An organoid drug sensitivity quantification analysis method based on morphological texture fingerprint

Abstract

The application discloses a kind of organoid drug sensitivity quantitative analysis methods based on morphological texture fingerprint, it is related to drug sensitivity quantitative analysis technical field, comprising the following steps: obtaining the bright field microscopic image and fluorescence microscopic image of 96 well plate organoid culture system;Using three-level degradation segmentation method to obtain the instance mask of each organoid;Extract multi-dimensional enhanced phenotype feature vector;Obtain the drug response metric value of each organoid instance;Generate drug sensitivity analysis report.The application can realize the accurate inference of the live and dead state of organoid without the need for fluorescent labeling, thereby realizing the full-automation process from image input to efficacy report, thereby improving the efficiency and repeatability of high-throughput organoid drug sensitivity screening.

Description

Technical Field

[0001] This invention relates to the field of drug sensitivity quantitative analysis technology, and in particular to an organoid drug sensitivity quantitative analysis method based on morphological texture fingerprints. Background Technology

[0002] Organoids are miniature organ-like structures that self-assemble from stem cells or tumor tissue in a three-dimensional matrix. They can highly mimic the morphology, function, and drug response characteristics of in vivo tissues, and have become a cutting-edge platform for tumor drug sensitivity screening. However, in practical applications, high-throughput 96-well plate organoid drug sensitivity experiments face the following core bottlenecks:

[0003] In vitro cultured organoids exhibit high heterogeneity in size (tens to thousands of pixels), shape (round to irregular budding), and grayscale (low and uneven contrast with the background). Existing general-purpose cell segmentation algorithms (such as Cellpose and StarDist) have low recall rates on bright-field images of organoids, while fluorescence-dependent staining segmentation schemes are phototoxic and destructive to live cells and cannot be used for continuous dynamic monitoring.

[0004] The gold standard for determining organoid viability relies on fluorescent labeling (e.g., green fluorescent protein (GFP) for live cells, red fluorescent protein (RFP) for dead cells, or propidium iodide (PI) for dead cells). However, fluorescence imaging is time-consuming and potentially toxic to organoids, and clinical samples often only have bright-field images. Current technology lacks an effective method for inferring organoid viability from bright-field morphological features under fluorescence-free conditions.

[0005] The effects of different drugs on organoid morphology are reflected in multiple dimensions (area changes, shape regularity, texture roughness, edge sharpness, etc.). Traditional methods only use a single indicator such as manually measuring diameter or area to evaluate, which cannot capture the global phenotypic fingerprint features of drug effects.

[0006] A single 96-well plate generates dozens to hundreds of high-resolution images of the entire plate (3072×4096 pixels). The traditional workflow of manually annotating and analyzing each well is slow and has poor repeatability.

[0007] Therefore, there is an urgent need for a fully automated analysis method that integrates deep learning and multidimensional feature engineering, and can achieve organoid instance segmentation, liveness / death determination, and drug efficacy quantification under the condition of only bright field image input. Summary of the Invention

[0008] The purpose of this invention is to provide a method for quantitative analysis of organoid drug sensitivity based on morphological texture fingerprints, so as to solve the problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a method for quantitative analysis of organoid drug sensitivity based on morphological texture fingerprints, comprising the following steps:

[0010] S1. Obtain bright-field microscopic images of the 96-well plate organoid culture system and optional paired green fluorescent protein and red fluorescent protein fluorescence microscopic images.

[0011] S2. The bright-field microscopic image is segmented using a three-level degradation segmentation method to obtain the mask of each organoid instance;

[0012] S3. For each organoid instance obtained from the segmentation, extract the multidimensional enhanced phenotypic feature vector from the bright-field micrograph.

[0013] S4. Based on the multidimensional enhanced phenotypic feature vector, calculate the drug response metric for each organoid instance to reflect the degree of phenotypic change of the organoid under drug action.

[0014] S5. Generate a drug sensitivity analysis report based on the drug response metrics.

[0015] Preferably, in step S2, the three-level degradation separation method specifically includes:

[0016] The bright-field microscopic image is automatically cascaded and segmented according to a predetermined priority order of the first-level deep learning segmentation algorithm, the second-level deep learning general segmentation algorithm, and the third-level classical computer vision segmentation algorithm.

[0017] If the current level segmentation algorithm outputs a valid result, it returns directly; otherwise, it automatically downgrades to the next level algorithm to ensure that the organoid instance mask can be obtained under all input conditions.

[0018] Preferably, in step S2:

[0019] The first-level deep learning segmentation algorithm is the YOLOv8-seg instance segmentation model, with a confidence threshold and a cross-union ratio threshold for non-maximum suppression configured.

[0020] The second-level deep learning general segmentation algorithm is the Cellpose model, and the Cellpose model adopts an automatic downgrade calling method that executes the Cyto3 model, Cyto2 model and Cyto model in order of priority. When the segmentation of the previous model fails, the next model is automatically activated.

[0021] The third-level classical computer vision segmentation algorithm includes background correction, Gaussian smoothing, Otsu global thresholding, morphological cleaning, distance transformation and watershed separation of adhering instances, and area filtering, all performed sequentially.

[0022] Preferably, in step S3, the multidimensional enhanced phenotypic feature vector includes:

[0023] 12-dimensional basic morphological texture features: region area, roundness, solidity, aspect ratio, eccentricity, average gray intensity, standard deviation of gray intensity, mean gray value after background correction, standard deviation of gray value after background correction, Sobel edge density, texture entropy, and local contrast.

