A virus detection method based on deep learning and surface-enhanced Raman spectroscopy

Abstract

The application discloses a virus detection method based on deep learning and surface enhanced Raman spectroscopy, which comprises the following steps: pretreating a target sample collected and mixing the target sample with a colloidal gold solution to obtain a mixed solution; performing SERS measurement on the mixed solution to obtain Raman spectrum data of the mixed solution; performing denoising processing on the Raman spectrum data; performing data enhancement processing on the Raman spectrum data to construct a Raman spectrum data set; constructing a virus classification model based on deep learning, and training the virus classification model by using a training data set; and performing virus classification detection by using the trained virus classification model. The application effectively deducts noise by performing cosmic noise deduction, SNIP baseline correction and SG convolution smoothing denoising on Raman spectrum, and obtains a spectrum graph with a better signal-to-noise ratio; in the classification algorithm for virus detection, the model is optimized, so that the model is more suitable for the spectrum data in the application, and a higher accuracy is obtained.

Description

Technical Field

[0001] This invention belongs to the technical field of virus detection, specifically relating to a virus detection method based on deep learning and surface-enhanced Raman spectroscopy. Background Technology

[0002] Viruses are tiny, simple, non-cellular organisms that lack a complete cellular structure, contain only one type of nucleic acid (DNA or RNA), and must parasitize living cells to multiply through replication. Most viruses are composed of nucleic acids and proteins, existing somewhere between life and non-life. Rapid and effective virus detection methods are crucial for suppressing viral transmission and protecting lives and property. Currently, traditional virus detection methods are broadly classified into two categories: nucleic acid detection and immunoassay. Nucleic acid detection can obtain the specific nucleic acid sequence of a virus, but the process is complex and time-consuming. For example, collecting throat or nasal swabs usually takes a long time, and the storage and transportation of nucleic acids are also critical steps. RNA viruses are very unstable and easily degrade during transportation. Throat and nasal swabs contain a large number of impurities during virus detection, so it is necessary to remove these impurities and purify the viral sequence. Each step in these processes is critical; any error in any step will result in the inability to detect viral nucleic acid. Immunoassay is an indirect method of virus detection, primarily relying on the search for virus-specific antibodies to prove the presence of the virus. However, in the early stages of viral infection, antibodies have a short duration of action and disappear quickly, making immunoassay difficult to use. Testing in the later stages of viral infection may cause patients to miss the optimal treatment window. Both nucleic acid and immunoassay tests are time-consuming, resource-intensive, and require extensive professional training for testing personnel. Therefore, developing a rapid and accurate viral detection method is of significant practical importance.

[0003] In 1928, CV Raman discovered in experiments that when light passes through a transparent medium and is scattered by molecules, the frequency of the light changes; this phenomenon is called Raman scattering. Later that year, this phenomenon was also observed in the Soviet Union and France. In the early days of Raman spectroscopy, its use was limited due to its weak signal and the requirement that the sample be sufficiently large, colorless, free of impurities, and free of fluorescence. In 1974, Fleischmann et al. obtained high-quality Raman spectra of pyridine adsorbed on the surface of a rough silver electrode using an electrochemical roughening method. This effect is called surface-enhanced Raman spectroscopy (SERS). SERS is a surface-sensitive technique that enhances Raman scattering by molecules or plasma adsorbed on rough metal surfaces, with enhancement factors reaching up to 10. 6 ~10 9This means that the technology can detect single molecules. Since water has almost no Raman peaks and does not interfere with the detection results, Raman spectroscopy is very suitable for research in the field of biological viruses. Therefore, surface-enhanced Raman spectroscopy holds promise as a novel, rapid, and accurate method for virus detection. Currently, there is already a large body of research on surface-enhanced Raman spectroscopy for virus detection.

[0004] Existing technologies utilize gold nanoparticles as an enhancing substrate to significantly enhance the Raman signal of the spike protein of the novel coronavirus, with a detection limit of 1 mmol / L.

[0005] Existing technology has designed a handheld novel coronavirus detection device based on Raman spectroscopy. In actual airport testing, this device achieved a false negative rate of 3.8% and a false positive rate of 0.1%. Its accuracy is comparable to that of polymerase chain reaction (PCR) testing, but it is faster and less expensive.

[0006] With the rapid improvement of computing power and the powerful feature extraction capabilities of deep learning, numerous deep learning models have been proposed, and artificial intelligence is being widely used in an increasing number of fields. Therefore, combining Raman spectroscopy with deep learning for virus detection is gradually attracting the attention of researchers.

