Machine learning-based quality assessment method and system for ecms-stem cell constructs

Abstract

The application relates to the technical field of image analysis, and discloses an ECM-stem cell construct evaluation method and system based on machine learning. The method comprises the following steps: acquiring an image sequence of an ECM-stem cell construct through equal-interval focus shooting, and calculating a projection area as a basic morphological index; determining a structure stability distance based on a change rate, and extracting image texture features and spatial distribution at the distance; dividing a complete region of the construct into a plurality of sub-regions, and calculating structure features of texture statistics between the sub-regions; and inputting the structure features into a pre-trained machine learning model, and outputting a quality score of the construct. The application can quickly, non-destructively and objectively evaluate the quality of the ECM-stem cell construct.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image analysis, in particular to an ECM-stem cell construct quality evaluation method and system based on machine learning. BACKGROUND

[0002] With the rapid development of regenerative medicine and tissue engineering technology, ECM-stem cell constructs, as three-dimensional functional structures that simulate the natural tissue microenvironment, have shown broad application prospects in disease treatment, tissue repair, and drug screening. However, their quality evaluation still faces many technical challenges. Current mainstream methods mainly rely on traditional biological detection techniques, such as histological analysis and biochemical component detection. These methods usually require destructive sampling of the construct, which prevents continuous monitoring or subsequent use of the same construct, severely limiting its feasibility in long-term quality tracking and clinical applications.

[0003] ECM-stem cell constructs are key products in tissue engineering and regenerative medicine, and their quality evaluation is a core link to ensure the safety and effectiveness of clinical applications. Currently, this field mainly relies on traditional biological detection methods, which have significant limitations. First, methods such as histological analysis usually require destructive sampling of the construct, making it impossible to continuously monitor and subsequently use the same construct, which makes long-term quality tracking difficult. Second, existing microscopic observation methods highly depend on the subjective experience of the operator, resulting in a lack of objectivity and repeatability in the evaluation results. In addition, conventional methods can only obtain single-dimensional morphological indicators such as projected area, and cannot comprehensively quantify key information such as internal texture and structural heterogeneity from the images. Although machine learning techniques have been introduced into medical image analysis, in the application of this field, how to stably and automatically extract features from conventional bright-field images that can reliably reflect the internal structural complexity of three-dimensional constructs and establish a precise quality evaluation model remains a technical challenge to be solved.

[0004] Therefore, the present application provides an ECM-stem cell construct quality evaluation method and system based on machine learning. SUMMARY

[0005] The embodiments of the present application provide the following technical solutions:

[0006] Step S1, continuously photographing the ECM-stem cell construct at multiple defocus distances at equal intervals to obtain a defocus image sequence, and calculating the projected area of the ECM-stem cell construct as a basic morphological indicator for each image of the defocus image sequence;

[0007] Step S2, calculating the rate of change of the basic morphological indicator with respect to the defocus distance, and taking the defocus distance at which the rate of change first becomes less than a preset threshold as the structural stability distance.

[0008] Step S3: Obtain the defocused image of the ECM-stem cell construct taken at the structurally stable distance, extract the complete region of the ECM-stem cell construct from it, calculate the texture features based on the complete region, obtain the spatial distribution of the texture features in the complete region, divide the complete region into two sub-regions, calculate the statistics of the texture features for each sub-region, and calculate the structural features representing the structural relationship between the sub-regions based on the statistics between different sub-regions.

[0009] Step S4: Pre-train the machine learning model, input the structural features into the machine learning model, and output the quality score of the ECM-stem cell construct by the machine learning model.

[0010] Compared with the prior art, the beneficial effects of the present invention are at least as follows:

[0011] The technical solution provided in this application achieves quality assessment without destructive processing of the ECM-stem cell construct by analyzing image sequences taken at different defocus distances, supporting continuous monitoring and subsequent use of the construct. It employs image processing and machine learning models to automatically extract and analyze morphological and texture features, effectively reducing the differences in subjective human judgment and ensuring the consistency and repeatability of assessment results. By comprehensively considering the spatial distribution of projected area and texture features, as well as the structural relationships between sub-regions, a more comprehensive and in-depth quantitative assessment of the construct's quality is achieved. The structural stability distance is determined by calculating the rate of change of morphological indicators, and features are extracted based on this, ensuring that the obtained features are insensitive to minute changes in focal length, significantly improving data reliability. By dividing the complete region into different sub-regions and calculating their structural features, the functional spatial differences within the construct can be transformed into quantifiable data, providing a more discriminative input for the machine learning model. Model training is conducted using image data with known quality labels taken at stable distances, enabling the final model to accurately predict the quality score of the ECM-stem cell construct based on structural features. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a schematic diagram of an embodiment of a machine learning-based ECM-stem cell construct quality assessment method in this application.

