Cell image segmentation method

Abstract

The invention discloses a cell image segmentation method, and relates to the technical field of cell analysis, and the method comprises the following steps: S1, image acquisition; s2, processing optical artifacts generated by imaging of the phase contrast microscope; s3, processing interference of impurities and debris in the cell image; in the method, a phase contrast microscope and an optical coherence tomography microscope are combined to obtain a cell image, three-dimensional structure information of an OCT image is utilized to assist in correcting artifacts generated by imaging of the phase contrast microscope, and the artifacts are processed based on phase recovery and filtering and texture feature matching and repairing. According to the method, the interference of impurities and debris is removed by applying a diffusion and repair method based on a partial differential equation and a machine learning classification and reconstruction method, so that the cell image quality is effectively improved, the interference of artifacts and impurities on subsequent analysis is reduced, and a reliable image data basis is provided for cell image segmentation and cell morphology analysis.

Description

Technical Field

[0001] The present invention relates to the technical field of cell analysis, in particular to a cell image segmentation method. Background Art

[0002] In the research of cell biology and other related fields, accurate segmentation of cell images plays a vital role in gaining a deeper understanding of cell structure and function and conducting many subsequent analysis tasks (such as cell morphology research, cell physiological process monitoring, etc.). However, there are currently many technical challenges in acquiring cell images and performing image segmentation.

[0003] When using phase contrast microscopy to obtain images of cells during division, the imaging principle of phase contrast microscopy itself has limitations. It uses the phase difference of light to enhance the contrast. In this process, some optical artifacts (artifacts, aperture, etc.) are easily generated, resulting in uneven image quality and interfering with subsequent segmentation. For example, the aperture may blur the edge of the cell, and the artifact may be misjudged as a cell structure, increasing the difficulty of segmentation.

[0004] Moreover, it is difficult to achieve an absolutely pure cell culture environment. Impurities in the solution and tissue debris around the cells will be recorded during imaging, and it is difficult to completely remove them before acquiring the image. Therefore, the acquired cell images contain additional interfering components such as solution impurities and tissue debris. These interfering factors will result in poor original quality of the cell image, leading to subsequent segmentation based on image features, which may easily mistake artifacts, impurities, etc. as cell parts, or make cell boundaries unclear.

[0005] In view of this, a cell image segmentation method is provided to overcome the above problems. Summary of the invention

[0006] The object of the present invention is to provide a cell image segmentation method to solve the problems raised in the above background technology.

[0007] In order to solve the above technical problems, the present invention provides a cell image segmentation method, comprising the following steps: S1. Image acquisition: Use a combination of phase contrast microscopy and optical coherence tomography microscopy to acquire images, and use the three-dimensional structural information of OCT images to assist in the correction of phase contrast microscopy images; S2. Processing of optical artifacts generated by phase contrast microscopy imaging: including collaborative processing based on phase recovery and frequency domain characteristics, and deep matching and intelligent restoration based on texture features; Based on the collaborative processing of phase recovery and frequency domain characteristics: including the introduction of phase recovery and system characteristics fusion algorithm, frequency domain filtering guided by three-dimensional structure information; that is, using iterative algorithms to process artifacts in the frequency domain combined with imaging system characteristics; Deep matching and intelligent restoration based on texture features: including accurate extraction and analysis of multi-scale texture features, image restoration based on intelligent matching and dynamic restoration; that is, restoration from the perspective of texture using a pre-established standard texture feature database; S3. Processing impurities and debris interference in cell images: including the construction of anisotropic diffusion model based on cell morphology perception, level set image restoration guided by cell prior knowledge, multi-dimensional feature depth extraction and classifier training, and impurity region reconstruction based on feature distribution perception.

[0008] Furthermore, in the introduction of the phase recovery and system characteristic fusion algorithm: For phase contrast microscopy images, a two-dimensional fast Fourier transform is first performed to accurately convert the image from the spatial domain to the frequency domain; On this basis, by deeply analyzing the phase information in the frequency domain, combining the unique optical transfer function characteristics of the phase contrast microscope imaging system, and using the iterative phase recovery algorithm, the phase distortion caused by phase difference imaging is accurately estimated and restored. During each iteration, not only is the image amplitude information strictly constrained in the spatial domain according to the original image amplitude, but the phase information is also finely updated in the frequency domain according to the imaging system characteristics and the current iteration status. Through multiple iterations, the real artifact-free phase distribution is gradually approached, thereby achieving accurate correction of artifacts such as aperture and artifacts caused by phase anomalies in the image.

