Bone marrow smear high-power microscope intelligent scanning method

Abstract

The invention discloses a high-power microscope intelligent scanning method for a bone marrow smear, which comprises the following steps: S1, automatically dividing a scanning area: acquiring a bone marrow smear image by adopting a front camera, and selecting an optimal scanning area through a multi-dimensional quality evaluation model; s2, intelligent partitioning and focusing point distribution: dividing an optimal scanning area into N * M sub-areas, determining P focusing sampling points by adopting a layered sampling strategy, and establishing a three-dimensional coordinate mapping table; s3, multi-point focusing data acquisition: according to a preset coordinate mapping table, driving an objective table and an objective lens assembly to move to each sampling point in sequence, and constructing an N * M * P-dimensional focusing data matrix; s4, performing focal plane fitting by a least square method: performing three-dimensional plane fitting on the acquired data; and S5, intelligent scanning imaging: calculating the theoretical optimal focal length of each position in the scanning path according to the fitting function, and driving the objective lens assembly to adjust the focal length in real time for scanning imaging. According to the invention, the scanning area identification accuracy and the image definition uniformity are automatically improved in the whole process.

Description

Technical Field

[0001] The present invention relates to the technical field of microscope imaging, and in particular to a high-power microscope intelligent scanning method for bone marrow smears based on automatic area division and multi-point focusing. Background Art

[0002] Bone marrow cytology is a key and fundamental tool for diagnosing hematologic diseases and is widely used clinically for laboratory testing and therapeutic evaluation of hematologic and related diseases. Bone marrow cells are the core component of the hematopoietic system. Bone marrow cells include hematopoietic stem cells, progenitor cells of various lineages, immature cells, and mature cells at various stages of differentiation. A bone marrow smear primarily contains the following cell types: erythroid cells, granulocytes, lymphocytes, monocytes, megakaryocytes, plasma cells, and non-hematopoietic cells. The number, proportion, morphology, and maturity distribution of various cell lineages in the bone marrow are important indicators for assessing hematopoietic function. Cell classification and enumeration based on cell morphology analysis facilitate the diagnosis and differential diagnosis of various hematologic diseases and provide clues for the diagnosis of difficult and complex conditions. Among hematologic malignancies, acute leukemias are characterized by an increased proportion of primitive and immature cells with abnormal morphology, chronic leukemias are characterized by abnormal proliferation of various lineages and an increase in mature cells, and myelodysplastic syndromes (MDS) present with bone marrow clonality and dysplastic hematopoiesis. In hematopoietic dysfunction, the bone marrow may show a significant decrease in the number of one or more hematopoietic lineages. Secondary bone marrow lesions include extramedullary tumor cell infiltration caused by metastatic tumors, inflammatory cell infiltration caused by infectious diseases, and characteristic bone marrow cells caused by metabolic diseases. Bone marrow hematopoietic assessment provides critical morphological information for accurate disease diagnosis, personalized treatment, and efficacy observation.

[0003] In clinical practice, clinical hematology physicians use optical microscopes to read and diagnose bone marrow smears. They usually need to complete the classification and counting of 200-500 nucleated cells, as well as the counting of special cells such as megakaryocytes. The general operating procedures for laboratory microscope reading are as follows: first observe the overall condition of the smear under a low-power microscope (10×), evaluate the quality of the sample and the proliferation of bone marrow cells, count the megakaryocytes, observe for abnormal cells or cell clusters, browse the entire slide and select areas with uniform cell distribution for the next step of nucleated cell classification and counting; then, perform cell lineage identification, classification and counting under a high-power microscope (100×) to provide quantitative and qualitative basis for disease diagnosis. The technical problems faced by manual microscope detection are as follows: 1. The identification of scanning areas for classification and counting relies on manual experience and judgment, which is highly subjective. There is a lack of objective and unified regional quality assessment standards, resulting in differences in the representativeness of the scanning areas selected by different operators, affecting the repeatability and accuracy of the test results.

[0004] 2. Traditional single-point focus or fixed focus point distribution cannot adapt to the thickness variations of bone marrow smear samples. Within the thickness fluctuation range of 5-15μm, the problem of uneven image clarity is prominent, which directly affects the recognition and accurate interpretation of key morphological details such as cell nuclear-cytoplasmic ratio and particle distribution.

[0005] 3. Uneven thickness distribution during sample preparation and slight slide tilt (usually within ±1°) can cause significant focus shift when scanning a large field of view. This can lead to significant differences in imaging quality across different regions within the same field of view, and blurred images in some areas can affect cell classification accuracy.

[0006] 4. In the digitization of microscope images, the entire scanning process is time-consuming. The automation integration from sample loading, area selection, focus adjustment to image acquisition is not high, which cannot meet the efficiency requirements of modern clinical testing for high-throughput and rapid testing, restricting the overall work efficiency of the testing department and the diagnostic effectiveness of blood system diseases.

