i is the weight of the i-th feature, b is the bias term, and exp is the exponential function.

[0085] Specifically, when constructing a fusion convolutional neural network model based on network training, it includes: performing parallel convolution operations on image information according to multiple groups of convolution kernels of different scales to extract the multi-scale color, texture and structural features of the image; aggregating multi-scale feature maps based on a feature fusion mechanism, which includes feature splicing and channel attention weighting; inputting the fused features into the regression output layer, and outputting the trust score based on the Sigmoid activation function; using historical annotated data as the supervision label, and performing supervised training based on minimizing the loss function between the network output trust score and the true trustworthy label, where the loss function is a binary cross entropy loss function.

[0086] Specifically, using historical annotated data as supervisory labels and performing supervised training based on minimizing the loss function between the network output trust score and the true credible label, the method includes: taking the historical annotated data as supervisory label input to construct a training sample set composed of multi-scale image features; forward propagating the image samples based on the fused convolutional neural network model to obtain the corresponding trust score output; normalizing the trust score to the [0,1] interval based on the Sigmoid function as the credibility probability value predicted by the model; based on a training mechanism with the goal of minimizing the loss function between the network output trust score and the true credible label in the historical annotated data, the fused convolutional neural network is supervised trained, and the model parameters are iteratively updated based on the backpropagation and gradient optimization algorithms.

[0087] As can be understood, image information is convolved in parallel using multiple sets of convolution kernels of different scales to extract color, texture, and structural features. Multi-scale convolution operations enable the model to capture multi-level information in the image, from local details to global structure, ensuring comprehensive capture of features of varying sizes and complexities. This multi-scale feature extraction approach makes the network more robust against different types of coagulation specimens, thereby improving assessment accuracy. Next, the multi-scale feature maps are aggregated through a feature fusion mechanism, which includes feature concatenation and channel-wise attention weighting. Feature concatenation directly merges feature maps from different scales to ensure the network utilizes information from all scales. Channel-wise attention weighting dynamically adjusts each channel of the feature map based on its contribution to the final classification result by calculating a weight for each channel. This mechanism enables the network to adaptively weight features at different scales to highlight those that are most influential in determining the confidence score, improving network accuracy and robustness. After feature fusion, the fused features are input to the regression output layer, which outputs a normalized confidence score using a sigmoid activation function. The confidence score, constrained to the interval [0, 1], represents the probability that a coagulation specimen is qualified. In practical applications, higher confidence scores indicate that the specimen's quality more closely meets the preset standards, while lower confidence scores indicate potential quality issues. The calculation of the confidence score relies not only on the image features themselves but also on historically annotated data, ensuring consistency between the assessment results and the actual specimen quality. To train the model, historically annotated data is used as supervisory labels. Model optimization is performed by minimizing the binary cross-entropy loss function between the network's output confidence score and the true, trusted labels. The cross-entropy loss function quantifies the difference between the predicted results and the true labels, enabling the network to adjust its parameters with each iteration, gradually improving accuracy. Through backpropagation and gradient optimization algorithms, the network continuously optimizes its weight parameters, ultimately learning a model that accurately assesses coagulation specimen quality. Through this training mechanism, the fused convolutional neural network model is able to learn the complex characteristics of coagulation specimen quality from a large amount of historically annotated data, thereby accurately outputting confidence scores. When the confidence score predicted by the model is higher than the preset threshold, it indicates that the coagulation sample is qualified and no secondary evaluation is required; when the confidence score is lower than the threshold, the model will automatically recommend a secondary evaluation, thereby ensuring the accuracy and efficiency of sample screening.

[0088] Specifically, based on the relationship between the evaluated trust score and the configured preset trust score, determining whether to perform a secondary evaluation on the coagulation sample includes: if the trust score is higher than or equal to the preset trust score, determining not to perform a secondary evaluation on the coagulation sample; if the trust score is lower than the preset trust score, determining to perform a secondary evaluation on the coagulation sample.