[0014] Figure 2This is a schematic diagram of one embodiment of a machine learning-based ECM-stem cell construct quality assessment system in this application. Detailed Implementation

[0015] This application provides a machine learning-based method and system for quality assessment of ECM-stem cell constructs. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0016] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of a machine learning-based ECM-stem cell construct quality assessment method in this application includes:

[0017] Step S1: Take multiple out-of-focus images of the ECM-stem cell construct at equal intervals to obtain an out-of-focus image sequence. For each image in the out-of-focus image sequence, calculate the projected area of ​​the ECM-stem cell construct as a basic morphological indicator.

[0018] Specifically, the ECM-stem cell construct is a three-dimensional scaffold filled with stem cells and ECM fibers. In bright-field imaging, each defocused image captures the projection of all structural information onto a two-dimensional plane after light passes through different depths of the construct. When the defocus distance is small, the image mainly contains surface information, and a large number of deep structures are not visualized, so the measured projection area is small. As the defocus distance increases, more deep scattering signals are included in the projection, and the measured projection area gradually increases. When the shooting distance is large enough, the light beam has penetrated the entire construct, and its macroscopic outline is maximized on the two-dimensional plane. After that, even if the defocus distance continues to increase, the projection area will not change significantly, but will fluctuate slightly around a stable value. This region is the stable plateau region. Within the stable plateau region, the projection area is not sensitive to small changes in focal length, so the calculated morphological features are reliable.

[0019] To obtain a reliable measurement benchmark for capturing the macroscopic morphology of the ECM-stem cell construct, the ECM-stem cell construct was first photographed continuously at multiple equally spaced defocus distances to obtain a defocus image sequence. The defocus distance refers to the axial distance between the ECM-stem cell construct and the ideal focal plane of the optical imaging system. Multiple defocus distances covered the surface and deep regions of the ECM-stem cell construct to fully capture its complete features. For each image in the defocus image sequence, the projected area of ​​the ECM-stem cell construct was calculated as a basic morphological indicator. The specific calculation method will be explained in detail later.

[0020] Step S2: Calculate the rate of change of the basic morphological index relative to the defocus distance, and take the defocus distance where the rate of change first falls below the preset threshold as the structural stability distance.

[0021] Specifically, in order to determine the structural stability distance that most accurately reflects the true macroscopic structure of the ECM-stem cell construct from multiple images with different defocus distances, when the defocus distance is large enough, even if the defocus distance is further increased, the projected area will no longer change significantly, but will fluctuate slightly around a stable value. This small fluctuation area is the stability zone. Within the stability zone, the projected area is not sensitive to small changes in defocus distance, so the calculated projected area is reliable and stable. As the defocus distance increases, the captured image will also become severely out of focus and blurry. Excessive blurring will lead to a decrease in the accuracy of the extracted boundary. Therefore, in order to find the starting point of the stability zone, that is, the point closest to the focus position that meets the stability conditions, at which stable measurement results can be obtained and the negative impact of optical blurring can be minimized, the rate of change of the basic morphological index relative to the defocus distance is first calculated, that is, the ratio of the change of the basic morphological index with the defocus distance is calculated. The defocus distance at which the rate of change first falls below a preset threshold is taken as the structural stability distance.

[0022] Step S3: Obtain the defocused image of the ECM-stem cell construct taken at the structurally stable distance, extract the complete region of the ECM-stem cell construct from it, calculate the texture features based on the complete region, obtain the spatial distribution of the texture features in the complete region, divide the complete region into two sub-regions, calculate the statistics of the texture features for each sub-region, and calculate the structural features representing the structural relationship between the sub-regions based on the statistics between different sub-regions.

[0023] Specifically, to obtain fundamental data on structural features that can provide support for subsequent quality assessment, defocused images of the ECM-stem cell construct were acquired at a structurally stable distance. The complete region of the ECM-stem cell construct was extracted from the defocused images. After obtaining the complete region, texture features were calculated based on it, and the spatial distribution of these texture features within the complete region was also obtained. The specific methods for calculating texture features and obtaining their spatial distribution within the complete region will be explained in detail later. Subsequently, to transform the functional spatial heterogeneity within the ECM-stem cell construct into quantitative data that can be utilized by machine learning and statistical analysis, the complete region was divided into... The system is divided into two sub-regions. The specific method for dividing the sub-regions will be explained in detail later. In order to reduce the amount of training data for subsequent machine learning, the statistics of texture features are calculated for each sub-region. Based on the statistics between different sub-regions, structural features representing the structural relationships between sub-regions are calculated. A large number of texture feature values ​​are transformed into structural features, and low-level, trivial pixel information is transformed into high-level structural descriptors with clear physical and biological significance. This allows subsequent machine learning models to bypass complex image details and directly learn the core structural rules that determine the quality of ECM-stem cell constructs, thereby achieving a more robust, accurate, and reliable quality assessment.