[0024] 18-dimensional enhancement features: 7-dimensional grayscale distribution features, 5-dimensional GLCM texture features, 3-dimensional Sobel edge gradient statistics, 1-dimensional radial intensity ratio, and 2-dimensional multi-scale granularity.

[0025] Preferably, in step S4, the drug response metric includes the organoid survival probability, which is obtained through cross-modal knowledge transfer, specifically including:

[0026] When there are green fluorescent protein fluorescence images and red fluorescent protein fluorescence images paired with the bright-field microscopic image, a gold standard segmentation label and a life-or-death score are generated based on the paired fluorescence images. The multidimensional enhanced phenotypic feature vector is extracted on the same instance mask of the corresponding bright-field image. The ensemble tree model is trained with the life-or-death score as a supervision signal. A mapping function from bright-field morphological texture features to organoid survival probability is established, so that the survival status of each organoid can be directly predicted when only bright-field image input is available.

[0027] When the paired fluorescence image cannot be read, the sample is automatically downgraded from the gold standard mode to the silver standard mode, and a replacement label is generated using the third-level classical computer vision segmentation algorithm.

[0028] Preferably, the method for generating the gold standard segmentation label specifically includes:

[0029] Acquire paired green fluorescent protein fluorescence images and red fluorescent protein fluorescence images, and segment the fluorescence images to obtain an instance mask;

[0030] Calculate a live cell score for each instance, which is equal to the sum of the green fluorescent protein-corrected pixel values ​​within the masked region of that instance divided by the sum of the green fluorescent protein-corrected pixel values ​​and the red fluorescent protein-corrected pixel values;

[0031] According to the gold standard segmentation label for live cell scoring, cells are considered alive when the live cell score reaches the maximum threshold, dead when the live cell score reaches the minimum threshold, and apoptosis when the live cell score is between the maximum and minimum thresholds.

[0032] Preferably, in step S4, the drug response metric includes a weighted mortality score, which is calculated in the following manner:

[0033] Calculate the apoptotic pathway death sub-signals, mitochondrial functional death sub-signals, and morphological death sub-signals, respectively.

[0034] Weighted mortality scores are calculated through adjustable weight fusion.

[0035] When the weighted mortality score is greater than the maximum mortality score threshold and the normalized apoptosis intensity value is greater than the normalized threshold, death is determined.

[0036] Apoptosis is determined when the weighted mortality score is between the maximum mortality score threshold and the minimum mortality score threshold.

[0037] A person is considered alive when the weighted mortality score is less than or equal to the minimum mortality score threshold.

[0038] Preferably, in step S5, the drug sensitivity analysis report includes statistical comparative analysis results across experimental conditions, and the process of obtaining the statistical comparative analysis specifically includes:

[0039] The images were grouped according to the experimental conditions, and the descriptive statistics of each multidimensional enhanced phenotypic feature within each group were calculated.

[0040] For the two groups, perform the independent samples t-test and the Mann-Whitney U test;

[0041] For multiple group comparisons, one-way ANOVA and Kruskal-Wallis H test were performed, and post-hoc pairwise comparisons between each experimental group and the control group were conducted using the Kruskal-Wallis H test.

[0042] Calculate the Kruskal-Wallis H test;

[0043] Features that satisfy the condition of having a significance probability value less than the preset significance level and an absolute value of the Cohen d effect size greater than 0.5 are selected as significant change features for output.

[0044] Preferably, it also includes the preparation of automatic data for 96-well plates, specifically including:

[0045] Based on an 8×12 grid and shrinkage ratio, 96 holes are automatically located, and images of each single hole are cut from the high-resolution image of the whole board and scaled to a predetermined size.

[0046] For high-resolution full-panel images, a non-overlapping sliding window partitioning with a step size equal to the window size is used. After scaling the image blocks to the target resolution, content filtering is performed based on the target pixel ratio to exclude empty image blocks and abnormally dense image blocks.

[0047] A file path list is used to achieve proportional hierarchical random splitting of the training and validation sets, avoiding physical copying of files.

[0048] Preferably, it also includes at least one of active learning or CellTiter-Glo gold standard calibration;

[0049] The active learning specifically includes: selecting manually reviewed targets from unlabeled samples with the highest prediction uncertainty based on the Shannon entropy sampling strategy; triggering model retraining when the number of labeled samples reaches the minimum threshold; and retaining multiple recent historical model versions to support version rollback.

[0050] The CellTiter-Glo gold standard calibration specifically includes: using ordinal-preserving regression to perform first-level monotonic calibration on the original predicted values ​​of the model, and then using at least one of Ridge regression, Lasso regression, random forest or gradient boosting regression to perform leave-one-out cross-validation on the 24-dimensional image-level convergent features, and selecting the best model as the second-level calibrator based on minimizing the mean absolute error.

[0051] The technical effects and advantages of this invention are as follows:

[0052] This invention employs a three-level degradation segmentation strategy, which automatically cascades and calls segments according to a predetermined priority. When the current level algorithm fails to output a valid segmentation result, it automatically degrades to the next level algorithm until a valid output is obtained. This avoids the insufficient adaptability of a single segmentation algorithm to the diverse morphologies of organoids, thereby achieving a processing coverage rate of over 95% across a wide range of organoid morphologies. This ensures that usable segmentation results can be obtained under any input conditions. Furthermore, by recording the actual segmentation level used for subsequent analysis and confidence correction, quantitative traceability of segmentation quality is achieved.

