Digital enhanced Raman spectroscopy method

Abstract

The invention belongs to the field of optics, and discloses a digital enhanced Raman spectroscopy method, which comprises the following steps of: establishing a nonlinear characteristic mapping relation between a spontaneous Raman spectrum signal and a surface enhanced Raman spectrum signal by using a deep learning method, and greatly enhancing the Raman signal in a numerical calculation mode. And a surface-enhanced Raman spectrum signal equivalent to a surface-enhanced Raman spectrum signal after the Raman-enhanced substrate is added can be obtained under the condition that the Raman-enhanced substrate is not needed. According to the digital enhanced Raman spectrum method, a Raman enhanced substrate does not need to be introduced, so that complex operation and professional experiment conditions required by sample preparation are effectively avoided, the time efficiency and the signal intensity of Raman spectrum measurement can be greatly improved, and the digital enhanced Raman spectrum method has important significance for quickly acquiring high-quality Raman spectrum signals.

Description

Technical Field

[0001] The invention relates to the field of optical technology, and in particular to a digital enhanced Raman spectroscopy method. Background Art

[0002] Raman spectroscopy is a vibrational spectroscopy technique that obtains fingerprint information about molecular vibrations by measuring Raman scattering between photons and molecules, and then extracts chemical properties and composition information about samples non-invasively and label-free. It has been widely used in many fields such as materials science, chemistry, biomedicine, environmental testing, and food safety. However, due to the small Raman scattering cross section, the Raman signal is usually very weak, which brings great challenges to the subsequent Raman spectroscopy data processing and analysis, and also greatly affects the practical application of Raman spectroscopy. Therefore, it is of great significance and application value to obtain high-quality Raman spectroscopy data by enhancing the Raman signal intensity.

[0003] To enhance the Raman signal, the simplest and most direct method in actual Raman measurement is to increase the laser power, extend the exposure time, or combine the two. However, increasing the laser power or extending the exposure time will greatly increase the risk of sample damage, and extending the exposure time will significantly increase the measurement time, so it is difficult to achieve fast and non-destructive Raman detection of the sample. In addition, surface-enhanced Raman spectroscopy technology can enhance the Raman signal more significantly without increasing the laser power or extending the exposure time by mixing the Raman enhancement substrate into the sample, and the molecules are physically and chemically enhanced due to adsorption on rough surfaces such as gold and silver. The enhancement effect can reach 10,000 to 10 billion times. Different from the enhancement of all Raman signals by increasing the laser power or extending the exposure time, surface-enhanced Raman spectroscopy technology has a certain selectivity for the enhancement of Raman signals, and the enhancement effect for different Raman peaks is not completely consistent, that is, the spontaneous Raman signal without the addition of the enhancement substrate and the surface-enhanced Raman signal after the addition of the enhancement substrate are nonlinear feature mapping relationships. Although surface-enhanced Raman technology can significantly enhance Raman signals, it usually requires professional operators and professional experimental conditions to prepare samples, which greatly limits the practical application of surface-enhanced Raman technology.

[0004] In addition to the above experimental methods to enhance Raman signals, numerical calculation methods are also widely used to obtain Raman spectra with higher signal-to-noise ratios, but the focus of the task is mainly on filtering and denoising, that is, smoothing, filtering and other denoising processes are performed on the measured Raman spectrum data to improve its signal-to-noise ratio. In recent years, with the rapid development of machine learning and deep learning methods, a Raman signal reconstruction and recovery model has been established through supervised learning methods, that is, learning the feature mapping relationship between low-SNR Raman signals and high-SNR Raman signals. It has been successfully achieved to reconstruct and restore high-SNR Raman signals from low-SNR Raman signals. While removing noise, Raman signals can also be enhanced. However, it is currently limited to low-SNR and high-SNR Raman signals with linear feature mapping relationships, that is, when the enhancement effect of all Raman peaks is consistent, such as low-SNR Raman signals with short exposure time and high-SNR Raman signals with long exposure time. However, due to the nonlinear feature mapping relationship between spontaneous Raman signals and surface-enhanced Raman signals, it has not yet been possible to reconstruct and restore spontaneous Raman signals to surface-enhanced Raman signals with more significant enhancement effects. Summary of the invention

[0005] In view of the problem that existing methods are difficult to achieve rapid and lossless acquisition of high-quality Raman signals, the present invention proposes a digital enhanced Raman spectroscopy method. The method uses a deep learning method to learn the nonlinear feature mapping relationship between the spontaneous Raman signal without the addition of an enhanced substrate and the surface enhanced Raman signal after the addition of an enhanced substrate, establishes a digital enhancement model for the Raman signal, and reconstructs the spontaneous Raman signal to restore the surface enhanced Raman signal with a more significant enhancement effect. The method greatly enhances the Raman signal by numerical calculation, and can obtain a surface enhanced Raman spectroscopy signal equivalent to that after the addition of a Raman enhanced substrate without the need to prepare a Raman enhanced substrate, effectively avoiding the professional operation and experimental conditions required for the preparation of a Raman enhanced substrate, and saving detection time and consumables costs, which is conducive to the wider application of Raman detection technology.

[0006] The technical solution of the present invention is: a digital enhanced Raman spectroscopy method, which adopts a deep learning training strategy to establish and train a Raman signal digital enhancement model to obtain a nonlinear feature mapping relationship between a spontaneous Raman spectroscopy signal and a surface enhanced Raman spectroscopy signal; according to the nonlinear feature mapping relationship, the Raman signal is enhanced by numerical calculation, and a surface enhanced Raman spectroscopy signal equivalent to that after adding a Raman enhanced substrate is obtained without the need for a Raman enhanced substrate.

