A computer vision-based multiple myeloma bone marrow image recognition system

Abstract

The application discloses a multiple myeloma bone marrow image recognition system based on computer vision, and relates to the technical field of cell recognition.The application solves the problem that traditional manual microscopy has great limitations, accurately recognizes a burr change section through angle analysis on the edge profile of a feature region, reliably judges whether a cell to be detected is a tumor cell according to the comprehensive proportion of the burr change section on the overall edge profile, and calibrates and displays the tumor cell in a gray-scale image.The method effectively utilizes the significant difference between tumor cells and normal cells in the morphology of cell nuclei, realizes fast and accurate recognition of the tumor cells, provides intuitive and clear results for clinical diagnosis, helps doctors to timely and accurately judge multiple myeloma, and improves the diagnosis efficiency and treatment effect.The mean value processing of the gray value of the profile region and the radial ratio analysis can accurately lock the profile region associated with the cell nucleus from the cell to be detected, calibrate the profile region as a feature region, and provide a key cell nucleus feature basis for recognizing the tumor cells.

Description

Technical Field

[0001] This invention relates to the field of cell recognition technology, specifically to a computer vision-based bone marrow image recognition system for multiple myeloma. Background Technology

[0002] Multiple myeloma (MM), a common malignant tumor of the blood system, poses a serious threat to human health. According to statistics from the World Health Organization, the global incidence of MM is increasing year by year, and the age of onset is trending younger.

[0003] With the rapid development of computer technology, computer vision technology is increasingly widely used in the medical field. Computer vision can quickly process massive amounts of image data and accurately extract image features through algorithms, showing great potential in disease diagnosis. In the identification of myeloma bone marrow morphology, the use of computer vision technology to build an automated identification system is expected to break through the limitations of traditional manual microscopic examination, achieve rapid and accurate identification of myeloma cells, improve the efficiency and accuracy of myeloma diagnosis, buy valuable time for clinical treatment, and improve patient prognosis. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a computer vision-based bone marrow image recognition system for multiple myeloma, which solves the problem of significant limitations in traditional manual microscopic examination.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a computer vision-based multiple myeloma bone marrow image recognition system, comprising:

[0006] The high-definition microscopic image conversion end converts the acquired high-definition microscopic images into grayscale images.

[0007] On the grayscale feature processing end, the SobeL algorithm is used to confirm the gradient features associated with different pixels in the grayscale image. Based on the gradient features corresponding to different pixels in the grayscale image, cell outlines are confirmed and cell selection is performed to identify individual cells. The specific method is as follows:

[0008] The SobeL algorithm was used to determine the lateral gradient X associated with different pixels in the grayscale image. i and vertical gradient S i Where i represents different pixels, using: Confirm the gradient feature ZH corresponding to this pixel. i ;

[0009] Will satisfy ZH i Pixels with a value greater than or equal to Y1 are labeled as gradient pixels; otherwise, no labeling is performed and Y1 is a preset value.

[0010] Based on the gradient pixels identified sequentially within the grayscale image, several adjacent gradient pixels are connected to form a line outline, and the individual associated with this line outline is recorded as a single cell. This line outline belongs to the edge outline of this single cell.

[0011] In the polymorphic cell analysis, based on the different cell monomers identified within the grayscale image, the contour regions existing within each cell monomer are confirmed. Based on the number of contour regions, the cell monomer is labeled as either a unimorphic or polymorphic cell. The cell to be examined is then identified from within the unimorphic cells. The specific method is as follows:

[0012] The SobeL algorithm was used to identify the lateral gradient X associated with different pixels within different cell monomers. k and vertical gradient S k Where k represents different pixels within a single cell, using: Confirm the gradient feature ZH corresponding to this pixel. k ;

[0013] Will satisfy ZH k Pixels with a value greater than or equal to Y2 are labeled as gradient pixels; otherwise, no labeling is performed, and Y2 is a preset value, where Y2 < Y1.

