Organoid viability evaluation system based on cbam-yolov3 integrated algorithm and application

Abstract

The application provides an organoid viability evaluation system based on a CBAM-YOLOv3 integrated algorithm and an application, the system comprises an input display module and a detection module, the CBAM-YOLOv3 integrated algorithm is used to obtain the average diameter, the number and the product of the number and the average diameter of the organoid, meanwhile, the number of the organoids with good viability and the number of the aging organoids are counted based on the machine learning result, and the respective proportions are calculated.The application also provides built-in electronic equipment in a microscope observation device in an organoid culture process and a non-transient computer readable storage medium.The results show that the No.xDia.parameter shows good evaluation accuracy, which is highly correlated with the SA-beta-Gal positive rate and the number of passages, can accurately reflect the cell viability of the organoid, meanwhile, based on the CBAM-YOLOv3 integrated algorithm in the detection of the viability of the gastric organoid, the proportion of the organoids with good activity and the aging organoids can be output, which provides valuable reference for downstream experiments.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of human tissue organoid images, and relates to application of a CBAM-YOLOv3 integrated algorithm in organoid activity evaluation, in particular to an organoid activity evaluation system and method based on the CBAM-YOLOv3 integrated algorithm and a corresponding derivative. BACKGROUND

[0002] Gastric cancer is a common malignant tumor worldwide, especially in East Asian countries including China. Because the stomach is a deep internal organ, it is difficult to detect at the early stage due to the lack of specific symptoms, so most gastric cancers have reached the middle and advanced stages when they are diagnosed, and even lose the opportunity for surgical treatment. In the treatment of gastric cancer, in addition to surgical methods, drug therapy, radiotherapy, and molecular targeted therapy or immunotherapy become important adjuvant treatment methods. Gastric cancer is a tumor with uneven growth activity, complex histological classification, and high heterogeneity. Therefore, the sensitivity to the above adjuvant therapies varies greatly. Whether it is chemotherapy, radiotherapy, molecular targeted therapy or immunotherapy, it is hoped to maximize the treatment benefit but minimize the treatment side effects. How to strike a balance between efficacy and toxicity is the key to the treatment of malignant tumors today. Therefore, before or during treatment, appropriate means are often used to evaluate the possible treatment response of patients.

[0003] Patient-derived organoids have been established by the US FDA as a three-dimensional cell culture model that can replace animal experiments for new drug screening or drug sensitivity experiments. Since the successful construction of mouse intestinal organoids in 2009 internationally, organoids have been widely used and developed as an experimental model. Because organoids can well reproduce the various characteristics of their source tissues in terms of tissue structure, cell function, and even gene profile, they have become an important means for studying disease pathogenesis, environmental harmful substance interaction, and drug sensitivity. However, there is no uniform method for evaluating the quality of organoids.

[0004] Existing organoid activity evaluation systems are mostly based on human observation or chemical reagent detection. Human observation consumes a lot of time and is highly dependent on the experience level of the experimenter, and is not strongly generalizable. Chemical reagent detection is not sensitive enough for small sample size organoid samples, and the experimental process can cause organoid lysis, and the organoids cannot be cultured after the detection is completed.

[0005] In order to overcome the above-mentioned defects, artificial intelligence is also applied to organoid vitality evaluation in combination with its excellent advantages in the field of image recognition. For example, Chinese invention patents with patent numbers CN113283352A and CN115641262A both construct evaluation functions or loss functions, and optimize the functions during the training process. Finally, the organoid vitality is evaluated through a neural network model. Chinese invention patent with patent number CN115358973A discloses a method for evaluating organoid vitality based on ATP analysis method. A series of organoid images are obtained by image acquisition of each organoid culture well plate. All organoid images in the same well plate are spliced into a panoramic image, and a series of organoid panoramic images are obtained. The CTG luminescence method is used to detect the organoid vitality value of the organoid culture well plate.

[0006] However, the current artificial intelligence method has the problem of inconsistent standards. The samples included are all mouse organoid samples, and there is a lack of human organ samples for model training and evaluation. At the same time, there are obvious defects in the analysis method. For example, in CN113283352A, the range of organoids included is not comprehensive, and the function recognition of the organoids included in the image processing process is single. The "ignoring of unnecessary organoids (diameter less than 30 μm, loose shape of outer wall, poor internal necrosis) " limits the range of organoids in the organoid vitality identification process. In CN115641262A, only the identification of organoids in organoid images is proposed, and the organoid identification parameters are not related to the biological characteristics of organoids. There is no correlation analysis of organoid vitality, and high-vitality organoids and aging organoids are not detected and classified. In CN115358973A, the panoramic collection and analysis of organoid image data will ignore the precise morphological characteristics of individual organoids.

