Computer vision-based blood sample image intelligent recognition and classification system

Abstract

The application discloses a blood sample image intelligent recognition and classification system based on computer vision, relates to the technical field of pathological image analysis and computer vision, and comprises the following steps: acquiring an original collected image and performing nonlinear concentration mapping to generate a biochemical layered image containing measured concentration data; locking a cell center and dividing an initial jurisdiction through center response calculation to construct an ideal simulation single-cell layer; calculating a difference image of a simulation superimposed image and a measured image, iteratively correcting the simulation layer in a reverse direction by using concentration difference data until an independent single-cell layer containing recovered texture is obtained; and finally generating an overlapping complexity weight map based on the difference image, constructing a multi-dimensional feature data packet and inputting the multi-dimensional feature data packet into a convolutional neural network to output a classification result; the application effectively solves the recognition problem caused by cell overlapping in high-density samples by combining physical model decoupling with deep learning.

Description

Technical Field

[0001] This invention relates to the fields of pathological image analysis and computer vision technology, specifically to a computer vision-based intelligent recognition and classification system for blood sample images. Background Technology

[0002] In the field of clinical medical testing, microscopic analysis of blood smears is a key step in assessing human health and diagnosing hematological diseases. With the advancement of digital medical technology, the use of computer vision technology for automated scanning and intelligent recognition of blood smear images has become an important development trend to improve the efficiency of auxiliary diagnosis and reduce the intensity of manual labor.

[0003] Current automated identification schemes are mainly based on deep neural networks to model the features of acquired digital color images. By learning the differences in color, texture and geometric contours of different cells, they can automatically classify and count red blood cells, white blood cells and various components.

[0004] However, in the actual preparation and imaging process of blood smears, due to the limitations of smear thickness and the distribution characteristics of biological samples, cell overlap and occlusion phenomena are widespread in the field of view. When multiple targets are spatially stacked in the imaging optical path, the signal sensed by the detector is not a simple pixel pattern superposition, but a nonlinear physical coupling of the photon absorption characteristics of multiple targets. Since this superposition process changes the information expression of the pixel layer, if pattern matching is performed only from the visual texture level, the model will find it difficult to restore the true geometric shape and internal structural details of each occluded target from the physically mixed signals. This leads to the model being prone to feature attribution ambiguity or drastic fluctuations in recognition accuracy when processing high-density samples.

[0005] Therefore, how to solve the problem of accurately separating and restoring the biological characteristics of individuals in complex stacked backgrounds from the perspective of the physical mechanism of imaging has become the core technical bottleneck for improving the accuracy of intelligent blood testing. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a computer vision-based intelligent recognition and classification system for blood sample images.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows:

[0008] In a first aspect, the present invention discloses a computer vision-based intelligent recognition and classification system for blood sample images, comprising:

[0009] The image preprocessing module is used to acquire the raw images of blood samples, perform illumination correction and nonlinear concentration mapping on the raw images, and separate the mapped images into biochemically layered images containing nuclear channel data and cytoplasmic channel data; the biochemically layered images contain measured concentration data characterizing the distribution of cellular substances;

[0010] The jurisdiction division module is used to perform center response calculation on cell nuclear channel data, determine multiple cell center points, and divide the biochemical stratification image into multiple corresponding initial jurisdictions based on the cell center points;

[0011] The single-cell layer construction module is used to construct a corresponding simulated single-cell layer based on the distribution characteristics of the measured concentration data within each initial jurisdiction; the simulated single-cell layer contains simulated concentration data characterizing the cell under ideal non-overlapping conditions;

[0012] The difference calculation module is used to linearly overlay the simulated concentration data of all simulated single-cell layers to generate a synthetic image, and calculate the difference map between the synthetic image and the biochemical stratified image; the difference map contains the concentration difference data between the synthetic image and the biochemical stratified image at each pixel position.

[0013] The layer restoration module is used to reverse correct the simulated concentration data of each simulated single-cell layer using concentration difference data until the difference map meets the preset convergence condition, thereby obtaining an independent single-cell layer containing restored texture.

[0014] The feature construction and classification module is used to generate an overlap complexity weight map based on the difference map that meets the preset convergence conditions, and to construct a multi-dimensional feature data package for each cell, which includes an independent single-cell layer and an overlap complexity weight map; the multi-dimensional feature data package is input into a pre-trained convolutional neural network model, and the class classification results of each cell are output.

[0015] Secondly, this invention discloses a method for intelligent recognition and classification of blood sample images based on computer vision, comprising the following steps:

[0016] The raw images of blood samples are acquired, illumination correction and nonlinear concentration mapping are performed on the raw images, and the mapped images are separated into biochemically layered images containing nuclear channel data and cytoplasmic channel data; the biochemically layered images contain measured concentration data characterizing the distribution of cellular substances;

[0017] The center response calculation is performed on the cell nuclear channel data to determine multiple cell center points, and the biochemical stratification image is divided into multiple corresponding initial jurisdictions based on the cell center points;

[0018] Based on the distribution characteristics of measured concentration data within each initial jurisdiction, a corresponding simulated single-cell layer is constructed; the simulated single-cell layer contains simulated concentration data characterizing the cell under ideal non-overlapping conditions;

[0019] The simulated concentration data of all simulated single-cell layers are linearly superimposed to generate a synthetic image, and the difference map between the synthetic image and the biochemical stratified image is calculated; the difference map contains the concentration difference data between the synthetic image and the biochemical stratified image at each pixel location.

[0020] The simulated concentration data of each simulated single-cell layer is reverse-corrected using the concentration difference data until the difference map meets the preset convergence condition, thereby obtaining an independent single-cell layer containing restored texture.

[0021] An overlap complexity weight map is generated based on the difference map that meets the preset convergence conditions, and a multi-dimensional feature data package containing an independent single-cell layer and an overlap complexity weight map is constructed for each cell.

[0022] The multidimensional feature data package is input into a pre-trained convolutional neural network model, which outputs the classification results of each cell.

