An organoid brightfield image analysis method based on instance segmentation

Abstract

The application provides an organoid bright-field image analysis method based on instance segmentation, which inputs an organoid bright-field image into an organoid vertical model for instance segmentation to identify an instance graph corresponding to a living organoid cell; performs morphological feature extraction according to the instance graph to obtain geometric features and texture features of the living organoid cell; determines evaluation index data of a prediction curve including a cell survival rate and a drug concentration response according to quantization data of the geometric features and the texture features; generates a drug efficacy evaluation report according to the instance graph, the quantization data of the geometric features and the texture features, and the prediction curve; improves segmentation robustness in a complex scene, establishes a multi-dimensional morphological feature extraction index system, comprehensively considers multi-dimensional parameters, more comprehensively reflects drug action effects, integrates multifunction from image instance segmentation, feature extraction to dose-response curve fitting, and realizes an organoid drug sensitivity detection scheme with more information dimensions and cost benefits.

Description

TECHNICAL FIELD

[0001] Embodiments of the present application relate to the field of image analysis, and in particular to a method for bright field image analysis of organoids based on instance segmentation. BACKGROUND

[0002] As a three-dimensional cell culture model, organoids have important application value in drug screening, disease modeling and personalized medicine. The core of organoid drug efficacy evaluation lies in the quantitative analysis of the activity and morphological changes of organoids after drug action. Currently, there are two technical routes in the industry, namely the drug efficacy evaluation method based on biochemical detection and the drug efficacy evaluation method based on image analysis.

[0003] For the drug efficacy evaluation method based on biochemical detection, cell viability is reflected by detecting cell metabolic activity. The commonly used method is ATP detection. The brief operation process is as follows: first, the organoids are inoculated in a multi-well plate and exposed to different concentrations of the drug to be tested. After a certain period of culture, the detection reagent is added. Then the living cells in the organoids convert the reagent into a colored product, and the absorbance (OD value) is measured at a specific wavelength using an enzyme-labeled instrument. Then the cell survival rate is calculated according to the OD value: survival rate (%) = (experimental group OD value / control group OD value) × 100%. Finally, data analysis is performed, with drug concentration as the horizontal coordinate and cell survival rate as the vertical coordinate, to draw a dose-response curve, and the IC50 value is calculated by fitting the curve with a four-parameter Logistic model.

[0004] For the drug efficacy evaluation method based on image analysis, basic image analysis tools such as ImageJ and CellProfiler are used to perform rough analysis of organoids through traditional computer vision methods such as threshold segmentation and edge detection. Deep learning segmentation methods such as Mask R-CNN and other general instance segmentation models are tried for biological image analysis.

[0005] However, in the drug efficacy evaluation method based on biochemical detection, the enzyme-labeled instrument detection method has inherent destructive detection, which requires termination of culture and addition of chemical reagents, and cannot perform continuous observation and time series analysis on the same batch of organoids. At the same time, this detection method only provides overall metabolic activity data, with single information dimension, and cannot obtain important information such as morphological changes and spatial distribution heterogeneity of organoids. In terms of cost, special detection reagent kits need to be purchased, increasing the experimental cost. This method also has potential interference, as some drugs may themselves affect the metabolism of the detection reagent, causing false positive or false negative results. In addition, this method requires additional experimental operations, which are prone to introduce operational errors, biological errors, etc., affecting the stability of the results.

[0006] In the image analysis-based drug efficacy evaluation method, the contrast of bright field images is sensitive to illumination changes, the segmentation effect is unstable, and it is difficult to process overlapping and edge blurred organoids. At the same time, the automation degree of this method is low, a large number of parameter adjustments are needed, and it is difficult to realize high-throughput analysis. For general deep learning models, the pre-trained model is not optimized for organoid bright field images, and it performs poorly in complex scenarios. At the same time, the process is not complete enough, only the segmentation task is completed, and the subsequent drug efficacy evaluation and curve fitting functions are lacking. SUMMARY

[0007] The following is a summary of the subject matter described in detail herein.

[0008] The purpose of the present application is to at least partially solve one of the technical problems existing in the related art, and the embodiment of the present application provides an organoid bright field image analysis method based on instance segmentation.

