Imaging analysis method based on tumor organ chip

Abstract

The invention provides an imaging analysis method based on a tumor organoid chip, which comprises the following steps: constructing a chip containing an organoid culture unit, a vascularization co-culture unit and a drug gradient generation unit for culturing tumor organoids, constructing a blood vessel-organoid co-culture system and generating a continuous drug concentration gradient respectively; an optical, magnetic resonance and intravascular ultrasound multi-mode imaging system is adopted to collect organoid form, deep structure and vascular function images; imaging data is preprocessed, defocus blurring is solved through multi-focus fusion, and spatial correlation is achieved through noise suppression and multi-modal registration; extracting features such as organ-like area change rate and blood vessel proportion based on deep learning; and integrating imaging characteristics, tumor gene mutation and clinical data to construct a multi-dimensional matrix, and training a drug sensitivity prediction model, so that the problem that the clinical transformation value is seriously restricted due to the fact that a plurality of technical bottlenecks still exist in the current imaging analysis method for the tumor organ-like chip is solved.

Description

Technical Field

[0001] This invention relates to the field of imaging, and more particularly to an imaging analysis method based on tumor organoid microarrays. Background Technology

[0002] Tumor organ-on-a-chip technology, a revolutionary breakthrough in precision medicine in recent years, can realistically simulate the structural heterogeneity, vascularization characteristics, and microenvironment signaling networks of in vivo tumors by constructing a three-dimensional co-culture system containing tumor cells, vascular endothelial cells, stromal cells, and extracellular matrix. It has become an ideal in vitro research platform to replace traditional two-dimensional cell models and animal models. However, current imaging and analysis methods for tumor organ-on-a-chip still face many technical bottlenecks, severely limiting its clinical translational value. In terms of imaging modalities, existing technologies mostly rely on optical microscopy or fluorescence imaging, which can only acquire surface morphological information of organoids and cannot penetrate three-dimensional structures to capture deep blood vessel distribution, cell metabolic state, and drug penetration patterns. For example, traditional optical imaging is limited by light scattering effects, resulting in blurred images of the internal structures of organoids with diameters exceeding 200 μm, making it difficult to quantify key parameters such as blood vessel branch density and necrotic nucleus volume.

[0003] In the field of quantitative analysis algorithms, the blurred edges of organoids, the intricate vascular networks, and the dynamic morphological changes after drug intervention lead to insufficient accuracy in existing segmentation algorithms (such as traditional thresholding methods and basic U-Net models). Studies have shown that the error in calculating the organoid area change rate (OAC) using conventional algorithms can reach 15%-20%, severely affecting the accuracy of drug sensitivity assessment. Furthermore, there is a lack of specific analysis strategies for the differentiated morphological characteristics of microvessels and macrovessels in vascularized chips, making it difficult to accurately quantify vascular function parameters.

[0004] Regarding dynamic monitoring capabilities, existing methods mostly employ endpoint detection models, which cannot continuously track the real-time response of organoids during drug intervention. The lack of dynamic characteristics such as drug onset time and sensitivity decay patterns makes it impossible to establish a kinetic model of drug action, thus hindering the optimization of clinical dosing timing.

[0005] In terms of multi-dimensional data integration, current analytical methods are mostly limited to morphological features and have not achieved deep correlation between imaging data and omics data such as gene expression and protein levels. Due to their single feature dimension, drug sensitivity prediction models typically have a positive predictive value (PPV) of less than 70%, which is insufficient to meet the needs of precision medicine in clinical practice.

[0006] Therefore, it is necessary to provide a new imaging analysis method based on tumor organoid microarrays to solve the above-mentioned technical problems. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides an imaging analysis method based on tumor organoid microarrays, which solves the problem that current imaging analysis methods for tumor organoid microarrays still have many technical bottlenecks, severely restricting their clinical translational value.

[0008] The imaging analysis method based on tumor organoid microarrays provided by this invention includes the following steps: S1. Construct a tumor organoid chip, the chip including an organoid culture unit, a vascularization co-culture unit and a drug gradient generation unit, the organoid culture unit is used to culture tumor organoids, the vascularization co-culture unit is used to construct a co-culture system of vascular network and organoids, and the drug gradient generation unit is used to generate a continuous drug concentration gradient. S2. The imaging data of the tumor organoid chip is acquired using a multimodal imaging system, which includes an optical imaging module, a magnetic resonance imaging module, and an intravascular ultrasound module. The imaging data includes organoid morphology images, deep structure images, and vascular function images. S3. Preprocess the imaging data, including multi-focus fusion, noise suppression and multimodal registration. The multi-focus fusion is used to solve the defocus blur problem in three-dimensional culture. The noise suppression is used to reduce image noise interference. The multimodal registration is used to realize the spatial correlation of images of different modalities. S4. Based on deep learning algorithms, perform quantitative analysis on the preprocessed imaging data to extract organoid morphological features, vascular network features and multimodal functional features. The organoid morphological features include area change rate (OAC), and the vascular network features include vascular proportion, number of branch nodes and average vascular length. S5. Construct a drug sensitivity prediction model. The model is generated based on a multi-dimensional feature matrix. The multi-dimensional feature matrix includes the imaging features extracted in step S4, tumor driver gene mutation data, and clinical features. The drug sensitivity prediction model is used to assess the sensitivity of tumor organoids to drugs.