[0007] Existing technologies add hidden layers to the Convolutional Neural Network (CNN) framework, giving the model powerful graphics processing capabilities.

[0008] Existing technologies, combined with multiple linear discriminant analysis (LDA), have achieved an accuracy rate of 95.65% in the qualitative examination of foodborne pathogens.

[0009] Existing technologies utilize one-dimensional CNNs combined with various data preprocessing methods to achieve rapid and accurate qualitative classification of estrogen powder Raman spectra.

[0010] Existing deep learning-based virus detection research relies on real throat swab Raman spectra. However, obtaining large quantities of real virus samples is difficult, and these samples often contain numerous impurities. Consequently, the obtained Raman spectra contain significant noise. Furthermore, during the acquisition of Raman spectra from real virus samples, it cannot be guaranteed that every point contains the virus; that is, it cannot be guaranteed that the spectra used for model training will all contain virus-containing Raman spectra. However, deep learning models require training data consisting of data samples with precise labels, which significantly impacts the detection accuracy of deep learning models. Summary of the Invention

[0011] The purpose of this invention is to address the aforementioned shortcomings in the prior art by providing a virus detection method based on deep learning and surface-enhanced Raman spectroscopy, thereby solving the problems of existing Raman spectral data containing a large amount of noise and the inability to guarantee that the spectra used for model training are all Raman spectra containing viruses.

[0012] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0013] A virus detection method based on deep learning and surface-enhanced Raman spectroscopy includes the following steps:

[0014] S1. Preprocess the collected target sample, and mix the preprocessed target sample with the prepared colloidal gold solution to obtain a mixed solution;

[0015] S2. Perform SERS measurements on the mixed solution to obtain the Raman spectral data of the mixed solution;

[0016] S3. Perform cosmic noise subtraction, baseline correction, and SG convolution smoothing and denoising on the Raman spectral data in sequence.

[0017] S4. Perform data augmentation on the Raman spectral data to construct a labeled Raman spectral dataset, and divide the Raman spectral dataset into a training dataset and a validation dataset.

[0018] S5. Construct a virus classification model based on deep learning, and train the virus classification model using a training dataset;

[0019] S6. Use the trained virus classification model to perform virus classification and detection.

[0020] Furthermore, the target samples collected in step S1 include protein samples of the target virus and throat swab samples.

[0021] Further, step S1 preprocesses the collected target samples, including:

[0022] The protein samples and pharyngeal swab samples were subjected to ultrafiltration centrifugation and pre-concentration.

[0023] Each sample had an initial volume of 450 μL. Each ultrafiltration centrifugation was performed at 4°C for 15 min with a centrifugal force of 14000 × g. A total of 4 ultrafiltration centrifugations were performed to obtain a protein sample or pharyngeal swab sample with a volume of 20 μL.

[0024] Further, the preparation of the mixed solution in step S1 includes:

[0025] Add 200 mL of ultrapure water and 8 mL of 1% chloroauric acid to a 250 mL triangular flask;

[0026] The solution in the triangular flask was heated and stirred using a heated magnetic stirrer until it boiled.

[0027] Add 480 μL of 20% trisodium citrate to the boiled solution at once, then heat to boiling for 5 minutes until the solution color changes from pale yellow to grayish black, and finally to wine red.

[0028] The wine-red solution was cooled naturally to obtain a colloidal gold solution.

[0029] 1 μL of protein sample and pharyngeal swab sample were dropped onto a gold-plated Si substrate, followed by 1 μL of colloidal gold solution and thorough mixing. The mixture was then dried in a validated biosafety cabinet.

[0030] Further, step S2 includes:

[0031] For protein samples, the laser power was set to 5mW, the integration time to 6s, and the wavelength range to 600–1680 cm⁻¹ was selected. -1 Raman spectroscopy measurements were performed within the spectral range to obtain Raman spectral data of the protein sample;

[0032] For pharyngeal swab samples, SERS spectra were acquired using fast Raman mapping mode with a laser power of 5 mW and a single-point integration time of 0.5 s. A 20 × 20 discrete point plot was generated for each sample, with a point spacing of 2 μm, and the plots were selected from 600 to 1700 cm⁻¹. -1 Raman spectroscopy measurements were performed within the spectral range to obtain Raman spectral data of the throat swab sample.

