A Raman spectroscopy-based serum data classification method and system

Abstract

The present invention discloses a serum data classification method and system based on Raman spectroscopy, which relates to the technical field of medical spectral analysis. The present invention adopts TransU-Net and a frequency-domain attention mechanism, combines time-domain and frequency-domain features, and significantly improves the fluorescence noise separation ability; constructs a dual decoder to retain key Raman peak information and effectively improves the classification accuracy; constructs a metabolite-Raman peak relationship graph, and through a multi-level classification model, accurately identifies target tumor-related components, reduces manual intervention, solves problems such as serious fluorescence interference, insufficient feature extraction, and low classification accuracy in traditional Raman spectroscopy analysis, and provides efficient and reliable technical support for early tumor diagnosis and precision medicine.

Description

Technical Field

[0001] The present invention relates to the technical field of medical spectral analysis, and particularly to a method and system for classifying serum data based on Raman spectroscopy. Background Art

[0002] Raman spectroscopy technology has been widely used in medical spectral analysis due to its non-destructive and rapid detection advantages. However, in practical applications, it faces four major problems: fluorescence interference, limitations in feature extraction, singularity of classification models, and lack of metabolite correlation analysis, which limit its application effect in early tumor diagnosis. Traditional classification methods have significant deficiencies in aspects such as fluorescence removal effect, feature expression ability, model generalization ability, and metabolic network relationship modeling. They cannot effectively separate fluorescence noise from Raman signals, resulting in inaccurate feature extraction, low classification accuracy, poor robustness, and inability to provide biological explanations, making it difficult to meet the analysis requirements of complex clinical data and the rapid and effective detection requirements of medical staff for patients' serum. Summary of the Invention

[0003] The purpose of the present invention is to provide a method and system for classifying serum data based on Raman spectroscopy to improve the above technical problems.

[0004] To achieve the above invention purpose, the embodiments of the present invention provide the following technical solutions:

[0005] A method for classifying serum data based on Raman spectroscopy is provided, which includes:

[0006] Obtain the serum sample of the patient and perform sample preprocessing and Raman spectroscopy measurement to obtain the corresponding original Raman spectroscopy data;

[0007] Perform fluorescence removal processing and data enhancement on the original Raman spectroscopy data to obtain Raman spectroscopy data;

[0008] Extract features from the Raman spectroscopy data to obtain Raman data;

[0009] Input the Raman data and the original Raman spectroscopy data into a tumor recognition and classification model to obtain a tumor classification result; the tumor classification result includes tumor and non-tumor.

[0010] Further, the fluorescence removal process uses a generative adversarial network; the generative adversarial network includes a generator and a discriminator; the generator uses a TransU-Net network; the TransU-Net network includes an encoder and a decoder connected in series; the encoder includes a first convolutional block, a second convolutional block, a third convolutional block, a Transformer module, and a fourth convolutional block connected in series in sequence; the structures of the first convolutional block to the third convolutional block are exactly the same; the fourth convolutional block includes two convolutional layers; the decoder includes a Raman decoding module and a fluorescence noise decoding module in parallel; the Raman decoding module includes a first transposed convolutional block, a second transposed convolutional block, a third transposed convolutional block, and a fourth transposed convolutional block connected in series in sequence and with exactly the same structure; the fluorescence noise decoding module includes a DWConv layer, a fifth convolutional block, and a Conv5 layer connected in series in sequence; the fifth convolutional block includes two convolutional layers;

[0011] The first convolutional block includes a Conv1 layer, a Conv2 layer, a Maxpool layer, and an FFT layer connected in series in sequence; the first transposed convolutional block includes a Conv3 layer, a Conv4 layer, an Upsampling layer, and an IFFT layer connected in series in sequence;

[0012] The Transformer module includes N Transformer layers connected in series; each Transformer layer includes a Layer Norm1 layer, an MSA layer, a Layer Norm2 layer, and an MLP layer connected in series in sequence.