[0007] The existing patent application number CN202110584308 is titled "An automatic leveling and autofocusing method for bone marrow smear scanning." The intelligent analysis technology used in bone marrow cytology diagnosis relies on the electronic collection of bone marrow smears. Its microscope scanning technology has the following technical defects: 1. Inadequate scanning leveling accuracy and inefficiency: The aforementioned technology performs leveling based solely on the distance measurement data of the left and right boundaries of the selected area and verification using a single random point, lacking a holistic assessment of the entire smear plane. When the smear exhibits local tilt or complex unevenness (such as a bulge in the middle area or warped edges), the leveling method based on a limited number of sampling points cannot accurately reflect the true planar state of the smear, resulting in incomplete leveling. Furthermore, when the random point distance measurement values ​​are inconsistent with the boundary data, the leveling process must be repeated multiple times, significantly increasing time costs. This makes leveling inefficient, especially when the smear has high unevenness.

[0008] 2. The scanning focus strategy is single and has poor adaptability: The above technology uses a single distance reference focusing method, with the distance measurement value of the first cell as the reference distance, assuming that the surface height of all cells in the scanning area is consistent. However, due to the unevenness of the preparation process and cell distribution, the actual thickness of bone marrow smears usually fluctuates by 5-50μm. The single distance reference cannot adapt to local height changes, resulting in out-of-focus blur in cell images in some areas. At the same time, this method lacks direct evaluation of image clarity and relies solely on distance consistency to judge the focus state. Focus deviation is prone to occur under complex sample conditions such as cell overlap, uneven staining, or background interference.

[0009] 3. High Hardware Dependence and Cost: The aforementioned technology places extremely high demands on hardware precision, requiring a high-precision laser rangefinder and a precision robotic arm (with a Z-axis step length of up to 25nm). These requirements place stringent demands on the equipment's resolution, stability, and calibration accuracy, significantly increasing equipment cost and maintenance complexity. Furthermore, the laser rangefinder's accuracy is susceptible to factors such as the smear surface reflectivity and the optical properties of the dye material. Measurement errors can occur between samples, impacting the consistency and reliability of leveling and focusing. Summary of the Invention

[0010] The present invention addresses the key technical difficulties in existing bone marrow smear microscopy scanning, namely, insufficient scanning leveling accuracy, a single focusing strategy with strong hardware dependence, and low scanning efficiency. A high-magnification microsmear intelligent scanning method based on automatic area division and multi-point focusing technology is proposed.

[0011] The technical solution adopted by the present invention to solve its technical problem is: A bone marrow smear high-power microscope intelligent scanning method comprises the following steps: Step S1. Automatically divide the scanning area: Use the front camera to obtain the full image of the bone marrow smear, and use edge detection and region segmentation algorithms to automatically identify the effective boundaries of the smear, excluding interference areas such as the slide edge, labels, and bubbles.

[0012] A multidimensional quality assessment model was established to comprehensively analyze the parameters of background uniformity and staining quality, and automatically select the optimal scanning area with uniform cell distribution, low overlap and clear staining, so as to improve the accuracy and repeatability of bone marrow cytology diagnosis.

[0013] Furthermore, the multidimensional quality assessment model works collaboratively through four core modules: a cell density assessment module, an overlap degree quantification module, a background uniformity assessment module, and a staining quality assessment module.

[0014] The cell density assessment module screens for areas of moderate density; the overlap quantification module uses morphological analysis to ensure low overlap; the background uniformity assessment module analyzes grayscale distribution to exclude areas of uneven staining; and the staining quality assessment module analyzes Wright's staining using the HSV color model. A weighted scoring mechanism is established, automatically selecting the scan area with the highest overall score through normalization.

[0015] Step S2. Intelligent zoning and focus point distribution: The optimal scanning area determined in step S1 is intelligently divided into N×M equal sub-regions. A layered sampling strategy is used in each sub-region to determine P focus sampling points. The system automatically establishes a three-dimensional coordinate mapping table of the focus points, records the X, Y plane coordinates and estimated Z-axis focal length information of each sampling point, and ensures high-precision focus coverage of the entire area.

[0016] Further, the P focus sampling points adopt a multi-level distribution strategy: including 1 center point to ensure regional representation, 4 corner points to cover the boundary focus change, and several randomly distributed points to supplement local thickness fluctuation information, and the uniformity and coverage integrity of the sampling points are ensured through a spatial distribution optimization algorithm, thereby improving the focus accuracy and stability of subsequent image acquisition.

[0017] Step 3. Multi-point focus data acquisition: according to the three-dimensional coordinate mapping table of the focus points preset in step S2, the objective table and the objective lens assembly are driven to move to each sampling point position in turn, and the best focus value is quickly obtained at each position through the automatic focusing model, while the corresponding three-dimensional coordinate information (x, y, z) is recorded, wherein the z value represents the best focus position of the point. All the collected focus data are constructed into an NXMXP-dimensional focus data matrix, which provides complete spatial reference data for subsequent focal plane interpolation calculation and real-time focus adjustment.