[0023] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0024] 1. By introducing the Lambert-Beer law and color deconvolution technology, image pixels are converted into measured concentration data with physical superposition characteristics; and by constructing a closed-loop mechanism of simulated single-cell layers and differential feedback iterative correction, the concentration difference data between the synthetic image and the measured image is used to drive the model for reverse numerical compensation; this decoupling method based on the conservation of material concentration can accurately separate the mixed signals of the overlapping area into independent single-cell layers from the physical level, effectively restoring the true geometric shape and material distribution of the occluded target, thus solving the technical bottleneck of cell counting and classification difficulties in high-density smears;

[0025] 2. The asymmetric compression vector is calculated using the spatial gradient features of the overlapping complexity weight graph; and the morphological decay factor of the radial basis function is directionally modulated accordingly to generate a dynamic decay coefficient that conforms to the physical compression state. This mechanism enables the simulation layer to automatically adjust to anisotropic irregular shapes (such as ellipses or polygons) according to the surrounding deconstruction uncertainty, thereby avoiding the misidentification of normal cells deformed by physical crowding as atypical pathological cells and significantly reducing the false positive rate in clinical diagnosis.

[0026] 3. A weight map of overlapping complexity was generated to quantitatively decouple uncertainty and input into the convolutional neural network as the fourth feature channel. By introducing a spatial attention mechanism into the network, the weight map was used to generate an attention mask to adaptively suppress convolutional features. This design gives the model the cognitive ability to avoid heavy and light areas, that is, to automatically reduce the weight of high-overlapping and high-error regions during the feature extraction stage and focus on the core physical features with high decoupling confidence. In this way, while retaining key pathological information, the impact of imaging artifacts and overlapping interference on classification decisions is suppressed to the greatest extent.

[0027] 4. A four-channel multidimensional feature data package containing independent nucleus / cytoplasm concentration data, original grayscale texture, and decoupled confidence weights was constructed. By backfilling the accumulated texture residuals obtained during the iteration process into the reconstruction layer, the original visual information of fine structures such as cell nucleoli and granules was preserved, and the material distribution information after physical purification was combined. This fusion of multimodal data enabled the neural network to make comprehensive judgments based on both biophysical properties (concentration / thickness) and morphological properties (texture / gloss), thus achieving refined classification of various blood cells (especially structurally complex white blood cells). Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is an overall block diagram of the system according to Embodiment 1 of the present invention;

[0030] Figure 2 This is a schematic diagram of the construction of multi-dimensional feature data packets and the CNN attention mechanism in Embodiment 1 of the present invention;

[0031] Figure 3 This is an overall block diagram of the method in Embodiment 2 of the present invention;

[0032] Figure 4 This is a flowchart illustrating the process of the method in Embodiment 2 of the present invention. Detailed Implementation

[0033] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] Application Overview: In the field of modern medical testing and pathological diagnosis, especially in the microscopic morphological analysis of blood smears, the clarity of the internal structure of cells and the definiteness of the nucleoplasmic boundary are regarded as key indicators for measuring the accuracy of identification. The acquisition of such high-quality morphological features is essentially a process of nonlinear attenuation of photon energy after absorption by matter at the physical optics level. That is, the imaging light penetrates the medium with specific biochemical components (such as chromatin of the cell nucleus and granules of the cytoplasm), and converts light energy into absorbance signal according to the Lambert-Beer law, thereby leaving a light density mapping with specific concentration distribution characteristics on the sensor target surface.

[0035] However, existing technologies lack a verification mechanism for the physical consistency between the input optical path superposition mechanism and the output pixel intensity distribution, resulting in the inability to accurately identify the widespread cell stacking and nonlinear aliasing problems in high-density smears. Stacking and aliasing manifests as multiple cells spatially overlapping along the optical axis, but existing algorithms often rely solely on surface texture features for pattern matching, ignoring the principle of additivity of optical density. Nonlinear aliasing, on the other hand, occurs when cells deform due to compression, yet the system still uses a rigid template for matching, leading to a failure to establish a strict biochemical correspondence between the edge gradient of the overlapping area and the inherent contour of a single cell. This results in missegmentation or ambiguous feature attribution of complex adhered targets, thereby affecting the confidence level of cell classification and the objectivity of clinical diagnosis.

[0036] For example, in scenarios such as leukemia screening or white blood cell classification under high magnification, when a lymphocyte partially covers a neutrophil, conventional deep learning systems can only capture the morphological outline of the mixture, but cannot distinguish whether the pixel values ​​in this area are caused by high-density nuclear-cytoplasmic aggregation or by the superposition of double membrane structures. Furthermore, when the bottom layer cells undergo asymmetric deformation due to compression by the top layer cells, the system only records the irregular boundary after fusion, failing to detect abnormal transitions in nuclear optical density values ​​and physical discontinuities in edge gradients. Specifically, the system often misclassifies such physical overlaps as atypical lymphocytes or binucleated cells, giving false positive feedback, or incorrectly classifies normal cells deformed by compression as pathological cells. This leads to the model continuously solidifying erroneous morphological perceptions and failing to form decoupled perceptions that conform to the principles of photophysical transmission.

[0037] If the above problems are not addressed, intelligent diagnostic systems will continue to lose their ability to pathologically distinguish complex samples. Specifically, the failure to decouple physical overlap will cause the model to rely excessively on local textures, leading to feature extraction paths deviating from the principles of biological material distribution, thereby weakening the ability to detect rare lesion cells. At the same time, the failure to correct for compression deformation will cause inaccurate calculation of morphological parameters (such as nucleocytoplasmic ratio and roundness), making cell classification unable to present true biological attributes, ultimately causing diagnostic results to lose their due clinical reference value. Thus, the lack of physical cognition will systematically hinder the breakthrough of precision in fully automated blood smear analysis technology, affecting the achievement of precision medicine goals.

