WO2025236925A1 - Method and apparatus for detecting circulating tumor cells, device, and medium - Google Patents

Abstract

The present application relates to the technical field of biomedicine, and discloses a method and an apparatus for detecting circulating tumor cells, a device, and a medium. The method comprises: acquiring a target optical image of an object under detection, and using a multi-scale feature map of the target optical image to obtain a region-of-interest feature map; inputting the region-of-interest feature map into a target multi-level mining detection model, wherein the target multi-level mining detection model comprises a plurality of detection heads; and grouping self-attention regions of interest in the region-of-interest feature map into initial positive samples and initial negative samples via a self-attention mechanism module in the target multi-level mining detection model, using the detection heads to screen the initial positive samples and the initial negative samples to obtain a target positive sample containing circulating tumor cells, and determining the target positive sample as a circulating tumor cell detection result for the object under detection. Using the model to perform automated circulating tumor cell detection can improve the efficiency and accuracy of detecting circulating tumor cells and reduce the time required for detection.

Description

Methods, devices, equipment and media for detecting circulating tumor cells Technical Field

This invention relates to the field of biomedical technology, and in particular to methods, devices, equipment and media for detecting circulating tumor cells.

Background Technology

Cancer cells invade the body, potentially in the early stages of tumor development. Metastasis from the primary tumor to distant vital organs is a leading cause of cancer-related death. Circulating tumor cells (CTCs) are cancer cells shed from the primary tumor in the early stages of development and can spread to lymph nodes, bone marrow, or peripheral blood. Numerous studies have shown that CTCs can be used to predict cancer cell metastasis, and even disease progression and survival time in early-stage cancer patients. They can serve as biomarkers for cancer cells in clinical trials, helping to identify therapeutic targets and drug resistance mechanisms.

However, the detection and identification of CTCs is challenging due to their scarcity and heterogeneity; their concentration in peripheral blood is very low, with only 1 to 10 cells per 10 ml in most cancer patients. Current detection methods include immunofluorescence, flow cytometry, and nucleic acid-based methods; however, these methods are time-consuming and subjective, leading to a high false-negative rate.

In summary, improving the efficiency and accuracy of circulating tumor cell detection is a problem that needs to be solved in this field.

Summary of the Invention

In view of this, the purpose of this invention is to provide a method, apparatus, device, and medium for detecting circulating tumor cells, thereby improving the efficiency and accuracy of circulating tumor cell detection. The specific solution is as follows:

In a first aspect, this application discloses a method for detecting circulating tumor cells, comprising:

Acquire a target optical image of the object to be detected, and use the multi-scale feature map of the target optical image to obtain the feature map of the region of interest.

The feature map of the region of interest is input into the multi-level target mining and detection model; wherein, the multi-level target mining and detection model contains multiple detection heads;

The self-attention mechanism module in the target multi-level mining detection model divides each self-attention region of interest in the feature map of the region of interest into initial positive samples and initial negative samples. The detection heads are used to screen out target positive samples containing circulating tumor cells from the initial positive samples and the initial negative samples, and the target positive samples are determined as the circulating tumor cell detection results of the target object.

Optionally, acquiring the target optical image of the object to be detected includes:

A stained digital slide of the object to be tested is acquired, and initial optical images of the digital slide under different shooting conditions are acquired.

Adjust the pixel size of each of the initial optical images to obtain the target optical image.

Optionally, obtaining the region of interest feature map using the multi-scale feature map of the target optical image includes:

Convolutional features of the target optical image are extracted using a backbone network based on a feature pyramid network to obtain a multi-scale feature map;

The multi-scale feature map is input into the region proposal network to obtain the region of interest feature map.

Optionally, before inputting the region of interest feature map into the target multi-level mining detection model, the method further includes:

The detection heads in the initial multi-level mining detection model are trained using the training dataset to obtain the detection heads in the current multi-level mining detection model.

The target multi-level mining detection model is obtained by back-optimizing each detection head of the current multi-level mining detection model based on the minimum loss function.