[0018] Further, the automatic focusing model adopts a CNN binary classification architecture, constructs a convolutional neural network, inputs the current focus image block, automatically extracts key features such as cell edge sharpness and texture details through multi-layer convolution and pooling operations, and performs binary classification in the output layer to determine the focus or defocus state.

[0019] The model is trained using a large amount of labeled data of bone marrow smears at different focal distances, learns the best focus feature representation, realizes fast and accurate focus quality evaluation, and the single inference time is controlled within 50 ms, meeting the real-time focus efficiency requirement.

[0020] Step S4. Least squares focal plane fitting: the least squares method is used to fit all the focus point data collected in step S3, and based on the focal plane function obtained by fitting, the system can calculate the theoretical best focus value of any position (x, y) in the scanning area, and generate a continuous and smooth focal plane distribution model.

[0021] Further, the three-dimensional plane fitting establishes a focal plane equation z=ax2+by2+cxy+dx+ey+f: the fitting parameters are solved by the least squares optimization algorithm, wherein a and b are the quadratic curvature coefficients in the X and Y directions, c is the cross-term coefficient reflecting the distortion degree of the focal plane, d and e are the inclination gradients in the X and Y directions, and f is the baseline height offset.

[0022] By minimizing the residual sum of squares ∑(z_measured-z_predicted)2, a continuous three-dimensional focal plane model is established, and the focus prediction and real-time compensation of any position in the scanning area are realized.

[0023] Step S5. Intelligent scanning imaging: according to the quadratic surface focal plane function fitted in step S4, the theoretical best focus of each position in the scanning path is calculated, and the objective lens assembly is driven to adjust the focus in real time for high-power scanning imaging.

[0024] A prediction-correction mechanism is adopted, that is, the focal length value of the target position is first predicted according to the focal plane function for coarse adjustment, and then fine correction is performed through real-time image clarity evaluation to ensure that the focal length deviation is controlled within the range of ±0.1μm.

[0025] High-definition images of each field of view are acquired in sequence according to the preset scanning path to form a complete image sequence, and each image maintains the best focus state.

[0026] The beneficial effects of the present invention are: The scanning area recognition accuracy of the present invention reaches over 95%, effectively eliminating the subjective differences in manual film reading selection and improving the efficiency and accuracy of bone marrow cytology diagnosis; the quadratic surface focal plane fitting algorithm improves the image clarity uniformity by more than 60%, ensuring the accurate identification of cell morphological characteristics; the intelligent scanning process integration significantly improves the overall scanning efficiency and meets the clinical high-throughput detection needs; the system automation level is significantly improved, and no manual intervention is required in the entire process from sample loading to image acquisition, providing reliable technical support for the standardized diagnosis of blood diseases and hematopoietic function assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 The present invention provides a flowchart of intelligent scanning of bone marrow smears using a high-power microscope.

[0028] Figure 2 It is the intelligent partition and focus point distribution map.

[0029] Figure 3 This is a multi-point focus data collection diagram.

[0030] Figure 4 is a schematic diagram of focal plane fitting.

[0031] Figure 5 This is a diagram of the structure of a high-power microscope. DETAILED DESCRIPTION

[0032] The following is combined with Figure 1-5 The technical solution of the present invention is further described in detail: like Figure 1 As shown in FIG, the flow chart describes the bone marrow smear high-power microscopic intelligent scanning method of the present invention, which includes five core steps (S1-S5) to form a complete intelligent scanning imaging system. Figure 1 As shown, the entire process is executed sequentially, and the output of each step serves as the input of the next step to ensure the continuity and accuracy of the scanning process.

[0033] Step S1. Automatically divide the scanning area. The implementation steps are as follows: S1.1 Panoramic image acquisition and preprocessing, including: S1.1.1 Start the front camera system and set standardized imaging parameters; S1.1.2 uses an adaptive exposure control algorithm to ensure uniform brightness of panoramic images; S1.1.3 perform image correction processing, including geometric distortion correction, color normalization, and noise filtering; S1.1.4 Establish a pixel-to-physical size mapping relationship to provide a benchmark for subsequent density calculations.

[0034] S1.2 Intelligent boundary identification and interference elimination, including: S1.2.1 Establish an interference recognition model based on HSV color space and texture features for automatic detection. Specifically: S1.2.1.1. Slide edge (based on high brightness and low saturation features); S1.2.1.2. Label area (based on regular geometric shapes and text textures); S1.2.1.3. Bubble area (based on circular contour and optical properties) S1.2.2 Generate a binary mask to mark the effective scanning area.