[0007] YOLOv3 is a kind of YOLO system, and CBAM is a kind of attention mechanism. After the integration of the two, there are related applications in image analysis, target tracking, identification and positioning, and even student classroom behavior detection and analysis in different environments. However, there is no report on its application in the biological field. SUMMARY

[0008] In view of the important role of organoids in the field of biological and medical research and the current research gap, the present application develops a CBAM-YOLOv3 integrated algorithm by training the target detection method in artificial intelligence with the microscope micrographs generated during the organoid culture process, and constructs an organoid vitality evaluation system based on the algorithm.

[0009] YOLOv3 is a target detection model, and target detection (object detection) relates to the field of computer vision in artificial intelligence. The target detection algorithm is a kind of expansion algorithm of convolutional neural network (CNN) in the field of computer vision, and its purpose is to determine whether a given target exists in a candidate picture. Taking a photographic picture in daily life as an example, a computer model can be trained by giving a target such as a person, an animal, a car, etc. Then the candidate picture is input into the model for target detection analysis, and the model will point out the person, animal, car, etc. in the analyzed image, and the target detection prediction box can be returned in the feedback area of the picture, and the confidence of the candidate target is labeled above the prediction box. The most important evaluation index of the target detection algorithm is the average precision (AP), and the AP value is between 0 and 1. The greater the AP value, the more accurate the model detects the target. The mAP refers to the average value (Mean) of AP.

[0010] CBAM is one of the attention mechanisms, specifically a convolutional block attention module (CBAM), which is a kind of algorithm module combining spatial attention mechanism and channel attention mechanism. If this kind of module is properly integrated with other algorithms, the performance of the original algorithm can be improved, and the model can be further optimized.

[0011] The present application integrates the performance of CBAM with the target detection model YOLOv3 to construct a CBAM-YOLOv3 integrated model for organoid viability prediction.

[0012] In the process of studying the viability of organoids by using the CBAM-YOLOv3 integrated algorithm, the average number of organoids, the average diameter of organoids, the diameter of single cells in the organoid cell suspension, etc. in the microscopic pictures of the microscope are continuously collected as basic evaluation indexes. When judging that the cells are aging and the viability is poor, the commonly used SA-β-Gal staining in cell biology research is used as a basic reference.

[0013] Through a large number of experimental tests, the candidate organoids are precisely classified as being in good vigor or being in aging in the microscopic pictures under a microscope, and meanwhile, the proportions of the organoids in different vigor degrees are synchronously given in the output pictures of the CBAM-YOLOv3.

[0014] In a first aspect, the application provides an organoid vigor evaluation system, comprising:

[0015] An input display module is configured to input microscopic images of the organoids to be detected and display the detection results.

[0016] A detection module is configured to obtain the average diameter of the organoids, the number of the organoids, and the product of the number of the organoids and the average diameter based on the CBAM-YOLOv3 integrated algorithm, and to count the numbers of the organoids in good vigor and the organoids in aging and calculate the respective proportions based on the machine learning results.

[0017] A storage module is configured to store various information during the operation of the organoid vigor evaluation system.

[0018] A control module is configured to control the normal operation of the input display module, the detection module, and the storage module.

[0019] Preferably, the organoids include gastric organoids, small intestinal organoids, and colorectal organoids, and the gastric organoids are taken as representatives for evaluation of the activity degree in the specific embodiments of the application.

[0020] Preferably, the microscopic images of the organoids are collected as follows: when the organoids are cultured for 10-14 days, the microscopic images of the organoids are collected under 10 times magnification of the objective lens of an inverted microscope, and are stored in TIF format after adjustment of pixels or directly.

[0021] The input display module displays the average diameter of the organoids, the number of the organoids, the product of the number of the organoids and the average diameter, and the respective proportions of the organoids in good vigor and the organoids in aging.

[0022] Preferably, in the detection module, the CBAM-YOLOv3 integrated algorithm is configured to add an attention mechanism module CBAM before a feature pyramid network of a backbone feature extraction network Darknet-53 of a YOLOv3 model.

