Microfluidic nanochip and ai analysis-based circulating tumor cell automatic detection device and application

Abstract

The application belongs to the technical field of biological intelligent detection, and particularly relates to a circulating tumor cell automatic detection device based on a microfluidic nanochip and AI analysis and application; the device comprises a microfluidic chip module, an optical acquisition module, an image recognition module, a feedback control module and a liquid path control module, and can realize efficient enrichment of circulating tumor cells in peripheral blood, image recognition and dynamic flow control adjustment. By constructing a change trend model of an image frame sequence, a cell state change is recognized in real time and a flow rate and a channel control instruction are output, so that the recognition accuracy and the sample utilization rate are synergistically optimized. The device is suitable for tumor early screening, efficacy monitoring and recurrence early warning scenes.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of biological intelligent detection, and particularly relates to a circulating tumor cell automatic detection device based on a microfluidic nanochip and AI analysis and application. BACKGROUND

[0002] Under the background that cancer early screening and individualized treatment are increasingly valued, circulating tumor cells (CTCs) as rare markers in blood have high diagnostic, prognostic evaluation and efficacy monitoring value. Traditional CTC detection methods mainly rely on immunomagnetic bead capture, flow cytometry or immunofluorescence staining and other means, and have problems of low sensitivity, complicated operation, limited throughput and large subjective interpretation deviation, and are difficult to meet the needs of high-throughput, automation and standardization of clinical application, especially in the CTC phenotype with obvious diversity and heterogeneity, and the traditional technology is more difficult to balance the accuracy of multiple recognition.

[0003] As a new generation of biological analysis platform, the microfluidic nanochip has been widely used in cell sorting and analysis scenarios due to its precise control of liquid flow and high integrability of structure and function. By constructing multi-channel, gradient structure, surface functionalization or mesoscopic physical structure on the chip, the enrichment of CTCs in size, charge and adhesion characteristics can be realized. In particular, combined with nanostructure modification technology, the cell recognition efficiency and capture specificity can be significantly improved. However, the images and multi-modal signals output by the chip still need manual analysis or simple software processing, and a truly closed-loop intelligent detection system has not been formed.

[0004] Artificial intelligence technology has shown strong advantages in medical image analysis and pattern recognition, therefore, constructing an automatic CTC detection device integrating microfluidic nanochip technology and AI intelligent analysis algorithm has become an important development direction in the current tumor liquid biopsy field, which can realize the full-process automation from sample loading, cell capture, image acquisition to intelligent recognition, and provide a more efficient and reliable solution for early diagnosis and dynamic monitoring of cancer. SUMMARY

[0005] In view of the above problems, the purpose of the present application is to provide a circulating tumor cell automatic detection device based on a microfluidic nanochip and AI analysis, comprising:

[0006] A microfluidic chip module, comprising a transparent cover plate, an intermediate structure layer and a substrate support layer, a plurality of independent sample channels, enrichment regions and waste liquid outlets are arranged in the intermediate structure layer, and the inner surface of the enrichment region is modified with circulating tumor cell specific recognition molecules;

[0007] An optical acquisition module, comprising an excitation light source, a light filter assembly and an image sensor, arranged below the microfluidic chip module, for acquiring multi-spectral image information in the enrichment region;

[0008] An image recognition module, including an image preprocessing unit, a feature extraction network and a recognition inference module, is used to extract cell morphology parameters and fluorescence distribution parameters from images and output recognition results;

[0009] A feedback control module is used to generate adjustment signals based on image recognition results, including flow rate adjustment, channel switching and residence time control instructions;

[0010] A liquid path control module, including a micro-flow pump, a pressure control unit and an electromagnetic valve group, is used to receive feedback instructions and adjust the flow behavior of samples in the chip.

[0011] As a preferred technical solution, the plurality of sample channels of the microfluidic chip module are at least 4 structurally symmetric microchannels, and each inlet is provided with an automatic switching valve, and each channel is connected to an independent enrichment area;

[0012] The channel width is 50-100 μm, the depth is controlled within the range of 30-80 μm, and a curved flow guide cavity is provided at the intersection of the channels to slow down the impact of cells and improve the adsorption efficiency of target cells.

[0013] As a preferred technical solution, the enrichment area is internally provided with a metal nanostructure array, the array includes a periodically arranged gold nanorod, silver nanosphere or carbon nanotaper array, the structure spacing is 100-200 nm, and the coverage area is more than 80% of the entire enrichment area.