[0035] S1.3 Multi-dimensional quality assessment and optimal region selection, including: S1.3.1 Cell density assessment module, specifically: S1.3.1.1 Use the watershed algorithm to segment cell nuclei; S1.3.1.2 Calculate the number of cells per unit area, selecting areas with a density of 50-150 cells / mm²; S1.3.1.3 Establish a density uniformity index to assess the consistency of cell distribution within a region; S1.3.2 Overlap quantification module, specifically: S1.3.2.1 Analysis of cell morphological characteristics using ellipse fitting algorithm S1.3.2.2 Calculate the overlap ratio to ensure a low overlap rate S1.3.2.3 Introducing overlap severity weights to prioritize areas with slight overlap S1.3.3 Background uniformity assessment module, specifically: S1.3.3.1 Analyze the grayscale histogram distribution characteristics and calculate the standard deviation and coefficient of variation; S1.3.3.2 Detect local uneven staining and exclude areas of gradient and patchy abnormalities; S1.3.3.3 Establish a background consistency scoring mechanism; S1.3.4 Staining quality assessment module, specifically: S1.3.4.1 Analyze Wright's staining effects based on the HSV color model; S1.3.4.2 Assess the staining contrast between the nucleus (blue-purple) and cytoplasm (pink); S1.3.4.3 Quantify color saturation and color purity indicators; S1.3.5 Optimal region selection module; S1.3.5.1 Comprehensive scoring mechanism S1.3.5.1.1 Establish a weighted scoring model: Cell density assessment weight: w1 = 0.25, overlap quantification weight: w2 = 0.30, background uniformity assessment weight: w3 = 0.25, staining quality assessment weight: w4 = 0.20 S1.3.5.1.2 Calculate the overall regional quality score: Total score = w1 × density score + w2 × overlap score + w3 × background score + w4 × staining score S1.3.5.3 Determination of the optimal area S1.3.5.3.1 Ranking selection: Sort by comprehensive score from high to low; S1.3.5.3.2 Area quantity control: Select the top N optimal areas (usually N = 3-5) based on the testing requirements; S1.3.5.3.3 Quality Verification: Perform secondary quality confirmation on selected areas.

[0036] Step S2. Intelligent partitioning and focus point distribution: Step S2. Intelligent Partitioning and Focus Point Distribution S2.1 Intelligent partitioning of optimal scanning area S2.1.1 Dynamically determine the N×M grid based on the area of ​​the region; S2.1.2 Establish sub-region index codes and boundary coordinate records; S2.1.3 Ensure the uniformity of size of each sub-area (error < 5%).

[0037] S2.2 Stratified Sampling Strategy and Focus Determination S2.2.1 Standard configuration P = 6-11 sampling points / sub-area; S2.2.2 Multi-level point distribution: 1 center point (weight 0.4-0.5) + 4 corner points (weight 0.1-0.15) + several random points; S2.2.3 Use minimum spacing constraints to ensure uniform spatial distribution.

[0038] S2.3 Three-dimensional coordinate mapping table construction S2.3.1 Establish an XYZ coordinate system (accuracy: XY ± 1 μm, Z ± 0.5 μm); S2.3.2 Estimate the Z-axis focal length of each point based on the slide thickness model; S2.3.3 Generate a mapping table containing seven parameters including sampling point ID, absolute / relative coordinates, and weight coefficients.

[0039] Step S3. Multi-point focus data acquisition S3.1 Precise Positioning and Stage Control S3.1.1 reads the three-dimensional coordinate mapping table output by S2 and accesses N²×M sampling points in sequence; S3.1.2 Drive the stage XY axis to precisely move to the target position (positioning accuracy ±0.5μm); S3.1.3 Move the objective lens assembly to the estimated focal length position on the Z axis as the focus starting point; S3.1.4 Establish a motion path optimization algorithm to minimize movement time and mechanical wear.

[0040] S3.2 CNN Autofocus Model Implementation S3.2.1 Autofocus model built using CNN binary classification architecture; S3.2.2 Input the image block of the current focal length (size 3208 × 2200 pixels); S3.2.3 Multi-layer convolution and pooling operations extract key features such as cell edge clarity and texture details; S3.2.4 Output layer binary classification judgment: focus state (confidence > 0.8) or defocus state; S3.2.5 Single inference time ≤ 50ms, ensuring real-time performance.

[0041] S3.3 Focus data matrix construction: S3.3.1 Record the final three-dimensional coordinates (x, y, z) and focus confidence of each sampling point; S3.3.2 construct an N × M × P dimensional focus data matrix and store it in order by sub-region and sampling point number; S3.3.3 Perform data quality checks and flag outliers and low-confidence points.

[0042] Step S4. Least squares focal plane fitting S4.1 Quadratic surface fitting model construction S4.1.1 Establish the equation of the quadratic surface: z = ax² + by² + cxy + dx + ey + f; S4.1.2 Parameter physical meaning: a and b are the quadratic curvature coefficients in the X and Y directions; c is the cross-term coefficient (focal plane distortion); d and e are the tilt gradients; f is the reference offset; S4.1.3 Construct the linear equation system Aθ = b, where θ = [a, b, c, d, e, f]ᵀ is the parameter vector to be determined; S4.1.4 Use the normal equation θ = (AᵀA)⁻¹Aᵀb to solve for the optimal parameters.