[0013] Further, the training process of the generative adversarial network includes:

[0014] Obtain the original Raman spectrum training data and its labels;

[0015] Input the original Raman spectrum training data into the encoder, and process it through the first convolutional block to the third convolutional block in sequence, and respectively output the first spectral encoding data to the third spectral encoding data at the corresponding levels;

[0016] Input the third spectral encoding data into the Transformer module, and output the spectral encoding key data;

[0017] Input the spectral encoding key data into the fourth convolutional block, and output the fourth spectral encoding data;

[0018] Input the first spectral encoding data to the fourth spectral encoding data into the Raman decoding module, and output the Raman spectrum training decoded data, that is, the generated training samples of the generator;

[0019] Input the first spectral encoding data to the fourth spectral encoding data into the fluorescence noise decoding module, and obtain the fluorescence noise spectral data;

[0020] Input the generated training samples and their labels into the discriminator to obtain the corresponding discriminative training results;

[0021] Based on the discriminative training results, the generated training samples, and the fluorescence noise spectral data, calculate the loss function; the loss function includes a reconstruction loss function, an adversarial loss function, and a fluorescence supervision loss function;

[0022] Based on the loss function, adjust the weight parameters of the adversarial generation network.

[0023] Furthermore, the processing process of the first convolutional block includes:

[0024] Input the original Raman spectral training data into the Conv1 layer, and output the initial convolutional spectral training data;

[0025] Input the initial convolutional spectral training data into the Conv2 layer, and output the convolutional spectral training data;

[0026] Input the convolutional spectral training data into the Maxpool layer, and output the spectral training pooling data;

[0027] Input the spectral training pooling data into the FFT layer, and output the spectral training frequency-domain data;

[0028] Concatenate the spectral training pooling data and the spectral training frequency-domain data to obtain the first spectral encoding data.

[0029] Furthermore, the processing process of the first deconvolutional block of the Raman decoding module is:

[0030] Concatenate the third spectral encoding data and the fourth spectral encoding data to obtain the spectral encoding concatenated data;

[0031] Input the spectral encoding concatenated data into the Conv3 layer, and output the first convolutional spectral decoding data;

[0032] Input the first convolutional spectral decoding data into the Conv4 layer, and output the second convolutional spectral decoding data;

[0033] Input the second convolutional spectral decoding data into the Upsampling layer, and output the spectral sampling decoding data;

[0034] Input the spectral sampling decoding data into the IFFT layer, and output the spectral time-domain decoding data;

[0035] Concatenate the spectral time-domain decoding data and the spectral sampling decoding data, and output the first spectral decoding data.

[0036] Furthermore, the processing process of the first Transformer layer in the Transformer module includes:

[0037] Input the third spectral encoded data into the Layer Norm1 layer, and output the normalized third spectral encoded data;

[0038] Input the normalized third spectral encoded data into the MSA layer, and output the spectral encoded tensor data;

[0039] Concatenate the spectral encoded tensor data and the third spectral encoded data to obtain the spectral encoded tensor concatenated data;

[0040] Input the spectral encoded tensor concatenated data into the Layer Norm2 layer, and output the normalized spectral encoded tensor concatenated data;

[0041] Input the normalized spectral encoded tensor concatenated data into the MLP layer, and output the first initial key data of spectral encoding;

[0042] Concatenate the spectral encoded tensor concatenated data and the first initial key data of spectral encoding to obtain the first key data of spectral encoding.

[0043] Further, the tumor recognition and classification model includes an initial classification module and a final classification module; the initial classification module includes a parallel XGBoost model and a 1D-ResNet model; the final classification module includes a three-layer graph convolutional network, a self-attention layer, and a classification layer connected in series in sequence; the classification layer includes a pooling layer and a fully connected classification layer connected in series.

[0044] Further, the input of the Raman data and the original Raman spectral data into the tumor recognition and classification model to obtain the tumor classification result includes:

[0045] Input the Raman data into the XGBoost model, and output the first initial classification and recognition result;

[0046] Input the Raman data and the original Raman spectral data into the 1D-ResNet model, and output the second initial classification and recognition result;

[0047] Based on the first initial classification and recognition result and the second initial classification and recognition result, calculate the initial tumor classification result through a weighted voting mechanism;

[0048] Based on the initial tumor classification result, obtain the corresponding metabolite and construct the corresponding relationship graph;

[0049] Input the Raman data corresponding to the initial tumor classification result and the relationship graph into the final classification module, and output the tumor classification result.