[0014] As a preferred technical solution, the image preprocessing unit in the image recognition module includes image denoising, edge sharpening and signal normalization processing procedures, the feature extraction network adopts a convolutional neural network structure, extracts parameters of cell size, fluorescence signal intensity, nuclear-cytoplasmic ratio and background gray difference dimensions in the target area, and is used to realize high-confidence identification of CTC and non-target cells.

[0015] As a preferred technical solution, the control strategy generation unit in the feedback control module constructs a dynamic trend identification based on image frame sequences, analyzes the change trend of target cells in the spatial position, morphological characteristics and fluorescence distribution intensity dimensions in the continuous image frames, judges the image state, and generates corresponding feedback control instructions according to the judgment result, and the specific control strategies include:

[0016] T1, image mutation response control:

[0017] When the centroid offset of the target cell in the continuous frame is greater than 10 μm, or the edge area change rate exceeds 20%, or the average gray value change amplitude of the main fluorescence channel exceeds 15%, it is identified as an image mutation state; the control strategy generation unit outputs a flow rate reduction instruction, reduces the current flow rate to 50-70% of the original rate, and sends a hysteresis holding instruction to extend the residence time of the current sample in the enrichment region by not less than 1.0 s, to restore the image stable state;

[0018] T2, stable state switching control:

[0019] When the average position offset of the target cell in the continuous 3 frames of images is less than 5 μm, the boundary overlap degree is higher than 80%, the fluorescence gray fluctuation rate is lower than 5%, and the recognition confidence variation amplitude is less than 5%, it is determined that the image recognition is in a stable state; the control strategy generation unit outputs a channel switching preparation signal, and resets the flow rate to the default value to advance the next sample segment loading;

[0020] T3, high-density cell recognition adjustment:

[0021] When the number of target cells identified in a single frame exceeds the set upper threshold value (such as the CTC count in each frame of image exceeds 10), and the edge overlap degree between cells exceeds 60%, it is determined that the high-density accumulation state, the control strategy generation unit outputs a periodic intermittent flow control signal, and pauses the flow for 1.0 s every 5 frames, for relieving cell accumulation and overlapping interference, and introducing a refocusing scanning instruction to optimize the image imaging area;

[0022] T4, background noise enhancement suppression:

[0023] When the background gray mean value of the continuous frame image rises by more than 20%, or the standard deviation rises by more than 25%, and the target cell edge detection failure rate exceeds 30%, it is determined that the background interference is enhanced; the control strategy generation unit triggers the multi-band fluorescence channel switching strategy, and synchronously adjusts the excitation light source power to 120% of the current value, returns to the original power after 3 frames, and prolongs the single frame exposure time by not less than 100 ms, for improving the signal contrast;

[0024] T5, weak signal target compensation recognition:

[0025] When the fluorescence signal peak value of the target cell marked by the image recognition module is lower than the gray threshold value 50, and the target edge recognition confidence is between 60-80%, and appears continuously for 3 frames, it is determined that it is a low expression weak signal unit; the control strategy generation unit marks the region as "increased sampling object", executes sub-region flow rate pause 0.8-1.2 s, and triggers image resampling process and model confidence correction mechanism;

[0026] T6, continuous misjudgment correction and frame compensation control:

[0027] When the model identifies a pseudo-target cell with incomplete structure or mismatched morphology in 5 consecutive image frames, and the center is inconsistent with the main recognition channel, accompanied by frequent switching of the recognition category more than 3 times, the false recognition correction mechanism is triggered; the control strategy generation unit records the image sequence number, feeds back the "false recognition flag" to the image recognition module, and pauses the current channel flow for 2.0 s, loads the cached image of the latest valid frame, performs inter-frame interpolation correction and reconstructs the recognition input sequence.

[0028] The control strategy generation unit establishes a real-time instruction channel with the liquid path control module through a high-speed bus interface, and the feedback response period of the updated image state vector is controlled within 50 ms, ensuring the real-time closed-loop control capability of recognition-adjustment-repair.

[0029] As a preferred technical solution, the micro flow pump in the liquid path control module supports adjustment in the range of 0.5-10.0 muL / min, with a minimum resolution of 0.1 muL / min, and the opening and closing response time of the electromagnetic valve is not more than 50 ms. The pressure control unit is internally provided with a silicon-based tension balance diaphragm for maintaining the pressure difference between the enrichment areas constant at 50-120 Pa.