[0009] In an embodiment of the present application, an organoid bright field image analysis method based on instance segmentation comprises:

[0010] Obtain organoid bright field images of multiple drugs at different concentrations;

[0011] Input the organoid bright field images into a trained organoid vertical model for instance segmentation to identify instance maps corresponding to surviving organoid cells;

[0012] According to the instance maps, morphological feature extraction is performed to obtain geometric features and texture features of the surviving organoid cells;

[0013] Determine evaluation index data according to the quantitative data of the geometric features and the texture features, the evaluation index data including a predicted curve of cell viability and drug concentration response;

[0014] Generate a drug efficacy evaluation report according to the instance maps corresponding to the surviving organoid cells, the quantitative data of the geometric features and the texture features, and the predicted curve of cell viability and drug concentration response.

[0015] According to certain embodiments of the present application, the organoid vertical model comprises a feature extraction network, a region proposal network, a region of interest alignment network, and a mask prediction head;

[0016] The organoid bright field images are input into the trained organoid vertical model for instance segmentation to identify instance maps corresponding to surviving organoid cells, comprising:

[0017] Convolve the organoid bright field images through the feature extraction network to obtain feature maps;

[0018] Generate a region of interest according to the feature maps through the region proposal network;

[0019] aligning features of the region of interest by the region of interest alignment network;

[0020] predicting a segmentation mask of the aligned region of interest by the mask prediction head to obtain a binary mask, and identifying an instance graph corresponding to the surviving organoid cells according to the binary mask.

[0021] According to certain embodiments of the present application, the geometric features include basic geometric features and complex geometric features; the basic geometric features include the area, the perimeter, and the circularity of the surviving organoid cells; and the complex geometric features include the concave-convex degree ratio of the surviving organoid cells and the width-height ratio of the mask bounding box.

[0022] According to certain embodiments of the present application, the texture features include the average black intensity,

[0023] The average black intensity is expressed as: μ_I = (1 / A)∑∑I(i,j)·M(i,j);

[0024] In the formula, μ_I is the average brightness, A is the area of the region of interest, I(i,j) is the pixel intensity value at position (i,j) in the image, and M(i,j) is the mask value at position (i,j) in the image.

[0025] According to certain embodiments of the present application, the cell viability is calculated according to the following formula: relative survival rate = (total area of the experimental group / total area of the negative control group) × 100%; wherein the total area of the experimental group and the total area of the negative control group are determined according to the quantification data of the geometric features.

[0026] According to certain embodiments of the present application, the determination of the evaluation index data according to the quantification data of the geometric features and the texture features comprises:

[0027] The first drug effectiveness auxiliary data as the evaluation index data is obtained according to the circularity change trend, the concave-convex degree ratio change trend, and the width-height ratio change trend of the mask bounding box.

[0028] According to certain embodiments of the present application, the determination of the evaluation index data according to the quantification data of the geometric features and the texture features comprises:

[0029] The second drug effectiveness auxiliary data as the evaluation index data is obtained according to whether the cell viability and the circularity change trend, the concave-convex degree ratio change trend, and the width-height ratio change trend of the mask bounding box present a synergistic change.

[0030] According to certain embodiments of the present application, the determination of the evaluation index data according to the quantification data of the geometric features and the texture features comprises:

[0031] The multi-dimensional pharmacodynamic score model is constructed as the evaluation index data by assigning weights to the quantized data of different geometric features and texture features.

[0032] According to some embodiments of the present application, the prediction curve of cell viability in response to drug concentration is obtained by curve fitting of a three-parameter Logistic model and a four-parameter Logistic model.

[0033] The three-parameter Logistic model is expressed as:

[0034] y = Bottom + (Top - Bottom) / (1 + 10^((LogIC50 - x) ×(-1))),

[0035] The four-parameter Logistic model is expressed as:

[0036] y = Bottom + (Top - Bottom) / (1 + 10^((LogIC50 - x) × HillSlope));

[0037] In the formula, y is the relative survival rate, x is the logarithm of drug concentration, Top is the asymptote on the curve, Bottom is the asymptote under the curve, LogIC50 is the logarithm value of half-inhibitory concentration, and HillSlope is the Hill slope.

[0038] According to some embodiments of the present application, the curve fitting quality is evaluated by a coefficient of determination,

[0039] The coefficient of determination is expressed as:

[0040] R²= 1-SS_res / SS_tot;

[0041] In the formula, R² is the coefficient of determination, SS_res is the residual sum of squares, and SS_tot is the total sum of squares.