[0007] The Raman signal digital enhancement model is implemented by the following steps:

[0008] Step 1: Establish a paired Raman signal training data set;

[0009] Using the same Raman spectrometer and the same parameter settings, the spontaneous Raman signal from the same sample and the surface-enhanced Raman signal after mixing into the Raman-enhanced substrate are measured respectively; the sample is replaced and the above Raman signal measurement process is repeated to obtain a large number of paired spontaneous Raman signals and surface-enhanced Raman signals, and a paired Raman signal training data set is established;

[0010] Step 2: training a Raman signal digital enhancement model from spontaneous Raman signal to surface enhanced Raman signal;

[0011] Based on the paired Raman signal training data set in step 1, a Raman signal digital enhancement model from spontaneous Raman signal to surface enhanced Raman signal is trained using a deep learning network and a corresponding deep learning training strategy; the Raman signal digital enhancement model uses the Raman Spectral Transfer adaptive loss function RST loss Solve the overfitting problem of deep learning method training and obtain the nonlinear feature mapping relationship between spontaneous Raman signal and surface enhanced Raman signal;

[0012] Step 3, reconstructing and restoring the surface enhanced Raman signal from the spontaneous Raman signal;

[0013] The Raman signal digital enhancement model trained in step 2 is used to take the newly measured spontaneous Raman signal as input and output an enhanced digitally enhanced Raman signal, wherein the enhanced digitally enhanced Raman signal is close to the actually measured surface enhanced Raman signal in terms of accuracy and signal-to-noise ratio.

[0014] The adaptive loss function RST loss The structural similarity and absolute average error are used as specific optimization targets to ensure that the digital enhanced Raman signal and the actual surface enhanced Raman signal maintain similar overall shape and intensity of characteristic peaks, and the gradient loss is introduced to optimize the calculation efficiency; the adaptive loss function RST loss The formula is as follows:

[0015] RST loss =1-(MAE loss +α×Grad+β×SSIM) (1)

[0016]

[0017] Among them, α and β are adaptive trainable parameters, ranging from (0,1); N is the number of spectral samples; y i is the intensity value of the surface enhanced Raman signal; is the intensity value of the digitally enhanced Raman signal; G i is the predicted gradient of the surface-enhanced Raman signal; is the predicted gradient of the digitally enhanced Raman signal; μx is the average value of the surface enhanced Raman signal; μ y is the average value of the digitally enhanced Raman signal; is the variance of the SERS signal, is the variance of the digitally enhanced Raman signal; σ xy is the covariance of the surface enhanced Raman signal and the digital enhanced Raman signal; C1 and C2 are calculated based on the maximum intensity value x of each surface enhanced Raman signal. max and the maximum intensity value of the digitally enhanced Raman signal Adaptive selection to avoid the situation where the denominator is zero.

[0018] The deep learning strategy training method is supervised learning, unsupervised learning or hybrid supervised learning. The first half of the hybrid supervision is supervised learning, and the second half is unsupervised learning (no labels are set to minimize the loss between the output of the first half and the output of the second half), but the training is carried out simultaneously. The digital enhanced Raman signal generated in the first half actually has a high signal-to-noise ratio, and the second half is further optimized (adjusting the signal strength) to improve the signal-to-noise ratio.

[0019] The deep learning network is a convolutional neural network and its variants, a generative adversarial network, a Transformer architecture or a Diffusion architecture.

[0020] The deep learning network is a Transformer architecture. When performing hybrid supervised learning, a multi-scale spectral feature fusion multi-head attention mechanism is designed to explore the correlation between global and local spectral features; the multi-scale spectral feature fusion multi-head attention mechanism is specifically as follows: it includes two parallel paths, one path uses a single multi-head self-attention mechanism to extract global information; the other path uses three groups of separable convolutions with different convolution kernel scales to extract different spectral local information, which are activated by the ReLU function and used as the query, key and value vectors required for calculating the multi-head self-attention mechanism to obtain the integrated local features; after layer normalization and activation processing, the global information is fused with the local information in a unit addition manner.

[0021] In the Transformer architecture of the deep learning network, a dual-channel fusion feedforward network is formed by using a separable convolutional path and a spatial-positional attention mechanism to further enhance feature representation;

[0022] The spatial-position attention mechanism includes a spatial attention module and a position attention module; the output of the spatial attention module is the input of the position attention module;

[0023] The spatial attention module is as follows: first, the feature signals extracted by the multi-scale spectral feature fusion multi-head attention mechanism are subjected to maximum pooling and average pooling to obtain the maximum value and average value of each channel respectively; the obtained maximum value feature vector and average value feature vector are concatenated along the channel dimension and processed through a separable convolution layer to generate the weight of the feature matrix; by multiplying and weighting with the original features input to the multi-scale spectral feature fusion multi-head attention mechanism, specific spatial features are enhanced and irrelevant information is suppressed;

[0024] The position attention module is as follows: first, the input is transformed through three identical separable convolutions to generate three sets of new feature representations, then the Softmax function is used to calculate the similarity scores between the first two sets of features, and the third set of features is weighted using the similarity scores to obtain contextual enhanced features based on position similarity; the contextual enhanced features are linearly scaled and added to the output of the spatial attention module to obtain the position attention weight, and finally the output of the spatial attention module is multiplied and weighted by the position attention weight to complete the spatial position feature fusion.

[0025] Structural similarity, root mean square error, and peak signal-to-noise ratio are used to quantitatively evaluate the accuracy and signal-to-noise ratio of digitally enhanced Raman signals.

[0026] The beneficial effect of the present invention is that a digital enhanced Raman spectroscopy method is proposed, which can greatly enhance the Raman signal by numerical calculation. Without the need to prepare a Raman enhanced substrate, a surface enhanced Raman spectroscopy signal equivalent to that after adding a Raman enhanced substrate can be obtained, which effectively avoids the professional operation and experimental conditions required for the preparation of the Raman enhanced substrate, saves detection time and consumables costs, effectively solves the key bottleneck limiting the practical application of Raman detection technology, and has important application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is a schematic diagram of the actual measured paired spontaneous and surface-enhanced Raman signals. (a) is spontaneous Raman; (b) is surface-enhanced Raman.

