Rapid circulating tumor cell detection method based on deep learning

Abstract

The invention provides a circulating tumor cell rapid detection method based on deep learning, and the method comprises the steps: firstly taking a RetinaNet network structure as a basis, employing RepVGG as a backbone network Backbone, employing a multi-branch structure during training, and employing a single-path structure during reasoning; a route from top to bottom is added on the basis of the FPN, and semantic information of a bottom layer and semantic information of a top layer are fused; simplifying a detection head Head, and optimizing a classification and regression sub-network; the loss function is reasonably optimized for different stages; and processing a new image collected from the biological sample by using the rapid detection model, and outputting a CTCs detection result. Cell recognition, detection and classification are carried out on the fluorescent microscopic image, the rapid detection model takes into account a plurality of performance indexes such as cell detection efficiency and accuracy, has relatively high accuracy and relatively strong robustness, and solves the problems of low circulating tumor cell detection speed, low accuracy and the like in the prior art.

Description

Technical Field

[0001] The present invention relates to the field of intersectional technology of artificial intelligence and biomedicine, mainly to the technical field of combining deep learning and image processing, and in particular to a method for rapid detection of circulating tumor cells based on deep learning. Background Art

[0002] Circulating tumor cells (CTCs) refer to a small number of tumor cells that shed from solid tumor lesions (primary or metastatic lesions) and enter the peripheral blood circulation, surviving after escaping immune killing. They are closely related to tumor staging, recurrence, metastasis, and drug resistance. As a typical biomarker in liquid biopsy, CTCs have broad application prospects in early tumor diagnosis, prognosis, and efficacy monitoring. CTCs are very rare in the patient's blood. It is usually necessary to separate and enrich CTCs in the patient's blood, then stain and prepare slides, and finally detect and analyze them on microscopic images. At present, pathologists generally use manual reading to detect CTCs and count them accurately. This manual detection method is not only inefficient, but also affected by subjective factors such as the pathologist's technical level and experience. In recent years, with the continuous expansion of the demand for automated detection of CTCs, researchers have introduced traditional image processing algorithms into CTCs image detection, such as distance transformation, morphological algorithms, and maximum stable value area detection. However, these methods have poor specificity and low sensitivity, resulting in low detection accuracy. In addition, researchers have applied machine learning to the detection and counting of CTCs. Machine learning has also been applied to CTC detection due to its advantages such as speed and ability to overcome noise. Most studies use physical properties such as size, density, deformability or adhesion of tumor cells for feature extraction, and then select different classifiers for training and classification to finally obtain the detection results. However, this method requires manual feature extraction and relies too much on the knowledge and experience of experts, resulting in large deviations in detection.

[0003] Deep learning has been widely used in biomedical image analysis, and has achieved good results in medical image recognition, segmentation and classification, and has been successfully applied to cancer diagnosis and tissue identification. Compared with traditional manual feature extraction methods, image processing methods based on convolutional neural networks (CNNs) can automatically learn complex data features, can be trained end-to-end, and reduce subjectivity and consistency bias. At present, only a few studies have explored the use of deep learning for CTCs detection. For example, He et al. built a new CNN network based on AlexNet, and used fluorescence in situ hybridization (imFISH) to achieve CTCs detection and counting on fluorescently labeled images, reflecting the advantages of deep learning in image feature extraction; Guo et al. used some classic CNN models, including VGG16, VGG19, ResNet18, AlexNet, and ResNet50, to perform binary classification tasks on multiple fluorescently labeled CTCs images, achieving a high accuracy rate. The model structure proposed by the above studies is relatively simple, and only involves single target detection or binary classification problems. The model performance cannot meet the requirements of clinical applications.

