A bone marrow biopsy image blur removing method for cell structure recovery

Abstract

The application relates to the technical field of image data processing, and discloses a bone marrow biopsy image blur removing method for cell structure recovery, which comprises the following steps: firstly, reliable low-frequency information in a blurred image is purified as a structural blueprint by applying a low-pass filter, and cell clusters in the structural blueprint are adaptively decoupled; then, the statistical characteristics of the structural blueprint are used as instructions to guide a programmed algorithm to redraw idealized cell structures; finally, errors caused by artifacts are iteratively removed by analyzing the redrawing residuals, the reliable low-frequency information in the image is used as control instructions, high-frequency details are generated in a forward construction mode, and therefore the risk of introducing error structures caused by guessing in a traditional repair method is avoided, the method realizes adaptive processing of cell clusters and artifacts through an inherent check and correction mechanism, and the logical integrity and final reliability of the reconstruction result in biology are guaranteed.

Description

TECHNICAL FIELD

[0001] The application relates to a marrow biopsy image blur removal method for cell structure recovery, and belongs to the technical field of image data processing. BACKGROUND

[0002] The core idea of the current technical solution for image blur problems has long been based on a common theoretical foundation, that is, through mathematical model reverse deduction and repair of the degraded image signal in the imaging process, so as to maximize the approximation of a physically clear original image assumed to exist at the pixel level. Whether it is a traditional method based on deconvolution or a deep learning method based on data driving in recent years, the essence is optimization and evolution within the signal repair framework. This approach has achieved recognized technical effects in processing natural landscape or macro object images.

[0003] However, when this classic processing method is applied to a special type of information composed of extremely complex biomedical images, such as marrow biopsy images processed through a series of complex physical and chemical processes such as slicing, fixing, and staining, the inherent principle limitation begins to appear. The reason is that the blur of this type of image is not simply optical diffraction or motion defocus. It is a process of physical irreversible damage and entropy increase of high-frequency biological structure information such as chromatin texture in the nucleus. At this time, still insisting on the goal of true restoration at the pixel level will inevitably lead the entire processing flow into a fundamental dilemma, that is, forcibly reconstructing high-frequency details in the region of incomplete information. The result is bound to be a speculation based on algorithm model, rather than restoration. This speculation is extremely easy to generate an algorithm illusion that looks clear but is completely wrong biologically in the absence of sufficient constraints. This introduces a misdiagnosis risk much higher than the information loss caused by image blur itself for clinical applications that rely on morphology for accurate diagnosis.

[0004] Specifically, the prior art mainly has the following deficiencies: 1. The traditional inverse operation method based on a simplified mathematical model, such as a fuzzy kernel, is difficult to accurately model the complex nonlinear physical and chemical degradation process, and the repair result is often accompanied by ringing effect and noise amplification, which destroys the original reliable low-frequency macrostructure information in the image; 2. Although the modern deep learning method has strong learning ability, under the premise of lacking a large amount of, high-quality, and strictly paired clear and fuzzy bone marrow biopsy image dataset, the generated high-frequency details are easy to deviate from the true biological characteristics, and there is a serious risk of repairing the morphological characteristics of one kind of cell into another kind of cell; 3. The existing technical path regards the cell image as a non-discriminatory pixel matrix for processing, and lacks a way to integrate the prior knowledge about cell morphology mastered by experts in the field into the image processing process in a non-data-driven and principled manner, thereby providing a reliable constraint for the generation of high-frequency information. Therefore, how to avoid the dilemma between pixel fidelity and diagnostic reliability when processing such physically damaged biological images, and develop a method that does not aim to restore the truth, but uses the reliable macro information remaining in the image as a guide to directly construct a clear and unambiguous cell morphology image in a deterministic manner consistent with biological prior knowledge, has become a technical problem to be solved by the present application. SUMMARY

[0005] The present application provides a bone marrow biopsy image blur removal method for cell structure recovery, which mainly aims to solve the problem that the prior art is trapped in the dilemma between pixel fidelity and diagnostic reliability when processing physically damaged biological images due to the pursuit of pixel-level repair, and cannot effectively improve the image clarity while ensuring the accuracy of biological characteristics.

[0006] To achieve the above purpose, the present application provides a bone marrow biopsy image blur removal method for cell structure recovery, which comprises the following steps:

[0007] Step a, Gaussian blur processing is applied to the input blurred bone marrow biopsy image and threshold segmentation is performed to obtain a plurality of structure blueprint regions representing cell positions and macroscopic contours; then, the mean and standard deviation are calculated based on the area distribution of the plurality of structure blueprint regions, and an abnormal fusion blueprint region with an area exceeding the mean plus a constant coefficient multiplied by the standard deviation is identified; for each abnormal fusion blueprint region, a plurality of local brightness minimum points are found as seed points in the corresponding region of the input blurred bone marrow biopsy image, and the seed points are used to divide the abnormal fusion blueprint region into a plurality of independent sub-blueprint regions by a watershed algorithm;

[0008] Step b, for any structural blueprint region or sub-blueprint region obtained after step a, first, smooth continuous edges are calculated based on the geometric moments of the region; then, the average gray value and the gray variance of the region in the input blurred bone marrow biopsy image are extracted as texture control parameters, and the texture density and contrast of the programmed texture generation algorithm are set according to the texture control parameters, and internal textures are generated in the smooth continuous edges;