[0038] Example 1:

[0039] like Figures 1-2 As shown, a computer vision-based intelligent recognition and classification system for blood sample images includes:

[0040] The image preprocessing module is used to acquire the raw images of blood samples, perform illumination correction and nonlinear concentration mapping on the raw images, and separate the mapped images into biochemically layered images containing nuclear channel data and cytoplasmic channel data; the biochemically layered images contain measured concentration data characterizing the distribution of cellular substances;

[0041] Before acquiring the raw images of the blood samples, this embodiment pre-constructs and configures a staining feature matrix. Specifically, this matrix is ​​not a randomly generated numerical table, but rather a logical physical correlation between the light absorption characteristics of a specific light source spectrum in the microscopic imaging system and commonly used blood staining agents (such as Wright-Gymsa stain). The construction process of the staining feature matrix involves: measuring the absorbance coefficients of different staining components in the red, green, and blue spectral channels using a spectral analyzer or by sampling a large number of samples of a single component (such as regions containing only cell nuclei and regions containing only cytoplasm), statistically determining these coefficients, and integrating them into a two-dimensional mapping structure. In this matrix, each column represents the absorption intensity vector of a staining component (such as cell nucleus or cytoplasm) in different color spaces, while each row corresponds to the spectral response channel of the imaging sensor. The reason for using a pre-constructed matrix instead of real-time extraction is to ensure that the unmixed biochemical channel data has strict physical consistency when processing different batches of blood samples, thereby avoiding recognition model failure due to variations in staining depth in individual samples.

[0042] During actual system operation, the image preprocessing module first acquires the raw image of the blood sample using a microscopic acquisition device. It is worth noting that due to the non-uniformity of the light source distribution and optical path loss, the pixel values ​​of the raw acquired image often combine the intensity characteristics of the light source with the absorption characteristics of the cells. Based on this, the system simultaneously retrieves a pre-calibrated blank background image. This image is obtained by imaging a clean slide area in a sample-free state, recording the ideal incident light distribution of the current imaging system. Specifically, the system performs a pixel-by-pixel comparison between the raw acquired image and the blank background image. Specifically, for any coordinate point in the image... Its transmittance diagram The calculation formula is as follows:

[0043] ;

[0044] in, Represents the pixel intensity at the corresponding coordinates in the original acquired image. This represents the reference intensity of the blank background image at the corresponding coordinates. Through this comparison operation, the system effectively removes background noise caused by uneven light source distribution.

[0045] Furthermore, to convert the transmittance signal into a linear signal reflecting the distribution and concentration of biological substances, the system performs a nonlinear concentration mapping, i.e., a negative logarithmic transformation, on the generated transmittance map based on the Lambert-Beer law. Physically, the absorbance of a substance (i.e., the value after concentration mapping) is linearly positively correlated with the thickness through which light passes and the substance concentration, while transmittance decays exponentially. In this embodiment, the concentration image... The calculation process is expressed as follows: ;

[0046] After this transformation, the pixel values ​​in the image are no longer simply visual brightness components ranging from 0 to 255, but rather transformed into measured concentration data with physical superposition properties. Based on this linear superposition property, the system further performs color deconvolution processing on the concentration image using a preset staining feature matrix.

[0047] In its implementation, color deconvolution is essentially a linear demixing process. Since each pixel in the concentration image is actually a composite representation of the light absorbed by the nuclear components, cytoplasmic components, and residual background components, the system projects the mixed concentration signals into a mutually orthogonal biochemical space by performing an inverse matrix operation on the staining feature matrix. In this space, the image is separated into a biochemically layered image containing nuclear channel data and cytoplasmic channel data. Assuming that at a certain pixel, its mapped concentration in the red, green, and blue channels is respectively... The measured concentration data of the decoupled nuclear channels Measured concentration data of cytoplasmic channels The mapping logic can be described as follows:

[0048] ;

[0049] here, The inverse matrix of the predefined coloring feature matrix. This represents background residue during the unmixing process. In this way, the system achieves preliminary purification of the physical information of the occluded target, enabling the subsequent jurisdictional division module to directly locate the target based on the purified nuclear components. In practical applications, if the nuclear channel concentration in a certain region is significantly higher than the preset baseline absorption threshold, it indicates the presence of high-density chromatin distribution in that region, providing a precise basis of measured concentration data for subsequent central response calculations.

[0050] The jurisdiction division module is used to perform center response calculation on cell nuclear channel data, determine multiple cell center points, and divide the biochemical stratification image into multiple corresponding initial jurisdictions based on the cell center points;

[0051] After the biochemical stratification image is generated, the jurisdiction division module performs in-depth feature analysis on the measured concentration data to locate the physical location of cells and establish a preliminary spatial index.

[0052] Specifically, before performing the segmentation, this embodiment pre-constructs and configures a multi-scale Gaussian difference operator, which is commonly used in computer vision to extract blobular structures in images. This operator is chosen because cell nuclei in blood samples appear as concentration clusters with specific scales in the cell nucleus channel data of biochemical layered images. The multi-scale Gaussian difference operator can effectively suppress background texture interference and highlight the central response features of cell nuclei by simulating filtering processes at different spatial frequencies. Based on this, the system first acquires the measured concentration data of cell nucleus channel data contained in the biochemical layered image. And using two different spatial scale parameters and The Gaussian smoothing kernel performs multi-scale Gaussian difference operations. Specifically, for coordinate points... Its response value in the central response diagram The calculation process can be described as follows:

[0053] ;

[0054] in, and These represent the standard deviations respectively. and Gaussian smoothing operator, This represents the convolution operation. The resulting central response map, generated through this operation, can characterize the drastic changes in substance concentration and shows a distinct energy peak at the physical center of the cell nucleus.

[0055] Furthermore, the system generates a central response diagram. The system performs a local maximum search. In the actual processing flow, it compares the response value of each coordinate point and its neighborhood. If the response value of a point is the largest in its neighborhood and meets a pre-set response threshold, then that point is identified as the cell center point. It is worth noting that the response threshold is set based on a dynamic benchmark of measured concentration data to exclude small-scale fluctuations caused by dye precipitation or impurities. Through this step, the system identifies a set of cell center points containing N discrete coordinates. .

[0056] Based on the locked cell center points, the system further performs spatial region partitioning. This embodiment introduces the Thiessen polygon algorithm as the partitioning rule. Its core logic is to assign consecutive pixel planes to the nearest seed point by calculating the Euclidean distance, given a set of seed points. The Thiessen polygon algorithm is introduced because it mathematically ensures that each pixel has the strongest geometric correlation with its corresponding cell center point, thus providing a reasonable initialization boundary for subsequent unmixing logic. Specifically, for any pixel coordinates in the biochemical layered image... The system calculates its distance to the center point of all cells. The distance is calculated and a minimum determination is made to divide it into the corresponding initial jurisdiction. The initial jurisdiction index map generated by this process The following logical relationship must be satisfied:

[0057] ;

[0058] After this step, the pixel regions of the biochemical stratification image are divided into independent initial jurisdictions, each containing one and only one cell centroid. This structured region division not only provides a location reference for the subsequent construction of single-cell layers but also lays the logical foundation for handling the concentration allocation of overlapping areas. Finally, the jurisdiction division module outputs a data package containing the boundaries of each initial jurisdiction and the corresponding set of cell centroids, serving as the basic input for subsequent single-cell layer modeling steps.