[0004] After CTCs are captured, detection and identification become the key and difficult points. Although the traditional CTCs detection method of manual film reading can detect CTCs, it has a low level of automation, is time-consuming, has low accuracy, and is affected by subjective errors. Applying deep learning methods to CTCs detection has the following advantages: (1) Using deep learning methods to train and identify CTC sample images is more efficient and accurate; (2) Compared with traditional CTC detection methods, the CTC detection model based on convolutional neural networks does not require prior knowledge of complex mechanisms and has higher accuracy and reliability; (3) Using deep learning networks for training, prediction, and classification can eliminate personal subjective errors to a large extent.

[0005] At present, the deep learning models used in CTCs detection are relatively simple, with poor sensitivity and accuracy, and cannot meet the requirements of clinical application.

[0006] The above-mentioned record of background technical knowledge is intended to help ordinary technicians in this field understand the existing technologies that are relatively close to the present invention, and at the same time facilitate the understanding of the inventive concept and technical solution of the present invention. It should be clear that in the absence of clear evidence that the above-mentioned content has been disclosed before the filing date of this patent application, the above-mentioned background technology should not be used to evaluate the novelty of the technical solution of this application. Summary of the invention

[0007] Technical issues In order to solve the above problems, the purpose of the present invention is to provide a method for rapid detection of circulating tumor cells based on deep learning, propose a new rapid detection model for rapid detection of CTCs, and relate to a method for rapid detection of tumor cells based on deep learning and image processing, which performs cell recognition, detection and classification on fluorescent microscopic images. The rapid detection model takes into account multiple performance indicators such as cell detection efficiency and accuracy, has high accuracy and strong robustness, eliminates personal subjective errors to a large extent, and solves the problems of slow detection speed and low accuracy of circulating tumor cells in the prior art.

[0008] Technical Solution That is, the present invention includes the following technical solutions.

[0009] Solution 1: A method for constructing a rapid detection model for circulating tumor cells, including: (1) Based on the RetinaNet network structure, the backbone network uses RepVGG. During training, a multi-branch structure is used to enhance feature extraction. During reasoning, the idea of ​​structural reparameterization is used to transform it into a single-channel structure, which reduces the memory usage and speeds up the reasoning speed. (2) Based on the traditional feature graph pyramid network FPN, a top-down route is added to fuse the semantic information of the bottom layer and the top layer, thereby compensating and enhancing the positioning information; (3) Simplify the detection head, optimize the classification and regression sub-networks, reduce the number of serial convolutions and channels on the basis of the traditional RetinaNet detection head, and adopt large separable convolutions; while reducing model parameters and computational complexity, increase the receptive field of the feature map, thereby enhancing the model detection performance; (4) For different stages, reasonably optimize the loss functions: Smooth L1 loss, GIoU loss and varifocalloss; obtain a fast detection model.

[0010] Furthermore, the multi-branch structure of step (1) includes at least one of a Conv-BN structure and a BN structure.

[0011] Furthermore, a weight adjustment mechanism is set in the step (1) to dynamically adjust the weights of different branch structures; the weight adjustment mechanism enables the model to automatically adjust the weights of each branch structure according to the different characteristics of the input data, and dynamically select which branches are active at a specific moment. This selection can be based on the characteristics of the data or the needs of the current training stage, which helps to enhance the generalization ability of the model, thereby improving the adaptability and efficiency of the model. The multi-branch structure needs to process more parameters and calculations during training, which will make the training process more complicated and time-consuming. In addition, since the multi-branch structure needs to save the intermediate results of each branch until the last step of the fusion operation, this will increase the memory access cost during reasoning, further slowing down the processing speed. Further verification shows that for the same batch of training set data, after setting the weight adjustment mechanism, it helps to reduce the training time by at least 10% while keeping the model accuracy unchanged, so that the rapid detection model of circulating tumor cells has the advantages of high accuracy and fast reasoning speed when performing circulating tumor cell detection, and can achieve efficient cell multi-classification recognition and CTCs detection.