[0009] Step c, all internal textures generated in step b and the corresponding smooth continuous edges are combined and fused with the image background to generate a preliminary sharpening image; then, the difference between the preliminary sharpening image and the input blurred bone marrow biopsy image is calculated to generate a redrawn residual image; then, the gradient energy of the residual region corresponding to each combination in the redrawn residual image is analyzed, and the energy threshold is determined based on the gradient energy distribution of all residual regions; if there is a residual region whose gradient energy exceeds the energy threshold, the structural blueprint region corresponding to the residual region is removed from the current processing flow, and steps b and c are executed based on the remaining structural blueprint regions; if the gradient energy of all residual regions does not exceed the energy threshold, the current preliminary sharpening image is output as the final sharpening image.

[0010] Preferably, in the internal texture generation step of step b, the texture control parameters are subjected to adaptive normalization based on the image background matrix before being used to set the texture density and contrast of the programmed texture generation algorithm; the adaptive normalization is specifically: first, identify the background matrix region other than all structural blueprint regions, and calculate its global average gray value ; then, the average gray value is compensated by the formula to obtain the effective gray value , which is used to set the texture density, where is a background gray reference system constant.

[0011] Preferably, in the internal texture generation step of step b, the generation process of the internal texture is spatially modulated; specifically: first, the corresponding part of the input blurred bone marrow biopsy image in the structural blueprint region or independent sub-blueprint region is extracted, and its gray value is subjected to linear mapping processing to generate a two-dimensional texture control field; then, when executing the programmed texture generation algorithm, for each pixel point in the region, the texture density and contrast parameters of the algorithm are dynamically determined by the pixel value at the corresponding position in the texture control field.

[0012] Preferably, the Gaussian kernel size and the standard deviation value used in the Gaussian blur process are set to a range sufficient to smooth the high frequency texture details within a single nucleus, but insufficient to cause two independent cells with a distance of one standard cell diameter to be recognized as a single connected region after segmentation.

[0013] Preferably, the procedural texture generation algorithm is a fractal noise algorithm, and the relationship between the average gray value and the texture density is set to an inverse relationship, and the relationship between the gray variance and the contrast is set to a proportional relationship.

[0014] Preferably, the constant coefficient is a value greater than 2; the energy threshold is determined by calculating the cumulative distribution function of the gradient energy of all residual regions, and setting the energy value at the 95th percentile position point of the function as the energy threshold.

[0015] Preferably, the smooth continuous edge is obtained by performing energy functional minimization iteration on the initial contour of the structural blueprint region or the contour of the independent sub-blueprint region using the active contour model.

[0016] Preferably, the threshold segmentation is an adaptive threshold segmentation using Otsu method.

[0017] Preferably, the fusion adopts Poisson image editing technology to ensure that the transition region between the generated internal texture and the corresponding smooth continuous edge and the image background is continuous in the gradient domain.

[0018] Preferably, the internal texture generated by the procedural texture generation algorithm has an inverse relationship between the texture density and the pixel value at the corresponding position in the texture control field, and a proportional relationship between the contrast and the pixel value at the corresponding position in the texture control field.

[0019] Compared with the prior art, the present application has the following beneficial effects:

[0020] 1. By applying low-pass filtering to the blurred image to obtain a structural blueprint containing only position and contour macro information, this eliminates unreliable high-frequency details and noise in the image as interference, rather than attempting to repair them; subsequently, the geometric and gray statistical characteristics of the structural blueprint region are used as control parameters to guide a procedural texture generation process to draw new cell structures with idealized edges and internal textures at the original position, which converts the blurred low-frequency macro region information in the original image into instructions for subsequent high-frequency detail deterministic generation, so that the clarity of the final image no longer depends on the reverse deduction of lost information, but is derived from a positive construction process based on reliable information guidance, thereby avoiding the technical risk of introducing incorrect structures due to guesswork repair when processing incomplete biological images in the prior art.

[0021] 2、The technical scheme of the present application further comprises a mechanism for checking and correcting the intermediate product itself during the processing process. When the structural blueprint extraction step produces an abnormal fusion blueprint area with a size exceeding the normal statistical range, the process does not stop or directly process, but the abnormal size itself is used as a trigger signal. This signal calls the local brightness minimum point information corresponding to the area in the original blurred image. These residual detail information, which is smoothed out in the strong filtering, is now used as the seed point of the watershed algorithm to divide the abnormal fusion blueprint area into multiple independent sub-blueprint areas. This nonlinear processing path that calls different information level data across different processing stages makes the method adaptively switch from macroscopic morphological recognition to more detailed structure decoupling when facing complex scenes such as high-density cell clusters, thereby restoring the clarity of individual cells while maintaining the accuracy of the original spatial topological relationship and counting information of the cell population.