[0023] The detection module first performs machine learning to train the CBAM-YOLOv3 integrated algorithm. The input image is divided into a training set, a validation set and a test set, and the ratio is 7:2:1.

[0024] Before training, a plurality of organ microcosm microscope microscopic images of organ cell groups in the organ culture process are labeled. The active organ and the aging organ are labeled as "Active" or "Aging" respectively by using LabelImg, and then the labeled information is saved in XML format.

[0025] The training process is as follows:

[0026] (1) The number of classifications is the number of classified lesions plus the background parameter, and the number of classifications is 2, which is composed of organoids and background;

[0027] (2) The training parameters include ModelCheckpoint, which saves the training model with the best validation set loss function value; ReduceLROnPlateau, which automatically adjusts the learning rate. When the validation set loss function value does not decrease for 5 periods, the learning rate is adjusted to 1 / 2 of the original; EarlyStopping, which stops training when the validation set loss function does not decrease for 15 periods; Batchsize, which is preferably 8;

[0028] (3) In order to improve the training efficiency, the transfer learning strategy is used. The pre-trained weight trained on the ImageNet data set is loaded into the model for training;

[0029] (4) Rough training by freezing training: freeze the feature extraction layer in the first 184 layers of the CBAM-YOLOv3 integrated algorithm, and set the initial learning rate to 1x10 -3 , and train the model after 184 layers.

[0030] (5) Fine training by unfreezing training: unfreeze the first 184 layers, and set the initial learning rate to 1x10 -4 .

[0031] The detection process of the validation set and the test set is the same as the detection process of the actual application. Based on the CBAM-YOLOv3 integrated algorithm, the average diameter of the organoids, the number of organoids and the product of the number of organoids and the average diameter are obtained. At the same time, based on the machine learning result, the number of active organoids and aging organoids is counted and the proportion of each is calculated.

[0032] In a second aspect of the present application, a derivative of the organoid viability evaluation system is provided: first, an electronic device built into a microscope observation device during organoid culture is provided, including a microscope image storage, an image processor, and a computer program of the organoid viability evaluation system stored on the storage and executable on the processor.

[0033] After the organoid viability evaluation system is completed, machine learning is performed, and the system is built into a microscope. After image capture on an inverted microscope, the organoid culture can be directly processed and evaluated for viability.

[0034] Secondly, an electronic device, such as a computer, loaded with the organoid viability evaluation system is provided, which is communicatively connected to an inverted microscope, acquires corresponding images, and processes and evaluates the viability of the images.

[0035] Thirdly, a non-transitory computer-readable storage medium having a computer program for implementing the organoid viability evaluation system detection process is provided, and the computer program is executed by a processor to implement the organoid viability evaluation steps.

[0036] In a third aspect of the present application, an organoid viability evaluation method is provided, which includes the following four parts:

[0037] S1, a culture and subculture method of human organs;

[0038] S2, microscope microscopic image acquisition and processing of human organs;

[0039] S3, a gastric organoid viability detection method based on CBAM-YOLOv3 integrated algorithm;

[0040] S4, gastric organoid viability detection and classification application based on CBAM-YOLOv3 integrated algorithm.

[0041] The simple steps of each step are as follows:

[0042] S1, establishment and subculture method of organoids

[0043] S1-1 primary culture

[0044] Fresh biopsy or surgical resection specimens are obtained, washed thoroughly, cut into small pieces, then digested in tissue digestion solution; collect the cell suspension, mix the cell suspension with the organoid culture scaffold material at a ratio of 1:1, and place it in a multi-well culture plate, incubate in a 37℃ incubator, then add organoid complete culture medium to each well, and culture in a 37℃, 5%CO2 incubator; subculture or freeze storage is performed according to the need when cultured to the 14th day.

[0045] S1-2 subculture

[0046] Human organ culture to the 10th to 14th day, using TrypLE TM Express Enzyme digestion, collect cell suspension, add Advanced DMEM / F12 culture medium to the system at 500 cells / μl, resuspend the precipitate; mix the cell suspension with the organoid culture scaffold material at 1:1 and place it in a multi-well culture plate, incubate in a 37°C incubator, and after the matrigel is fully solidified, add organoid complete culture medium to each well, and incubate in a 37°C, 5% CO2 incubator,

[0047] S2, microscopic image acquisition and processing of human organ

[0048] S2-1 Microscopic image acquisition of human organ: When the organoid is subcultured to appear cell clusters, the microscopic image of the organoid is collected under an inverted microscope and stored in TIF format;

[0049] S2-2 Microscopic image of human organ organoid cell cluster labeling: Use LabelImg tool with "Organoid" as the label to mark the organoid cell clusters in the thousands of microscopic microscopic images captured during the organoid culture process. In order to train the CBAM-YOLOv3 integrated algorithm, the LabelImg is marked as "Active" or "Aging" for well-living organoids and aging organoids, respectively. The image is marked with a prior box, and then the marked information is saved as XML format.