[0024] The method for extracting the complete region and the steps for identifying white pixel regions in the process of calculating the projected area of ​​the ECM-stem cell construct are the same. First, the image is enhanced at the edges, then a binary image is generated, and then the binary image is optimized. After that, connectivity analysis is performed to filter noise and identify the interconnected white pixel regions, which are then taken as the complete region.

[0025] Step S4: Pre-train the machine learning model, input the structural features into the machine learning model, and output the quality score of the ECM-stem cell construct by the machine learning model.

[0026] Specifically, a machine learning model is pre-trained, which can output a corresponding quality score based on structural features. The previously calculated structural features are input into the pre-trained machine learning model, and the machine learning model outputs the quality score of the ECM-stem cell construct.

[0027] In one specific embodiment, calculating the projected area of ​​the ECM-stem cell construct includes the following steps:

[0028] Edge enhancement is performed on the original defocused image to obtain a boundary-enhanced image. Adaptive threshold segmentation is performed on the boundary-enhanced image to generate a binary image. Morphological closing operation is performed on the binary image to obtain an optimized binary image. Connectivity analysis is performed on the optimized binary image to filter out noise regions, identify all interconnected white pixel regions, count the total number of pixels in the white pixel regions, and use the total number of pixels as the projected area.

[0029] Specifically, to accurately calculate the projected area of ​​the ECM-stem cell construct, the original defocused image is first enhanced at the edges to obtain a boundary-enhanced image. Edge enhancement can use edge enhancement algorithms such as Sobel, Prewitt, or Lapkacian. By enhancing the edges, the subtle contrast differences between the construct and the background, which are difficult for the human eye to distinguish, are transformed into gradient signals that are easy for the algorithm to recognize, significantly highlighting the overall outline of the construct. To clearly divide the boundary-enhanced image into two parts, the construct region and the background region, adaptive thresholding is performed on the boundary-enhanced image to generate a binary image. Adaptive thresholding is different from the global thresholding method. The adaptive threshold is calculated independently for each pixel based on the pixel value distribution of its local neighborhood. The calculation principle is as follows: traverse each pixel in the image, compare its gray value with the local average value of its local neighbors centered on it. If the gray value is lower than the local average value, set it to black; otherwise, set it to white.

[0030] To correct the contour discontinuity caused by imperfect threshold segmentation and to fill the voids that may be caused by texture or noise inside the construct, a complete and smooth construct region mask is obtained. Morphological closing operation is performed on the binary image to obtain an optimized binary image. The closing operation is a continuous operation of linear dilation followed by erosion. The calculation principle is as follows: using a structuring element of a specific shape and size (such as a 3×3 rectangle), a dilation operation is first performed: the structuring element is slid on the binary image. If the structuring element intersects with the white area of ​​the binary image, the center point of the structuring element is set to white. The dilation operation can fill the small voids and breaks in the white area, making its boundary expand outward. Then, an erosion operation is performed: the structuring element is slid on the binary image. Only when the structuring element is completely contained in the white area is the center point of the structuring element set to white. Otherwise, the center point of the structuring element is set to black. The erosion operation can eliminate small isolated noise points and smooth the boundary, offsetting the over-expansion caused by dilation.

[0031] To ensure the specificity of feature extraction and make the calculated area strictly correspond to the ECM-stem cell construct, connectivity analysis was performed on the optimized binary image to identify all interconnected white regions. In addition, noise (such as interference from irrelevant regions like bubbles or the edges of the culture dish) was filtered out according to a preset area threshold, retaining the main region representing the ECM-stem cell construct. Then, the total number of pixels in the white region was counted, and the total number of pixels was used as the projected area of ​​the ECM-stem cell construct.

[0032] In one specific embodiment, the rate of change of the basic morphological index relative to the defocus distance is calculated, and the defocus distance at which the rate of change first falls below a preset threshold is taken as the structural stability distance. This specifically includes the following steps:

[0033] For each pair of adjacent images in the defocus image sequence, their basic morphological indices are obtained, and the rate of change of the two basic morphological indices is calculated to obtain a rate of change sequence. Each rate of change in the rate of change sequence is compared with a pre-set rate of change threshold. The first rate of change that is less than the rate of change threshold is called the target rate of change. The two images corresponding to the target rate of change are obtained, and the two defocus distances corresponding to the two images are obtained. The minimum defocus distance is selected as the structural stability distance.

