Cell detection method and system for glomerular pathological section image

Abstract

The invention relates to the technical field of glomerular pathological section cell detection, and discloses a cell detection method and system for a glomerular pathological section image, and the method and system are used for dynamically adjusting a detection threshold value based on the morphological characteristics of cells in the glomerular pathological section image so as to solve the problem that the cell boundary is fuzzy. The contrast enhancement degree is regulated and controlled according to the non-uniform illumination characteristic of the image, so that the detail loss area in the low-quality image is optimized; adjusting classification model parameters according to different types of glomerular lesion features so as to reduce the probability of occurrence of false detection and missing detection; and setting the number of iterations of the segmentation algorithm according to the cell overlapping degree to realize complete separation of dense cells. Through the scheme of the embodiment of the invention, the problem of how to regulate and control the detection threshold value according to the morphological characteristics of the cells in the glomerular pathological section so as to solve the problem of fuzzy cell boundary can be solved, and the phenomena of false detection and missing detection are reduced by taking the lesion specific marker in the training data set as input guidance.

Description

Technical Field

[0001] The present invention relates to the technical field of cell detection in glomerular pathological slice images, and in particular to a cell detection method and system for glomerular pathological slice images. Background Art

[0002] A cell detection method and system for glomerular pathology slice images is a technical means for automatically identifying and analyzing cell structures in glomerular pathology images. Its goal is to achieve accurate positioning and classification of cells through computer vision algorithms.

[0003] However, this method still faces several challenges: First, when cell boundaries are blurred, how to dynamically adjust the detection threshold based on cell morphological characteristics; second, uneven illumination can lead to loss of details in low-quality images, so the degree of contrast enhancement needs to be adjusted to optimize image clarity; third, different types of glomerular lesions may lead to false detection and missed detection, requiring adaptive adjustment of classification model parameters; fourth, densely packed cells are difficult to completely separate, and the number of segmentation algorithm iterations needs to be adjusted according to the degree of cell overlap; fifth, weak signal cells may not be effectively detected due to the weak expression intensity of specific pathological markers, which requires improving detection sensitivity by adjusting the label weights. The quality of solving these problems directly affects the accuracy and reliability of the detection method. Summary of the Invention

[0004] In order to solve the above technical problems, the present invention provides the following technical solutions: A cell detection method and system for glomerular pathological section images, comprising: The detection threshold is dynamically adjusted based on the cell morphological characteristics in glomerular pathological slice images to solve the problem of blurred cell boundaries. The contrast enhancement level is controlled according to the uneven illumination characteristics of the image to optimize the areas of lost detail in low-quality images. The classification model parameters are adjusted according to the characteristics of different glomerular lesions to reduce the probability of false detection and missed detection. The number of segmentation algorithm iterations is set according to the degree of cell overlap to achieve complete separation of densely packed cells. The method of dynamically adjusting the detection threshold based on the cell morphological characteristics in the glomerular pathological section image includes: obtaining the average grayscale value I of the target cell area, setting the initial detection threshold Th0, and dynamically adjusting the detection threshold by the formula Th=Th0+α*(II0), where I0 is the reference grayscale value and α is the threshold sensitivity factor. If the adjusted threshold Th exceeds the preset maximum threshold Tmax, Th is limited to Tmax.

[0005] Preferably, regulating the contrast enhancement degree according to the uneven illumination characteristics of the image includes: Count the local brightness variance Var_local from the input image; Initialize the enhancement factor γ0; Update the enhancement factor γ according to the following formula: γ0 + β * log (Var_local / V0), where V0 is the standard deviation value and β is the enhancement strength coefficient; If the calculated result is lower than the minimum enhancement factor γmin, γ is limited to γmin.

[0006] Preferably, after determining the local brightness variance, the method includes: Select an appropriate pixel window size S=[1,9] according to the variance range; For low-brightness areas, a higher enhancement ratio γlow=γ_high+δ is applied, where δ is the low-light compensation increment; Use adaptive gamma correction to optimize brightness; Repeatedly check whether the optimized image meets the global consistency test criteria.

[0007] Preferably, adjusting the classification model parameters according to the characteristics of different types of glomerular lesions includes: Obtain more fine-grained data features F=[F1,F2,Fk] of the lesions, representing the specific pattern of each type of lesion; The best hyperparameter set P_best={P_min,P_mean,} is obtained through training; Combined with the formula Wij=f(P_best,Fj), use the weight function to assign the weight value Wij of each feature, f is the predefined mapping rule; Reconstruct the neural network structure using the updated weight set.

[0008] Preferably, the following constraints are added to ensure that the model adapts to the balance of samples of multiple disease categories: If the number of sample labels Y is severely unevenly distributed, the balanced loss Loss_balanced=Sum[(N_max / Ncw_i)] is introduced, where Nc represents the number of samples per category and w_i represents the element in the weight vector of each sample; Additional sampling is performed on the minority class to achieve a balanced number; Substitute the adjusted new dataset into the optimizer training process; Test the final output accuracy and stability performance indicators and correct model bias.

[0009] Preferably, the method for improving the cell separation effect includes setting the following segmentation logic: Estimate the radius of a single cell Rc and calculate the possible maximum coverage density Dmax=Rc 2 ; Set the number of iterations N_iters to satisfy the relationship ceil((A_image / D) / k) where A_image is the area of ​​the entire image, D=Dcell is the effective segmentation unit, and the k value determines the degree of conservatism; If the difference in segmentation area between two iterations is less than the convergence limit Err, the iterative calculation is terminated directly; Otherwise, continue the refinement steps until all specified rounds are completed.