[0050] S3, gastric organoid vitality detection based on CBAM-YOLOv3 integrated algorithm

[0051] Based on the CBAM-YOLOv3 integrated algorithm (i.e. the combination of CBAM and YOLOv3 before the backbone feature extraction network Darknet-53 and feature pyramid network (FPN) of YOLOv3 model), the average diameter of organoids (Average diameter), the number of organoids (Organoid number) and the product of the number of organoids and the average diameter (No. × Dia.) are obtained, and the number of well-living organoids and aging organoids is counted based on the machine learning results and the proportion of each is calculated.

[0052] S4, application

[0053] S4-1 About the product of the number of organoids and the average diameter (No. × Dia.)

[0054] After repeated experiments, it is shown that with the increase of the number of passages, the No. x Dia. parameter gradually decreases (Pearson r = -0.7988, p = 0.0004), and the No. x Dia. parameter is negatively correlated with the SA-β-Gal staining positive rate (Spearman r = -0.7344, p = 0.0018).

[0055] Therefore, No. x Dia. can be used to judge the vitality of the same type of organ with different passage numbers, and the larger the value means the stronger the vitality of the organoid. However, the No. x Dia. values of the same generation of organoids from different tissues or different individuals are quite different, and it is difficult to directly judge according to the values.

[0056] S4-2 Application of CBAM-YOLOv3 integrated algorithm for predicting organoids with good vitality or aging

[0057] Using the CBAM-YOLOv3 integrated algorithm we constructed, the mAP for detecting organoids with good vitality is 88.02%, the mAP for detecting aging organoids is 77.21%, and the overall mAP value is 82.62%. When the CBAM-YOLOv3 integrated algorithm is used to predict the microscopic pictures of candidate organoids, the output prediction results can show the proportion of organoids with good vitality and aging. After repeated experiments, it is shown that with the increase of the number of passages of organoids, the proportion of aging organoids gradually increases (Spearman r = 0.7591, p = 0.0037), and the SA-β-Gal staining positive rate also gradually increases (Pearson r = 0.8825, P = 0.0001).

[0058] The beneficial guarantees and effects of the present application are as follows:

[0059] 1. The prior art still lacks real-time evaluation indicators for the growth vitality of human gastric organoids. The present application integrates the CBAM attention mechanism module with the YOLOv3 target detection model to form a new CBAM-YOLOv3 integrated algorithm, which solves the problem of accurate and real-time evaluation of the growth state of organoids. By using SA-β-GAL staining as a reference for comparative analysis, the evaluation effect of the CBAM-YOLOv3 integrated algorithm is highly correlated with the SA-β-GAL staining results.

[0060] 2. Based on the CBAM-YOLOv3 integrated algorithm in the evaluation of the vitality of gastric organoids, the No. x Dia. parameter shows good evaluation accuracy, and this parameter is highly correlated with the SA-β-Gal positive rate and the number of passages, and can accurately reflect the cell vitality of organoids.

[0061] 3. The CBAM-YOLOv3 integrated algorithm based on the application can output the proportion of good activity organoids and aging organoids in the detection of gastric organoid activity, providing valuable reference for downstream experiments.

[0062] 4. The CBAM-YOLOv3 integrated algorithm constructed in the application is a proprietary algorithm and related analysis device.

[0063] 5. The microscopic image dataset of the captured human gastric organoids in the application is a proprietary dataset. BRIEF DESCRIPTION OF DRAWINGS

[0064] Figure 1 Figure 1 is a structural diagram of the CBAM-YOLOv3 integrated algorithm.

[0065] Figure 2 Figure 2 is the output parameter of the CBAM-YOLOv3 integrated algorithm, including the average diameter of organoids, the number of organoids, and the product of the number and diameter (No. x Dia.).