[0059] The single-cell layer construction module is used to construct a corresponding simulated single-cell layer based on the distribution characteristics of the measured concentration data within each initial jurisdiction; the simulated single-cell layer contains simulated concentration data characterizing the cell under ideal non-overlapping conditions;

[0060] After completing the initial spatial division of the jurisdiction, the single-cell layer construction module establishes a mathematical model for each identified cell center point that satisfies the ideal non-overlapping state in terms of physical distribution, thereby providing a benchmark for subsequent overlapping deconstruction.

[0061] Before executing the specific modeling process, this embodiment pre-configures a nucleation modeling rule based on radial symmetry. This modeling rule is adopted because, from a cell biology perspective, the distribution of cell body material in most blood cells (such as red blood cells and various types of white blood cells) exhibits a radial decay pattern from the center to the edge in macroscopic imaging. By introducing a Gaussian kernel function with high smoothness, the independent material distribution morphology of a single cell without external compression interference can be effectively simulated, and the decay parameters can be adjusted to accommodate cell variants of different diameters or shapes.

[0062] In actual operation, the single-cell layer construction module first retrieves the initial jurisdiction index map and cell center point set output by the jurisdiction division module. Since the edge regions of the initial jurisdiction often physically correspond to overlapping areas between cells, their concentration values ​​are affected by multi-target coupling and cannot reflect the true properties of the individual cells. Therefore, within each initial jurisdiction, the system determines a confidence sampling area based on geometric constraints, using the corresponding cell center point as the origin. Specifically, the confidence sampling area is limited to a set of pixels that satisfy a preset radius constraint, meaning that the Euclidean distance from any pixel within the area to the cell center point is strictly less than the minimum geometric distance from that center point to the boundary of its initial jurisdiction. Through this geometric filtering mechanism, the system can avoid boundary regions with high overlap risk and extract the inherent physical feature information of each cell's core area.

[0063] Furthermore, the system acquires the measured concentration data of the nuclear channels and the cytoplasmic channels within the confidence sampling region, respectively. For each channel c of each cell i, the system calculates the average concentration within its region. The system then extracts the coordinates of the cell center point. and average concentration As input parameters, a mapping relationship from pixel coordinates to concentration values ​​is established using a preset radial basis function, and the corresponding simulated concentration data is calculated for each coordinate point in the initial jurisdiction.

[0064] In practice, the simulated concentration data is calculated using the following formula:

[0065] ;

[0066] in, Represents the pixel coordinates within the initial jurisdiction; This represents the simulated concentration data of the i-th simulated single-cell layer in channel c (which can be either the nuclear channel or the cytoplasmic channel); A pre-defined morphological attenuation factor is used to control the degree of shrinkage at the edge of the cell model. Assuming that for a neutrophil, the mean concentration of its nuclear channels in the confidence sampling region is... The morphological attenuation factor is 0.85. When set to 12.0, the system will generate a concentration layer that smoothly decays from the center outwards throughout the entire initial jurisdiction based on the above formula.

[0067] Finally, by calculating the nucleus and cytoplasm channels separately, a simulated single-cell layer containing information from both channels was generated. This layer not only numerically preserves the biochemical concentration characteristics of the single cell but also establishes an ideal physical reference in terms of spatial distribution. This output provides necessary data support for the subsequent difference calculation module to reverse-correct and restore cell texture by comparing the synthesized image with the measured image.

[0068] The difference calculation module is used to linearly overlay the simulated concentration data of all simulated single-cell layers to generate a synthetic image, and calculate the difference map between the synthetic image and the biochemical stratified image; the difference map contains the concentration difference data between the synthetic image and the biochemical stratified image at each pixel position.

[0069] After the single-cell layer is constructed, the difference calculation module further quantitatively evaluates the deviation between the idealized distribution generated by the mathematical model and the physical real distribution captured by the sensor, and generates feedback signals to guide the subsequent layer recovery.

[0070] Before executing the specific overlay and comparison logic, this embodiment pre-configures a linear concentration overlay rule, which is based on the physical optics fundamentals followed by the image preprocessing module of this invention. Since the transmittance signal has been converted into measured concentration data at the absorbance level using Lambert-Beer's law during the preprocessing stage, according to the principle of physical additivity of absorbance, when multiple overlapping targets exist in space, the total concentration value generated at the imaging point should be strictly equal to the linear algebraic sum of the independent concentration values ​​of each target. Based on this, the difference calculation module determines the overlay synthesis logic as a linear accumulation operation, aiming to simulate the theoretical synthesized value that the imaging system should detect if all current simulated single-cell layers are physically real distributions. This logic construction eliminates nonlinear interference and ensures the physical accuracy of subsequent difference quantization.

[0071] During actual system operation, the difference calculation module first obtains all simulated single-cell layers output by the single-cell layer construction module. These layers are presented as a set. The data exists in the form of [database name], and each layer contains simulated concentration data for nuclear and cytoplasmic channels. Specifically, for any coordinate point in the image... The difference calculation module performs linear superposition processing on the nuclear and cytoplasmic channels respectively, generating corresponding synthetic images. The synthetic image displays the synthetic concentration value in a specific channel c (representing either the nuclear or cytoplasmic channel). The calculation formula is as follows:

[0072] ;

[0073] in, This represents the simulated concentration data of the i-th simulated single-cell layer at the corresponding coordinates and channels. Through this step, the system generates a multi-channel composite image that is perfectly aligned with the original biochemical stratification image in terms of physical dimensions.