[0014] Based on the gradient pixels identified sequentially within a single cell, several adjacent gradient pixels are connected to identify the corresponding contour region. If there is only one contour region within this single cell, this single cell is labeled as a monomorphic cell. If there are multiple contour regions within this single cell, it is labeled as a polymorphic cell.

[0015] Based on the edge contour associated with the monomorphic cell, the cell center point belonging to this monomorphic cell is first identified. A set of contour points on the edge contour is randomly selected and recorded as principal points. The principal points are connected to the cell center point to identify a set of lines. This line is then extended to the other side of the edge contour to identify the first set of dividing lines. A perpendicular line belonging to this dividing line is then constructed, with both endpoints of the perpendicular line located on the edge contour. The evaluation ratio of the dividing line and the perpendicular line is confirmed: the evaluation ratio = the length of the longer line segment ÷ the length of the shorter line segment. The longer line segment is the longest line segment between the dividing line and the perpendicular line, and the same applies to the shorter line segment. The dividing line is rotated based on this cell center point. Based on the rotation process, the maximum evaluation ratio is confirmed. If the maximum evaluation ratio is ≥1.5, this monomorphic cell is labeled as the cell to be tested; otherwise, no labeling is performed.

[0016] In the feature region labeling stage, the SobeL algorithm is used to sequentially confirm other contour regions within the cell under test, lock the contour region associated with the cell nucleus, and then perform specific feature region labeling again. The specific method is as follows:

[0017] The SobeL algorithm and the set threshold are used to sequentially confirm other contour regions existing in the cells to be examined, with different contour regions corresponding to different thresholds;

[0018] The mean value of several sets of gray values ​​associated with different pixels in different contour areas is processed to confirm the regional features associated with the corresponding contour areas. Contour areas whose regional features belong to a set interval are marked as suspected regions, and their set interval is a preset interval.

[0019] If there is only one suspected region, then this suspected region is marked as a feature region;

[0020] If multiple suspected regions exist, the radial ratio of these regions is used to confirm their existence through a verification unit, followed by feature analysis to pinpoint the feature regions.

[0021] Based on the edge contour of the suspected area, the center point of the area belonging to this suspected area is locked. Then, the same processing method used for confirming the cells to be tested for single-state cells is adopted to construct the dividing line and vertical line in the suspected area, and the evaluation ratio belonging to this suspected area is confirmed. From several sets of evaluation ratios, the largest evaluation ratio is selected and recorded as the standard feature of this suspected area. From several sets of standard features, the value closest to 1 is selected and the suspected area associated with the selected value is recorded as the feature area.

[0022] The tumor cell identification and labeling terminal identifies and marks the characteristic regions marked within the cell under test by identifying spiky segments on the edge contour of the characteristic regions. Based on the overall proportion of these spiky segments on the overall edge contour, it confirms whether the cell under test is a tumor cell and displays the labeling. The specific method is as follows:

[0023] Based on the edge contour of the feature region, identify the feature center point belonging to this feature region, and randomly select a set of contour points from the edge contour as the starting point. Then, determine the angle of the contour segment between the starting point and the adjacent points in a clockwise direction.

[0024] Connect the starting point to the feature center point to confirm the first angle line. Then connect the starting point to the adjacent point to confirm the second angle line. Confirm the angle J between the first angle line and the second angle line. If 85°≤J≤95°, then mark the part of the contour segment between the starting point and the adjacent point as the normal segment. Otherwise, mark it as the burr variation segment.

[0025] Then, using the same processing method, the normal segment or burr change segment is marked for the subsequent contour segments in sequence, until the overall edge contour of the feature area is completely marked.

[0026] Confirm the total length L1 of several spur change segments, and then confirm the total length L2 of the feature region edge contour. If (L1÷L2)≤0.2, no calibration is performed. Otherwise, the cell to be examined is calibrated as a tumor cell and calibrated simultaneously in the grayscale image.