[0010] Preferably, the conditions for determining the degree of overlap of newly added cells include: Record the intersection-over-union ratio of the overlapping area O_ij=A_overlap / (A_i+A_j); The case where O_ij>=5 is defined as severe overlap; For this type, a more sophisticated strategy is adopted to increase the iterative segmentation step size; When all relevant parts reach the ideal structure, further operations are stopped to save resource consumption.

[0011] Preferably, the expansion of a weak signal enhancement mechanism to support identification of more complex situations includes: Measure the expression intensity of the marker of interest, C_expr; The additional bonus score is obtained using the formula Δweight=(e^(κ*|C_base / C_target|)+ζ,where κζ are the exponential and basic gain parameters, C_base represents the baseline concentration, and C_target is the target expected expression rate).

[0012] Preferably, the improved process outside the detection module includes: Implement a second-order verification algorithm based on the preliminary marking to ensure reliable result screening; Perform cross-correlation analysis on all potential candidate units to reduce the risk of false positive interference; Special emphasis is placed on adding geometric contour-assisted correction techniques at ambiguous boundary locations; Finally, information from multiple dimensions is integrated to form a comprehensive evaluation system to improve the overall system efficiency.

[0013] Preferably, the enhanced automatic calibration part is implemented as follows: Design the automatic calibration loop structure Loop_AutoCalibrate; At the beginning of each iteration, the latest state parameter set Para_set_t is retrieved and the subsequent working parameters are updated accordingly: Para_set_new=fusion(Para_Set_T); If the fluctuation amplitude of the output value of the evaluation function Score_Evaluation remains below the ε criterion for n consecutive rounds, the convergence exit mechanism is confirmed to be triggered, where ε defines the error accuracy; The final configuration file after the verification is saved is used to start the continuation task next time.

[0014] Compared with the prior art, the present invention has the following beneficial effects: 1. By dynamically adjusting the detection threshold based on the cell morphological features in glomerular pathological section images to solve the problem of blurred cell boundaries, regulating the contrast enhancement degree according to the uneven illumination characteristics of the image to optimize the areas with lost details in low-quality images, adjusting the classification model parameters according to the characteristics of different types of glomerular lesions to reduce the probability of false detection and missed detection, and setting the number of iterations of the segmentation algorithm according to the degree of cell overlap to achieve complete separation of dense cells, the solution of the embodiments of the present disclosure can solve the problem of how to regulate the detection threshold according to the cell morphological features in glomerular pathological sections to solve the problem of blurred cell boundaries.

[0015] 2. By adjusting key variables in gamma correction or other adaptive gain algorithms to enhance the contrast between low-light and dark areas, restore details and ensure the accuracy of subsequent analysis, use lesion-specific markers in the training dataset as input guidance, and flexibly adjust configuration items such as the activation function form, inter-layer connection weights or hyperparameter values ​​in the model to ensure that the predicted output is highly consistent with the actual annotation, thereby reducing false detection and missed detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 Schematic diagram of the cell detection method and system arrangement process of glomerular pathological section images of the present invention; Figure 2 This is a schematic diagram of the threshold image optimization process of the present invention; Figure 3 This is a schematic diagram of the neural network structure flow of the present invention; Figure 4 It is a flow chart of the numerical system of the present invention. DETAILED DESCRIPTION

[0017] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0018] See also Figure 1 - Figure 4The present invention provides a cell detection method and system for glomerular pathology slide images. Prior to preprocessing the glomerular pathology slides, several key steps and corresponding challenges must be identified. These steps include dynamically adjusting the detection threshold based on morphological features, controlling contrast enhancement based on image illumination, adjusting classification model parameters based on different lesion characteristics, setting the number of segmentation algorithm iterations based on cell overlap, and regulating label weights.

[0019] First, in the first step, the core of the detection method is to solve the problem of blurred cell boundaries in glomeruli by dynamically adjusting the detection threshold. This step uses the morphological characteristics of specific cells in the image, including size, shape and texture information as the main reference factors to dynamically determine the threshold range of detection. The key to this strategy is to combine prior knowledge and real-time calculated statistics to determine the area that is most likely to represent the cell. For example, in one embodiment, an initial probability model can be constructed by measuring the grayscale variation range of a series of annotated cell edges, and the detection boundary can be estimated based on this. If an area with insignificant boundary changes is detected (such as some blurred edges or shadow effects), the threshold will be gradually lowered to include more suspected areas, and the reliability of the results will be further verified by other filters at a later stage.

[0020] Secondly, to address detail loss in low-quality images, a contrast-adaptive enhancement strategy is employed to further explore the image characteristics of the input pathology slides. This method first analyzes the global brightness distribution curve to identify shadows or bright spots caused by unbalanced image illumination. It then introduces a weighted enhancement factor to adjust the contrast intensity of different sub-regions. For example, if certain areas appear dark due to insufficient staining but retain weak characteristic signals, the pixel contrast stretching amplitude can be appropriately increased for these local sub-blocks. For areas with normal or overexposure, only the existing differences need to be appropriately maintained to avoid excessive noise amplification. This refined processing ultimately enables better recovery of subtle information hidden in the background.

[0021] The third step focuses on reducing false positives and missed detections in different types of lesions. For this purpose, a multi-dimensional learning model architecture is specially designed to allow users to pre-train several category-specific network branches for a variety of common types of lesions, and then use the switching structure to select a model instance that best suits the characteristics of the current sample set to participate in the workflow control parameter adjustment task. In actual application, for example, when the system encounters a proliferative lesion sample, it may be found that this type of data emphasizes the judgment rules of cell size heterogeneity, so the relevant weight distribution will focus on reflecting this characteristic direction. Similarly, in the context of non-proliferative lesions, a relatively simple judgment mode may be enabled to increase speed while ensuring a certain accuracy.