[0022] 3、The present application further extends the robustness of the image processing flow from dealing with optical blur itself to recognizing and calibrating common physical artifacts in the manufacturing process. On the one hand, after generating the final image, the system calculates the difference between it and the original input image to obtain a redrawn residual image, and analyzes the residual structural features of each redrawn area. A residual area that presents strong structural features will cause its corresponding initial blueprint to be identified as a low-confidence area generated by non-cellular structure errors such as staining artifacts, and will be excluded in subsequent iterative redrawing. On the other hand, the system also identifies the background matrix area widely existing in the image, and uses its average gray level as an internal reference to measure the overall staining level of the current slice, to normalize the gray scale statistical properties of each cell blueprint. The combination of these two mechanisms enables the entire image processing flow not only to have error correction ability to verify the authenticity of the generated results, but also to have adaptive ability to the staining process fluctuations. All judgments and calibrations rely on the data of the image itself, without any external prior knowledge or training data set. BRIEF DESCRIPTION OF DRAWINGS

[0023] Fig. 1 The logical flowchart of the blurred removal and iterative correction method of the present application;

[0024] Fig. 2 The functional deployment architecture diagram of the image processing server of the present application;

[0025] Fig. 3 The dynamic interaction timing diagram of the core processing module of the present application. DETAILED DESCRIPTION

[0026] In order to make the technical solutions and advantages of the present application clearer, the technical solutions of the present application will be described in detail below in conjunction with specific embodiments, but it should be understood that these descriptions are only to show the preferred embodiments of the present application and are not intended to limit the protection scope of the present application.

[0027] The present application discloses a bone marrow biopsy image blur removal method for cell structure recovery, which is configured as a sequentially executed image data processing flow, including three main stages of structure blueprint extraction and decoupling, programmed redrawing based on blueprint guidance, and iterative checking and image fusion based on redrawing residual error, which work together to convert the input blurred bone marrow biopsy image, a two-dimensional pixel matrix, into a cell morphology image with clear structure; when processing a digital bone marrow biopsy image whose cell nucleus boundary is unclear and internal texture details are lost due to the combined action of optical diffraction and chemical staining, existing processing methods may introduce false structures when attempting to reversely recover the damaged high-frequency information, to deal with this situation, the present application first applies Gaussian blur processing to the input blurred bone marrow biopsy image in step a, this processing is a controlled low-pass filtering operation that smooths and filters out unreliable high-frequency noise and blurred details in the image, thereby purifying low-frequency information containing only cell location, approximate size and outline; the numerical setting of the Gaussian kernel size and standard deviation for this Gaussian blur processing is based on a technical trade-off, that is, the values should be sufficient to smooth the high-frequency texture inside a single cell nucleus, but not enough to identify two independent cells with a standard cell diameter as a single connected region in subsequent segmentation, a procedure for determining this parameter is as follows: first, obtain the average diameter of the standard cell under the current imaging system through image analysis , then set the standard deviation to a value that satisfies , such as , and the Gaussian kernel size is correspondingly set to cover the range of ; after Gaussian blur processing, the system immediately uses the Otsu adaptive threshold segmentation technique to binarize the image, generating multiple binary connected domains with pixel values of 0 or 1, these connected domains are the structure blueprint areas representing cell location and outline, this operation materializes the low-frequency information in the original image into a discrete data structure that can be directly called by subsequent processing steps.

[0028] Furthermore, in the face of high-density cell clusters that may cause multiple adjacent cells to be identified as a single connected domain after low-pass filtering, the system starts a structure decoupling processing flow after obtaining the structure blueprint area set, which first calculates the statistical mean and standard deviation of the area distribution of multiple structure blueprint areas, and identifies structure blueprint areas with an area exceeding the mean plus a constant factor abnormal fusion blueprint region, i.e. the region satisfying where the constant coefficient is set to a value greater than 2, for example 2.5, to ensure statistical significance of the identification; for each identified abnormal fusion blueprint region, the system returns to the corresponding region of the input fuzzy bone marrow biopsy image, searches for a plurality of local intensity minima in the region, and before performing the structural decoupling of the abnormal fusion blueprint region in step a, the system first performs a signal verification and seed point refinement operation on the corresponding part of the input fuzzy bone marrow biopsy image, which first calculates the contrast-to-noise ratio of the region where is the average gray level in the region, and are the average gray level and the gray level standard deviation of the global background of the image, respectively, only when the value of exceeds a minimum processable threshold value determined in the system calibration phase, the subsequent processing is triggered; then, after identifying all the local intensity minima in the region, the system calculates the depth of each of the minima, i.e. the gray level difference between the point and the overflow point of the attractor basin in which it is located, and only keeps the minima with a depth greater than the gray level standard deviation of the region multiplied by a calibration coefficient as the final set of seed points for the watershed algorithm; these minima, corresponding to the staining centers of the individual cells in the fuzzy cell clusters, serve as markers indicating the positions of the individual cells, and the system then uses these local intensity minima as seed points to perform the watershed algorithm on the binary mask of the abnormal fusion blueprint region to segment the abnormal fusion blueprint region into a plurality of independent sub-blueprint regions, in this way, a morphological abnormality feature is used as a control signal to trigger a fine structural segmentation, thereby helping to maintain the accuracy of cell counting and the independence of individual morphology in high-density regions.