[0066] Figure 3 Figure 3 shows the relationship between the No. x Dia. parameter, the SA-β-Gal positive rate, and the number of passages, and shows that the output parameter No. x Dia. of the CBAM-YOLOv3 integrated algorithm is negatively correlated with the number of passages (A) and the SA-β-Gal staining positive rate (B).

[0067] Figure 4 Figure 4 is an example of SA-β-GAL staining of good activity organoids and aging organoids.

[0068] Figure 5 Figure 5 shows the proportion of good activity organoids and aging organoids in the output parameter of the CBAM-YOLOv3 integrated algorithm.

[0069] Figure 6 Figure 6 shows the correlation analysis of the proportion of good activity organoids and aging organoids in the output parameter of the CBAM-YOLOv3 integrated algorithm, the SA-β-Gal staining positive rate, and the number of passages, and shows that the classification results of the CBAM-YOLOv3 integrated algorithm for aging organoids are positively correlated with the number of passages (A) and the SA-β-Gal staining positive rate (B). DETAILED DESCRIPTION

[0070] The application will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the application and not to limit the scope of the application.

[0071] Related terms used in the present application:

[0072] Organoid: a three-dimensional (3D) tissue cell culture system. It mainly relies on stem cells in the tissue to differentiate, self-organize into a structure similar to the tissue of origin in a suitable culture environment, have one or more cells of the tissue of origin, and have certain functions of the tissue of origin cells.

[0073] Attention mechanism: a data processing method in computer machine learning, used to learn and calculate the contribution of input data to output data. It can be embedded in an artificial intelligence network to greatly improve the performance of the analysis model with small amount of calculation and parameters.

[0074] Convolutional block attention module (CBAM): a module that combines spatial attention mechanism and channel attention mechanism. This module can multiply attention maps with input feature maps along independent channels and spatial dimensions to optimize adaptive feature matching algorithms.

[0075] YOLOv3: a target detection algorithm. Target detection algorithm is a kind of expansion algorithm of convolutional neural network (CNN) in computer vision field. Its purpose is to determine whether a given target exists in a candidate picture. YOLOv3 uses Darknet-53 model as the backbone network, and introduces feature pyramid network (FPN) module, thereby improving the detection performance of the model for small targets to be detected.

[0076] Below, taking human gastric organoids as an example, the culture of organoids and the evaluation of organoid viability using CBAM-YOLOv3 integrated algorithm are described. CBAM-YOLOv3 integrated algorithm adds attention mechanism module CBAM before the backbone feature extraction network Darknet-53 and feature pyramid network (FPN) of YOLOv3 model. Its algorithm structure diagram is shown in Figure 1 .

[0077] I. Culture and subculture of human gastric organoids

[0078] 1. Primary culture of human gastric organoids

[0079] 1) The tissue to be tested is taken from the tissue subjected to gastroscopy or surgical resection due to gastric cancer: the surgical resection sample is cut within 30 minutes after being removed from the body. The collected tissue is stored in Advanced DMEM / F12 culture medium (12634010, ThermoFisher Scientific);

[0080] 2) Tissue washing: Wash thoroughly with PBS containing Penicillin-Streptomycin (C0222, Beyotime) and Puromycin (A1113802, Thermo Fisher Scientific), and the total washing times are not less than 4 times;

[0081] 3) Tissue cutting and digestion into cell suspension: Cut the tissue into 1-2 mm3 size using ophthalmic scissors, and then use a mixed tissue digestion solution containing 1-2 mg / mL Collagenase I (40507ES60, Yeasen), 0.5-2.5 mg / mL Collagenase IV (40510ES60, Yeasen), and 0.6-2.4 U / mL Dispase II (40104ES60, Yeasen) to digest at 37°C constant temperature water bath for 1-2 hours;

[0082] 4) Filtration of tissue residues: The tissue residue fragments that cannot be completely digested can be filtered using a 70-100 μm cell filter, and the cell suspension is centrifuged at 1000 rpm for 5 minutes to obtain a cell pellet. The cell pellet is mixed with organoid scaffold material (such as Matrigel, 356231, Corning) at a ratio of 1:1, and inoculated in a 24-well plate at a volume of 50 μL, and then taken out after 30 min of solidification at 37°C. Add 500 μL of complete medium, and place in a 37°C, 5% CO2 incubator for organoid culture.