[0033] Furthermore, step S3 uses a cosmic noise subtraction method to process the Raman spectral data, including the following sub-steps:

[0034] A1. Design two center-aligned decision windows and a baseline window. The decision window is 20 points wide, corresponding to the maximum value of typical muon noise broadening. The baseline window is 42 points wide, and the baseline is considered to remain stable within the width of the baseline window.

[0035] A2. Calculate the mean and standard deviation of the baselines outside the decision window and inside the baseline window, respectively;

[0036] A3. Set the decision threshold to 20 times the standard deviation; if the intensity of the center point of the decision window exceeds the decision threshold, then treat the center point as a cosmic ray noise point and set the value of the center point to be equal to the baseline mean.

[0037] A4. Shift the decision window and baseline window along the Raman spectrum, and repeat steps A2 and A3 until the cosmic ray noise of the entire Raman spectrum is removed.

[0038] Furthermore, the background subtraction method SNIP is used to perform baseline correction processing on the Raman spectral data after cosmic noise subtraction;

[0039] SG convolution smoothing denoising was applied to the Raman spectral data after baseline correction.

[0040] Further, step S4 includes:

[0041] Raman spectra of high concentrations of viral characteristic proteins were randomly selected, multiplied by a scaling factor of 0.1 to 1, and superimposed with blank substrate spectra obtained in the laboratory to form a positive Raman spectrum dataset with positive labels.

[0042] 80% of the real negative samples were used as the negative training dataset, and the remaining 20% ​​were used as the validation dataset; the artificially synthesized Raman spectrum positive dataset was divided into two parts, with 80% used as the positive training set and the remaining 20% ​​used as the validation dataset.

[0043] Furthermore, the virus classification models in step S5 include MLP classification models, CNN classification models, and ResNet classification models.

[0044] The virus detection method based on deep learning and surface-enhanced Raman spectroscopy provided by this invention has the following beneficial effects:

[0045] 1. This invention promotes the development of surface-enhanced Raman spectroscopy in the field of biological detection technology and provides a technical reference for rapid and accurate virus detection. The Raman spectra collected in the laboratory are subjected to cosmic noise subtraction, SNIP baseline correction, and SG convolution smoothing to effectively remove noise and obtain spectra with good signal-to-noise ratio. In the classification algorithm for virus detection, the model is optimized to make it more suitable for the spectral data in this paper, thereby achieving a higher accuracy.

[0046] 2. The reinforcing substrate used in this invention is a combination substrate of gold nanoparticles and gold-plated Si. After the virus sample is dropped onto the reinforcing substrate and dried in a biosafety cabinet, a Raman spectroscopy is used to perform Raman surface scanning on each sample to avoid the influence of uneven virus distribution.

[0047] 3. In the process of preprocessing Raman spectral data, this invention employs a cosmic ray noise reduction algorithm based on a sliding window; at the same time, it uses the SNIP algorithm, a background reduction algorithm widely used in the field of energy spectrum analysis, for baseline correction, and uses SG convolution smoothing for noise reduction to complete the preprocessing of spectral data.

[0048] 4. This invention uses high-concentration viral protein spectra superimposed with blank substrate spectra to generate artificially synthesized samples, which are then used to train neural network models. The artificially synthesized sample spectral data is used to train MLP, CNN, and ResNet neural network models. In the classification models, the MLP model achieves a detection accuracy of 92.9% for SARS-CoV-2 and 93.8% for H1N1 virus; the CNN model achieves a detection accuracy of 94.1% for SARS-CoV-2 and 95.0% for H1N1 virus; and the ResNet model achieves a detection accuracy of 95.3% for SARS-CoV-2 and 96.3% for H1N1 virus. Attached Figure Description

[0049] Figure 1 This is the original Raman spectrum.

[0050] Figure 2 This is the Raman spectrum after deducting cosmic ray noise in this invention.

[0051] Figure 3 This is the Raman spectrum after baseline correction according to the present invention.

[0052] Figure 4 This is the Raman spectrum after SG smoothing convolution according to the present invention.

[0053] Figure 5 The present invention provides the Raman spectra of the SARS-CoV-2 S protein and the HA and NA proteins of influenza A virus.

[0054] Figure 6 This is the MLP classification model of the present invention.

[0055] Figure 7 This is the CNN classification model of the present invention.

[0056] Figure 8 This is the ResNet classification model of the present invention.

