A method for evaluating the viability of single CTC cells based on deep learning

Abstract

This invention discloses a method for evaluating the viability of a single CTC cell based on deep learning. The method includes the following steps: 1) Preprocessing initial image data of CTC cells in different states to obtain individual CTC cell feature images, constructing datasets corresponding to state classification labels, and randomly dividing the datasets for each state into training, validation, and test sets; 2) Inputting the training and validation sets into the EffficientNet-B0 model for viability interval determination training to obtain the optimal model; 3) Inputting the training and validation sets into the YOLOv5s model for single CTC cell localization and CTC cell viability value prediction training to obtain the optimal model; 4) Inputting the cell image data to be tested into the optimal EffficientNet-B0 and YOLOv5s models respectively to calculate the viability value of a single CTC cell. The advantages are high efficiency, speed, and accuracy.

Description

Technical Field

[0001] This invention relates to a method for assessing the viability of CTC cells, and more particularly to a method for assessing the viability of a single CTC cell based on a deep learning approach. Background Technology

[0002] Circulating tumor cells (CTCs) are closely related to cancer metastasis and recurrence. They are a small number of tumor cells that detach from the primary or metastatic lesion and enter the peripheral blood, escaping immune killing and surviving. As a novel tumor marker, CTCs are increasingly being used for auxiliary tumor staging, prognostic assessment, postoperative recurrence detection, and efficacy evaluation. Currently, the main indicator for CTC detection and monitoring is quantity, but CTC viability is a promising potential indicator. CTC viability reflects the physiological state and function of CTCs, revealing their survival, proliferation rate, and metabolic status. This can provide richer reference information for early tumor detection and diagnosis, accurate diagnosis of metastatic patients, monitoring postoperative tumor recurrence and metastasis, assessing the sensitivity of antitumor drugs and patient prognosis, and selecting individualized treatment strategies. Therefore, developing an efficient and rapid method for assessing CTC viability is crucial.

[0003] Traditional methods for assessing cell viability, including the MTT assay, MTS assay, CCK-8 assay, and flow cytometry, are widely accepted and often considered the standard for evaluation. These techniques typically rely on chemical staining reagents, require a certain staining time, and may be toxic to cells, affecting long-term viability observation. To address these challenges, several innovative cell viability assays have emerged in recent years, including cell membrane potential assays, lensless shadow imaging techniques, and electrochemical assays. However, both traditional methods and these emerging techniques largely focus on viability assessment at the cell population level, indirectly assessing cell viability by calculating the ratio of live cells to total cells. This often fails to achieve precise evaluation at the single-cell level, especially in the field of CTC research. Typically, only 1–10 CTCs are present per milliliter of blood, compared to 5 × 10⁶ CTCs per milliliter. 9 7 × 10 red blood cells and 7 × 10 6 The extreme rarity of CTCs in the blood, with only a few white blood cells, makes accurate assessment of CTC activity particularly difficult.

[0004] Deep learning possesses powerful feature learning capabilities. Through deep, nonlinear neural network structures, it effectively combines low-level features from massive amounts of data to form dense, high-level semantic abstractions, thereby efficiently and accurately extracting deep-level features from the data. It has been widely applied in fields such as image processing, natural language understanding, and speech recognition. As an important technology in computer vision, deep learning has made breakthroughs in the biomedical field, particularly in image analysis, image segmentation, and assisted diagnosis and treatment. While there have been some reports on the application of deep learning in cell viability assessment, such as distinguishing between live and dead cells and assessing cell metabolic capacity through cell images, it remains difficult to accurately determine the viability of individual cells. Currently, there are no reports of using deep learning methods to directly assess the viability of individual cells based on cell morphology features. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide an efficient, fast and accurate method for evaluating the viability of a single CTC cell based on deep learning.

[0006] The technical solution adopted by this invention to solve the above-mentioned technical problems is as follows: a method for evaluating the viability of a single CTC cell based on deep learning, characterized by comprising the following steps:

[0007] (1) Data collection and image preprocessing

[0008] Initial images of CTC cells in different states were acquired using fluorescence electron microscopy. The initial CTC cell images were then processed using the cell segmentation tool Cellpose to outline the morphological contours of individual cells and obtain mask files. The mask files were then imported into ImageJ software to obtain the coordinate information of individual cells. The coordinates of individual cells were then used to perform image segmentation to obtain the feature images of individual CTC cells. A dataset corresponding to the state classification labels was constructed, and the dataset for each state was randomly divided into training set, validation set, and test set.