[0053] This invention utilizes paired green fluorescent protein and red fluorescent protein fluorescence images to generate gold standard segmentation labels. Using the multidimensional morphological texture fingerprint features of the same instance on a bright field image as input, an ensemble tree classification model is trained to establish a mapping function from bright field morphological texture features to organoid survival probability. This enables the direct prediction of the survival status of each organoid when only a single bright field input is available, thus enabling the determination of organoid viability under fluorescence-free conditions.

[0054] This invention can systematically quantify organoid phenotypes from six dimensions: morphological features, gray-level distribution features, gray-level co-occurrence matrix texture features, edge gradient statistical features, radial intensity ratio features, and multi-scale granularity features. It also uses a dual-testing mechanism of two-group and multi-group comparisons, combined with Cohen's d effector for joint screening, thereby achieving a systematic quantitative method and statistical significance assessment of the effect of drugs on organoid morphology, which can significantly reduce the risk of false positives.

[0055] This invention utilizes a fully automated data preparation pipeline for 96-well plates. Based on automatic grid positioning and boundary shrinking, it extracts images of each well. A sliding window is used to segment and filter the high-resolution image of the entire plate, automatically generating gold or silver standard segmentation labels and completing dataset construction. This achieves automated processing of the entire 96-well plate image from well extraction, pseudo-label generation, dataset construction to deep learning model training, thus improving the data processing efficiency of high-throughput drug sensitivity screening. Attached Figure Description

[0056] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention, but do not constitute a limitation thereof. In the drawings:

[0057] Figure 1 This is a flowchart of the overall method of the present invention.

[0058] Figure 2 This is a flowchart of the three-level degradation segmentation process of the present invention.

[0059] Figure 3 This is a flowchart of the death score determination process of the present invention.

[0060] Figure 4 This is a flowchart of the automatic data preparation process for the 96-well plate of the present invention. Detailed Implementation

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] This invention provides, for example Figures 1-4 The method for quantitative analysis of organoid drug sensitivity based on morphological texture fingerprints, as shown, includes the following steps:

[0063] S1. Obtain bright-field microscopic images of the 96-well plate organoid culture system. For some experimental wells, simultaneously acquire the corresponding paired green fluorescent protein (GFP) and red fluorescent protein (RFP) fluorescence images. The bright-field images have a resolution of 3072 pixels by 4096 pixels, are stored in TIFF format, and have a grayscale depth of 16 bits.

[0064] S2. For the bright-field microscopic image obtained in step S1, automatic cascaded segmentation is performed according to a predetermined priority order: the first-level deep learning segmentation algorithm, the second-level general deep learning segmentation algorithm, and the third-level classical computer vision segmentation algorithm. First, the first-level YOLOv8-seg instance segmentation model is attempted, with a confidence threshold of 0.25 and a non-maximum suppression cross-union ratio threshold of 0.45. Inference is performed to obtain instance masks. If the first-level model detects at least one valid instance, the first-level segmentation result is directly used as the output, and the segmentation level identifier is recorded. If the first level does not detect a valid instance, it automatically downgrades to the second-level Cellpose general segmentation model, attempting to load sub-models for inference in the priority order of cyto3, cyto2, and cyto. If the second level still has no valid output, it automatically downgrades to the third-level classical computer vision segmentation algorithm, sequentially performing background correction, Gaussian smoothing, Otsu global thresholding, morphological cleaning, distance transformation, watershed separation of adhered instances, and area filtering to ensure that organoid instance masks can be obtained under any input conditions.

[0065] S3. For each organoid instance obtained from step S2, extract a multidimensional enhanced phenotypic feature vector from the bright-field micrograph. This feature vector contains 12 basic morphological texture features and 18 enhanced features, for a total of 30 dimensions. The features are organized into a feature vector in a predetermined order.

[0066] S4. Based on the 30-dimensional enhanced phenotypic feature vector extracted in step S3, calculate the drug response metric for each organoid instance. When paired fluorescence images are available, a cross-modal knowledge transfer approach is used to obtain the organoid survival probability; that is, a gold standard live / dead label is generated based on the fluorescence image, and a gradient boosting classification model is trained using bright-field features as input and fluorescence live / dead labels as supervisory signals. The trained model is then used to predict the survival probability of each instance as the drug response metric. When only bright-field images are available, the survival probability is directly output from the trained model. Simultaneously, a weighted mortality score is calculated based on apoptosis pathway signals, mitochondrial function signals, and morphological signals as a supplementary indicator to the drug response metric.

[0067] S5. Based on the drug response metrics obtained in step S4, generate a drug sensitivity analysis report. The report includes descriptive statistics for each experimental condition, inter-group significance test results, and multi-dimensional visualization output. Features that meet the criteria of a significance probability value less than a preset significance level and an absolute value of the Cohen d effect size greater than 0.5 are marked as significant change features and output to the analysis report.

[0068] This embodiment presents a complete closed-loop process through steps S1 to S5, from image acquisition, three-level degradation segmentation, multi-dimensional feature extraction, drug response metric calculation to pharmacodynamic analysis report generation, verifying its complete feasibility and smoothness in a real organoid drug sensitivity testing scenario. It can automatically complete the entire processing chain from raw images to pharmacodynamic evaluation reports without manual intervention, thus providing an automated analysis method for high-throughput drug sensitivity screening.