[0050] Further, the obtaining the corresponding metabolite based on the initial tumor classification result and constructing the corresponding relationship graph includes:

[0051] Based on the initial tumor classification result, obtain the corresponding metabolite;

[0052] Perform peak position matching on the metabolite and the initial tumor classification result to obtain the corresponding peak position matching result;

[0053] Based on the peak position matching result, use the metabolite interaction network to screen relevant pathways;

[0054] Take the Raman peaks corresponding to the metabolite and the initial tumor classification result as nodes, and take the relevant pathways and the mapping relationship between the Raman peaks corresponding to the initial tumor classification result and the metabolite as edges to construct metabolite nodes, Raman peak nodes, metabolite-Raman peak edges and metabolite-metabolite edges, and obtain the corresponding relationship graph.

[0055] A serum data classification system based on Raman spectroscopy is provided, which includes:

[0056] A Raman spectroscopy acquisition and processing module for obtaining the serum sample of the patient and performing sample pretreatment and Raman spectroscopy measurement to obtain the corresponding original Raman spectroscopy data;

[0057] A Raman spectroscopy enhancement module for performing fluorescence removal processing and data enhancement on the original Raman spectroscopy data to obtain Raman spectroscopy data;

[0058] A Raman feature extraction module for extracting features from the Raman spectroscopy data to obtain Raman data;

[0059] A tumor classification module for inputting the Raman data and the original Raman spectroscopy data into a tumor recognition and classification model to obtain a tumor classification result.

[0060] The beneficial effects of the present invention are:

[0061] The present invention adopts TransU-Net and a frequency domain attention mechanism, combines time domain and frequency domain features, and significantly improves the fluorescence noise separation ability; constructs a dual decoder to retain key Raman peak information and effectively improves the classification accuracy; constructs a metabolite-Raman peak relationship graph, and through a multi-level classification model, accurately identifies target tumor-related components, reduces manual intervention, solves the problems of serious fluorescence interference, insufficient feature extraction, low classification accuracy, etc. in traditional Raman spectroscopy analysis, and provides efficient and reliable technical support for precision medicine. Description of the Drawings

[0062] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, without creative efforts, other related drawings can also be obtained based on these drawings.

[0063] Figure 1 It is the flowchart of the method in the embodiment of the present invention;

[0064] Figure 2 It is the structural diagram of the TransU-Net network in the embodiment of the present invention;

[0065] Figure 3 It is the structural diagram of the first convolutional block and the first transposed convolutional block in the embodiment of the present invention;

[0066] Figure 4 It is the structural diagram of the Transformer module in the embodiment of the present invention;

[0067] Figure 5 It is the structural diagram of the tumor recognition and classification model in the embodiment of the present invention;

[0068] Figure 6 It is the Raman spectrogram corresponding to IHC0 and IHC3 in the embodiment of the present invention;

[0069] Figure 7 It is the system structural diagram in the embodiment of the present invention. Detailed implementation manners

[0070] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Usually, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed present invention, but only represents the selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative efforts belong to the scope of protection of the present invention.

[0071] Please refer to Figure 1 , a method for classifying serum data based on Raman spectroscopy provided in this embodiment includes:

[0072] S1. Obtain the serum sample of the patient and perform sample pretreatment and Raman spectroscopy measurement to obtain the corresponding original Raman spectral data;

[0073] S2. Perform fluorescence removal processing and data enhancement on the original Raman spectroscopy data to obtain Raman spectroscopy data;

[0074] In Raman spectroscopy analysis, fluorescence background interference is a common problem. Since Raman scattering signals are usually very weak, while fluorescence background signals may be several orders of magnitude stronger than Raman scattering signals, this will cause the Raman scattering signals to be masked, making it difficult to identify characteristic peaks, thus affecting the quality of the data and the accuracy of subsequent analysis. In addition, the original Raman spectroscopy data itself has high-dimensional and complex non-linear relationships, which increases the difficulty of feature extraction and classification. Traditional feature extraction methods often cannot fully express these complex features, especially when the number of samples is limited, the model is prone to overfitting problems. Therefore, the S2 includes:

