An organoid evaluation method and system based on artificial intelligence

Abstract

The application provides an organoid evaluation method and system based on artificial intelligence, and relates to the technical field of biomedical image analysis.The method comprises the following steps: S1, collecting a bright field layer scanning image sequence of a new sample to obtain a bright field image; S2, an organoid two-dimensional segmentation model automatically outputs a mask result; S3, a three-dimensional voxel space is constructed using a three-dimensional reconstruction process; S4, a feature vector is formed based on the number of organoids and statistical indicators; S5, the structured feature vector is input into a solidified organoid evaluation model; and a comprehensive evaluation result for the new sample is output.The system comprises an image data acquisition module, a model training unit, an automatic analysis and prediction unit, and a data management and result output module.The application integrates a complete technical process of intelligent segmentation of layer scanning images, high-precision three-dimensional reconstruction, multi-dimensional feature quantization, and machine learning model evaluation, and realizes automatic, objective, and standardized evaluation of the quality of organoids.

Description

Technical Field

[0001] This invention relates to the field of biomedical image analysis technology, specifically to an artificial intelligence-based organoid evaluation method and system. Background Technology

[0002] Organoids, as advanced three-dimensional cell culture models capable of mimicking the structure and function of human organs in vitro, have demonstrated significant application value in cutting-edge biomedical fields such as regenerative medicine, disease mechanism research, and personalized drug screening. With the rapid development of organoid technology, how to quickly, accurately, and in a standardized manner evaluate their growth status, maturity, and batch-to-batch consistency has become a key technical bottleneck that must be addressed to propel this technology from laboratory research to large-scale application.

[0003] Currently, methods for evaluating the quality of organoids still face several challenges in terms of objectivity, efficiency, and standardization: Existing evaluation methods largely rely on traditional two-dimensional microscopic images. Researchers typically estimate the growth status of organoids by observing images of a single focal plane and manually or semi-automatically measuring a few two-dimensional parameters such as the diameter. This approach cannot accurately reflect the complex and irregular three-dimensional spatial morphology and true volume of organoids, resulting in limited accuracy and information dimensions in quantitative analysis.

[0004] Segmentation and recognition of organoid images, especially on low-contrast bright-field images, largely rely on manual operation, which is not only inefficient but also makes it difficult to guarantee consistency of results among different operators. Furthermore, the final quality level determination is highly dependent on the personal experience of domain experts, lacking a unified and reproducible objective standard, which limits the horizontal comparison and standardized promotion of experimental results. Summary of the Invention

[0005] The purpose of this invention is to provide an artificial intelligence-based organoid evaluation method and system to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: On the one hand, this invention provides an artificial intelligence-based organoid evaluation method, the specific steps of which include: S1. For new samples that require organoid evaluation, acquire a bright-field tomographic image sequence of the new sample to obtain a bright-field image. S2. Input the bright field image of each layer of the new sample into the organoid two-dimensional segmentation model in sequence. The organoid two-dimensional segmentation model will automatically output the masking result. S3. Using a 3D reconstruction process, stack the mask results of each layer in Z-axis order and construct a 3D voxel space. S4. Divide the organoids in the three-dimensional voxel space into multiple diameter intervals, and count the number of organoids in each interval. At the same time, calculate at least one of the following statistical indicators for the new sample: organoid volume, number, surface area, equivalent ellipsoid axis, object bounding box size, sphericity, inner sphere radius, height, maximum thickness, maximum cross-sectional area, and maximum inner disk radius, and form a structured feature vector. S5. Input the structured feature vectors into the solidified organoid evaluation model; the organoid evaluation model analyzes the structured feature vectors based on the complex mapping relationship learned internally from morphological features to organoid growth status, maturity or quality level, and finally outputs a comprehensive evaluation result for the new sample.

[0007] Furthermore, the new sample contains several bright-field images at different depths, and records the corresponding interlayer step distance and pixel size; The masking results include a determination of whether each pixel belongs to an organoid or the background; For the two-dimensional segmentation results in the inference stage, a post-processing strategy is adopted, which includes: A threshold area is set, and connected regions with fewer than the threshold area number are filtered as noise; isolated points are removed using morphological opening operations; and Gaussian smoothing kernels are used to smooth the mask edges.

[0008] Furthermore, in step S3, the steps for constructing the three-dimensional voxel space are as follows: S31. Construct a three-dimensional coordinate system, and map the pixel coordinates in the two-dimensional mask to the layer index of the layer they belong to into three-dimensional voxel coordinates. The value of the z-axis in the three-dimensional voxel coordinates is equal to the product of the layer index and the inter-layer step distance. S32. Perform three-dimensional connected component analysis on the masking results to identify the three-dimensional objects of all independent organoids. The three-dimensional connected component analysis is based on voxel neighborhood relations and determines whether they are the same three-dimensional object based on the fact that the overlapping areas of adjacent layer masks meet the preset physical connectivity constraints. S33. Based on the number of voxels contained in the object, the interlayer step distance, and the pixel size, calculate the physical volume of the independent organoids, and calculate the equivalent diameter and related geometric features one by one to obtain the total number and volume of organoids in the new sample, as well as the distribution of each organoid in three-dimensional space.

[0009] Furthermore, in step S1, the processing steps for the bright-field tomographic image sequence are as follows: S11. Collect multimodal microscopic image data containing organoid information. For each experimental sample, use a microscopic imaging system to perform layer scanning to acquire several images of different focal planes. Each layer of different focal plane images is a bright field image. At the same time, acquire the corresponding stained image at the same field of view. S12. The bright-field image and staining image of each experimental sample are organized in depth order by layer to form a bright-field tomographic image sequence and the corresponding staining sequence, and the physical spacing and pixel spacing of each layer are recorded; the physical spacing and pixel spacing are uniformly stored in a standard format file. S13. Perform data cleaning and preprocessing on the acquired bright-field and stained images.