[0069] In a preferred embodiment, the specific process of three-level degradation instance segmentation is as follows:

[0070] First, the preprocessed grayscale image is converted into a three-channel pseudo-color image, and a lightweight YOLOv8-seg instance segmentation model fine-tuned with organoid data is loaded for inference. During inference, the confidence threshold is set to 0.25, and the cross-union threshold for non-maximum suppression is set to 0.45. The backbone network of the model extracts features layer by layer from the input image, the neck network fuses multi-scale features, and the head network predicts bounding boxes and instance masks simultaneously. The binary mask of each instance is extracted from the prediction results and filtered according to the area range. If at least one valid instance exists after filtering, the segmentation label image is directly output and the segmentation level is recorded as Level 1, ending the segmentation process. If no valid instance is detected, the process automatically proceeds to Level 2 segmentation.

[0071] Then, after the first-level segmentation fails, the second-level Cellpose model segmentation is automatically activated. First, image brightness inversion is determined by obtaining the average grayscale value of the region inwards by 10% from each edge and the average grayscale value of the central region by 50%. If the edge average is greater than 1.05 times the central average, the organoid is identified as a dark target, and the image pixel values ​​are inverted to make the dark target a bright target, thus adapting to the training data distribution characteristics of the Cellpose model. Next, automatic diameter estimation is performed by obtaining the area of ​​each connected region in the image, multiplying the median area by 1.5, and taking the square root as the estimated diameter. When loading the Cellpose model, it is tried sequentially according to the priority order of the cyto3 sub-model, cyto2 sub-model, and cyto sub-model. Segmentation inference is performed using the currently highest priority available sub-model, with the flow field threshold set to 0.4 and the cell probability threshold set to 0.0. After inference, post-processing is performed on the segmentation results: first, small holes with an area less than 50 pixels are filled; then, instances are filtered by area range; finally, large-area adhered instances are separated using a secondary watershed separation based on distance transformation. If at least one valid instance is found after processing, the second-level segmentation result is used as the final output, and the segmentation level is recorded as second level. If no valid instance is found, the process automatically proceeds to the third-level segmentation.

[0072] Finally, if no valid instances are detected in the second stage, the third stage of classical computer vision segmentation is initiated as a fallback. First, background correction is performed. When the overall brightness of the organoid is lower than the background, morphological grayscale closing operations are used to estimate the background and subtract it from the original image; otherwise, opening operations are used to estimate the background. Then, Gaussian smoothing is applied to the corrected image, prioritizing global thresholding for binarization, and downgrading to adaptive local thresholding if this fails. Subsequently, morphological closing and opening operations are performed sequentially to remove noise and excessively small objects. Local maxima are detected using distance transform as markers, and the watershed algorithm is used to separate adhered instances. Finally, the segmentation result is output after area filtering.

[0073] This embodiment verifies the effectiveness and robustness of the three-level degradation segmentation strategy on organoid bright-field images by sequentially cascading the first-level YOLOv8-seg, the second-level Cellpose, and the third-level classical computer vision segmentation algorithm. When the first-level deep learning model encounters a special-shaped organoid outside the training data distribution, it automatically degrades to the second level. When the second level also fails, the third-level classical algorithm based on low-level image features serves as a fallback, ensuring usable segmentation results can be obtained under any input conditions. This achieves processing coverage of over 95% across diverse organoid morphologies. Furthermore, by recording the actual segmentation levels used, subsequent analysis can be performed on confidence levels based on differences in the quality of the segmentation methods, enabling traceable management of segmentation accuracy. Additionally, the automatic brightness reversal mechanism for dark targets allows the Cellpose model to adapt to diverse image features where the organoid's brightness direction differs from the background.

[0074] In a preferred embodiment, the specific process of extracting multidimensional enhanced phenotypic feature vectors includes:

[0075] Based on the instance mask, calculate the morphological features of the region, such as area, perimeter, roundness, solidity, aspect ratio, and eccentricity; calculate the average gray intensity, gray intensity standard deviation, average gray intensity after background correction, and gray standard deviation after background correction within the mask region; calculate the edge density, texture entropy, and local contrast within the mask region.

[0076] Extract the gray value sequence of all foreground pixels within the mask area, and calculate the skewness, kurtosis, 10th percentile, 25th percentile, 75th percentile, 90th percentile, and interquartile range.

[0077] Within the instance bounding box, the gray level is quantized to 32 levels. The gray-level co-occurrence matrix is ​​calculated based on distances of 1 and 3, and angles of 0, 45, 90, and 135 degrees. After symmetric normalization, contrast, homogeneity, energy, correlation, and difference are extracted respectively, and the mean of each angle is taken.

[0078] The gradient magnitude of the entire image is calculated based on the edge detection operator, and the mean, standard deviation and maximum value of the gradient magnitude are extracted within the mask region;

[0079] Calculate the normalized distance from each pixel within the instance to the centroid, divide the inner and outer circles with half of the normalized radius as the boundary, and calculate the ratio of the average gray level of the inner circle to the average gray level of the outer circle.

[0080] Calculate the mean local variance at two scales: 3 pixels and 5 pixels.

[0081] When conducting cross-condition drug comparisons, the above 30-dimensional characteristics are normalized using standard scores to eliminate dimensional differences.