[0030] The application also provides an application of the detection device in tumor liquid biopsy, comprising:

[0031] S1, sample loading and enrichment: inject the human peripheral blood sample after red blood cell lysis and filtration into the microfluidic chip, and guide it into the enrichment area through the liquid path control system, wherein the area is provided with surface modified recognition molecules and nano enhanced structures for selectively adsorbing circulating tumor cells and enhancing fluorescence response;

[0032] S2, image acquisition and recognition: the image acquisition module acquires continuous bright field and fluorescence images, and the recognition module extracts cell morphology and signal characteristics, and outputs target cell classification results, position information and recognition confidence;

[0033] S3, feedback control adjustment: the feedback control module judges the cell mutation or stable trend according to the image state, and outputs flow rate adjustment, residence time adjustment or channel switching instructions to drive the liquid path module to dynamically control the sample flow parameters;

[0034] S4, result output and uploading: the recognition results and control trajectory records are stored and a detection report is generated, which is uploaded to the diagnosis platform through the interface for assisting in analyzing tumor evolution and treatment response.

[0035] As a preferred technical solution, in step S2, the image feature vector extracted by the image recognition module includes the following five types of parameters:

[0036] Edge size of target cell, calculated as the average of long and short axis length, unit: μm;

[0037] Nuclear-cytoplasmic ratio, calculated as the ratio of the gray value of the center region of the main fluorescence channel to the gray value of the cell outline;

[0038] Mean gray value and standard deviation of fluorescence channel, used to represent signal intensity and uniformity, respectively;

[0039] Edge sharpness score, calculated based on the variation rate of local gradient histogram;

[0040] Mean gray value and standard deviation of background noise, used to estimate signal-to-noise ratio;

[0041] The five types of feature parameters are merged into a unified state vector after being extracted by a convolutional network, and are input into a feedback control module together with the output confidence score of the recognition inference model, for real-time control of the generation of adjustment strategies.

[0042] As a preferred technical solution, in step S4, the recognized CTC cell image, classification label, time stamp, and recognition confidence data are stored in a local database, a multi-dimensional detection report is generated, and is uploaded to a cloud information system through a graphical user interface or a remote interface of the detection device.

[0043] As a preferred technical solution, in step S4, the multi-dimensional detection report contains the following information fields:

[0044] Target cell recognition time stamp, recorded as image frame number and system real-time time;

[0045] Cell position coordinates and classification label, generated based on the image recognition center point and model classification result;

[0046] Fluorescence channel gray value distribution map, drawing the signal intensity curve of each channel and the background interference analysis result;

[0047] Sample channel number and flow rate trajectory record, storing the flow rate change of each channel in time sequence;

[0048] Recognition state curve, reflecting the change trend of recognition confidence over time in the whole process;

[0049] The above information is output in PDF report and structured JSON data dual format.

[0050] The beneficial effects achieved by the present application are:

[0051] The application constructs a multi-channel symmetric structure inside a microfluidic chip and embeds a metal nano array, so that circulating tumor cells can still maintain high recognition resolution under the condition of multi-sample parallel enrichment, effectively improve the imaging intensity and edge definition of weak fluorescent signals, and solve the problem of recognition ability decline of existing single-channel CTC enrichment equipment in high-throughput detection scenarios.

[0052] The application first introduces an image feature driven dynamic feedback control module in the detection device, converts the indexes such as cell displacement, morphological change and fluorescence fluctuation extracted from the image frame sequence into real-time flow parameter control signals, realizes the bidirectional closed-loop adjustment of the recognition state and the sample flow state, and significantly improves the confidence and imaging stability of the target cell recognition.

[0053] The application records and generates a multi-dimensional detection report with time stamp index for the recognition output and control response process, combines the recognition behavior trajectory and the flow rate control path, can realize the traceability and optimization of the detection process, and provides a reliable cell behavior analysis basis for the clinic, has real-time auxiliary diagnosis value in multiple scenes such as postoperative recurrence monitoring and immune response evaluation. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 It is a schematic diagram of the system structure of the application;

[0055] Figure 2 It is a schematic diagram of the method flow of the application. DETAILED DESCRIPTION

[0056] In order to deepen the understanding of the application, the application will be further described in combination with examples, and the examples are only used to explain the application and do not constitute a limitation on the protection scope of the application.