[0083] 2. Subculture of human gastric organoids

[0084] 1) Digestion of organoids: Remove the culture medium from well-grown organoids, and add 500 μL TrypLE Express Enzyme. Every 15 min, use a gun head to blow thoroughly until the organoids are dissociated into a cell suspension. Add serum-containing medium to mix thoroughly to terminate digestion, and centrifuge at 1500 rpm for 10 min at room temperature. Remove the supernatant; TM 2) Mix the cell pellet with organoid scaffold material (such as Matrigel, 356231, Corning) at a ratio of 1:1 to prepare a 500 cell / μl suspension, and inoculate in a 24-well plate at a volume of 50 μL per well. Incubate in a 37°C incubator for 30 min, and then take out the culture plate after the Matrigel solidifies. Add 500 μL of organoid complete medium, and place in a 37°C incubator for culture;

[0085] 2) Mix the cell pellet with organoid scaffold material (such as Matrigel, 356231, Corning) at a ratio of 1:1 to prepare a 500 cell / μl suspension, and inoculate in a 24-well plate at a volume of 50 μL per well. Incubate in a 37°C incubator for 30 min, and then take out the culture plate after the Matrigel solidifies. Add 500 μL of organoid complete medium, and place in a 37°C incubator for culture;

[0086] 3) Subculture or freeze storage as needed when cultured to the 14th day.

[0087] II. Microscope microscopic image acquisition and processing of human gastric organoids

[0088] 1. Microscope microscopic image acquisition of human gastric organoids. The organoid culture plate cultured to the 10th day was taken out, and the image was collected under the inverted microscope (objective lens 10 times magnification), and stored in TIF format.

[0089] 2. Annotation of organoid clusters in microscope microscopic images of human gastric organoids. Used in model training, using LabelImg tool with "Organoid" as label, a total of 2000 microscope microscopic images collected from 14 samples were labeled, a total of 6220 organoid clusters were labeled. In the study of organoid activity classification, the active organoids and aging organoids were labeled as "Active" or "Aging" respectively using LabelImg. The organoid clusters were labeled with prior box, and the label information was saved in XML format.

[0090] III. Detection of organoid activity using CBAM-YOLOv3 integrated algorithm

[0091] 1. Model training based on CBAM-YOLOv3 integrated algorithm

[0092] 1) Parameter setting

[0093] Number of classes (Num class) setting: number of classified lesions + 1 (background parameter); In this experiment, Num class = 2 (classified into two categories, organoids and background: Organoid and Backgroud).

[0094] Training parameters: including ModelCheckpoint (saving the training model with the best validation set loss function value (Val loss)), ReduceLROnPlateau (learning rate automatic adjustment, when the validation set loss function value (Val loss) does not decrease for 5 cycles, the learning rate is adjusted to 1 / 2 of the original), EarlyStopping (when the validation set loss function (Val loss) does not decrease for 15 cycles, stop training) and Batch size (batch size, the number of pictures input into the model each time, in this experiment, Batch size is set to 8).

[0095] 2) Load pre-trained model. To improve training efficiency, use transfer learning strategy, load pre-trained weight (Yolo_weights.h5) trained on ImageNet dataset into this model for training.

[0096] 3) Input image size adjustment (Input shape). Adjust the input image size to 416x416 pixels.

[0097] 4) Training validation set division. The input image is divided into training set, validation set and test set according to the proportion of 7:2:1 (ensure 90% for model training and validation, 10% for model test).

[0098] 5) Freeze training. Freeze the first 184 layers of CBAM-YOLOv3 integrated algorithm, mainly freeze the feature extraction layer, set the initial learning rate (Learning rate, lr) to 1x10 -3 Since the initial learning rate is relatively large, this part of the training is rough training, the purpose is to train the model after 184 layers.

[0099] 6) Unfreeze training. Unfreeze the first 184 layers, set the initial learning rate to 1x10 -4 The initial learning rate is relatively small, which can fine-tune the model.

[0100] 2, CBAM-YOLOv3 integrated algorithm model training results

[0101] 1) The model training results of CBAM-YOLOv3 integrated algorithm are: mAP 93.20%, F10.86, recall rate 82.88%, accuracy 88.43%;

[0102] 2) CBAM-YOLOv3 integrated algorithm can output parameters including organoid average diameter (Averagediameter), organoid number (Organoid number) and the product of organoid number and average diameter (No. x Dia.), see Figure 2 .

