3D printing microdroplet microfluidic immunoassay method based on YOLO-Drop

Abstract

The invention discloses a YOLO-Drop-based 3D printing droplet microfluidic immunoassay method, which organically integrates a high-precision microfluidic chip preparation technology, a droplet-level immunoreaction system and a fluorescence imaging acquisition and deep learning detection algorithm. And full-chain closed-loop processing from droplet generation, fluorescence signal acquisition, image recognition, result statistics to concentration quantitative analysis is realized. Through deep combination of chip physical parameters and model input features, the method not only significantly improves the accuracy and stability of droplet identification and concentration quantification, but also enhances the robustness of the algorithm under complex sample conditions through multi-scale feature fusion and a signal dual screening mechanism. Meanwhile, the optimized YOLO-Drop model supports embedded efficient reasoning, so that the whole analysis process has the characteristics of real-time performance and high throughput, and the method is suitable for rapid quantitative detection of various biomarkers.

Description

Technical Field

[0001] This invention relates to the field of biological detection and analysis technology, and more specifically, to a 3D-printed microdroplet microfluidic immunoassay method based on YOLO-Drop. Background Technology

[0002] Microdroplet microfluidic immunoassay technology has become an important tool in the life sciences and clinical diagnostics fields in recent years. By conducting reactions in droplets with volumes of microliters or even nanoliters, it significantly reduces reagent consumption and improves reaction efficiency. This technology is particularly suitable for the sensitive detection of low-abundance biomarkers and has shown broad application prospects in various detection tasks such as inflammatory markers, tumor markers, and infectious diseases. However, traditional droplet immunoassay workflows rely on manual microscope readings or inefficient batch processing algorithms, which are difficult to meet the needs of high-throughput and high-precision analysis. On the one hand, manual interpretation is highly subjective and inefficient, and cannot handle large-scale rapid detection in clinical or research settings; on the other hand, traditional image processing methods are prone to false positives and false negatives when droplet sizes are uneven, fluorescence intensity varies greatly, or complex background interference exists, resulting in insufficient accuracy and stability of analytical results.

[0003] ddELISA (digital droplet enzyme-linked immunosorbent assay), as a digital droplet immunoassay method, combines traditional ELISA with a microdroplet digital platform, enabling independent reactions of target molecules within a single droplet. This significantly improves detection sensitivity and dynamic range, making it one of the core technologies for ultrasensitive quantitative analysis. However, existing ddELISA platforms still have significant shortcomings: First, droplet generation, identification, and statistics largely rely on manual or semi-automated processes, resulting in long detection times, high labor costs, and difficulty in achieving batch consistency. Second, existing algorithms have limited ability to recognize droplet fluorescence signals in complex backgrounds, especially under low signal-to-noise ratio conditions, leading to a high false positive rate. Third, there is a lack of intelligent algorithm models that can be closely integrated with chip physical parameters and droplet volume information, making it difficult to achieve closed-loop optimization of the detection chain.

[0004] In recent years, deep learning-based target detection methods have been gradually introduced into droplet immunoassay tasks. The typical YOLO (YouOnlyLookOnce) series of algorithms, with their high accuracy and speed, excels in multi-target recognition. However, directly applying general detection models to droplet immunoassay scenarios still faces the following challenges: First, small droplet targets have weak features and blurred boundaries, making them easily ignored by general models; second, the fluorescence signal intensity of droplets varies significantly, and illumination and background interference under different batches or sample conditions can reduce the model's generalization ability; third, general algorithms lack integration with experimental parameters, failing to directly output core quantitative indicators such as droplet count, volume, and concentration, resulting in results that only reach the recognition level and cannot meet the urgent needs of clinical and research fields for quantitative analysis.

[0005] Therefore, there is an urgent need for a 3D-printed droplet microfluidic immunoassay method based on YOLO-Drop to solve these problems. Summary of the Invention

[0006] The purpose of this invention is to solve the technical problems mentioned in the background section above, and to provide a solution based on...