[0009] Furthermore, the construction of the tumor organoid chip in step S1 includes: S11. Organoid preparation: Collect human tumor tissue samples, digest them, embed them in matrix gel, and culture them for 7-14 days to form tumor organoids. The diameter of the tumor organoids is 200-500 μm. S12. Construction of vascular co-culture system: Human umbilical vein endothelial cells (HUVEC) are seeded into the microarray vascular channel and dynamically cultured for 7-9 days using a swing perfusion instrument to form a dense vascular network. Vascular growth factor is added to induce vascular endothelial cells to permeate into organoid matrix gel. S13. Drug gradient generation unit integration: A tree-shaped microfluidic gradient generator is adopted, and the channel structure is optimized through COMSOL simulation to achieve a continuous concentration gradient of 0.1-10 times the maximum blood drug concentration. The gradient generator is connected to the upper channel of the chip and the fluid flow rate is controlled by an injection pump.

[0010] Furthermore, the acquisition process of the multimodal imaging system in step S2 includes: S21. Optical Imaging Module Acquisition: An inverted fluorescence microscope combined with a high-sensitivity CMOS camera is used to acquire bright-field images and fluorescence images. The fluorescence images are used to distinguish between live and dead cells by live cell labeling and dead cell labeling. S22. Magnetic Resonance Imaging Module Acquisition: A 9.4T small animal MRI system was used, employing a multi-contrast quantitative imaging sequence with parameters set to TR=500ms, TE=20ms, FOV=15×15mm, and slice thickness of 0.1mm to acquire T1-weighted, T2-weighted, and diffusion-weighted imaging data. S23. Intravascular ultrasound module acquisition: A 40MHz high-frequency ultrasound catheter is used for the vascularization chip. The catheter retraction is controlled by a precision displacement platform with a retraction speed of 0.5mm / s and a frame rate of 30 frames / second to acquire cross-sectional images of the blood vessels. S24. Dynamic time-series acquisition: Multimodal imaging was performed at 0h, 6h, 12h, 24h, 48h, and 72h after drug intervention to record the dynamic changes of organoids and blood vessels.

[0011] Furthermore, the preprocessing described in step S3 includes: S31. Multi-focus fusion: The Laplacian pyramid algorithm is used to decompose and weightedly fuse optical images of five different focal planes in the same field of view. The weighted fusion is based on the local variance criterion. S32. Noise suppression: Non-local mean filtering is applied to the magnetic resonance imaging images; S33. Multimodal registration: Based on the maximum mutual information algorithm, rigid registration of magnetic resonance imaging and intravascular ultrasound images is performed with optical images as reference, and the registration error is controlled within 10μm.

[0012] Furthermore, the organoid morphological feature extraction in step S4 includes: S411, Organoid Segmentation: An improved ResUNet model is adopted. The model introduces a dilated residual convolution module in the encoder part with an inflation rate of 2 and 4. A non-linear attention mechanism (ULA-Block) is added in the decoder part. The training parameters are batch size=8, learning rate=0.001, 200 iterations, and the Dice coefficient of the segmentation result is ≥0.94. S412, OAC value calculation: Calculate the endpoint OAC (OAC1) and the individual OAC (OAC2), where:

[0013]

[0014]

[0015] (T0, C0) represent the organoid area at the time of drug administration in the experimental group and the blank group, respectively. t C t These are the organoid areas of the experimental group and the control group at time t, respectively, where A0 is the organoid area at the time of drug administration, and A... t Let be the area of ​​the organoid at time t; S413. Comprehensive assessment: When OAC1 ≥ 60% and OAC2 < 20%, the drug is considered effective; when OAC1 < 60% and OAC2 ≥ 20%, the drug is considered ineffective; all other cases are considered potential susceptibility.