[0022] To address the difficulty of distinguishing individual units within densely packed clusters, a recursive subdivision mechanism is proposed to achieve complete separation. The basic concept is to sequentially check the relationships between adjacent points for significant interference before making merge or split decisions. In other words, whenever a significant overlap within a preliminary cluster is detected, a new round of refinement is triggered. For example, if a first rough block split reveals a significant number of unseparated target clusters, a second, more rigorously constrained partitioning operation is automatically restarted until the expected conditions are met.

[0023] The final step involves addressing marker expression intensity. Here, we introduce a method for enhancing the capture of targets with low signal strength. This involves introducing a floating weight coefficient to connect the strengths and weaknesses of each category's attributes. This is then combined with a formula to predict the reaction characteristics of external chemical substances to achieve sensitivity compensation. For example, if a test scenario shows that a specific antibody-stained area has faded to a level that is difficult to detect, the corresponding sensitivity ratio within this range is immediately increased to increase the chance of recognition, thereby effectively improving the capture rate.

[0024] The present invention dynamically adjusts the detection threshold based on the cell morphological characteristics in the glomerular pathological section image. This method mainly includes the following specific steps: obtaining the average grayscale value of the target cell area; setting the initial detection threshold; adjusting the detection threshold through a formula and limiting it when it exceeds the preset upper limit.

[0025] The first step is to obtain the average grayscale value, I, of the target cell region. This is achieved by calculating the average grayscale value of each pixel within the cell region. Grayscale values ​​reflect image brightness and, in this scenario, measure the light and dark characteristics of the cell region. This step ensures that the detection threshold can be adjusted based on the actual cell morphology.

[0026] The second step is to set the initial detection threshold Th0. The value of Th0 is typically pre-set based on empirical data or experimental results. This threshold serves as a fixed starting point, upon which subsequent steps will dynamically adjust. The range of Th0 can be flexibly set based on the specific characteristics of the medical image, and the optimal value is typically determined through experimentation. It provides an initial reference point for the dynamic threshold.

[0027] The third step is to dynamically adjust the detection threshold through the formula Th=Th0+α*(I-I0). Among them, I0 is a fixed reference grayscale value, which represents the standard grayscale of the ideal cell area; α is the threshold sensitivity factor, which usually ranges from 0.1 to 1. The optimal value needs to be selected based on the detection accuracy requirements and error analysis. The meaning of this formula is to compare the average grayscale value I of the current cell area with the reference value I0. The difference is amplified by α and applied to the initial detection threshold to form a new dynamic threshold Th. This formula is designed to make the threshold automatically adapt to different cell morphologies and lighting conditions, thereby improving the accuracy and robustness of recognition.

[0028] The fourth step is to limit the detection threshold Th when it exceeds the preset maximum threshold Tmax. This is to prevent unreasonably high thresholds in extreme cases from interfering with cell detection or missing important details. For example, when Th>Tmax, the limiting threshold Th is set to Tmax to ensure that it does not exceed the acceptable range.

[0029] Taking the detection of epithelial cells in glomerular pathology images as an example, in one embodiment, the average grayscale value in the target cell area is measured to be I=85. Assuming that the reference grayscale value is I0=70, the initial detection threshold is Th0=50, the sensitivity factor α is set to 0.8, and the maximum allowable threshold Tmax=100. Substituting into the above formula, the new detection threshold Th=50+0.8*(85-70)=62. It can be seen that this threshold can dynamically match the actual image characteristics well and ensure detection accuracy. In addition, if the adjustment result shows that Th exceeds 100, it will be limited to Tmax to ensure rationality. Specifically, this method can provide more accurate data support in complex medical diagnoses.

[0030] The present invention regulates the degree of contrast enhancement based on the uneven illumination characteristics of an image. The following specific steps are described first: the first step is to calculate the local brightness variance Var_local from the input image; the second step is to initialize the enhancement factor γ0; the third step is to update the enhancement factor γ according to the formula γ=γ0+β*log(Var_local / V0), where V0 is the standard deviation value and β is the enhancement strength coefficient; the last step is to determine whether it is lower than the minimum enhancement factor γmin. If the condition is met, γ is forcibly limited to γmin.

[0031] By calculating the local brightness variance Var_local from the input image, this operation aims to assess the brightness uniformity of each image region. This requires calculating the brightness variation within a local region to measure the impact of uneven lighting. Next, in the second step, the enhancement factor γ0 is initialized. This enhancement factor represents the baseline improvement in image contrast. Its initialization ensures that subsequent algorithms can dynamically adjust based on the initial state. For the formula update in the third step, γ=γ0+β*log(Var_local / V0), where Var_local represents the brightness variance of the current image block. V0 is a standard value typically selected based on historical statistical data. β is a parameter that controls the sensitivity of the enhancement, typically set between [0.5 and 1.5]. The optimal value depends on the actual data distribution. For example, if the image lighting is very uneven, this formula can increase contrast to highlight details. The fourth step is to prevent excessive contrast reduction and information loss, ensuring minimal effective adjustment.

[0032] In one embodiment, when processing glomerular pathology slices, for example, detecting specific cell types. Assuming that a slice contains more dark areas but also has a bright background area, in actual calculations, it may be found that a certain local Var_local reaches a larger value, such as 30, and the preset standard deviation value V0 is set to 10, and the enhancement coefficient β=1 and the initial value γ0=1 are used. After substitution, the new γ=1+1*log(30 / 10) is calculated to be approximately equal to 2, indicating that the contrast enhancement needs to be further increased. If the new γ value obtained at this time is much smaller than the minimum limit value γmin=0.8 set, it is forcibly increased to 0.8 to avoid the phenomenon of too low contrast damaging visual resolution.