[0074] Based on the generated synthetic image, the difference calculation module further calculates the difference map between it and the biochemical stratification image. In specific implementation, the system uses the measured concentration data representing the distribution of cellular substances in the biochemical stratification image as a baseline value, subtracts the theoretical synthetic value from the synthetic image, and thus extracts the concentration difference data at each pixel location. For coordinate points... Concentration difference data in channel c The calculation process can be described as follows:

[0075] ;

[0076] in, This represents the measured concentration data of the biochemical stratified image at the corresponding location and channel. It's worth noting that the generated difference map contains residual information between the synthetic image and the biochemical stratified image at each pixel location. For example, a positive concentration difference at a certain coordinate point indicates that the current simulated single-cell layer set has not fully explained the physical concentration at that location, suggesting the possible existence of overlooked cell textures or overlapping edges; conversely, a negative value indicates that the simulation model overestimates at the boundaries.

[0077] Finally, the difference calculation module outputs a difference map containing concentration differences in nuclear and cytoplasmic channels. This difference map not only quantifies the inconsistency between the current model and the physical reality, but also serves as a feedback field carrying spatial location information, providing a precise driving force for the subsequent layer restoration module to use this concentration difference data to reverse-correct the simulated concentration data. Through this data flow based on synthetic alignment, the system achieves the transition from idealized modeling to realistic pathological texture restoration.

[0078] The layer restoration module is used to reverse correct the simulated concentration data of each simulated single-cell layer using concentration difference data until the difference map meets the preset convergence condition, thereby obtaining an independent single-cell layer containing restored texture.

[0079] After the difference map is generated, the layer restoration module uses an iterative optimization method to feed back the complex texture information and overlap correction amount contained in the difference map into the initially established mathematical model, thereby achieving the approximation from the idealized model to the real physical morphology of cells.

[0080] This embodiment pre-configures an iterative correction algorithm based on variational inference. The algorithm is designed to utilize the residuals from physical imaging as a driving force to propel the morphological evolution of the simulated single-cell layer. Notably, a smoothing operator is pre-defined in the system. With update step size parameter The reason for introducing the smoothing operator is... This is because, in the initial state, the concentration value of the simulated single-cell layer may be zero at certain coordinate points (such as outside the initial boundary of the jurisdiction). Without this operator, the logic of allocating differences based on concentration ratios would result in the allocation weights for these regions always being zero, preventing the cell model from growing to the initially defined geometric boundaries. By setting a very small smoothing operator (in this embodiment, it can be set according to the noise baseline), this problem is mitigated. (At the order of magnitude), the system allows concentration difference data to be injected into regions where the original value was zero, thereby achieving dynamic recovery of irregular cell edges or pseudopodia.

[0081] In a single iteration of the system's actual operation, the layer restoration module retrieves the difference map output by the difference calculation module and performs numerical compensation for the simulated concentration data in the nucleus and cytoplasm channels for each simulated single-cell layer. Specifically, the system first calculates the coordinates of each pixel. Spatial allocation weight This weight determines how the total concentration difference data at the current coordinate point should be allocated to the cells participating in the overlap. The calculation process for the spatial allocation weight follows the formula below:

[0082] ;

[0083] Where N is the total number of the initial jurisdiction, that is, the total number of cells in the field of view participating in the calculation; This represents the simulated concentration data of the i-th simulated single-cell layer in channel c (nucleus or cytoplasm channel) at the t-th iteration. This formula reflects that the higher the simulated concentration at the current point or the closer it is to the cell center, the greater the proportion of differential compensation it receives.

[0084] Based on the calculated spatial allocation weights, the system updates the simulated concentration data for each layer according to the following iterative correction formula:

[0085] ;

[0086] Where t is the index of the current iteration number. The preset update step size parameter is used to control the correction magnitude of each iteration to ensure algorithm convergence; This represents the concentration difference data in channel c of the difference map. The formula contains... The function performs a maximum value operation, which physically ensures that the corrected simulated concentration data is always non-negative, consistent with the physical property of the substance's absorbance.

[0087] Taking a specific overlapping pixel as an example, suppose the difference in nuclear channel concentration at this point in the difference map is 0.12, indicating that the model underestimates the true concentration at this point. If this point involves two overlapping cells, where the current simulated concentration of cell A is significantly higher than that of cell B, then according to the weight allocation logic described above, cell A will receive the majority of this 0.12 difference value. After one iterative correction, the simulated concentration data of cell A at this point will be increased, thus making its texture features closer to the measured concentration distribution in the biochemical stratification image.

[0088] The above correction process is repeated until the difference map meets the preset convergence conditions, such as the root mean square error of the differences across the entire map being less than a set threshold, or reaching the preset maximum number of iterations. At this point, the layer after multiple physical feedback corrections is no longer a smooth geometric model, but an independent single-cell layer containing rich pathological details and realistic morphological features. This output completely solves the information aliasing problem caused by physical cell overlap, providing a data foundation for subsequent extraction of high-purity cell features and accurate deep learning classification.

[0089] The feature construction and classification module is used to generate an overlap complexity weight map based on the difference map that meets the preset convergence conditions, and to construct a multi-dimensional feature data package for each cell, which includes an independent single-cell layer and an overlap complexity weight map; the multi-dimensional feature data package is input into a pre-trained convolutional neural network model, and the class classification results of each cell are output;

[0090] After the layer restoration module completes the iterative approximation of the cell's physical morphology, the feature construction and classification module structurally integrates the independent signals obtained from the physical layer decoupling, the original visual texture, and the uncertainty quantification indicators in the decoupling process to construct a multidimensional tensor containing rich physical semantics, and uses a pre-trained deep learning model to achieve accurate determination of cell type.

[0091] Before performing specific feature construction, this embodiment pre-builds and configures a convolutional neural network model. This model is not a general image classification network, but a deep architecture customized for the four-channel input data of this invention. Specifically, the model's input layer is configured to receive data in a dimension of... Tensor data, in which For the preset cutting size (e.g.) (pixels). The model's hidden layers employ an alternating stacking of multi-layer convolutional layers and pooling layers to extract semantic features at different levels. Specifically, a spatial attention module is designed within the hidden layers. This module utilizes the fourth channel data of the input tensor as a weight mask to weighted suppress the preceding convolutional feature maps, thereby reducing the interference of high-uncertainty regions on classification decisions. In this embodiment, this mechanism is not implemented through adaptive learning to generate the mask, but rather by directly introducing prior knowledge from the physical layer. Specifically, assuming the... The feature map output by the convolutional layer is The corresponding downsampled fourth channel weight data is The feature map after attention weighting The calculation formula is:

[0092] ;

[0093] in, This represents element-wise multiplication. This is the suppression coefficient (e.g., 0.8). The function maps weights to a non-linear region. This computational logic forces the network to reduce highly overlapping complexity regions (i.e., ...) during feature extraction. The activation response of areas with higher values ​​is used to achieve the avoidance of key points in physical guidance.