[0012] Furthermore, in step (4), an IoU-Aware Class Quartification Score (IACS) which can simultaneously represent the object existence confidence and positioning accuracy is used during the training process to produce a more accurate detection level in the dense object detector.

[0013] Furthermore, in step (4), the loss function is reasonably optimized as follows: for bounding box regression, the loss function in the early training stage is dominated by Smooth L1 loss; the loss function after the bounding box prediction is stabilized is dominated by GIoU loss; for target classification, varifocal loss is used.

[0014] Furthermore, the method for constructing the rapid detection model of circulating tumor cells also includes: The rapid detection model was trained using a VOC format annotated dataset obtained using the following steps: CTCs were separated and enriched from the blood of tumor patients, multi-channel fluorescence images were collected using a fluorescence microscope after immunofluorescence staining, the fluorescence images were preprocessed to make various cells easier to distinguish visually, and the images were manually annotated to indicate the location and category of each cell to obtain a VOC format annotated dataset.

[0015] Furthermore, the preprocessing of the fluorescence image includes: A median filter was applied to the fluorescence image to reduce salt and pepper noise; Use bilinear downsampling to reduce shot noise; The pixel intensities of each image were normalized to have the same mean and standard deviation.

[0016] Option 2: A rapid detection model for circulating tumor cells obtained by the construction method described above.

[0017] Option three, a method for rapid detection of circulating tumor cells for non-disease diagnosis and treatment based on deep learning, comprising applying the aforementioned rapid detection model for circulating tumor cells to process new images collected from biological samples and outputting CTCs detection results.

[0018] Furthermore, the biological sample is selected from blood or lymph.

[0019] Option 4: Application of the aforementioned deep learning-based rapid detection method for circulating tumor cells for non-disease diagnosis and treatment in biological sample detection.

[0020] Furthermore, the biological sample is selected from blood or lymph.

[0021] Scheme 5: Application of the rapid detection model of circulating tumor cells obtained by the aforementioned construction method in biological sample detection.

[0022] Furthermore, the biological sample is selected from blood or lymph.

[0023] Option six, a computer device, comprising a memory, a processor, a communication interface and a communication bus; wherein the memory, the processor and the communication interface communicate with each other via the communication bus; the memory is used to store computer programs; the processor is used to execute the computer programs stored in the memory, and when the processor executes the computer program, it implements at least one step of the aforementioned method for constructing a circulating tumor cell rapid detection model or the method for rapid detection of circulating tumor cells based on deep learning.

[0024] Option 7, a computer-readable storage medium, having a computer program stored thereon, which, when executed by a processor, implements at least one step of the aforementioned method for constructing a rapid detection model for circulating tumor cells or the method for rapid detection of circulating tumor cells based on deep learning.

[0025] Based on the common knowledge in this field, the above-mentioned preferred conditions can be combined with each other to obtain a specific implementation method.

[0026] Beneficial Effects According to the present invention, based on the RetinaNet network structure, the backbone network Backbone adopts RepVGG, and on the basis of the traditional feature graph pyramid network FPN, a top-down routing is added to fuse the semantic information, simplify the detection head Head, optimize the classification and regression sub-networks, and reasonably optimize the loss function for different stages. The rapid detection model is trained with a labeled data set in the VOC format, thereby obtaining a circulating tumor cell rapid detection model, which has the advantages of high accuracy and fast reasoning speed when performing circulating tumor cell detection, can realize efficient cell multi-classification recognition and CTCs detection, eliminate personal subjective errors to a large extent, and solve the problems of slow circulating tumor cell detection speed and low accuracy in the prior art.

[0027] The present invention adopts the above technical solution to achieve the above purpose, which makes up for the shortcomings of the prior art and has reasonable design and convenient operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to make the above and / or other purposes, features, advantages and examples of the present invention more obvious and easy to understand, the drawings required for use in the specific implementation of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without paying creative work.