[0043] S4.2 Least Squares Optimization and Model Validation S4.2.1 Optimize the fitting parameters by minimizing the sum of squared residuals ∑(z measured - z predicted)²; S4.2.2 Calculate the goodness of fit R² and root mean square error RMSE to assess model quality; S4.2.3 Perform cross-validation and randomly retain 20% of the data points for model validation; S4.2.4 Generate a distribution plot of the fitted residuals to identify areas of systematic bias.

[0044] S4.3 Generation of continuous focal plane distribution model S4.3.1 The package fitting parameters are the focal length prediction function f(x,y) = ax² + by² + cxy + dx + ey +f; S4.3.2 Establish the focal length gradient field and calculate ∇f = (∂f / ∂x, ∂f / ∂y) for motion trajectory optimization; S4.3.3 Set the model validity range and confidence interval, and mark the low reliability areas.

[0045] Step S5. Intelligent scanning imaging S5.1 Scanning Path Planning and Focal Length Prediction S5.1.1 Import the quadratic focal plane function f(x,y) = ax² + by² + cxy + dx + ey+ f output from S4; S5.1.2 Plan a serpentine scanning path to minimize stage movement distance and time; S5.1.3 Calculate the theoretical optimal focal length value for each field of view position along the scan path; S5.1.4 Establish a focus change gradient prediction to optimize the focus adjustment strategy between consecutive fields of view.

[0046] S5.2 Predictive-corrective focus mechanism S5.2.1 Coarse adjustment stage: Rapidly adjust the Z-axis position of the objective lens based on the predicted value of the focal plane function; S5.2.2 Fine-tuning stage: fine-tuning using the autofocus model algorithm; S5.2.3 Ensure that the final focal length deviation is controlled within the range of ±0.1 μm; S5.2.4 Single focus time ≤ 100ms to ensure scanning efficiency.

[0047] S5.3 High-magnification scanning imaging execution S5.3.1 Drive the stage to move along a predetermined path, and adjust the focal length of the objective lens assembly synchronously; S5.3.2 Acquire high-resolution images at each field of view; S5.3.3 Implement real-time monitoring of image quality and automatically retake blurry or abnormal images; S5.3.4 Record the spatial coordinates, focal length, and quality score of each image.

[0048] S5.4 Image sequence integration and quality control S5.4.1 Arrange the image sequence in scanning order and establish an image-coordinate index table; S5.4.2 Perform statistical analysis of image quality and calculate average clarity and success rate; S5.4.3 Mark low-quality areas that require rescanning and perform supplemental scans.

[0049] like Figure 2 An intelligent partitioning method for the optimal scanning area is demonstrated, which is evenly divided into N×M sub-areas. A hierarchical sampling strategy is used in each sub-area to determine P focus sampling points to optimize the focus accuracy and coverage.

[0050] 2.1 Determination of the Optimal Scanning Area As shown in the figure, in the panoramic image of the bone marrow smear, the red rectangular frame marked ① identifies the optimal scanning area automatically divided in step S1. This area has the following characteristics: 2.1.1 The cell density is moderate and evenly distributed; 2.1.2 The degree of overlap is below the set threshold of 15%; 2.1.3 The background staining is uniform and there is no obvious interference; 2.1.4 The Wright staining effect is clear and the cell morphology is complete.

[0051] 2.2 Intelligent Partitioning Strategy As shown in the figure marked ②, the optimal scanning area is intelligently divided according to the preset N×M grid pattern: 2.2.1 Using an equal segmentation algorithm, the rectangular scanning area is divided into multiple sub-areas of equal area; 2.2.2 The size of each sub-region is dynamically adjusted according to the microscope field of view and overlap requirements; 2.2.3 The sub-area boundaries are divided by regular grid lines to ensure the continuity and coverage integrity of the scanning path; 2.2.4 Grid division takes into account the stage movement accuracy and imaging overlap rate to optimize scanning efficiency.

[0052] 2.3 Layered Sampling Point Distribution The blue area shown in the figure marked ③ shows the focus sampling point distribution strategy within a single sub-area: 2.3.1 Grid Distribution: M focus sampling points are distributed in each sub-area according to a regular grid pattern. 2.3.2 Stratified Sampling Architecture: 2.3.2.1 First layer: 4 corner sampling points (marked in red), covering the focal length variation at the sub-region boundary; 2.3.2.2 Second layer: edge midpoint sampling points, supplementing boundary focal length gradient information; 2.3.2.3 The third layer: internal grid points, ensuring the continuity of focal length distribution within the area.