[0034] Specifically, the greater the defocus distance, the blurrier the captured image becomes due to severe defocus. Therefore, in order to find a stable starting point, that is, the point closest to the focus position that satisfies the stability condition, for each pair of adjacent images in the defocus image sequence, the corresponding basic morphological indicators, that is, their projected areas, are obtained, and the rate of change of the two basic morphological indicators is calculated. Since there are multiple images in the defocus image sequence, multiple rates of change can be obtained to form a rate of change sequence. Each rate of change in the rate of change sequence is compared with a pre-set rate of change threshold, and the first rate of change that is less than the rate of change threshold is obtained. This rate of change is called the target rate of change. A rate of change is calculated based on the basic morphological indicators of two images. Therefore, the two images corresponding to the target rate of change are obtained, and the two defocus distances corresponding to the two images are obtained. The minimum defocus distance is selected as the structurally stable distance. The structurally stable distance represents the optimal balance distance that can obtain stable measurement results and minimize the negative impact of optical blur. In other words, the image captured based on the structurally stable distance is both clear and can accurately reflect the true and complete structure of the ECM-stem cell construct.

[0035] In one specific embodiment, calculating texture features based on the complete region and obtaining the spatial distribution of texture features in the complete region specifically includes the following steps:

[0036] A sliding window is set up and slids within the complete area at a preset step size. For each sliding area block, a first calculation and a second calculation are performed to obtain multiple local texture feature values. These multiple local texture feature values ​​are associated with the center pixel of the corresponding area block to obtain the spatial distribution of the complete area.

[0037] Specifically, texture is a regional feature and cannot be calculated from a single pixel. Therefore, a fixed-size sliding window (e.g., 64×64 pixels) is set. The sliding window starts from the top left corner of the complete region and slides at a preset step size (e.g., 1 pixel). After each slide, the region block corresponding to the sliding window is subjected to a first calculation and a second calculation. The first calculation refers to describing the roughness, contrast, and orderliness of the image by statistically analyzing the spatial relationships between pixels. The process of the first calculation will be explained in detail later. The second calculation refers to transforming the image from the spatial domain to the frequency domain and analyzing its periodic structure and frequency components. The process of the second calculation will be explained in detail later. Multiple local texture feature values ​​are obtained through the first and second calculations. These local texture feature values ​​are the texture features. In order to integrate the calculated local texture feature values ​​into a global image to form a feature map that can be used in subsequent steps, the multiple local texture feature values ​​calculated for each sliding window are assigned to the center pixel of the corresponding sliding window. When the sliding window has traversed the entire complete region, a new image is obtained. Each pixel value of the new image represents the texture feature of the local region centered on that pixel. The above method will transform visual texture features into objective, quantifiable numerical data through local window scanning, providing a data foundation for subsequent analysis and quality assessment.

[0038] In one specific embodiment, the first calculation is performed, which specifically includes the following steps:

[0039] For the current region block, calculate the gray-level co-occurrence matrix of the region block based on the gray value of each pixel in the region block, and calculate the contrast, energy, entropy and correlation of the region block based on the gray-level co-occurrence matrix.

[0040] Specifically, to quantify image features such as contrast, uniformity, and orderliness by statistically analyzing the spatial relationships between pixels, a gray-level co-occurrence matrix (GLCM) is calculated for each region based on the gray value of each pixel. The algorithm for the GLCM is existing technology and will not be explained in detail here. Based on the GLCM, the contrast, energy, entropy, and correlation of the region are calculated. Contrast measures the sharpness and gray-level variation of the region; high contrast indicates strong gray-level transitions, corresponding to clear, high-contrast ECM fiber boundaries or distinct boundaries between cells and pores. Low contrast indicates a homogeneous and smooth region, possibly an amorphous ECM or culture medium area. Energy measures the uniformity of image texture; high energy values ​​indicate... A few large values ​​in the grayscale co-occurrence matrix indicate a very regular and uniform texture, such as neatly arranged and oriented ECM fiber bundles. Low energy values ​​indicate a messy and random texture, such as the presence of a broken ECM structure or randomly distributed cell clusters. Entropy measures the complexity or randomness of information contained in a region. High entropy values ​​indicate complex and unnecessary textures, corresponding to intricate fiber networks, highly porous sponge-like structures, or disordered cell distributions. Low entropy values ​​indicate simple and orderly textures. Correlation measures the linear dependence between pixel pairs. High correlation values ​​indicate the existence of a linear, directional texture pattern in a specified direction, corresponding to highly oriented ECM fibers. Low correlation values ​​indicate that the texture has no directionality.