[0029] Figure 1 A schematic diagram showing the construction route of a rapid detection model for circulating tumor cells; Figure 2 A schematic diagram of the backbone network Backbone representing the circulating tumor cell rapid detection model; Figure 3 Schematic diagram of the feature pyramid FPN structure representing the rapid detection model of circulating tumor cells; Figure 4 A schematic diagram of a detection head representing a rapid detection model for circulating tumor cells; Figure 5 Schematic diagram showing the cell allocation of the dataset; Figure 6 Indicates an example of a fluorescent image and a test result (where A indicates an image example, B indicates a test result, R (red) indicates CTC, G (green) indicates a normal cell, and U (blue) indicates a cell nucleus); Figure 7 Represents the Epoch loss curve; Figure 8 Indicates validation set Map curve 1; Fig. 9 Represents the validation set Map curve 2. DETAILED DESCRIPTION

[0030] Those skilled in the art can refer to the content of this article and appropriately replace and / or modify the process parameters to achieve the same. However, it should be particularly noted that all similar replacements and / or modifications are obvious to those skilled in the art and are considered to be included in the present invention. The products and preparation methods described in the present invention have been described through preferred examples. It is obvious that relevant personnel can modify or appropriately change and combine the products and preparation methods described herein without departing from the content, spirit and scope of the present invention to implement and apply the technology of the present invention.

[0031] Unless otherwise defined, technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the invention belongs. The present invention uses the methods and materials described herein; however, other suitable methods and materials known in the art may also be used. The materials, methods, and examples described herein are illustrative only and are not intended to be limiting. All publications, patent applications, patents, provisional applications, database entries, and other references mentioned herein are incorporated herein by reference in their entirety. In the event of a conflict, the present specification, including definitions, shall prevail.

[0032] Unless specifically stated, the materials, methods, and examples described herein are illustrative only and not limiting. Although methods and materials similar or equivalent to those described herein can be used in the practice or testing of the present invention, suitable methods and materials are still described herein.

[0033] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present invention. At the same time, the embodiments in this application and the features in the embodiments can be combined with each other without conflict.

[0034] It should be understood that any technical solution claimed for protection in the present invention does not involve the diagnosis and treatment of diseases.

[0035] In order to facilitate understanding of the embodiments of the present invention, the abbreviations and key terms that may be involved in the embodiments of the present invention are first explained or defined. For undefined abbreviations or key terms, they are commonly understood by those skilled in the art.

[0036] In addition, the test methods used in the examples are conventional methods unless otherwise specified; the materials, reagents, etc. used are all commercially available unless otherwise specified. The reagents or instruments used without indicating the manufacturer are all conventional products that can be obtained commercially. All the publications and other references mentioned in this article are incorporated herein by reference in their entirety.

[0037] The present invention is described in detail below Embodiment 1: This embodiment provides a circulating tumor cell rapid detection model that does not involve disease diagnosis and treatment methods. It detects CTCs based on deep learning. Specifically, it builds an efficient target detection model based on the classic RetinaNet network structure. The construction route of the circulating tumor cell rapid detection model based on deep learning is as follows: Figure 1 As shown, first, the CTCs original data set is obtained, including blood sample preparation, CTCs separation and enrichment, immunofluorescence staining, image acquisition and preprocessing, which is used as the original data for deep learning model training and testing, and the data set is divided; then, a deep learning model is established, and the model is trained, tested and evaluated.