[0042] The above scheme has at least the following beneficial effects: a non-destructive organoid pharmacodynamic evaluation method is provided, which makes it possible to continuously observe the same batch of samples based on bright field image analysis, while avoiding the use of chemical reagents. A high-precision organoid instance segmentation algorithm is realized, which is optimized for bright field image characteristics and improves the segmentation robustness in complex scenes. A multi-dimensional morphological feature extraction index system is established, which comprehensively evaluates parameters such as area, perimeter, circularity, and number, and more comprehensively reflects the effect of drug action. An end-to-end automated analysis process is constructed, which integrates complete functions from image instance segmentation, feature extraction to dose-response curve fitting. A more information-dimension and cost-effective organoid drug sensitivity detection scheme is realized, which provides technical support for drug screening and personalized medicine. BRIEF DESCRIPTION OF DRAWINGS

[0043] The accompanying drawings are used to provide a further understanding of the technical solutions of the present application, and constitute a part of the specification, and are used to explain the technical solutions of the present application together with the embodiments of the present application, and do not constitute a limitation on the technical solutions of the present application.

[0044] Figure 1 is a step diagram of the organoid bright field image analysis method;

[0045] Figure 2 is a sub-step diagram of step S200;

[0046] Figure 3 is a structural diagram of the organoid vertical class model. DETAILED DESCRIPTION

[0047] In order to make the purposes, technical solutions and advantages of the present application more clear, the present application is further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and do not limit the present application.

[0048] It should be noted that although the functional modules are divided in the device schematic diagram, and the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a manner different from the module division in the device or the order in the flowchart. The terms "first", "second", etc. in the specification, claims or above-described drawings are used to distinguish similar objects, and do not necessarily describe a specific order or sequence.

[0049] The embodiments of the present application are further described below in combination with the drawings.

[0050] The embodiments of the present application provide an organoid bright field image analysis method based on instance segmentation. It relates to deep learning visual technology and vertical class application model development, and in particular to a method of extracting morphological features of organoid bright field images and establishing a mathematical mapping thereof with drug efficacy.

[0051] Reference Figure 1 , the organoid bright field image analysis method comprises the following steps:

[0052] Step S100, acquiring organoid bright field images of a plurality of drugs at different concentrations;

[0053] Step S200, inputting the organoid bright field images into a trained organoid vertical class model for instance segmentation to identify an instance graph corresponding to the surviving organoid cells;

[0054] Step S300, performing morphological feature extraction according to the instance graph to obtain geometric features and texture features of the surviving organoid cells;

[0055] Step S400, determine evaluation index data according to the quantization data of geometric features and texture features, the evaluation index including a predicted curve of cell viability and drug concentration response;

[0056] Step S500, generate a pharmacodynamic evaluation report according to the instance graph corresponding to the organoid cell, the quantization data of geometric features and texture features, and the predicted curve of cell viability and drug concentration response.

[0057] In this embodiment, a non-destructive organoid pharmacodynamic evaluation method is provided, which makes it possible to continuously observe the same batch of samples based on bright field image analysis, while avoiding the use of chemical reagents. An organoid instance segmentation algorithm with high precision is implemented, which is optimized for bright field image characteristics to improve the segmentation robustness in complex scenarios. A multi-dimensional morphological feature extraction index system is established to comprehensively evaluate parameters such as area, perimeter, circularity, and number, providing a more comprehensive reflection of drug effect. An end-to-end automated analysis process is constructed, integrating complete functions from image instance segmentation, feature extraction to dose-response curve fitting. A more information-dimensioned and cost-effective organoid drug sensitivity detection scheme is realized, providing technical support for drug screening and personalized medicine.

[0058] For step S100, the organoids are inoculated in a multi-well plate and exposed to different concentrations of different drugs to be tested. After a certain period of culture, bright field images of the organoids are taken by allowing a transmission beam (direct electron) to pass through the objective diaphragm imaging in the transmission electron microscope.

[0059] For step S200, the organoid vertical model includes a feature extraction network, a region proposal network, a region of interest alignment network, and a mask prediction head. The organoid vertical model is based on a convolutional neural network of mask regions, which is adapted to various morphological images of organoids and compatible with texture heterogeneity requirements. By inputting a labeled organoid bright field image dataset for training, the surviving organoids in the image are identified, and finally the organoid vertical model is obtained for organoid instance segmentation under bright field images.