[0010] Furthermore, in step S11, the steps for obtaining the stained image are as follows: A multimodal annotation strategy using chromatogram-assisted annotation and linked bright-field map mapping is adopted to view the bright-field image and the corresponding chromatogram of the same layer and display them on the same interface; Taking advantage of the strong dye signal and clear boundaries in stained images, automatic and semi-automatic mask generation tools are provided. The mask generation tools include image segmentation functions based on global thresholding, adaptive local thresholding, and multi-channel fusion. By adjusting the threshold and smoothing parameter settings, an initial binary mask covering the organoid region can be quickly obtained.

[0011] Furthermore, in step S2, the training steps for the organoid two-dimensional segmentation model are as follows: After obtaining the labeled dataset of the initial binary mask, a two-dimensional segmentation model for organoids is constructed and trained. The two-dimensional segmentation model for organoids adopts one of the following structures: a convolutional neural network with context modeling capability and a Transformer hybrid structure. The input is a single-layer bright field image, and the output is a probability map or a binary mask of the same size as the input. During training, a composite loss function is used to calculate the loss. The composite loss function is obtained by weighting the cross-entropy loss function and the Dice loss function. By setting data augmentation strategies and performing iterative training, including at least one of geometric transformation and pixel transformation, an organoid two-dimensional segmentation model on a bright field image is obtained.

[0012] Furthermore, in step S3, the steps of the 3D reconstruction process are as follows: After obtaining the initial binary mask for each layer, the multiple initial binary masks are stacked in depth order and reconstructed in 3D using an image processing library. For one experimental sample, the bright field images of layers 1-N are processed by the organoid 2D segmentation model to obtain N binary mask images. Based on the Z-axis step and X and Y pixel spacing recorded during acquisition, the binary mask images are mapped to a unified physical coordinate system. Each binary mask image is stored in a 3D voxel grid in the form of voxels to form a 3D binary volume. Based on the 3D binary volume, all non-intersecting organoid 3D objects are identified. The number of voxels contained in each organoid 3D object is calculated, and the physical volume and 3D set parameters are obtained by combining the size of the voxels in physical space.

[0013] Furthermore, in step S5, the processing steps of the organoid evaluation model are as follows: Domain experts label each experimental sample with corresponding evaluation tags based on experimental standards or experience; the structured feature vectors are paired with the evaluation tags to form the training dataset for the evaluation model. Gradient boosting decision trees, random forests, or multilayer perceptrons were selected as machine learning models to fit the training dataset. The stability of the model was verified by K-fold cross-validation, and the number of trees, maximum depth, or learning rate hyperparameters were optimized using grid search to finally obtain a fixed organoid evaluation model. The organoid evaluation model outputs results based on the structured feature vectors and evaluation labels of new samples. The output results are presented to users intuitively through a visual interface; or the calculated output results are automatically compared with preset quality control standards to assist users in making decisions, transforming abstract analytical data into clear and quantifiable experimental decision-making basis.

[0014] Furthermore, in step S4, the diameter interval is divided in the following manner: Based on the biological characteristics of organoid developmental stages, a fixed diameter range is pre-defined; Alternatively, statistical analysis of organoid diameter data from historical samples can be performed, and the K-Means clustering algorithm can be used to automatically determine the boundaries of densely distributed intervals.

[0015] On the other hand, the present invention provides an artificial intelligence-based organoid evaluation system for the aforementioned artificial intelligence-based organoid evaluation method, comprising: The image data acquisition module is responsible for interacting with the microscopic imaging equipment to acquire the tomographic image data of the organoids and obtain bright-field tomographic image sequences and stained tomographic image sequences. The model training unit is responsible for offline training of all the artificial intelligence models required by this system; the model training unit includes a proxy annotation module, a segmentation model training module, and an evaluation model training module. The automated analysis and prediction unit is responsible for end-to-end fully automated analysis and prediction of new samples; the automated analysis and prediction unit calls the two-dimensional segmentation module, the three-dimensional reconstruction and feature quantization module, and the organoid evaluation model; The data management and results output module is responsible for the overall data management and results presentation of the system.

[0016] Compared with the prior art, the beneficial effects of the present invention are: This invention acquires multi-layer bright-field image sequences of organoid samples and uses a deep learning segmentation model to perform automated, high-precision pixel-level recognition of each layer of the image. Then, the two-dimensional segmentation results are accurately reconstructed in a three-dimensional space with physical size calibration. This achieves a fundamental shift from traditional two-dimensional approximate measurement to the accurate calculation of the true physical volume, equivalent diameter, and other three-dimensional geometric features of each independent organoid, providing an unprecedented high-dimensional, high-fidelity morphological data foundation for subsequent objective evaluation. This invention statistically integrates and condenses the massive morphological parameters obtained after 3D reconstruction into standardized numerical feature vectors. It then uses machine learning algorithms to train an organoid evaluation model, enabling the model to learn and reproduce the complex mapping relationship between objective quantitative indicators and evaluation conclusions such as excellent, good, poor, qualified, and unqualified from the perspective of experts in the field. This transforms the implicit knowledge that originally relied on expert experience and was difficult to express into a solidified, computer-executable, standardized, and explicit model. This invention deeply integrates the precise quantification of the three-dimensional morphology of two originally relatively independent processes with experience-based quality evaluation, thus constructing a new data-driven intelligent evaluation paradigm. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall method flow of the present invention; Figure 2 This is a schematic diagram of the overall system structure of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0019] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0020] Example 1: Please see Figure 1This embodiment provides an artificial intelligence-based organoid evaluation method, the specific steps of which include: Step S1: For new samples that require organoid evaluation, acquire the bright-field tomographic image sequence of the new samples to obtain bright-field images; Step S2: Input the bright field image of each layer of the new sample into the organoid two-dimensional segmentation model in sequence. The organoid two-dimensional segmentation model will automatically output the masking result. Step S3: Using the 3D reconstruction process, stack the mask results of each layer in Z-axis order and construct a 3D voxel space. Step S4: Divide the organoids in the three-dimensional voxel space into multiple diameter intervals. The diameter intervals are determined based on the biological characteristics of the organoid development stage or the size distribution histogram of the sample population. Count the number of organoids in each interval. At the same time, calculate the statistical indicators of the organoid volume, number, surface area, equivalent ellipsoid axis, object bounding box size, sphericity, inner sphere radius, height, maximum thickness, maximum cross-sectional area, and maximum inner disk radius of the new sample to form a structured feature vector. Step S5: Input the structured feature vector into the solidified organoid evaluation model; the organoid evaluation model analyzes the structured feature vector based on the complex mapping relationship learned internally from morphological features to organoid growth status, maturity or quality level, and finally outputs a comprehensive evaluation result for the new sample.