[0029] The internal texture generation stage in step b aims to solve the technical problem of how to generate a new cell structure with clear edges and reasonable internal texture for a blueprint region containing only contour information, for this, for each structural blueprint region or independent sub-blueprint region obtained in step a, the system first performs edge redrawing, specifically, the system calculates the center of mass and the principal axis direction of the region based on its geometric moments, and based on this, performs energy functional minimization iteration through the active contour model to finally converge to a smooth continuous edge curve; in order to determine the internal energy parameters of the active contour model used in step b in the edge fitting process of different cell morphologies, the system first calculates the area of the region when processing each structural blueprint region or independent sub-blueprint region, with the perimeter and a dimensionless morphological complexity factor is derived based on this and the elastic parameter will be adjusted according to this factor, with the specific adjustment relationship being and wherein, is the morphological complexity factor of the current region, and and is the reference coefficient value determined by processing the reference data set covering the typical cell morphology and minimizing the edge fitting error during the system offline parameter calibration stage; then, the system performs internal texture redrawing, which converts the reliable gray scale statistical properties in the original blurred image into instructions for controlling the programmed texture generation algorithm. The system first extracts the average gray value of the blueprint region in the input blurred bone marrow biopsy image and the gray scale variance as the initial texture control parameters. To cope with the possible gray value drift caused by dyeing process fluctuations between different slices, the system also uses an adaptive normalization process based on the image background matrix. This process first identifies the background matrix region by performing a difference set operation on the entire image mask and all structural blueprint region masks, and calculates its global average gray value Then, the average gray value is compensated by the formula to obtain the effective gray value wherein is a preset reference system constant representing the background gray value under ideal dyeing conditions. This normalization process uses as an internal reference to eliminate the impact of dyeing differences between slices on the absolute value of and thus improve the stability of as a biological state representation.

[0030] Subsequently, the system uses a fractal noise algorithm as the programmed texture generation algorithm, and uses the normalized effective gray value and the gray scale variance to set the texture density and contrast of the algorithm. The relationship between the average gray value and the texture density is set to be inversely proportional, while the relationship between the gray scale variance and the contrast is set to be directly proportional. This means that a more deeply colored original blurred region (lower value) has a larger internal gray scale variation (higher Cells with higher values (i.e. higher intensity) will be redrawn as structures with thicker and higher-contrast internal textures; to reproduce the spatial heterogeneity of complex nuclear internal textures, the internal texture generation process is configured to be spatially modulated, specifically, the system first extracts the corresponding part of the input blurred bone marrow biopsy image within the structural blueprint region or independent sub-blueprint region, and linearly maps its grayscale values, for example, normalized to a two-dimensional texture control field, and when executing the fractal noise algorithm, for each pixel point in the region, the algorithm's texture density and contrast parameters are dynamically determined by the pixel value at the corresponding position in the texture control field, where the region with higher pixel value in the texture control field corresponds to the generation of low-frequency and high-amplitude fractal noise, while the region with lower pixel value corresponds to the generation of high-frequency and low-amplitude fractal noise, in this way, the local grayscale information of the original image is used as the parameter field to control the generation process, so that the redrawn process can reproduce the texture distribution pattern related to the original cell state.

[0031] Finally, in step c, the system combines all the internal textures and the corresponding smooth continuous edges generated in step b, and fuses them with the image background to generate a preliminary clear image, the technical problem faced in this fusion process is to avoid the generation of artificial seams between the newly generated structures and the background, for this purpose, the system uses the Poisson image editing technique, by solving a Poisson equation to ensure that the transition area between the generated internal texture and the corresponding smooth continuous edge and the image background is continuous in the gradient domain, so as to realize smooth fusion; At the same time, common physical artifacts in the production process, such as dye precipitated particles, may be mistakenly identified as structural blueprints and redrawn as morphologically clear false cells due to their size and density characteristics. To deal with this situation, the present application establishes a verification and iterative correction mechanism based on redrawn residuals. After generating a preliminary clear image, the system first calculates the difference between the preliminary clear image and the input blurred bone marrow biopsy image to generate a redrawn residual image. The judgment basis is that the correctly redrawn true fuzzy cells have residual mainly showing diffuse unstructured morphology, while the incorrectly redrawn physical artifacts have residual showing strong structured features. Accordingly, the system then analyzes the gradient energy of the residual area corresponding to each combination in the redrawn residual image, and determines an adaptive energy threshold based on the gradient energy distribution of all residual areas. The determination procedure of the threshold is as follows: calculate the cumulative distribution function of the gradient energy of all residual areas, and set the energy value at the 95th percentile position point of the function as the energy threshold. If there is a residual area with gradient energy exceeding the energy threshold, the system removes the initial structural blueprint area corresponding to the residual area from the set to be processed, and returns to perform the idealized structure redrawing of step b and the image fusion of step c. This iterative loop continues until the gradient energy of all remaining residual areas does not exceed the energy threshold. At this time, the system outputs the current preliminary clear image. Through this closed-loop feedback, the processing flow can verify and correct the intermediate results, thereby improving the reliability of the final output image.