Optionally, the step of dividing each self-attention region of interest in the feature map of the region of interest into initial positive samples and initial negative samples through the self-attention mechanism module in the multi-level target mining detection model includes:

The self-attention mechanism module of the initial detection head in the target multi-level mining detection model generates each self-attention region of interest in the feature map of the region of interest, and divides each self-attention region of interest into initial positive samples and initial negative samples according to the intersection-union ratio of each self-attention region of interest.

Optionally, the step of generating each self-attention region of interest (ROI) of the ROI feature map through the self-attention mechanism module of the initial detection head in the target multi-level mining detection model includes:

The self-attention mechanism module of the initial detection head in the target multi-level mining detection model performs matrix transformation on the feature map of the region of interest to obtain the target matrix. The target matrix is then transformed into a query vector, a key vector, and a value vector. The SoftMax function is used to convert the product of the query vector and the key vector into a self-attention score. The matrix product between the self-attention score and the value vector is obtained. The matrix product is then convolved to obtain the dimension-restored matrix. The dimension-restored matrix and the feature map of the region of interest are then added point by point to obtain each self-attention region of interest in the feature map of the region of interest.

Optionally, the step of using each of the detection heads to screen for target positive samples containing circulating tumor cells from the initial positive samples and the initial negative samples includes:

Based on a preset positive and negative sample ratio, and using a sampler to select the current initial sample from the initial positive sample and the initial negative sample, the difficult negative samples in the current initial sample are removed to obtain the current target sample. Then, the current target sample is input to the current detection head to obtain the next positive sample selected by the current detection head from the current target sample.

If the current detection head is the last detection head, then the next positive sample is determined as the target positive sample containing circulating tumor cells;

If the current detection head is not the last detection head, then based on the preset positive and negative sample ratio, the sampler is used to select the next initial sample from the next positive sample and the initial negative sample, and the next initial sample is updated to the current initial sample. Then, the process jumps back to the step of removing the difficult-to-distinguish negative samples from the current initial sample.

Secondly, this application discloses a circulating tumor cell detection device, comprising:

The feature map acquisition module is used to acquire the target optical image of the object to be detected, and to acquire the region of interest feature map using the multi-scale feature map of the target optical image.

The feature map input module is used to input the feature map of the region of interest into the target multi-level mining detection model; wherein, the target multi-level mining detection model includes multiple detection heads;

The detection result determination module is used to divide each self-attention region of interest in the feature map of the region of interest into initial positive samples and initial negative samples through the self-attention mechanism module in the multi-level target mining detection model, and to use each of the detection heads to screen out target positive samples containing circulating tumor cells from the initial positive samples and the initial negative samples, and to determine the target positive samples as the circulating tumor cell detection result of the object to be detected.

Thirdly, this application discloses an electronic device, including:

Memory, used to store computer programs;

A processor is configured to execute the computer program to implement the steps of the aforementioned disclosed method for detecting circulating tumor cells.

Fourthly, this application discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the steps of the aforementioned disclosed circulating tumor cell detection method.

The beneficial effects of this application are as follows: This application acquires a target optical image of the object to be detected, and uses the multi-scale feature map of the target optical image to obtain a region of interest feature map; the region of interest feature map is input into a target multi-level mining detection model; wherein, the target multi-level mining detection model contains multiple detection heads; through the self-attention mechanism module in the target multi-level mining detection model, each self-attention region of interest in the region of interest feature map is divided into initial positive samples and initial negative samples, and each of the detection heads is used to screen out target positive samples containing circulating tumor cells from the initial positive samples and the initial negative samples, and the target positive samples are determined as the circulating tumor cell detection result of the object to be detected. Therefore, this application inputs the feature map of the region of interest of the object to be detected into the target multi-level mining detection model to obtain the final target positive sample containing circulating tumor cells, that is, to obtain the circulating tumor cell detection result of the object to be detected. In other words, using the model for automated circulating tumor cell detection can improve the efficiency of circulating tumor cell detection and reduce the detection time. Furthermore, the target multi-level mining detection model contains multiple detection heads, which can improve the classification ability of the target multi-level mining detection model. It can screen out the target positive sample containing circulating tumor cells from the initial positive sample and the initial negative sample that are not accurate enough, thereby improving the accuracy of circulating tumor cell detection.