[0016] Furthermore, the vascular network feature extraction in step S4 includes: S421, Vessel segmentation: The U-Net model improved with multi-scale transition module (MSA-Block) is adopted, which integrates features of different scales to capture microvascular branch details, with a segmentation accuracy of ≥0.92. S422. Topology Analysis: The vascular skeleton image is analyzed using a breadth-first search algorithm to calculate the vascular proportion, number of connected components, number of branch nodes, average vascular length, and average vascular width. The vascular proportion is the ratio of the number of vascular pixels to the total number of pixels in the field of view.

[0017] Furthermore, the multimodal functional feature extraction in step S4 includes: S431. Magnetic resonance imaging features: Extract the necrotic nucleus volume, apparent diffusion coefficient (ADC value), and mean T1-weighted signal intensity. The necrotic nucleus volume is obtained based on high-signal region segmentation of the T2-weighted image. S432. Intravascular ultrasound features: Extracting the diameter of the blood vessel lumen, the thickness of the vessel wall, and the area of ​​calcified plaques, wherein the calcified plaques are marked from the intravascular ultrasound image using an automatic recognition algorithm; S433. Feature selection: LASSO regression is used to remove redundant features, and 15 key features are retained to construct a multi-dimensional feature matrix.

[0018] Furthermore, the construction of the drug sensitivity prediction model in step S5 includes: S51. Data integration: Collect multi-center sample data, including imaging features extracted in step S4, tumor driver gene mutation and expression data detected by PCR, and clinical characteristics such as patient age and tumor stage. S52. Model Training: The Stacking ensemble learning framework is adopted. The basic models include Random Forest (n_estimators=200), Gradient Boosting Tree (GBDT, learning_rate=0.1), and Support Vector Machine (SVM, kernel function is RBF). Logistic Regression is used as a meta-classifier to integrate the output of the basic models. S53. Model optimization: Search for optimal parameters using the Bayesian optimization algorithm. The objective function is to maximize the cross-validation AUC. The optimized model has an AUC ≥ 0.9. S54. Dynamic Validation: Fit the drug response curve using time-series data and calculate the half-maximal inhibition time (T0). 50 ), to assess the speed at which the drug takes effect.

[0019] Furthermore, the system integration includes: Hardware integration: The optical imaging, magnetic resonance imaging, and intravascular ultrasound modules are mechanically linked through a motion control card, with a positioning accuracy of ±1μm, enabling automatic chip alignment; Software platform: The host computer software is developed based on C#, including modules for image acquisition, preprocessing, algorithm analysis, and result visualization; Data Management: The system uses a MySQL database to store imaging data, feature parameters, and prediction results, and supports multi-terminal access and data traceability.

[0020] The beneficial effects of this invention are: 1. This invention is the first to integrate optical imaging, magnetic resonance imaging (MRI), and intravascular ultrasound (IU) technologies for tumor organoid microarray analysis, enabling comprehensive monitoring from micrometer-level surface morphology to millimeter-level deep structures. Optical imaging captures dynamic changes in cell activity, the MRI module quantifies the volume of necrotic nuclei and ADC values ​​(reflecting apoptosis), and the IU assesses vascular functional integrity. The complementary data from these three modalities solve the problem that traditional single imaging techniques cannot penetrate three-dimensional structures.

[0021] 2. To address the complex morphologies of organoids and blood vessels, innovative and improved ResUNet and Attention-Res-U-Net models were developed. These models incorporate dilated convolution, attention mechanisms, and multi-scale fusion strategies, demonstrating significant advancements over the traditional U-Net model. The proposed multi-focus fusion algorithm solves the defocusing problem in 3D culture, keeping the OAC calculation error within 5%. Furthermore, through LASSO feature selection and ensemble learning models, efficient integration of multi-dimensional features is achieved, significantly outperforming existing single-feature models.

[0022] 3. A 72-hour dynamic temporal imaging analysis strategy was established, introducing the concept of half-inhibition time for the first time. This parameterized pharmacokinetic characteristics distinguished between rapidly acting and slowly acting drugs, providing experimental evidence for optimizing clinical dosing regimens. Multimodal registration technology enabled spatial correlation between optical-MRI-IVUS data, allowing for precise matching of morphological features with functional parameters (such as vascular permeability and cell metabolism), thus expanding the depth and breadth of tumor microenvironment research.