[0027] This invention provides a computer vision-based system for identifying bone marrow morphology in multiple myeloma. Compared with existing technologies, it has the following advantages:

[0028] The polymorphic cell analysis end can accurately distinguish between monomorphic and polymorphic cells based on the number of contour regions inside the cell monomer. Through a unique segmentation line and vertical line evaluation method, it can effectively screen out suspected tumor cells from monomorphic cells, providing precise targets for subsequent in-depth analysis and helping to improve the targeting of tumor cell diagnosis.

[0029] By utilizing the Sobe I algorithm and empirical threshold setting, combined with mean processing of gray values ​​in the contour region and radial ratio analysis, the contour region associated with the cell nucleus in the cell under test can be accurately located and marked as a feature region, providing key cell nucleus feature basis for tumor cell identification and improving the accuracy and reliability of tumor cell identification.

[0030] By performing angular analysis on the edge contour of the characteristic region, the burr-like change segments are accurately identified. Based on the comprehensive proportion of the burr-like change segments on the overall edge contour, it is reliably determined whether the cell to be examined is a tumor cell, and it is marked and displayed in the grayscale image. This method effectively utilizes the significant differences in the nuclear morphology between tumor cells and normal cells to achieve rapid and accurate identification of tumor cells, providing intuitive and clear results for clinical diagnosis, helping doctors to make timely and accurate judgments of multiple myeloma, and improving diagnostic efficiency and treatment effectiveness. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the principle framework of the present invention;

[0032] Figure 2 This is a schematic diagram illustrating the determination of the angle between angle line 1 and angle line 2 in this invention. Detailed Implementation

[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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] Please see Figure 1This application provides a computer vision-based multiple myeloma bone marrow image recognition system, including a high-definition microscopic image conversion terminal, a grayscale feature processing terminal, a polymorphic cell analysis terminal, a feature region calibration terminal, a verification unit, and a tumor cell recognition calibration terminal;

[0035] Among them, the high-definition microscopic image conversion end, grayscale feature processing end, polymorphic cell analysis end, and feature region calibration end are electrically connected from the output node to the input node in sequence, and the verification unit is bidirectionally connected to the feature region calibration end, and the feature region calibration end is electrically connected to the input node of the tumor cell identification calibration end.

[0036] The high-definition microscopic image conversion unit converts the acquired high-definition microscopic images to grayscale. Based on preset weights and the RGB values ​​of corresponding points, it confirms the grayscale value of the corresponding points and converts the high-definition microscopic images into grayscale images. Specifically, the high-definition microscopic images are generally obtained using a high-resolution microscope imaging system equipped with a 50-megapixel professional microscope camera, which can clearly capture the details of bone marrow smears. It is equipped with automatic focus and intelligent exposure adjustment functions to ensure stable, high-quality images are acquired at different magnifications (such as 400X, 1000X).

[0037] The specific method for converting grayscale images is as follows:

[0038] Based on the confirmed high-resolution photomicrograph, the RGB values ​​associated with the corresponding pixel points are confirmed. Then, the grayscale value HD associated with the corresponding pixel point is confirmed using the formula: HD = 0.299R + 0.587G + 0.114B. The grayscale value is adjusted based on the grayscale value associated with different pixel points to convert the high-resolution photomicrograph into a grayscale image.

[0039] Specifically, the method of converting grayscale values ​​is quite common in existing technologies, so it will not be elaborated on here. Its RGB values ​​are three sets of values, each set of values ​​is assigned a different weight, so that the grayscale values ​​of pixels with different features can be adjusted.

[0040] In the grayscale feature processing section, based on the grayscale values ​​associated with different pixels within the grayscale image, the SobeL algorithm is used to confirm the gradient features associated with the corresponding pixels. Based on the gradient features corresponding to different pixels within the grayscale image, cell outlines are confirmed and cells are selected. Specifically, the pixels are sorted in a nine-grid arrangement, so each pixel is surrounded by eight groups of pixels, and each pixel has a corresponding grayscale value. By employing the SobeL algorithm to assign different weight factors to different pixels, the associated weight factors are different when confirming the horizontal and vertical gradients. Since the method of confirming the gradient values ​​corresponding to pixels is common in existing technologies, it will not be elaborated further here. The SobeL algorithm uses two 3×3 convolution kernels: one for calculating the horizontal gradient, denoted as Gx; and the other for calculating the vertical gradient, denoted as Gy. Different gradients correspond to different convolution kernels. Based on different convolution kernels and the grayscale values ​​associated with different pixels, the horizontal and vertical gradients associated with the center pixel can be confirmed.