[0089] In the above-described embodiment, standardized collection and preprocessing of coagulation specimen image information creates a uniformly high-quality input data source, providing a stable foundation for subsequent automated analysis and improving overall anti-interference capabilities and adaptability, making it particularly suitable for batch testing scenarios. Secondly, by extracting multi-dimensional features such as color, texture, and morphology from the image and combining them with the Mahalanobis distance for a quantitative assessment, each coagulation specimen can be scientifically and objectively assessed for quality using a unified metric. Furthermore, by constructing a mapping relationship between the degree of overlap between a preset feature model and sample features, this solution achieves a shift from "empirical judgment" to "data-driven" quality assessment, significantly improving the consistency and traceability of quality assessment and reducing the risk of human error. Finally, a fused convolutional neural network is introduced to analyze the confidence score of the primary assessment results, providing self-assessment capabilities. By training the mapping relationship between image features and labels through deep learning, the system can automatically perceive the credibility of the assessment results and determine whether to conduct a secondary review based on the confidence score. This mechanism not only enhances the robustness of processing border samples or complex samples, but also achieves the optimal allocation of resource utilization. That is, while ensuring screening accuracy, it avoids unnecessary repeated processing of high-confidence samples, significantly improving the intelligence level and operational efficiency of the entire process.

[0090] In another preferred embodiment based on the above embodiment, Figure 2 As shown, this embodiment provides a coagulation specimen quality screening and classification system based on image recognition, including:

[0091] an image acquisition module configured to acquire image information of each coagulation specimen in the batch to be screened and perform image preprocessing on the image information;

[0092] An evaluation module electrically connected to the image acquisition module, the evaluation module being configured to extract characteristic information of each coagulation specimen from the pre-processed image information, and to perform an evaluation of the coagulation specimen based on the characteristic information and configured preset characteristic information;

[0093] The analysis module is electrically connected to the evaluation module. The analysis module is configured to perform a trust score evaluation on the primary evaluation result based on a fused convolutional neural network, and determine whether to control the evaluation module to perform a secondary evaluation on the coagulation sample based on the relationship between the evaluated trust score and the configured preset trust score: if the trust score is higher than or equal to the preset trust score, the analysis module determines not to control the evaluation module to perform a secondary evaluation on the coagulation sample; if the trust score is lower than the preset trust score, the analysis module determines to control the evaluation module to perform a secondary evaluation on the coagulation sample.

[0094] It is understandable that the image recognition-based coagulation specimen quality screening and classification system and method in the above-mentioned embodiments of the present invention have the same beneficial effects and will not be described in detail.

[0095] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of a complete hardware embodiment, a complete software embodiment, or a combination of software and hardware embodiments. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0096] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems) and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0097] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0098] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0099] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. A method for screening and classifying coagulation specimen quality based on image recognition, characterized in that: include: Obtaining image information of each coagulation specimen in the batch to be screened, and performing image preprocessing on the image information; Extracting characteristic information of each coagulation specimen from the pre-processed image information, and performing an evaluation on the coagulation specimen based on the characteristic information and the configured preset characteristic information; The confidence score of the first evaluation result is evaluated based on the fusion convolutional neural network, and the relationship between the evaluated confidence score and the configured preset confidence score is used to determine whether to perform a second evaluation on the coagulation sample: If the confidence score is higher than or equal to the preset confidence score, it is determined that the coagulation specimen will not be reassessed; If the confidence score is lower than the preset confidence score, a second evaluation of the coagulation specimen is performed.

2. The method for screening and classifying coagulation specimen quality based on image recognition according to claim 1, wherein: When obtaining image information of each coagulation specimen in the batch to be screened, including: Obtaining the mean value of the R, G, and B channels in the image information of each coagulation sample, and determining a deviation index in the image information of each coagulation sample based on the mean value of the R, G, and B channels; Based on the relationship between the deviation index of the image information and the configured preset deviation threshold, it is determined whether the coagulation sample needs to obtain image information a second time: When the deviation index is lower than or equal to the preset deviation threshold, it is determined that the coagulation sample does not require secondary acquisition of image information; When the deviation index is higher than the preset deviation threshold, it is determined that the coagulation sample needs to obtain image information a second time.