[0023] 4. The developed integrated imaging analysis system automates the entire process, from automatic chip alignment and multimodal data acquisition to intelligent analysis and result visualization. Operation time is reduced from 48 hours using traditional methods to within 72 hours, with a 5-fold increase in throughput. Simultaneously, the system supports multi-center data sharing and model iteration, providing drug development companies with an efficient screening platform and accelerating the development of new anti-tumor drugs. Attached Figure Description

[0024] Figure 1 The overall flowchart of the imaging analysis method based on tumor organoid chips provided by the present invention is shown below. Figure 2 A flowchart for constructing tumor organoid chips is provided for this invention; Figure 3 This invention provides a flowchart of the acquisition process for a multimodal imaging system. Figure 4 This is a flowchart of the preprocessing process described in step S3 of the present invention; Figure 5 This is a flowchart of the vascular network feature extraction process of the present invention; Figure 6 This is a flowchart of the multimodal functional feature extraction process of the present invention; Figure 7 The system integration flowchart provided for this invention. Detailed Implementation

[0025] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0026] Please refer to the following: Figure 1 , Figure 2 , Figure 3 , Figure 4 , Figure 5 , Figure 6 as well as Figure 7 ,in Figure 1 The overall flowchart of the imaging analysis method based on tumor organoid chips provided by the present invention is shown below. Figure 2 A flowchart for constructing tumor organoid chips is provided for this invention; Figure 3 This invention provides a flowchart of the acquisition process for a multimodal imaging system. Figure 4This is a flowchart of the preprocessing process described in step S3 of the present invention; Figure 5 This is a flowchart of the vascular network feature extraction process of the present invention; Figure 6 This is a flowchart of the multimodal functional feature extraction process of the present invention; Figure 7 The system integration flowchart provided for this invention.

[0027] In the specific implementation process, such as Figures 1-7 As shown, it includes the following steps: S1. Construct a tumor organoid microarray. The microarray includes an organoid culture unit, a vascularization co-culture unit, and a drug gradient generation unit. The organoid culture unit is used to culture tumor organoids, the vascularization co-culture unit is used to construct a co-culture system of vascular networks and organoids, and the drug gradient generation unit is used to generate a continuous drug concentration gradient. S2. Imaging data of tumor organoid chips are acquired using a multimodal imaging system. The multimodal imaging system includes an optical imaging module, a magnetic resonance imaging module, and an intravascular ultrasound module. The imaging data includes organoid morphology images, deep structure images, and vascular function images. S3. Preprocess the imaging data, including multi-focus fusion, noise suppression and multimodal registration. Multi-focus fusion is used to solve the defocus blur problem in 3D culture, noise suppression is used to reduce image noise interference, and multimodal registration is used to realize the spatial correlation of images of different modalities. S4. Based on deep learning algorithms, quantitative analysis is performed on the preprocessed imaging data to extract organoid morphological features, vascular network features and multimodal functional features. Organoid morphological features include area change rate (OAC), and vascular network features include vascular proportion, number of branch nodes and average vascular length. S5. Construct a drug sensitivity prediction model. The model is generated based on a multi-dimensional feature matrix, which includes the imaging features extracted in step S4, tumor driver gene mutation data, and clinical features. The drug sensitivity prediction model is used to assess the sensitivity of tumor organoids to drugs.

[0028] The construction of tumor organoid microarrays in step S1 includes: S11. Organoid preparation: Collect human tumor tissue samples, digest them, embed them in matrix gel, and culture them for 7-14 days to form tumor organoids with a diameter of 200-500 μm. S12. Construction of vascular co-culture system: Human umbilical vein endothelial cells (HUVEC) are seeded into the microarray vascular channel and dynamically cultured for 7-9 days using a swing perfusion instrument to form a dense vascular network. Vascular growth factor is added to induce vascular endothelial cells to permeate into organoid matrix gel. S13. Drug gradient generation unit integration: A tree-shaped microfluidic gradient generator is adopted. The channel structure is optimized through COMSOL simulation to achieve a continuous concentration gradient of 0.1-10 times the maximum blood drug concentration. The gradient generator is connected to the upper channel of the chip and the fluid flow rate is controlled by an injection pump.

[0029] Example 1: Construction and Validation of Tumor Organoid Microarrays 1.1 Sample Source Five surgical specimens of head and neck squamous cell carcinoma were collected (with informed consent from the patients). The samples were processed within 4 hours after collection to maintain tissue viability.

[0030] 1.2 Organoid Culture Cut the tumor tissue into 1mm pieces 3 Small pieces were added to a digestion solution containing collagenase IV (2 mg / mL) and hyaluronidase (1 mg / mL), and digested at 37°C with shaking for 40 min. The mixture was then filtered through a 70 μm filter to obtain a single-cell suspension. After cell counting, the suspension was mixed with Matrigel at a 1:3 ratio, and 50 μL was seeded into each well of a 48-well plate. After solidification at 37°C for 30 min, organoid culture medium was added, and the plate was cultured for 10 days, changing the medium every 2 days, to form organoids with a diameter of 300-400 μm.