[0007] YOLO-Drop's 3D-printed droplet microfluidic immunoassay method includes the following steps:

[0008] S1. Chip Fabrication: Microfluidic chips are fabricated using 3D printing combined with soft molding processes to obtain a chip structure with channel dimensions that meet the requirements for droplet generation, and the channel height H, droplet diameter D, and droplet volume V are defined. d The parameter range;

[0009] S2, Immunoreactivity Loading: Immunoload with a target protein concentration of C... t The sample is mixed with capture antibody magnetic beads and detection reagents to form an immune complex, and the complex is injected into the chip channel as a dispersion phase.

[0010] S3, Droplet Generation: Uniform monodisperse droplets are formed at the oil-water interface of the chip, encapsulating the immune reaction system within the droplets, allowing the reaction signal S... d Formed within the droplet;

[0011] S4. Fluorescence Signal Acquisition: The droplet is imaged using a fluorescence microscopy system with multiple channels to obtain the original droplet image matrix I(x,y) and then pixel normalization is performed to obtain I0. n (x,y);

[0012] S5, YOLO-Drop Feature Extraction: Normalized image I n Inputting (x, y) into the YOLO-Drop model, the image is segmented and embedded to generate segmented feature vectors P.i And extract a multi-layer feature map set {C} through the backbone network. s}, perform multi-scale feature fusion to generate fused feature F l ;

[0013] S6. Candidate Droplet Identification: Based on Fusion Feature F l Identify candidate droplet locations and calculate confidence level p j A collection of candidate positive droplets is obtained;

[0014] S7. Confidence Optimization: Combining candidate droplet confidence levels p j and droplet reaction signal S d The weights are optimized using the focus loss function, and the signal is normalized to screen the set of positive droplets J. + ;

[0015] S8. Droplet Statistics and Analysis: Count the number of positive droplets N positive and the total number of droplets N total, And combined with the single droplet volume V d Calculate the total positive volume V p ositive and total reaction volume V total and the proportion of positive droplets R p ;

[0016] S9. Quantitative Concentration Analysis: Combining standard curve parameters A, B, C0, D, k, b with the input concentration C t According to the positive rate R p and average intensity Fusion calculation of target protein concentration C;

[0017] S10. Result Output and Real-time Control: The calculated target protein concentration C and analysis curve are output through a graphical interface, and the total analysis delay T is constrained. total Achieving rapid and high-precision immunoassay in under 5 seconds.

[0018] As a preferred embodiment of the present invention, the S1 chip fabrication steps include:

[0019] S11. Draw the channel structure in CAD software and export the STL file;

[0020] S12. After slicing the STL file, import it into the 3D printer and print it layer by layer using UV-cured photosensitive resin to obtain a high-precision resin mold.

[0021] S13. Mix PDMS prepolymer and crosslinking agent at a ratio of 10:1 and mold them. After vacuum degassing and heat curing, peel them off to obtain a microfluidic chip structure.

[0022] S14. After oxygen plasma treatment, the chip structure layer is bonded to the glass substrate;

[0023] And using the droplet volume formula:

[0024]

[0025] Determine the volume V of a single droplet d Where H: channel height, in μm; D: droplet diameter, in m; V d : Volume of a single droplet; π: Pi, a constant.

[0026] As a preferred embodiment of the present invention, the S2 immune response loading step includes:

[0027] S21. Add the target protein antigen to the magnetic bead solution containing the conjugated capture antibody and incubate to form a complex;

[0028] S22. After washing the magnetic beads, add biotinylated detection antibodies to form a sandwich immune structure;

[0029] S23. Add enzyme-coupled reagent, continue incubation and wash to remove excess reagent;

[0030] S24. The immune complex is suspended in the fluorescent substrate solution and injected into the oil-water interface of the chip as a dispersed phase to generate droplets.

[0031] As a preferred embodiment of the present invention, the S4 fluorescence signal acquisition step includes:

[0032] S41. Place the chip under a fluorescence microscopy imaging system and set the excitation wavelength to 561nm and the standard gain.