[0009] Furthermore, in the frequency domain filtering guided by the three-dimensional structure information: With the help of the three-dimensional information of the cell interior and surrounding environment provided by OCT images, the spectral characteristics of the cell images are analyzed in the frequency domain. Based on the adaptive frequency domain filter guided by the OCT three-dimensional structure, the frequency components related to artifacts are accurately identified according to the information of the cell interior and surrounding environment revealed by the OCT images. Then, the frequency components related to the identified artifacts are efficiently suppressed while retaining the frequency information corresponding to the cell structure to the greatest extent; Finally, the filtered frequency domain image is converted back to the spatial domain through inverse fast Fourier transform to obtain an image with preliminary artifacts removed. At this time, the edges and internal structures of the cells are clearer, effectively reducing the interference of artifacts such as artifacts and aperture on subsequent analysis.

[0010] Furthermore, in the accurate extraction and analysis of multi-scale texture features: A composite method combining gray-level co-occurrence matrix and local binary pattern is used to calculate the texture feature vectors of cell images at different scales. A texture feature database is constructed by performing the same texture feature extraction operation on a large number of standard myocardial cell images without artifacts that have been pre-acquired under simulated ideal imaging conditions and strictly verified by professional humans.

[0011] Furthermore, in image restoration based on intelligent matching and dynamic restoration: Intelligently match the texture feature vector of the phase contrast microscope image with artifacts with the standard features in the texture feature database, and find the most similar standard texture feature and its corresponding standard image area through a matching strategy based on a machine learning algorithm; Then, using the texture information of these standard image areas, with the help of an image restoration algorithm based on block matching and with dynamic adjustment capabilities, similar texture blocks are replaced or fused to the corresponding areas of the current image with artifacts, and the cell texture parts damaged by artifacts such as artifacts and aperture are finely repaired, further significantly improving the image quality, making the overall cell structure more complete and clear, and more in line with the real cell morphology.

[0012] Furthermore, in the construction of anisotropic diffusion model based on cell morphology perception: An anisotropic diffusion model based on partial differential equations is constructed for cell images. The diffusion coefficient of the model is adaptively adjusted according to the gradient information of the image and the morphological characteristics of the cells. In the direction of the cell boundary, diffusion is inhibited according to the clarity and morphological characteristics of the cell boundary; while inside the cell and in the impurity area, diffusion is not inhibited according to the continuity of the internal cell structure and the distribution of impurities.

[0013] Furthermore, in the level set image restoration guided by cell prior knowledge: First, based on the shape of the cell, the internal structural characteristics of the cell, and known prior knowledge, a level set function is initialized to accurately represent the segmentation contours of the cell area and the background area; Then, by defining an energy functional on the level set function, the variational method is used to solve the minimum value of the energy functional, driving the level set curve to evolve according to the morphological characteristics and internal structure information of the cell, so that the curve gradually approaches the real cell boundary, while filling the defective area, restoring the complete morphology of the cell, and completely removing the interference of impurities and debris; The energy functional includes but is not limited to a fidelity term based on image data, a regularization term that controls the smoothness of the level set curve according to cell morphology characteristics, and a guidance term based on the internal structure of the cell.

[0014] Furthermore, in multi-dimensional feature depth extraction and classifier training: Extracting features of multiple dimensions from cell images, including but not limited to implicit color-related features mined by color space transformation methods, shape features obtained by calculating geometric features of cell contours, texture features obtained by fusion texture feature extraction methods, and multi-scale features based on wavelet transform; A deeply optimized support vector machine classifier is trained using a large number of labeled image samples containing impurities and pure cell areas, which can accurately distinguish cell areas from impurity areas with the help of multi-dimensional features.

[0015] Furthermore, in the impurity region reconstruction based on feature distribution perception: Input the cell image to be processed into the trained SVM classifier to obtain the accurate classification result of whether each pixel belongs to the cell area or the impurity area; For pixels determined to be impurity areas, a reconstruction method based on interpolation and texture synthesis based on multi-dimensional feature perception is used according to the multi-dimensional feature distribution of pixels in the surrounding cell areas. The impurity areas are filled with pixel values ​​and textures based on multi-dimensional feature perception that match the surrounding cell areas. This allows the entire cell image to remove impurities while maintaining the consistency and authenticity of the cell structure.