[0031] 1.3 Fabrication of Vascularized Chips The chip is made using a PDMS three-layer process and fabricated using standard photolithography. Upper channel: Tree gradient generator with 6 branches, forming a concentration gradient of 0.1, 0.3, 1, 3, and 10 Cmax at an inlet flow rate of 0.5 μL / min; The middle culture chamber has a diameter of 6 mm and a depth of 7 mm, with a microcolumn array (50 μm in diameter and 100 μm in spacing) at the bottom to enhance organoid attachment; Lower vascular channel: 300 μm wide and 150 μm deep, coated with fibronectin (10 μg / mL) to promote endothelial cell adhesion.

[0032] 1.4 Cultivation and Validation Organoids were seeded into the middle culture chamber, then into the lower channel, and cultured in a swing perfusion apparatus for 6 days. Validation was performed on day 6. CD31 immunofluorescence staining: The vascular network is continuous, with intact branches and a coverage of >70%; Barrier function test: fluorescein sodium (100 μM) was added to the vascular channel, and the fluorescence intensity of the organoid culture chamber was detected after 30 min. The permeability was <3% / h. Pressure test: No leakage was observed after 24 hours of continuous operation at a fluid pressure of 0.3 kPa.

[0033] The acquisition process of the multimodal imaging system in step S2 includes: S21. Optical Imaging Module Acquisition: An inverted fluorescence microscope combined with a high-sensitivity CMOS camera is used to acquire bright-field images and fluorescence images. The fluorescence images are used to distinguish between live and dead cells by live cell labeling and dead cell labeling. S22. Magnetic Resonance Imaging Module Acquisition: A 9.4T small animal MRI system was used, employing a multi-contrast quantitative imaging sequence with parameters set to TR=500ms, TE=20ms, FOV=15×15mm, and slice thickness of 0.1mm to acquire T1-weighted, T2-weighted, and diffusion-weighted imaging data. S23. Intravascular ultrasound module acquisition: A 40MHz high-frequency ultrasound catheter is used for the vascularization chip. The catheter retraction is controlled by a precision displacement platform with a retraction speed of 0.5mm / s and a frame rate of 30 frames / second to acquire cross-sectional images of the blood vessels. S24. Dynamic time-series acquisition: Multimodal imaging was performed at 0h, 6h, 12h, 24h, 48h, and 72h after drug intervention to record the dynamic changes of organoids and blood vessels.

[0034] The preprocessing in step S3 includes: S31. Multi-focus fusion: The Laplacian pyramid algorithm is used to decompose and weightedly fuse optical images of five different focal planes in the same field of view. The weighted fusion is based on the local variance criterion. S32. Noise suppression: Non-local mean filtering is applied to the magnetic resonance imaging images; S33. Multimodal registration: Based on the maximum mutual information algorithm, rigid registration of magnetic resonance imaging and intravascular ultrasound images is performed with optical images as reference, and the registration error is controlled within 10μm.

[0035] Step S4, organoid morphology feature extraction, includes: S411, Organoid Segmentation: An improved ResUNet model is adopted. The model introduces a dilated residual convolution module in the encoder part with an inflation rate of 2 and 4. A non-linear attention mechanism (ULA-Block) is added in the decoder part. The training parameters are batch size=8, learning rate=0.001, 200 iterations, and the Dice coefficient of the segmentation result is ≥0.94. S412, OAC value calculation: Calculate the endpoint OAC (OAC1) and the individual OAC (OAC2), where:

[0036]

[0037]

[0038] (T0, C0) represent the organoid area at the time of drug administration in the experimental group and the blank group, respectively. t C t These are the organoid areas of the experimental group and the control group at time t, respectively, where A0 is the organoid area at the time of drug administration, and A... t Let be the area of ​​the organoid at time t; S413. Comprehensive assessment: When OAC1 ≥ 60% and OAC2 < 20%, the drug is considered effective; when OAC1 < 60% and OAC2 ≥ 20%, the drug is considered ineffective; all other cases are considered potential susceptibility.

[0039] Step S4, vascular network feature extraction, includes: S421, Vessel segmentation: The U-Net model improved with multi-scale transition module (MSA-Block) is adopted, which integrates features of different scales to capture microvascular branch details, with a segmentation accuracy of ≥0.92. S422. Topology Analysis: The vascular skeleton image is analyzed using a breadth-first search algorithm to calculate the vascular proportion, number of connected components, number of branch nodes, average vascular length, and average vascular width. The vascular proportion is the ratio of the number of vascular pixels to the total number of pixels in the field of view.