[0032] In an image data processing task, a digitized bone marrow biopsy image is inputted, which presents composite degenerative features: firstly, the image suffers from mixed blur due to non-ideal conditions of optical imaging and sample preparation process, leading to diffusion of cell boundaries and loss of high-frequency detail information such as chromatin texture inside the nucleus; secondly, there is a high-density cell cluster in a key region of the image, with multiple cells physically close together, and the blurred image signals of which are spatially overlapped; thirdly, there is a physical artifact formed by dye crystallization next to the cell cluster, which has similar size and gray value to the nucleus, and its edge is sharper than that of the real blurred cell, which brings technical challenges to image processing, i.e. the conventional operation aiming to sharpen the blurred cell will equally enhance the artifact, while the operation aiming to smooth the artifact will exacerbate the blur of the cell cluster; to process this image, the aforementioned technical solution is deployed and executed, after the process is started, firstly, step a is executed, a Gaussian blur processing is applied to the input blurred bone marrow biopsy image, which filters out all high-frequency information in the original image, including blurred boundaries between cells, unreliable textures inside the cells and sharp edges of the physical artifact, and only retains the low-frequency spatial position and contour information of each object; after adaptive threshold segmentation by Otsu method, the system obtains a set of structure blueprints, among which the high-density cell cluster is identified as an abnormal fused blueprint region with an area far exceeding the mean value, and the physical artifact is identified as an independent structure blueprint region with a size similar to that of the normal cell blueprint; at this time, the system triggers the structure decoupling mechanism, which takes the identified abnormal fused blueprint region as the processing object, calls the pixel data of the original blurred image in this region, and locates multiple local brightness minimum points as seed points, and then uses these seed points to segment the abnormal fused blueprint region into multiple independent sub-blueprint regions by watershed algorithm, here, the Gaussian blur processing provides a prerequisite for the execution of the structure decoupling mechanism, the former unambiguously identifies the complex region that needs to be analyzed in detail by filtering out high-frequency details, and the latter relies on this identification to call back the higher-resolution information in the original image to complete the segmentation of this region.

[0033] Subsequently, the system enters the step b of generating internal texture stage, for each structure blueprint region and the segmented sub-blueprint region, according to its geometric moment and gray statistical characteristics, a new cell structure with smooth continuous edge and programmed internal texture is generated in the original position, and enters step c, all the newly generated structures and the background are fused by Poisson image editing technology to generate a preliminary clear image, in this image, each cell in the original cell cluster position is independently and clearly redrawn, but at the same time, the structure blueprint transformed from the physical artifact is also redrawn as a structure with smooth edge and internal texture; At this time, the process does not end, but starts the iterative checking mechanism based on the redrawn residual, the system differentiates the preliminary clear image from the original input blurred image to generate a redrawn residual image, and analyzes the gradient energy of each redrawn region; The analysis result shows that the area redrawn by the real blurred cell presents a diffuse low gradient energy form, while the structure redrawn by the physical artifact presents a strong, structured and high gradient energy feature, because there is structural inconsistency between the high-frequency sharp edge of the artifact and the smooth edge of the newly generated cell; The system determines the energy threshold according to the gradient energy cumulative distribution function of all residual regions, and removes the structure blueprint generated by the artifact from the set to be processed when the gradient energy exceeds the threshold, then the system takes the purified blueprint set as input and re-executes step b and step c; In the final output image, the high-density cell cluster is analyzed as multiple independent and complete cell units, the total number of cells and the spatial topological relationship are maintained, and the position of the physical artifact in the original image is smoothly processed as a background region without introducing misleading cell structures. This processing flow converts the blurred image enhancement into a forward construction process based on low-frequency information guidance, and introduces a structure consistency-based check after construction, thereby addressing the technical challenge between image clarity improvement and structure accuracy guarantee under the condition of incomplete information.

[0034] Example 2: To objectively evaluate the performance of the technical solution of the present invention under different image degradation conditions, this example constructs and executes a comparative numerical experiment. The experimental results show that when processing images with blurred cell clusters or artifacts, the output results of the method of the present invention are superior to two representative existing technical methods in terms of structural accuracy and object recognition integrity. This experiment uses a benchmark dataset containing 100 clear, high-resolution bone marrow biopsy images as a reference, and applies three standardized digital degradation methods to each benchmark image in a procedural manner to generate test samples with known true values: Class A test images are subjected to only medium-intensity Gaussian blur; Class B test images are subjected to mixed blur and high-density cell clusters are artificially created in the image; Class C test images are subjected to mixed blur and high-density cell clusters are artificially created in the image. In this experiment, physical artifacts with typical staining precipitate characteristics were placed in the image. One sample group and two control groups were set up. Control group A used a Wiener filtering-based deconvolution algorithm, while control group B used a U-Net network pre-trained on a general image set. The experiment used multi-dimensional metrics for evaluation. In terms of peak signal-to-noise ratio (PSNR) and structural similarity metrics, which measure pixel-level fidelity, the two control groups performed better than the sample group. For example, in the C-class test image, control group B had a PSNR of 26.1 dB, while the sample group had 22.8 dB, which is consistent with the technical path of this invention, which does not aim for pixel-level accurate restoration. However, different results were observed in metrics measuring structural accuracy. The structural integrity score of the sample group showed a higher score. The Integrity Score (SIS) remained above 0.93 in all test categories, while the SIS value of the control group dropped below 0.65 when processing Class B images containing clusters. The performance difference was even more significant in object recognition and artifact suppression. When processing Class B images, the object detection rate of the sample group of this invention was 96.5%, while that of the control groups A and B was 65.3% and 72.8%, respectively. This difference is mainly due to the structural decoupling mechanism included in the technical solution of this invention. When processing Class C images containing artifacts, the artifact suppression rate (ASR) of the sample group of this invention reached 97.2%, while that of the control groups A and B was 15.4% and 23.7%, respectively. This performance difference is directly attributed to the iterative verification machine based on redrawing residuals included in the technical solution of this invention. The experimental data are summarized in Table 1.