[0016] Compared with the prior art, the present invention has the following beneficial effects: Reduce artifact interference: By acquiring images using a combination of phase contrast microscopy and optical coherence tomography microscopy, and using the method of collaborative processing of phase recovery and frequency domain characteristics, the artifacts generated by phase contrast microscopy imaging are accurately corrected and suppressed from the phase angle and frequency domain with the help of the three-dimensional structural information of the OCT image. At the same time, based on the method of deep matching of texture features and intelligent repair, the cell texture part destroyed by artifacts is further repaired from the perspective of texture features, making the edge and internal structure of the cell clearer, effectively reducing the interference of artifacts such as artifacts and aperture on subsequent analysis, avoiding problems such as blurred cell edges and artifacts being misjudged as cell structures due to artifacts, improving image quality, and laying the foundation for accurate segmentation of cell images.

[0017] Removal of impurities and debris interference: In dealing with impurities and debris interference in cell images, a diffusion and repair method based on the collaboration of partial differential equations and cell morphology is used to construct an anisotropic diffusion model to adaptively adjust the diffusion coefficient according to the cell morphology. It can effectively remove small impurity particles while retaining the sharpness of cell boundaries. For larger defect areas, a level set image repair method guided by cell prior knowledge is used to fill and restore them. In addition, based on the method of deep classification and fine reconstruction of machine learning, the support vector machine classifier trained with multi-dimensional features is used to accurately distinguish between cell areas and impurity areas, and then the impurity areas are filled through reconstruction methods based on interpolation and texture synthesis based on multi-dimensional feature perception. While removing impurities, the continuity and authenticity of the cell structure are maintained, solving the problem of blurred cell boundaries and mistaking impurities for cell parts due to interfering components such as solution impurities and tissue debris, further improving the original quality of cell images.

[0018] Combination and utilization of imaging equipment: Use a phase contrast microscope and an optical coherence tomography microscope (OCT) to acquire images, and make full use of the three-dimensional structural information of the OCT image to assist in the correction of the phase contrast microscope image when processing artifacts. From the perspective of the utilization of imaging principles, an effective method is achieved. The phase recovery algorithm is combined with the optical transfer function characteristics of the phase contrast microscope imaging system and the multi-dimensional information provided by the OCT image for processing, showing the segmentation advantage of cell images.

[0019] Processing artifacts and impurities: The methods based on phase recovery and filtering and the methods based on texture feature matching and repair are used to conduct in-depth processing from the perspectives of phase, texture, etc. that are more in line with the characteristics of cell images; in terms of processing impurities, the diffusion and repair methods based on partial differential equations and the methods based on machine learning classification and reconstruction are used. By building complex and adaptive models and targeted reconstruction methods, and using image registration fusion and deep learning models to process artifacts, it is possible to more effectively deal with the interference problems of artifacts and impurities in cell images. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is a schematic diagram of the cell image segmentation method of the present invention. DETAILED DESCRIPTION

[0021] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0022] See also Figure 1 , the present invention provides a technical solution: Example of a cell image segmentation method: Experimental materials and equipment: Cell samples: Cardiomyocyte samples in the division period are selected and cultured in a cell culture medium whose ingredients have been strictly screened and ultra-fine filtered; this is to prevent initial impurities from mixing into the cell samples to the greatest extent possible, laying the foundation for the subsequent acquisition of high-quality cell images.

[0023] Microscope equipment: A unique combination of phase contrast microscopy and optical coherence tomography (OCT) is used to obtain cell images. Phase contrast microscopy is used to obtain preliminary images of cells during division. However, due to the inherent limitations of its imaging principle, the images obtained will inevitably have imaging artifacts such as apertures and artifacts caused by the phase difference of light. At the same time, impurities and tissue debris around the cells will interfere with the image because the culture environment is difficult to achieve absolute purity. Optical coherence tomography (OCT) performs tomographic imaging of the same cardiomyocyte sample from different angles and depths. It can provide extremely clear and detailed information on the internal structure of the cell and the cell and the surrounding environment at a three-dimensional level. This information will play a key auxiliary role in the subsequent processing of phase contrast microscopy image artifacts and impurity interference.

[0024] Computing Devices: The computer is equipped with a high-performance graphics processor (GPU) with powerful computing power, and an image processing algorithm library customized for the image processing needs of this experiment is installed. At the same time, the computer is also installed with analysis software specifically used for medical image processing, which has the function of in-depth analysis of cell image characteristics.