[0057] Figure 9 This invention is used for MLP detection of real COVID-19 samples.

[0058] Figure 10 This invention is used for MLP detection of real influenza A samples.

[0059] Figure 11 This invention uses CNN to detect real COVID-19 samples.

[0060] Figure 12 This invention uses CNN to detect real H1N1 influenza samples.

[0061] Figure 13 This invention uses ResNet to detect real COVID-19 samples.

[0062] Figure 14 This invention uses ResNet to detect real influenza A samples.

[0063] Figure 15 This is a flowchart of the virus detection method based on deep learning and surface-enhanced Raman spectroscopy of the present invention. Detailed Implementation

[0064] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0065] Example 1

[0066] This embodiment presents a virus detection method based on deep learning and surface-enhanced Raman spectroscopy. This method provides an effective approach for rapid and accurate virus detection. (Refer to...) Figure 15 Specifically, it includes:

[0067] Step S1: Preprocess the collected target sample, and mix the preprocessed target sample with the prepared colloidal gold solution to obtain a mixed solution, which specifically includes the following:

[0068] The target samples collected in this embodiment include protein samples of the target virus and throat swab samples.

[0069] As a preferred embodiment, the target virus is preferably SARS-CoV-2 or influenza A virus;

[0070] For SARS-CoV-2, the target protein is the spike protein (S protein); for influenza A virus, there are two target proteins: hemagglutinin (HA) protein and neuraminidase (NA) protein. The S protein, HA protein, and NA protein of influenza A virus used in this example were purchased from China National Biotec Group Co., Ltd.

[0071] All clinical samples were pharyngeal swab samples, which were divided into two parts after collection. One part was used for nucleic acid testing as the true value of the result, and the other part was used for Raman spectroscopy. The COVID-19 clinical samples were collected in December 2022 at Sichuan Science City Hospital, including 35 patients who tested positive for COVID-19 nucleic acid and 50 healthy individuals who tested negative. The H1N1 clinical samples were collected in September 2022 at Sichuan Science City Hospital, with 30 confirmed H1N1 patients as the positive group and 50 healthy individuals as the healthy control group.

[0072] Preprocessing of the collected target samples, namely sample purification and pre-concentration, includes:

[0073] Protein and clinical samples contain buffer solutions or viral transport medium (VTM) solutions, which may contain salts or other micromolecular compounds that can interfere with Raman measurements. Therefore, ultrafiltration centrifugation should be performed to remove the buffer or VTM solution and to pre-concentrate the protein and clinical samples before Raman measurement.

[0074] Specifically, the protein samples and pharyngeal swab samples were subjected to ultrafiltration centrifugation and pre-concentration.

[0075] Each sample had an initial volume of 450 μL. Each ultrafiltration centrifugation was performed at 4°C for 15 min with a centrifugal force of 14000 × g. A total of 4 ultrafiltration centrifugations were performed to obtain a protein sample or pharyngeal swab sample with a volume of 20 μL.

[0076] The preparation of the enhanced substrate includes:

[0077] The Raman spectra of all samples in this embodiment were collected on a gold-plated Si substrate with a coating thickness of 100 nm and a roughness of 0.5 nm. The specific fabrication process is as follows:

[0078] Add 200 mL of ultrapure water and 8 mL of 1% chloroauric acid to a 250 mL triangular flask;

[0079] The solution in the triangular flask was heated and stirred using a heated magnetic stirrer until it boiled.

[0080] Add 480 μL of 20% trisodium citrate to the boiled solution at once, then heat to boiling for 5 minutes until the solution color changes from pale yellow to grayish black, and finally to wine red.

[0081] The wine-red solution was cooled naturally to obtain a colloidal gold solution.

[0082] 1 μL of protein sample and pharyngeal swab sample were dropped onto a gold-plated Si substrate, followed by 1 μL of colloidal gold solution and thorough mixing. The mixture was then dried in a validated biosafety cabinet.

[0083] Step S2: Perform SERS measurements on the mixed solution to obtain the Raman spectral data of the mixed solution, which specifically includes the following:

[0084] In this embodiment, Raman spectra were measured using a Zolihan RTS2 Raman spectrometer equipped with a 785nm laser to avoid fluorescence interference. Wavenumber calibration was first performed using a Si sample before measurement. Raman signals were acquired using a ×50 objective telephoto lens.