[0028] Figure 2 This is a schematic diagram of the DCGAN network architecture. (a) is the generator; (b) is the discriminator.

[0029] Figure 3 Figure 1 is a schematic diagram of the CycleGAN network structure. (a) is the generator; (b) is the discriminator.

[0030] Figure 4 This is a schematic diagram of the Auto encoder network architecture.

[0031] Figure 5This is a schematic diagram of the Diffusion architecture. (a) is the network framework; (b) is the HNF structure; (c) is the bridge structure.

[0032] Figure 6 This is a schematic diagram of the deep learning network architecture using the Transformer architecture.

[0033] Figure 7 It is a schematic diagram of the multi-head attention mechanism for fusing multi-scale spectral features.

[0034] Figure 8 It is a schematic diagram of the spatial-position attention mechanism.

[0035] Fig. 9 This is a schematic diagram of the Transformer architecture.

[0036] Fig.10 Schematic diagram of digital enhanced Raman signal results based on different deep learning methods. (a) is the actual measured surface enhanced Raman signal; (b) is DCGAN; (c) is Auto encoder; (d) is Cycle Gan; (e) is Diffusion; (f) is Transformer.

[0037] Fig.11 Schematic diagram of the digitally enhanced Raman signal results of different deep learning methods based on the traditional loss function mean square error MSE. (a) is DCGAN; (b) is Auto encoder; (c) is Cycle Gan; (d) is Diffusion; (e) is Transformer.

[0038] Fig.12 Schematic diagram of digitally enhanced Raman spectroscopy method. DETAILED DESCRIPTION

[0039] The following is a detailed description of the specific implementation of the present invention using a set of actually measured spontaneous and surface enhanced Raman spectroscopy data in conjunction with the accompanying drawings.

[0040] Example 1

[0041] In this example, the spontaneous and surface-enhanced Raman signals collected from the serum of cancer patients were randomly sampled in a stratified manner according to a ratio of 3:1:1 and divided into training, validation and test sets. Figure 2 As shown, combined with the adaptive loss function RST lossThe deep learning architecture is trained using the training and validation sets and the test set is used to test the trained Raman signal digital enhancement model. The reconstructed digitally enhanced Raman signal is evaluated based on its similarity to the real surface enhanced Raman signal in the test set. The specific evaluation indicators include structural similarity, root mean square error and peak signal-to-noise ratio.

[0042] The specific implementation steps are as follows:

[0043] Step 1: Under the same experimental conditions, 2800 pairs of spontaneous and surface-enhanced Raman signals from cancer patient serum were collected to establish a paired Raman signal data set, such as Figure 1 As shown in Figure 2, the training dataset is randomly stratified and sampled in a ratio of 3:1:1 and further divided into training, validation, and test sets.

[0044] Step 2: Based on the training set of 1680 pairs and the validation set of 560 pairs of spontaneous Raman signals and surface-enhanced Raman signals in step 1, a DCGAN deep learning model architecture is established, such as Figure 2 As shown. Combined with the adaptive loss function RST loss And unsupervised deep learning strategy, training the Raman signal digital enhancement model from spontaneous Raman signal to surface enhanced Raman signal, learning the nonlinear feature mapping relationship between spontaneous Raman spectroscopy signal and surface enhanced Raman spectroscopy signal. The specific situation is as follows:

[0045] (1) The loss function of the generator network is the adaptive loss function RST loss , the formula is as follows:

[0046] RST loss =1-(MAE loss +α×Grad+β×SSIM) (1)

[0047]

[0048] Among them, α and β are adaptive trainable parameters, ranging from (0,1); N is the number of spectral samples; y i and are the intensity values ​​of surface enhanced Raman signal and digital enhanced Raman signal respectively; G i and are the predicted gradients of the surface enhanced Raman signal and the digital enhanced Raman signal, respectively; μ x and μ y are the average values ​​of surface enhanced Raman signals and digital enhanced Raman signals, respectively; and are the variance of the surface enhanced Raman signal and the variance of the digital enhanced Raman signal respectively; σ xyis the covariance of the surface enhanced Raman signal and the covariance of the digital enhanced Raman signal; C1 and C2 are calculated based on the maximum intensity value x of each surface enhanced Raman signal and digital enhanced Raman signal. max or Adaptive selection to avoid the situation where the denominator is zero.

[0049] (2) The loss function of the discriminator network is the cross entropy loss function;

[0050] (3) The generator adopts unsupervised training and alternates training for 500 rounds;

[0051] (4) Use the ADAM optimizer to correct the gradient deviation;

[0052] (5) The batch size and learning rate were optimized. When the batch size was 32 and the learning rate was 2×10 -4 Get the best performance when

[0053] Step 3: Using the Raman signal digital enhancement model trained in step 2, the low signal-to-noise ratio spontaneous Raman signal in the test set divided in step 1 is enhanced into a high signal-to-noise ratio digitally enhanced Raman signal. The representative digitally enhanced Raman signal results are shown in the figure. Fig.10 (b) as shown.

[0054] The quantitative evaluation index results of the digitally enhanced Raman signal based on the DCGAN Raman signal digital enhancement model are shown in Table 1. Among them, the SSIM is 0.9782, the RMSE is 0.0306, and the PSNR is 30.1251, which shows that the digitally enhanced Raman spectroscopy method has an excellent enhancement effect. Fig.10 (a) and Fig.10 (b) The digitally enhanced Raman signal is close to the actual measured surface enhanced Raman signal in terms of accuracy and signal-to-noise ratio. In addition, on the Intel(R) Core(TM) i9-14900KF CPU and NVIDIA GeForce RTX4090 GPU, the time required to enhance a spontaneous Raman test set is about 7.5×10 -3 The time required for the preparation and incubation of the enhanced substrate for surface enhanced Raman signal measurement is usually about 1 hour or more. Therefore, this embodiment can significantly save the consumption of consumables and time costs required to obtain high-quality Raman signals.