[0025] Implementation steps: S1. Image acquisition: First, the prepared cardiomyocyte samples were imaged using a phase contrast microscope to obtain original images of the cells during division. Due to the imaging principle of the phase contrast microscope, these original images must contain artifacts such as apertures and artifacts caused by the phase difference of light, and there are impurities and tissue debris around the cells because the culture environment is not completely pure, which interferes with the image quality.

[0026] At the same time, the optical coherence tomography (OCT) microscope was started synchronously to perform tomography imaging of the same cardiomyocyte sample from multiple different angles and depths. The acquired OCT images can provide rich information on the internal structure of cells and the relationship between cells and the surrounding environment at the three-dimensional level. This information will serve as an important auxiliary basis for the subsequent processing of phase contrast microscopy imaging artifacts and cell image impurity interference.

[0027] S2. Dealing with optical artifacts produced by phase contrast microscopy imaging: Method based on coordinated processing of phase recovery and frequency domain characteristics: Phase recovery and system characteristic fusion algorithm: For phase contrast microscope images, a two-dimensional fast Fourier transform (2DFFT) is first performed to accurately convert the image from the spatial domain to the frequency domain. On this basis, by deeply analyzing the phase information in the frequency domain and combining the optical transfer function (OTF) characteristics unique to the phase contrast microscope imaging system, an iterative phase recovery algorithm is used to accurately estimate and restore the phase distortion caused by phase difference imaging. In each iteration, not only is the amplitude information of the image strictly constrained in the spatial domain based on the amplitude of the original image, but the phase information is also finely updated in the frequency domain based on the characteristics of the imaging system and the current iteration status. Through multiple iterations, the real artifact-free phase distribution is gradually approached, thereby achieving accurate correction of artifacts such as aperture and artifacts caused by phase anomalies in the image.

[0028] Frequency domain filtering guided by three-dimensional structural information: With the help of detailed information on the inside and surrounding environment of cells provided by OCT images at the three-dimensional level, the spectral characteristics of cell images are further analyzed in the frequency domain. Through an adaptive frequency domain filter guided by the three-dimensional structure of OCT (different from the common notch filter based on the statistical characteristics of the local spectrum, it takes into account the three-dimensional spatial information provided by the OCT image), according to the internal structure of the cell and the surrounding environment information revealed by the OCT image, the frequency components related to artifacts are accurately identified (by analyzing the characteristics such as spectral energy distribution and frequency peak, and combining the three-dimensional spatial position relationship in the OCT image to determine the abnormal frequency segment). Then, these identified artifact-related frequency components are efficiently suppressed, while retaining the frequency information corresponding to the cell structure to the greatest extent. Finally, the filtered frequency domain image is converted back to the spatial domain through the inverse fast Fourier transform (IFFT) to obtain an image with preliminary artifact removal. At this time, the edges and internal structures of the cells are clearer, effectively reducing the interference of artifacts such as artifacts and apertures on subsequent analysis.

[0029] Method based on texture feature depth matching and intelligent restoration: Precise extraction and analysis of multi-scale texture features: Perform multi-scale, all-round texture feature extraction operations on phase contrast microscope images. A composite method that integrates the improved gray-level co-occurrence matrix (GLCM) and the local binary pattern (LBP) is used to carefully calculate the texture feature vectors of cell images at different scales, comprehensively covering multiple aspects of texture information such as contrast, correlation, energy, and direction. By performing the same texture feature extraction operation on a large number of standard cardiomyocyte images without artifacts that have been pre-acquired under simulated ideal imaging conditions and strictly verified by professional humans, a texture feature database with rich content and precise features is constructed. From the construction method of this database and the texture feature extraction method used, it is concluded that this database pays more attention to the in-depth mining of the texture features of cell images.

[0030] Image restoration based on intelligent matching and dynamic restoration: Intelligently match the texture feature vector of the phase contrast microscope image with artifacts with the standard features in the texture feature database, and find the most similar standard texture features and their corresponding standard image areas through a matching strategy based on a machine learning algorithm. Then, using the texture information of these standard image areas, with the help of an image restoration algorithm based on block matching and with dynamic adjustment capabilities (the restoration parameters can be dynamically adjusted according to the local features of the image), similar texture blocks are cleverly replaced or merged into the corresponding areas of the current image with artifacts, and the cell texture parts damaged by artifacts such as artifacts and apertures are finely restored, further significantly improving the image quality, making the overall cell structure more complete and clear, and more in line with the real cell morphology.