[0082] This embodiment extracts multidimensional enhanced phenotypic feature vectors, enabling the use of morphological parameters such as region area, roundness, and solidity to reflect changes in the regularity of organoid volume and shape in drug response; gray-level distribution features such as skewness, kurtosis, and percentiles to reflect the response of optical density distribution features to drug action; five texture quantities—contrast, homogeneity, energy, correlation, and variability—of the gray-level co-occurrence matrix to reflect changes in the uniformity and regularity of gray-level distribution within the organoid; the mean, standard deviation, and maximum value of the Sobel edge gradient to reflect the response of organoid boundary sharpness to drug action; the radial intensity ratio to reflect changes in the differences between the organoid core and peripheral structures; and the mean of multi-scale local variance to reflect changes in the scale information of internal substructures. This allows for quantitative comparison of the global phenotypic effect patterns of drug action, overcoming the limitations of traditional methods that assess drug efficacy using only a single indicator.

[0083] In a preferred embodiment, the specific implementation process of generating gold standard labels using paired fluorescence images and training a bright-field classification model includes:

[0084] Search for triplet files with the same number suffix in the bright field image directory, green fluorescent protein image directory, and red fluorescent protein image directory according to the experimental data number, and establish a pairing mapping relationship;

[0085] For each pair of green fluorescent protein and red fluorescent protein images, the following processing is performed: The pixel values ​​of the two channels are added point by point to obtain a fused fluorescence signal map. Morphological opening operations are performed on the fused fluorescence signal map to estimate the background. Negative values ​​are truncated by subtracting the estimated background image from the original image to achieve background correction. After Gaussian smoothing with a standard deviation of 2, binarization is performed using the Otsu method to obtain a mask of the fluorescent positive region. Then, morphological closing and opening operations are performed to finely clean the mask boundaries, removing isolated objects with an area less than 20 pixels. A distance transformation is performed on the mask, and a local maximum is found on the distance map using a search window of 15×15 pixels as a seed point. Adhesive instances are split using the watershed algorithm and filtered according to an area threshold of 20-5000 pixels. The binary mask contour of each instance is extracted into a polygon, and a polygon approximation algorithm is used to simplify the contour points. The coordinates of each vertex are normalized to the 0-1 interval to generate a gold standard segmentation label.

[0086] For each fluorescence segmentation instance, background correction is performed on both the green fluorescent protein (GFP) and red fluorescent protein (RFP) channels. The instance mask is expanded outwards three times, and the outer ring region is selected. The average pixel value within this outer ring region is calculated as the background estimate. The background estimate is subtracted from each pixel value within the instance mask region, with negative values ​​truncated to zero, to obtain the corrected foreground pixel values. When calculating the live cell score, the sum of the GFP-corrected pixel values ​​is divided by the sum of the GFP-corrected pixel values ​​and the red fluorescent protein-corrected pixel values. A live cell score ≥ 0.6 is considered alive; a live cell score ≤ 0.35 is considered dead; and a live cell score > 0.35 and < 0.6 is considered apoptotic.

[0087] On the bright-field image at the same location, a 30-dimensional enhanced phenotypic feature vector was extracted using the same instance mask as for fluorescence segmentation. The extracted features were used as columns of the input matrix, with the fluorescence gold standard's live / dead label as the classification target variable and the continuous live cell score as the regression target variable. The gradient boosting algorithm was used to train the classifier and regressor separately, and the model performance was evaluated using five-fold cross-validation. After training, the classifier and regressor were saved as persistent files using object serialization technology.

[0088] If the paired fluorescence image cannot be read normally due to file compression corruption or other reasons, the sample will be automatically downgraded from the gold standard mode to the silver standard mode. That is, the use of fluorescence image for segmentation will be abandoned, and instead a three-level downgrade segmentation method will be used on the corresponding bright field image to generate segmentation labels, and the source of the labels will be marked as the silver standard.

[0089] In subsequent inference scenarios with only bright-field image input, the persistent classifier model is directly loaded, and 30-dimensional enhanced phenotypic features are extracted from the input bright-field image before the model is called for prediction. The survival probability and live / dead category of each organoid instance can be output without fluorescence staining.

[0090] This embodiment utilizes paired green and red fluorescent protein fluorescence images to generate high-precision gold-standard segmentation labels and continuous-value live cell scores. It extracts 30-dimensional enhanced phenotypic features of the same instance from corresponding bright-field images. Using the fluorescent live cell scores as a supervisory signal, it trains an ensemble tree model, establishing a reliable mapping function from bright-field morphological texture features to organoid survival probability. This allows for direct prediction of survival status during subsequent pure bright-field inference without the need for fluorescence staining. Furthermore, through a multi-level fault-tolerant reading mechanism for corrupted fluorescent TIFF files and a strategy of automatically downgrading the gold standard to the silver standard, it avoids data waste caused by the loss of the entire sample due to individual file corruption, thereby improving the efficiency of utilizing valuable experimental data.

[0091] In a preferred embodiment, the specific process of calculating the weighted mortality score includes:

[0092] The background correction intensity of apoptosis markers was normalized to a preset upper limit, and combined with the apoptosis positive determination value and the nuclear texture entropy normalization value, and fused according to a preset weight to obtain the death sub-signal of the apoptosis pathway.

[0093] The mitochondrial functional death sub-signal is obtained by fusing the mitochondrial membrane potential loss degree and mitochondrial punctate aggregation degree according to a preset weight.