[0033] S42. Automatically acquires fluorescence channel, bright field channel and overlay image, with a resolution of 2048×2048 pixels;

[0034] S43. Output a JPG file with standardized naming and record metadata information.

[0035] As a preferred technical solution of the present invention, S5 normalizes the image I. n (x,y) is segmented and embedded to obtain the segmented feature vector P. i The data is then input into the backbone network to obtain a multi-layer feature map set C. s :

[0036]

[0037] Among them, I n (x,y): Normalized droplet image matrix, B i : The i-th image block region; P i: Embedded feature vector of the i-th image block; M: Total number of image blocks; C s : Multilayer feature map set {C3,C4,C5,C6}.

[0038] As a preferred technical solution of the present invention, for multi-layer features {C s} Perform weighted fusion to generate fusion feature F l The candidate box b is output by the detection head. j and confidence level p j :

[0039] (b j ,s j ) = Head(F l ),p j =σ(s j )

[0040] Among them, F l : Layer l fusion features; w ls : The fusion weights satisfy ∑w ls =1;, C s : Multilayer feature map set; b j : The bounding box of the j-th candidate droplet; s j : Logit score corresponding to candidate droplet; p j : Confidence level of candidate droplets.

[0041] As a preferred technical solution of the present invention, S7 combines confidence level p j and droplet reaction signal S d Optimize the filtering process:

[0042]

[0043] J + ={j∈J∣p j ≥τ∧μ j ≥θ}

[0044] Among them, y j : Candidate droplet tags; α, γ: Focal loss parameters; S d (j): The intensity of the reaction signal of the j-th droplet; S d,max The largest reaction signal among all droplets; M(b) j ): The set of pixels corresponding to candidate droplets; μ j : Signal intensity after normalization; τ,θ: Threshold; J + : Positive droplet collection.

[0045] As a preferred technical solution of the present invention, the total number of candidates N is counted. total With the number of positive droplets Npositive And calculate the positive rate R. p and average positive intensity

[0046]

[0047] Where, N total : Total number of candidate droplets; N positive : Total number of positive droplets; R p : Proportion of positive droplets; The average normalized intensity of positive droplets.

[0048] As a preferred technical solution of the present invention, the positive ratio R is utilized. p and average intensity Fusion calculation of target protein concentration C:

[0049]

[0050] in, From average strength Calculated concentration; From the positive rate R p Calculated concentration; A,D,B,C0: four-parameter Logistic curve parameters obtained by fitting standard samples; k,b: proportional mapping parameters obtained by fitting standard samples; η: fusion weight coefficient; C: target protein concentration.

[0051] As a preferred technical solution of the present invention, the total delay T of the analysis chain total It consists of the time delays of each component and satisfies:

[0052] T total =T acq +T embed +T fuse +T det +T sel +T stat +T quant ≤5s

[0053] Among them, T acq Image acquisition and normalization time; T embed : Block embedding and backbone extraction time; T fuse Multi-scale feature fusion time; T det : Time taken for candidate droplet generation; T sel Positive screening takes time; T stat Droplet statistics time; T quant Time consumed in quantitative concentration analysis; T total Total analysis time.

[0054] Beneficial effects:

[0055] First, this invention is the first to deeply integrate the physical parameters of microfluidic chips with deep learning detection algorithms, establishing a parameterized closed-loop processing logic for the entire chain of droplet generation, fluorescence acquisition, feature extraction, and result analysis. By introducing key parameters such as chip channel size, droplet volume, and reaction signal into the model, the detection process is upgraded from simple image recognition to intelligent quantitative analysis, breaking through the technical bottleneck of "separation of recognition and quantification" in traditional methods, and significantly improving the accuracy and reliability of the analysis.

[0056] Secondly, this invention innovatively improves the YOLO-Drop deep detection model to address the challenges of small target size, large signal differences, and complex background interference in droplet immunoassay. By optimizing the block embedding strategy, multi-scale feature fusion structure, and a dual screening mechanism of confidence and signal, the model maintains stable recognition and analysis performance in complex sample environments. This truly achieves deep customization for droplet immunoassay tasks, filling the technical gap where general-purpose algorithms struggle to achieve high-precision recognition in such specialized scenarios.