[0031] It is necessary to add that: Algorithm principle: The core of this image restoration algorithm based on block matching and with dynamic adjustment capabilities is to use the similarity between image blocks to repair areas damaged by artifacts (artifacts, apertures, etc.). It is based on the assumption that in a cell image, despite the influence of artifacts caused by imaging, the texture and other features of the unaffected normal area have certain repeatability and similarity. By finding similar image blocks, the texture features of these intact areas can be replaced or merged into the damaged area to achieve restoration.

[0032] The difference between this algorithm and conventional image restoration algorithms based on block matching is that it has dynamic adjustment capabilities. Conventional algorithms often use fixed matching rules and restoration parameters, such as fixed block size, fixed similarity measurement criteria, and fixed fusion weights. When faced with cell images that have complex textures and are interfered by artifacts, the restoration effect may not be good. However, this algorithm can dynamically adjust the restoration parameters according to the local features of the image. It can flexibly change the size of the matching block, select the most appropriate similarity measurement method, and adjust the weights during fusion based on local features such as the texture complexity around the area to be restored, the grayscale changes of the image, and the characteristics of the cell structure.

[0033] For example, in cell images, for local areas with fine textures and clear cell structure boundaries, if it is found that matching with smaller image blocks can more accurately find similar textures, the algorithm will dynamically reduce the size of the matching blocks; at the same time, if the grayscale changes here are relatively gentle, the algorithm may adjust the similarity measurement standard accordingly, focusing more on the similarity judgment of features such as texture direction; when performing texture fusion repair, the fusion weight is dynamically adjusted according to the degree of influence of the surrounding cell structure on the repaired effect, so that the repaired texture is better integrated with the surrounding cell structure, ensuring that the overall morphology of the repaired cells is more natural and realistic, reducing repair traces, and further improving image quality, making it more in line with the image quality requirements of subsequent analysis (such as cell image segmentation, cell morphology analysis, etc.).

[0034] Specific implementation steps: Feature vector extraction: First, local feature vectors are extracted from phase contrast microscope images with artifacts and standard image regions in the texture feature database. For each local region where the pixel to be repaired is located in the image (for example, a rectangular region is defined as an image block with the pixel as the center according to the preset initial block size), the corresponding texture feature vector is calculated according to the established fusion texture feature extraction method (a composite method that integrates the improved GLCM and the local binary pattern (LBP)). This vector covers multiple aspects of texture information such as contrast, correlation, energy, and direction, so as to describe the texture characteristics of the image block.

[0035] Smart matching process: The texture feature vector corresponding to each area to be repaired in the image to be repaired is matched with the standard feature vector in the texture feature database. The matching strategy is based on advanced machine learning algorithms (for example, it can be based on algorithms such as support vector machines or neural networks for feature space mapping and similarity measurement). By calculating the distance between feature vectors (such as Euclidean distance, cosine similarity, etc.) or the matching score output by the machine learning model, the standard texture feature vector and its corresponding standard image area that are most similar to the texture feature vector of the area to be repaired are found. In this process, some matching parameters are dynamically adjusted according to the local features of the image around the area to be repaired. For example, if the surrounding texture changes a lot, appropriately increase the feature vector dimension or adjust the weight parameter in the similarity measurement function to more accurately find similar blocks in complex texture environments.

[0036] Dynamic repair operations: After determining the most similar standard image area, the texture information of these areas is used to repair the area to be repaired. According to the local characteristics such as the size of the area to be repaired, the boundary of the surrounding cell structure, and the continuity of the texture, the repair parameters are dynamically adjusted to perform the repair operation. For example, if the area to be repaired is small and located near the cell boundary, in order to avoid affecting the clarity of the boundary, the size of the texture block used for replacement will be appropriately reduced, and the weight during fusion will be dynamically adjusted according to the grayscale change trend at the boundary, so that the replaced texture can naturally transition to the texture of the surrounding undamaged cell area; if the area to be repaired is large and is in an area with relatively uniform texture inside the cell, the texture block size may be appropriately increased to improve the repair efficiency. At the same time, according to the overall texture direction consistency of the area and other characteristics, the fusion method is optimized to ensure the consistency and authenticity of the repaired texture, so that the texture structure of the entire cell is more complete, clear, and more in line with the real cell morphology.