[0055] Combined with the existing technical solutions - using the traditional quantitative loss function mean square error MSE comparison. The difference from Example 1 is that the loss function of the generator network is the mean square error MSE. When the batch size is 64 and the learning rate is 2×10 -4The best performance is obtained when the DCGAN Raman signal digital enhancement model under the MSE loss function is used. The quantitative evaluation index results of the digitally enhanced Raman signal are shown in Table 2. Among them, SSIM is 0.9374, RMSE is 0.0354, and PSNR is 28.7054. Compared with the adaptive loss function RST designed in this invention, loss The reconstruction indicators under the condition are: SSIM is 0.9782, RMSE is 0.0306, PSNR is 30.1251, and the digital enhanced Raman signal is inferior to the adaptive loss function RST in terms of accuracy and signal-to-noise ratio. loss The reason for this is that the mean square error (MSE) only emphasizes minimizing the reconstruction error and ignores the overall feature shape and intensity. The adaptive loss function RST loss The structural similarity, absolute average error and gradient loss are comprehensively considered to enhance the accuracy of the characteristic morphology in the signal, so that each peak and change are better preserved, thereby improving the overall effect of spectral reconstruction. Therefore, this comparative example can clearly prove that the adaptive loss function RST loss The optimization effect of digitally enhanced Raman signal is shown in Figure 2.

[0056] Example 2

[0057] In this example, spontaneous and surface-enhanced Raman signals collected from cancer patient serum were randomly sampled in a stratified manner according to a ratio of 3:1:1 and divided into training, validation, and test sets. Figure 3 As shown, combined with the adaptive loss function RST loss The CycleGAN architecture is trained using the training and validation sets and the test set is used to test the trained Raman signal digital enhancement model. The reconstructed digitally enhanced Raman signal is evaluated based on its similarity to the real surface enhanced Raman signal in the test set. The specific evaluation indicators include structural similarity, root mean square error and peak signal-to-noise ratio.

[0058] The specific implementation steps are as follows:

[0059] Step 1: Under the same experimental conditions, 2800 pairs of spontaneous and surface-enhanced Raman signals from cancer patient serum were collected to establish a paired Raman signal data set, such as Figure 1 As shown in Figure 2, the training dataset is randomly stratified and sampled in a ratio of 3:1:1 and further divided into training, validation, and test sets.

[0060] Step 2: Based on the training set of 1680 pairs and the validation set of 560 pairs of spontaneous and surface-enhanced Raman signals in step 1, the CycleGAN deep learning model architecture is established, as shown in Figure 3 As shown. Combined with the adaptive loss function RST lossAnd unsupervised deep learning strategy, training the Raman signal digital enhancement model from spontaneous Raman signal to surface enhanced Raman signal, learning the nonlinear feature mapping relationship between spontaneous Raman spectroscopy signal and surface enhanced Raman spectroscopy signal. The specific situation is as follows:

[0061] (1) The loss function of the generator network is the adaptive loss function RST loss , the formula is as follows:

[0062] RST loss =1-(MAE loss +α×Grad+β×SSIM) (1)

[0063]

[0064] Among them, α and β are adaptive trainable parameters, ranging from (0,1); N is the number of spectral samples; y i and are the intensity values ​​of surface enhanced Raman signal and digital enhanced Raman signal respectively; G i and are the predicted gradients of the surface enhanced Raman signal and the digital enhanced Raman signal, respectively; μ x and μ y are the average values ​​of surface enhanced Raman signals and digital enhanced Raman signals, respectively; and are the variance of the surface enhanced Raman signal and the variance of the digital enhanced Raman signal respectively; σ xy is the covariance of the surface enhanced Raman signal and the covariance of the digital enhanced Raman signal; C1 and C2 are calculated based on the maximum intensity value x of each surface enhanced Raman signal and digital enhanced Raman signal. max or Adaptive selection to avoid the situation where the denominator is zero.

[0065] (2) The loss function of the discriminator network is the cycle consistency loss function;

[0066] (3) The generator adopts unsupervised training and alternates training for 500 rounds;

[0067] (4) Use the ADAM optimizer to correct the gradient deviation;

[0068] (5) The batch size and learning rate were optimized. When the batch size was 32 and the learning rate was 2.5×10 -4 Get the best performance when

[0069] Step 3: Using the Raman signal digital enhancement model trained in step 2, the spontaneous Raman signal with low signal-to-noise ratio in the test set divided in step 1 is enhanced into a digitally enhanced Raman signal with high signal-to-noise ratio. The representative digitally enhanced Raman signal results are shown in the figure. Fig.10 (c) as shown.

[0070] The quantitative evaluation index results of the digitally enhanced Raman signal based on the CycleGAN Raman signal digital enhancement model are shown in Table 1. Among them, the SSIM is 0.9812, the RMSE is 0.0232, and the PSNR is 33.0058, which shows that the digitally enhanced Raman spectroscopy method has an excellent enhancement effect. Fig.10 (a) and Fig.10 (c) The digitally enhanced Raman signal is close to the actual measured surface enhanced Raman signal in terms of accuracy and signal-to-noise ratio. In addition, on an Intel(R) Core(TM) i9-14900KF CPU and NVIDIA GeForce RTX 4090 GPU, the time required to enhance a spontaneous Raman test set is approximately 7.7×10 -3 The time required for the preparation and incubation of the enhanced substrate for surface enhanced Raman signal measurement is usually about 1 hour or more. Therefore, this embodiment can significantly save the consumption of consumables and time costs required to obtain high-quality Raman signals.