[0085] For protein samples, the laser power was set to 5mW, the integration time to 6s, and the wavelength range to 600–1680 cm⁻¹ was selected. -1 The spectral range, which contains most of the protein information, is used to perform Raman spectroscopy measurements to obtain the Raman spectral data of the protein sample.

[0086] For pharyngeal swab samples, SERS spectra were acquired using rapid Raman mapping mode with a laser power of 5 mW and a single-point integration time of 0.5 s. A 20 × 20 discrete point plot was generated for each sample, with a point spacing of 2 μm to ensure the detection of proteins and viruses and to avoid spectral overlap. The target area was selected from 600 to 1700 cm⁻¹. -1 Raman spectroscopy measurements were performed within the spectral range to obtain Raman spectral data of the throat swab sample.

[0087] Step S3: Perform cosmic noise subtraction, baseline correction, and SG convolution smoothing denoising on the Raman spectral data in sequence;

[0088] As the principles of Raman spectroscopy state, Raman spectroscopy is a vibrational spectrum generated by energy exchange between molecules or in crystal lattices. According to the principle of energy exchange, the energy of a laser beam changes after passing through a substance, thus altering its wavelength and producing a Raman shift. The Raman shift is an inherent property of matter. Because Raman spectra are often affected by fluorescence interference, random noise, and cosmic noise, leading to errors in subsequent identification processes, effective spectral preprocessing is crucial.

[0089] Depend on Figure 1 It is known that the Raman spectra obtained in the laboratory contain a lot of noise, resulting in complex peaks and valleys, and are subject to severe substrate drift due to fluorescence effects. Therefore, the preprocessing of Raman spectra is necessary and important.

[0090] The cosmic noise contained in Raman spectroscopy originates from high-energy proton streams in the universe. When these protons pass through the atmosphere, they collide with atomic nuclei in the air, causing nuclear fission and producing a series of particles that form cosmic rays. These rays possess extremely strong penetrating power. Because CCDs are particularly sensitive to cosmic rays, they generate extremely strong and narrow cosmic ray noise when reading signals generated by cosmic rays. This embodiment designs a cosmic noise subtraction algorithm based on a sliding window, as follows:

[0091] Step A1: Design two center-aligned decision windows and baseline windows. The decision window width is 20 points, corresponding to the maximum value of typical muon noise broadening; the baseline window width is 42 points, and the baseline is considered to remain stable within the width of the baseline window.

[0092] Step A2: Calculate the mean and standard deviation of the baselines outside the decision window and inside the baseline window, respectively;

[0093] Step A3: Set the decision threshold to 20 times the standard deviation; if the intensity of the center point of the decision window exceeds the decision threshold, then the center point is regarded as a cosmic ray noise point, and the value of the center point is set to equal the baseline mean.

[0094] Step A4: Shift the decision window and baseline window along the Raman spectrum, and repeat steps A2 and A3 until the cosmic ray noise of the entire Raman spectrum is removed.

[0095] Figure 2 The Raman spectrum after cosmic ray subtraction is obtained from... Figure 1 and Figure 2 The comparison shows that cosmic ray noise has the characteristics of extremely strong and narrow signal. After subtracting cosmic ray noise, the peak and trough characteristics of the Raman spectrum become more obvious. Therefore, the subtraction of cosmic ray noise is very necessary.

[0096] Because laser excitation of the analyte introduces environmental and equipment noise during experiments, especially fluorescence noise interference which can cause baseline drift, many baseline subtraction algorithms have been proposed and applied. However, due to limitations such as the complexity of polynomial methods and the need for extensive parameter fine-tuning, this embodiment applies the widely used background subtraction method SNIP (Signal-Nearest Interference Proportioning) in gamma-ray spectroscopy analysis to baseline correction, achieving better baseline correction results. Figure 3 As shown.

[0097] Random noise in Raman spectroscopy is generated by thermal noise from the CCD and during data acquisition. Ambient lighting in the laboratory, laser plasma lines, etc., can also contribute to random noise. Therefore, this embodiment selects the most effective SG convolutional smoothing algorithm for noise reduction. The convolution coefficients of SG convolutional smoothing are independent of the spectral data, depending only on the sliding window size w and the highest degree k of the polynomial fitting. The spectra after SG smoothing convolutional filtering with k=3 and w=13 are as follows: Figure 4 As shown.