[0060] Reference Figure 2 and Figure 3 The organoid bright field image is input into the trained organoid vertical model for instance segmentation, and the instance graph corresponding to the surviving organoid cells is identified, including the following steps:

[0061] Step S210, convolve the organoid bright field image through the feature extraction network to obtain a feature map;

[0062] Step S220, generate a region of interest from the feature map through the region proposal network;

[0063] Step S230, align the features of the region of interest through the region of interest alignment network;

[0064] In step S240, a binary mask is obtained by predicting the segmentation mask of the region of interest of the aligned feature through the mask prediction head, and an instance graph corresponding to the living organ cell is identified according to the binary mask.

[0065] For the feature extraction network, a feature pyramid network ResNet-50-FPN is used as the backbone network. For the calculation of the feature map, let the input image be I ∈ R^(H×W×3), and the feature map F ∈ R^(H / s×W / s×C) is obtained after the convolution layer, where I is the input image, containing the original image data; H and W are the height and width of the input image; 3 is the number of color channels (RGB) of the input image; F is the feature map after the convolution operation, which is the feature representation of the input image; s is the downsampling rate, which controls the reduction ratio of the feature map size; C is the number of channels of the feature map, and the feature dimension of each spatial position.

[0066] For the region proposal network, a sliding window is used on the feature map to generate a region of interest (ROI), and k candidate boxes are generated for each anchor. A softmax classifier is used to determine whether the target is included.

[0067] For the region of interest alignment network, in order to solve the quantization error problem of ROI pooling, bilinear interpolation is used to accurately align the features, and finally a fixed-size feature map is output.

[0068] For the mask prediction head, a fully convolutional network (FCN) is used to predict the segmentation mask of each ROI, and finally a binary mask M ∈ {0,1}^(m×m) is output, indicating whether each pixel belongs to a high-rate organ. According to the binary mask, an instance graph corresponding to the living organ cell is identified.

[0069] Vertical fine-tuning training is performed according to the morphological characteristics of the organoids. Through the special collection and annotation of the organoid dataset, transfer learning is performed, which significantly improves the recognition accuracy of organoid instances in complex backgrounds, especially for the segmentation of fuzzy and mutually adhered organoids, which is significantly better than the general segmentation model.

[0070] It should be noted that in other embodiments, other models can also be used for instance segmentation, such as YOLO or Faster R-CNN detection framework combined with traditional image segmentation algorithms. Such solutions have an advantage in detection speed, but are not as good as the Mask R-CNN framework in terms of segmentation accuracy, especially for irregularly shaped organoid edge segmentation.

[0071] For step S300, morphological feature extraction is performed according to an example diagram to obtain geometric features and texture features of the surviving organoid cells.

[0072] The geometric features include basic geometric features and complex geometric features; the basic geometric features include area, perimeter and circularity of the surviving organoid cells; and the complex geometric features include concave-convex degree ratio of the surviving organoid cells and aspect ratio of the mask bounding box.

[0073] The area is represented as A = ∑∑M(i,j), where M is a binary mask. The perimeter is represented as P = ∑√(Δx²+ Δy²), which is calculated based on the contour point sequence. The circularity is represented as C = 4πA / P², which reflects the degree of shape close to a circle.

[0074] The concave-convex degree ratio is represented as R_conv = A / A_conv, where A_conv is the convex hull area. The aspect ratio is represented as AR = w_bbox / h_bbox, which is the ratio of the width to the height of the bounding box.

[0075] The texture features include average black solidity, which is represented as μ_I = (1 / A)∑∑I(i,j)·M(i,j).

[0076] In the formula, μ_I is the average brightness, A is the area of the region of interest, I(i,j) is the pixel intensity value at position (i,j) in the image, and M(i,j) is the mask value at position (i,j) in the image.

[0077] A morphological feature combination specifically for organoid drug efficacy evaluation is established, including geometric features (area, perimeter, circularity, convex area ratio, concave-convex degree ratio and aspect ratio) and texture features (average black solidity). The feature system accurately quantifies the morphological changes of organoids through mathematical formulas, providing comprehensive and objective data basis for drug efficacy evaluation.

[0078] For step S400, evaluation index data is determined according to the quantized data of the geometric features and the texture features.

[0079] On the one hand, a correlation mechanism between the features and the drug efficacy evaluation is constructed.

[0080] Area feature as the core pharmacodynamic indicator is directly used for relative survival rate calculation as a direct correlation application. Cell viability is calculated according to the following formula: relative survival rate = (total area of experimental group / total area of negative control group) x 100%; wherein the total area of experimental group and the total area of negative control group are determined according to the quantitative data of geometric features; total area of experimental group = ∑ (single organoid area); total area of negative control group = average organoid area of control group. It should be noted that the reduction in area directly reflects the cell killing effect caused by the drug, which is the main data source for dose-response curve fitting.