[0041] The specific method for cell selection is as follows:

[0042] The SobeL algorithm was used to determine the lateral gradient X associated with different pixels in the grayscale image. i and vertical gradient S i Where i represents different pixels, using: Confirm the gradient feature ZH corresponding to this pixel. i ;

[0043] Will satisfy ZH i Pixels with a value greater than or equal to Y1 are labeled as gradient pixels; otherwise, no labeling is performed, and Y1 is a preset value. The specific value of Y1 is determined by the operator based on experience. Y1 is generally between 150 and 200 and is determined in advance by the operator.

[0044] Based on the gradient pixels identified sequentially within the grayscale image, several adjacent gradient pixels are connected to form a set of line contours. The individual associated with this line contour is recorded as a single cell, and this line contour belongs to the edge contour of this single cell.

[0045] Specifically, on the outline of each cell, the associated outline points are all gradient points. There are obvious gray-scale changes between the cell and the medullary fluid, and the corresponding gradient pixel changes are also quite obvious. Then, there are corresponding gradient outline points on the outer side of the cell, which are the associated edge outlines.

[0046] In the polymorphic cell analysis, based on the different cell monomers identified within the grayscale image, the contour regions (i.e., vacuolar regions) existing within each cell monomer are confirmed. Based on the number of contour regions, the cell monomer is labeled as either a unimorphic or polymorphic cell. The cell to be examined is then identified from within the unimorphic cell. A unimorphic cell can be understood as a single cell, i.e., a complete cell entity. A polymorphic cell can be understood as multiple cells clustered together, resulting in overlapping edge contours and a larger contour, but each cell contains multiple corresponding vacuoles, thus creating multiple contour regions. The specific method for labeling cell monomers is as follows:

[0047] The SobeL algorithm was used to identify the lateral gradient X associated with different pixels within different cell monomers. k and vertical gradient S k Where k represents different pixels within a single cell, using: Confirm the gradient feature ZH corresponding to this pixel. k ;

[0048] Will satisfy ZH k Pixels with a value ≥ Y2 are labeled as gradient pixels; otherwise, no labeling is performed, and Y2 is a preset value. Y2 < Y1 because the gray value evaluation benchmark for Y1 is the gray value feature between the cell and the medullary fluid, which has a stronger feature. Y2 is the gray value feature between the vacuoles and the cytoplasm, which has a weaker feature. Therefore, Y2 < Y1.

[0049] Based on the gradient pixels identified sequentially within a single cell, several adjacent gradient pixels are connected to identify the corresponding contour region. If there is only one contour region within this single cell, this single cell is labeled as a monomorphic cell. If there are multiple contour regions within this single cell, it is labeled as a polymorphic cell.

[0050] Based on the edge contour associated with the monomorphic cell, the cell center point belonging to this monomorphic cell is first identified (by combining the edge contour with a two-dimensional coordinate system, identifying the two-dimensional coordinates of different points within the contour, then averaging several two-dimensional coordinates to identify the mean coordinates, and marking the identified mean coordinates within the monomorphic cell to identify the cell center point). A set of contour points on the edge contour is randomly selected and recorded as principal points. The principal points are connected to the cell center point to identify a set of lines, and these lines are extended to the other side of the edge contour to identify the first set of dividing lines. Then, a perpendicular line belonging to this dividing line is constructed, with both endpoints of the perpendicular line also located on the edge contour. The ratio of the dividing line to the perpendicular line is identified: the ratio = the length of the longer segment ÷ the length of the shorter segment, where the longer segment is the longest segment between the dividing line and the perpendicular line, and the same applies to the shorter segment. This ensures that the dividing line is based on the cell center point. The centroid rotates, and based on the rotation process, the maximum rating ratio is determined. If the maximum rating ratio is ≥1.5, the monomorphic cell is labeled as the cell to be tested; otherwise, no labeling is performed. For example, there is a clear difference in appearance between normal cells and tumor cells: normal cells are relatively round, while tumor cells are relatively elliptical, and the degree of ellipticity is quite large. Therefore, the maximum rating ratio is determined by confirming the proportion of the length of the internal line segments. Based on the rotation process of the corresponding dividing line, its perpendicular line also rotates. Thus, the dividing line and the perpendicular line can confirm the specific elliptical state of the monomorphic cell and confirm the specific rating ratio. This allows for the initial screening of the corresponding cells to be tested. Further analysis of the cells to be tested is needed to determine whether they are tumor cells, which requires verification analysis of the internal cell nuclei.