Attached Figure Description

To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

Figure 1 is a flowchart of a circulating tumor cell detection method disclosed in this application;

Figure 2 is a schematic diagram of a specific self-attention region of interest acquisition disclosed in this application;

Figure 3 is a schematic diagram of a specific multi-level target mining and detection model disclosed in this application;

Figure 4 is a schematic diagram of the evaluation of a specific multi-level mining detection model disclosed in this application;

Figure 5 is a schematic diagram of the structure of a circulating tumor cell detection device disclosed in this application;

Figure 6 is a structural diagram of an electronic device disclosed in this application.

Detailed Implementation

The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

Cancer cells invade the body, potentially in the early stages of tumor development. Metastasis from the primary tumor to distant vital organs is a leading cause of cancer-related deaths. Circulating tumor cells (CTCs) are cancer cells shed from the primary tumor in the early stages of cancer development and can spread to lymph nodes, bone marrow, or peripheral blood. Numerous studies have shown that CTCs can be used to predict cancer cell metastasis, and even disease progression and survival time in early-stage cancer patients. They can serve as biomarkers for cancer cells in clinical trials, identifying therapeutic targets and drug resistance mechanisms.

However, the detection and identification of CTCs is challenging due to their scarcity and heterogeneity; their concentration in peripheral blood is very low, with only 1 to 10 cells per 10 ml in most cancer patients. Current detection methods include immunofluorescence, flow cytometry, and nucleic acid-based methods; however, these methods are time-consuming and subjective, leading to a high false-negative rate.

Therefore, this application provides a circulating tumor cell detection scheme to improve the efficiency and accuracy of circulating tumor cell detection.

Referring to Figure 1, this application discloses a method for detecting circulating tumor cells, including:

Step S11: Acquire the target optical image of the object to be detected, and use the multi-scale feature map of the target optical image to obtain the region of interest feature map.

In this embodiment, acquiring the target optical image of the object to be tested includes: acquiring a stained digital slide of the object to be tested, and acquiring initial optical images of the digital slide under different shooting conditions; adjusting the pixel size of each initial optical image to obtain the target optical image. The object to be tested is peripheral blood from the patient's testing site, which may contain circulating tumor cells. The object to be tested is stained with H&E (Hematoxylin and eosinstains) or methylene blue to obtain a stained digital slide. The digital slide is then photographed under an optical microscope at a fixed magnification under different shooting conditions to obtain initial optical images. For example, the digital slide is horizontally flipped, vertically flipped, and the image brightness is adjusted before photographing to obtain multiple initial optical images. The short side of each initial optical image is then uniformly adjusted to 800 pixels to obtain the target optical image of the object to be tested, thereby reducing the computational load.

In this embodiment, obtaining the region of interest (ROI) feature map using the multi-scale feature map of the target optical image includes: extracting the convolutional features of the target optical image using a backbone network based on a feature pyramid network to obtain a multi-scale feature map; and inputting the multi-scale feature map into a region proposal network to obtain a region of interest (ROI) feature map. The target optical image is input into a backbone network based on a feature pyramid network (FPN), which extracts the convolutional features of the input image to obtain a multi-scale feature map. This multi-scale feature map is then input into a region proposal network (RPN) to obtain the region of interest (ROI) feature map, i.e., ROIs. It is important to note that the ROI feature map includes the ROI and its corresponding alignment.

Step S12: Input the feature map of the region of interest into the target multi-level mining detection model; wherein the target multi-level mining detection model contains multiple detection heads.

In this embodiment, before inputting the region of interest feature map into the target multi-level mining detection model, the method further includes: training each detection head in the initial multi-level mining detection model using a training dataset to obtain each detection head of the current multi-level mining detection model; and performing reverse optimization on each detection head of the current multi-level mining detection model based on the minimum loss function to obtain the target multi-level mining detection model.