[0035] Table 1: Performance comparison of different image processing methods on different degraded images.

[0036]

[0037] The experimental data shows that the technical solution of the present invention sacrifices some pixel-level similarity to improve the performance of cell object detection, segmentation, morphological reconstruction and artifact suppression. This result confirms that the present invention, by abandoning the reverse recovery of unreliable high-frequency information and instead adopting a forward construction based on reliable low-frequency information and combined with iterative verification, can generate images that are more accurate and reliable in terms of structural logic, thereby providing higher quality input data for subsequent image analysis tasks.

[0038] Example 3: This example combines Figs. 1 to 3 This paper describes a method for blur removal in bone marrow biopsy images aimed at restoring cell structure. Fig. 1 As shown, the system begins with a blurry bone marrow biopsy image containing cell clusters and artifacts. It then enters a structural blueprint extraction stage, using low-pass filtering to purify cell location and macroscopic contour information. Based on this, the system makes a judgment. If an abnormal fusion blueprint region is identified, an adaptive decoupling process for cell clusters is initiated, using a watershed algorithm to segment the fusion region into independent sub-blueprints. Subsequently, all independent blueprints or sub-blueprints enter a blueprint-guided procedural redrawing stage to generate new cell structures with idealized edges and internal structures. Next, in the image fusion stage, the redrawn structure is seamlessly fused with the background to generate a preliminary clear image. Then, the system enters an iterative verification stage based on redrawing residuals. It identifies and eliminates erroneous redrawing caused by artifacts by calculating the residual gradient energy. A core condition is that the gradient energy of all residual regions does not exceed a threshold. If not, the erroneous blueprint is eliminated and the system returns to the procedural redrawing stage, forming a closed-loop correction iteration until all regions meet the condition. Finally, a clear and unambiguous clear cell morphology image is output.

[0039] like Fig. 2As shown, the figure takes the digital microscope imaging system as the source of the image, transmits the original image to the core image processing server, which contains an image receiving module, which functions to receive the input blurred bone marrow biopsy image data; a core processing engine, which functions to execute the complete algorithm process of the application, including extracting and decoupling the structural blueprint, programmatic redrawing based on blueprint guidance, and iterative checking according to the redrawing residual; a result generation module, which functions to fuse the final cell structure that passes the checking with the image background, generate a final image with clear structure and output; during the processing, the server reads and writes data and calls models with the data and model storage server, which centrally manages the original image database, the clear image database, the texture model benchmark library and the system configuration file, and the processed clear image can be read by the clinical reading workstation for subsequent diagnostic analysis by the image reading software, in addition, the architecture also includes a parameter calibration workstation, which updates the models and configurations in the data and model storage server through a special offline parameter calibration program, thereby ensuring the adaptability and continuous optimization of the processing performance of the entire system.

[0040] As shown, Fig. 3 As shown, the entire process is started by the user inputting a blurred bone marrow biopsy image, the image processing system first calls the blueprint extraction module, which sequentially performs operations such as applying Gaussian blur processing, performing threshold segmentation, generating structural blueprint regions, identifying abnormal fusion regions, and performing structural decoupling, and passes the processed blueprint region set to the redrawing module, which then processes each blueprint, including calculating smooth continuous edges, extracting texture control parameters, and generating internal textures, and passes the redrawing results to the checking and fusion module, which is responsible for combining each part with the edges, fusing the image background, generating a preliminary clear image, and further calculating the redrawing residual and analyzing the gradient energy. At this time, if the system determines that there is a residual above the threshold, it will issue a command to the redrawing module to return to processing, forming an iterative correction loop, until all residuals meet the requirements, the checking and fusion module finally outputs the final clear image, and the image processing system returns the processing result to the user, thereby completing a complete image processing session.

[0041] Example 4: In an engineering application, a bone marrow biopsy image deblurring method for cell structure restoration needs to be deployed to a new digital microscope system with different optical resolution and imaging characteristics. To ensure the performance of the algorithm module, a standardized offline parameter calibration procedure needs to be performed to determine a set of working parameters for the new system, thereby eliminating the uncertainty introduced by manual parameter adjustment. To start the calibration procedure, a calibration dataset containing at least 50 representative images collected by the new system is prepared. The images in the dataset need to cover isolated single cells, different degrees of cell cluster aggregation, and physical artifacts generated during the preparation process. The technician needs to label all cells and artifacts in the images as a reference for subsequent calculations.

[0042] The first step of the procedure is to determine the standard deviation used in the Gaussian blur processing in step a The technical trade-off of the set value is that a too small value cannot filter out the internal texture of the cell, thereby interfering with the generation of the structure blueprint, while a too large value will cause adjacent cells to be over-smoothed and incorrectly merged; to determine its value, the system takes the labeled single cell area as the object, iterates within a value interval set according to the average diameter of the standard cell with a step size of In each iteration, the current value is used to blur and segment all isolated cell images, and the average intersection-over-union between the segmented blueprint outline and the labeled outline is calculated. Finally, the value that maximizes the average intersection-over-union is determined as the standard deviation of the imaging system.