[0094] The model's output layer is set to a fully connected layer connected to a Softmax activation function, with the number of output nodes corresponding to the total number of cell types to be classified (e.g., neutrophils, lymphocytes, monocytes, etc.). Before system deployment, the model had already undergone parameter training on a training set containing a large number of labeled samples. To accommodate the rotational invariance of blood cells, online data augmentation strategies were employed during training, including random rotation (0-360 degrees) and horizontal flipping. The model was trained using the Adam optimizer, with an initial learning rate set to [value missing]. The learning rate is decayed using a cosine annealing strategy. The loss function employs weighted cross-entropy loss to balance the sample imbalance between different cell types (such as fewer eosinophils and more neutrophils).

[0095] During actual system operation, the feature construction and classification module first generates an overlap complexity weight map based on the final state output by the layer restoration module. Since some concentration difference data (usually corresponding to extremely complex overlap regions or imaging artifacts) may remain in the difference map during iterative convergence that cannot be interpreted by the model, the system utilizes this residual data to decouple uncertainties. Specifically, the system acquires the concentration difference data of cell nuclear channels when preset convergence conditions are met. Concentration difference data with cytoplasmic channels And calculate the magnitude at each pixel position. Overlap complexity weight map The formula for generating it is as follows:

[0096] ;

[0097] in, This is a normalization function used to map the modulus value to the [0,1] interval. The generated weight map numerically reflects the deconstruction error of the corresponding pixel region; a higher value indicates a lower reliability of the physical reconstruction of that region.

[0098] Furthermore, to address the asymmetric deformation of cells caused by physical compression in high-density overlapping regions, the system performs adaptive cell morphology correction processing after generating the overlap complexity weight map and before constructing the multidimensional feature data package. In actual processing, the system first uses the center point of each cell... Extract the overlapping complexity weight graph with the origin as the origin. In the preset neighborhood (e.g. The spatial gradient distribution characteristics within a pixel window. Specifically, the system uses gradient operators (such as the Sobel operator) to calculate the gradient vector of each pixel in the neighborhood. The direction of the gradient vector indicates the direction in which the overlap complexity increases the fastest, and physically corresponds to the direction from which the adjacent cells squeeze the current cell.

[0099] Furthermore, the system calculates the numerical integral intensity of the spatial gradient distribution characteristics in different radial directions. The system divides the neighborhood space into K discrete radial sectors (e.g., K=8, covering 0 to 1). (Angle), and calculate the weighted sum of the magnitudes of the gradient vectors falling within each sector. The specific calculation logic is as follows:

[0100] ;

[0101] in, This represents the set of pixels within the k-th radial sector. For gradient magnitude, It is a distance weight function (the closer to the center, the greater the weight).

[0102] Based on this, the system determines the asymmetric compression vector experienced by the corresponding cell. The vector at a specific angle The larger the modulus length, the higher the structural uncertainty of the cell in that direction, that is, the stronger the physical deformation pressure.

[0103] Subsequently, the system utilizes the asymmetric extrusion vector to apply a preset morphological attenuation factor. Directional modulation is performed to obtain a dynamic attenuation coefficient that varies with the radial direction. The modulation process follows a negative correlation mapping rule, meaning that the greater the squeezing pressure, the smaller the attenuation coefficient (indicating more severe edge contraction). The calculation formula is as follows:

[0104] ;

[0105] in, The initial global morphological decay factor. The preset modulation sensitivity coefficient, This is the normalized value of the compressive strength in this radial direction.

[0106] Based on the calculated dynamic attenuation coefficient The system recalculates the simulated concentration data corresponding to each coordinate point within the initial jurisdiction to reconstruct the simulated single-cell layer. The reconstructed simulated concentration data... The calculation logic is updated as follows:

[0107] ;

[0108] in Through this step, the cell model, which was originally isotropic (circular), is corrected to an anisotropic distribution (such as elliptical or irregular polygonal) that conforms to the actual physical compression state.

[0109] Ultimately, the system will reconstruct the simulated single-cell layer. As the input source for subsequent multidimensional feature data package construction steps, to replace the initially generated simulated single-cell layer. This replacement operation ensures that the first and second channels of the input convolutional neural network contain accurate morphological features corrected by physical feedback, thereby significantly improving the model's robustness in recognizing aberrant red blood cells or crushed white blood cells.

[0110] It is worth noting that, in order to preserve the texture details acquired by the layer restoration module during the iteration process, the system linearly superimposes the cumulative texture residuals calculated in the previous steps (i.e., the difference between the simulated layer and the pure Gaussian basis at the end of the iteration) onto the reconstructed simulated single-cell layer. Therefore, the independent single-cell layer that ultimately serves as the input source for the multidimensional feature data package actually incorporates the accurately physical contours that have been adaptively corrected (by...). The system provides both real-world cell texture features (provided by accumulated residuals) and real-world cell texture features, thus ensuring the completeness of the input data.

[0111] Then, the system constructs a multidimensional feature data package, using each determined cell center point. Using the geometric center, cut off preset dimensions on the corresponding data source planes (e.g., ...). Image data blocks (pixels). To fuse physical and visual features, the system performs tensor stacking processing on the extracted data blocks in a preset order, generating a four-dimensional tensor containing four feature channels. Specifically:

[0112] The first channel is taken from the simulated concentration data of the independent single-cell layer under the cell nucleus channel output by the layer restoration module. This data provides the pure nucleus morphology after removing the occlusion.

[0113] The second channel is taken from simulated concentration data of an independent single-cell layer under the cytoplasm channel, which reflects the complete distribution of the cytoplasm.

[0114] The third channel originates from the raw acquired image obtained by the image preprocessing module. The system performs color space conversion on the raw color image (e.g., ...). The grayscale mapping data is obtained, and the corresponding data blocks are extracted to preserve the original visual texture details of the cells (such as nucleoli, granules, etc.).