[0041] In one specific embodiment, the second calculation is performed, which specifically includes the following steps:

[0042] The power spectrum is obtained by performing a two-dimensional fast Fourier transform on the current region block. The average energy of the power spectrum in the rings of different radii around the center is calculated, as well as the average energy of the power spectrum in the sector regions of different angles.

[0043] Specifically, to transform the image from the spatial domain to the frequency domain and analyze its periodic structure and frequency components, a two-dimensional fast Fourier transform is performed on the region block corresponding to the current sliding window. This transforms the image from the spatial domain to the frequency domain, resulting in a power spectrum. The center of the power spectrum represents low-frequency components (regions with gradual changes), while regions far from the center represent high-frequency components (regions with rapid changes). The distribution pattern of bright spots reveals the periodicity and direction of the texture. Radial energy characteristics are obtained by calculating the average energy of the power spectrum in rings of different radii around the center and performing radial distribution analysis. The structural scale of the construct can be obtained through radial energy characteristics. Energy concentrated in the low-frequency rings indicates a coarse texture with large pores or wide fibers, while energy distributed in the high-frequency rings indicates a fine texture with dense fine pores or a fine fiber network. Angular energy characteristics are obtained by calculating the average energy of the power spectrum in fan-shaped regions at different angles and performing angular distribution analysis. Angular energy characteristics represent the structural directionality of the construct. If the energy is concentrated in a specific angular range, it indicates that the texture has a strong orientation in that direction, i.e., the arrangement direction of the ECM fibers. If the energy is evenly distributed across all angles, it indicates that the structure is non-directional.

[0044] The process of calculating the average energy of the power spectrum across rings of different radii around the center is illustrated with a simple example. Assuming the region is 5×5, a Fourier fast transform is performed on the region to obtain the corresponding power spectrum P: P = [[0.8, 1.2,2.1,1.2,0.8],[1.2,8.5, 15.3,8.5,1.2],[2.1, 15.3, 89.2, 15.3,2.1],[1.2,8.5, The power spectrum is 15.3, 8.5, 1.2], [0.8, 1.2, 2.1, 1.2, 0.8]]. The power value at the center of this power spectrum is the largest, corresponding to the low-frequency components in the spatial domain. The power spectrum is symmetrically distributed, reflecting the structural symmetry of the original image. The power values ​​decrease from the center outwards, indicating fewer high-frequency components. Using the center of the power spectrum as the origin, a polar coordinate system is established to obtain the distance matrix D (Euclidean distance to the center): D = [[2.83, 2.24, 2.00, 2.24, 2.83], [2.24, 1.41, 1.00, 1.41, 2.24], [2.00, 1.00, 0.00, 1.00, 2.00], [2.24, 1.41, 1.00, 1.41, 2.24], [2.83, Given the values ​​2.24, 2.00, 2.24, 2.83, we assume three rings: Ring 1 (radius [0-1)), Ring 2 (radius [1-2)), and Ring 3 (radius [2-3)). Calculate the average energy of each ring. Ring 1 contains only the center point, with a power value of 89.2 and an average energy of 89.2 / 1 = 89.2. Ring 2 contains all points within the range [1-2), with power values ​​of 8.5, 15.3, 8.5, 15.3, 15.3, 8.5, 15.3, 8.5, and an average energy of (8.5 + 15.3 + 8.5 + 15.3 + 15.3 + 8.5 + 15.3 + 8.5) / 8 = 11.9, the total energy of ring 3 is 25.6, and the average energy is 1.6. In this example, ring 1 has the highest average energy, indicating that the ECM structure is dominated by low-frequency components, corresponding to a large-scale fiber skeleton. The average energy of ring 2 is medium, reflecting medium-scale fiber details. Ring 3 has the lowest energy, indicating that there is less fine texture and high-frequency noise.

[0045] To illustrate the calculation of the average energy of the power spectrum across different angular sector regions using a simple example, with the center of the power spectrum as the origin, the angle (in radians) at each point is calculated: The angle matrix θ is: θ = [[2.36, 2.68, 3.14, -2.68, -2.36], [2.03, 2.36, 3.14, -2.36, -2.03],[1.57, 1.57, 0.00, -1.57, -1.57],[1.11, 0.79, 0.00, -0.79, -1.11],[0.79, 0.46, 0.00, The values ​​are -0.46, -0.79, and the system is divided into four 90-degree sectors: Sector 1: -45° to 45° (horizontal); Sector 2: 45° to 135° (vertical); Sector 3: 135° to 225° (negative horizontal); Sector 4: 225° to 315° (negative vertical). The power values ​​in Sector 1 are: 89.2, 15.3, 2.1, 8.5, 1.2, 8.5, 1.2, with an average energy of 125 / 7 = 17.86. The power values ​​in Sector 2 are: 15.3, 15.3, 8.5, 2.1, 2.1, with an average energy of 43.4 / 5 = 8.66. The power values ​​in Sector 3 are: 1.2, 0.8, 1.2, ... 2.1, the average energy is 5.3 / 4=1.33. The power values ​​in sector 4 are: 0.8, 1.2, 1.2, 1.2, 2.1, with an average energy of 6.5 / 5=1.3. In this example, sector 1 has the highest energy, indicating that the ECM fibers have a strong dominant orientation in the horizontal direction. Sector 2 has medium energy, indicating that there is also a certain degree of fiber arrangement in the vertical direction. Sectors 3 and 4 have very low energy (about 1.3), indicating that there is very little fiber content in the diagonal direction.