[0021] In this embodiment, the preset diameter range can be defined as follows: based on the biological characteristics of organoid growth and development cycles, a diameter less than 50 μm is defined as the cell cluster stage (without cavity structure); 50 μm-150 μm is defined as the logarithmic growth phase (cavities begin to form); and greater than 150 μm is defined as the mature stage (structure complete). Alternatively, based on statistical methods, organoid diameter data from no fewer than 1000 historical samples are collected, a size distribution histogram is plotted, and the K-Means clustering algorithm is used to automatically determine the boundaries of densely distributed intervals as the preset diameter ranges.

[0022] As a preferred embodiment, the new sample contains several layers of bright field images at different depths, with a value of more than 3 layers, and the corresponding interlayer step distance and pixel size are recorded. The masking results include the determination of whether each pixel belongs to an organoid or the background. For the two-dimensional segmentation results in the inference stage, a post-processing strategy is adopted, which includes: setting an area threshold and filtering connected regions with fewer pixels than the area threshold as noise; removing isolated points using morphological opening operations; and smoothing the mask edges using a Gaussian smoothing kernel to improve the connectivity and geometric rationality of the masking results. The area threshold is the number of pixels corresponding to the minimum physical area of ​​the organoid in the image, calculated based on the pixel size and the empirical minimum diameter.

[0023] In a preferred embodiment, the morphological opening operation uses a circular structuring element with a radius of 3-5 pixels; the Gaussian smoothing kernel has a size of 5×5 pixels and a standard deviation of 1.0.

[0024] It should be noted that by performing post-processing such as filtering and smoothing on the multi-layer segmentation mask, segmentation noise and artifacts can be effectively removed, ensuring the geometric accuracy of the 3D reconstruction and thus improving the reliability of the final quantitative analysis.

[0025] As a preferred embodiment, the steps for constructing the three-dimensional voxel space are as follows: Step S31: Construct a three-dimensional coordinate system. Combine the pixel coordinates in the two-dimensional mask with the layer index to map them into three-dimensional voxel coordinates. The value of the z-axis in the three-dimensional voxel coordinates is equal to the product of the layer index and the inter-layer step distance. Step S32: Perform three-dimensional connected component analysis on the mask results to identify the three-dimensional objects of all independent organoids. The three-dimensional connected component analysis is based on voxel neighborhood relations and determines whether they are the same three-dimensional object based on the fact that the overlapping areas of adjacent layer masks meet the preset physical connectivity constraints. Step S33: Based on the number of voxels contained in the object, the interlayer step distance and the pixel size, calculate the physical volume of the independent organoids, and calculate the equivalent diameter and related geometric features one by one to obtain the total number of organoids, the total volume and the distribution of each organoid in three-dimensional space in the new sample. It should be noted that the voxel neighborhood relationship mentioned in step S32 is 26-neighborhood; the preset physical connectivity constraint is that the overlapping projection area of ​​the masks between adjacent layers is greater than 10% of the average area of ​​a single-layer organoid.

[0026] The construction of three-dimensional voxel space is not a simple stacking, but a mapping strictly based on physical dimensions. Specifically, let the interlayer step distance be... (e.g., 5μm), pixel physical size is (e.g., 0.65 μm / pixel). In connected component analysis, the 26-neighborhood of a voxel V(x,y,z) is defined as: All voxels within the range. A pre-defined physical connectivity constraint is introduced when determining whether they belong to the same 3D object. Only when the mask overlap area between adjacent layers satisfies the physical connectivity constraint (i.e., the projected overlap area in the Z-axis direction is greater than 10% of the average area of ​​a single organoid slice) are they classified as the same 3D object. This effectively avoids reconstruction breaks caused by interlayer drift or sample jitter.

[0027] This embodiment combines physical size calibration with three-dimensional connected component analysis to accurately identify and separate each independent three-dimensional organoid from the layer-scan mask. It solves the problem that traditional two-dimensional observation cannot accurately count and is difficult to handle irregular or overlapping objects. It realizes the accurate calculation of key geometric features such as the true physical volume and equivalent diameter of a single organoid, providing a reliable data foundation for objectively and deeply evaluating the growth status and heterogeneity distribution of organoids.