[0071] Combined with the existing technical solutions - using the traditional quantitative loss function mean square error MSE comparison. The difference from Example 2 is that the loss function of the generator network is the mean square error MSE. When the batch size is 64 and the learning rate is 2×10 -4 The best performance is obtained when the CycleGAN Raman signal digital enhancement model under the MSE loss function is used. The quantitative evaluation index results of the digitally enhanced Raman signal are shown in Table 2. Among them, SSIM is 0.9499, RMSE is 0.0339, and PSNR is 29.3723. Compared with the adaptive loss function RST designed in this invention, loss The reconstruction indicators under the condition are: SSIM is 0.9812, RMSE is 0.0232, PSNR is 33.0058. The digital enhanced Raman signal is inferior to the adaptive loss function RST in terms of accuracy and signal-to-noise ratio. loss The reason for this is that the mean square error (MSE) only emphasizes minimizing the reconstruction error and ignores the overall feature shape and intensity. The adaptive loss function RST loss The structural similarity, absolute average error and gradient loss are comprehensively considered to enhance the accuracy of the characteristic morphology in the signal, so that each peak and change are better preserved, thereby improving the overall effect of spectral reconstruction. Therefore, this comparative example can clearly prove that the adaptive loss function RST lossThe optimization effect of digitally enhanced Raman signal is shown in Figure 2.

[0072] Example 3

[0073] In this example, the spontaneous and surface enhanced Raman signals collected from the serum of cancer patients were randomly sampled in a stratified manner according to a ratio of 3:1:1 and divided into training, validation and test sets. loss The Autoencoder architecture is trained using the training and validation sets and the test set is used to test the trained Raman signal digital enhancement model. The reconstructed digitally enhanced Raman signal is evaluated based on its similarity to the real surface enhanced Raman signal in the test set. The specific evaluation indicators include structural similarity, root mean square error and peak signal-to-noise ratio.

[0074] The specific implementation steps are as follows:

[0075] Step 1: Under the same experimental conditions, 2800 pairs of spontaneous and surface-enhanced Raman signals from cancer patient serum were collected to establish a paired Raman signal data set, such as Figure 1 As shown in Figure 2, the training dataset is randomly stratified and sampled in a ratio of 3:1:1 and further divided into training, validation, and test sets.

[0076] Step 2: Based on the training set of 1680 pairs and the validation set of 560 pairs of spontaneous and surface-enhanced Raman signals in step 1, the Auto encoder deep learning model architecture is established, as follows: Figure 4 As shown. Combined with the adaptive loss function RST loss And supervised deep learning strategy, train the Raman signal digital enhancement model from spontaneous Raman signal to surface enhanced Raman signal, and learn the nonlinear feature mapping relationship between spontaneous Raman spectroscopy signal and surface enhanced Raman spectroscopy signal. The specific situation is as follows:

[0077] (1) The loss function of the Auto encoder is the adaptive loss function RST loss , the formula is as follows:

[0078] RST loss =1-(MAE loss +α×Grad+β×SSIM) (1)

[0079]

[0080] Where: α and β are adaptive trainable parameters, ranging from (0,1); N is the number of spectral samples; y i and are the intensity values ​​of surface enhanced Raman signal and digital enhanced Raman signal respectively; Gi and are the predicted gradients of surface enhanced Raman signal and digital enhanced Raman signal, respectively; μ x and μ y are the average values ​​of surface enhanced Raman signal and digital enhanced Raman signal, respectively; and are the variances of the surface enhanced Raman signal and the digital enhanced Raman signal, respectively; σ xy is the covariance of the surface enhanced Raman signal and the digital enhanced Raman signal; C1 and C2 are calculated based on the maximum intensity value x of each surface enhanced Raman signal and digital enhanced Raman signal. max or Adaptive selection to avoid the situation where the denominator is zero.

[0081] (2) Supervised training was adopted, and 500 rounds of training were conducted;

[0082] (3) Use the ADAM optimizer to correct the gradient deviation;

[0083] (4) The batch size and learning rate were optimized. When the batch size was 64 and the learning rate was 2×10 -4 Get the best performance when

[0084] Step 3: Using the Raman signal digital enhancement model trained in step 2, the spontaneous Raman signal with low signal-to-noise ratio in the test set divided in step 1 is enhanced into a digitally enhanced Raman signal with high signal-to-noise ratio. The representative digitally enhanced Raman signal results are shown in the figure. Fig.10 (d) as shown.

[0085] Based on the Raman signal digital enhancement model of Auto encoder, the quantitative evaluation index results of the digitally enhanced Raman signal are shown in Table 1. Among them, SSIM is 0.9858, RMSE is 0.0275, and PSNR is 31.9076, which shows that the digital enhanced Raman spectroscopy method has an excellent enhancement effect. Fig.10 (a) and Fig.10 (d) The digitally enhanced Raman signal is close to the actual measured surface enhanced Raman signal in terms of accuracy and signal-to-noise ratio. In addition, on an Intel(R) Core(TM) i9-14900KF CPU and NVIDIA GeForce RTX 4090 GPU, the time required to enhance a spontaneous Raman test set is approximately 6.6×10 -3 The time required for the preparation and incubation of the enhanced substrate for surface enhanced Raman signal measurement is usually about 1 hour or more. Therefore, this embodiment can significantly save the consumption of consumables and time costs required to obtain high-quality Raman signals.