[0046] The above-mentioned annular analysis shows that the ECM structure is dominated by low-frequency, large-scale components, which is consistent with the characteristics of an ordered fiber network. The sector analysis shows the strong orientation of the fibers in the horizontal direction, providing a quantitative basis for evaluating the structural anisotropy of the ECM construct. These frequency domain features can effectively distinguish different ECM structural modes and provide important input features for machine learning-based quality assessment.

[0047] In one specific embodiment, the complete region is divided into multiple sub-regions, which specifically includes the following steps:

[0048] Map the entire region onto a two-dimensional coordinate system, calculate the average value of the x-coordinate and the average value of the y-coordinate of each pixel in the entire region, round up the average value of the x-coordinate and the average value of the y-coordinate to obtain two coordinate values, take the pixel corresponding to these two coordinate values ​​as the geometric center of the entire region, calculate the distance from each pixel in the entire region to the geometric center, normalize all distance values, and divide the entire region into two different sub-regions based on the normalized distances.

[0049] Specifically, to transform the functional spatial heterogeneity within the ECM-stem cell construct into quantitative data that can be utilized by machine learning and statistical analysis, assuming that the distance from the black core of the ECM-stem cell construct is a key factor determining the local microenvironment, the geometric center of the construct is calculated. To calculate the geometric center, the entire region is mapped onto a two-dimensional coordinate system, and the average value of the x-coordinate and y-coordinate of each pixel in the entire region is calculated. Assuming that the average value is not an integer, the average value of the x-coordinate and y-coordinate is rounded up to obtain two coordinate values. The pixels corresponding to these two coordinate values ​​are taken as the geometric center of the entire region. Then, the distance from each pixel in the entire region to the geometric center is calculated. To eliminate the influence of the absolute size of constructs of different sizes and to unify the region division standard, all distance values ​​are normalized. Based on the normalized distance, the entire region is divided into two different sub-regions. A distance threshold is set, and pixels with a normalized distance less than or equal to the first distance threshold are classified into a sub-region as the first sub-region, and pixels with a normalized distance greater than the distance threshold are classified into a sub-region as the second sub-region.

[0050] In one specific embodiment, the structural features representing the structural relationships between different sub-regions are calculated based on statistics between different sub-regions, specifically including the following steps:

[0051] For each sub-region, extract the texture feature values ​​of all pixels belonging to the sub-region, and calculate the statistics of all texture feature values, including the mean and standard deviation.

[0052] For each texture feature value, the ratio of the corresponding mean and the difference of the standard deviation between the two sub-regions are calculated as structural features.

[0053] Specifically, texture features are pixel-level. An image has many pixels, thus generating tens of thousands of feature values. Machine learning based on massive, high-dimensional, and fragmented data is inefficient and prone to introducing noise, leading to model overfitting. Therefore, by calculating structural features, performing region statistics and relationship extraction on the first feature, data aggregation and information convergence are achieved. First, for each sub-region, the texture feature values ​​of all pixels belonging to the sub-region are extracted, and the statistics of all texture feature values ​​are calculated, including the mean and standard deviation. For example, for the texture feature value of contrast, the statistics for the first sub-region are Ca1 and Cs1, and the statistics for the second sub-region are Ca2 and Cs2. As a set of structural features, Ca1 / Ca2 and Cs1-Cs2, if Ca1 / Ca2 is greater than 1, it means that the core fibers are clearer than the edges, which may mean that the edge ECM is degraded. The ratio of Ca1 / Ca2 directly quantifies the ECM density gradient from the center to the edge. If the difference between Cs1 and Cs2 is large, it means that the structure of the first region is much more complex and disordered than the second sub-region, and the corresponding building quality may not be very good. The structural features calculated based on other texture feature values ​​can evaluate the quality of the structure from other dimensions. Performing the same calculation on all texture features yields multiple sets of structures. Combining all the calculated sets of structural features is used as the structural features of the structural relationship between sub-regions.