[0057] Example one: circulating tumor cell automatic detection device based on microfluidic nanochip and AI analysis.

[0058] As shown in Figure 1 The application provides a circulating tumor cell automatic detection device based on microfluidic nanochip and AI analysis, which is suitable for high-throughput and high-precision automatic enrichment and recognition of circulating tumor cells (CTCs) in peripheral blood samples in liquid biopsy scenarios. The device combines microstructure design, nano-enhanced structure, image recognition algorithm and feedback control mechanism, realizes intelligent linkage of the whole process from sample guidance, cell capture, image processing to fluid dynamic control.

[0059] The device mainly includes the following five core functional modules: microfluidic chip module, optical acquisition module, image recognition module, feedback control module and liquid path control module.

[0060] Each of the above modules is deployed in a functional hierarchical manner inside the device framework, and is physically coupled and data-linked through a high-speed digital communication bus and an optical interface.

[0061] The microfluidic chip module is the core sample processing component of the device, and has a three-layer composite structure, including a transparent cover plate, an intermediate structural layer, and a base support layer from top to bottom.

[0062] Intermediate structural layer design: This layer is molded using PDMS material, and has four symmetrically independent microchannels inside. The channel width is 50-100 pm, the depth is 30-80 pm, and the length is about 25 mm. The inlet of each channel is provided with an automatic switching microvalve, which can be connected to a multi-channel sampling system for parallel sample processing.

[0063] Enrichment area structure: Each channel has an enrichment cavity in the middle section, which is widened to 150 pm and has a length of 3 mm. The surface of the cavity is treated by plasma and then introduced with a polyvinyl alcohol support, and grafted with tumor-related recognition antibodies such as EpCAM, Vimentin, and CD45 to form a CTC-specific binding layer.

[0064] The area is provided below with a metal nano-enhanced structure, including a gold nanorod array (diameter 80 nm, spacing 180 nm), which has an effective coverage area of more than 80%, and can enhance the binding capacity of target cells and improve the fluorescence signal response.

[0065] Flow guide cavity design: To prevent cells from being impacted into the enrichment area and causing adhesion failure, a curvature transition section is designed at the interface between the channel and the enrichment cavity to form a curved flow guide cavity, which effectively slows down and uniformly distributes the liquid flow.

[0066] The optical acquisition module is vertically arranged below the microfluidic chip and is optically aligned with the enrichment area. The module includes the following units:

[0067] Excitation light source: Integrates a high-power LED array and a multi-band filter, supports green (FITC), red (TRITC), and ultraviolet (DAPI) excitation wavelength switching, has automatic intensity adjustment capability, and the maximum power is 150 mW, with a minimum adjustable unit of 2 mW.

[0068] Filter assembly: Three-channel filter units are connected in parallel, and the filter is quickly switched through an electrically controlled rotating platform to accurately capture the corresponding fluorescence signal.

[0069] Image sensor: A cooled sCMOS camera is used, with a resolution of 2048x2048, a frame rate of up to 60 fps, and a signal-to-noise ratio of more than 45 dB, with high sensitivity and low noise image acquisition capability.

[0070] The image signal is transmitted in real time to the image recognition module through the USB3.0 interface for processing.

[0071] The image recognition module is constructed based on a deep neural network structure and specifically consists of the following three sub-modules:

[0072] The image preprocessing unit performs background fitting removal, artifact suppression and edge enhancement processing on the input image, and filters out non-biological signal areas based on local gradient variation analysis.

[0073] The feature extraction network adopts a multi-scale convolution structure, extracts cell size (μm), nucleus-to-cytoplasm ratio (gray distribution ratio), edge sharpness (local gradient amplitude), fluorescence intensity distribution (peak value and mean value), signal-to-noise ratio, etc. after inputting a single frame of image, and outputs a high-dimensional state vector.

[0074] The recognition inference module is modeled based on ResNet-50 backbone and attention mechanism, fuses three consecutive images and performs discriminative inference, and outputs recognition categories (CTC / white blood cells / non-cells), center coordinate positions and recognition confidence scores (between 0 and 1).

[0075] The module can support GPU accelerated inference, and the average single frame recognition processing time delay is less than 20 ms.