[0033] The above content demonstrates the adaptation process and its logical basis for the special scenario of pathological sections, enabling the system to achieve intelligent adjustment under different lighting conditions, thereby accurately capturing weak characteristic signals for medical judgment purposes.

[0034] After determining the local brightness variance, the present invention further includes the following steps: First, a specific value S = [1, 9] is selected from the candidate pixel window size based on the range of the local brightness variance. The S parameter here represents the size of the analyzed area in the image, ranging from a single pixel to a square window with a side length of 9 pixels. The selection of the optimal value depends on the contrast requirements between the cell features and the surrounding environment in the actual image. Smaller S values ​​are suitable for capturing subtle but clear features, while larger S values ​​are more suitable for smoothing complex backgrounds. By analyzing the relationship between the local brightness variance and the target cell structure, an optimal window can be selected to balance detail preservation and noise filtering.

[0035] Next, for low-brightness areas, the parameter γlow = γ_high + δ is set. In this formula, γ_high represents a correction factor for normal brightness areas, typically ranging from [0.8 to 1.2] to moderately adjust the brightness distribution. δ is a compensation increment parameter ranging from [0.1 to 0.5]. This increment is specifically used to enhance low-brightness areas, ensuring that key features in pathology slides are not obscured by low light levels. The formula is designed to allow for more significant enhancement in low-brightness areas without compromising the consistency of the overall visual effect.

[0036] Furthermore, adaptive gamma correction is used to achieve optimized global brightness balance. In this process, each small area of ​​the image is independently mapped to a new brightness value based on the statistical properties of the selected window size S. This step aims to reduce the impact of uneven illumination while maintaining the clarity of spatial details in the image.

[0037] Finally, the results of the above processing are verified to meet the global consistency test criteria. This includes checking whether the standard deviation of the image brightness distribution, texture continuity, and the effectiveness of key cell outline extraction meet preset thresholds. Only after passing the consistency check does the image pre-processing phase end and proceed to the next step of cell detection.

[0038] In one example, suppose the initial variance calculation results for a glomerular pathology slice image show that the target cell is located in a localized dark area and the background has strong uneven illumination. Therefore, the window size S=5 is set to balance detail capture and smoothing. The γlow coefficient is then increased by δ=0.3 to enhance the representation of low-brightness cell areas. Adaptive gamma correction is applied to improve brightness uniformity across the slice. Finally, an inspection reveals a significant decrease in standard deviation and a clearer outline, confirming that the image processing has achieved the expected quality standards.

[0039] The present invention adjusts the classification model parameters based on the characteristics of different types of glomerular lesions. First, a more fine-grained data feature F = [F1, F2, Fk] of the lesions is obtained. The key to this step is to extract unique patterns from various lesion types in pathological slide images. Each pattern represents an attribute of a specific lesion area, such as the degree of inflammatory cell aggregation or fibrosis level, within a range defined by specific experimental data. The goal is to construct a complete and detailed feature set as the basis for further processing.

[0040] Next, the optimal hyperparameter set P_best = {P_min, P_mean, ...} is obtained through training. This involves selecting the specific configuration values ​​from the candidate hyperparameters that best suit the task at hand, such as the learning rate or number of layers in the deep learning network, to ensure optimal detection model performance. Typically, these hyperparameters require verification and adjustment based on multiple experiments.

[0041] Then, a weight function is used to assign a weight value, Wij, to each feature using the formula Wij = f(P_best, Fj). In this formula, f represents a predefined mapping rule designed to account for the weighted influence of hyperparameters on the importance of different lesion characteristics. The parameter Wij energy measures the weighted contribution of the j-th lesion to the overall feature system, with values ​​ranging from [0 to 1]. The optimal state occurs only when the assignment truly reflects the actual lesion dynamics. This setting is based on the fact that different lesion characteristics have different impacts on the model results, so appropriately setting this weight helps optimize the final output quality.

[0042] The final step is to reconfigure the neural network architecture using the updated weight set. Specifically, the new weights are used to adjust the original network connections and computational layer settings to adapt to the latest feature distribution requirements, making the entire system more suitable for identifying different types of glomerular lesions.

[0043] In one example, consider a case study of focal segmental glomerulosclerosis. Initially, relevant lesion features are extracted, such as the degree of capillary wall thickening F1 or the level of hyaluronic acid deposition F2. Based on experimental data, an appropriate learning rate is selected to construct the optimal hyperparameter combination P_best. Based on this, a linear recursive mapping rule f is formulated to define the strength of the association between lesion features. The weights assigned to each feature are then used to rebuild a CNN architecture that is more appropriate for this scenario, significantly improving lesion classification accuracy. Specifically, a normalization method may be employed to ensure that the weights remain appropriately constant, facilitating accurate evaluation and correction of network parameter configuration.

[0044] The present invention adds the following constraints to ensure the model adapts to the balance of samples from multiple disease categories: if the number of sample labels Y is severely unevenly distributed, a specific formula is introduced to recalculate the model loss; additional sampling is then performed on minority class samples to address data imbalance; the resulting new dataset is integrated into the training optimization process; and the final results are tested and adjusted to correct any potential biases. These operations are intended to improve the recognition of multiple diseases.

[0045] First, when the distribution of the number of sample labels Y is significantly unbalanced, it is necessary to calculate a balanced loss value to guide the model to converge to a learning strategy that is more suitable for the characteristics of each category. Specifically, the formula Loss_balanced=Sum[(N_max / Nc*w_i)] is used, where the parameter N_max is the number of samples of the largest single category among all categories, ranging from the maximum of the sum of the classification labels in the training set; and the parameter Nc indicates the total number of individual samples actually owned by each category, and its value is non-negative and varies with the disease category. The weight coefficient w_i is a scalar numerical array member corresponding to each sample itself. The core idea of ​​the definition of the formula is to amplify the impact of errors in rare categories on the overall loss, so that the network can focus more on these difficult-to-capture information during backpropagation.