An initial multi-level mining detection model with multiple detection heads is constructed. The initial detection head (i.e., the first-level detection head) includes a self-attention mechanism module, while the remaining detection heads do not. All detection heads also include a sampler classifier ht and a localization parameter regressor ft . During the iterative training of the initial multi-level mining detection model using the training dataset, the parameters of each detection head are optimized by minimizing the loss function. The losses of all detection heads are then aggregated to train the model. The specific formula is shown below:
L(x t ,g)=L cls (h t (x t ),y t )+λ[y t ≥1]L loc (f t (x t ,b t ),g);

In the formula, g represents the true value assigned to x ^t , λ＝1 is the weighting coefficient, [·] is the index function, and b ^t is the predicted box coordinates of the parametric regressor f _t-1 (x ^t-1 , b ^t-1 ) of the previous detector or the suggestion box coordinates generated by the RPN network. The type of coordinates used can be determined based on t.

Step S13: The self-attention mechanism module in the target multi-level mining detection model divides each self-attention region of interest in the feature map of the region of interest into initial positive samples and initial negative samples. The detection heads are used to screen out target positive samples containing circulating tumor cells from the initial positive samples and the initial negative samples, and the target positive samples are determined as the circulating tumor cell detection results of the object to be detected.

In this embodiment, the step of dividing the self-attention regions of interest (ROIs) of the region of interest feature map into initial positive samples and initial negative samples through the self-attention mechanism module in the multi-level target mining detection model includes: generating each ROI of the region of interest feature map through the self-attention mechanism module of the initial detection head in the multi-level target mining detection model, and dividing each ROI into initial positive samples and initial negative samples according to the intersection-over-union (IoU) ratio of each ROI. The self-attention mechanism module of the initial detection head generates each ROI of the region of interest feature map and calculates the intersection-over-union (IoU) ratio of each ROI. Based on the IoU ratio, each ROI is divided into initial positive samples and initial negative samples, as detailed below:

In the formula, y is the sample label, g _y is the class label of the ground value g, x is a proposal (i.e., the region of interest for self-attention), u is the IoU threshold, and y ^t is the label x ^t given by the formula for separating positive and negative samples.

If the Intersection over Union (IoU) ratio is less than the threshold u, it is considered an initial negative sample; if the IoU ratio is not less than the threshold u, it is considered an initial positive sample. In other words, predicted boxes with an IoU ratio not less than the threshold are classified as positive samples, and their corresponding labels are positive samples. Boxes with an IoU ratio less than the threshold are classified as negative samples, and their corresponding labels are negative samples. It can be understood that the initial positive samples contain difficult-to-classify positive samples (i.e., incorrectly labeled positive samples) and easy-to-classify positive samples (i.e., correctly labeled positive samples), and the initial negative samples contain difficult-to-classify negative samples (i.e., incorrectly labeled negative samples) and easy-to-classify negative samples (i.e., correctly labeled negative samples).

In this embodiment, the step of generating each self-attention region of interest (ROI) of the ROI feature map through the self-attention mechanism module of the initial detection head in the target multi-level mining detection model includes: performing matrix transformation on the ROI feature map through the self-attention mechanism module of the initial detection head in the target multi-level mining detection model to obtain a target matrix, and converting the target matrix into a query vector, a key vector, and a value vector; using the SoftMax function to convert the product of the query vector and the key vector into a self-attention score; obtaining the matrix product between the self-attention score and the value vector; performing convolution on the matrix product to obtain a dimension-restored matrix; and adding the dimension-restored matrix and the ROI feature map point by point to obtain each self-attention ROI of the ROI feature map. For example, Figure 2 shows a specific schematic diagram of self-attention ROI acquisition. The specific process of self-attention ROI acquisition is as follows:

1) The self-attention mechanism module performs matrix transformation on the feature map of the region of interest to obtain the target matrix. That is, the number and size of the feature map of the region of interest are represented as F _N and BN×C×h×w, where B is the number of images input to the neural network at one time, C is the number of channels, h is the height of the feature, and w is the width of the feature.

2) Transform the target matrix into a query vector, a key vector, and a value vector, as shown in the following formula:
q＝Reshape(f _{Q：conv 1×1} (F _N ));
k=Reshape(f _{K：conv 1×1} (F _N ));
v＝Reshape(f _{V：conv 1×1} (F _N ));

In the formula, q, k, and v represent the query vector, key vector, and value vector, respectively; f(·) represents the convolution operation; and the Reshape(·) function flattens the feature matrix into a vector.