[0051] In the feature region labeling stage, the SobeL algorithm is used to sequentially identify other contour regions within the cell under test. From these identified contour regions, the contour region associated with the cell nucleus is located and its feature regions are specifically labeled. The specific labeling method is as follows:

[0052] The SobeL algorithm and the set threshold are used to sequentially confirm other contour regions existing in the cell to be examined. Different contour regions correspond to different thresholds, and the values ​​of the thresholds are determined by the operator based on experience.

[0053] The mean value of several sets of gray values ​​associated with different pixels in different contour areas is processed to confirm the regional features associated with the corresponding contour areas. Contour areas whose regional features belong to a set interval are marked as suspected regions. The set interval is a preset interval, which is pre-determined based on the gray value features of the cell nucleus, and generally takes the value [80, 100].

[0054] If there is only one suspected region, then this suspected region is marked as a feature region;

[0055] If multiple suspected regions exist, the radial ratio of these regions is used to confirm their existence through a verification unit, followed by feature analysis to pinpoint the feature region. The specific method for pinpointing the region is as follows:

[0056] Based on the edge contour of the suspected region, the center point of the region belonging to this suspected region is located. Then, using the same processing method as for confirming the cell to be examined in unimorphic cells, dividing lines and perpendicular lines are constructed within the suspected region. The evaluation ratio belonging to this suspected region is confirmed, and the largest evaluation ratio is selected from several sets of evaluation ratios. The largest evaluation ratio is recorded as the standard feature of this suspected region. From several sets of standard features, the value closest to 1 is selected, and the suspected region associated with the selected value is recorded as the feature region. Specifically, the corresponding cell nucleus is a spherical body. In most biological cells, the cell nucleus is shaped like a sphere. The nucleus appears spherical or nearly spherical. This is because a spherical structure allows the nucleus to have the largest volume within a limited space, which is conducive to the uniform distribution of genetic material (such as chromosomes) within the nucleus and provides a relatively stable environment for various physiological activities within the nucleus. For example, the nuclei of human liver cells and nerve cells are mostly spherical, and the rating associated with their spherical shape is generally the closest value to 1. Inside the cell body, apart from the nucleus being the closest to a circle, other contour areas are irregular regions, so 1 can be directly used as the associated numerical feature for selection.

[0057] The tumor cell identification and labeling end identifies spiky segments on the edge contour of the marked characteristic regions within the cell to be examined, and confirms whether the cell to be examined is a tumor cell based on the overall proportion of these spiky segments on the overall edge contour. The specific method for confirmation is as follows:

[0058] Combination Figure 2 Based on the edge contour of the feature region, the feature center point belonging to this feature region is identified (by combining the edge contour with a two-dimensional coordinate system, the two-dimensional coordinates of different points within the contour are identified, and then the average of several two-dimensional coordinates is calculated to identify the mean coordinates. The identified mean coordinates are then marked within the feature region to identify the feature center point). A set of contour points is randomly selected from the edge contour and recorded as the starting point. The angle between the starting point and the adjacent points is then determined in a clockwise direction.