[0043] The second step of the procedure is to calibrate the constant coefficient used to identify abnormal merged blueprint areas The technical trade-off of the set value is that a too low value may misjudge a normal single cell with a large size as a cluster, while a too high value may miss a real cluster; to calibrate this coefficient, the system first processes all single cell images in the calibration dataset using the determined in the previous step, and calculates the area mean and standard deviation of all generated structure blueprint areas, then the system processes all images containing cell clusters in the dataset, and iterates different values within an interval of with a step size of 0.1, in each iteration, according to the criterion ​​identify abnormal fusion blueprint area, and calculate the true positive rate and false positive rate of the identified, finally, in order to ensure that the false positive rate is lower than a preset threshold, in this embodiment, the threshold is 1%, the true positive rate is maximized The value is determined as a constant coefficient .

[0044] The third step of the procedure is to verify the energy threshold determination method for removing artifacts in step c. The system uses the calibrated And , a complete processing flow is performed on all images in the calibration data set until the preliminary clarification image and the redrawn residual image are generated. Then, the system respectively calculates the gradient energy of the residual area generated by the real cells and the physical artifacts, and organizes them into two independent distributions. By comparing the two distributions, it can be seen that the gradient energy distribution corresponding to the artifacts is significantly biased towards the high energy area. By mixing the gradient energy of all residual areas and calculating the 95th percentile position point of the cumulative distribution function, an energy threshold can be determined, which can distinguish most of the artifact residual areas from most of the real cell residual areas. After the calibration procedure, the obtained parameter set And the aforementioned energy threshold determination method based on cumulative distribution function are fixed as the configuration file special for this new imaging system, so that in the subsequent regular processing flow, the system can be based on this configuration to run stably and automatically.

[0045] Example 5: Before the image processing method is put into regular use, a one-time offline data calibration procedure needs to be performed to establish the reference model library required for the internal programmed texture generation. The procedure starts with collecting a group of confirmed cell images, covering various cell types including primitive granulocytes and mature lymphocytes. Then, in an interactive parameter adjustment environment, for each cell type, the system operator adjusts the core parameters of the fractal noise algorithm, such as the frequency fractal dimension amplitude and the octave, with reference to the corresponding confirmed image, until the visually generated texture in terms of particle fineness, density and contrast, etc. matches the chromatin distribution pattern of the nucleus in the reference image. The confirmed parameter combination is immediately stored in a lookup table and indexed and associated with the gray mean and variance of the reference cell image. By repeating this process for all cell types, the system finally generates a texture model reference library containing multiple parameter sets that can be called according to the statistical characteristics of the input image.

[0046] To ensure that the adaptive normalization processing based on the image background matrix in the method remains stable under different staining process conditions, the background gray reference system constant calibration, the process requires first adopting a target staining protocol, preparing a batch of reference slides with stable staining characteristics under standardized process control, and digitizing them using the target imaging system; then, the system automatically identifies and extracts all background matrix regions without cell structures from the batch of reference digital images, and calculates the global average of the pixel gray scale values of these regions, which is subsequently stored as a constant corresponding to the specific staining protocol When processing subsequent regular images prepared using the same staining protocol, the system calls this pre-calibrated value to perform normalization calculation so that the gray scale compensation is always based on a stable reference baseline associated with a specific process.

[0047] Example 6: Before deploying to a specific working environment, the method of the present application needs to perform a one-time baseline model establishment and pre-checking procedure to adapt to the imaging and process characteristics of the target environment; the procedure first establishes a reference model library for the programmed texture generation step, which starts with collecting a set of confirmed category cell images covering multiple key cell types, and for each cell type, the system operator adjusts the core parameters of the fractal noise algorithm such as the frequency fractal dimension amplitude and octave in an interactive parameter adjustment environment with the corresponding confirmed image as a reference, until the visually features such as particle fineness, density and contrast of the programmed generated texture match the chromatin distribution pattern of the nucleus in the reference image, the set of confirmed parameter combinations are immediately stored in a lookup table and indexed and associated with the gray scale mean and variance of the reference cell image, by repeating this process for all cell types, a texture model reference library containing multiple sets of parameter sets is finally generated.

[0048] The procedure also includes calibrating the background gray scale reference system constant in adaptive normalization processing, and solidifying the active contour model and signal quality pre-checking logic; for each specific staining work procedure, the operator prepares a batch of reference slides using the procedure under standardized process control, the system then automatically identifies and calculates the global average of the pixel gray scale of all background matrix regions in the batch of slide images, and stores the value as a constant corresponding to the staining procedure ; at the same time, the procedure determines the energy functional of the active contour model used in step b to be a form that minimizes the sum of the pixel gray scale variance inside and outside the contour, and finally, to deal with the case of low input signal quality, the procedure establishes a signal quality pre-checking step that calculates the contrast-to-noise ratio of each input image, if the ratio is lower than a pre-set minimum acceptable threshold, the processing flow is not triggered, and the image is marked as unqualified, this series of pre-checking procedures provides stable baseline models and clear working boundaries for the regular automated operation of the method.