[0094] The morphological death sub-signal is obtained by fusing nuclear fragmentation degree, area shrinkage penalty, shape irregularity and nuclear texture entropy according to preset weights.

[0095] The apoptosis pathway death sub-signal is multiplied by the first weight, the mitochondrial functional death sub-signal is multiplied by the second weight, and the morphological death sub-signal is multiplied by 0.2, and then summed to obtain the comprehensive death score. The sum of the first weight and the second weight is 1. Under the default configuration, the apoptosis pathway weight is 0.6 and the mitochondrial functional weight is 0.4.

[0096] When the overall mortality score is greater than 0.45 and the apoptosis intensity is greater than 0.3, or the overall mortality score is greater than 0.65, or the apoptosis is positive and the area is less than 300, death is determined.

[0097] Apoptosis is defined as follows: a comprehensive mortality score greater than 0.25 and not greater than 0.45; or apoptosis is considered true and the instance area is greater than or equal to 300 pixels; or mitochondrial membrane potential loss is greater than 0.2 and less than or equal to 0.55, and the normalized apoptosis intensity value is greater than 0.15. Survival is defined as a comprehensive mortality score less than 0.25; all other cases are considered uncertain. Each state is assigned a corresponding confidence level.

[0098] In a three-dimensional parameter space consisting of the apoptosis death threshold range of 0.25-0.65, the apoptosis threshold range of 0.15-0.4, and the apoptosis pathway weight range of 0.4-0.8, a grid search is performed with the composite score as the optimization objective to automatically select the optimal parameter combination.

[0099] This embodiment comprehensively assesses cell death from three dimensions: apoptosis pathway signals, mitochondrial functional signals, and morphological signals. A comprehensive death score is calculated through adjustable weighted fusion, reducing the risk of misjudgment due to a single signal source and providing multiple pieces of evidence to support the determination of cell death. Furthermore, the three-category classification logic of apoptosis, death, and survival distinguishes between intermediate apoptotic states and clearly defined dead / live states, thus providing more refined biological information for drug efficacy evaluation than a binary classification approach.

[0100] In a preferred embodiment, the drug sensitivity analysis report includes statistical comparative analysis results across experimental conditions, and the process of obtaining the statistical comparative analysis specifically includes:

[0101] Bright-field images were divided into two groups according to experimental conditions: a control group and a drug treatment group, with each group containing 6 images;

[0102] For all instances within the two groups, calculate the mean, standard deviation, median, first quartile, third quartile, interquartile range, minimum and maximum values, and sample size for each of the 30 feature dimensions.

[0103] Independent samples t-tests, Mann-Whitney U tests, and Cohen d-effect size calculations were performed on each of the 30 feature dimensions. The independent samples t-test was used to calculate the test statistic and the corresponding significance probability value. The Mann-Whitney U test was used to calculate the test statistic and the corresponding significance probability value, which served as the basis for priority reporting. The Cohen d-effect size was calculated by dividing the difference between the two group means by the pooled standard deviation, thereby quantifying the actual magnitude of the change.

[0104] Features that simultaneously satisfy the Mann-Whitney U test with a significance probability value less than 0.05 and an absolute value of the Cohen d effect size greater than 0.5 are selected as significant change features for output.

[0105] When three or more experimental conditions are involved, one-way ANOVA and Kruskal-Wallis H test are performed for global significance testing, and Mann-Whitney U test is used for post-hoc pairwise comparisons between each experimental group and the control group. The same screening criteria are used to output the significant change characteristics.

[0106] This embodiment provides a dual verification mechanism that combines parametric and nonparametric tests, thus balancing testing power and robustness. Furthermore, the strategy of jointly screening using Cohen's d effect size and significance probability ensures that the identified changes are not only statistically significant but also possess a moderate to high magnitude of actual biological effect. This effectively avoids the problem of small, meaningless changes being judged as significant even with a large sample size when relying solely on significance probability.

[0107] In a preferred embodiment, the complete process for automated data preparation of a 96-well plate includes:

[0108] Based on an 8×12 grid layout and shrinkage rate, the row and column spacing between adjacent holes is calculated. After determining the center coordinates of all 96 holes, a square region of predetermined side length is cut out using the center coordinates of each hole as the center. This region is then scaled to 256×256 pixels using bilinear interpolation and saved as a single-hole image according to a predetermined naming rule. Holes with a grayscale variance less than a predetermined threshold are identified as empty holes and automatically skipped.

[0109] For high-resolution full-panel images, a non-overlapping sliding window with a stride equal to the window size is used for partitioning. The window size is set to 1024×1024 pixels, and each full-panel image can be partitioned into 6 image blocks. Each image block is scaled to a uniform resolution of 256×256 pixels using region interpolation, and the grayscale values ​​are normalized from 16-bit precision to 8-bit precision. The target pixel percentage is calculated for each image block, and filtering is performed. If the target pixel percentage is less than 0.003, it is determined to be an empty image block and skipped; if the target pixel percentage is greater than 0.8, it is determined to be an abnormal image block and skipped; if the number of instances in the image block is zero, it is skipped; empty image blocks with too low a percentage, abnormal image blocks with too high a percentage, and image blocks without any instances are excluded.

[0110] The filtered image patches and their corresponding label files are organized in YOLO format. A file path list method is used to achieve proportional hierarchical random splitting of the training and validation sets: generating lists of training and validation files, with the training set accounting for 85% and the validation set for 15%. The random seed is fixed at 42 to ensure reproducibility. This avoids physical copying of files, achieving zero-copy data management.