[0057] Finally, the technical solution proposed in this invention possesses high scalability and versatility. It can serve as a high-precision detection platform for a single indicator, and through parameter optimization, it can also achieve simultaneous detection and quantitative analysis of multiple indicators and types of biomarkers. This structured and modular technical framework not only expands the application boundaries of droplet immunoassay technology but also provides a sustainable upgrade path for future multi-dimensional point-of-care testing and precision medical diagnosis, demonstrating significant innovative value and industrial application potential. Attached Figure Description

[0058] Figure 1 The flowchart shows the 3D-printed droplet microfluidic immunoassay method based on YOLO-Drop proposed in this invention.

[0059] Figure 2 This is a flowchart of the immune reaction loading step in the YOLO-Drop-based 3D-printed droplet microfluidic immunoassay method proposed in this invention;

[0060] Figure 3 This is a flowchart of the fluorescence signal acquisition steps in the YOLO-Drop-based 3D-printed droplet microfluidic immunoassay method proposed in this invention. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0062] See Figures 1-3 As shown, this invention proposes a 3D-printed microdroplet microfluidic immunoassay method based on YOLO-Drop. This method organically integrates high-precision microfluidic droplet preparation technology, an intradroplet immunoreaction system, fluorescence imaging acquisition technology, and a deep learning detection algorithm, enabling highly sensitive immunoassay and quantitative concentration output of target proteins within seconds. The entire analysis process consists of droplet preparation, immunoreaction loading, image acquisition, feature extraction, positive screening, statistical analysis, and concentration calculation, with each step interconnected to form a closed-loop logic.

[0063] First, in the microfluidic chip fabrication stage, CAD software was used to design a three-dimensional channel structure that met the droplet generation requirements, and the STL file was exported for 3D printing. A high-precision chip mold was obtained by layer-by-layer molding using UV-cured resin. Then, PDMS prepolymer and crosslinking agent were mixed in a specific ratio and poured into the mold. After vacuum degassing and curing, the mold was peeled off to obtain a microfluidic chip structure with a smooth surface and precise dimensions. To ensure droplet consistency, the channel height H and droplet diameter D, and the volume V of a single droplet were strictly controlled during chip design. d Through formula Calculations are performed, where π is the constant of pi. This structure ensures that the droplet diameter is within the range of 8-80 μm, and the volume of a single droplet is maintained at the nL level, which meets the requirements of reaction encapsulation and ensures the stability of subsequent signal recognition.

[0064] Subsequently, the sample to be analyzed is introduced into the immunoreaction loading module. The sample binds to the capture antibody magnetic beads under mixed conditions to form a specific complex. Then, the detection antibody and enzyme conjugate are added sequentially to form a sandwich-style immunoreaction system, which is then washed to remove non-specific impurities. At this point, the sample contains the target protein at a concentration of C. t This parameter will be used as a calibration reference in subsequent concentration calculations and corrections. The above reaction system is injected as a dispersed phase into the microfluidic chip channel, and is sheared at the oil-water interface of the chip to generate uniform monodisperse droplets. This allows the immune reaction system to be uniformly encapsulated inside each droplet, and the reaction continues under isothermal conditions to form a stable reaction signal S. d .

[0065] After the reaction was complete, the chip was placed under a fluorescence microscopy imaging system to acquire images. By setting the excitation wavelength to 561 nm, standard gain, and multi-channel acquisition mode, the system simultaneously recorded fluorescence signal images and bright-field images, achieving an image resolution of 2048×2048 pixels, clearly presenting the fluorescence signal distribution of a single droplet. To eliminate differences in light intensity and exposure conditions between batches, the acquired raw image matrix I(x,y) was normalized using the following formula:

[0066]

[0067] Among them, I min and I max These represent the minimum and maximum pixel values ​​of the current image, respectively. The normalized image pixel values ​​are fixed in the range of 0 to 1, which helps improve the stability and recognition accuracy of subsequent algorithms.