[0009] (2) Optimization of the CTC cell viability range output model

[0010] A. Input the training set and validation set obtained in step (1) into the EffficientNet-B0 model to perform training for the 100-75%, 75-50%, 50-25%, and 25-0% vitality intervals to obtain the optimal model weight data; load the optimal model weight data into the EffficientNet-B0 model.

[0011] B. Input the test set obtained in step (1) into the EffficientNet-B0 model obtained in step (2)A to perform image classification performance testing, and output the vitality interval of CTC images in the test set. If the preset vitality value of CTC images in the test set falls within the output vitality interval, the optimal EffficientNet-B0 model is obtained; if the preset vitality value of CTC images in the test set falls outside the output vitality interval, repeat step (2)A until the optimal EffficientNet-B0 model is obtained.

[0012] (3) Optimization of CTC cell localization and viability prediction model

[0013] A. Input the training set and validation set obtained in step (1) into the YOLOv5s model to train the localization of individual CTC cells in cell images and the prediction of CTC cell viability values, and obtain the optimal model weight data; load the optimal model weight data into the YOLOv5s model;

[0014] B. The test set obtained in step (1) is input into the YOLOv5s model for image classification performance testing. The location and viability value of CTC cells in the test set are output. If the preset viability value of CTC cells in the test set matches the output viability value, the optimal YOLOv5s model is obtained. If the preset viability value of CTC cells in the test set does not match the output viability value, step (3)A is repeated until the optimal YOLOv5s model is obtained.

[0015] (4) Assessment of viability of individual CTC cells

[0016] A. Images of the CTC-containing sample were acquired using a fluorescence electron microscope. These images were then input into the optimal EffficientNet-B0 model, which outputs the CTC cell viability range, with upper and lower values ​​designated as v1 and v2 respectively.

[0017] B. Input the cell image data obtained in step (4)A into the optimal YOLOv5s model, output the detection probability P1 based on the v1 vitality value dataset, output the detection probability P1 based on the v2 vitality value dataset, and calculate the vitality value of a single CTC cell using v1*P1+v2*P2.

[0018] Furthermore, the initial images of CTC cells in different states in step (1) were obtained as follows: HeLa cells were stimulated with exogenous reactive oxygen species to change their viability. First, CTC cells were incubated in 2 mL of culture medium for 18 h in a 5% CO2, 37℃ environment, and 1.5 mL of the upper culture medium was aspirated. Well-grown, adherent CTC cells were added to 1 mL of 1×PBS solution containing H2O2 at concentrations of 0 mmol / L, 1 mmol / L, 8 mmol / L, 15 mmol / L, and 30 mmol / L, respectively, and incubated for 40 min to obtain CTC cells in five classification states i, ii, iii, iv, and v, with corresponding viability values ​​of 100%, 75%, 50%, 25%, and 0%, respectively.

[0019] Furthermore, the ratio of the number of training set, validation set and test set mentioned in step (1) is 6:3:1.

[0020] Furthermore, the parameters of the EffficientNet-B0 model in step (2) are set as follows: the number of samples in each training batch is specified as 32, the total number of training rounds is set to 100, and the number of categories in the dataset is 5.

[0021] Furthermore, in step (2), the core structural module of the EffficientNet-B0 model is the Mobile Flip Bottleneck Convolution (MBConv) module. The Mobile Flip Bottleneck Convolution module performs a 1×1 pointwise convolution on the input CTC image and changes the output channel dimension according to the expansion ratio, and then performs a k×k depthwise convolution. When compression and activation operations are introduced, the Mobile Flip Bottleneck Convolution will be performed after the depthwise convolution, and the original channel dimension will be restored at the end of the 1×1 pointwise convolution. Then, connection deactivation and input skip connections are performed to give the model a random depth.

[0022] Furthermore, in step (3), the YOLOv5s model includes a Conv module and a C3 module. The input is upsampled twice and downsampled twice by the Conv module. Then, the C3 module is used to extract features. The C3 module defines a neural network module for implementing the CSP Bottleneck structure.

[0023] Compared with existing technologies, the advantages of this invention are as follows: This invention provides a method for evaluating the viability of individual CTC cells based on deep learning. It collects images of CTCs with different viability levels, trains a deep learning model, extracts their morphological features, and then applies this to the identification of individual CTC cells and the evaluation of single-cell viability. The deep learning model constructed using EfficientNet-B0 and YOLOv5s exhibits good predictive performance and has the advantages of being label-free, non-toxic, and accurately evaluating the viability of individual CTCs, thus possessing significant potential application value in the medical field. Attached Figure Description

[0024] Figure 1 For (A) the relationship between H2O2 concentration and cell viability determined by CCK-8 assay, (B) cell image of state i, (C) cell morphology of state ii, (D) cell morphology of state iii, (E) cell morphology of state iv, (F) cell morphology of state v, scale bar, 10 μm.