[0076] The feedback control module is responsible for generating flow adjustment strategies according to the image recognition results to form a real-time closed-loop feedback control. It contains a strategy mapping unit, state determination logic and control signal encoder:

[0077] Image mutation recognition (T1): when the centroid offset in the image state is >10 μm or the gray scale change is >15%, the flow rate is reduced by 50-70%, and the residence time is extended to 1.0-1.5 s;

[0078] Image stable state (T2): when the recognition confidence of three consecutive frames is >95% and the boundary overlap is >85%, it is determined to be stable, the default flow rate is restored, and the channel switching preparation is started;

[0079] Low signal enhancement strategy (T3): when the recognition cell fluorescence peak is lower than the threshold, the light source power is increased to 130%, and the exposure time is increased by 20 ms;

[0080] High-density recognition state (T4): when the number of CTCs exceeds 8 per frame, the flow is paused for 1 s, and image refocusing and slow flow scanning are performed.

[0081] All control instructions are transmitted to the liquid control module through the high-speed CAN bus, and the response time is not more than 50 ms.

[0082] The liquid control module serves as an execution unit and specifically includes:

[0083] Micro-pump unit: Stepping micro-pump, supporting flow rate adjustment range of 0.5-10.0 μL / min, adjustment accuracy of 0.1 μL / min, and driving signal sent by feedback control module.

[0084] Pressure control unit: Adopting silicon membrane tension balancing structure design, adjusting pressure difference between front and back of channel to be between 50-120 Pa, monitoring pressure difference change through bidirectional detection bridge system and feeding back closed-loop adjustment.

[0085] Electromagnetic valve group: Used for sample channel selection switching, response time less than 50 ms, having self-locking and state feedback capabilities, supporting high-speed switching and one-way control between channels.

[0086] The device continuously runs for 4 hours under the condition of simulating blood samples, and the average time required for each detection is not more than 15 minutes, the CTC recognition accuracy is stably maintained at more than 92%, the recognition frame proportion with a confidence score greater than 0.95 is more than 88%. In the diluted blood sample mixed with 100-1000 CTC / ml, the detection recovery rate is as high as 85% or more, and the misjudgment rate is less than 5%.

[0087] Compared with the traditional "image recognition + static flow control" device, the present application realizes dynamic response control driven by image state, and has significant advantages in heterogeneity CTC recognition accuracy, sample throughput utilization rate and system adaptive adjustment capability, and is suitable for efficient recognition of rare tumor marker cells in clinical liquid biopsy.

[0088] Embodiment two: Application method of circulating tumor cell automatic detection device based on microfluidic nanochip and AI analysis in tumor liquid biopsy.

[0089] As shown in Figure 2 The embodiment provides an application method of a circulating tumor cell automatic detection device based on a microfluidic nanochip and AI analysis in tumor liquid biopsy.

[0090] The method takes human peripheral blood samples as detection objects, combines a high-throughput microfluidic sample enrichment platform, an image recognition and feedback control algorithm module, realizes automatic processing of the whole process from sample loading, image recognition to dynamic fluid regulation and detection result output, and is suitable for clinical scenes such as early tumor screening, postoperative recurrence monitoring and immune therapy response evaluation.

[0091] The method mainly includes the following steps:

[0092] S1, sample loading and enrichment:

[0093] Firstly, the collected human peripheral blood samples were pretreated. Red blood cells were removed by using red blood cell lysis solution (such as ACK buffer), and most of the debris and free nucleic acid components were removed by low-speed centrifugation, followed by 0.45 μm microporous membrane filtration to exclude large size background interference.

[0094] The processed plasma cell mixture was injected into the sample injection port of the microfluidic chip by an automatic sampler.

[0095] The microfluidic chip used in this embodiment is a four-channel symmetric structure, each channel has a width of 80 μm and a depth of 60 μm, and the enrichment area is about 3 mm long. The middle part of the chip is the enrichment area, which is provided with a gold nanorod array and a silver nanosphere dot array, and the arrangement period is 150 nm, covering more than 85% of the total surface of the enrichment area.

[0096] At the same time, the surface of the enrichment area is covalently grafted with specific antibodies such as anti-EpCAM and anti-Vimentin to achieve selective capture of heterogeneous CTCs.

[0097] The sample flow is adjusted by a microfluidic pump in the liquid path control system, and the initial flow rate is set to 2.0 μL / min, which maintains stable flow within a pressure difference of 50-70 Pa, ensuring effective contact and binding between cells and the surface recognition layer.