[0046] Then, given that the relatively small number of minority class data may lead to underfitting or neglect, specialized techniques are used to enhance the sampling step of these minority targets to expand the corresponding size to approach the level of the majority group. For example, data augmentation methods commonly used in deep learning frameworks such as flipping or rotating images can be used to accomplish this task.

[0047] Furthermore, the newly supplemented database was formally integrated into a new round of iterations within the regular optimizer. Specifically, this ensured that the mini-batch composition used in each gradient update incorporated the aforementioned improvements, ensuring that the entire system gradually gravitated toward the optimal solution search path that satisfies a more equitable distribution, rather than being pulled off course by unilateral over-dominance.

[0048] The final step, after completing all the aforementioned configuration steps, is to enter the performance test. This involves measuring two key aspects: detection accuracy and sufficient consistency of results. Based on any identified issues, adjustments are made to the original assumptions, such as the pre-set bias factors. In one example, suppose that while evaluating the effectiveness of an automated screening solution for glomerular lesions, under-localization of a particular type of abnormal cell is observed. Therefore, the associated parameter settings should be adjusted until the overall performance reaches the ideal state.

[0049] In order to improve the cell separation effect of the present invention, the various steps are first listed: estimating the radius Rc of a single cell and calculating the possible maximum coverage density Dmax=Rc accordingly 2 ; Set the number of iterations N_iters that satisfies ceil((A_image / D) / k) through A_image (full image area), D=Dcell (effective segmentation unit) and k value; if the difference in segmentation area between two iterations is less than the convergence limit Err, terminate the iterative calculation early; otherwise, continue to refine until all specified rounds are completed.

[0050] The first step is to estimate the radius of a single cell Rc and define the maximum possible coverage density Dmax=Rc based on this value. 2 . Among them, Rc represents half of the diameter of the cell in the image, and its value comes from the actual measurement of typical cell morphology in pathological sections. Dmax represents the maximum ratio that these round cells can densely cover under theoretical circumstances, that is, the area of ​​the circle occupied by a single cell is used as the reference unit. In this way, the degree of cell arrangement can be preliminarily evaluated. Since the podocytes in the glomerulus are irregular in shape but close to spherical features, this estimate has certain guiding significance. Usually, the value of Rc ranges from 5 to 10 pixels, and the optimal choice depends on the magnification ratio and staining clarity of the specific microscope equipment.

[0051] The second step involves setting the number of iterations, N_iters, using the relationship ceil((A_image / D) / k). The parameter A_image represents the total volume of the complete pathological slice image to be inspected, while D is the basic measurement value Dcell for the actual cell area division, which represents the ideal size of each valid separation. The coefficient k determines conservatism; the smaller it is, the more cautious it is about the possibility of potential unidentified cells. The recommended value is between 0.7 and 1.3. This formula is used to quantify how many times it is necessary to traverse to ensure that most target cells are covered by the analysis. Assuming that the input image size is approximately 1024×768, that is, there is an area of ​​approximately 786,432 square pixels. If D is set to a circular area with a radius of approximately 8 pixels per block, then approximately 50 rounds of iterations (calculated based on standard values) are required to meet the full capture requirements while saving redundant loop operations.

[0052] The third rule states that if the difference in the total area updated between two consecutive rounds is less than a specified threshold, Err, the decomposition process is terminated early. Otherwise, the process is run through the specified number of times. Err serves as a metric for monitoring the degree of convergence. If the newly added area is already very small relative to the overall area, there is no point in expanding further. For example, in one embodiment, if the initial total representation area S = 5000 only increases by 5 or fewer pixels after several generations, the separation task can be considered essentially complete without further performance increases.

[0053] For example, in a specific application scenario, when processing a digital scan of glomerular tissue containing numerous foot-process-like structures, the team first selected an appropriate pixel level to match the characteristics of various markers at the actual biological scale. The aforementioned framework was then applied to a progressively progressive process, from large-scale screening to localized precision correction, until the distribution status of each independent unit under optimal resolution was recorded and output as a result file for subsequent interpretation and research by medical experts. This entire approach scientifically, efficiently, and stably promoted the achievement of target functional improvements and demonstrated superior performance.

[0054] The present invention adds a new condition for judging the degree of cell overlap: First, the formula for calculating the intersection-over-union ratio of the overlapping area is O_ij=A_overlap / (A_i+A_j). Among them, A_overlap refers to the actual area of ​​the overlapping area of ​​​​the two cells, and A_i and A_j are the area values ​​​​of individual cell objects. In this scenario, all values ​​​​are positive. O_ij represents the degree of overlap. Theoretically, the value range is [0,0.5]. In actual operation, it is usually preferred not to exceed a specific value to balance accuracy and performance. The purpose of establishing this formula is to quantitatively evaluate the relative relationship between any two adjacent or touching cells, thereby assisting in the formulation of subsequent segmentation rules.

[0055] Second, we define cases where O_ij >= 5 as severe overlap, employing a more refined segmentation strategy and iteratively adjusting the step size parameter. This setting is intended to identify tightly connected cell groups caused by crowding or pathology. For example, a pathology slide image might contain multiple abnormally enlarged glomerular cells adhered to each other; this situation would be classified as a special case and handled accordingly. Since the intersection-over-union ratio mentioned above normally does not exceed 0.5, setting a threshold of 5 here is actually a hypothetical case of highly dense expression, suggesting that the algorithm's sensitivity needs to be recalibrated to be effective under more extreme conditions.