[0086] Combined with the existing technical solutions - using the traditional quantitative loss function MSE comparison. The difference from Example 3 is that the loss function of the generator network is the mean square error MSE. When the batch size is 128 and the learning rate is 2.5×10 -4 The best performance is obtained when the Auto encoder Raman signal digital enhancement model based on the MSE loss function is used. The quantitative evaluation index results of the digital enhanced Raman signal are shown in Table 2. Among them, SSIM is 0.9414, RMSE is 0.0346, and PSNR is 29.2035. Compared with the adaptive loss function RST designed in this invention, loss The reconstruction indicators under the condition are: SSIM is 0.9782, RMSE is 0.0306, PSNR is 30.1251, and the digital enhanced Raman signal is inferior to the adaptive loss function RST in terms of accuracy and signal-to-noise ratio. loss The reason for this is that the mean square error (MSE) only emphasizes minimizing the reconstruction error and ignores the overall feature shape and intensity. The adaptive loss function RST loss The structural similarity, absolute average error and gradient loss are comprehensively considered to enhance the accuracy of the characteristic morphology in the signal, so that each peak and change are better preserved, thereby improving the overall effect of spectral reconstruction. Therefore, this comparative example can clearly prove that the adaptive loss function RST loss The optimization effect of digitally enhanced Raman signal is shown in Figure 2.

[0087] Example 4

[0088] In this example, spontaneous and surface enhanced Raman signals collected from cancer patient serum were randomly sampled in a stratified manner according to a ratio of 3:1:1 and divided into training, validation and test sets. A Diffusion network based on the Homogeneous Neural Filters (HNF) structure was constructed, combined with an adaptive loss function RST loss The Diffusion architecture is trained using the training and validation sets and the test set is used to test the trained Raman signal digital enhancement model. The reconstructed digitally enhanced Raman signal is evaluated based on its similarity to the real surface enhanced Raman signal in the test set. The specific evaluation indicators include structural similarity, root mean square error and peak signal-to-noise ratio.

[0089] The specific implementation steps are as follows:

[0090] Step 1: Under the same experimental conditions, 2800 pairs of spontaneous and surface-enhanced Raman signals from cancer patient serum were collected to establish a paired Raman signal data set, such as Figure 1As shown in Figure 2, the training dataset is randomly stratified and sampled in a ratio of 3:1:1 and further divided into training, validation, and test sets.

[0091] Step 2: Based on the 1680 pairs of training sets and 560 pairs of validation sets of spontaneous and surface-enhanced Raman signals in step 1, a Diffusion deep learning model architecture is established, such as Figure 5 As shown. Combined with the adaptive loss function RST loss And supervised deep learning strategy, train the Raman signal digital enhancement model from spontaneous Raman signal to surface enhanced Raman signal, and learn the nonlinear feature mapping relationship between spontaneous Raman spectroscopy signal and surface enhanced Raman spectroscopy signal. The specific situation is as follows:

[0092] (1) The loss function of Diffusion is the adaptive loss function RST loss , the formula is as follows:

[0093] RST loss =1-(MAE loss +α×Grad+β×SSIM) (1)

[0094]

[0095] Where: α and β are adaptive trainable parameters, ranging from (0,1); N is the number of spectral samples; y i and are the intensity values ​​of surface enhanced Raman signal and digital enhanced Raman signal respectively; G i and are the predicted gradients of the surface enhanced Raman signal and the digital enhanced Raman signal, respectively; μ x and μ y are the average values ​​of surface enhanced Raman signals and digital enhanced Raman signals, respectively; and are the variance of the surface enhanced Raman signal and the variance of the digital enhanced Raman signal respectively; σ xy is the covariance of the surface enhanced Raman signal and the covariance of the digital enhanced Raman signal; C1 and C2 are calculated based on the maximum intensity value x of each surface enhanced Raman signal and digital enhanced Raman signal. max or Adaptive selection to avoid the situation where the denominator is zero.

[0096] (2) Supervised training was adopted, and 400 rounds of training were conducted;

[0097] (3) Use the ADAM optimizer to correct the gradient deviation;

[0098] (4) The minimum value of the noise scheduling parameter setting is β0 = 0.0001, and the maximum value is βT =0.5;

[0099] (5) The batch size and learning rate were optimized. When the batch size was 8 and the learning rate was 8×10 -4 Get the best performance when

[0100] Step 3: Using the Raman signal digital enhancement model trained in step 2, the spontaneous Raman signal with low signal-to-noise ratio in the test set divided in step 1 is enhanced into a digitally enhanced Raman signal with high signal-to-noise ratio. The representative digitally enhanced Raman signal results are shown in the figure. Fig.10 (e) as shown.

[0101] Based on the Raman signal digital enhancement model of Diffusion, the quantitative evaluation index results of the digitally enhanced Raman signal are shown in Table 1. Among them, SSIM is 0.9895, RMSE is 0.0194, and PSNR is 35.2820, which shows that the digital enhanced Raman spectroscopy method has an excellent enhancement effect. Fig.10 (a) and Fig.10 (e) The digitally enhanced Raman signal is close to the actual measured surface enhanced Raman signal in terms of accuracy and signal-to-noise ratio. In addition, on an Intel(R) Core(TM) i9-14900KF CPU and NVIDIA GeForce RTX 4090 GPU, the time required to enhance a spontaneous Raman test set is 1.0×10 -2 The time required for the preparation and incubation of the enhanced substrate for surface enhanced Raman signal measurement is usually about 1 hour or more. Therefore, this embodiment can significantly save the consumption of consumables and time costs required to obtain high-quality Raman signals.