[0115] The fourth channel takes the weight data at the corresponding position from the overlapping complexity weight map generated above, and serves as a guiding signal for the decoupling confidence.

[0116] Ultimately, the system will construct the multidimensional feature data package. The input is fed into a pre-trained convolutional neural network model. During forward propagation, the model uses the physical features of the first and second channels for morphological analysis, the third channel to supplement texture details, and automatically adjusts the region of interest for feature extraction based on the weights of the fourth channel. The model's output layer outputs a probability vector. ,in This represents the probability that the cell belongs to the k-th category. By comparing the values ​​in the probability vector, the system determines the category with the highest probability as the classification result for each cell, thus completing the entire intelligent processing process from complex overlapping images to accurate cell classification.

[0117] In summary, this embodiment transforms the physical phenomenon of cell overlap into computable concentration difference data, and uses this difference data as a feedback signal to drive the simulation model for nonlinear adaptive correction and texture backfilling. The system successfully decouples the originally spatially overlapping complex light signals into single-cell layers with independent biological significance. This deconstruction based on physical mechanisms not only eliminates the interference of visual occlusion on texture features, but also provides clear confidence guidance for subsequent classification through the overlap complexity weight map, thereby ensuring that the convolutional neural network can focus on the restored real pathological features. Ultimately, it fundamentally solves the problem of decreased recognition accuracy caused by physical stacking in high-density blood samples, and significantly improves the clinical application value of the system.

[0118] Example 2:

[0119] like Figures 3-4 As shown, the intelligent recognition and classification method for blood sample images based on computer vision includes the following steps:

[0120] The raw images of blood samples are acquired, illumination correction and nonlinear concentration mapping are performed on the raw images, and the mapped images are separated into biochemically layered images containing nuclear channel data and cytoplasmic channel data; the biochemically layered images contain measured concentration data characterizing the distribution of cellular substances;

[0121] The center response calculation is performed on the cell nuclear channel data to determine multiple cell center points, and the biochemical stratification image is divided into multiple corresponding initial jurisdictions based on the cell center points;

[0122] Based on the distribution characteristics of measured concentration data within each initial jurisdiction, a corresponding simulated single-cell layer is constructed; the simulated single-cell layer contains simulated concentration data characterizing the cell under ideal non-overlapping conditions;

[0123] The simulated concentration data of all simulated single-cell layers are linearly superimposed to generate a synthetic image, and the difference map between the synthetic image and the biochemical stratified image is calculated; the difference map contains the concentration difference data between the synthetic image and the biochemical stratified image at each pixel location.

[0124] The simulated concentration data of each simulated single-cell layer is reverse-corrected using the concentration difference data until the difference map meets the preset convergence condition, thereby obtaining an independent single-cell layer containing restored texture.

[0125] An overlap complexity weight map is generated based on the difference map that meets the preset convergence conditions, and a multi-dimensional feature data package containing an independent single-cell layer and an overlap complexity weight map is constructed for each cell.

[0126] The multidimensional feature data package is input into a pre-trained convolutional neural network model, which outputs the classification results of each cell.

[0127] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

[0128] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0129] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A computer vision-based intelligent recognition and classification system for blood sample images, characterized in that, include: The image preprocessing module is used to acquire the original acquisition image of the blood sample, perform illumination correction and nonlinear concentration mapping on the original acquisition image, and separate the mapped image into a biochemically layered image containing cell nuclear channel data and cytoplasmic channel data. The biochemical stratification image contains measured concentration data characterizing the distribution of cellular substances; The jurisdiction division module is used to perform center response calculation on the cell nuclear channel data, determine multiple cell center points, and divide the biochemical stratification image into multiple corresponding initial jurisdictions based on the cell center points; The single-cell layer construction module is used to construct a corresponding simulated single-cell layer based on the distribution characteristics of the measured concentration data within each initial jurisdiction; the simulated single-cell layer contains simulated concentration data characterizing the cell in an ideal non-overlapping state; The difference calculation module is used to linearly overlay the simulated concentration data of all simulated single-cell layers to generate a synthetic image, and calculate the difference map between the synthetic image and the biochemical stratified image; the difference map contains the concentration difference data between the synthetic image and the biochemical stratified image at each pixel position. The layer restoration module is used to reverse correct the simulated concentration data of each simulated single-cell layer using the concentration difference data until the difference map meets the preset convergence condition, thereby obtaining an independent single-cell layer containing restored texture. The feature construction and classification module is used to generate an overlap complexity weight map based on the difference map that satisfies the preset convergence condition, and to construct a multi-dimensional feature data package for each cell that includes the independent single-cell layer and the overlap complexity weight map; The multidimensional feature data package is input into a pre-trained convolutional neural network model, which outputs the classification results of each cell.

2. The intelligent recognition and classification system for blood sample images based on computer vision according to claim 1, characterized in that: The process of generating the biochemical stratified image includes: Obtain a pre-calibrated blank background image; The original acquired image is compared pixel by pixel with the blank background image to obtain a transmittance map. A concentration image is generated by performing a negative logarithmic transformation on the transmittance map based on the Lambert-Beer law. The concentration image is subjected to color deconvolution processing using a preset staining feature matrix to decouple and obtain the measured concentration data of the nuclear channel and the measured concentration data of the cytoplasmic channel.

3. The intelligent recognition and classification system for blood sample images based on computer vision according to claim 1, characterized in that: The initial jurisdictional division process includes: Perform multi-scale Gaussian difference operation on the cell nuclear channel data to generate a central response map characterizing the degree of drastic change in substance concentration; Local maxima are extracted from the central response map, and the local maxima that satisfy a preset response threshold are determined as the cell center point. Using the cell center point as the seed point, a Thiessen polygon is generated, dividing the pixel region of the biochemical layered image into mutually independent initial jurisdictions.