[0111] This embodiment utilizes an automatic well location algorithm based on an 8×12 grid and shrinkage ratio, eliminating the need for manual labeling of well coordinates and achieving fully automated extraction from the entire plate image to individual well images. Simultaneously, a non-overlapping sliding window partitioning strategy systematically transforms the high-resolution entire plate image into uniformly sized image blocks, while dual threshold filtering based on target pixel proportions ensures the quality of the data entering the dataset. A dual-standard labeling strategy maximizes the comprehensive utilization of valuable fluorescence pairing data and a large amount of bright-field data. This eliminates the need for manual intervention throughout the entire process from inputting the high-resolution entire plate image to outputting the YOLO training dataset, thereby improving the efficiency of high-throughput 96-well plate drug sensitivity testing data processing and reducing repetition rates.

[0112] In a preferred embodiment, it further includes at least one of active learning or CellTiter-Glo gold standard calibration;

[0113] The specific process of active learning is as follows:

[0114] Run the current model to predict unlabeled instances and obtain the predicted probabilities for each category. Calculate the Shannon entropy of the predicted probability distribution for each instance; a higher entropy value indicates greater uncertainty in the model's prediction of that instance. Sort the instances by entropy value from highest to lowest, and select a predetermined number of the top-ranked instances to submit to the user for manual annotation.

[0115] After users submit their annotation results, the system checks the cumulative number of manually annotated samples in the database. When the number of new samples reaches a preset minimum sample size threshold, the model retraining process is triggered, merging all historical annotated data with the new annotated data and retraining the gradient boosting classifier.

[0116] Retain the three most recently trained model versions. If performance degradation is detected with the new model, you can choose to roll back to any historical version. When more than three versions are stored, delete the oldest historical version in chronological order.

[0117] The specific process for CellTiter-Glo gold standard calibration is as follows:

[0118] Ordinal-preserving regression was used to perform monotonicity calibration on the original predicted values ​​of the model output. The output range was set to a lower limit of 0 and an upper limit of 1.2 to ensure that the calibrated predicted values ​​maintained monotonic consistency with the CTG measured values ​​and did not change the relative ranking between samples.

[0119] 24-dimensional image-level convergent features were extracted for each image. Leave-one-out cross-validation was employed, and four models—ridge regression, lasso regression, random forest regression, and gradient boosting regression—were used for training and prediction. The optimal regression model was selected as the second-level calibrator based on the minimum mean absolute error. Evaluation metrics included Pearson correlation coefficient, Spearman rank correlation coefficient, mean absolute error, root mean square error, and threshold accuracy within each predetermined tolerance range.

[0120] The active learning mechanism in this embodiment utilizes the Shannon entropy sampling strategy to prioritize the samples with the highest uncertainty in the current model prediction and submit them for manual review, thereby maximizing information gain with minimal annotation cost. At the same time, the CellTiter-Glo gold standard calibration provides a complete quantitative mapping path from the model's original prediction to the gold standard-comparable calibration value, thereby improving the absolute accuracy of the prediction results.

[0121] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for quantitative analysis of organoid drug sensitivity based on morphological texture fingerprints, characterized in that, Includes the following steps: S1. Obtain bright-field microscopic images of the 96-well plate organoid culture system and optional paired green fluorescent protein and red fluorescent protein fluorescence microscopic images. S2. The bright-field microscopic image is segmented using a three-level degradation segmentation method to obtain the mask of each organoid instance; S3. For each organoid instance obtained from the segmentation, extract the multidimensional enhanced phenotypic feature vector from the bright-field micrograph. S4. Based on the multidimensional enhanced phenotypic feature vector, calculate the drug response metric for each organoid instance to reflect the degree of phenotypic change of the organoid under drug action. S5. Generate a drug sensitivity analysis report based on the drug response metrics.

2. The method for quantitative analysis of organoid drug sensitivity based on morphological texture fingerprints according to claim 1, characterized in that, In step S2, the three-level degradation separation method specifically includes: The bright-field microscopic image is automatically cascaded and segmented according to a predetermined priority order of the first-level deep learning segmentation algorithm, the second-level deep learning general segmentation algorithm, and the third-level classical computer vision segmentation algorithm. If the current level segmentation algorithm outputs a valid result, it returns directly; otherwise, it automatically downgrades to the next level algorithm to ensure that the organoid instance mask can be obtained under all input conditions.

3. The method for quantitative analysis of organoid drug sensitivity based on morphological texture fingerprints according to claim 2, characterized in that, In step S2: The first-level deep learning segmentation algorithm is the YOLOv8-seg instance segmentation model, with a confidence threshold and a cross-union ratio threshold for non-maximum suppression configured. The second-level deep learning general segmentation algorithm is the Cellpose model, and the Cellpose model adopts an automatic downgrade calling method that executes the Cyto3 model, Cyto2 model and Cyto model in order of priority. When the segmentation of the previous model fails, the next model is automatically activated. The third-level classical computer vision segmentation algorithm includes background correction, Gaussian smoothing, Otsu global thresholding, morphological cleaning, distance transformation and watershed separation of adhering instances, and area filtering, all performed sequentially.