[0037] Through such a dynamic adjustment process, the image restoration algorithm can better adapt to the complex situations presented by cell images due to artifacts, effectively improve the quality of image restoration, and provide a guarantee for more accurate subsequent cell image-related analysis.

[0038] S3. Dealing with impurities and debris in cell images Diffusion and repair method based on partial differential equations (PDE) and cell morphology: Construction of anisotropic diffusion model based on cell morphology perception: An anisotropic diffusion model based on partial differential equations is constructed for cell images. The diffusion coefficient of this model is not simply set, but is adaptively adjusted according to the gradient information of the image and the morphological characteristics of the cells. Specifically, in the direction of the cell boundary, diffusion is strictly suppressed (to ensure clear boundaries) based on the clarity and morphological characteristics of the cell boundary; while in flat areas such as the interior of the cell and impurity areas, diffusion can be carried out effectively based on the continuity of the internal cell structure and the distribution of impurities. The form of the diffusion equation (taking a two-dimensional image as an example): ; in, Indicates that over time Varying image intensity function, is the gradient operator, is an adaptive diffusion coefficient, which is determined according to the gradient amplitude of the image and the morphological characteristics of the cell, such as: , is a function that is dynamically adjusted according to cell morphology and impurity distribution (it can be designed to be similar to: ; of the form, where Indicates that cells The morphological characteristics of the location, is a threshold parameter determined by experience or experiment). By numerically solving the partial differential equation (a numerical solution such as the finite difference method can be used), the cell image is diffused, which can effectively remove small impurity particles while retaining the sharpness of the cell boundary and significantly reducing impurity interference.

[0039] Level set image restoration based on cell prior knowledge: For large defective areas in cell images caused by impurities and debris (for example, cell boundaries are blocked and disconnected by impurities), a level set image restoration method based on cell prior knowledge is adopted. First, based on the approximate shape of the cell, the internal structural characteristics of the cell, and the known prior knowledge, a level set function is initialized to accurately represent the segmentation contours of the cell area and the background area. Then, by defining an energy functional on the level set function (this energy functional includes a fidelity term based on image data, a regularization term that controls the smoothness of the level set curve according to the cell morphological characteristics, and a guidance term based on the internal structure of the cell, which is significantly different from the energy functional in the existing patent), the variational method is used to solve the minimum value of the energy functional, and the level set curve is driven to evolve according to the morphological characteristics and internal structural information of the cell, so that the curve gradually approaches the real cell boundary, while filling the defective area, perfectly restoring the complete morphology of the cell, and completely removing the interference of impurities and debris.

[0040] Methods based on deep classification and fine reconstruction of machine learning: Deep extraction of multi-dimensional features and fine training of classifiers: Extract multi-dimensional features from cell images, including but not limited to implicit color-related features mined through innovative color space transformation methods (although cell images are usually grayscale images, this method can obtain richer color-related information), shape features obtained by accurately calculating the geometric features of cell contours (such as perimeter, area, circularity, etc.), texture features obtained by the above-mentioned fusion texture feature extraction method (such as GLCM and LBP features), and multi-scale features based on wavelet transform, etc. A large number of labeled image samples containing impurities and pure cell areas (these samples are labeled through professional manual labeling combined with simulated generation of impurity images, and the labeling process strictly follows cell biology expertise) are used to train a deeply optimized support vector machine (SVM) classifier, so that it can accurately distinguish between cell areas and impurity areas with multi-dimensional features, and the training method of this classifier and the feature extraction method used are significantly different from the classifier training in existing patents.

[0041] Impurity region reconstruction based on feature distribution perception: The cell image to be processed is input into the trained SVM classifier to obtain the accurate classification result of each pixel belonging to the cell region or the impurity region. For the pixels determined to be impurity regions, according to the multi-dimensional feature distribution of the pixels in the surrounding cell regions (by statistically analyzing the feature mean, variance, covariance and other information of the cell pixels in the neighborhood), a reconstruction method based on multi-dimensional feature perception interpolation and texture synthesis is adopted to fill the impurity region with pixel values ​​and textures based on multi-dimensional feature perception that match the surrounding cell regions, so that the entire cell image can remove impurities while maintaining the coherence and authenticity of the cell structure.