[0098] contrast Figure 3 and Figure 4 It is evident that a smoother spectrum is obtained after SG convolution smoothing. The fingerprint characteristics of Raman spectra are determined by Raman shifts; therefore, the loss of sharp details in the peaks and troughs of Raman spectra does not affect the subsequent feature extraction and classification process of neural network models.

[0099] Step S4: Perform data augmentation on the Raman spectral data to construct a labeled Raman spectral dataset, and divide the Raman spectral dataset into a training dataset and a validation dataset, which specifically includes the following:

[0100] Deep learning is a technology built on massive amounts of data. The training data consists of two parts: the first part is a blank substrate obtained from the laboratory and different concentrations of viral characteristic proteins; the second part is real positive and negative samples collected clinically. In this embodiment, each sample is scanned at 400 points to obtain 400 Raman spectra. In the process of obtaining Raman spectra with low concentrations of characteristic proteins, some points may not contain the characteristic protein, making it impossible to assign a definitive positive label to this spectrum. Therefore, in order to obtain a dataset with clear positive labels, this embodiment performs data augmentation processing, as follows:

[0101] Raman spectra of high concentrations of viral characteristic proteins were randomly selected, multiplied by a scaling factor of 0.1 to 1, and superimposed with blank substrate spectra obtained in the laboratory to form a positive Raman spectrum dataset with positive labels.

[0102] 80% of the real negative samples were used as the negative training dataset, and the remaining 20% ​​were used as the validation dataset. The artificially synthesized Raman spectrum positive dataset was divided into two parts, with 80% used as the positive training set and the remaining 20% ​​used as the validation dataset, ensuring that the spectra in the validation set would not appear in the training dataset.

[0103] Raman spectral characteristics of the novel coronavirus and influenza A virus in this embodiment

[0104] In this embodiment, Raman spectra of two viruses were used for detection, targeting their characteristic proteins. For the novel coronavirus, the target protein was the S protein, which is widely distributed on the surface of the virus and is key to its attack on the human body, thus serving as one of the biomarkers for detecting the novel coronavirus. For influenza A virus, the targets were the HA and NA proteins, which are two antigens on the influenza A virus and the origin of the virus's abbreviation H1N1.

[0105] Figure 5 The Raman spectra of the SARS-CoV-2 S protein and the HA and NA proteins of influenza A virus are shown. It can be seen that the Raman spectra of the HA and NA proteins of influenza A virus are highly similar, while the Raman spectrum of the SARS-CoV-2 S protein shows some differences. Figure 5 The green shaded area highlights the differences in the Raman spectra of the three proteins, with the most significant differences occurring between 10¹⁰ and 10⁶⁰ cm⁻¹. -1 This regional variation is due to differences in the content of amino acids such as phenylalanine, arginine, glycine, isoleucine, lysine, and tyrosine in these protein molecules. Other differences include 1450 cm⁻¹. -1 The nearby area is caused by the bending vibration of CH2; for example, at 1650cm -1 The difference in the surrounding region stems from variations in the α-helical structure of amide I. The assignments of the peak positions in the Raman spectrum are as follows: Figure 5 As shown.

[0106] Step S5: Construct a virus classification model based on deep learning and train the virus classification model using a training dataset. The virus classification model includes an MLP classification model, a CNN classification model, and a ResNet classification model.

[0107] This embodiment obtains a large amount of training data with precise labels based on the data augmentation method in step S4. Three classification models—Multilayer Perceptron (MLP), Convolutional Neural Network (CNN), and Residual Neural Network (ResNet)—are built using the TensorFlow machine learning library. 80% of the Raman spectroscopy data is used as the training set for the models, and the weights in the classification models are updated and optimized to achieve better training results. The remaining 20% ​​of the Raman spectroscopy data is used as the test set to test the accuracy of the model's classification predictions.

[0108] MLP classification model

[0109] The model structure in this embodiment is as follows: Figure 6As shown. Since both the hidden and output layers in an MLP are fully connected layers, MLPs possess strong feature representation capabilities and generalization abilities, enabling them to solve nonlinear problems. The multilayer perceptron model established in this embodiment achieves an accuracy of 92.9% in detecting SARS-CoV-2 virus and 93.8% in detecting H1N1 virus.