[0028] In step S3, the 3D reconstruction process is as follows: After obtaining the initial binary mask for each layer, an image processing library (ITK, VTK, SimpleITK, or a similar 3D image processing library) is used to stack the multiple binary masks in depth order and perform 3D reconstruction. For one experimental sample, the bright-field images of layers 1-N are processed by the organoid 2D segmentation model to obtain N binary mask images. Based on the Z-axis step and X and Y pixel spacing recorded during acquisition, the binary mask images are mapped to a unified physical coordinate system. Each binary mask image is stored as a voxel in a 3D voxel grid to form a 3D binary volume, where voxels with a value of 1 represent organs occupied and voxels with a value of 0 represent the background. Based on the 3D binary volume recognition... Identify all non-overlapping 3D organoid objects, calculate the number of voxels contained in each 3D organoid object, and combine the size of the voxels in physical space (in this embodiment, the product of the X and Y pixel spacing and the Z-axis step) to obtain the physical volume and 3D set parameters. In this embodiment, the 3D set parameters include, but are not limited to, the number of organoids, the volume and distribution of different diameter ranges, the physical volume of each organoid, the equivalent sphere diameter, the maximum / minimum diameter, the aspect ratio and other scale-related features, and form structured quality parameter features accordingly. The traditional method of roughly estimating the volume based on the diameter of a single layer is completely replaced. The volume and 3D structure of organoids can be represented and stored in a physically accurate and reusable way. This embodiment accurately reconstructs a three-dimensional binary volume from a binary mask image by integrating physical space calibration; it completely replaces the traditional coarse estimation based on two-dimensional diameter with voxel-level calculation, realizing accurate and reproducible quantitative analysis of the true volume and complex three-dimensional morphology of organoids, providing a reliable morphological gold standard for label-free dynamic monitoring, and greatly improving the depth and accuracy of the research.

[0029] In a preferred embodiment, the processing steps of the bright-field tomographic image sequence in step S1 are as follows: Step S11: Collect multimodal microscopic image data containing organoid information. For each experimental sample, use a microscopic imaging system to perform layer scanning to acquire several images of different focal planes. The number of images of different focal planes is more than 3. Each layer of different focal plane images is a bright field image. At the same time, acquire the corresponding stained image at the same field of view. Step S12: The brightfield image and staining image of each experimental sample are organized layer by layer in depth order to form a brightfield tomographic image sequence and the corresponding staining sequence. The physical spacing of each layer (in this embodiment, the physical spacing is represented by the Z-direction step) and the pixel spacing (in this embodiment, the pixel spacing is represented by the X and Y direction calibration) are recorded. The physical spacing and pixel spacing are uniformly stored in a standard format file to ensure that they can be read and processed without loss in the subsequent preprocessing and reconstruction stages. Step S13: Perform data cleaning and preprocessing on the bright field image and the stained image to ensure the quality of the data used for annotation and training. For each experimental sample, extract the maximum value of the stained image in each field of view of the experimental sample. Divide all images in all fields of view by this maximum value to amplify the visual effect of the darker stained image. This embodiment acquires spatially registered multimodal tomographic images and uniformly records physical calibration parameters. Combined with standardized preprocessing, it constructs a high-quality training dataset with accurate physical dimensions. This not only provides a high-quality image source for subsequent efficient annotation, but also ensures the authenticity and accuracy of the final 3D reconstruction and quantitative analysis results from the data source, laying a solid foundation for the reliability of the entire automated system.

[0030] In a preferred embodiment, the steps for obtaining the stained image in step S11 are as follows: This software employs a multimodal annotation strategy that combines staining image-assisted annotation with linked bright-field image mapping. Users can view bright-field images and their corresponding staining images at the same layer, displayed on a single interface. Leveraging the strong dye signals and clear boundaries in staining images, the software provides automatic and semi-automatic mask generation tools. These tools include image segmentation functions based on global thresholding, adaptive local thresholding, and multi-channel fusion. By adjusting threshold and smoothing parameters, initial binary masks covering organoid regions can be quickly obtained. To address unavoidable detail errors in automatic segmentation, the software also provides histograms and other pixel value distribution information, allowing users to refine the initial binary mask on the staining image. After confirming the initial binary mask on the staining image, it is mapped in real-time to the corresponding bright-field image according to the registration relationship. This data serves as the label data for the bright-field image and is saved in standard mask format or NumPy array format. This approach avoids the extremely difficult manual outlining on bright-field images while ensuring the geometric accuracy of the initial binary mask in bright-field space, achieving a high-quality, efficient, and scalable annotation system. It should be noted that the multimodal annotation strategy, which employs stained image assistance and bright field image mapping, cleverly utilizes the advantage of clear boundaries in stained images to efficiently generate accurate binary masks through a human-machine collaboration approach and automatically maps them to bright field image labels. This overcomes the core pain point of the difficulty in directly annotating bright field images, improving annotation efficiency and quality by several orders of magnitude, and enabling the large-scale production of high-quality training data, providing a solid guarantee for the accurate training of subsequent models. The values ​​of the global threshold, adaptive local threshold, adjustment threshold, and smoothing parameter are not fixed but need to be dynamically adjusted according to the specific situation. The settings are as follows: Different experimental samples, such as staining agent concentration, cell type, imaging equipment, and exposure time, will result in vastly different brightness, contrast, and noise levels in the images; a fixed threshold cannot be applied to all situations. The operator will dynamically adjust these values ​​using sliders or input boxes based on the segmentation effect previewed on the screen in real time, until the most satisfactory initial binary mask is obtained; The smoothing parameter affects the noise level, i.e. the kernel size of the Gaussian blur, which in turn affects the selection of the optimal threshold. The global threshold is suitable for simple images with strong contrast between the background and the target. During operation, the user will adjust the value between 0 and 255 (for 8-bit images) to find the point that can most clearly distinguish the cells from the background.