[0053] 2.4 3D Coordinate Mapping Table Construction The system automatically creates a 3D coordinate mapping table for the focus point, specifically including: 2.4.1 Plane coordinate recording: record the X and Y plane coordinates of each sampling point, with the absolute coordinates of the stage movement as the reference; 2.4.2 Estimating Z-axis information: Based on the prior knowledge of smear thickness distribution, estimate the initial focal length position of each sampling point; 2.4.3 Index coding system: Assign a unique coding identifier to each sampling point to facilitate subsequent data management and retrieval.

[0054] like Figure 3 The multi-point focus data acquisition process is demonstrated. A preset focus point coordinate mapping table is used to drive the stage and objective lens assembly to move precisely to each sampling point. An autofocus model is used to quickly obtain the optimal focal length value, and an N×M×P-dimensional focus data matrix is ​​constructed to provide an accurate spatial reference for subsequent focal plane interpolation and real-time focus adjustment.

[0055] S3.1. Process initialization This step initiates the multi-point focus data acquisition process. The system first reads the focus point coordinate mapping table preset in step S2. This mapping table contains the two-dimensional coordinate information (xi,yi) of all predefined sampling points within the sampling area, where i represents the sampling point number, i = 1, 2, ..., N × M. Simultaneously, the system initializes the stage control module and the objective lens assembly control module to ensure that all hardware components are ready.

[0056] S3.2. Sampling area settings The system determines the spatial range of data acquisition based on preset sampling area parameters. The sampling area is defined by the starting coordinates (x0, y0) and the ending coordinates (xmax, ymax), forming a rectangular sampling area. The system verifies the validity of the sampling area, ensuring that all sampling point coordinates are within the effective range of the stage's motion to avoid mechanical conflicts or out-of-bounds operations.

[0057] S3.3. Multi-point traversal acquisition and autofocus execution The system sequentially traverses each sampling point to collect data according to the preset focus point coordinate mapping table. The system first drives the stage's XY-axis stepper motors to precisely move the stage to the current sampling point coordinates (xi, yi). During this movement, the system monitors the stage's position feedback in real time to ensure that positioning accuracy meets preset requirements (typical accuracy is ±1μm). After the stage reaches the target position, the system drives the objective lens assembly along the Z-axis to the preset starting focus position. This starting position is typically selected near the estimated optimal focal length to shorten subsequent autofocus time.

[0058] The system executes an autofocus procedure at the current sampling point. The objective lens assembly is fine-tuned along the Z axis, capturing an image at each Z position. The system then evaluates the clarity of each image. Using the autofocus model, the system analyzes the image clarity at different Z positions and finds the Z position corresponding to the maximum value of the clarity evaluation function, which is the optimal focal length position zi for the current sampling point. This process utilizes an optimized hill-climbing algorithm to ensure submicron focal accuracy.

[0059] Step S3.4. 3D coordinate data recording The system records the complete three-dimensional coordinate information (xi, yi, zi) of the current sampling point, where zi represents the optimal focal length position of the point. All data is stored in a preset data format, including the sampling point number, XY coordinates, Z-axis focal length value, image clarity evaluation value, and acquisition timestamp.

[0060] Step S3.5. Collection completion judgment The system determines whether data collection of all preset sampling points has been completed. The judgment condition is: whether the current sampling point number i is equal to the total number of sampling points N×M. If the collection is not completed (i <N×M),系统返回步骤S3.3,继续遍历下一个采样点;如果采集完成(i=N×M),系统进入步骤S3.6。

[0061] Step S3.6. Focus data matrix construction The system constructs all collected focus data into an N×M×P dimensional focus data matrix F(x,y,z), where: N×M×P represents the total number of sampling points in the sampling area P represents the data dimension of each sampling point, including coordinate information, focal length value, clarity evaluation value, etc. The matrix element F(i,j,k) represents the k-th dimension data of the sampling point in the i-th row and j-th column. This data matrix provides a complete spatial reference data basis for subsequent focal plane interpolation calculation and real-time focus adjustment.

[0062] Step S3.7. End of process Once the multi-point focus data acquisition process is complete, the system outputs a complete focus data matrix for subsequent use by the microscope's autofocus control system. This data matrix serves as input to the focal plane interpolation algorithm, enabling high-precision, real-time focus adjustment across the entire field of view.

[0063] like Figure 4 The figure shows the focus distribution model obtained by fitting the three-dimensional plane of all focus point data within the scanning area using the least squares method. The horizontal axis (X) ranges from -70,000 to -60,000, the vertical axis (Y) ranges from -39,000 to -37,600, and the vertical axis (Z) ranges from 16,140 to 16,280, all in microns.

[0064] Figure 4 The golden surface in the middle represents the continuous focal plane model obtained by least squares fitting. This focal plane strictly follows the quadratic surface equation z=ax²+by²+cxy+dx+ey+f. It can be clearly observed from the figure: Overall morphological characteristics of the focal plane: The fitted focal plane presents obvious quadratic surface features, with significant curvature changes in both the X and Y directions, and the Z value presents a continuous and smooth distribution within the scanning area.