3. The method for screening and classifying coagulation specimen quality based on image recognition according to claim 2, characterized in that: When image information is preprocessed, it includes: Remove noise from image information based on median filtering, and enhance the local contrast of image information based on the denoised image information; The brightness and contrast of the image information are normalized based on histogram equalization.

4. The method for screening and classifying coagulation specimen quality based on image recognition according to claim 1, wherein: Extracting characteristic information of each coagulation specimen from the pre-processed image information includes: The pre-processed image information is segmented based on threshold segmentation, and the effective analysis area of ​​the coagulation specimen is extracted; Based on the extracted effective analysis area, color feature information and texture feature information of the effective analysis area are obtained, wherein the color feature information includes the channel mean, standard deviation and color deviation index of the image pixels in the effective analysis area calculated in the RGB color space; the texture feature information includes the grayscale change pattern of the local structure of the area extracted based on the local binary pattern; Based on the color feature information, the color balance features and potential color cast features of the image in the effective analysis area are obtained; Based on the texture feature information, the surface structural features and precipitation particle distribution features of the coagulation specimen area in the effective analysis area are obtained; Based on the color balance characteristics, potential color cast characteristics, surface structure characteristics and precipitation particle distribution characteristics of the image in the effective analysis area, the shape characteristics of the plasma layer, the shape characteristics of the cell layer, and the distribution characteristics and proportion characteristics between the plasma layer and the cell layer in the effective analysis area are obtained.

5. The method for screening and classifying coagulation specimen quality based on image recognition according to claim 4, characterized in that: When a coagulation specimen is evaluated based on the characteristic information and the configured preset characteristic information, it includes: Obtaining a shape feature of a preset plasma layer, a shape feature of a preset cell layer, a preset distribution feature and a preset ratio feature between the plasma layer and the cell layer in the preset feature information; Based on the Mahalanobis distance, the degree of overlap between the shape features of the plasma layer, the shape features of the cell layer, the distribution features and the ratio features between the plasma layer and the cell layer in the feature information and the preset shape features of the plasma layer, the preset shape features of the cell layer, the preset distribution features and the preset ratio features between the plasma layer and the cell layer is obtained: Among them, D M is the overlap between the feature information and the preset feature information, Fsample is the feature information, where Fsample=[f1,f2,f3,f4] T , f1 is the shape feature of the plasma layer, f2 is the shape feature of the cell layer, f3 is the distribution feature between the plasma layer and the cell layer, and f4 is the ratio feature between the plasma layer and the cell layer; Fref = [f1 ref ,f2 ref ,f3 ref ,f4 ref ] T , f1 ref is the shape feature of the preset plasma layer, f2 ref To preset the shape characteristics of the cell layer, f3 ref is the preset distribution feature between the plasma layer and the cell layer, f4 ref is the preset ratio feature between the plasma layer and the cell layer, Σ is the covariance matrix between the feature information and the preset feature information; According to the relationship between the degree of coincidence and the preset degree of coincidence, the primary score of the coagulation specimen is determined, and based on the primary score, whether the coagulation specimen is qualified is determined: When the primary score is lower than the preset qualified score, the coagulation specimen is determined to be unqualified; When the primary score is higher than or equal to the preset qualified analysis, the coagulation specimen is determined to be qualified.