[0025] Figure 2 To evaluate the activity of CTCs treated with different concentrations of H2O2, the AF concentrations were 20 mmol / L, 12 mmol / L, 10 mmol / L, 6 mmol / L, 5 mmol / L and 0.6 mmol / L respectively.

[0026] Figure 3 Cell viability at different H2O2 concentrations was simultaneously measured using a deep learning model and the CCK-8 method.

[0027] Figure 4 Bland-Altman diagrams for deep learning models and the CCK-8 method;

[0028] Figure 5 The performance evaluation of the deep learning model jointly constructed by EfficientNet-B0 and YOLOv5s in this invention includes (A) precision, (B) recall, (C) F1 score, (D) AP, and (E) mAP.

[0029] Figure 6 The CTC activity of the deep learning model constructed in this invention is measured using real samples. Scale bar, 10 μm. Detailed Implementation

[0030] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. I. Specific Implementation Methods

[0032] 1. Data collection and image preprocessing

[0033] Human cervical cancer cells (HeLa) were selected as the research subject. HeLa cells were stimulated with exogenous reactive oxygen species (ROS) to alter their viability. First, HeLa cells were incubated in 2 mL of culture medium for 18 h at 37°C with 5% CO2, and then 1.5 mL of the supernatant was removed. Well-grown, adherent HeLa cells were then added to 1 mL of 1×PBS solution containing H2O2 at concentrations of 0 mmol / L, 1 mmol / L, 8 mmol / L, 15 mmol / L, and 30 mmol / L, respectively, and incubated for 40 min to obtain five taxonomic states (i, ii, iii, iv, and v) of CTC cells, with corresponding viability values ​​of 100%, 75%, 50%, 25%, and 0%, respectively.

[0034] Initial images of CTC cells in corresponding states were acquired using fluorescence electron microscopy. The initial CTC cell images were then processed using the cell segmentation tool Cellpose to outline the morphological contours of individual cells and obtain mask files. The mask files were imported into ImageJ software to obtain the coordinate information of individual cells. The coordinates of individual cells were then used to segment the images to obtain the feature images of individual CTC cells. A dataset corresponding to the state classification labels was constructed, with 33358, 39952, 17594, 28821, and 23078 individual cell images for states i, ii, iii, iv, and v, respectively. The dataset for each state was randomly divided into training, validation, and test sets in a 6:3:1 ratio.

[0035] 2. Optimization of the CTC cell viability range output model

[0036] A. Input the training set and validation set obtained in step 1 into the EffficientNet-B0 model to perform training for the 100-75%, 75-50%, 50-25%, and 25-0% vitality intervals to obtain the optimal model weight data; load the optimal model weight data into the EffficientNet-B0 model.

[0037] B. Input the test set obtained in step 1 into the EffficientNet-B0 model obtained in step 2A to perform image classification performance testing. Output the vitality interval of CTC images in the test set. If the preset vitality value of the CTC images in the test set falls within the output vitality interval, the optimal EffficientNet-B0 model is obtained. If the preset vitality value of the CTC images in the test set falls outside the output vitality interval, repeat step 2A until the optimal EffficientNet-B0 model is obtained.

[0038] The above process begins with parameter settings: for example, specifying the number of samples in each training batch as 32, which is the number of samples used each time the model is updated; setting the total number of training rounds to 100, which is the number of times the entire training dataset is traversed; and setting the number of classes in the dataset to 5, which is used to specify the number of output classes of the model.

[0039] The core structural module of the EffficientNet-B0 model is the Mobile Flip Bottleneck Convolution (MBConv) module. The Mobile Flip Bottleneck Convolution module performs a 1×1 pointwise convolution on the input CTC image and changes the output channel dimension according to the expansion ratio, followed by a k×k depthwise convolution. When compression and activation operations are introduced, the Mobile Flip Bottleneck Convolution is performed after the depthwise convolution, and the original channel dimension is restored at the end of the 1×1 pointwise convolution. Then, connection deactivation and input skip connections are performed to give the model a random depth.