[0075] S2-1. Use a generative adversarial network to perform fluorescence removal processing on the original Raman spectroscopy data to obtain processed Raman spectroscopy data;

[0076] The generative adversarial network includes a generator and a discriminator; as Figure 2 shown, the generator uses a TransU-Net network; the TransU-Net network includes an encoder and a decoder connected in series; the encoder includes a first convolutional block, a second convolutional block, a third convolutional block, a Transformer module, and a fourth convolutional block connected in series in sequence; the structures of the first convolutional block to the third convolutional block are exactly the same; the fourth convolutional block includes two convolutional layers; the decoder includes a Raman decoding module and a fluorescence noise decoding module in parallel; the Raman decoding module includes a first transposed convolutional block, a second transposed convolutional block, a third transposed convolutional block, and a fourth transposed convolutional block connected in series in sequence and with exactly the same structure; the fluorescence noise decoding module includes a DWConv layer, a fifth convolutional block, and a Conv5 layer connected in series in sequence; the fifth convolutional block includes two convolutional layers;

[0077] as Figure 3 shown, the first convolutional block includes a Conv1 layer, a Conv2 layer, a Maxpool layer (maximum pooling layer), and an FFT layer (Fourier transform layer) connected in series in sequence; the first transposed convolutional block includes a Conv3 layer, a Conv4 layer, an Upsampling layer (upsampling layer), and an IFFT layer (inverse Fourier transform layer) connected in series in sequence;

[0078] as Figure 4 shown, the Transformer module includes N Transformer layers connected in series; each Transformer layer includes a Layer Norm1 layer, an MSA (multi-head attention mechanism) layer, a Layer Norm2 layer, and an MLP layer (multi-layer perceptron) connected in series in sequence. In this embodiment, N is 3.

[0079] The training process of the adversarial generative network includes:

[0080] S2-1-1. Obtain the original Raman spectrum training data and its label; the label is the original Raman spectrum training data after removing fluorescence noise.

[0081] S2-1-2. Input the original Raman spectrum training data into the encoder, and process it through the first convolutional block to the third convolutional block in sequence, and respectively output the first spectral encoding data to the third spectral encoding data at the corresponding levels. Specifically, input the original Raman spectrum training data into the first convolutional block to obtain the first spectral encoding data; input the first spectral encoding data into the second convolutional block to obtain the second spectral encoding data; input the second spectral encoding data into the third convolutional block to obtain the third spectral encoding data.

[0082] Please refer to Figure 3 for the processing process of the first convolutional block:

[0083] S2-1-2-1. Input the original Raman spectrum training data into the Conv1 layer, and output the initial convolutional spectrum training data.

[0084] S2-1-2-2. Input the initial convolutional spectrum training data into the Conv2 layer, and output the convolutional spectrum training data.

[0085] S2-1-2-3. Input the convolutional spectrum training data into the Maxpool layer, and output the spectral training pooling data.

[0086] S2-1-2-4. Input the spectral training pooling data into the FFT layer, and output the spectral training frequency domain data.

[0087] S2-1-2-5. Concatenate the spectral training pooling data and the spectral training frequency domain data to obtain the first spectral encoding data.

[0088] The second convolutional block and the third convolutional block adopt the same method as S2-1-2-1 to S2-1-2-5, so they will not be elaborated here.

[0089] This method introduces a frequency domain attention mechanism (fast Fourier transform) in the first convolutional block to the third convolutional block, enhances the ability of the adversarial generative network to distinguish fluorescence (low frequency) and Raman peaks (high frequency), avoids the limitation of traditional convolution that only focuses on local time domain features, and concatenates the spectral training pooling data and the spectral training frequency domain data to retain multi-scale information.

[0090] S2-1-3. Input the third spectral encoding data into the Transformer module, and use N Transformer layers to extract and capture global dependencies and perform feature transformation, and output the spectral encoding key data.