[0110] CNN classification model

[0111] Since Raman spectroscopy is one-dimensional data, this embodiment uses one-dimensional convolution to construct a CNN classification model, the structure of which is as follows: Figure 7 As shown. Due to the characteristics of the data in this embodiment, no huge memory expansion occurs during the convolution process. Therefore, in the construction of the CNN classification model, the pooling layer and the normalization layer in the CNN convolution model are removed. This preserves the most Raman spectral peak features, resulting in a better classification model. The improved CNN classification model achieves an accuracy of 94.1% in detecting SARS-CoV-2 virus and 95.0% in detecting H1N1 virus.

[0112] ResNet classification model

[0113] As neural networks become deeper, overfitting may occur. Therefore, fusing the input with the convolutional features to extract further features ensures no feature loss and prevents overfitting. Thus, a ResNet classification model was built, with the following structure: Figure 8 As shown, the ResNet classification model achieved an accuracy of 95.3% in detecting the SARS-CoV-2 virus and 96.3% in detecting the H1N1 virus.

[0114] Step S6: Use the trained virus classification model to perform virus classification and detection.

[0115] In this embodiment, 400 spectra were collected for each sample during Raman spectroscopy to avoid missed detections due to uneven virus distribution. Therefore, a threshold needs to be set in actual sample validation. When the number of positive points in the detected sample is greater than or equal to a certain threshold, the sample is considered positive; otherwise, it is considered negative.

[0116] In the real SARS-CoV-2 sample testing, 35 real positive samples and 50 real negative samples were tested. The classification accuracy was highest when the positive point threshold was set to 40. Figure 9 This is a schematic diagram illustrating the number of positive cases in real SARS-CoV-2 samples tested by the MLP model. The MLP model detected 33 positive cases and 46 negative cases, with a detection accuracy of 92.9%. Figure 11This diagram illustrates the number of positive cases in real SARS-CoV-2 samples tested using a CNN model. The CNN model detected 33 positive cases and 47 negative cases, achieving an accuracy rate of 94.1%. Figure 13 This is a schematic diagram illustrating the number of real SARS-CoV-2 positive cases tested using ResNet. The ResNet model detected 34 positive cases and 47 negative cases, with a detection accuracy of 95.3%.

[0117] In the real H1N1 sample test, 30 real H1N1 positive samples and 50 real negative samples were tested. The classification accuracy was highest when the positive point threshold was set to 150. Figure 10 This is a schematic diagram illustrating the number of positive H1N1 samples tested by the MLP model. The MLP model detected 28 positive cases and 47 negative cases, with a detection accuracy of 93.8%. Figure 12 This diagram illustrates the number of positive H1N1 samples tested using a CNN model. The CNN model detected 28 positive cases and 48 negative cases, achieving an accuracy rate of 95.0%. Figure 14 This diagram illustrates the number of positive H1N1 samples tested with ResNet. The ResNet model detected 28 positive cases and 49 negative cases, achieving an accuracy of 96.3%.

[0118] This invention is a virus detection technology based on deep learning and surface-enhanced Raman spectroscopy. It effectively removes noise by performing cosmic noise subtraction, SNIP baseline correction, and SG convolution smoothing on Raman spectra collected in the laboratory, resulting in a spectrum with a good signal-to-noise ratio. Furthermore, the accuracy of each algorithm in the virus detection classification algorithm remains above 92.9%.

[0119] This invention utilizes a ResNet-based detection model, which not only acquires more spectral features but also avoids overfitting during training. Therefore, the ResNet-based classification model achieved the highest detection accuracies of 95.3% and 96.3% for real samples of SARS-CoV-2 and H1N1, respectively.

[0120] Although specific embodiments of the invention have been described in detail with reference to the accompanying drawings, this should not be construed as limiting the scope of protection of this patent. Various modifications and variations that can be made by a person skilled in the art without inventive effort within the scope described in the claims still fall within the scope of protection of this patent.

Claims

1. A virus detection method based on deep learning and surface-enhanced Raman spectroscopy, characterized in that, Includes the following steps: S1. Preprocess the collected target sample, and mix the preprocessed target sample with the prepared colloidal gold solution to obtain a mixed solution; S2. Perform SERS measurements on the mixed solution to obtain the Raman spectral data of the mixed solution; S3. Perform cosmic noise subtraction, baseline correction, and SG convolution smoothing noise reduction on the Raman spectral data in sequence. S4. Perform data augmentation on the Raman spectral data to construct a labeled Raman spectral dataset, and divide the Raman spectral dataset into a training dataset and a validation dataset. S5. Construct a virus classification model based on deep learning, and train the virus classification model using a training dataset; S6. Use the trained virus classification model to perform virus classification and detection.