[0081] A calculation method is proposed to use the total area of organoids as a proxy indicator for survival rate, and a mathematical model of relative survival rate = total area of experimental group / total area of control group x 100% is established. This model avoids the destructive detection defect of the traditional enzyme label method, and realizes continuous observation and time series analysis of the same batch of samples.

[0082] According to the change trend of circularity, the change trend of concave-convex degree ratio and the change trend of the aspect ratio of the mask bounding box, the first drug effectiveness auxiliary data is obtained as the evaluation index data, which is used as an indirect correlation and verification application. Morphological features are used as pharmacodynamic auxiliary evaluation indicators. It should be noted that for the change trend of circularity, when the drug is effective, the organoids often lose their normal spherical structure and the circularity value decreases. For the concave-convex degree ratio trend, the drug may cause the surface of the organoids to roughen, and the concave-convex degree ratio value abnormally fluctuates. For the aspect ratio change trend, a specific drug may cause the organoids to change in shape orientation, and the aspect ratio changes regularly.

[0083] According to whether the cell viability and the change trend of circularity, the change trend of concave-convex degree ratio and the change trend of the aspect ratio of the mask bounding box present a synergistic change, the second drug effectiveness auxiliary data is obtained as the evaluation index data. Specifically, a feature consistency checking mechanism is established. When the area index corresponding to the cell viability shows drug efficacy, it is simultaneously checked whether the circularity change trend, the concave-convex degree ratio change trend and the aspect ratio change trend of the mask bounding box present a synergistic change. By establishing a multi-feature change pattern library, false positives or abnormal reactions can be identified. For example, if the area decreases but the circularity increases, it may indicate physical contraction rather than cell death, which needs to be further verified.

[0084] Different geometric and texture feature quantitative data are assigned weights to construct a multi-dimensional pharmacodynamic score model as evaluation index data. Specifically, the pharmacodynamic score model is represented as: S = w1·ΔA + w2·ΔC + w3·ΔR_conv+ w4·Δμ_I, where Δ represents the change rate relative to the control group, and w1, w2, w3 and w4 are the weight coefficients of each feature. This scoring model can be used for more detailed drug potency ranking and mechanism analysis.

[0085] These morphological features together constitute a multi-dimensional evidence system for organoid pharmacodynamic evaluation, not only providing quantitative pharmacodynamic data, but also providing important clues for understanding drug action mechanisms, realizing the leap from simple survival rate calculation to comprehensive pharmacodynamic analysis.

[0086] The evaluation indicators include the prediction curve of cell viability and drug concentration response.

[0087] The prediction curve of cell viability and drug concentration response is obtained by curve fitting of three-parameter Logistic model and four-parameter Logistic model.

[0088] Among them, the three-parameter Logistic model is expressed as:

[0089] y = Bottom + (Top - Bottom) / (1 + 10^((LogIC50 - x) ×(-1)));

[0090] The four-parameter Logistic model is expressed as:

[0091] y = Bottom + (Top - Bottom) / (1 + 10^((LogIC50 - x) × HillSlope));

[0092] In the formula, y is the relative survival rate, x is the logarithm of drug concentration (log10(concentration)), Top is the asymptote on the curve (maximum response), Bottom is the asymptote below the curve (minimum response), LogIC50 is the logarithmic value of half-inhibitory concentration, HillSlope is the Hill slope, reflecting the steepness of the curve. It should be noted that in the three-parameter Logistic model, HillSlope is fixed at -1, reducing one degree of freedom and improving the stability of fitting.

[0093] Nonlinear least squares method is used for fitting (Levenberg-Marquardt algorithm). Nonlinear least squares method (NLS) is a standard method for fitting nonlinear models. When fitting 3PL and 4PL models by NLS, the least squares method aims to minimize the difference between predicted values and true data, usually represented by the sum of squares of errors: . Where is the actual observation, is the predicted value based on the model and parameters , are the fitted parameters (such as the bottom, top, IC50 of 3PL, etc.). By minimizing the error, the parameters ​The curve fitting algorithm is used to achieve this goal, which combines the Gauss Newton method and the gradient descent method to avoid the slow convergence problem of gradient descent, while speeding up the fitting process. The algorithm adjusts the parameters by iteration until the error is minimized, thereby obtaining the best parameters of the fitted curve.