[0040] Example 2: Multimodal Imaging Data Acquisition 2.1 Drug intervention Two chemotherapy drugs, paclitaxel (Taxol) and cisplatin (Cisplatin), were selected. A concentration gradient (0.1-10Cmax) was applied through an upper gradient generator. Three replicate chips were set up for each drug. An equal amount of culture medium was added to the blank control group.

[0041] 2.2 Imaging Process 2.2.1 Optical Imaging: The Operetta CLS high-content imaging system was used with a 20× objective lens for bright-field imaging of organoid morphology. Fluorescence channels (Ex494 / Em517 for Calcein-AM, Ex535 / Em617 for PI) were used to capture the distribution of live and dead cells. The exposure time per field of view was 200ms, and three non-overlapping fields of view were acquired at each time point.

[0042] 2.2.2 Magnetic Resonance Imaging Module Imaging: Using a μMR9.4T system, T1, T2 weighted and DWI images were acquired using MTP sequences at a scan time of 21 min / scan. Three-dimensional images were reconstructed using Paravision software, and ROIs (organoid regions) were manually delineated.

[0043] 2.2.3 Intravascular ultrasound module imaging: The Boston Scientific Polaris system was used. The 40MHz catheter was retracted along the vascular channel (at a speed of 0.5mm / s) to acquire cross-sectional images. The images were taken every 24 hours, and the lumen diameter was measured using the accompanying software.

[0044] 2.3 Dynamic Data Recording Record data at 0h, 6h, 12h, 24h, 48h, and 72h: Optical images: organoid area, live / dead cell fluorescence intensity ratio; MRI data: Necrotic nucleus volume (high signal area in T2-weighted image), ADC value; IVUS data: vessel lumen diameter and vessel wall thickness.

[0045] Step S4, multimodal functional feature extraction, includes: S431. Magnetic resonance imaging features: Extract the necrotic nucleus volume, apparent diffusion coefficient (ADC value), and mean T1-weighted signal intensity. The necrotic nucleus volume is obtained based on the high-signal region segmentation of the T2-weighted image. S432. Intravascular ultrasound features: Extract the diameter of the blood vessel lumen, the thickness of the vessel wall, and the area of ​​calcified plaques. Calcified plaques are marked from the intravascular ultrasound images using an automatic recognition algorithm. S433. Feature selection: LASSO regression is used to remove redundant features, and 15 key features are retained to construct a multi-dimensional feature matrix.

[0046] Step S5 involves constructing the drug sensitivity prediction model, which includes: S51. Data integration: Collect multi-center sample data, including imaging features extracted in step S4, tumor driver gene mutation and expression data detected by PCR, and clinical characteristics such as patient age and tumor stage. S52. Model Training: The Stacking ensemble learning framework is adopted. The basic models include Random Forest (n_estimators=200), Gradient Boosting Tree (GBDT, learning_rate=0.1), and Support Vector Machine (SVM, kernel function is RBF). Logistic Regression is used as a meta-classifier to integrate the output of the basic models. S53. Model optimization: Search for optimal parameters using the Bayesian optimization algorithm. The objective function is to maximize the cross-validation AUC. The optimized model has an AUC ≥ 0.9. S54. Dynamic Validation: Fit the drug response curve using time-series data and calculate the half-maximal inhibition time (T0). 50 ), to assess the speed at which the drug takes effect.

[0047] Example 4: Construction and Validation of Drug Susceptibility Prediction Model 3.1 Dataset Construction We collected organoid microarray data from several patients, including datasets and validation sets, containing response results for various drugs such as paclitaxel and cisplatin.

[0048] 3.2 Model Training Using Stacking Ensemble Learning: Base models: Random Forest (n_estimators=200), GBDT (max_depth=8), SVM (C=10); Meta-classifier: Logistic Regression; The optimized model has an AUC of 0.92, accuracy of 88%, sensitivity of 85%, and specificity of 90%.

[0049] 3.3 Verification Results Internal validation: 10-fold cross-validation AUC = 0.91 ± 0.03; External validation: Independent test set AUC=0.89, PPV=87%; Dynamic prediction: Paclitaxel T 50 =22h, Cisplatin T 50 >72h, consistent with clinical efficacy observation.

[0050] System integration includes: Hardware integration: The optical imaging, magnetic resonance imaging, and intravascular ultrasound modules are mechanically linked through a motion control card, with a positioning accuracy of ±1μm, enabling automatic chip alignment; Software platform: The host computer software is developed based on C#, including modules for image acquisition, preprocessing, algorithm analysis, and result visualization; Data Management: The system uses a MySQL database to store imaging data, feature parameters, and prediction results, and supports multi-terminal access and data traceability.