[0103] CBAM-YOLOv3 integrated algorithm and YOLOv3 target detection algorithm parameters and results comparison see table 1:

[0104] Table 1 CBAM-YOLOv3 integrated algorithm and YOLOv3 target detection algorithm main parameter comparison

[0105]

[0106] Note: mAP, mean average precision of each class average precision;

[0107] F1, harmonic mean of precision and recall (F1=2Recall x Precision / (Recall+Precision);

[0108] Recall is the proportion of correctly identified positive samples out of the total number of positive samples (Recall = TP / (TP + FN)).

[0109] Precision is the proportion of correctly identified positive samples out of the total number of identified samples (Precision = TP / (TP + FP)).

[0110] TP (True positive) represents a true positive, meaning the algorithm predicted it as a positive example and the prediction was correct (True).

[0111] FN (False negative) represents a false negative, meaning the algorithm predicted a negative example and made a false prediction.

[0112] FP (False positive) represents a false positive, meaning the algorithm predicts a positive result but the prediction is incorrect (False).

[0113] IV. Validation and Application of the CBAM-YOLOv3 Integrated Algorithm in Detecting Gastric Organoid Viability

[0114] 1. SA-β-GAL staining

[0115] Organoids were tested using a β-galactosidase staining kit (C602, Beyotime, China), which is commonly used in cell senescence assessment, as a comparative reference.

[0116] 1) Digestion of organoids: After culturing organoids to day 14, remove the culture medium and add 500 μL TrypLE. TM Express Enzyme digestive enzymes were thoroughly pipetted every 15 minutes until the organoids were dissociated into a cell suspension. Serum-containing culture medium was then added and thoroughly mixed to terminate the digestion reaction. The mixture was centrifuged at 1500 rpm for 10 minutes at room temperature, and the supernatant was removed.

[0117] 2) Washing: Wash cells once with PBS buffer and centrifuge at 1500 rpm for 5 min;

[0118] 3) Fixation: Add 1 ml of β-galactosidase staining fixative and fix at room temperature for 15 min;

[0119] 4) Washing: Centrifuge the cell suspension at 1500 rpm for 5 min, remove the fixative, and wash the cells 3 times with PBS buffer for 3 min each time, centrifuging at 1500 rpm for 5 min each time.

[0120] 5) Staining: remove PBS buffer, add 0.5ml staining working solution (5ul of beta-galactosidase staining solution A, 5ul of beta-galactosidase staining solution B, 465ul of beta-galactosidase staining solution B, 25ul of X-Gal solution);

[0121] 6) Incubate at 37℃ overnight;

[0122] 7) Staining result observation and photography: remove part of the stained cells, drop onto a glass slide or 6-well plate, count and take photos under a general optical microscope.

[0123] 2. Application of the No. x Dia. parameter in the CBAM-YOLOv3 integrated algorithm to evaluate the activity of organoids.

[0124] The No. x Dia. parameter in the CBAM-YOLOv3 integrated algorithm showed a significant negative correlation with the SA-beta-Gal staining positive rate (Pearson r = -0.7988, p = 0.0004) and the number of passages (Spearman r = -0.7344, p = 0.0018), indicating that this parameter can well reflect the cell activity of organoids. Figure 3

[0125] 3. Application of the CBAM-YOLOv3 integrated algorithm to classify organoids with good activity and aging organoids.

[0126] The CBAM-YOLOv3 integrated algorithm can output the proportion of organoids with good activity and aging organoids in the microscope microscopic pictures. The predicted proportion of aging organoids showed a significant positive correlation with the number of organoid passages (Spearman r = 0.7591, p = 0.0037) and the SA-beta-Gal staining positive rate (Pearson r = 0.8825, P = 0.0001 Figures 4-6 ).

[0127] The above has specifically described examples of the present application, but the present application is not limited to the described embodiments, and those skilled in the art can make equivalent modifications or replacements without deviating from the spirit of the present application, and these equivalent modifications or replacements are all included in the scope defined by the claims of the present application.​