[0038] The specific steps for constructing the circulating tumor cell rapid detection model are as follows: (1) The backbone network uses RepVGG, such as Figure 2 As shown in the figure, a multi-branch structure is used to enhance feature extraction during training, and during reasoning, the idea of ​​structural reparameterization is used to turn it into a single-channel structure, which reduces the memory usage and speeds up the reasoning speed. (2) Based on the traditional feature pyramid FPN structure, a top-down route is added to fuse the semantic information of the bottom layer with the top layer, thereby compensating and enhancing the positioning information. Figure 3 As shown; (3) The detection head is simplified and the classification and regression sub-networks are optimized. Figure 4 As shown in the figure, based on the traditional RetinaNet detection head, the number of serial convolutions and channels is reduced, and large separable convolutions are used to reduce model parameters and computational complexity while increasing the receptive field of the feature map, thereby enhancing the model detection performance. At the same time, the corresponding loss function is optimized; (4) Finally, the loss function is optimized. For bounding box regression, Smooth L1 loss is combined with GIoU loss. In the early training stage, the loss function is dominated by Smooth L1 loss to accelerate model convergence. After the bounding box prediction is stable, GIoU loss can make the positioning more accurate. For object classification, varifocal loss is used. This loss function uses Focal loss to solve the imbalance of positive and negative samples. During the training process, IoU-Aware Class Quartification Score (IACS) is used to simultaneously represent the confidence of object existence and positioning accuracy, so as to produce more accurate detection levels in dense object detectors.

[0039] Obtain the labeled dataset to train the fast detection model through the following steps: Step 1: Isolate and enrich CTCs from the blood of tumor patients, perform immunofluorescence staining, and finally use a fluorescence microscope to collect multi-channel fluorescence images.

[0040] Step 2: Preprocess the fluorescence images to make the various cells easier to distinguish visually. First, a median filter is applied to the fluorescence images to reduce salt and pepper noise; then, bilinear downsampling is used to reduce shot noise; finally, the pixel intensity of each image is normalized to have the same mean and standard deviation.

[0041] Step 3: After the image processing is completed, use LabelImg to manually annotate the image, mark the location and category of each cell, and obtain a labeled dataset in VOC format. Then, divide the obtained CTCs dataset into training set, validation set and test set; Step 4: Use the data set obtained in step 3 to train, test and optimize the target detection model. Finally, deploy the optimized model to the target platform, process new images collected from the blood of tumor patients, and output CTCs detection results.

[0042] This embodiment uses a public data set for model training and testing. The data set is divided into the following two parts: training set and test set. The training set is used for model training, and the test set is used for model evaluation.

[0043] The production process of this dataset: This dataset uses a bile duct cancer cell line, and the specific distribution is as follows Figure 5 As shown; CTCs were then mixed with normal cells at a ratio of 1:1, CTCs were stained with deep red fluorescence (Cytopainter ab 176736), normal cells were stained with green fluorescence (Cytopainter ab 176735), and cell nuclei were stained with blue fluorescence (Hoechst), as shown Figure 6 As shown. Finally, the cells were dropped onto a 96-well plate and imaged using an Opera Phenix instrument (Perkin Elmer). Ultimately, 1,200 fluorescence images of CTCs and normal cells were obtained. Each image consisted of 1,080 × 1,080 pixels and contained an average of 20 to 30 individual cells.

[0044] The above dataset was used for fast detection model training. The model training strategy uses transfer learning and frozen training. Using the pre-trained model on ImageNet, the backbone network was first frozen for 50 epochs on the dataset, and then unfrozen for 250 epochs, i.e., a total of 300 epochs were trained. The batch size (batchsize) was set to 4, the Adam optimizer was used, the learning rate used cosine decay (cos), and the learning rate was set to 3e-4. The results are shown in Figure 2. Figure 7 and Figure 8 As shown in the figure, both the training and validation losses converge quickly, and a higher Map value is achieved on the validation set. This shows that compared with classic models such as Faster RCNN, SSD, and RetinaNet, the fast detection model of the present invention converges significantly faster, and the Map value of the training process is significantly improved.