[0042] Summarize: Reduce artifact interference: By acquiring images using a combination of phase contrast microscopy and optical coherence tomography (OCT) microscopy, and using the method of collaborative processing of phase recovery and frequency domain characteristics, the artifacts (such as aperture, artifacts, etc.) generated by phase contrast microscopy imaging are accurately corrected and suppressed from the phase angle and frequency domain with the help of the three-dimensional structural information of the OCT image. At the same time, based on the method of deep matching of texture features and intelligent repair, the cell texture part destroyed by artifacts is further repaired from the perspective of texture features, making the edge and internal structure of the cell clearer, effectively reducing the interference of artifacts such as artifacts and aperture on subsequent analysis, and avoiding problems such as blurred cell edges and artifacts being misjudged as cell structures due to artifacts, thereby improving image quality and laying the foundation for accurate segmentation of cell images.

[0043] Removal of impurities and debris interference: In dealing with impurities and debris interference in cell images, a diffusion and repair method based on partial differential equations (PDE) and cell morphology is used to construct an anisotropic diffusion model to adaptively adjust the diffusion coefficient according to the cell morphology. It can effectively remove small impurity particles while retaining the sharpness of cell boundaries. For larger defect areas, a level set image repair method guided by cell prior knowledge is used to fill and restore them. In addition, based on the method of deep classification and fine reconstruction of machine learning, the support vector machine (SVM) classifier trained with multi-dimensional features is used to accurately distinguish between cell areas and impurity areas, and then the impurity areas are filled through interpolation and texture synthesis based on multi-dimensional feature perception. While removing impurities, the continuity and authenticity of the cell structure are maintained, solving the problem of blurred cell boundaries and mistaking impurities for cell parts due to interfering components such as solution impurities and tissue debris, further improving the original quality of cell images.

[0044] Combination and utilization of imaging equipment: Use a phase contrast microscope and an optical coherence tomography microscope (OCT) to acquire images, and make full use of the three-dimensional structural information of the OCT image to assist in the correction of the phase contrast microscope image when processing artifacts. From the perspective of the utilization of imaging principles, an effective method is achieved. The phase recovery algorithm is combined with the optical transfer function characteristics of the phase contrast microscope imaging system and the multi-dimensional information provided by the OCT image for processing, showing the segmentation advantage of cell images.

[0045] Processing artifacts and impurities: The methods based on phase recovery and filtering and the methods based on texture feature matching and repair are used to conduct in-depth processing from the perspectives of phase, texture, etc. that are more in line with the characteristics of cell images; in terms of processing impurities, the diffusion and repair methods based on partial differential equations (PDE) and the methods based on machine learning classification and reconstruction are used. By building complex and adaptive models and targeted reconstruction methods, and using image registration fusion and deep learning models (such as GAN, CNN, etc.) to process artifacts, it is possible to more effectively deal with the interference problems of artifacts and impurities in cell images.

[0046] Assist in subsequent cell-related research: Through the above-mentioned effective solution to the problems of cell image artifacts and impurity interference, the quality of cardiomyocyte images has been significantly improved, providing a more reliable image data foundation for subsequent cell-related research work (such as cell image segmentation, cell morphology analysis, etc.), which will help improve the accuracy and scientific nature of these research works and promote the further development of the field of cell research.

Claims

1. A cell image segmentation method, characterized in that: The following steps are involved: S1. Image acquisition: Use a combination of phase contrast microscopy and optical coherence tomography microscopy to acquire images, and use the three-dimensional structural information of OCT images to assist in the correction of phase contrast microscopy images; S2. Processing of optical artifacts generated by phase contrast microscopy imaging: including collaborative processing based on phase recovery and frequency domain characteristics, and deep matching and intelligent restoration based on texture features; Based on phase recovery and frequency domain characteristics collaborative processing: This includes the introduction of phase recovery and system characteristic fusion algorithms, and frequency domain filtering guided by three-dimensional structural information; that is, using iterative algorithms to process artifacts in the frequency domain in combination with imaging system characteristics; Deep matching and intelligent restoration based on texture features: including accurate extraction and analysis of multi-scale texture features, image restoration based on intelligent matching and dynamic restoration; that is, restoration from the perspective of texture using a pre-established standard texture feature database; S3. Processing impurities and debris interference in cell images: including the construction of anisotropic diffusion model based on cell morphology perception, level set image restoration guided by cell prior knowledge, multi-dimensional feature depth extraction and classifier training, and impurity region reconstruction based on feature distribution perception.