[0031] In a preferred embodiment, the training steps of the organoid two-dimensional segmentation model in step S2 are as follows: After obtaining the labeled dataset of the initial binary mask, a two-dimensional organoid segmentation model is constructed and trained. The organoid segmentation model preferably adopts a U-Net network structure or a Swin-Transformer structure with residual connections. The input of the organoid segmentation model is a single-layer bright-field image, and the output is a probability map or a binary mask of the same size as the input. During the training process, a composite loss function is used, which is obtained by weighting the cross-entropy loss function and the Dice loss function. By setting data augmentation strategies such as random cropping, rotation, flipping, brightness perturbation, gamma correction and other geometric transformations or pixel transformations and iterative training, an organoid two-dimensional segmentation model on the bright-field image is obtained, so that the organoid two-dimensional segmentation model can complete the automatic identification and contour extraction of organs without relying on the color map during inference. In this embodiment, the organoid 2D segmentation model is used to learn the mapping relationship from the bright-field image to the organoid segmentation mask. The organoid 2D segmentation model can also employ TransUNet, UNet++, DeepLab series, Swin-UNet, or other hybrid structures combining convolutional neural networks and Transformers with context modeling capabilities. A composite loss function can enhance the sensitivity of the organoid 2D segmentation model in organoid boundary regions. Data augmentation strategies improve the robustness of the organoid 2D segmentation model to different imaging conditions and organoid morphologies.

[0032] In this embodiment, to address the issue of blurred organoid boundaries, this disclosure improves the standard U-Net network by introducing an attention gate mechanism at the skip connection to suppress the weights of background features and focus on the organoid edge features. The training parameters are set as follows: the input image size is uniformly adjusted to 512x512 pixels; the composite loss function... Assign the Dice loss function Higher weights are used to address the foreground-background pixel imbalance. The optimizer chosen is AdamW with an initial learning rate of 1e-4 and a cosine annealing learning rate decay strategy. Specific parameters for the data augmentation strategy are: random rotation angle range [-180°, 180°], random horizontal flip probability 0.5, random vertical flip probability 0.5, brightness jitter range ±20%, and contrast jitter range ±20%. The training termination condition is: the total loss of the composite loss function on the validation set does not decrease within 10 consecutive epochs.

[0033] This embodiment achieves accurate and automated segmentation of organoids in bright-field images by training and optimizing a U-Net encoder-decoder network, combined with a composite loss function and data augmentation strategy. It successfully eliminates the dependence on staining labels, transforming a time-consuming manual task into an efficient machine vision process. This provides core technical support for achieving label-free, high-throughput, and long-term dynamic monitoring of organoids, and has significant application value.

[0034] In a preferred embodiment, the processing steps of the organoid evaluation model in step S5 are as follows: After completing the 3D reconstruction and obtaining the physical volume and 3D ensemble parameters of each organoid, the voxel count and distribution of the organoids are statistically analyzed at the experimental sample level. The 3D objects of all independent organoids in each experimental sample are counted to obtain the total number of organoids in the experimental sample. Based on the equivalent sphere diameter of the total number of organoids, the organoids are divided into intervals to obtain the number of intervals. In this embodiment, the number of intervals is selectively set to five diameter intervals: less than 50μm, 50-100μm, 100-150μm, 150-200μm, and greater than 200μm. The number of organoids in each interval is counted to characterize the organoid size distribution spectrum. The organoid size distribution spectrum reflects, to a certain extent, the growth status, maturity, and heterogeneity of organoids in the experimental sample, and has important reference value for subsequent organoid evaluation. It compresses the organoid information originally scattered in multiple layers of images into a set of structured numerical features, providing a foundation for constructing an organoid evaluation dataset. Domain experts label each experimental sample with corresponding evaluation tags based on experimental standards or experience (in this embodiment, the evaluation tags include: growth status divided into three levels: excellent, good, and poor; or quality control judgment as qualified and unqualified); the structured feature vectors are paired with the evaluation tags to form the training dataset of the evaluation model; Specifically, the following steps are taken: Prepare a batch of representative experimental samples, setting the number of experimental samples to 200. Perform the complete S1-S4 process on each experimental sample to obtain the corresponding structured feature vector. Organize 2-3 domain experts to independently evaluate the original images or culture records of the experimental samples and provide evaluation labels according to pre-defined standards. An ordered grading system is defined, with growth status categorized into three levels: excellent, good, and poor. A binary classification label is defined, with quality control judgments as qualified and unqualified. Drug sensitivity is graded as highly sensitive, moderately sensitive, and insensitive. A continuous score range is defined, set to 0.0-10.0 points. For evaluation labels from multiple experts, strategies such as majority voting or averaging are used to determine the final gold standard label to reduce individual subjective bias. The feature vector of each experimental sample is paired with the final determined evaluation label to form a complete training dataset, with each row containing one feature vector and one label. In this embodiment, machine learning models are selected, including gradient boosting trees, random forests, linear or nonlinear regression, multilayer perceptrons, etc., to fit the training dataset and establish an intelligent mapping from structured feature vectors to evaluation results. The stability of the model is verified by K-fold cross-validation, and the number of trees, maximum depth, or learning rate hyperparameters are optimized by grid search to finally obtain a solidified high-performance organoid evaluation model. Specifically, LightGBM is one of the fastest and most efficient algorithms for processing tabular data, making it very suitable for high-throughput analysis scenarios. As an excellent implementation of Gradient Boosting Tree (GBM), it typically achieves top performance on various classification and regression tasks. It uses a histogram-based algorithm, resulting in lower memory consumption. It can output feature importance, helping to understand which morphological features organoid evaluation models primarily rely on for judgment. The constructed training dataset is divided into a training set (70%–80% in this embodiment) and a test set (20%–30% in this embodiment) for training the model and evaluating the final performance. Within the training set, K-fold cross-validation (K=5 or 10 in this embodiment) is used to find the optimal hyperparameters of the organoid evaluation model and evaluate the generalization ability of the organoid evaluation model; effectively preventing the model from overfitting on specific data partitions; Automated methods such as grid search or random search are used to find the hyperparameter combination that enables the organoid evaluation model to perform best in cross-validation within a preset parameter range. Using the found optimal hyperparameters, the organoid evaluation model is retrained on the entire training set to obtain the final fixed organoid evaluation model. The performance of the finalized organoid evaluation model was evaluated on a test set that was never used for training, and objective performance metrics were obtained to confirm the reliability of the organoid evaluation model. The organoid evaluation model outputs results based on the structured feature vectors and evaluation labels of new samples. The output results are presented to users intuitively through a visual interface; or the calculated output results are automatically compared with preset quality control standards to assist users in making decisions, transforming abstract analytical data into clear and quantifiable experimental decision-making basis. The trained organoid evaluation model and all parameters are saved as a file. In this embodiment, the joblib or pickle library is used. The solidified model file is integrated into the organoid evaluation module of the automated analysis and evaluation unit. In practical applications, the organoid evaluation model only needs to load this model file to quickly infer new feature vectors and output evaluation results.