6. The method for screening and classifying coagulation specimen quality based on image recognition according to claim 5, characterized in that: When determining the primary score of a coagulation specimen based on the relationship between the degree of coincidence and the preset degree of coincidence, the following steps are included: Determine a primary score of the coagulation sample based on the relationship between the coincidence degree and the configured first preset coincidence degree and second preset coincidence degree: When the coincidence degree is lower than the first preset coincidence degree, the primary score of the coagulation sample is determined to be L3; When the coincidence degree is higher than or equal to the first preset coincidence degree and the coincidence degree is lower than the second preset coincidence degree, the primary score of the coagulation sample is determined to be L2; When the coincidence degree is higher than or equal to the second preset coincidence degree, the primary score of the coagulation sample is determined to be L1; The first preset overlap is smaller than the second preset overlap, and L1<L2<L3.

7. The method for screening and classifying coagulation specimen quality based on image recognition according to claim 6, characterized in that: When evaluating the trust score of an assessment result based on a fused convolutional neural network, it includes: A fusion convolutional neural network model is constructed based on network training. The characteristic information of the coagulation specimen with an unqualified evaluation result and the preprocessed image are substituted into the fusion convolutional neural network model. Feature extraction is performed on the characteristic information of unqualified coagulation specimens and preprocessed images, and multi-scale image features are extracted based on the main network branch and auxiliary network branch of the fusion convolutional neural network model; Based on the Sigmoid function and multi-scale image features, the confidence score of the unqualified coagulation specimen is determined: Among them, T is the confidence score of unqualified coagulation specimen, x i is the image feature extracted at the i-th scale, w i is the weight of the i-th feature, b is the bias term, and exp is the exponential function.

8. The method for screening and classifying coagulation specimen quality based on image recognition according to claim 7, characterized in that: When building a fusion convolutional neural network model based on network training, it includes: Perform parallel convolution operations on image information based on multiple groups of convolution kernels of different scales to extract the multi-scale color, texture and structural features of the image; Aggregate multi-scale feature maps based on feature fusion mechanism, which includes feature concatenation and channel attention weighting; The fused features are input into the regression output layer, and the confidence score is output based on the Sigmoid activation function; Using historical annotated data as supervision labels, supervised training is performed based on minimizing the loss function between the network output trust score and the true credible label, where the loss function is a binary cross entropy loss function.

9. The method for screening and classifying coagulation specimen quality based on image recognition according to claim 8, characterized in that: When using historically labeled data as supervisory labels, supervised training is performed based on minimizing the loss function between the network output trust score and the true trustworthy label, including: The historical annotated data is used as the supervision label input to construct a training sample set consisting of multi-scale image features; Perform forward propagation on the image samples based on the fused convolutional neural network model to obtain the corresponding trust score output; The trust score is normalized to the interval [0,1] based on the Sigmoid function as the credibility probability value predicted by the model; Based on the training mechanism that aims to minimize the loss function between the network output trust score and the true credible labels in the historical annotated data, the fused convolutional neural network is supervised trained, and the model parameters are iteratively updated based on the backpropagation and gradient optimization algorithms.

10. A coagulation specimen quality screening and classification system based on image recognition, applicable to the coagulation specimen quality screening and classification method based on image recognition according to any one of claims 1 to 9, characterized in that: include: an image acquisition module configured to acquire image information of each coagulation specimen in the batch to be screened and perform image preprocessing on the image information; An evaluation module electrically connected to the image acquisition module, the evaluation module being configured to extract characteristic information of each coagulation specimen from the pre-processed image information, and to perform an evaluation of the coagulation specimen based on the characteristic information and configured preset characteristic information; The analysis module is electrically connected to the evaluation module. The analysis module is configured to perform a trust score evaluation on the primary evaluation result based on a fused convolutional neural network, and determine whether to control the evaluation module to perform a secondary evaluation on the coagulation sample based on the relationship between the evaluated trust score and the configured preset trust score: if the trust score is higher than or equal to the preset trust score, the analysis module determines not to control the evaluation module to perform a secondary evaluation on the coagulation sample; if the trust score is lower than the preset trust score, the analysis module determines to control the evaluation module to perform a secondary evaluation on the coagulation sample.

Images