3) Use the SoftMax function (normalization function) to convert the product of the query vector and the key vector into a self-attention score. The specific formula is shown below:
score = δ(q* ^kT );

In the formula, δ(·) represents the SoftMax function, which is a probability distribution that maps the scores to the range [0, 1].

4) Obtain the matrix product between the self-attention score and the value vector. Perform convolution on the matrix product to obtain the dimension-restored matrix. Add the dimension-restored matrix and the region of interest feature map point by point to obtain the self-attention regions of interest in the region of interest feature map. The specific formula is as follows:
F′ _N ＝f _{Re:conv 1×1} (Reshape ^-1 (score*v))+F _N ;

In the formula, _{fRe:conv 1×1} is the convolution operation used to recover the original channel dimension of the feature space, and _F′N is the attention-based region of interest (i.e., AttROIs).

In this embodiment, the step of using each of the detection heads to screen out target positive samples containing circulating tumor cells from the initial positive samples and the initial negative samples includes: based on a preset positive-negative sample ratio, using a sampler to screen out the current initial sample from the initial positive samples and the initial negative samples, removing difficult-to-distinguish negative samples from the current initial sample to obtain the current target sample, and then inputting the current target sample into the current detection head to obtain the next positive sample screened out by the current detection head from the current target sample; if the current detection head is the last detection head, then the next positive sample is determined as the target positive sample containing circulating tumor cells; if the current detection head is not the last detection head, based on the preset positive-negative sample ratio, using the sampler to screen out the next initial sample from the next positive sample and the initial negative samples, updating the next initial sample to the current initial sample, and then jumping back to the step of removing difficult-to-distinguish negative samples from the current initial sample.

For example, Figure 3 shows a schematic diagram of a specific multi-level target mining detection model. The number of initial positive samples is much smaller than the number of initial negative samples. A balanced positive-negative sampler is used to randomly generate positive and negative samples from all proposals at a specified ratio. Based on a preset positive-negative sample ratio, the sampler filters the current initial samples from the initial positive and negative samples. Difficult-to-distinguish positive and negative samples from the initial positive and negative samples are preferentially selected as the current initial samples. Difficult-to-distinguish negative samples are removed from the current initial samples to obtain the current target sample. This means that the detection head in subsequent stages corrects the classification results predicted in previous stages, performing multiple identifications and corrections on difficult-to-distinguish samples to optimize detection performance. In one step, the current target sample is input into the current detection head. The current detection head selects the next positive sample from the current target sample. If the current detection head is not the last detection head, then based on the preset positive and negative sample ratio, the next initial sample is selected from the next positive sample and the initial negative sample using a sampler. The next initial sample is then updated to the current initial sample. Then, the process jumps back to the step of removing the difficult negative samples from the current initial sample. In other words, the difficult sample prediction boxes of the previous stage detection head are preferentially used as the proposal input for the next stage detection head, and positive and negative samples are extracted proportionally by the positive and negative sample sampler. The multi-stage detection heads correspond to different sample distributions and gradually improve the quality of the proposals. The specific formula is as follows:
f(x,b)＝f _T *f _T-1 *…*f _f (x,b);

In the formula, T is the total number of ladder levels, T-1 is the total number of ladder levels minus 1, x is a proposal, and b is the coordinate of the proposal.

In this embodiment, the multi-level target mining detection model includes two modules: a self-attention module for enhancing the ROI and a hard sample mining sampler. In the first stage of training, the self-attention mechanism module is used to improve the performance and accuracy of target detection, while the hard sample sampler is used in non-initial detection heads to identify and correct hard samples multiple times to optimize detection performance.