[0068] Normalized Image I n After inputting (x, y) into the YOLO-Drop model, the first step is to enter the image block embedding module. This module divides the entire image into multiple small regions B. i The pixel information within each block is converted into a corresponding vector P by the embedding network. i The formula is

[0069]

[0070] Among them, P i Let represent the feature vector of the i-th block, and M represent the total number of blocks. These block feature vectors are input into the backbone network for multi-layer convolutional feature extraction, outputting a set of multi-scale feature maps.

[0071]

[0072] Among them, C s This represents the feature map of the s-th layer, which contains spatial and semantic information at different scales.

[0073] Next, a bidirectional weighted feature pyramid (BiFPN) structure is used to fuse multi-scale features to obtain the fused feature F. l :

[0074]

[0075] Among them, w ls These are the weighting coefficients of features at each scale, satisfying ∑w ls =1, Resize(·) is the size alignment function.

[0076] The fused features are processed by the detection head to output the location boxes of the candidate droplets. j and the corresponding logit value s j The confidence level p of each candidate droplet is obtained by transforming it using the Sigmoid function. j :

[0077] (b j ,s j ) = Head(F l ),p j =σ(s j )

[0078] The set of all candidate droplets is denoted as J.

[0079] To improve recognition accuracy, a focus loss function is used to optimize the network training process. The formula is as follows:

[0080]

[0081] Among them, y j This represents the true label of the sample, and α and γ are adjustment parameters.

[0082] At the same time, the normalized fluorescence intensity μ of each candidate droplet was... j The calculation is as follows:

[0083]

[0084] Among them, S d (j) represents the reaction signal intensity of the j-th droplet, S d,max M(b) represents the maximum signal intensity of this batch of droplets. j ) represents the set of pixel regions corresponding to the candidate droplets, combined with the confidence level p. j With normalized intensity μ j Joint screening yielded a set of positive droplets:

[0085] J + ={j∈J∣p j ≥τ∧μ j ≥θ}

[0086] Where τ and θ are the confidence threshold and intensity threshold, respectively. Joint screening can effectively reduce false positives and improve the accuracy and robustness of detection.

[0087] After positive droplet identification, statistical analysis of the number and volume of droplets is performed. Total number of candidate droplets N total and the number of positive droplets N positive They are respectively:

[0088] N total =|J|,N positive =|J + |

[0089] Positive droplet ratio R p Calculate using the following formula:

[0090]

[0091] Meanwhile, based on the single droplet volume V d Calculate the total volume V of the positive droplets. positive and total reaction volume V total :

[0092] V total =N total ·Vd V positive =N positive ·V d

[0093] Concentration-quantitative analysis is based on two core indicators: the proportion of positive droplets R. p and the average normalized intensity of positive droplets

[0094] First, the concentration is estimated based on the average intensity using a four-parameter Logistic fitting formula:

[0095]

[0096] Then, a linear estimation was performed using the proportion of positive droplets:

[0097]

[0098] Finally, the two are weighted and fused together, and combined with the input sample concentration C. t Perform deviation correction:

[0099]

[0100] Where η and δ are the fusion weighting coefficient and correction coefficient, respectively, and C is the final output target concentration value. This fusion strategy can comprehensively utilize intensity information, positive ratio information, and sample input information, significantly improving the accuracy of concentration prediction.

[0101] To verify the feasibility and superiority of the YOLO-Drop-based 3D-printed microdroplet microfluidic immunoassay method proposed in this invention, the research team first independently designed and fabricated a microfluidic chip with a channel height of approximately 50 μm and a droplet diameter of approximately 40 μm. According to the formula... Calculations show that the volume of a single droplet is approximately 33.5 nL, a size parameter that ensures droplet uniformity and encapsulation stability.