Claims

1. An organoid viability assessment system, characterized in that, include: The input display module takes a microscopic image of the organoid to be tested as input and displays the test results. The detection module uses the CBAM-YOLOv3 ensemble algorithm to obtain the average diameter of organoids, the number of organoids, and the product of the number of organoids and the average diameter. Simultaneously, based on machine learning results, it statistically analyzes the number of viable organoids and aging organoids and calculates their respective proportions. The CBAM-YOLOv3 ensemble algorithm adds an attention mechanism module, CBAM, between the Darknet-53 backbone feature extraction network and the feature pyramid network in the YOLOv3 model. The detection module first performs machine learning, dividing the input image into a training set, a validation set, and a test set in a 7:2:1 ratio. Before training, the organoid cell clusters in the organoid microscope microscopic images were labeled. Viable organoids and aging organoids were marked as "Active" or "Aging" respectively using LabelImg, and then the labeling information was saved in XML format. During training, the number of classifications is the sum of the number of lesions and the background parameter. The number of classifications is 2, consisting of organoids and background. Training parameters include ModelCheckpoint, which saves the trained model with the best validation set loss function value; ReduceLROnPlateau, which automatically adjusts the learning rate, reducing it to half its original value if the validation set loss function value does not decrease for 5 epochs; EarlyStopping, which stops training if the validation set loss function value does not decrease for 15 epochs; and Batchsize. To improve training efficiency, a transfer learning strategy is used, in which pre-trained weights trained on the ImageNet dataset are loaded into this model for training. Coarse training was performed using frozen training: the feature extraction layers of the first 184 layers of the CBAM-YOLOv3 ensemble algorithm were frozen, and the initial learning rate was set to 1×10. -3 Training is performed on models with 184 layers or more; Fine-tuning was performed using unfrozen training: the first 184 layers were unfrozen, and the initial learning rate was set to 1×10⁻⁶. -4 .

2. The organoid viability assessment system according to claim 1, characterized in that, Also includes: The storage module stores various information during the operation of the organoid vitality assessment system; The control module controls the normal operation of the input display module, detection module, and storage module.

3. The organoid viability assessment system according to claim 1, characterized in that: in, The organoids include gastric organoids, small intestinal organoids, and colorectal organoids.

4. The organoid viability assessment system according to claim 1, characterized in that: in, The method for acquiring microscopic images of organoids is as follows: When the organoids are cultured for 10 to 14 days, microscopic images of the organoids are acquired under 10x magnification with an inverted microscope objective and stored in TIF format. The input display module displays the average diameter of organoids, the number of organoids, the product of the number of organoids and the average diameter, the proportion of viable organoids, and the proportion of aging organoids.

5. A built-in electronic device in a microscope observation apparatus during organoid culture, characterized in that, The system includes a microscope image storage device, an image processor, and a computer program for an organoid viability assessment system stored in the storage device and capable of running on the processor, the organoid viability assessment system being as described in any one of claims 1 to 4.

6. A method for assessing organoid viability using the organoid viability assessment system of claim 1, characterized in that, Includes the following steps: A. Microscopic image acquisition and processing of organoids When the organoids were cultured for 10 to 14 days, microscopic images of the organoids were acquired under 10x magnification with an inverted microscope objective, and the pixels of the images were adjusted and stored. B. Organoid viability testing The average diameter of organoids, the number of organoids, and the product of the number of organoids and the average diameter are obtained based on the CBAM-YOLOv3 ensemble algorithm. At the same time, the number of viable organoids and aging organoids are statistically analyzed and their respective proportions are calculated based on machine learning results.

7. The organoid viability assessment method according to claim 6, characterized in that: in, In step A, the methods for establishing and passageing organoids are as follows: (1) Primary culture Obtain fresh biopsy or surgically removed specimens, thoroughly clean them, mince the tissue, and then digest them in tissue digestion solution; collect the cell suspension, mix the cell suspension with organoid culture scaffold material at a 1:1 ratio, place it in a multi-well culture plate, and incubate at 37°C. After the matrix gel has fully solidified, add organoid complete culture medium to each well and incubate at 37°C in a 5% CO2 incubator; after 14 days of culture, passage or cryopreservation may be performed as needed. (2) Subculture Human organs were cultured for 10-14 days, then digested with TrypLE™ Express Enzyme, and the cell suspension was collected. The precipitate was resuspended in Advanced DMEM / F12 medium at a ratio of 500 cells / μl. The cell suspension was then mixed with organoid culture scaffold material at a 1:1 ratio and placed in multi-well plates. The plates were incubated at 37°C. After the matrix gel had fully solidified, organoid complete culture medium was added to each well, and the plates were incubated at 37°C in a 5% CO2 incubator. The image has 416×416 pixels.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the organoid viability assessment steps as described in claim 6.

Images