The beneficial effects of this application are as follows: This application acquires a target optical image of the object to be detected, and uses the multi-scale feature map of the target optical image to obtain a region of interest feature map; the region of interest feature map is input into a target multi-level mining detection model; wherein, the target multi-level mining detection model contains multiple detection heads; through the self-attention mechanism module in the target multi-level mining detection model, each self-attention region of interest in the region of interest feature map is divided into initial positive samples and initial negative samples, and each of the detection heads is used to screen out target positive samples containing circulating tumor cells from the initial positive samples and the initial negative samples, and the target positive samples are determined as the circulating tumor cell detection result of the object to be detected. Therefore, this application inputs the feature map of the region of interest of the object to be detected into the target multi-level mining detection model to obtain the final target positive sample containing circulating tumor cells, that is, to obtain the circulating tumor cell detection result of the object to be detected. In other words, using the model for automated circulating tumor cell detection can improve the efficiency of circulating tumor cell detection and reduce the detection time. Furthermore, the target multi-level mining detection model contains multiple detection heads, which can improve the classification ability of the target multi-level mining detection model. It can screen out the target positive sample containing circulating tumor cells from the initial positive sample and the initial negative sample that are not accurate enough, thereby improving the accuracy of circulating tumor cell detection.

For example, Figure 4 shows a specific evaluation diagram of a multi-level mining detection model. In order to quantitatively evaluate the detection performance, the accuracy of the multi-level mining detection model is evaluated, including average accuracy, receiver operating characteristic curve (ROC) analysis, and ablation experiments, as detailed below:

1) Evaluating the performance of multi-level cell detection models using average precision: Considering the need to reduce false positives and balance cell recall in cell detection tasks, when evaluating model performance, bounding boxes with a predicted probability greater than the threshold τ = 0.6 can be retained as the final prediction result. Average precision (AP) can simultaneously reflect the recall and detection precision for each class, and is defined as follows:

In the formula, k represents the number of CTC cells or CTC-like cells detected, P(k) represents the precision up to the kth element in the list, Δr(k) represents the recall from the (k-1)th to the kth item, TP, FP, and FN represent the number of predicted boxes for true positives, false positives, and false negatives, respectively. For multi-class detection tasks, the mAP value is also used as an indicator, which is the average AP value for all classes.

2) Utilize the free-response operating characteristic curve (FROC) to evaluate the performance of the multi-level mining detection model: In the ROC curve, the vertical axis plots sensitivity and the horizontal axis plots 1-specificity. The FROC curve replaces 1-specificity with the number of false positives per image, which can reflect the model's ability to detect within the allowed number of false positives.

3) Evaluation of the performance of the multi-level mining detection model using ablation experiments: Based on the Swin-Transformer backbone network, ablation experiments were used to evaluate the impact of hyperparameters in the multi-level mining detection model. First, the Cascade R-CNN network was set as the baseline method. The proposed self-attention module and the hard sample mining sampler were added to the baseline model, and the model combining these two modules was evaluated simultaneously. Further, ablation experiments were conducted using hyperparameter settings, i.e., setting multiple IoU thresholds. In the original Cascade R-CNN, higher-quality positive samples were selected by progressively increasing the IoU thresholds at each stage of the training samples, for example, using the following three thresholds: {0.5, 0.5, 0.5}, {0.5, 0.6, 0.7}, and {0.6, 0.7, 0.8}.

The accuracy and reliability of the multi-level target mining detection model can be verified by evaluating the average accuracy, free acceptor operation characteristic curve analysis, and ablation experiments.

Referring to Figure 5, this application discloses a circulating tumor cell detection device, comprising:

Feature map acquisition module 11 is used to acquire the target optical image of the object to be detected, and to acquire the region of interest feature map using the multi-scale feature map of the target optical image;

The feature map input module 12 is used to input the feature map of the region of interest into the target multi-level mining detection model; wherein, the target multi-level mining detection model includes multiple detection heads;

The detection result determination module 13 is used to divide each self-attention region of interest in the feature map of the region of interest into initial positive samples and initial negative samples through the self-attention mechanism module in the multi-level target mining detection model, and to use each of the detection heads to screen out target positive samples containing circulating tumor cells from the initial positive samples and the initial negative samples, and to determine the target positive samples as the circulating tumor cell detection result of the object to be detected.