[0059] Connect the starting point to the feature center point to confirm the first angle line. Then connect the starting point to the adjacent point to confirm the second angle line. Confirm the angle J between the first angle line and the second angle line. If 85°≤J≤95°, then mark the part of the contour segment between the starting point and the adjacent point as the normal segment. Otherwise, mark it as the burr variation segment.

[0060] Then, using the same processing method, the normal segment or burr change segment is marked for the subsequent contour segments in sequence, until the overall edge contour of the feature area is completely marked.

[0061] Confirm the total length L1 of several spur change segments, and then confirm the total length L2 of the feature region edge contour. If (L1÷L2)≤0.2, no calibration is performed. Otherwise, the cell to be examined is calibrated as a tumor cell and calibrated simultaneously in the grayscale image.

[0062] Specifically, the difference between tumor cells and normal cells lies in the fact that the nucleus of tumor cells has a distinct region. In normal cells, the nucleus is relatively round and smooth, without any protrusions or spiky features. However, the nucleus of tumor cells has a large number of spiky protrusions on its surface. These features are clearly different from those of normal cells, so tumor cells can be quickly identified.

[0063] Some of the data in the above formulas are numerical calculations with dimensions removed, and the contents not described in detail in this specification are all prior art known to those skilled in the art.

[0064] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A computer vision-based bone marrow morphology recognition system for multiple myeloma, characterized in that, include: The high-definition microscopic image conversion end converts the acquired high-definition microscopic images into grayscale images. On the grayscale feature processing end, the SobeL algorithm is used to confirm the gradient features associated with different pixels in the grayscale image. Based on the gradient features corresponding to different pixels in the grayscale image, cell outlines are confirmed and cell selection is performed to identify individual cells. The specific method is as follows: The SobeL algorithm was used to determine the lateral gradient X associated with different pixels in the grayscale image. i and vertical gradient S i Where i represents different pixels, using: Confirm the gradient feature ZH corresponding to this pixel. i ; Will satisfy ZH i Pixels with a value greater than or equal to Y1 are labeled as gradient pixels; otherwise, no labeling is performed and Y1 is a preset value. Based on the gradient pixels identified sequentially within the grayscale image, several adjacent gradient pixels are connected to form a line outline, and the individual associated with this line outline is recorded as a single cell. This line outline belongs to the edge outline of this single cell. The polymorphic cell analysis end identifies the contour regions within the cell monomers based on the different cell monomers marked in the grayscale image. Based on the number of contour regions, the cell monomers are labeled as monomorphic or polymorphic cells, and the cells to be examined are identified from the monomorphic cells. In the feature region labeling stage, the SobeL algorithm is used to sequentially confirm other contour regions within the cell under test, lock the contour region associated with the cell nucleus, and then perform specific feature region labeling again. The specific method is as follows: The SobeL algorithm and the set threshold are used to sequentially confirm other contour regions existing in the cells to be examined, with different contour regions corresponding to different thresholds; The mean value of several sets of gray values ​​associated with different pixels in different contour areas is processed to confirm the regional features associated with the corresponding contour areas. Contour areas whose regional features belong to a set interval are marked as suspected regions, and their set interval is a preset interval. If there is only one suspected region, then this suspected region is marked as a feature region; If there are multiple suspected areas, the radial ratio of the multiple suspected areas is confirmed by the verification unit, and then feature analysis is performed to lock the feature area. The tumor cell identification and labeling end identifies the spiky variation segments on the edge contour of the characteristic region based on the marked characteristic region within the cell to be examined. Based on the comprehensive proportion of the spiky variation segments on the overall edge contour, it confirms whether the cell to be examined is a tumor cell and displays the labeling.

2. The computer vision-based bone marrow morphology recognition system for multiple myeloma according to claim 1, characterized in that, The high-definition microscopic image conversion terminal converts grayscale images in the following specific way: Based on the confirmed high-resolution photomicrograph, the RGB values ​​associated with the corresponding pixel points are confirmed. Then, the grayscale value HD associated with the corresponding pixel point is confirmed using the formula: HD=0.299R+0.587G+0.114B. The grayscale value is adjusted based on the grayscale value associated with different pixel points to convert the high-resolution photomicrograph into a grayscale image.