2. The cell image segmentation method according to claim 1, characterized in that: In the introduction of phase recovery and system characteristics fusion algorithm: For phase contrast microscopy images, a two-dimensional fast Fourier transform is first performed to accurately convert the image from the spatial domain to the frequency domain; Analyze the phase information in the frequency domain, combine the optical transfer function characteristics unique to the phase contrast microscope imaging system, and use the iterative phase recovery algorithm to estimate and restore the phase distortion caused by phase difference imaging; During each iteration, not only is the image amplitude information strictly constrained in the spatial domain according to the original image amplitude, but the phase information is also finely updated in the frequency domain according to the imaging system characteristics and the current iteration status.

3. The cell image segmentation method according to claim 1, characterized in that: In frequency domain filtering guided by three-dimensional structural information: With the help of the three-dimensional information of the cell interior and surrounding environment provided by OCT images, the spectral characteristics of the cell images are analyzed in the frequency domain. Based on the adaptive frequency domain filter guided by the three-dimensional structure of OCT, the frequency components related to artifacts are identified according to the information of the cell interior and surrounding environment revealed by the OCT images. Then, the frequency components related to the identified artifacts are suppressed while retaining the frequency information corresponding to the cell structure; Finally, the filtered frequency domain image is converted back to the spatial domain through inverse fast Fourier transform to obtain an image with preliminary artifacts removed.

4. The cell image segmentation method according to claim 1, characterized in that: In the accurate extraction and analysis of multi-scale texture features: A composite method combining gray-level co-occurrence matrix and local binary pattern is used to calculate the texture feature vectors of cell images at different scales. A texture feature database is constructed by performing the same texture feature extraction operation on a large number of standard myocardial cell images without artifacts that have been pre-acquired under simulated ideal imaging conditions and strictly verified by professional humans.

5. The cell image segmentation method according to claim 1, characterized in that: In image restoration based on intelligent matching and dynamic restoration: Intelligently match the texture feature vector of the phase contrast microscope image with artifacts with the standard features in the texture feature database, and find the most similar standard texture feature and its corresponding standard image area through a matching strategy based on a machine learning algorithm; Then, using the texture information of these standard image areas, with the help of an image restoration algorithm based on block matching and with dynamic adjustment capabilities, similar texture blocks are replaced or merged into the corresponding areas of the current image with artifacts.

6. The cell image segmentation method according to claim 1, characterized in that: In the construction of anisotropic diffusion model based on cell morphology perception: An anisotropic diffusion model based on partial differential equations is constructed for cell images. The diffusion coefficient of the model is adaptively adjusted according to the gradient information of the image and the morphological characteristics of the cells. In the direction of the cell boundary, diffusion is inhibited according to the clarity and morphological characteristics of the cell boundary; while inside the cell and in the impurity area, diffusion is not inhibited according to the continuity of the internal cell structure and the distribution of impurities.

7. The cell image segmentation method according to claim 1, characterized in that: In level set image restoration guided by cell prior knowledge: First, based on the shape of the cell, the internal structural characteristics of the cell, and known prior knowledge, a level set function is initialized to accurately represent the segmentation contours of the cell area and the background area; Then, the energy functional is defined on the level set function; the variational method is used to solve the minimum value of the energy functional, and the level set curve is driven to evolve according to the morphological characteristics and internal structure information of the cell; The energy functional includes but is not limited to a fidelity term based on image data, a regularization term that controls the smoothness of the level set curve according to cell morphology characteristics, and a guidance term based on the internal structure of the cell.

8. The cell image segmentation method according to claim 1, characterized in that: In multi-dimensional feature depth extraction and classifier training: Extracting features of multiple dimensions from cell images, including but not limited to implicit color-related features mined by color space transformation methods, shape features obtained by calculating geometric features of cell contours, texture features obtained by fusion texture feature extraction methods, and multi-scale features based on wavelet transform; A deeply optimized support vector machine classifier is trained using a large number of labeled image samples containing impurities and pure cell areas, which can accurately distinguish cell areas from impurity areas with the help of multi-dimensional features.

9. The cell image segmentation method according to claim 1, characterized in that: In feature distribution-aware impurity region reconstruction: Input the cell image to be processed into the trained SVM classifier to obtain the accurate classification result of whether each pixel belongs to the cell area or the impurity area; For pixels determined to be impurity areas, according to the multi-dimensional feature distribution of pixels in the surrounding cell areas, a reconstruction method based on interpolation and texture synthesis based on multi-dimensional feature perception is adopted to fill the impurity areas with pixel values ​​and textures based on multi-dimensional feature perception that match the surrounding cell areas.

Images