The beneficial effects of this application are as follows: This application acquires a target optical image of the object to be detected, and uses the multi-scale feature map of the target optical image to obtain a region of interest feature map; the region of interest feature map is input into a target multi-level mining detection model; wherein, the target multi-level mining detection model contains multiple detection heads; through the self-attention mechanism module in the target multi-level mining detection model, each self-attention region of interest in the region of interest feature map is divided into initial positive samples and initial negative samples, and each of the detection heads is used to screen out target positive samples containing circulating tumor cells from the initial positive samples and the initial negative samples, and the target positive samples are determined as the circulating tumor cell detection result of the object to be detected. Therefore, this application inputs the feature map of the region of interest of the object to be detected into the target multi-level mining detection model to obtain the final target positive sample containing circulating tumor cells, that is, to obtain the circulating tumor cell detection result of the object to be detected. In other words, using the model for automated circulating tumor cell detection can improve the efficiency of circulating tumor cell detection and reduce the detection time. Furthermore, the target multi-level mining detection model contains multiple detection heads, which can improve the classification ability of the target multi-level mining detection model. It can screen out the target positive sample containing circulating tumor cells from the initial positive sample and the initial negative sample that are not accurate enough, thereby improving the accuracy of circulating tumor cell detection.

Furthermore, embodiments of this application also provide an electronic device. Figure 6 is a structural diagram of an electronic device 20 according to an exemplary embodiment; the content in the figure should not be construed as limiting the scope of this application.

Figure 6 is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Specifically, it may include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input/output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the circulating tumor cell detection method performed by the electronic device disclosed in any of the foregoing embodiments.

In this embodiment, the power supply 23 is used to provide operating voltage for various hardware devices on the electronic device; the communication interface 24 can create a data transmission channel between the electronic device and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input/output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

The processor 21 may include one or more processing cores, such as a quad-core processor or an octa-core processor. The processor 21 may be implemented using at least one hardware form selected from DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). The processor 21 may also include a main processor and a coprocessor. The main processor, also known as a CPU (Central Processing Unit), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, the processor 21 may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the screen. In some embodiments, the processor 21 may also include an AI (Artificial Intelligence) processor, which is used to handle computational operations related to machine learning.

In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored on it include operating system 221, computer program 222 and data 223, etc., and the storage method can be temporary storage or permanent storage.

The operating system 221 manages and controls the various hardware devices and computer programs 222 on the electronic device to enable the processor 21 to perform calculations and processing on the massive amounts of data 223 in the memory 22. The operating system can be Windows, Unix, Linux, etc. The computer program 222, in addition to including a computer program capable of performing the circulating tumor cell detection method executed by the electronic device as disclosed in any of the foregoing embodiments, may further include computer programs capable of performing other specific tasks. The data 223 may include data received by the electronic device from external devices, as well as data collected by its own input/output interface 25.

Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned method for detecting circulating tumor cells. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application. The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly in hardware, software modules executed by a processor, or a combination of both. The software module may be located in random access memory (RAM), memory, read-only memory (ROM), electrically programmable EPROM (EPROM), electrically erasable programmable read-only memory (EEPROM), register, hard disk, removable disk, CD-ROM (Compact Disc Read-Only Memory), or any other form of storage medium known in the art.

Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

The foregoing has provided a detailed description of the circulating tumor cell detection method, apparatus, equipment, and medium provided by the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims (10)