[0102] The standardized ddELISA platform, built upon this foundation, is used to generate and collect droplet fluorescence image data, resulting in a high-density training and validation dataset containing 5,574 images and nearly 1 million droplet annotation instances. Droplet types include single magnetic bead-signaled droplets, single magnetic bead-no-signaled droplets, droplets without magnetic beads, multiple magnetic bead-signaled droplets, and multiple magnetic bead-no-signaled droplets. This diverse droplet structure not only enhances the richness of the dataset but also significantly improves the model's generalization ability and robustness.

[0103] After data collection, the research team used the GPU-accelerated X-AnyLabeling tool for automatic labeling, and ensured that the labeling accuracy reached more than 99% through manual review and refinement. The final dataset contained an average of 180-200 droplets per image, with a balanced distribution of categories.

[0104] To verify the suitability and advantages of this dataset in droplet detection tasks, it was compared with mainstream object detection datasets. The results show that although the dataset of this invention has fewer images than large-scale datasets such as COCO and DOTA, it has a higher annotation density, richer single-image droplet instances, and the label categories are highly matched with the droplet analysis task, making it particularly suitable for the niche scenario of microdroplet immunoassay. The comparison results are shown in the table below:

[0105] Dataset Number of images Number of annotations Number of categories COCO (2017) 118，000 ～1.5M 80 DOTA (V2.0) 11，268 2.8M 18 VisDrone (2019) 10，209 540K 10 SKU-110K 11，762 1.7M 1 TinyPerson 7，608 70K 1 OursDataset 5，574 ～1M 5

[0106] Based on this high-quality dataset, YOLOv8 was used as the base network, and an improved YOLO-Drop architecture was introduced for training. The Adam optimizer was used during training, with an initial learning rate of 1×10⁻⁶. -3 The batch size is 32, and the training converges after 300 iterations.

[0107] The mean accuracy (mAP@0.5) on the validation set reached 98.15%, and the prediction results on 550 validation images showed a high linear correlation with the manually labeled results, with a correlation coefficient R0. 2 The mean square error is 0.993, and the mean square error is less than 0.014, which fully demonstrates that the model has extremely high detection accuracy and stability.

[0108]

[0109]

[0110] With a total processing time significantly less than the 5-second real-time constraint, it achieves rapid detection throughout the entire process from image input to concentration output, providing real-time advantages for clinical and field applications.

[0111] During the concentration-quantitative analysis validation phase, the research team selected IL-6 standard solutions with different concentration gradients, performed regression fitting using the standard curves generated by the platform, and determined the results based on the positive droplet ratio R. p and the average intensity of positive droplets Substitute into the following formula to perform the calculation:

[0112]

[0113] Then, the formula was corrected by fusion: The final output concentration was obtained. Experimental results show that, within the range of 0.05 fM to 10 fM, the goodness of fit R between the predicted and actual concentrations is [value missing]. 2 The detection sensitivity reached 0.998, with a minimum of 0.015 fM. Recovery rates in serum sample matrices remained between 94.0% and 116.96%, with deviations all below ±6%. Detection results for some concentration gradient samples are shown in the table below:

[0114]

[0115]

[0116] These results fully demonstrate that the method of the present invention maintains high detection sensitivity while also possessing excellent quantitative accuracy and stability.

[0117] Furthermore, in multiple batches of experiments, the method of this invention maintained high consistency and low error fluctuation in the detection of complex serum samples, demonstrating strong robustness and anti-interference ability. This performance advantage makes the method suitable not only for research and clinical laboratory environments, but also for diverse applications such as rapid screening, point-of-care testing, and emergency testing.

[0118] In summary, through the verification of the above embodiments, the 3D-printed microdroplet microfluidic immunoassay method based on YOLO-Drop proposed in this invention exhibits significant advantages in terms of sensitivity, accuracy, real-time performance, and robustness. The detection limit can reach 0.015 fM, the detection accuracy exceeds 98%, and the concentration fit goodness R0 is [value missing]. 2 With a sensitivity greater than 0.99 and a total detection time of less than 3 seconds, this invention is suitable for rapid, high-throughput detection scenarios. It is not only applicable to the highly sensitive quantitative detection of the inflammatory marker IL-6, but can also be extended to the detection of other proteins, nucleic acids, or multiplex biomarkers, demonstrating broad application prospects.