[0098] S2, image acquisition and recognition:

[0099] During the enrichment process, the image acquisition module continuously acquires bright field images and multiple fluorescence channel images (such as DAPI, FITC, TRITC) in the enrichment area. The image acquisition frame rate is 30 fps, and the image size is 2048×2048 pixels. The image data is transmitted in real time to the image recognition module.

[0100] The image recognition module includes an image preprocessing network, a feature extraction network, and a recognition inference module. To ensure the accuracy of weak signal cell recognition, the image preprocessing network first performs background fitting removal, edge enhancement, and artifact suppression operations; then enters the multi-scale convolution network for feature extraction.

[0101] In this embodiment, the image recognition module extracts the following five types of image features to form a state vector:

[0102] Target cell edge size: the average length of the major and minor axes is calculated using elliptical fitting boundary, with the unit of μm, reflecting the actual outline of the cell;

[0103] Nucleus-cytoplasm ratio: calculated by the average gray value of the central nucleus region in the main fluorescence channel and the average gray value of the overall cell outline region, representing the change in nucleus proportion;

[0104] Mean and standard deviation of fluorescence channel: Statistics the concentration and uniformity of gray value in different channels, representing the expression strength and distribution uniformity of the marker;

[0105] Edge sharpness score: Based on local gradient histogram, the variation rate of target edge pixels is analyzed to quantify the boundary quality.

[0106] Mean of background gray value and standard deviation of background noise: used to estimate the background signal-to-noise ratio, to assist in determining whether the target is disturbed or masked.

[0107] The above five types of features are uniformly encoded to form a state vector, which is jointly input into the recognition inference network, and the semantic features extracted by the deep convolutional backbone network are fused, and the target cell class (CTC / white blood cell / pseudo-image), spatial coordinate position (x, y) and recognition confidence (between 0 and 1) are obtained through the full connection output layer.

[0108] S3, feedback control adjustment:

[0109] The recognition module outputs a frame of recognition result, and the feedback control module judges its state and generates a strategy. In this embodiment, the control strategy is classified and responded according to the image change trend:

[0110] When the centroid offset of the target cell in the continuous frames is greater than 10 μm, the edge area change rate is more than 20%, or the main fluorescence gray value change amplitude is more than 15%, it is determined that the image is in a mutation state;

[0111] When the cell position offset in the continuous 3 frames of images is less than 5 μm, the boundary overlap degree is more than 80%, the fluorescence gray fluctuation is less than 5%, and the recognition confidence changes less than 5%, it is considered that the image recognition is stable.

[0112] In the mutation state, the system reduces the current flow rate to 60% of the original value, and prolongs the residence time of the sample in the enrichment area by 1.2 s; in the stable recognition state, the output channel switching instruction is output and the flow rate is reset to the default value.

[0113] In addition, when the image recognition module detects weak fluorescence signal (gray value less than 50) in the same area for 3 continuous frames and the confidence is between 0.6 and 0.8, the feedback module marks the area as an "enhanced target area", adjusts the LED excitation power to 130% of the original value, and prolongs the exposure time by 20 ms to improve the imaging contrast.

[0114] All control instructions are transmitted in real time to the liquid control module through a digital interface (such as CAN or SPI), and the control response period is less than 50 ms, ensuring the closed-loop adjustment capability of the image-liquid flow feedback link.

[0115] S4, result output and upload:

[0116] While the identification and adjustment process is ongoing, the system stores the identification results of each frame of image into a local database. Each record includes the image segment of the target cell, the identification category label, the center coordinate position, the identification confidence value, the time stamp (frame number and real time), and the channel number.

[0117] In this embodiment, the final output multidimensional detection report contains the following contents:

[0118] CTC identification time stamp sequence: indicating the frame number and actual time when the CTC is first identified;

[0119] Cell spatial position and classification information: including (x, y) coordinates, category, and channel label;

[0120] Fluorescence gray scale distribution map: showing the average intensity and maximum intensity of cells and background in each channel;

[0121] Flow rate control trajectory graph: drawing the flow rate adjustment curve of the sample in the entire detection period;

[0122] Confidence change curve graph: showing the time axis evolution trend of the confidence of all identified cells.

[0123] The above report is automatically generated by the system software in PDF file and output in JSON structured data format, and the data is transmitted to a remote server or HIS system through a USB interface, a Wi-Fi module, or a hospital internal local area network for follow-up analysis and remote auxiliary judgment by doctors.