A method for detecting circulating tumor cells, characterized in that it includes: Acquire a target optical image of the object to be detected, and use the multi-scale feature map of the target optical image to obtain the feature map of the region of interest. The feature map of the region of interest is input into the multi-level target mining and detection model; wherein, the multi-level target mining and detection model contains multiple detection heads; The self-attention mechanism module in the target multi-level mining detection model divides each self-attention region of interest in the feature map of the region of interest into initial positive samples and initial negative samples. The detection heads are used to screen out target positive samples containing circulating tumor cells from the initial positive samples and the initial negative samples, and the target positive samples are determined as the circulating tumor cell detection results of the target object. The method for detecting circulating tumor cells according to claim 1, characterized in that acquiring the target optical image of the object to be detected includes: A stained digital slide of the object to be tested is acquired, and initial optical images of the digital slide under different shooting conditions are acquired. Adjust the pixel size of each of the initial optical images to obtain the target optical image. The method for detecting circulating tumor cells according to claim 1, characterized in that, the step of obtaining the region of interest feature map using the multi-scale feature map of the target optical image includes: Convolutional features of the target optical image are extracted using a backbone network based on a feature pyramid network to obtain a multi-scale feature map; The multi-scale feature map is input into the region proposal network to obtain the region of interest feature map. The method for detecting circulating tumor cells according to claim 1, characterized in that, before inputting the feature map of the region of interest into the target multi-level mining detection model, it further includes: The detection heads in the initial multi-level mining detection model are trained using the training dataset to obtain the detection heads in the current multi-level mining detection model. The target multi-level mining detection model is obtained by back-optimizing each detection head of the current multi-level mining detection model based on the minimum loss function. The method for detecting circulating tumor cells according to claim 1, characterized in that, the step of dividing each self-attention region of interest in the feature map of the region of interest into initial positive samples and initial negative samples by the self-attention mechanism module in the multi-level target mining detection model includes: The self-attention mechanism module of the initial detection head in the target multi-level mining detection model generates each self-attention region of interest in the feature map of the region of interest, and divides each self-attention region of interest into initial positive samples and initial negative samples according to the intersection-union ratio of each self-attention region of interest. The circulating tumor cell detection method according to claim 5, characterized in that, the step of generating each self-attention region of interest of the region of interest feature map through the self-attention mechanism module of the initial detection head in the target multi-level mining detection model includes: The self-attention mechanism module of the initial detection head in the target multi-level mining detection model performs matrix transformation on the feature map of the region of interest to obtain the target matrix. The target matrix is then transformed into a query vector, a key vector, and a value vector. The SoftMax function is used to convert the product of the query vector and the key vector into a self-attention score. The matrix product between the self-attention score and the value vector is obtained. The matrix product is then convolved to obtain the dimension-restored matrix. The dimension-restored matrix and the feature map of the region of interest are then added point by point to obtain each self-attention region of interest in the feature map of the region of interest. The method for detecting circulating tumor cells according to claim 6, characterized in that, the step of using each of the detection heads to screen out target positive samples containing circulating tumor cells from the initial positive samples and the initial negative samples includes: Based on a preset positive and negative sample ratio, and using a sampler to select the current initial sample from the initial positive sample and the initial negative sample, the difficult negative samples in the current initial sample are removed to obtain the current target sample. Then, the current target sample is input to the current detection head to obtain the next positive sample selected by the current detection head from the current target sample. If the current detection head is the last detection head, then the next positive sample is determined as the target positive sample containing circulating tumor cells; If the current detection head is not the last detection head, then based on the preset positive and negative sample ratio, the sampler is used to select the next initial sample from the next positive sample and the initial negative sample, and the next initial sample is updated to the current initial sample. Then, the process jumps back to the step of removing the difficult-to-distinguish negative samples from the current initial sample. A circulating tumor cell detection device, characterized in that it comprises: The feature map acquisition module is used to acquire the target optical image of the object to be detected, and to acquire the region of interest feature map using the multi-scale feature map of the target optical image. The feature map input module is used to input the feature map of the region of interest into the target multi-level mining detection model; wherein, the target multi-level mining detection model includes multiple detection heads; The detection result determination module is used to divide each self-attention region of interest in the feature map of the region of interest into initial positive samples and initial negative samples through the self-attention mechanism module in the multi-level target mining detection model, and to use each of the detection heads to screen out target positive samples containing circulating tumor cells from the initial positive samples and the initial negative samples, and to determine the target positive samples as the circulating tumor cell detection result of the object to be detected. An electronic device, characterized in that it comprises: Memory, used to store computer programs; A processor for executing the computer program to implement the steps of the circulating tumor cell detection method as described in any one of claims 1 to 7. A computer-readable storage medium is characterized in that it is used to store a computer program; wherein, when the computer program is executed by a processor, it implements the steps of the circulating tumor cell detection method as described in any one of claims 1 to 7.