[0040] 3. Optimization of the CTC cell localization and viability prediction model

[0041] A. Input the training set and validation set obtained in step 1 into the YOLOv5s model to train the localization of individual CTC cells in cell images and the prediction of CTC cell viability values, and obtain the optimal model weight data; load the optimal model weight data into the YOLOv5s model;

[0042] B. The test set obtained in step 1 is input into the YOLOv5s model for image classification performance testing. The location and viability values ​​of CTC cells in the test set are output. If the preset viability values ​​of CTC cells in the test set match the output viability values, the optimal YOLOv5s model is obtained. If the preset viability values ​​of CTC cells in the test set do not match the output viability values, step 3A is repeated until the optimal YOLOv5s model is obtained.

[0043] The YOLOv5s model described above includes a Conv module and a C3 module. The Conv structure performs two upsampling and two downsampling processes on the input: the upsampling process doubles the size of the feature map; the downsampling process uses a convolutional layer to reduce the size of the feature map by half in the first downsampling and a convolutional layer to reduce the size of the feature map by half again in the second downsampling.

[0044] The C3 module is then used to extract features. The C3 module defines a neural network module for implementing the CSP Bottleneck structure. CSPdarknet first extracts three different layers of feature maps feat1, feat2 and feat3 through the backbone model. Then, according to different conditions, a series of convolution operations are performed on the extracted features in order to carry out the next step of feature processing and fusion.

[0045] 4. Assessment of individual CTC cell viability

[0046] A. The test sample containing CTC cells is captured by fluorescence electron microscopy to obtain the cell image data. The cell image data is input into the optimal EffficientNet-B0 model to output the viability range of CTC cells, with the upper and lower values ​​being v1 and v2 respectively.

[0047] B. Input the cell image data obtained in step 4A into the optimal YOLOv5s model, output the detection probability P1 based on the v1 vitality value dataset, output the detection probability P1 based on the v2 vitality value dataset, and calculate the vitality value of a single CTC cell according to the formula v1*P1+v2*P2.

[0048] The detection probability mentioned above is an inherent output value of the YOLOv5s model. It compares the target CTC image with images in the v1 or v2 (e.g., 25%) classification dataset to observe how similar the target image is to the images in the dataset. The output probability value is usually between 0 and 1.

[0049] II. Analysis of Experimental Results

[0050] 1. Assess viability based on cell morphology

[0051] HeLa cells were cultured in culture flasks containing DMEM medium (containing 10% fetal bovine serum, 100 U / mL streptomycin, and 100 U / mL penicillin) at 5% CO2 and 37°C. Afterward, the old medium was discarded, and the cells were washed with 1 mL of 1×PBS. Then, 1 mL of 0.25% trypsin digest was added to the culture flasks for 1.5 min, the trypsin digest was discarded, and the flasks were incubated at 5% CO2 and 37°C for 1 min. Then, 2 mL of fresh medium was added, and the cells were repeatedly pipetted to suspend them. Finally, the cells were transferred to 35 mm × 12 mm culture dishes for later use. HeLa cell viability was altered by stimulating them with exogenous reactive oxygen species (ROS). First, HeLa cells were incubated in 2 mL of medium at 5% CO2 and 37°C for 18 h, and then 1.5 mL of the supernatant was aspirated. Subsequently, 1 mL of 1×PBS solution containing a certain concentration of H2O2 was added and incubated for 40 min to investigate the relationship between H2O2 concentration and cell viability as determined by the CCK-8 assay.

[0052] The steps for determining cell viability using the CCK-8 assay are as follows:

[0053] (1) Well-adhered cells were digested with trypsin to obtain a cell suspension and counted. In a 96-well plate, 100 μL of cell suspension was seeded into each well to obtain approximately 1000 cells.

[0054] (2) Place the culture plate in a 5% CO2, 37℃ culture environment for 2 hours to pre-culture the cells to allow them to adhere;

[0055] (3) Remove 75 μL of the upper culture medium and add 50 μL of 1×PBS solution containing different concentrations of H2O2 to the wells and incubate for 40 min;

[0056] (4) Transfer the suspension mixture in the well to a 2 mL filter centrifuge tube, centrifuge at 800×g for 5 min, redisperse the cells on the filter membrane in 100 μL of cell culture medium and transfer them to the original well;

[0057] (5) Add 10 μL of CCK-8 solution and continue culturing for 2.5 h;

[0058] (6) Measure the absorbance (A) at 450 nm using an ELISA reader, and calculate the cell viability value using the CCK-8 assay according to the following formula:

[0059] Cell viability = [(A (实验组) -A (空白组) ) / (A (对照组) -A (空白组) )]×100%

[0060] A (实验组) : Absorbance of experimental wells (cells treated with CCK-8 solution and H2O2);

[0061] A (对照组) : Absorbance of control wells (CCK-8 solution, cells);

[0062] A (空白组) : Absorbance of blank well (CCK-8 solution).