[0056] Third, once all parts involved in the classification have been optimally reconstructed, redundant processing steps are immediately terminated, thus reducing computational waste. This means that once the detection system has accurately distinguished all cells that are judged to be severely overlapping, any unnecessary recursive actions are stopped, ensuring efficiency while reducing energy consumption. In one embodiment, if a group of highly dense cell blocks has been clearly divided after several iterations, even if there is a theoretical possibility that another verification is needed, it will be proactively abandoned to save system resource overhead.

[0057] In summary, calculating the intersection-over-union ratio (IoU) determines whether a significant overlap exists, allowing for specific responses and timely action to ensure cost-effectiveness and efficiency. These principles are crucial for improving image analysis of glomerular pathology sections, particularly in enabling precise control of complex sample processing.

[0058] The present invention extends a weak signal enhancement mechanism to support the identification of more complex situations, measures the expression intensity C_expr of the marker of interest, and obtains additional bonus points through the specific formula Δweight=(e^(κ*|C_base / C_target|)+ζ).

[0059] The first step is to measure the expression intensity of the marker of interest, C_expr. During this stage, regions within the image labeled with the target biomarker (e.g., glomerular-associated antigen) are analyzed to determine their specific expression level, C_expr. This value directly determines the basic input for the subsequent reward score calculation. For example, in a glomerular pathology section, the expression of Podocin protein in podocytes may be a core evaluation metric. Therefore, C_expr specifically represents the intensity of the biomarker detected within a specific image region or group of image regions.

[0060] The second step is to calculate the bonus score using the formula Δweight = (e^(κ*|C_base / C_target|) + ζ). Four parameters, κ, ζ, C_base, and C_target, are introduced in this process. κ is an exponential control parameter that determines the sensitivity of the formula and typically ranges from 0 to 5, with 2 being the optimal empirical value. ζ represents the base gain, ensuring a certain degree of influence even when C_expr is low. Its value typically ranges from 0 to 1, with 0.5 being a desirable default. C_base is the baseline concentration used for comparison and can be set to the average marker expression level in normal renal tissue under known conditions. C_target is the target expression rate, which is the ideal standard value that the marker should have under ideal conditions. The significance of this formula lies in providing a dynamic compensation adjustment component, Δweight, based on the relationship between the actual measured value and the expected and baseline values. This design is necessary to accommodate two extreme cases: when the actual measured value is far below the ideal value, a large increment is given to enhance the discernible signal, while when it approaches or exceeds the expected value, a gradual convergence is used to avoid distortion amplification effects.

[0061] The third step adds the additional bonus score Δweight from the previous step to the original score, resulting in the final confidence score ConfidenceFinal. This means that the basic scoring result BaseScore previously generated by the model output is supplemented and optimized to form a more accurate and effective evaluation value. In one embodiment, if the initial classification scores a certain morphological abnormality as 70 points, but further calculation based on the local labeling intensity shows that Δweight reaches +25 points, the final confidence score will be 95 points, thus more accurately identifying those subtle pathological changes that are easy to miss but are actually very important.

[0062] Suppose a slide from a patient with chronic sclerosing glomerulonephritis is analyzed, and suspected early signs of injury are observed. These microlesions are often accompanied by trace overexpression of glial fibrosis factors. Although the overall expression is not significant, they still require special attention to prevent missed diagnoses or misdiagnoses. Specifically, a preliminary prediction yields a BaseScore of 65. Furthermore, the intensity of the target protein GFAP at the corresponding location, C_expr, is relatively low, approximately equal to 2.1×C_base and much smaller than C_target. Using the aforementioned conversion rules, Δweight ≈ (e^(2*|2.1|)) + 0.5, or approximately 46, is determined to be highly probable for this cell, and the significantly improved ConfidenceFinal score reflects the improvement in the correctness of the conclusion.

[0063] The improved detection module of the present invention includes the following additional process steps. First, a second-order verification algorithm is implemented on the basis of preliminary marking to ensure the screening of reliable results. This step refers to the use of a more accurate and targeted algorithm on the results generated after the initial recognition to reconfirm whether the potential target unit is correct. This step ensures high-reliability detection output by setting strict verification rules. For example, in glomerular pathological sections, several target areas that may be mesangial cells are obtained after preliminary inspection, and then the validity of these targets is further judged by mathematical expressions based on grayscale features and spatial distribution characteristics to eliminate marking points with higher errors.

[0064] Specifically, the false positive reduction mechanism is implemented through cross-correlation analysis of all candidate cells. This involves logically comparing the spatial and physical properties of multiple adjacent candidate cells, removing target regions that are inconsistent with the actual situation due to isolation, noise interference, and other factors. In one embodiment, if a candidate region exhibits an abnormal morphology and lacks support from surrounding cell connectivity, it is likely a false positive. Cross-correlation verification can eliminate this interference factor.

[0065] The issue of blurred boundaries is addressed with geometric contour-assisted correction technology. For candidate objects in complex scenarios, such as those near tissue edges, a new correction algorithm based on image geometric feature extraction is added. This technology relies on a formula definition, where parameter A represents the average brightness gradient in the boundary region, typically ranging from 0 to 1, with 0.3 to 0.5 being the preferred range to balance sensitivity and robustness. B is the normalized value of a local shape index (such as the roundness ratio). Using this formula optimizes boundary location accuracy, thereby reducing precision loss due to incorrect boundary determination.

[0066] The comprehensive evaluation system is an efficient evaluation framework established by integrating the data collected in the above stages, with the aim of improving the overall detection efficiency. The dimensions considered in this part include but are not limited to factors such as pixel quality score and unit connectivity score. In specific practice, a weight model equation can be established, such as W=a*QualityMetric+b*ConnectivityScore, where QualityMetric represents the single-cell clarity measurement value and ConnectivityScore refers to the quantitative result of the connection strength between units. When setting a and b, the different size relationships of their contributions are taken into account, and the optimal values ​​are selected to be approximately 0.7 and 0.3 respectively to reflect the focus. In this way, the entire detection module can not only have better control over the details, but also the overall operating effect has achieved the expected goals.