[0035] To ensure the objectivity of the training data, this invention establishes a standardized "training dataset annotation process for organoid evaluation models": (1) Independent labeling by multiple experts: Each sample is independently scored or classified by 3 senior biological experts; (2) Inter-Rater Reliability: The Cohen's Kappa coefficient between expert scores is calculated. If the Kappa coefficient is greater than 0.85, the labeling is considered to be in good agreement, and the final label is determined by majority voting; if the Kappa coefficient is less than 0.85, a fourth expert is introduced for arbitration or an expert discussion meeting is organized until a consensus is reached. (3) Establishment of the gold standard: For samples that are difficult to judge, the morphological labeling is corrected by using subsequent ATP activity detection or pathological section results as the "gold standard".

[0036] To verify the generalization ability and stability of this scheme, this embodiment was not only validated on colorectal cancer tumor organoids (200 cases), but also extended to validate normal liver organoids (50 cases) and brain organoids (50 cases). Experimental results show: (1) Adaptability to different morphologies: For dense liver organoids and hollow tumor organoids, the Dice coefficient of the two-dimensional segmentation model remains above 0.88. (2) Different imaging conditions: Under imaging conditions of low contrast (30% reduction in light intensity) and high noise (increased ISO gain), the volume calculation error fluctuation is less than 5% thanks to the data augmentation strategy. This shows that the "structured feature vectors" (such as sphericity, roughness, etc.) extracted by this method can effectively capture the common morphological features of different types of organoids and have good cross-class robustness.

[0037] It should be noted that the organoid types used in this embodiment are merely illustrative and do not constitute any limitation on the scope of protection of this invention. Those skilled in the art should understand that the evaluation method and system provided by this invention are also applicable to other types of organoids, including but not limited to: organoids derived from different tissues and organs, such as brain organoids, lung organoids, kidney organoids, stomach organoids, pancreatic organoids, and intestinal organoids; organoids derived from different species, such as human organoids, mouse organoids, and porcine organoids; organoids from different disease models, such as tumor organoids, inflammatory disease organoids, and hereditary disease organoids; and organoids obtained at different culture stages or under different culture conditions. Any technical solution based on the technical concept of this invention that combines bright-field tomography images with three-dimensional reconstruction for organoid quality evaluation falls within the scope of protection of this invention.

[0038] Example 2: refer to Figure 2This embodiment provides an artificial intelligence-based organoid evaluation system, including: an image data acquisition module, a model training unit, an automated analysis and prediction unit, and a data management and result output module; The image data acquisition module is responsible for interacting with the microscopic imaging equipment to acquire the computed tomography image data of the organoids and obtain bright-field computed tomography image sequences and stained computed tomography image sequences. Input the raw image stream from the microscopic imaging system and the user-defined acquisition parameters. In this embodiment, the parameters include the field of view, number of layers, step distance, etc. Output the spatially registered bright-field tomographic image sequence and stained tomographic image sequence of the training mode; and the bright-field tomographic image sequence of the test sample of the inference mode; and the physical size parameters associated with the image sequence, including the X / Y pixel spacing and the Z-axis step distance; The model training unit is responsible for offline training of all the artificial intelligence models required by this system; the model training unit includes a proxy annotation module, a segmentation model training module, and an evaluation model training module. The proxy annotation module generates label data for training efficiently and with high quality through a strategy of coloring map assistance and bright field map mapping. The input training samples consist of a sequence of stained tomographic images; an integrated image segmentation algorithm is used to process high-contrast stained images using adaptive thresholding to generate an initial two-dimensional segmentation mask; a human-computer interaction interface is provided for users to refine the mask; finally, the mask that has been confirmed to be correct is used as a label. Output a pixel-accurate two-dimensional segmentation mask sequence that corresponds one-to-one with the bright field image sequence; The segmentation model training module is responsible for training a deep learning model for automatically segmenting organoids from bright-field images; The input consists of a bright-field tomographic image sequence provided by the image data acquisition module and a corresponding two-dimensional segmentation mask sequence generated by the proxy annotation module. The input bright-field tomographic image sequence and the two-dimensional segmentation mask sequence are constructed into a training pair. Data augmentation strategies are applied. Iterative training is performed using a preset neural network architecture and a composite loss function until convergence. Output the trained and solidified 2D segmentation model; The evaluation model training module is responsible for fitting the training dataset and establishing an intelligent mapping from structured feature vectors to evaluation results; it optimizes the model hyperparameters through cross-validation and grid search methods, and finally obtains a solidified high-performance organoid evaluation model. Input a batch of structured 3D morphological feature vectors of training samples. The 3D morphological feature vectors need to be obtained by running a part of the inference process once. Evaluation labels corresponding to 3D morphological feature vectors; pairing 3D morphological feature vectors with evaluation labels to form a training set; using a pre-defined machine learning algorithm for fitting and training; Output the trained and solidified organoid evaluation model; The automated analysis and prediction unit is responsible for end-to-end fully automated analysis and prediction of new samples; the automated analysis and prediction unit calls the two-dimensional segmentation module, the three-dimensional reconstruction and feature quantization module, and the organoid evaluation model; The 2D segmentation module is used to load and execute the trained 2D segmentation model; Input a sequence of bright-field tomographic images of the sample to be tested; output a predicted two-dimensional segmentation mask sequence. The 3D reconstruction and feature quantization module is used to realize the transformation from 2D mask to 3D objective quantitative indicators, which is another core innovation of this invention; The input consists of a predicted mask sequence generated by the 2D segmentation module and physical size parameters provided by the image data acquisition module. The 2D mask sequence is stacked into a 3D voxel model based on the physical size parameters. 3D connected component analysis and an optional 3D watershed algorithm are applied to identify and separate all independent organoid 3D objects. For each object, its physical volume, equivalent sphere diameter, and other 3D morphological parameters are accurately calculated. Statistical analysis is performed at the sample level to generate a structured 3D morphological feature vector consistent with the input dimension of the prediction model.