[0051] Example 4: System Integration and Application The developed imaging analysis system is implemented as follows: Automated operation: Automatic chip loading, multimodal imaging linkage, and data analysis are completed with one click, with a total operation time of less than 4 hours; Results visualization: Generate organoid morphology change curves, vascular network topology maps, and drug sensitivity prediction heatmaps.

[0052] This invention constructs a complete tumor organoid microarray imaging analysis method through multimodal imaging fusion, intelligent algorithm innovation, and system integration optimization. It significantly improves the quantitative accuracy and clinical translation value of in vitro models, providing strong technical support for precision oncology and drug development.

[0053] The circuits and controls involved in this invention are all existing technologies and will not be described in detail here.

[0054] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. An imaging analysis method based on tumor organoid microarrays, characterized in that, Includes the following steps: S1. Construct a tumor organoid chip, the chip including an organoid culture unit, a vascularization co-culture unit and a drug gradient generation unit, the organoid culture unit is used to culture tumor organoids, the vascularization co-culture unit is used to construct a co-culture system of vascular network and organoids, and the drug gradient generation unit is used to generate a continuous drug concentration gradient. S2. The imaging data of the tumor organoid chip is acquired using a multimodal imaging system, which includes an optical imaging module, a magnetic resonance imaging module, and an intravascular ultrasound module. The imaging data includes organoid morphology images, deep structure images, and vascular function images. S3. Preprocess the imaging data, including multi-focus fusion, noise suppression and multimodal registration. The multi-focus fusion is used to solve the defocus blur problem in three-dimensional culture. The noise suppression is used to reduce image noise interference. The multimodal registration is used to realize the spatial correlation of images of different modalities. S4. Based on deep learning algorithms, perform quantitative analysis on the preprocessed imaging data to extract organoid morphological features, vascular network features and multimodal functional features. The organoid morphological features include area change rate (OAC), and the vascular network features include vascular proportion, number of branch nodes and average vascular length. S5. Construct a drug sensitivity prediction model. The model is generated based on a multi-dimensional feature matrix. The multi-dimensional feature matrix includes the imaging features extracted in step S4, tumor driver gene mutation data, and clinical features. The drug sensitivity prediction model is used to assess the sensitivity of tumor organoids to drugs.

2. The imaging analysis method based on tumor organoid microarrays according to claim 1, characterized in that, The construction of the tumor organoid microarray in step S1 includes: S11. Organoid preparation: Collect human tumor tissue samples, digest them, embed them in matrix gel, and culture them for 7-14 days to form tumor organoids. The diameter of the tumor organoids is 200-500 μm. S12. Construction of vascular co-culture system: Human umbilical vein endothelial cells (HUVEC) are seeded into the microarray vascular channel and dynamically cultured for 7-9 days using a swing perfusion instrument to form a dense vascular network. Vascular growth factor is added to induce vascular endothelial cells to permeate into organoid matrix gel. S13. Drug gradient generation unit integration: A tree-shaped microfluidic gradient generator is adopted, and the channel structure is optimized through COMSOL simulation to achieve a continuous concentration gradient of 0.1-10 times the maximum blood drug concentration. The gradient generator is connected to the upper channel of the chip and the fluid flow rate is controlled by an injection pump.

3. The imaging analysis method based on tumor organoid microarrays according to claim 2, characterized in that, The acquisition process of the multimodal imaging system in step S2 includes: S21. Optical Imaging Module Acquisition: An inverted fluorescence microscope combined with a high-sensitivity CMOS camera is used to acquire bright-field images and fluorescence images. The fluorescence images are used to distinguish between live and dead cells by live cell labeling and dead cell labeling. S22. Magnetic Resonance Imaging Module Acquisition: A 9.4T small animal MRI system was used, employing a multi-contrast quantitative imaging sequence with parameters set to TR=500ms, TE=20ms, FOV=15×15mm, and slice thickness of 0.1mm to acquire T1-weighted, T2-weighted, and diffusion-weighted imaging data. S23. Intravascular ultrasound module acquisition: A 40MHz high-frequency ultrasound catheter is used for the vascularization chip. The catheter retraction is controlled by a precision displacement platform with a retraction speed of 0.5mm / s and a frame rate of 30 frames / second to acquire cross-sectional images of the blood vessels. S24. Dynamic time-series acquisition: Multimodal imaging was performed at 0h, 6h, 12h, 24h, 48h, and 72h after drug intervention to record the dynamic changes of organoids and blood vessels.