3. The computer vision-based bone marrow morphology recognition system for multiple myeloma according to claim 1, characterized in that, The specific method for identifying cell monomers as unimorphic or polymorphic cells in the polymorphic cell analysis terminal is as follows: The SobeL algorithm was used to identify the lateral gradient X associated with different pixels within different cell monomers. k and vertical gradient S k Where k represents different pixels within a single cell, using: Confirm the gradient feature ZH corresponding to this pixel. k ; Will satisfy ZH k Pixels with a value greater than or equal to Y2 are labeled as gradient pixels; otherwise, no labeling is performed, and Y2 is a preset value, where Y2 < Y1. Based on the gradient pixels identified sequentially within a single cell, several adjacent gradient pixels are connected to identify the corresponding contour region. If there is only one contour region within a single cell, the single cell is labeled as a monomorphic cell; if there are multiple contour regions within a single cell, it is labeled as a polymorphic cell.

4. The computer vision-based bone marrow morphology recognition system for multiple myeloma according to claim 3, characterized in that, The specific method for identifying the target cell from unimorphic cells in the polymorphic cell analysis terminal is as follows: Based on the edge contour associated with the monomorphic cell, the cell center point belonging to this monomorphic cell is first identified. A set of contour points on the edge contour is randomly selected and recorded as principal points. The principal points are connected to the cell center point to confirm a set of lines. This line is then extended to the other side of the edge contour to confirm the first set of dividing lines. A perpendicular line belonging to this dividing line is then constructed, with both endpoints of the perpendicular line located on the edge contour. The evaluation ratio of the dividing line and the perpendicular line is confirmed: the evaluation ratio = the length of the longer line segment ÷ the length of the shorter line segment, where the longer line segment is the longest line segment between the dividing line and the perpendicular line, and the same applies to the shorter line segment. The dividing line is rotated based on this cell center point. Based on the rotation process, the maximum evaluation ratio is confirmed. If the maximum evaluation ratio is ≥1.5, this monomorphic cell is labeled as the cell to be tested; otherwise, no labeling is performed.

5. A computer vision-based bone marrow morphology recognition system for multiple myeloma according to claim 4, characterized in that, The specific method by which the verification unit confirms the radial ratio is as follows: Based on the edge contour of the suspected area, the center point of the area belonging to this suspected area is located. Then, the same processing method used to confirm the cells to be tested for single-state cells is adopted to construct the dividing line and vertical line in the suspected area, and the evaluation ratio belonging to this suspected area is confirmed. From several sets of evaluation ratios, the largest evaluation ratio is selected and recorded as the standard feature of this suspected area. From several sets of standard features, the value closest to 1 is selected, and the suspected area associated with the selected value is recorded as the feature area.

6. The computer vision-based bone marrow morphology recognition system for multiple myeloma according to claim 1, characterized in that, The tumor cell identification and calibration terminal confirms whether the cell under test is a tumor cell in the following specific way: Based on the edge contour of the feature region, identify the feature center point belonging to this feature region, and randomly select a set of contour points from the edge contour as the starting point. Then, determine the angle of the contour segment between the starting point and the adjacent points in a clockwise direction. Connect the starting point to the feature center point to confirm the first angle line. Then connect the starting point to the adjacent point to confirm the second angle line. Confirm the angle J between the first angle line and the second angle line. If 85°≤J≤95°, then mark the part of the contour segment between the starting point and the adjacent point as the normal segment. Otherwise, mark it as the burr variation segment. Then, using the same processing method, the normal segment or burr change segment is marked for the subsequent contour segments in sequence, until the overall edge contour of the feature area is completely marked. Confirm the total length L1 of several spur change segments, and then confirm the total length L2 of the feature region edge contour. If (L1÷L2)≤0.2, no calibration is performed. Otherwise, the cell to be examined is calibrated as a tumor cell and calibrated simultaneously in the grayscale image.

Images