[0049] In the application of processing images containing large volume and complex internal structure of cell nucleus, such as megakaryocytes or primitive cells, the chromatin of such cell nucleus often presents spatial heterogeneity, i.e. part of the region is chromatin loose bright area, and part is chromatin concentration dark area, if the global statistics is used for texture redrawing, the key internal texture distribution pattern will be lost; for this reason, the method of the present application starts the spatial modulation mechanism when performing the internal texture redrawing of step b, which first extracts and normalizes the corresponding part of the original blurred image in the structural blueprint area of the cell into a two-dimensional texture control field, then when generating each pixel by the fractal noise algorithm, the frequency and amplitude parameters are dynamically set by the corresponding value of the pixel in the texture control field, wherein the area with lower value in the control field will generate high frequency and low amplitude noise, simulating fine euchromatin, while the area with higher value will generate low frequency and high amplitude noise, simulating coarse heterochromatin, in this way, the finally redrawn nucleus not only has clear boundary, but also the spatial distribution of internal texture reproduces similar heterogeneity characteristics with the original blurred image, thus retaining more morphological details.

[0050] It is obvious for those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, and the present application can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application.

[0051] Finally, it should be noted that the above examples are only used to illustrate the technical solutions of the present application and are not limiting, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the present application.

Claims

1. A bone marrow biopsy image blur removal method for cell structure recovery, characterized by, The method comprises the following steps: Step a, Gaussian blur processing is applied to the input fuzzy bone marrow biopsy image and threshold segmentation is performed to obtain a plurality of structural blueprint regions representing cell positions and macroscopic contours; then, the mean value and standard deviation are calculated based on the area distribution of the plurality of structural blueprint regions to identify an abnormal fusion blueprint region with an area exceeding the mean value plus a constant coefficient multiplied by the standard deviation; For each abnormal fusion blueprint region, a plurality of local brightness minimum points are found in the corresponding region of the input fuzzy bone marrow biopsy image as seed points, and the seed points are used to divide the abnormal fusion blueprint region into a plurality of independent sub-blueprint regions by a watershed algorithm; Step b, for any structural blueprint region or sub-blueprint region obtained after step a, first, a smooth continuous edge is calculated based on the geometric moment of the region; then, the average gray value and gray variance of the region in the input fuzzy bone marrow biopsy image are extracted as texture control parameters, and the texture density and contrast of the programmed texture generation algorithm are set according to the texture control parameters to generate internal texture in the smooth continuous edge; Step c, all internal textures generated in step b and the corresponding smooth continuous edges are combined and fused with the image background to generate a preliminary clear image; then, the difference between the preliminary clear image and the input fuzzy bone marrow biopsy image is calculated to generate a redrawn residual image; then, the gradient energy of the residual region corresponding to each combination in the redrawn residual image is analyzed, and the energy threshold is determined based on the gradient energy distribution of all residual regions; If there is a residual region with gradient energy exceeding the energy threshold, the structural blueprint region corresponding to the residual region is removed from the current processing flow, and step b and step c are executed based on the remaining structural blueprint regions; If the gradient energy of all residual regions does not exceed the energy threshold, the current preliminary clear image is output as the final clear image.

2. The bone marrow biopsy image blurring removal method for cell structure recovery according to claim 1, wherein, In the step b of generating internal texture, the texture control parameters are subjected to adaptive normalization based on the background matrix of the image before being used to set the texture density and contrast of the programmed texture generation algorithm; the adaptive normalization is specifically as follows: first, the background matrix area other than all the structural blueprint areas is identified, and the global average gray value thereof is calculated ; then, the average gray value is compensated by the formula to obtain the effective gray value which is used to set the texture density, wherein is a background gray reference system constant. 3.The bone marrow biopsy image blurring removal method for cell structure recovery according to claim 1, wherein, In the internal texture generation step of step b, the generation process of the internal texture is spatially modulated; specifically: first, the corresponding part of the input fuzzy bone marrow biopsy image in the structural blueprint region or independent sub-blueprint region is extracted, and the gray value is linearly mapped to generate a two-dimensional texture control field; then, when executing the programmed texture generation algorithm, the texture density and contrast parameters of each pixel point in the region are dynamically determined by the pixel value at the corresponding position in the texture control field.

4. The bone marrow biopsy image blurring removal method for cell structure recovery according to claim 1, wherein, The programmed texture generation algorithm is a fractal noise algorithm, and the relationship between the average gray value and the texture density is set to be inversely proportional, and the relationship between the gray variance and the contrast is set to be proportional.

5. The bone marrow biopsy image blurring removal method for cell structure recovery according to claim 1, wherein, The constant coefficient is a value greater than 2; the energy threshold is determined by calculating the cumulative distribution function of the gradient energy of all residual regions, and the energy value at the 95th percentile position point of the function is set as the energy threshold.

6. The bone marrow biopsy image blurring removal method for cell structure recovery according to claim 1, wherein, The smooth continuous edge is obtained by the active contour model, taking the contour of the structural blueprint region or the independent sub-blueprint region as the initial contour, and performing energy functional minimization iteration.

7. The bone marrow biopsy image blurring removal method for cell structure recovery according to claim 1, wherein, The threshold segmentation is the Otsu adaptive threshold segmentation.

8. The bone marrow biopsy image blurring removal method for cell structure recovery according to claim 1, wherein, The fusion adopts Poisson image editing technology.

9. The bone marrow biopsy image blurring removal method for cell structure recovery according to claim 3, wherein, The internal texture generated by the procedural texture generation algorithm has a texture density inversely proportional to the pixel value of the corresponding position in the texture control field and a contrast proportional to the pixel value of the corresponding position in the texture control field.

Images