[0045] As a further optimization of the training mode, additional interference subsets are added to the training set used for model training, including: (1) fluorescence images that have not been preprocessed with median filters and bilinear downsampling, in which various cells are relatively difficult to distinguish visually; (2) the fluorescence images are first preprocessed independently for each channel using Morphology Filter and Denoising to retain the spherical structure and remove background noise, and then pixel intensity normalization is performed; (3) background color removal is first performed, and the darkest central pixel point is found through Kmean or brightness, and the pixel values ​​close to this central pixel point are lowered, so that the very dark area that should be the background is pulled into black, and then the pixel intensity is normalized; each interference subset accounts for 2-10% in the training set. When the proportion of each interference subset in the training set is set to 10%, the same training method is used to obtain the verification set results as shown below. Fig. 9 As shown in the Map curve 2, it can be seen that relative to Figure 8 After adding the interference subset to the training set, the model converges faster and the Map value increases faster during the training process. It can be seen that expanding the training set with new images from the interference subset is conducive to improving the generalization ability and robustness of the model and significantly improving the model performance. In addition, the setting of the weight adjustment mechanism also helps the model to adjust different branch structures calmly when facing input images that have never been seen before, thereby improving the generalization ability of the model.

[0046] In order to verify the detection ability of the fast detection model for CTCs, the performance of the fast detection model was evaluated using the test set, and compared with classic models such as Faster R-CNN, SSD, YOLOv3, RetianNet, and Ours. The evaluation results are shown in Table 1. The Map value of the fast detection model of the present invention reached 90.43% and the F1 score reached 0.882, indicating that the fast detection model of the present invention has better detection accuracy and faster model inference speed, and the comprehensive performance of the model is optimal.

[0047] Table 1 - Comparison of detection performance between the fast detection model and other classic models

[0048] Therefore, this embodiment also provides a rapid detection model for circulating tumor cells obtained by the aforementioned construction method.

[0049] A method for rapid detection of circulating tumor cells for non-disease diagnosis and treatment based on deep learning is also provided, comprising applying the aforementioned rapid detection model for circulating tumor cells to process new images collected from biological samples and outputting CTCs detection results. The biological sample is selected from blood or lymph fluid.

[0050] Also provided is the application of the aforementioned deep learning-based rapid detection method for circulating tumor cells for non-disease diagnosis and treatment in biological sample detection, which does not involve disease diagnosis and treatment. The biological sample is selected from blood or lymph.

[0051] Also provided is the application of the circulating tumor cell rapid detection model obtained by the aforementioned construction method in biological sample detection, which does not involve the diagnosis and treatment of diseases. The biological sample is selected from blood or lymph fluid.

[0052] Embodiment 2: A computer-readable storage medium is also provided, in which a computer program executable by a processor is stored. When the computer program is executed by the processor, at least one step of the specific construction method of the aforementioned circulating tumor cell rapid detection model is executed, and the same technical effect can be achieved. To avoid repetition, this embodiment will not be described in detail.

[0053] Embodiment 3: A computer device, comprising a memory, a processor, a communication interface and a communication bus; wherein the memory, the processor and the communication interface communicate with each other via the communication bus; the memory is used to store computer programs; the processor is used to execute the computer programs stored in the memory, and when the processor executes the computer program, at least one step in the specific construction method of the aforementioned circulating tumor cell rapid detection model is implemented, and the same technical effect can be achieved. To avoid repetition, this embodiment will not be repeated.

[0054] Computer readable media include permanent and non-permanent, removable and non-removable media, and can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PR AM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, read-only compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0055] The conventional techniques in the above embodiments are prior arts known to those skilled in the art, and thus will not be described in detail here.

[0056] The specific embodiments described herein are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

[0057] Although the present invention has been described in detail and some specific embodiments have been cited, it is obvious to those skilled in the art that various changes and modifications can be made without departing from the spirit and scope of the present invention.