[0063] The relationship between H2O2 concentration and cell viability as determined by the CCK-8 assay is as follows: Figure 1 As shown in Figure A, when the H2O2 concentration is 0, 1, 8, 15, and 30 mmol / L, the cell viability values ​​corresponding to states i, ii, iii, iv, and v are 100%, 75%, 50%, 25%, and 0%, respectively. Figure 1 As shown in BF, the treated CTC cells were classified into five states: state i, where the cells exhibited optimal growth due to good adhesion ( Figure 1 B); State ii, the cell membrane exhibits vacuolation, accompanied by budding of some cells ( Figure 1 C); State iii, cell volume shrinks, and apoptotic bodies begin to appear around each cell. Figure 1 D); In state iv, the volume of apoptotic bodies around the cells is significantly increased ( Figure 1 E); State v, cell membrane ruptures, cells cannot adhere to the wall and are in a suspended state. Figure 1 F). The results indicate that assessing cell viability based on cell morphology characteristics is feasible.

[0064] 2. Model Performance Evaluation

[0065] (1) Model accuracy

[0066] A deep learning model constructed using specific embodiments was used to analyze cell images of CTCs treated with different concentrations (20, 12, 10, 6, 5, 600 mmol / L) of H2O2 to analyze CTC viability. The results are as follows: Figure 2 As shown in (AF), each CTC in the image has a specific viability value output. The model results were then compared with the cell viability measured by CCK-8 assay, as shown below. Figure 3 As shown, the two methods exhibit high consistency in CTC activity analysis. Furthermore, as... Figure 4 The Bland-Altman analysis shows that the ordinates of all points fall within this range, and the average difference between this measurement and CCK-8 is -0.255. Furthermore, the 95% confidence intervals for the difference between the two measures are between -3.042 and 2.532, further demonstrating good consistency between the constructed deep learning model and the CCK-8 method. These results indicate that the deep learning model can accurately distinguish cells at different viability states and precisely determine the viability of each CTC based on cell morphology characteristics. These analytical results validate the effectiveness of the cell morphology-based deep learning model for cell viability assessment.

[0067] (2) Comparative Analysis

[0068] As shown in Table 1, compared with baseline models such as VGG-16, ResNet18, ResNeXt50_32X4D, and DenseNet-121, the EfficientNet-B0 model used in the experiment can extract cell state features and achieve accurate classification of cell images, showing the best performance across all evaluation metrics.

[0069] Table 1 compares the accuracy and loss rate of different CNN models in classifying cell viability.

[0070]

[0071] Then, a series of standard metrics commonly used to evaluate machine learning models were applied to assess the overall performance of the deep learning model jointly built by EfficientNet-B0 and YOLOv5s. The results of the performance metrics are shown in [link to performance metrics]. Figure 5 As shown, this includes an accuracy rate of 90.12% ( Figure 5 A) Recall rate 42.99% Figure 5 B) The F1 score is close to 1 ( Figure 5 C), AP 82.31% Figure 5 D) and mAP 82.31% ( Figure 5 E). The above analysis results all indicate that the model has good predictive performance.

[0072] III. Determination of CTC activity in real samples

[0073] Cell images of whole blood samples were acquired and imported into a constructed deep learning model to analyze the viability of CTCs in whole blood. The results are as follows: Figure 6 As shown in (AF), the model can quickly and accurately identify CTCs in whole blood samples and output viability values ​​based on cell morphology. The results indicate that the deep learning-based cell viability assessment model can be applied to the analysis of real whole blood samples.

[0074] The foregoing description is not intended to limit the invention, nor is the invention limited to the examples given. Any changes, modifications, additions, or substitutions made by those skilled in the art within the scope of the invention should also be considered within the protection scope of the invention.