[0094] It should be noted that in other embodiments, other curve fitting algorithms can also be used, such as polynomial fitting or spline interpolation models. Although these methods can achieve curve fitting, they lack pharmacological significance and cannot directly derive key pharmacodynamic parameters such as IC50, and have poor biological interpretation.

[0095] The quality of curve fitting is evaluated by the coefficient of determination, which is represented as: R²= 1-SS_res / SS_tot; where R² is the coefficient of determination, SS_res is the sum of squares of residuals, and SS_tot is the total sum of squares. The closer the R² value is to 1, the better the fitting degree.

[0096] At the same time, three-parameter Logistic model and four-parameter Logistic model are used for parallel fitting, where the three-parameter model fixes the HillSlope slope as -1 to improve the stability of fitting, and the four-parameter model allows the HillSlope slope to change freely to adapt to complex response patterns.

[0097] For step S500, a pharmacodynamic evaluation report is generated based on the instance graph corresponding to the surviving organoid cells, the live organoid mask selected by the frame, the quantification data of geometric features and texture features, and the predicted curve of cell viability and drug concentration response. An automatic end-to-end automatic processing flow is achieved.

[0098] In terms of system integration, a complete automatic process from raw image input to final pharmacodynamic report generation is constructed, integrating instance segmentation, feature extraction, curve fitting, and parameter calculation modules. The system supports Web online service mode, realizes multi-drug parallel processing and comparative analysis, and significantly improves the analysis efficiency and standardization degree.

[0099] Overall, first, the user uploads organoid brightfield images under different drugs and different concentrations. Then, the images are input into the model based on the Mask R-CNN base, which will perform an organoid instance segmentation task on each image. Next, the algorithm extracts and quantifies the morphological features of the identified organoid cells. After obtaining the quantification data, the cell viability-drug concentration response prediction curve is drawn. Finally, the instance segmented brightfield images, the live organoid mask selected by the frame, the morphological feature quantification data of the organoids in each image, and the predicted IC50 curve response graph are merged as a pharmacodynamic evaluation reference report output.

[0100] It should be noted that in other embodiments, integration can also be performed using separate software tool chains, such as ImageJ segmentation + Excel statistical analysis + GraphPad curve fitting. Although this scheme can complete the analysis process using existing tools, it requires a large amount of manual operation and data processing, cannot achieve automated analysis, and has the risk of introducing human error.

[0101] The following describes specific application embodiments.

[0102] Embodiment one: single drug analysis process.

[0103] Taking "Drug A" as an example, the specific implementation steps include:

[0104] Input: bright field images of Drug A at 6 concentration gradients (0.1, 1, 10, 100, 1000, 10000 μM);

[0105] Segmentation: use a deep learning model based on Mask R-CNN base to identify organoid instances in each image;

[0106] Feature extraction: calculate the area, perimeter, circularity, etc. of each organoid;

[0107] Survival rate calculation: calculate the relative survival rate at each concentration based on the uploaded negative control group;

[0108] Curve fitting: fit with 3PL and 4PL models respectively, and output the dose response curve graph for reference by experimenters;

[0109] Output: bright field images of organoids after instance segmentation, IC50 = 58.3 μM, AUC = 0.723, and morphological feature indicators (rendered and displayed in the form of a table), and a dose response curve graph.

[0110] Embodiment two: multi-drug comparison analysis.

[0111] For multiple drugs, first, each drug follows the analysis process of Embodiment one to output the corresponding analysis results. Then, the algorithm will integrate and output a total dose response curve graph for multiple drugs, facilitating reference by experimenters.

[0112] Simultaneous analysis of drugs A, B, and C, and sorting by Area Under the Curve (AUC) value: Drug A: AUC = 0.723; Drug B: AUC = 0.815; Drug C: AUC = 0.692.

[0113] Reference conclusion: drug C is the best, drug B is the worst. AUC is the area under the fitting curve, the algorithm is normalized, the maximum value is 1, the smaller the value, the smaller the area under the drug curve, which reflects that the organ is more sensitive to the drug and the killing is more effective.

[0114] The above is a specific description of the preferred embodiments of the present application, but the present application is not limited to the embodiments, and those skilled in the art can make various equivalent modifications or replacements without departing from the spirit of the present application. These equivalent modifications or replacements are all included in the scope defined by the claims of the present application.