2. The virus detection method based on deep learning and surface-enhanced Raman spectroscopy according to claim 1, characterized in that: The target samples collected in step S1 include protein samples of the target virus and throat swab samples.

3. The virus detection method based on deep learning and surface-enhanced Raman spectroscopy according to claim 1, characterized in that, Step S1 involves preprocessing the collected target samples, including: The protein samples and pharyngeal swab samples were subjected to ultrafiltration centrifugation and pre-concentration. Each sample had an initial volume of 450 μL. Each ultrafiltration centrifugation was performed at 4°C for 15 min with a centrifugal force of 14000 × g. A total of 4 ultrafiltration centrifugations were performed to obtain a protein sample or pharyngeal swab sample with a volume of 20 μL.

4. The virus detection method based on deep learning and surface-enhanced Raman spectroscopy according to claim 3, characterized in that, The preparation of the mixed solution in step S1 includes: Add 200 mL of ultrapure water and 8 mL of 1% chloroauric acid to a 250 mL triangular flask; The solution in the triangular flask was heated and stirred using a heated magnetic stirrer until it boiled. Add 480 μL of 20% trisodium citrate to the boiled solution at once, then heat to boiling for 5 minutes until the solution color changes from pale yellow to grayish black, and finally to wine red. The wine-red solution was cooled naturally to obtain a colloidal gold solution. 1 μL of protein sample and pharyngeal swab sample were dropped onto a gold-plated Si substrate, followed by 1 μL of colloidal gold solution and thorough mixing. The mixture was then dried in a validated biosafety cabinet.

5. The virus detection method based on deep learning and surface-enhanced Raman spectroscopy according to claim 2, characterized in that, Step S2 includes: For protein samples, the laser power was set to 5mW, the integration time to 6s, and the wavelength range to 600–1680 cm⁻¹ was selected. -1 Raman spectroscopy measurements were performed within the spectral range to obtain Raman spectral data of the protein sample; For pharyngeal swab samples, SERS spectra were acquired using fast Raman mapping mode with a laser power of 5 mW and a single-point integration time of 0.5 s. A 20 × 20 discrete point plot was generated for each sample, with a point spacing of 2 μm, and the plots were selected from 600 to 1700 cm⁻¹. -1 Raman spectroscopy measurements were performed within the spectral range to obtain Raman spectral data of the throat swab sample.

6. The virus detection method based on deep learning and surface-enhanced Raman spectroscopy according to claim 1, characterized in that, Step S3 uses a cosmic noise subtraction method to process the Raman spectral data, including the following sub-steps: A1. Design two center-aligned decision windows and a baseline window. The decision window width is 20 points, corresponding to the maximum value of typical muon noise broadening. The baseline window width is 42 points, and the baseline is considered to remain stable within this baseline window width; A2. Calculate the mean and standard deviation of the baselines outside the decision window and inside the baseline window, respectively; A3. Set the decision threshold to 20 times the standard deviation; if the intensity of the center point of the decision window exceeds the decision threshold, then treat the center point as a cosmic ray noise point and set the value of the center point to be equal to the baseline mean. A4. Shift the decision window and baseline window along the Raman spectrum, and repeat steps A2 and A3 until the cosmic ray noise of the entire Raman spectrum is removed.

7. The virus detection method based on deep learning and surface-enhanced Raman spectroscopy according to claim 6, characterized in that, The background subtraction method SNIP was used to perform baseline correction on the Raman spectral data after cosmic noise subtraction. SG convolution smoothing denoising was applied to the Raman spectral data after baseline correction.

8. The virus detection method based on deep learning and surface-enhanced Raman spectroscopy according to claim 1, characterized in that, Step S4 includes: Raman spectra of high concentrations of viral characteristic proteins were randomly selected, multiplied by a scaling factor of 0.1 to 1, and superimposed with blank substrate spectra obtained in the laboratory to form a positive Raman spectrum dataset with positive labels. 80% of the real negative samples were used as the negative training dataset, and the remaining 20% ​​were used as the validation dataset; the artificially synthesized Raman spectrum positive dataset was divided into two parts, with 80% used as the positive training set and the remaining 20% ​​used as the validation dataset.

9. The virus detection method based on deep learning and surface-enhanced Raman spectroscopy according to claim 1, characterized in that, The virus classification models in step S5 include MLP classification models, CNN classification models, and ResNet classification models.

Images