[0065] Curvature distribution law: The focal plane shows a monotonic change trend in the X direction from -70000 to -60000, and the Z value gradually decreases from about 16200 to 16150; in the Y direction from -39000 to -37600, it also shows a regular change, reflecting the field curvature characteristics of the lens in the laser processing system.

[0066] Verification of fitting accuracy: The scattered sampling points (marked with small dots) in the figure are highly consistent with the fitted surface, indicating that the least squares fitting algorithm effectively minimizes the residual sum of squares ∑(z_measured - z_predicted)², achieving high-precision focal plane modeling.

[0067] Physical meaning of parameters: The fitting results can be used to determine the specific values ​​of the coefficients in the focal plane equation. The quadratic curvature coefficients a and b reflect the degree of field curvature in the X and Y directions, the cross-term coefficient c reflects the distortion characteristics of the focal plane, the tilt gradients d and e characterize the overall tilt trend of the focal plane, and the reference height offset f determines the absolute position of the focal plane.

[0068] After the focal plane model is established, the system can calculate the corresponding theoretical optimal focal length value z in real time based on the coordinates (x, y) of any position in the scanning area, providing an accurate theoretical basis for subsequent dynamic focal length compensation.

[0069] like Figure 5The schematic diagram of the microscope system is shown below. The following describes each component in detail: Camera: Located at the top of the microscope system, its main function is to capture high-resolution images of the microscope field of view, transmit image data to the computer in real time for image processing and analysis, and support image quality assessment of the autofocus algorithm.

[0070] 2. Objective Lens: Installed on a rotatable objective lens turntable, it contains multiple objective lenses with different magnifications (such as 10x, 20x, 40x, 100x, etc.), responsible for optically magnifying the sample and can quickly switch between different magnifications according to observation needs.

[0071] 3. Motorized Stage: It has high-precision XY-axis electric control function, which can realize automatic scanning and precise positioning of samples. It supports programmed control and can move according to preset paths to ensure the accurate positioning of samples in the field of view.

[0072] 4. High-Power Illumination: Provides a strong illumination system to ensure sufficient image brightness during focusing, effectively reduce image noise, improve autofocus accuracy and stability, and provide the necessary lighting conditions for high-quality imaging.

[0073] 5. AutoZ-axis: Controls the vertical movement of the objective lens or stage to achieve precise focal length adjustment. It is the core component of the autofocus function and uses algorithmic control to automatically find and lock the optimal focal plane.

[0074] 6. Front camera: used to capture panoramic images of bone marrow smears, provide overall field of view information of the sample, and assist the system in global positioning and area selection.

[0075] 7. Front lighting: Provides uniform lighting for the front camera, ensuring the clarity and contrast of the panoramic image, and cooperates with the front camera to complete the pre-scanning and positioning of the sample.

[0076] This microscope system achieves a fully automated process from sample positioning, autofocus to image acquisition through the collaborative work of multiple components. It is particularly suitable for the automated analysis of medical samples such as bone marrow smears.

[0077] In summary, the main advantages of the technical solution of the present invention are as follows: 1. Intelligent area recognition with high precision: Edge detection and area segmentation algorithms are used to automatically identify the effective boundaries of smears. The accuracy of scanning area recognition is over 95%, effectively eliminating interference factors such as slide edges, labels, bubbles, etc., and eliminating subjective differences in manual selection.

[0078] 2. A comprehensive, multi-dimensional quality assessment system includes four core evaluation modules: cell density, overlap, background uniformity, and staining quality. A weighted scoring mechanism automatically selects the optimal scanning area to ensure uniform cell distribution, low overlap, and clear staining.

[0079] 3. The advanced intelligent zoning focus strategy intelligently divides the scanning area into N×M sub-areas. A multi-level sampling strategy (center point + corner points + random points) is used to ensure uniform coverage of sampling points. A complete 3D coordinate mapping table is established through a spatial distribution optimization algorithm to achieve high-precision focus coverage.

[0080] 4. The high-precision focal plane fitting algorithm uses the least squares method to perform three-dimensional plane fitting and establish the quadratic focal plane equation. This enables focal length prediction at any position within the scanning area, improving image clarity and uniformity by over 60%.

[0081] 5. An intelligent prediction-correction mechanism combines focal plane function prediction with real-time image clarity evaluation to ensure focus deviation is controlled within ±0.1μm, ensuring optimal focus for every image.

[0082] 6. High degree of automation throughout the entire process: From sample loading to image acquisition, no human intervention is required, significantly improving the system's automation level and meeting clinical high-throughput testing needs.

[0083] 7. Outstanding clinical application value: It provides reliable technical support for the standardized diagnosis of blood diseases and the evaluation of hematopoietic function, improves the accuracy, efficiency and repeatability of bone marrow cytological diagnosis, and has important clinical promotion value.