Claims

1. An instance segmentation based organoid brightfield image analysis method, characterized in that, The method comprises the following steps: Obtaining bright-field images of the organoids under different concentrations of multiple drugs; Inputting the bright-field images of the organoids into a trained organoid vertical model for instance segmentation to identify instance maps corresponding to the surviving organoid cells; Extracting morphological features according to the instance maps to obtain geometric features and texture features of the surviving organoid cells; Determining evaluation index data according to the quantization data of the geometric features and the texture features, wherein the evaluation index data comprises a predicted curve of the cell viability and the drug concentration response; Generating a pharmacodynamic evaluation report according to the instance maps corresponding to the surviving organoid cells, the quantization data of the geometric features and the texture features, and the predicted curve of the cell viability and the drug concentration response. The organoid vertical model comprises a feature extraction network, a region proposal network, a region of interest alignment network, and a mask prediction head; the step of inputting the bright-field images of the organoids into the trained organoid vertical model for instance segmentation to identify instance maps corresponding to the surviving organoid cells comprises the following steps: Convolving the bright-field images of the organoids through the feature extraction network to obtain feature maps; Generating regions of interest according to the feature maps through the region proposal network; Aligning the regions of interest through the region of interest alignment network to obtain aligned features; Predicting the segmentation mask of the regions of interest with the aligned features through the mask prediction head to obtain a binary mask, and identifying the instance maps corresponding to the surviving organoid cells according to the binary mask; The step of determining evaluation index data according to the quantization data of the geometric features and the texture features comprises the following steps: Obtaining first drug effectiveness auxiliary data as the evaluation index data according to the variation trend of the circularity, the variation trend of the concave-convex degree ratio, and the variation trend of the width-height ratio of the mask bounding box; The step of determining evaluation index data according to the quantization data of the geometric features and the texture features comprises the following steps: Obtaining second drug effectiveness auxiliary data as the evaluation index data according to whether the cell viability and the variation trend of the circularity, the variation trend of the concave-convex degree ratio, and the variation trend of the width-height ratio of the mask bounding box present a synergistic change.

2. The method of claim 1, wherein the method further comprises: The geometric features comprise basic geometric features and complex geometric features; the basic geometric features comprise the area, the perimeter, and the circularity of the surviving organoid cells; the complex geometric features comprise the concave-convex degree ratio of the surviving organoid cells and the width-height ratio of the mask bounding box.

3. The method of claim 1, wherein the method further comprises: The texture features comprise an average black solidity, The average black solidity is represented as: μ_I = (1 / A)∑∑I(i,j)·M(i,j); wherein μ_I is the average brightness, A is the area of the region of interest, I(i,j) is the pixel intensity value at position (i,j) in the image, and M(i,j) is the mask value at position (i,j) in the image.

4. The method of claim 2, wherein the method further comprises: The cell viability is calculated according to the following formula: relative survival rate = (total area of the experimental group / total area of the negative control group) × 100%; wherein the total area of the experimental group and the total area of the negative control group are determined according to the quantization data of the geometric features.

5. The method of claim 1, wherein the method further comprises: The step of determining evaluation index data according to the quantization data of the geometric features and the texture features comprises the following steps: The multi-dimensional pharmacodynamic score model is constructed by assigning weights to the quantified data of different geometric features and texture features as the evaluation index data.

6. The method of claim 4, wherein the method further comprises: The prediction curve of the cell viability and the drug concentration response is obtained by curve fitting of a three-parameter Logistic model and a four-parameter Logistic model. The three-parameter Logistic model is expressed as: y = Bottom + (Top - Bottom) / (1 + 10^((LogIC50 - x) ×(-1))), The four-parameter Logistic model is expressed as: y = Bottom + (Top - Bottom) / (1 + 10^((LogIC50 - x) × HillSlope)); In the formula, y is the relative survival rate, x is the logarithm of the drug concentration, Top is the asymptote on the curve, Bottom is the asymptote under the curve, LogIC50 is the logarithm value of the half-inhibitory concentration, and HillSlope is the Hill slope.

7. The instance segmentation-based organoid brightfield image analysis method according to claim 6, characterized in that, The curve fitting quality is evaluated by a coefficient of determination, The coefficient of determination is expressed as: R²= 1-SS_res / SS_tot; In the formula, R² is the coefficient of determination, SS_res is the residual sum of squares, and SS_tot is the total sum of squares.

Images