[0054] The above method performs the same statistical process on each texture feature to obtain the corresponding statistical quantity. Then, it calculates the relationship (ratio, difference) between sub-regions for each texture feature's statistical quantity as a structural feature. All structural features are combined to form a structural feature representing the structural relationship between sub-regions.

[0055] In one specific embodiment, the machine learning model is pre-trained, which includes the following steps:

[0056] Training images of ECM-stem cell constructs containing multiple known quality score labels are obtained. The training images are taken at structurally stable distances. Texture features and corresponding structural features are calculated based on the training images. Using the structural features of all training images and their corresponding quality score labels, a machine learning model is trained using a supervised machine learning algorithm.

[0057] Specifically, in order to obtain a machine learning model that can obtain corresponding quality scores based on structural features, training images of ECM-stem cell constructs containing multiple known quality score labels are obtained in advance. The training images are images of the corresponding ECM-stem cell constructs taken at the structural stability distance of the constructs corresponding to the ECM-stem cell constructs. Then, texture features are calculated for each training image, and structural features are calculated based on the texture features. Using the structural features of all training images and their corresponding quality score labels, a machine learning model is trained through a supervised machine learning algorithm to obtain a machine learning model that can accurately evaluate ECM-stem cell constructs based on structural features.

[0058] The above describes a machine learning-based ECM-stem cell construct quality assessment method in the embodiments of this application. The following describes a machine learning-based ECM-stem cell construct quality assessment system in the embodiments of this application. Please refer to [link to relevant documentation]. Figure 2 One embodiment of the machine learning-based ECM-stem cell construct quality assessment system in this application includes:

[0059] The first calculation module continuously captures images of the ECM-stem cell construct at multiple equally spaced defocus distances to obtain a defocus image sequence. For each image in the defocus image sequence, the projected area of ​​the ECM-stem cell construct is calculated as a basic morphological indicator.

[0060] The second calculation module calculates the rate of change of the basic morphological index relative to the defocus distance, and takes the defocus distance where the rate of change first falls below a preset threshold as the structural stability distance.

[0061] The feature extraction module acquires the construct image of the ECM-stem cell construct taken at a structurally stable distance, extracts the complete region of the ECM-stem cell construct from it, calculates texture features based on the complete region, obtains the spatial distribution of texture features in the complete region, divides the complete region into two sub-regions, calculates the statistics of texture features for each sub-region, and calculates structural features representing the structural relationship between sub-regions based on the statistics between different sub-regions.

[0062] The quality assessment module pre-trains a machine learning model, inputs structural features into the machine learning model, and outputs a quality score for the ECM-stem cell construct.

[0063] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0064] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0065] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A machine learning-based method for quality assessment of ECM-stem cell constructs, characterized in that, The method includes: Step S1: Take multiple images of the ECM-stem cell construct at equal intervals to obtain a defocused image sequence. For each image in the defocused image sequence, calculate the projected area of ​​the ECM-stem cell construct as a basic morphological indicator. Step S2: Calculate the rate of change of the basic morphological index relative to the defocus distance, and take the defocus distance where the rate of change first falls below the preset threshold as the structural stability distance. Step S3: Obtain the defocused image of the ECM-stem cell construct taken at a structurally stable distance, extract the complete region of the ECM-stem cell construct from it, calculate the texture features based on the complete region, obtain the spatial distribution of the texture features in the complete region, divide the complete region into two sub-regions, calculate the statistical measure of the texture features for each sub-region, and calculate the structural features representing the structural relationship between the sub-regions based on the statistical measure between different sub-regions. The division of the complete region into two sub-regions includes: mapping the complete region to a two-dimensional coordinate system, calculating the average value of the horizontal coordinate and the average value of the vertical coordinate of each pixel in the complete region, rounding the average value of the horizontal coordinate and the average value of the vertical coordinate up to obtain two coordinate values, taking the pixel corresponding to these two coordinate values ​​as the geometric center of the complete region, calculating the distance from each pixel in the complete region to the geometric center, normalizing all distance values, and dividing the complete region into two different sub-regions based on the normalized distances. Set a distance threshold, divide the pixels whose normalized distance is less than or equal to the first distance threshold into a sub-region as the first sub-region, and divide the pixels whose normalized distance is greater than the distance threshold into a sub-region as the second sub-region. Step S4: Pre-train the machine learning model, input the structural features into the machine learning model, and output the quality score of the ECM-stem cell construct by the machine learning model.