[0102] Combined with the existing technical solutions - using the traditional quantitative loss function MSE comparison. The difference from Example 4 is that the loss function of the generator network is the MSE. When the batch size is 128 and the learning rate is 1.5×10 -4 The best performance is obtained when the Diffusion Raman signal digital enhancement model under the MSE loss function is used. The quantitative evaluation index results of the digital enhanced Raman signal are shown in Table 2. Among them, SSIM is 0.9600, RMSE is 0.0389, and PSNR is 28.7505. Compared with the adaptive loss function RST designed in the present invention, loss The reconstruction indicators under the condition are: SSIM is 0.9895, RMSE is 0.0194, PSNR is 35.2820, and the digital enhanced Raman signal is inferior to the adaptive loss function RST in terms of accuracy and signal-to-noise ratio. lossThe reason for this is that the mean square error (MSE) only emphasizes minimizing the reconstruction error and ignores the overall feature shape and intensity. The adaptive loss function RST loss The structural similarity, absolute average error and gradient loss are comprehensively considered to enhance the accuracy of the characteristic morphology in the signal, so that each peak and change are better preserved, thereby improving the overall effect of spectral reconstruction. Therefore, this comparative example can clearly prove that the adaptive loss function RST loss The optimization effect of digitally enhanced Raman signal is shown in Figure 2.

[0103] Example 5

[0104] In this example, spontaneous and surface-enhanced Raman signals collected from cancer patient serum were randomly sampled in a stratified manner according to a ratio of 3:1:1 and divided into training, validation and test sets. Figure 7 , Figure 8 The schematic diagram builds the Transformer module, such as Fig. 9 As shown. Based on the Transformer module, the Transformer deep learning network architecture is further constructed as shown Figure 6 As shown in the figure, specifically, the encoder and decoder of the reconstruction subnetwork are composed of several Transformer modules, which decompose the spectrum downsampling and restore the upsampling step by step. In order to avoid the gradient vanishing problem as the Transformer module deepens, feature aggregation is performed through jump connections to reduce information loss in downsampling. Finally, the Raman spectrum features are preliminarily mapped to SERS using a fully connected layer and a Softplus activation function. After the correction subnetwork receives the preliminarily mapped digitally enhanced Raman signal, a similar Unet structure is used to reconstruct the spectrum. The preliminarily digitally enhanced Raman signal is extracted at multiple levels using a fully connected layer, and then the abstract features at each level are further refined and refined using the Transformer module. Finally, the refined features are fused with the upsampled output step by step.

[0105] Combined with adaptive loss function RST loss The Transformer deep learning network architecture is trained using the training and validation sets and the test set is used to test the trained Raman signal digital enhancement model. The reconstructed digitally enhanced Raman signal is evaluated based on its similarity to the real surface enhanced Raman signal in the test set. The specific evaluation indicators include structural similarity, root mean square error and peak signal-to-noise ratio.

[0106] The specific implementation steps are as follows:

[0107] Step 1: Under the same experimental conditions, 2800 pairs of spontaneous and surface-enhanced Raman signals from cancer patient serum were collected to establish a paired Raman signal data set, such as Figure 1 As shown in Figure 2, the training dataset is randomly stratified and sampled in a ratio of 3:1:1 and further divided into training, validation, and test sets.

[0108] Step 2: Based on the training set of 1680 pairs and the validation set of 560 pairs of spontaneous and surface-enhanced Raman signals in step 1, establish the Transformer deep learning model architecture, such as Figure 6 As shown. Combined with the adaptive loss function RST loss And the hybrid supervised deep learning strategy is used to train the Raman signal digital enhancement model from spontaneous Raman signal to surface enhanced Raman signal, and learn the nonlinear feature mapping relationship between spontaneous Raman spectroscopy signal and surface enhanced Raman spectroscopy signal. The specific situation is as follows:

[0109] (1) Train the reconstruction network (supervised) and the correction network (unsupervised) for 500 rounds simultaneously;

[0110] (2) The loss function of the reconstruction network (supervised) and the correction network (unsupervised) is the adaptive loss function RST loss , the formula is as follows:

[0111] RST loss =1-(MAE loss +α×Grad+β×SSIM) (1)

[0112]

[0113] Where: α and β are adaptive trainable parameters, ranging from (0,1); N is the number of spectral samples; y i and are the intensity values ​​of surface enhanced Raman signal and digital enhanced Raman signal respectively; G i and are the predicted gradients of the surface enhanced Raman signal and the digital enhanced Raman signal, respectively; μ x and μ y are the average values ​​of surface enhanced Raman signals and digital enhanced Raman signals, respectively; and are the variance of the surface enhanced Raman signal and the variance of the digital enhanced Raman signal respectively; σ xy is the covariance of the surface enhanced Raman signal and the covariance of the digital enhanced Raman signal; C1 and C2 are calculated based on the maximum intensity value x of each surface enhanced Raman signal and digital enhanced Raman signal. max or Adaptive selection to avoid the situation where the denominator is zero.

[0114] (3) Use the ADAM optimizer to correct the gradient deviation;

[0115] (4) The batch size and learning rate were optimized. When the batch size was 24 and the learning rate was 2×10 -4 Get the best performance when

[0116] (5) Adaptive learning rate decay and early stopping are used. If the validation loss of the Transformer does not decrease within 10 consecutive rounds of training, the learning rate will be automatically reduced to 80% of the previous value, with a lower limit of 10 -5 If the validation loss does not decrease for 100 consecutive epochs, training is terminated.

[0117] Step 3: Using the Raman signal digital enhancement model trained in step 2, the spontaneous Raman signal with low signal-to-noise ratio in the test set divided in step 1 is enhanced into a digitally enhanced Raman signal with high signal-to-noise ratio. The representative digitally enhanced Raman signal results are shown in the figure. Fig.10 (f) as shown.

[0118] The quantitative evaluation index results of the digitally enhanced Raman signal based on the Transformer Raman signal digital enhancement model are shown in Table 1. Among them, the SSIM is 0.9946, the RMSE is 0.0118, and the PSNR is 39.3677, which shows that the digitally enhanced Raman spectroscopy method has an excellent enhancement effect. Fig.10 (a) and Fig.10 (f) The digitally enhanced Raman signal is close to the actual measured surface enhanced Raman signal in terms of accuracy and signal-to-noise ratio. In addition, on an Intel(R) Core(TM) i9-14900KF CPU and NVIDIA GeForce RTX 4090 GPU, the time required to enhance a spontaneous Raman test set is 1.2×10 -2 The time required for the preparation and incubation of the enhanced substrate for surface enhanced Raman signal measurement is usually about 1 hour or more. Therefore, this embodiment can significantly save the consumption of consumables and time costs required to obtain high-quality Raman signals.