[0039] Output a feature vector describing the overall state of the organoid in the test sample; The organoid evaluation model is used to load and execute the trained organoid evaluation model, completing the final step from morphology to decision-making. Input the feature vector generated by the 3D reconstruction and feature quantization module; output the comprehensive evaluation result for the new sample.

[0040] The data management and results output module is responsible for the overall data management and results presentation of the system.

[0041] The system stores and versions the two-dimensional segmentation model and organoid evaluation model generated by the model training unit; stores and manages image data, labeled data, calculated features, and comprehensive evaluation results; and presents the comprehensive evaluation results to users in the form of visual charts and numerical reports, or automatically compares the calculated output results with preset quality control standards to assist users in making decisions, transforming abstract analytical data into clear and quantifiable experimental decision-making basis.

[0042] The data management and results output module is responsible for the overall data management and results presentation of the system.

[0043] This invention also provides an application scenario for organoid evaluation based on artificial intelligence. It mainly optimizes key steps and strategies of organoid experiments based on comprehensive evaluation results. Specific application scenarios include, but are not limited to: determining the optimal treatment time for organoids, judging the appropriate seeding cell density, assessing cell confluence status, clarifying drug dosage and treatment conditions, non-destructively evaluating the activity and quantity of organoids or cells, subsequent experimental requirements (in this embodiment, subsequent experiments include tumor organoid drug sensitivity testing, pathological identification, gene sequencing, etc.), non-destructive dynamic cell activity detection, and step setting for application scenarios based on comprehensive evaluation results. In addition, it can be extended to organoid quality assessment, culture condition optimization, phenotypic analysis, and experimental reproducibility verification, providing support for the standardization and systematic advancement of organoid research.

[0044] It should be noted that all calculation formulas in this application employ regression analysis, including but not limited to machine learning algorithms, to deeply analyze the collected parameters and identify their natural trends and interrelationships. Specialized software, such as Python's Scikit-learn library or the R language, is used to automatically generate mathematical models that match the data. Then, cross-validation and other methods are used to objectively evaluate the model performance, and continuous feedback and optimization are combined to ensure that the created formulas truly reflect the inherent laws of the data, thereby guaranteeing their effectiveness and accuracy. In all calculation formulas in this application, the parameters in each formula undergo dimensionless processing within a consistent range to ensure that different physical quantities are compared on the same scale; dimensionless processing techniques include, but are not limited to, Min-Max Normalization and Z-Score standardization. The technical solution of this invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments of this invention.

[0045] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0046] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An artificial intelligence-based organoid evaluation method, characterized in that, The specific steps include: S1. For new samples that require organoid evaluation, acquire a bright-field tomographic image sequence of the new sample to obtain a bright-field image. S2. Input the bright field image of each layer of the new sample into the organoid two-dimensional segmentation model in sequence. The organoid two-dimensional segmentation model will automatically output the masking result. S3. Using a 3D reconstruction process, stack the mask results of each layer in Z-axis order and construct a 3D voxel space. S4. Divide the organoids in the three-dimensional voxel space into multiple diameter intervals, and count the number of organoids in each interval. At the same time, calculate at least one of the following statistical indicators for the new sample: organoid volume, number, surface area, equivalent ellipsoid axis, object bounding box size, sphericity, inner sphere radius, height, maximum thickness, maximum cross-sectional area, and maximum inner disk radius, and form a structured feature vector. S5. Input the structured feature vectors into the solidified organoid evaluation model; the organoid evaluation model analyzes the structured feature vectors based on the complex mapping relationship learned internally from morphological features to organoid growth status, maturity or quality level, and finally outputs a comprehensive evaluation result for the new sample.

2. The artificial intelligence-based organoid evaluation method according to claim 1, characterized in that, The new sample contains several bright-field images at different depths, and records the corresponding interlayer step distance and pixel size; The masking results include a determination of whether each pixel belongs to an organoid or the background; For the two-dimensional segmentation results in the inference stage, a post-processing strategy is adopted, which includes: A threshold area is set, and connected regions with fewer than the threshold area number are filtered as noise; isolated points are removed using morphological opening operations; and Gaussian smoothing kernels are used to smooth the mask edges.