[0119] After training, the optimized YOLO-Drop model was deployed to the NVIDIA Jetson Orin Nano embedded inference platform. During inference, the system automatically recorded the time consumed at each stage using timestamps, from image acquisition, normalization, feature extraction, multi-scale feature fusion, candidate detection, positive screening, statistical analysis, to concentration calculation.

[0120] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A 3D-printed microdroplet microfluidic immunoassay method based on YOLO-Drop, characterized in that, Includes the following steps: S1. Chip Fabrication: Microfluidic chips are fabricated using 3D printing combined with soft molding processes to obtain a chip structure with channel dimensions that meet the requirements for droplet generation, and the channel height H, droplet diameter D, and droplet volume V are defined. d The parameter range; S2, Immunoreactivity Loading: Immunoload with a target protein concentration of C... t The sample is mixed with capture antibody magnetic beads and detection reagents to form an immune complex, and the complex is injected into the chip channel as a dispersion phase. S3. Droplet Generation: Uniform monodisperse droplets are formed at the oil-water interface of the chip, encapsulating the immune reaction system within the droplets, allowing the reaction signal S... d Formed within the droplet; S4. Fluorescence Signal Acquisition: The droplet is imaged using a fluorescence microscopy system with multiple channels to obtain the original droplet image matrix I(x,y), and pixel normalization is performed to obtain I0. n (x,y); S5, YOLO-Drop Feature Extraction: Normalized image I n Inputting (x, y) into the YOLO-Drop model, the image is segmented and embedded to generate segmented feature vectors P. i And extract a multi-layer feature map set {C} through the backbone network. s }, perform multi-scale feature fusion to generate fused feature F l ; S6. Candidate Droplet Identification: Based on Fusion Feature F l Identify candidate droplet locations and calculate confidence level p j A collection of candidate positive droplets is obtained; S7. Confidence Optimization: Combining candidate droplet confidence levels p j and droplet reaction signal S d The weights are optimized using the focus loss function, and the signal is normalized to screen the set of positive droplets J. + ; S8. Droplet Statistics and Analysis: Count the number of positive droplets N positive and the total number of droplets N total, And combined with the single droplet volume V d Calculate the total positive volume V p ositive and total reaction volume V total and the proportion of positive droplets R p ; S9. Quantitative Concentration Analysis: Combining standard curve parameters A, B, C0, D, k, b with the input concentration C t According to the positive rate R p and average intensity Fusion calculation of target protein concentration C; S10. Result Output and Real-time Control: The calculated target protein concentration C and analysis curve are output through a graphical interface, and the total analysis delay T is constrained. total Achieving rapid and high-precision immunoassay in under 5 seconds.

2. The YOLO-Drop-based 3D-printed microdroplet microfluidic immunoassay method according to claim 1, characterized in that, The S1 chip fabrication steps include: S11. Draw the channel structure in CAD software and export the STL file; S12. After slicing the STL file, import it into the 3D printer and print it layer by layer using UV-cured photosensitive resin to obtain a high-precision resin mold. S13. Mix PDMS prepolymer and crosslinking agent at a ratio of 10:1 and mold them. After vacuum degassing and heat curing, peel them off to obtain a microfluidic chip structure. S14. After oxygen plasma treatment, the chip structure layer is bonded to the glass substrate; And using the droplet volume formula: Determine the volume V of a single droplet d Where H: channel height, in μm; D: droplet diameter, in m; V d : Volume of a single droplet; π: Pi, a constant.