4. The imaging analysis method based on tumor organoid microarrays according to claim 3, characterized in that, The preprocessing described in step S3 includes: S31. Multi-focus fusion: The Laplacian pyramid algorithm is used to decompose and weightedly fuse optical images of five different focal planes in the same field of view. The weighted fusion is based on the local variance criterion. S32. Noise suppression: Non-local mean filtering is applied to the magnetic resonance imaging images; S33. Multimodal registration: Based on the maximum mutual information algorithm, rigid registration of magnetic resonance imaging and intravascular ultrasound images is performed with optical images as reference, and the registration error is controlled within 10μm.

5. The imaging analysis method based on tumor organoid microarrays according to claim 4, characterized in that, The organoid morphological feature extraction in step S4 includes: S411, Organoid Segmentation: An improved ResUNet model is adopted. The model introduces a dilated residual convolution module in the encoder part with an inflation rate of 2 and 4. A non-linear attention mechanism (ULA-Block) is added in the decoder part. The training parameters are batchsize=8, learning rate=0.001, and 200 iterations. The Dice coefficient of the segmentation result is ≥0.

94. S412, OAC value calculation: calculate the endpoint OAC (OAC1) and autologous OAC (OAC2), wherein: OAC1 (%) = (1 - ROA T / ROA C ) x 100% ROA C =(C t -C0) / C0 OAC2(%)=(A t -A0) / A0×100%(T0, C0) represent the organoid area at the time of drug administration in the experimental group and the blank group, respectively, and T0 represents the organoid area at the time of drug administration. t C t These are the organoid areas of the experimental group and the control group at time t, respectively, where A0 is the organoid area at the time of drug administration, and A... t Let be the area of ​​the organoid at time t; S413. Comprehensive assessment: When OAC1 ≥ 60% and OAC2 < 20%, the drug is considered effective; when OAC1 < 60% and OAC2 ≥ 20%, the drug is considered ineffective; all other cases are considered potential susceptibility.

6. The imaging analysis method based on tumor organoid microarrays according to claim 5, characterized in that, The vascular network feature extraction in step S4 includes: S421. Vessel segmentation: The U-Net model improved by the Multi-Scale Transition Module (MSA-Block) is adopted, which integrates features of different scales to capture microvascular branch details, with a segmentation accuracy of ≥0.

92. S422. Topology Analysis: The vascular skeleton image is analyzed using a breadth-first search algorithm to calculate the vascular proportion, number of connected components, number of branch nodes, average vascular length, and average vascular width. The vascular proportion is the ratio of the number of vascular pixels to the total number of pixels in the field of view.

7. The imaging analysis method based on tumor organoid microarrays according to claim 6, characterized in that, The multimodal functional feature extraction in step S4 includes: S431. Magnetic resonance imaging features: Extract the necrotic nucleus volume, apparent diffusion coefficient (ADC value), and mean T1-weighted signal intensity. The necrotic nucleus volume is obtained based on high-signal region segmentation of the T2-weighted image. S432. Intravascular ultrasound features: Extracting the diameter of the blood vessel lumen, the thickness of the vessel wall, and the area of ​​calcified plaques, wherein the calcified plaques are marked from the intravascular ultrasound image using an automatic recognition algorithm; S433. Feature selection: LASSO regression is used to remove redundant features, and 15 key features are retained to construct a multi-dimensional feature matrix.

8. The imaging analysis method based on tumor organoid microarrays according to claim 7, characterized in that, The construction of the drug sensitivity prediction model in step S5 includes: S51. Data integration: Collect multi-center sample data, including imaging features extracted in step S4, tumor driver gene mutation and expression data detected by PCR, and clinical characteristics such as patient age and tumor stage. S52. Model Training: The Stacking ensemble learning framework is adopted. The basic models include Random Forest (n_estimators=200), Gradient Boosting Tree (GBDT, learning_rate=0.1), and Support Vector Machine (SVM, kernel function is RBF). Logistic Regression is used as a meta-classifier to integrate the output of the basic models. S53. Model optimization: Search for optimal parameters using the Bayesian optimization algorithm. The objective function is to maximize the cross-validation AUC. The optimized model has an AUC ≥ 0.

9. S54. Dynamic Validation: Fit the drug response curve using time-series data and calculate the half-maximal inhibition time (T). 50 To assess the speed at which the drug takes effect.

9. The imaging analysis method based on tumor organoid microarrays according to claim 8, characterized in that, The system integration includes: Hardware integration: The optical imaging, magnetic resonance imaging, and intravascular ultrasound modules are mechanically linked through a motion control card, with a positioning accuracy of ±1μm, enabling automatic chip alignment; Software platform: The host computer software is developed based on C#, including modules for image acquisition, preprocessing, algorithm analysis, and result visualization; Data Management: The system uses a MySQL database to store imaging data, feature parameters, and prediction results, and supports multi-terminal access and data traceability.

Images