[0067] The enhanced automated calibration portion of this invention is implemented as follows, specifically comprising the following steps: designing an automatic calibration loop structure, Loop_AutoCalibrate; retrieving the latest state parameter set and generating new operating parameter settings; and triggering an exit mechanism if the output of the evaluation function Score_Evaluation satisfies fluctuation constraints within a certain number of rounds, and saving the final configuration file. The following is a detailed description of each step and its implementation.

[0068] First, define and design the automatic calibration loop structure, Loop_AutoCalibrate. This is a calibration mechanism based on closed-loop feedback, continuously optimizing parameters through multiple iterations until the preset accuracy requirements are met. This loop ensures a stable parameter calibration process with a high level of accuracy. In this step, it is important to clearly define the termination conditions for the iterations and avoid infinite loops.

[0069] Then, before each iteration begins, the system's latest state parameter Para_set_t is called, which is a data set in the form of multiple numerical arrays or dictionaries related to the current operating status. These parameters usually cover key setting values ​​used in glomerular image cell detection, such as brightness adjustment, edge sharpness coefficient, and target size threshold, etc., and the range may be between 0 and 1 or within a specific pixel unit interval (depending on the specific scenario). In order to update the working parameter combination Para_set_new=fusion(Para_Set_T), the above original input parameters are subjected to some kind of fusion calculation to obtain a new version setting. This step involves the design and selection of the mathematical function operation funsion(). For example, in a practical example, the weighted average formula \(P_i=\frac{(P_i^{Old}+K_i\cdotP_{Adjustment})}{C}) can be adopted, where \(P_i) represents the new estimate of a single configuration item, \(P_i^{Old}) represents the old data, Ki is the sensitivity factor at that position, and C is the overall normalization correction factor. At the optimal level, K is generally kept close to 0.6, ensuring that historical information is retained and adjustment opinions are appropriately incorporated.

[0070] Another check rule exists: if the evaluation result, Score_Evaluation, remains below the predefined error limit ε for N consecutive times (n rounds), the algorithm is deemed sufficiently accurate and the adjustment cycle ends. n represents the maximum number of generations allowed without improvement before the search attempt is terminated; a reasonable value for n is 5 to 15. ε represents the absolute tolerance limit for measurement accuracy. For example, if precision to two decimal places is required, it is set to ±0.01. This decision logic ensures that losses are avoided due to premature or late withdrawal.

[0071] Finally, the entire process is recorded, and the final valid global variable sequence is persistently stored in a configuration file (Config_file_end). This allows future restarts to resume execution at the point where the task was interrupted, without having to rebuild all intermediate data from scratch. For example, in a real-world application, these finalized hyperparameter sets can be directly saved in JSON format for later retrieval and loading, simplifying the loading process.

[0072] The present invention provides a comprehensive solution to the technical difficulties encountered in glomerular pathology slice images, thereby improving the accuracy, reliability, and sensitivity of cell detection. The following are the specific steps and technical solutions to the method.

[0073] 1. How to adjust the detection threshold according to the cell morphological characteristics in glomerular pathological sections to solve the problem of blurred cell boundaries: This invention introduces a dynamic adjustment mechanism to analyze the size, shape, and distribution characteristics of cells in the image, extract basic morphological features in the initial detection stage, and dynamically optimize the detection threshold parameters based on prior knowledge to avoid the missegmentation of cell edges or unclear boundaries due to fixed thresholds, thereby achieving more accurate contour capture.

[0074] 2. How to adjust the degree of contrast enhancement based on uneven image illumination to address the problem of detail loss in low-quality images: To improve detail loss caused by uneven image illumination, the present invention calculates an adaptive contrast enhancement model based on the local illumination distribution. By adjusting key variables in gamma correction or other adaptive gain algorithms, the contrast effect of low-light and dark areas is enhanced, details are restored, and the accuracy of subsequent analysis is ensured.

[0075] 3. How to adjust the classification model parameters according to the characteristics of different types of glomerular lesions to solve the problems of false positives and missed detections: Based on the characteristics of different types of glomerular lesions (such as IgA nephropathy or membranous nephropathy), a customizable classification model is pre-built. Using the lesion-specific markers in the training dataset as input guidance, the activation function form, inter-layer connection weights, or hyperparameter values ​​in the model can be flexibly adjusted to ensure that the predicted output is highly consistent with the actual annotation, thereby reducing false positives and missed detections.

[0076] 4. How to regulate the number of segmentation algorithm iterations based on the degree of cell overlap to address the problem of incomplete separation of densely packed cells: To address the difficulty in distinguishing individual cells due to significant cell overlap, an intelligent strategy is designed to automatically estimate the overlap ratio, rationally select the number of execution layers of the watershed method or deep learning model, and adapt the corresponding forward pass rounds (i.e., number of iterations) until all overlapping bodies are fully resolved.

[0077] 5. How to regulate marker weights based on the expression intensity of specific pathological markers to address the problem of insufficient sensitivity in weak signal cell detection: When faced with the problem of low concentrations of some markers and weak cell signals in the corresponding areas, it is proposed to use signal quantification indicators as a reference system to assign different importance coefficients to marker points. In specific implementation, the normalized relative brightness values ​​are used to redistribute the priority and contribution ratio of each detection unit in the target tracking process, ensuring that sensitive objects are retained first, greatly compensating for the shortcomings of the original solution that cannot effectively capture sparse events. The combined operation of these five steps makes the entire system highly targeted and universally adaptable, greatly improving the performance of medical diagnostic assistance.