[0119] Combined with the existing technical solutions - using the traditional quantitative loss function mean square error MSE comparison. The difference from Example 5 is that the loss function of the generator network is the mean square error MSE. When the batch size is 24 and the learning rate is 2×10 -4The best performance is obtained when the Transformer Raman signal digital enhancement model under the MSE loss function is used. The quantitative evaluation index results of the digital enhanced Raman signal are shown in Table 2. Among them, SSIM is 0.9553, RMSE is 0.0391, and PSNR is 28.9040. Compared with the adaptive loss function RST designed in this invention, loss The reconstruction indicators under the condition are: SSIM is 0.9946, RMSE is 0.0118, PSNR is 39.3677. The digital enhanced Raman signal is inferior to the adaptive loss function RST in terms of accuracy and signal-to-noise ratio. loss The reason for this is that the mean square error (MSE) only emphasizes minimizing the reconstruction error and ignores the overall feature shape and intensity. The adaptive loss function RST loss The structural similarity, absolute average error and gradient loss are comprehensively considered to enhance the accuracy of the characteristic morphology in the signal, so that each peak and change are better preserved, thereby improving the overall effect of spectral reconstruction. Therefore, this comparative example can clearly prove that the adaptive loss function RST loss The optimization effect of digitally enhanced Raman signal is shown in Figure 2.

[0120] Table 1 Quantitative average evaluation indicators of Raman signal enhancement by various deep learning methods.

[0121]

[0122] Table 2 Quantitative average evaluation index of Raman signal enhancement of various deep learning methods under the traditional mean square error loss function MSE

[0123]

Claims

1. A digital enhanced Raman spectroscopy method, characterized in that: A deep learning training strategy is used to establish and train a Raman signal digital enhancement model to obtain the nonlinear feature mapping relationship between spontaneous Raman spectroscopy signals and surface enhanced Raman spectroscopy signals. According to the nonlinear characteristic mapping relationship, the Raman signal is enhanced by numerical calculation, and a surface enhanced Raman spectroscopy signal equivalent to that after adding the Raman enhancement substrate is obtained without the need for a Raman enhancement substrate.

2. The digital enhanced Raman spectroscopy method according to claim 1, characterized in that: The Raman signal digital enhancement model is implemented by the following steps: Step 1: Establish a paired Raman signal training data set; The same Raman spectrometer and the same parameter settings are used to measure the spontaneous Raman signal from the same sample and the surface-enhanced Raman signal after mixing into the Raman-enhancing substrate. Replace the sample and repeat the above Raman signal measurement process to obtain a large number of paired spontaneous Raman signals and surface enhanced Raman signals, and establish a paired Raman signal training data set; Step 2: training a Raman signal digital enhancement model from spontaneous Raman signal to surface enhanced Raman signal; Based on the paired Raman signal training data set in step 1, a Raman signal digital enhancement model from spontaneous Raman signal to surface enhanced Raman signal is trained using a deep learning network and a corresponding deep learning training strategy; the Raman signal digital enhancement model uses the Raman Spectral Transfer adaptive loss function RST loss Solve the overfitting problem of deep learning method training and obtain the nonlinear feature mapping relationship between spontaneous Raman signal and surface enhanced Raman signal; Step 3, reconstructing and restoring the surface enhanced Raman signal from the spontaneous Raman signal; The Raman signal digital enhancement model trained in step 2 is used to take the newly measured spontaneous Raman signal as input and output an enhanced digitally enhanced Raman signal, wherein the enhanced digitally enhanced Raman signal is close to the actually measured surface enhanced Raman signal in terms of accuracy and signal-to-noise ratio.

3. The digital enhanced Raman spectroscopy method according to claim 2, characterized in that: The adaptive loss function RST loss The structural similarity and absolute average error are used as specific optimization targets to ensure that the digital enhanced Raman signal and the actual surface enhanced Raman signal maintain similar overall shape and intensity of characteristic peaks, and the gradient loss is introduced to optimize the calculation efficiency; the adaptive loss function RST loss The formula is as follows: RST loss =1-(MAE loss +α×Grad+β×SSIM) (1) Among them, α and β are adaptive trainable parameters, ranging from (0,1); N is the number of spectral samples; y i is the intensity value of surface enhanced Raman signal; is the intensity value of the digitally enhanced Raman signal; G i is the predicted gradient of the surface-enhanced Raman signal; is the predicted gradient of the digitally enhanced Raman signal; μ x is the average value of the surface enhanced Raman signal; μ y is the average value of the digitally enhanced Raman signal; is the variance of the SERS signal, is the variance of the digitally enhanced Raman signal; σ xy is the covariance of the surface enhanced Raman signal and the digital enhanced Raman signal; C1 and C2 are calculated based on the maximum intensity value x of each surface enhanced Raman signal. max and the maximum intensity value of the digitally enhanced Raman signal Adaptive selection to avoid the situation where the denominator is zero.

4. The digital enhanced Raman spectroscopy method according to claim 2, characterized in that: The deep learning strategy training method is supervised learning, unsupervised learning or mixed supervised learning.

5. The digital enhanced Raman spectroscopy method according to claim 2, characterized in that: The deep learning network is a convolutional neural network and its variants, a generative adversarial network, a Transformer architecture or a Diffusion architecture.

6. The digital enhanced Raman spectroscopy method according to claim 1, characterized in that: Structural similarity, root mean square error, and peak signal-to-noise ratio are used to quantitatively evaluate the accuracy and signal-to-noise ratio of digitally enhanced Raman signals.

Images