3. The YOLO-Drop-based 3D-printed microdroplet microfluidic immunoassay method according to claim 1, characterized in that, The S2 immune response loading step includes: S21. Add the target protein antigen to the magnetic bead solution containing the conjugated capture antibody and incubate to form a complex; S22. After washing the magnetic beads, add biotinylated detection antibodies to form a sandwich immune structure; S23. Add enzyme-coupled reagent, continue incubation and wash to remove excess reagent; S24. The immune complex is suspended in the fluorescent substrate solution and injected into the oil-water interface of the chip as a dispersed phase to generate droplets.

4. The YOLO-Drop-based 3D-printed microdroplet microfluidic immunoassay method according to claim 1, characterized in that, The S4 fluorescence signal acquisition step includes: S41. Place the chip under a fluorescence microscopy imaging system and set the excitation wavelength to 561nm and the standard gain. S42. Automatically acquires fluorescence channel, bright field channel and overlay image, with a resolution of 2048×2048 pixels; S43. Output a JPG file with standardized naming and record metadata information.

5. The YOLO-Drop-based 3D-printed microdroplet microfluidic immunoassay method according to claim 1, characterized in that, S5 normalizes image I n (x,y) is segmented and embedded to obtain the segmented feature vector P. i The data is then input into the backbone network to obtain a multi-layer feature map set C. s : Among them, I n (x,y): Normalized droplet image matrix, B i : The i-th image block region; P i : Embedded feature vector of the i-th image block; M: Total number of image blocks; C s : Multilayer feature map set {C3,C4,C5,C6}.

6. The YOLO-Drop-based 3D-printed droplet microfluidic immunoassay method according to claim 5, characterized in that, For multi-layer features {C s } Perform weighted fusion to generate fusion features The candidate box b is output by the detection head. j and confidence level p j : p j =σ(s j ) Among them, F l : Layer l fusion features; w ls : The fusion weights satisfy ∑w ls =1;, C s : Multilayer feature map set; b j : The bounding box of the j-th candidate droplet; s j : Logit score corresponding to candidate droplet; p j : Confidence level of candidate droplets.

7. The YOLO-Drop-based 3D-printed droplet microfluidic immunoassay method according to claim 1, characterized in that, The S7 is combined with confidence level p j and droplet reaction signal S d Optimize the filtering process: J + ={j∈J∣p j ≥τ∧μ j ≥θ} Among them, y j : Candidate droplet tags; α, γ: Focal loss parameters; S d (j): The intensity of the reaction signal of the j-th droplet; S d,max The largest reaction signal among all droplets; M(b) j ): The set of pixels corresponding to candidate droplets; μ j : Signal intensity after normalization; τ,θ: Threshold; J + : Positive droplet collection.

8. The YOLO-Drop-based 3D-printed microdroplet microfluidic immunoassay method according to claim 7, characterized in that, The total number of candidates N is statistically significant. total With the number of positive droplets N positive And calculate the positive rate R. p and average positive intensity N total =|J|,N positive =|J + |, Where, N total : Total number of candidate droplets; N positive : Total number of positive droplets; R p : Proportion of positive droplets; The average normalized intensity of positive droplets.

9. The YOLO-Drop-based 3D-printed droplet microfluidic immunoassay method according to claim 8, characterized in that, Using the positive rate R p and average intensity Fusion calculation of target protein concentration C: in, From average strength Calculated concentration; From the positive rate R p Calculated concentration; A,D,B,C0: four-parameter Logistic curve parameters obtained by fitting standard samples; k,b: proportional mapping parameters obtained by fitting standard samples; η: fusion weight coefficient; C: target protein concentration.

10. The YOLO-Drop-based 3D-printed droplet microfluidic immunoassay method according to claim 1, characterized in that, The total delay T of the analysis chain total It consists of the time delays of each component and satisfies: T total =T acq +T embed +T fuse +T det +T sel +T stat +T quant ≤5s Among them, T acq Image acquisition and normalization time; T embed : Block embedding and backbone extraction time; T fuse : Multi-scale feature fusion time; T det : Time taken for candidate droplet generation; T sel Positive screening takes time; T stat Droplet statistics time; T quant Time consumed in quantitative concentration analysis; T total Total analysis time.

Images