[0078] While the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and alterations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A cell detection method and system for glomerular pathological section images, characterized in that: include: The detection threshold is dynamically adjusted based on the cell morphological characteristics in glomerular pathological slice images to solve the problem of blurred cell boundaries. The contrast enhancement level is controlled according to the uneven illumination characteristics of the image to optimize the areas of lost detail in low-quality images. The classification model parameters are adjusted according to the characteristics of different glomerular lesions to reduce the probability of false detection and missed detection. The number of segmentation algorithm iterations is set according to the degree of cell overlap to achieve complete separation of densely packed cells. The method of dynamically adjusting the detection threshold based on the cell morphological characteristics in the glomerular pathological section image includes: obtaining the average grayscale value I of the target cell area, setting the initial detection threshold Th0, and dynamically adjusting the detection threshold by the formula Th=Th0+α*(II0), where I0 is the reference grayscale value and α is the threshold sensitivity factor. If the adjusted threshold Th exceeds the preset maximum threshold Tmax, Th is limited to Tmax.

2. A cell detection method and system for glomerular pathological section images according to claim 1, characterized in that: The adjusting of the contrast enhancement degree according to the uneven illumination characteristics of the image includes: Count the local brightness variance Var_local from the input image; Initialize the enhancement factor γ0; Update the enhancement factor γ according to the following formula: γ0 + β * log (Var_local / V0), where V0 is the standard deviation value and β is the enhancement strength coefficient; If the calculated result is lower than the minimum enhancement factor γmin, γ is limited to γmin.

3. A cell detection method and system for glomerular pathological section images according to claim 1, characterized in that: After determining the local brightness variance, the method further comprises: Select an appropriate pixel window size S=[1,9] according to the variance range; For low-brightness areas, a higher enhancement ratio γlow=γ_high+δ is applied, where δ is the low-light compensation increment; Use adaptive gamma correction to optimize brightness; Repeatedly check whether the optimized image meets the global consistency test criteria.

4. A cell detection method and system for glomerular pathological section images according to claim 1, characterized in that: The adjustment of classification model parameters according to the characteristics of different types of glomerular lesions includes: Obtain more fine-grained data features F=[F1,F2,Fk] of the lesions, representing the specific pattern of each type of lesion; The best hyperparameter set P_best={P_min,P_mean,} is obtained through training; Combined with the formula Wij=f(P_best,Fj), use the weight function to assign the weight value Wij of each feature, f is the predefined mapping rule; Reconstruct the neural network structure using the updated weight set.

5. The cell detection method and system for glomerular pathological section images according to claim 1, characterized in that: The following constraints are added to ensure that the model adapts to the balance of samples of multiple disease classes: If the number of sample labels Y is severely unevenly distributed, the balanced loss Loss_balanced=Sum[(N_max / Ncw_i)] is introduced, where Nc represents the number of samples per category and w_i represents the element in the weight vector of each sample; Additional sampling is performed on the minority class to achieve a balanced number; Substitute the adjusted new dataset into the optimizer training process; Test the final output accuracy and stability performance indicators and correct model bias.

6. A cell detection method and system for glomerular pathological section images according to claim 1, characterized in that: To improve the cell separation effect, the following segmentation logic is set: Estimate the radius of a single cell Rc and calculate the possible maximum coverage density Dmax=Rc 2 ; Set the number of iterations N_iters to satisfy the relationship ceil((A_image / D) / k) where A_image is the area of ​​the entire image, D=Dcell is the effective segmentation unit, and the k value determines the degree of conservatism; If the difference in segmentation area between two iterations is less than the convergence limit Err, the iterative calculation is terminated directly; Otherwise, continue the refinement steps until all specified rounds are completed.

7. A cell detection method and system for glomerular pathological section images according to claim 1, characterized in that: The conditions for judging the degree of overlap of newly added cells include: Record the intersection-over-union ratio of the overlapping area O_ij=A_overlap / (A_i+A_j); The case where O_ij>=5 is defined as severe overlap; For this type, a more sophisticated strategy is adopted to increase the iterative segmentation step size; When all relevant parts reach the ideal structure, further operations are stopped to save resource consumption.

8. The cell detection method and system for glomerular pathological section images according to claim 1, characterized in that: The expanded weak signal enhancement mechanism supports more complex identification situations, including: Measure the expression intensity of the marker of interest, C_expr; The additional bonus score is obtained using the formula Δweight=(e^(κ*|C_base / C_target|)+ζ,where κζ are the exponential and basic gain parameters, C_base represents the baseline concentration, and C_target is the target expected expression rate).

9. The cell detection method and system for glomerular pathological section images according to claim 1, characterized in that: The improved detection module external process includes: Implement a second-order verification algorithm based on the preliminary marking to ensure reliable result screening; Perform cross-correlation analysis on all potential candidate units to reduce the risk of false positive interference; Special emphasis is placed on adding geometric contour-assisted correction techniques at ambiguous boundary locations; Finally, information from multiple dimensions is integrated to form a comprehensive evaluation system to improve the overall system efficiency.

10. The cell detection system for glomerular pathological section images according to claim 1, characterized in that: The following implementations of enhanced automated calibration are included: Design the automatic calibration loop structure Loop_AutoCalibrate; At the beginning of each iteration, the latest state parameter set Para_set_t is retrieved and the subsequent working parameters are updated accordingly: Para_set_new=fusion(Para_Set_T); If the fluctuation amplitude of the output value of the evaluation function Score_Evaluation remains below the ε criterion for n consecutive rounds, the convergence exit mechanism is confirmed to be triggered, where ε defines the error accuracy; The final configuration file after the verification is saved is used to start the continuation task next time.

Images