3. The method of claim 2, wherein the method is based on artificial intelligence. In step S3, the steps for constructing the three-dimensional voxel space are as follows: S31. Construct a three-dimensional coordinate system, and map the pixel coordinates in the two-dimensional mask to the layer index of the layer they belong to into three-dimensional voxel coordinates. The value of the z-axis in the three-dimensional voxel coordinates is equal to the product of the layer index and the inter-layer step distance. S32. Perform three-dimensional connected component analysis on the masking results to identify the three-dimensional objects of all independent organoids. The three-dimensional connected component analysis is based on voxel neighborhood relations and determines whether they are the same three-dimensional object based on the fact that the overlapping areas of adjacent layer masks meet the preset physical connectivity constraints. S33. Based on the number of voxels contained in the object, the interlayer step distance, and the pixel size, calculate the physical volume of the independent organoids, and calculate the equivalent diameter and related geometric features one by one to obtain the total number and volume of organoids in the new sample, as well as the distribution of each organoid in three-dimensional space.

4. The method of claim 3, wherein the method is based on artificial intelligence. In step S1, the processing steps for the bright-field tomographic image sequence are as follows: S11. Collect multimodal microscopic image data containing organoid information. For each experimental sample, use a microscopic imaging system to perform layer scanning to acquire several images of different focal planes. Each layer of different focal plane images is a bright field image. At the same time, acquire the corresponding stained image at the same field of view. S12. The bright-field image and staining image of each experimental sample are organized in depth order by layer to form a bright-field tomographic image sequence and the corresponding staining sequence, and the physical spacing and pixel spacing of each layer are recorded; the physical spacing and pixel spacing are uniformly stored in a standard format file. S13. Perform data cleaning and preprocessing on the acquired bright-field and stained images.

5. The method of claim 4, wherein the method is based on artificial intelligence. In step S11, the steps for obtaining the stained image are as follows: A multimodal annotation strategy using chromatogram-assisted annotation and linked bright-field map mapping is adopted to view the bright-field image and the corresponding chromatogram of the same layer and display them on the same interface; Taking advantage of the strong dye signal and clear boundaries in stained images, automatic and semi-automatic mask generation tools are provided. The mask generation tools include image segmentation functions based on global thresholding, adaptive local thresholding, and multi-channel fusion. By adjusting the threshold and smoothing parameter settings, an initial binary mask covering the organoid region can be quickly obtained.

6. The method of claim 5, wherein the method is based on artificial intelligence. In step S2, the training steps of the organoid 2D segmentation model are as follows: After obtaining the labeled dataset of the initial binary mask, a two-dimensional segmentation model for organoids is constructed and trained. The two-dimensional segmentation model for organoids adopts one of the following structures: a convolutional neural network with context modeling capability and a Transformer hybrid structure. The input is a single-layer bright field image, and the output is a probability map or a binary mask of the same size as the input. During training, a composite loss function is used to calculate the loss. The composite loss function is obtained by weighting the cross-entropy loss function and the Dice loss function. By setting data augmentation strategies and performing iterative training, including at least one of geometric transformation and pixel transformation, an organoid two-dimensional segmentation model on a bright field image is obtained.

7. The method of claim 3, wherein the method is based on artificial intelligence. In step S3, the steps of the 3D reconstruction process are as follows: After obtaining the initial binary mask for each layer, the multiple initial binary masks are stacked in depth order and reconstructed in 3D using an image processing library. For one experimental sample, the bright field images of layers 1-N are processed by the organoid 2D segmentation model to obtain N binary mask images. Based on the Z-axis step and X and Y pixel spacing recorded during acquisition, the binary mask images are mapped to a unified physical coordinate system. Each binary mask image is stored in a 3D voxel grid in the form of voxels to form a 3D binary volume. Based on the 3D binary volume, all non-intersecting organoid 3D objects are identified. The number of voxels contained in each organoid 3D object is calculated, and the physical volume and 3D set parameters are obtained by combining the size of the voxels in physical space.

8. The method of claim 7, wherein the method is based on artificial intelligence. In step S5, the processing steps of the organoid evaluation model are as follows: Domain experts label each experimental sample with corresponding evaluation tags based on experimental standards or experience; the structured feature vectors are paired with the evaluation tags to form the training dataset for the evaluation model. Gradient boosting decision trees, random forests, or multilayer perceptrons were selected as machine learning models to fit the training dataset. The stability of the model was verified by K-fold cross-validation, and the number of trees, maximum depth, or learning rate hyperparameters were optimized using grid search to finally obtain a fixed organoid evaluation model. The organoid evaluation model outputs results based on the structured feature vectors and evaluation labels of new samples, and the output results are presented to users intuitively through a visualization interface; Alternatively, the calculated output can be automatically compared with preset quality control standards to assist users in making decisions, transforming abstract analytical data into clear and quantifiable experimental decision-making basis. 9.The artificial intelligence-based organoid evaluation method of claim 1, wherein, In step S4, the diameter range is divided in the following manner: Based on the biological characteristics of organoid developmental stages, a fixed diameter range is pre-defined; Alternatively, statistical analysis of organoid diameter data from historical samples can be performed, and the K-Means clustering algorithm can be used to automatically determine the boundaries of densely distributed intervals.

10. An artificial intelligence-based organoid evaluation system for performing the artificial intelligence-based organoid evaluation method according to any one of claims 1 to 9, characterized by, include: The image data acquisition module is responsible for interacting with the microscopic imaging equipment to acquire the tomographic image data of the organoids and obtain bright-field tomographic image sequences and stained tomographic image sequences. The model training unit is responsible for offline training of all the artificial intelligence models required by this system; the model training unit includes a proxy annotation module, a segmentation model training module, and an evaluation model training module. The automated analysis and prediction unit is responsible for performing end-to-end fully automated analysis and prediction on new samples; The automated analysis and prediction unit calls upon the two-dimensional segmentation module, the three-dimensional reconstruction and feature quantization module, and the organoid evaluation model; The data management and results output module is responsible for the overall data management and results presentation of the system.

Images