2. The method according to claim 1, characterized in that, Calculate the projected area of ​​the ECM-stem cell construct, including: Edge enhancement is performed on the original defocused image to obtain a boundary-enhanced image. Adaptive threshold segmentation is performed on the boundary-enhanced image to generate a binary image. Morphological closing operation is performed on the binary image to obtain an optimized binary image. Connectivity analysis is performed on the optimized binary image to filter out noise regions, identify all interconnected white pixel regions, count the total number of pixels in the white pixel regions, and use the total number of pixels as the projected area.

3. The method according to claim 1, characterized in that, Calculate the rate of change of basic morphological indicators relative to the defocus distance, and define the defocus distance at which the rate of change first falls below a preset threshold as the structurally stable distance, including: For each pair of adjacent images in the defocus image sequence, their basic morphological indices are obtained, and the rate of change of the two basic morphological indices is calculated to obtain a rate of change sequence. Each rate of change in the rate of change sequence is compared with a pre-set rate of change threshold. The first rate of change that is less than the rate of change threshold is called the target rate of change. The two images corresponding to the target rate of change are obtained, and the two defocus distances corresponding to the two images are obtained. The minimum defocus distance is selected as the structural stability distance.

4. The method according to claim 1, characterized in that, Texture features are calculated based on the entire region, and the spatial distribution of texture features within the entire region is obtained, including: A sliding window is set up and slids within the complete area at a preset step size. For each sliding area block, a first calculation and a second calculation are performed to obtain multiple local texture feature values. These multiple local texture feature values ​​are associated with the center pixel of the corresponding area block to obtain the spatial distribution of the complete area.

5. The method according to claim 4, characterized in that, The first calculation includes: For the current region block, calculate the gray-level co-occurrence matrix of the region block based on the gray value of each pixel in the region block, and calculate the contrast, energy, entropy and correlation of the region block based on the gray-level co-occurrence matrix.

6. The method according to claim 4, characterized in that, The second calculation includes: The power spectrum is obtained by performing a two-dimensional fast Fourier transform on the current region block. The average energy of the power spectrum in the rings of different radii around the center is calculated, as well as the average energy of the power spectrum in the sector regions of different angles.

7. The method according to claim 1, characterized in that, Structural features representing the structural relationships between different sub-regions are calculated based on statistics between them, including: For each sub-region, extract the texture feature values ​​of all pixels belonging to the sub-region, and calculate the statistics of all texture feature values, including the mean and standard deviation. For each texture feature value, the ratio of the corresponding mean and the difference of the standard deviation between the two sub-regions are calculated as structural features.

8. The method according to claim 1, characterized in that, Pre-trained machine learning models include: Training images of ECM-stem cell constructs containing multiple known quality score labels are obtained. The training images are taken at structurally stable distances. Texture features and corresponding structural features are calculated based on the training images. Using the structural features of all training images and their corresponding quality score labels, a machine learning model is trained using a supervised machine learning algorithm.

9. A machine learning-based ECM-stem cell construct quality assessment system, used to implement the machine learning-based ECM-stem cell construct quality assessment method as described in any one of claims 1-8, characterized in that, The system includes: The first calculation module continuously captures images of the ECM-stem cell construct at multiple equally spaced defocus distances to obtain a defocus image sequence. For each image in the defocus image sequence, the projected area of ​​the ECM-stem cell construct is calculated as a basic morphological indicator. The second calculation module calculates the rate of change of the basic morphological index relative to the defocus distance, and takes the defocus distance where the rate of change first falls below a preset threshold as the structural stability distance. The feature extraction module acquires defocused images of the ECM-stem cell construct taken at a structurally stable distance, extracts the complete region of the ECM-stem cell construct from it, calculates texture features based on the complete region, obtains the spatial distribution of texture features in the complete region, divides the complete region into two sub-regions, calculates the statistical measures of texture features for each sub-region, and calculates structural features representing the structural relationship between sub-regions based on the statistical measures between different sub-regions. The division of the complete region into two sub-regions includes: mapping the complete region to a two-dimensional coordinate system, calculating the average value of the horizontal coordinate and the average value of the vertical coordinate of each pixel in the complete region, rounding the average value of the horizontal coordinate and the average value of the vertical coordinate to obtain two coordinate values, taking the pixel corresponding to these two coordinate values ​​as the geometric center of the complete region, calculating the distance from each pixel in the complete region to the geometric center, normalizing all distance values, and dividing the complete region into two different sub-regions based on the normalized distances. A distance threshold is set, and pixels whose normalized distance is less than or equal to the first distance threshold are divided into a sub-region as the first sub-region, and pixels whose normalized distance is greater than the distance threshold are divided into a sub-region as the second sub-region. The quality assessment module pre-trains a machine learning model, inputs structural features into the machine learning model, and outputs a quality score for the ECM-stem cell construct.

Images