Claims

1. A bone marrow smear high-power microscope intelligent scanning method, characterized in that: The steps include: Step S1. Automatically divide the scanning area: Use the front camera to obtain the full image of the bone marrow smear, and use edge detection and region segmentation algorithms to automatically identify the effective boundaries of the smear, excluding interference areas such as the slide edge, labels, and bubbles; Establish a multi-dimensional quality assessment model that comprehensively analyzes parameters such as background uniformity and staining quality, automatically selecting the optimal scanning area with uniform cell distribution, low overlap rate and clear staining, and improving the accuracy and repeatability of bone marrow cytology diagnosis; Step S2. Intelligent Partitioning and Focus Point Distribution: The optimal scanning area determined in step S1 is intelligently divided into N×M equal sub-regions. Within each sub-region, a stratified sampling strategy is used to determine P focus sampling points. The system automatically creates a 3D coordinate mapping table for the focus points, recording the X and Y coordinates and estimated Z-axis focal length information of each sampling point to ensure high-precision focus coverage across the entire area. Step S3. Multi-point focus data acquisition: According to the three-dimensional coordinate mapping table of the focus points preset in step S2, the stage and objective lens assembly are driven to move precisely to each sampling point position in sequence. At each position, the optimal focal length value is quickly obtained through the autofocus model, and the corresponding three-dimensional coordinate information (x, y, z) is recorded. The z value represents the optimal focal length position of the point. All collected focus data are constructed into an N×M×P-dimensional focus data matrix to provide complete spatial reference data for subsequent focal plane interpolation calculations and real-time focus adjustment. The output layer performs binary classification to determine the focus or defocus state. The model is trained using a large amount of labeled data from bone marrow smears at different focal lengths to learn the optimal focus feature representation, enabling fast and accurate focus quality assessment. The single inference time is controlled within 50ms, meeting the requirements of real-time focus efficiency. Step S4. Least Squares Focal Plane Fitting: The least squares method is used to perform three-dimensional plane fitting on all focus point data collected in step S3. Based on the focal plane function obtained by fitting, the system can calculate the theoretical optimal focal length value at any position (x, y) within the scanning area and generate a continuous and smooth focal plane distribution model. Step S5. Intelligent scanning imaging: Calculate the theoretical optimal focal length of each position in the scanning path based on the quadratic surface focal plane function fitted in step S4, and drive the objective lens assembly to adjust the focal length in real time to perform high-magnification scanning imaging; A prediction-correction mechanism is used, which first predicts the focal length value at the target position based on the focal plane function for coarse adjustment, and then performs fine correction through real-time image clarity evaluation to ensure that the focal length deviation is controlled within the range of ±0.1μm; High-definition images of each field of view are acquired in sequence according to the preset scanning path to form a complete image sequence, and each image maintains the best focus state.

2. The method for intelligent scanning of bone marrow smears using a high-power microscope according to claim 1, characterized in that: In step S1, the multidimensional quality assessment model works in collaboration through four core modules: a cell density assessment module, an overlap quantification module, a background uniformity assessment module, and a staining quality assessment module; The cell density assessment module screens areas with moderate density; the overlap quantification module uses morphological analysis to ensure low overlap; the background uniformity assessment module analyzes grayscale distribution and excludes areas of uneven staining; the staining quality assessment module analyzes the Wright staining effect based on the HSV color model; and a weighted scoring mechanism is established to automatically select the scan area with the highest comprehensive score through normalization processing.

3. The method for intelligent scanning of bone marrow smears using a high-power microscope according to claim 1, characterized in that: In step S2, the P focus sampling points adopt a multi-level distribution strategy: including 1 center point to ensure regional representativeness, 4 corner points to cover boundary focal length changes, and several randomly distributed points to supplement local thickness fluctuation information. The uniformity and coverage integrity of the sampling points are ensured through a spatial distribution optimization algorithm, thereby improving the focus accuracy and stability of subsequent image acquisition.

4. The method for intelligent scanning of bone marrow smears using a high-power microscope according to claim 1, characterized in that: In step S3, the autofocus model adopts a CNN binary classification architecture to construct a convolutional neural network, inputs the image block of the current focal length, and automatically extracts key features of cell edge clarity and texture details through multi-layer convolution and pooling operations.

5. The method for intelligent scanning of bone marrow smears using a high-power microscope according to claim 1, characterized in that: In step S4, the three-dimensional plane fitting establishes the focal plane equation z=ax²+by²+cxy+dx+ey+f: the fitting parameters are solved by the least squares optimization algorithm, where a and b are the quadratic curvature coefficients in the X and Y directions, c is the cross-term coefficient reflecting the degree of focal plane distortion, d and e are the tilt gradients in the X and Y directions, and f is the reference height offset; by minimizing the sum of squares of the sampling point residuals ∑(z_measured-z_predicted)², a continuous three-dimensional focal plane model is established to achieve focal length prediction and real-time compensation at any position in the scanning area.

Images