[0058] Although the above-mentioned specific embodiments have been shown, described and pointed out the novel features applied to various embodiments, it should be understood that various omissions, substitutions and changes can be made to the form and details of the described device or method without departing from the spirit of the present disclosure. In addition, the above-mentioned various features and methods can be used independently of each other, or can be combined in various ways. All possible combinations and sub-combinations are intended to fall within the scope of the present disclosure. Many of the above-mentioned embodiments include similar components, and therefore, these similar components are interchangeable in different embodiments. Although the present invention has been disclosed in the context of certain embodiments and examples, it should be understood by those skilled in the art that the present invention can extend beyond the specifically disclosed embodiments to other alternative embodiments and / or applications and obvious modifications and equivalents thereof. Therefore, the present invention is not intended to be limited by the specific disclosure of the preferred embodiments herein.

[0059] Matters not covered in the present invention are all known technologies.

Claims

1. A method for constructing a rapid detection model for circulating tumor cells, characterized in that include: (1) Based on the RetinaNet network structure, the backbone network uses RepVGG. During training, a multi-branch structure is used to enhance feature extraction, and during reasoning, the idea of ​​structural reparameterization is used to transform it into a single-path structure; (2) Based on the traditional feature graph pyramid network FPN, a top-down route is added to fuse the semantic information of the bottom layer and the top layer; (3) Simplify the detection head, optimize the classification and regression sub-networks, reduce the number of serial convolutions and channels on the basis of the traditional RetinaNet detection head, and adopt large separable convolutions; (4) For different stages, reasonably optimize the loss functions: Smooth L1 loss, GIoU loss and varifocalloss; obtain a fast detection model.

2. The method for constructing a rapid detection model for circulating tumor cells according to claim 1, characterized in that: The multi-branch structure of step (1) includes at least one of a Conv-BN structure and a BN structure; and In the step (1), a weight adjustment mechanism is set to dynamically adjust the weights of different branch structures.

3. The method for constructing a rapid detection model for circulating tumor cells according to claim 1 or 2, characterized in that: In the step (4), the loss function is optimized reasonably as follows: for bounding box regression, the loss function in the early training stage is dominated by Smooth L1 loss; the loss function after the bounding box prediction is stabilized is dominated by GIoU loss; for target classification, varifocal loss is used.

4. The method for constructing a rapid detection model for circulating tumor cells according to claim 1 or 2, characterized in that: It also includes: training a rapid detection model using a VOC format annotated data set obtained by the following steps: separating and enriching CTCs from the blood of tumor patients, collecting multi-channel fluorescence images using a fluorescence microscope after immunofluorescence staining, preprocessing the fluorescence images to make various cells easier to distinguish visually, manually annotating the images to indicate the location and category of each cell, and obtaining a VOC format annotated data set.

5. The method for constructing a rapid detection model for circulating tumor cells according to claim 4, characterized in that: The preprocessing of the fluorescence image comprises: A median filter was applied to the fluorescence image to reduce salt and pepper noise; Use bilinear downsampling to reduce shot noise; The pixel intensities of each image were normalized to have the same mean and standard deviation.

6. A rapid detection model for circulating tumor cells obtained by the construction method according to any one of claims 1 to 5.

7. Application of the circulating tumor cell rapid detection model obtained by the construction method according to any one of claims 1 to 5 in biological sample detection.

8. A method for rapid detection of circulating tumor cells for non-disease diagnosis and treatment based on deep learning, characterized in that: The method comprises applying the circulating tumor cell rapid detection model described in claim 6 to process new images collected from biological samples and outputting CTCs detection results.

9. Application of the method for rapid detection of circulating tumor cells for non-disease diagnosis and treatment based on deep learning as described in claim 8 in biological sample detection.

10. A computer device, comprising a memory, a processor, a communication interface and a communication bus; wherein: The memory, processor, and communication interface communicate with each other through the communication bus; the memory is used to store computer programs; the processor is used to execute the computer programs stored in the memory, characterized in that: when the processor executes the computer program, it implements at least one step of the method for constructing a rapid detection model of circulating tumor cells according to any one of claims 1 to 5 or the method for rapid detection of circulating tumor cells for non-disease diagnosis and treatment based on deep learning according to claim 8.

Images