[0124] The method provided in this embodiment maintains high identification sensitivity, realizes adaptive control of sample flow through image-driven feedback adjustment mechanism, and has effective adjustment capability in the identification of abnormal, weak signal, image mutation, and the like. It has high CTC identification accuracy, high sample utilization efficiency, real-time system response, good engineering adaptability and clinical transformation prospect, and is particularly suitable for high-throughput automated detection of multiple marker heterogeneous CTCs.

[0125] The basic principles, main features and advantages of the present application are shown and described. Those skilled in the art should understand that the present application is not limited by the above embodiments, and the above embodiments and descriptions in the specification are only to illustrate the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. An automatic detection device for circulating tumor cells based on microfluidic nanochip and AI analysis, characterized in that, The application relates to a circulating tumor cell automatic detection device based on a microfluidic nanochip and AI analysis. The microfluidic chip module comprises a transparent cover plate, an intermediate structural layer and a substrate support layer, a plurality of independent sample channels, enrichment areas and waste liquid outlets are arranged in the intermediate structural layer, and the inner surface of the enrichment area is modified with a circulating tumor cell specific recognition molecule; The optical acquisition module comprises an excitation light source, a light filter assembly and an image sensor, is arranged below the microfluidic chip module and is used for acquiring multispectral image information in the enrichment area; The image recognition module comprises an image preprocessing unit, a feature extraction network and an identification inference module, is used for extracting cell morphology parameters and fluorescence distribution parameters from the image and outputs an identification result; The feedback control module is used for generating an adjustment signal based on the image recognition result, and comprises flow rate adjustment, channel switching and residence time control instructions; The control strategy generation unit in the feedback control module constructs a dynamic trend identification based on image frame sequences, analyzes the change trend of target cells in the spatial position, morphological characteristics and fluorescence distribution intensity dimensions in the continuous image frames, judges the image state, and generates corresponding feedback control instructions according to the judgment result, the control strategy generation unit establishes a real-time instruction channel with the liquid path control module through a high-speed bus interface, the feedback response period of the updated image state vector is controlled to be less than 50 ms, and the real-time closed-loop control capability of identification, adjustment and repair is ensured; The liquid path control module comprises a microfluidic pump, a pressure control unit and an electromagnetic valve group, is used for receiving feedback instructions and adjusting the flow behavior of samples in the chip.

2. The circulating tumor cell automatic detection device based on a microfluidic nanochip and AI analysis according to claim 1, wherein the plurality of sample channels of the microfluidic chip module are at least four structure-symmetrical microchannels, the inlet of each channel is provided with an automatic switching valve, and each channel is connected with an independent enrichment area. The channel width is 50-100 mu m, the depth is controlled within the range of 30-80 mu m, a curved flow guide cavity is arranged at the intersection of the channels, and the curved flow guide cavity is used for slowing down the impact of cells and improving the adsorption efficiency of target cells.

3. The circulating tumor cell automatic detection device based on a microfluidic nanochip and AI analysis according to claim 1, wherein a metal nanostructure array is arranged in the enrichment area, the array comprises periodically arranged gold nanorods, silver nanospheres or carbon nanotaper arrays, the structure spacing is 100-200 nm, and the coverage area is more than 80% of the entire enrichment area.

4. The circulating tumor cell automatic detection device based on a microfluidic nanochip and AI analysis according to claim 1, wherein the image preprocessing unit in the image recognition module comprises an image denoising, edge sharpening and signal normalization processing procedure, the feature extraction network adopts a convolutional neural network structure, extracts parameters in the dimensions of cell size, fluorescence signal intensity, nuclear-cytoplasmic ratio and background gray difference of a target area and is used for realizing high-confidence identification of CTCs and non-target cells. ​ ​ ​ 5. The automatic detection device for circulating tumor cells based on microfluidic nanochip and AI analysis according to claim 1, characterized in that, The micro flow pump in the liquid path control module supports adjustment in the range of 0.5-10.0 μL / min, with a minimum resolution of 0.1 μL / min, the opening and closing response time of the electromagnetic valve is not more than 50 ms, and the pressure control unit is internally provided with a silicon-based tension balance diaphragm for maintaining the pressure difference before and after the enrichment area constant at 50-120 Pa.

Images