Claims

1. A method for evaluating the viability of a single CTC cell based on deep learning, characterized in that... Includes the following steps: (1) Data collection and image preprocessing Initial images of CTC cells in different states were acquired using fluorescence electron microscopy. The initial CTC cell images were then processed using the cell segmentation tool Cellpose to outline the morphological contours of individual cells and obtain mask files. The mask files were then imported into ImageJ software to obtain the coordinate information of individual cells. The coordinates of individual cells were then used to perform image segmentation to obtain the feature images of individual CTC cells. A dataset corresponding to the state classification labels was constructed, and the dataset for each state was randomly divided into training set, validation set, and test set. (2) Optimization of the CTC cell viability range output model A. Input the training set and validation set obtained in step (1) into the EfficientNet-B0 model to perform vitality interval determination training of 100-75%, 75-50%, 50-25%, and 25-0% to obtain the optimal model weight data; load the optimal model weight data into the EfficientNet-B0 model. B. Input the test set obtained in step (1) into the EfficientNet-B0 model obtained in step (2)A to perform image classification performance testing, and output the vitality interval of CTC images in the test set. If the preset vitality value of CTC images in the test set falls within the output vitality interval, the optimal EfficientNet-B0 model is obtained; if the preset vitality value of CTC images in the test set falls outside the output vitality interval, repeat step (2)A until the optimal EfficientNet-B0 model is obtained. (3) Optimization of CTC cell localization and viability prediction model A. Input the training set and validation set obtained in step (1) into the YOLOv5s model to train the localization of individual CTC cells in cell images and the prediction of CTC cell viability values, and obtain the optimal model weight data; load the optimal model weight data into the YOLOv5s model; B. The test set obtained in step (1) is input into the YOLOv5s model for image classification performance testing. The location and viability value of CTC cells in the test set are output. If the preset viability value of CTC cells in the test set matches the output viability value, the optimal YOLOv5s model is obtained. If the preset viability value of CTC cells in the test set does not match the output viability value, step (3)A is repeated until the optimal YOLOv5s model is obtained. (4) Assessment of viability of individual CTC cells A. Images of the CTC-containing sample were acquired using a fluorescence electron microscope. These images were then input into the optimal EfficientNet-B0 model, which outputs the viability range of the CTC cells, with upper and lower limits as follows: v 1 and v 2; B. Input the cell image data obtained in step (4)A into the optimal YOLOv5s model, and according to... v The dataset with 1 vitality value outputs the detection probability. P 1. According to v The dataset with 2 vitality values ​​outputs the detection probability. P 1. Utilize v 1 * P 1+ v 2 * P 2. Calculate the viability value of a single CTC cell.

2. The method for evaluating the viability of a single CTC cell based on deep learning according to claim 1, characterized in that... The initial images of CTC cells in different states in step (1) were obtained as follows: HeLa cells were stimulated by exogenous reactive oxygen species to change their viability. First, CTC cells were incubated in 2 mL of culture medium for 18 h in a 5% CO2, 37 ℃ culture environment, and 1.5 mL of the upper culture medium was aspirated. Well-grown and adherent CTC cells were added to 1 mL of 1× PBS solution containing H2O2 at concentrations of 0 mmol / L, 1 mmol / L, 8 mmol / L, 15 mmol / L, and 30 mmol / L, respectively, and incubated for 40 min to obtain CTC cells in five classification states i, ii, iii, iv, and v, with corresponding viability values ​​of 100%, 75%, 50%, 25%, and 0%, respectively.

3. The method for evaluating the viability of a single CTC cell based on deep learning according to claim 1, characterized in that: The ratio of the number of training set, validation set and test set mentioned in step (1) is 6:3:

1.

4. The method for evaluating the viability of a single CTC cell based on deep learning according to claim 1, characterized in that... In step (2), the parameters of the EfficientNet-B0 model are set as follows: the number of samples in each training batch is specified as 32, the total number of training rounds is set to 100, and the number of classes in the dataset is 5.

5. The method for evaluating the viability of a single CTC cell based on deep learning according to claim 1, characterized in that: In step (2), the core structural module of the EfficientNet-B0 model is the Mobile Flip Bottleneck Convolution (MBConv) module. The Mobile Flip Bottleneck Convolution module performs a 1×1 pointwise convolution on the input CTC image and changes the output channel dimension according to the expansion ratio, and then performs a k×k depthwise convolution. When compression and activation operations are introduced, the Mobile Flip Bottleneck Convolution will be performed after the depthwise convolution. The original channel dimension is restored at the end of the 1×1 pointwise convolution. Then, connection deactivation and input skip connections are performed to give the model random depth.

6. The method for evaluating the viability of a single CTC cell based on deep learning according to claim 1, characterized in that: In step (3), the YOLOv5s model includes a Conv module and a C3 module. The input is upsampled twice and downsampled twice by the Conv module. Then, the C3 module is used to extract features. The C3 module defines a neural network module for implementing the CSPBottleneck structure.

Images