4. The intelligent recognition and classification system for blood sample images based on computer vision according to claim 1, characterized in that: The process of generating the simulated single-cell layer includes: Within each initial jurisdiction, with the corresponding cell center point as the origin, the set of pixels that satisfy the preset radius constraint is determined as the confidence sampling area; The measured concentration data of the nuclear channel and the measured concentration data of the cytoplasmic channel within the confidence sampling area were obtained respectively, and the average concentration of each channel was calculated. Using the coordinates of the cell center point and the average concentration as input parameters, a mapping relationship from pixel coordinates to concentration values ​​is established using a preset radial basis function. The simulated concentration data corresponding to each coordinate point within the initial jurisdiction is calculated to generate a simulated single-cell layer containing dual-channel information. The formula for calculating the simulated concentration data is: ； in, These are the pixel coordinates within the initial jurisdiction. Represents a nuclear channel or a cytoplasmic channel. For the i-th simulated single-cell layer, the simulated concentration data is in channel c. Let these be the coordinates of the center point of the i-th cell; The mean concentration of the measured concentration data in channel c is the confidence sampling area. This is the preset morphological attenuation factor.

5. The intelligent recognition and classification system for blood sample images based on computer vision according to claim 4, characterized in that: The concentration difference data is used to reverse-correct the simulated concentration data of each simulated single-cell layer until the difference map meets a preset convergence condition, specifically including: In a single iteration loop, based on the concentration difference data of each coordinate point in the difference map, combined with the preset step size parameter and the assigned weight, numerical compensation is performed on the simulated concentration data of each simulated single cell layer in the nucleus channel and cytoplasm channel to update the simulated single cell layer. The update process follows the iterative correction formula: ； Where t is the index of the current iteration number. This represents the simulated concentration data of the i-th simulated single-cell layer in channel c during the t-th iteration. This is the concentration difference data in channel c of the difference map. The preset update step size parameter, Spatial weights are assigned to control the proportion of concentration difference data. It is a function for maximizing the value; The spatial allocation weight The determination process includes: obtaining pixel coordinates respectively. At the t-th iteration, the simulated concentration data of each simulated single-cell layer is calculated, and the pixel coordinates of the i-th simulated single-cell layer are calculated. The concentration percentage at each location is determined; the concentration percentage is adjusted using a preset smoothing operator to generate the spatial allocation weight; the calculation formula for the spatial allocation weight is as follows: ； Where N is the total number of initial jurisdictions. This is a preset smoothing operator.

6. The intelligent recognition and classification system for blood sample images based on computer vision according to claim 1, characterized in that: The process of generating the overlapping complexity weight graph includes: Obtain the concentration difference data corresponding to the nuclear channel and the cytoplasmic channel when the preset convergence condition is met; Calculate the modulus of the concentration difference data of the nuclear channel and cytoplasmic channel at each pixel location to generate an initial weight map characterizing the distribution of local deconstruction error; The initial weight map is normalized to obtain the overlap complexity weight map, which is used to quantify the decoupling uncertainty of each pixel region.

7. The intelligent recognition and classification system for blood sample images based on computer vision according to claim 1, characterized in that: The construction process of the multidimensional feature data packet includes: Centered on the center point of each cell, image data blocks of a preset size are extracted at the corresponding spatial positions of the independent single-cell layer, the original acquired image, and the overlapping complexity weight map. The captured image data blocks are processed by tensor stacking in a preset order to generate a four-dimensional tensor containing four feature channels, which serves as the multi-dimensional feature data packet; the four feature channels include: The first channel is the simulated concentration data of the independent single-cell layer under the cell nucleus channel, used to provide the distribution characteristics of single-cell nucleus material after decoupling; The second channel is the simulated concentration data of the independent single-cell layer under the cytoplasm channel, used to provide the distribution characteristics of single-cell cytoplasm after decoupling; The third channel is grayscale mapping data obtained by performing color space conversion on the original acquired image at the corresponding position, which is used to provide the original visual texture features of the cell; The fourth channel contains the weight data of the overlapping complexity weight graph at the corresponding position, which is used to provide decoupling confidence guidance.

8. The intelligent recognition and classification system for blood sample images based on computer vision according to claim 4, characterized in that: After generating the overlapping complexity weight map and before constructing the multidimensional feature data package, an adaptive correction process for cell morphology is also performed: Extract the spatial gradient distribution features of the overlapping complexity weight map within a preset neighborhood of each cell center point; Calculate the numerical integral intensity of the spatial gradient distribution characteristics in different radial directions; The asymmetric compression vector of the corresponding cell is determined based on the numerical integral intensity. The asymmetric compression vector is used to characterize the physical deformation pressure of the cell in the overlapping region.

9. The intelligent recognition and classification system for blood sample images based on computer vision according to claim 8, characterized in that: The adaptive correction process also includes: The asymmetric compression vector is used to directionally modulate the morphological attenuation factor to obtain a dynamic attenuation coefficient that varies with the radial direction. Based on the dynamic attenuation coefficient, the simulated concentration data corresponding to each coordinate point is recalculated in order to reconstruct the simulated single-cell layer; The reconstructed simulated single-cell layer is used as the input source for the subsequent multidimensional feature data package construction step, replacing the initially generated simulated single-cell layer.

10. A computer vision-based intelligent recognition and classification method for blood sample images, characterized in that, Includes the following steps: The original acquisition image of the blood sample is acquired, illumination correction and nonlinear concentration mapping are performed on the original acquisition image, and the mapped image is separated into a biochemically layered image containing nuclear channel data and cytoplasmic channel data. The biochemical stratification image contains measured concentration data characterizing the distribution of cellular substances; The center response calculation is performed on the cell nuclear channel data to determine multiple cell center points, and the biochemical stratification image is divided into multiple corresponding initial jurisdictions based on the cell center points; Based on the distribution characteristics of measured concentration data within each initial jurisdiction, a corresponding simulated single-cell layer is constructed; the simulated single-cell layer contains simulated concentration data characterizing the cell under an ideal non-overlapping state; The simulated concentration data of all simulated single-cell layers are linearly superimposed to generate a synthetic image, and a difference map is calculated between the synthetic image and the biochemical stratified image; the difference map contains the concentration difference data between the synthetic image and the biochemical stratified image at each pixel position. The concentration difference data is used to reverse correct the simulated concentration data of each simulated single-cell layer until the difference map meets the preset convergence condition, thereby obtaining an independent single-cell layer containing restored texture. An overlap complexity weight map is generated based on the difference map that satisfies the preset convergence condition, and a multidimensional feature data package containing the independent single-cell layer and the overlap complexity weight map is constructed for each cell; The multidimensional feature data package is input into a pre-trained convolutional neural network model, which outputs the classification results of each cell.

Images