4. The method for quantitative analysis of organoid drug sensitivity based on morphological texture fingerprints according to claim 3, characterized in that, In step S3, the multidimensional enhanced phenotypic feature vector includes: 12-dimensional basic morphological texture features: region area, roundness, solidity, aspect ratio, eccentricity, average gray intensity, standard deviation of gray intensity, mean gray value after background correction, standard deviation of gray value after background correction, Sobel edge density, texture entropy, and local contrast. 18-dimensional enhancement features: 7-dimensional grayscale distribution features, 5-dimensional GLCM texture features, 3-dimensional Sobel edge gradient statistics, 1-dimensional radial intensity ratio, and 2-dimensional multi-scale granularity.

5. The method for quantitative analysis of organoid drug sensitivity based on morphological texture fingerprints according to claim 4, characterized in that, In step S4, the drug response metric includes the organoid survival probability, which is obtained through cross-modal knowledge transfer, specifically including: When there are green fluorescent protein fluorescence images and red fluorescent protein fluorescence images paired with the bright-field microscopic image, a gold standard segmentation label and a life-or-death score are generated based on the paired fluorescence images. The multidimensional enhanced phenotypic feature vector is extracted on the same instance mask of the corresponding bright-field image. The ensemble tree model is trained with the life-or-death score as a supervision signal. A mapping function from bright-field morphological texture features to organoid survival probability is established, so that the survival status of each organoid can be directly predicted when only bright-field image input is available. When the paired fluorescence image cannot be read, the sample is automatically downgraded from the gold standard mode to the silver standard mode, and a replacement label is generated using the third-level classical computer vision segmentation algorithm.

6. The method for quantitative analysis of organoid drug sensitivity based on morphological texture fingerprints according to claim 5, characterized in that, The specific methods for generating the gold standard segmentation label include: Acquire paired green fluorescent protein fluorescence images and red fluorescent protein fluorescence images, and segment the fluorescence images to obtain an instance mask; Calculate a live cell score for each instance, which is equal to the sum of the green fluorescent protein-corrected pixel values ​​within the masked area of ​​that instance divided by the sum of the green fluorescent protein-corrected pixel values ​​and the red fluorescent protein-corrected pixel values; According to the gold standard segmentation label for live cell scoring, cells are considered alive when the live cell score reaches the maximum score threshold, dead when the live cell score reaches the minimum score threshold, and apoptosis when the live cell score is between the maximum and minimum score thresholds.

7. The method for quantitative analysis of organoid drug sensitivity based on morphological texture fingerprints according to claim 1, characterized in that, In step S4, the drug response metric includes a weighted mortality score, which is calculated in the following manner: Calculate the apoptotic pathway death sub-signals, mitochondrial functional death sub-signals, and morphological death sub-signals, respectively. Weighted mortality scores are calculated through adjustable weight fusion. When the weighted mortality score is greater than the maximum mortality score threshold and the normalized apoptosis intensity value is greater than the normalized threshold, death is determined. Apoptosis is determined when the weighted mortality score is between the maximum mortality score threshold and the minimum mortality score threshold. A person is considered alive when the weighted mortality score is less than or equal to the minimum mortality score threshold.

8. The method for quantitative analysis of organoid drug sensitivity based on morphological texture fingerprints according to claim 1, characterized in that, In step S5, the drug sensitivity analysis report includes statistical comparative analysis results across experimental conditions. The process of obtaining the statistical comparative analysis specifically includes: The images were grouped according to the experimental conditions, and the descriptive statistics of each multidimensional enhanced phenotypic feature within each group were calculated. For the two groups, perform the independent samples t-test and the Mann-Whitney U test; For multiple group comparisons, one-way ANOVA and Kruskal-Wallis H test were performed, and post-hoc pairwise comparisons between each experimental group and the control group were conducted using the Kruskal-Wallis H test. Calculate the Kruskal-Wallis H test; Features that satisfy the condition of having a significance probability value less than the preset significance level and an absolute value of the Cohen d effect size greater than 0.5 are selected as significant change features for output.

9. The method for quantitative analysis of organoid drug sensitivity based on morphological texture fingerprints according to claim 1, characterized in that, This also includes the preparation of automated data for 96-well plates, specifically including: Based on an 8×12 grid and shrinkage ratio, 96 holes are automatically located, and images of each single hole are cut from the high-resolution image of the whole board and scaled to a predetermined size. For high-resolution full-panel images, a non-overlapping sliding window partitioning with a step size equal to the window size is used. After scaling the image blocks to the target resolution, content filtering is performed based on the target pixel ratio to exclude empty image blocks and abnormally dense image blocks. A file path list is used to achieve proportional hierarchical random splitting of the training and validation sets, avoiding physical copying of files.

10. The method for quantitative analysis of organoid drug sensitivity based on morphological texture fingerprints according to claim 1, characterized in that, It also includes at least one of active learning or CellTiter-Glo gold standard calibration; The active learning specifically includes: selecting manually reviewed targets from unlabeled samples with the highest prediction uncertainty based on the Shannon entropy sampling strategy; triggering model retraining when the number of labeled samples reaches the minimum threshold; and retaining multiple recent historical model versions to support version rollback. The CellTiter-Glo gold standard calibration specifically includes: using ordinal-preserving regression to perform first-level monotonic calibration on the original predicted values ​​of the model, and then using at least one of Ridge regression, Lasso regression, random forest or gradient boosting regression to perform leave-one-out cross-validation on the 24-dimensional image-level convergent features, and selecting the best model as the second-level calibrator based on minimizing the mean absolute error.

Images