[0091] Please refer to Figure 4 , the processing process of the first Transformer layer in the Transformer module includes:

[0092] S2-1-3-1: Input the third spectral encoding data into the Layer Norm1 layer, and output the normalized third spectral encoding data;

[0093] S2-1-3-2: Input the normalized third spectral encoding data into the MSA (Multi-Head Attention Mechanism) layer, capture the dependencies between different positions, splice the outputs of each attention head, and output the spectral encoding tensor data;

[0094] S2-1-3-3: Splice the spectral encoding tensor data and the third spectral encoding data to obtain the spectral encoding tensor splicing data;

[0095] S2-1-3-4: Input the spectral encoding tensor splicing data into the Layer Norm2 layer, and output the normalized spectral encoding tensor splicing data;

[0096] S2-1-3-5: Input the normalized spectral encoding tensor splicing data into the MLP (Multi-Layer Perceptron) layer for non-linear transformation to enhance the feature expression ability, and output the first spectral encoding initial key data;

[0097] S2-1-3-6: Splice the spectral encoding tensor splicing data and the first spectral encoding initial key data to obtain the first spectral encoding key data.

[0098] The processing processes of the Transformer layers of other layers are the same as those of the first Transformer layer, and the output of the last Transformer layer is the spectral encoding key data.

[0099] S2-1-4: Input the spectral encoding key data into the fourth convolutional block, and output the fourth spectral encoding data;

[0100] S2-1-5: Input the first spectral encoding data to the fourth spectral encoding data into the Raman decoding module, and output the Raman spectrum training decoding data, that is, the generated training samples of the generator (pure Raman signal false data without fluorescence noise);

[0101] The said S2-1-5 includes:

[0102] S2-1-5-1: Splice the third spectral encoding data and the fourth spectral encoding data to obtain the spectral encoding splicing data;

[0103] S2-1-5-2. Input the spectral encoding splicing data into the first deconvolution block, and output the first spectral decoding data;

[0104] Please refer to Figure 3 , where S2-1-5-2 includes:

[0105] Input the spectral encoding splicing data into the Conv3 layer, and output the first convolutional spectral decoding data;

[0106] Input the first convolutional spectral decoding data into the Conv4 layer, and output the second convolutional spectral decoding data;

[0107] Input the second convolutional spectral decoding data into the Upsampling layer, and output the spectral sampling decoding data;

[0108] Input the spectral sampling decoding data into the IFFT layer, and output the spectral time-domain decoding data;

[0109] Concatenate the spectral time-domain decoding data and the spectral sampling decoding data, and output the first spectral decoding data.

[0110] The second to fourth deconvolution blocks adopt the same method as S2-1-5-2.

[0111] S2-1-5-3. Concatenate the first spectral decoding data and the second spectral encoding data to obtain the first spectral decoding splicing data;

[0112] S2-1-5-4. Input the first spectral decoding splicing data into the second deconvolution block, and output the second spectral decoding data;

[0113] S2-1-5-5. Concatenate the second spectral decoding data and the first spectral encoding data to obtain the second spectral decoding splicing data;

[0114] S2-1-5-6. Input the second spectral decoding splicing data into the third deconvolution block, and output the third spectral decoding data;

[0115] S2-1-5-7. Input the third spectral decoding data into the fourth deconvolution block, and output the Raman spectrum training decoding data, that is, the processed Raman spectrum training data (generate training samples);

[0116] S2-1-6. Input the first to fourth spectral encoding data into the fluorescence noise decoding module to obtain the fluorescence noise spectral data;

[0117] Where S2-1-6 includes:

[0118] S2-1-6-1. Concatenate the third spectral encoded data and the fourth spectral encoded data to obtain spectral encoded concatenated data;

[0119] S2-1-6-2. Input the spectral encoded concatenated data into the DWConv layer and output the first fluorescence noise decoded data;

[0120] S2-1-6-3. Concatenate the first fluorescence noise decoded data and the second spectral encoded data to obtain the first fluorescence noise concatenated decoded data;

[0121] S2-1-6-4. Input the first fluorescence noise concatenated decoded data into the fifth convolutional block and output the second fluorescence noise decoded data;

[0122] S2-1-6-5. Concatenate the second fluorescence noise decoded data and the first spectral encoded data to obtain the second fluorescence noise concatenated decoded data;

[0123] S2-1-6-6. Input the second fluorescence noise concatenated decoded data into the Conv5 layer and output the fluorescence noise spectral data.

[0124] S2-1-7. Input the generated training samples and their labels into the discriminator to obtain the corresponding discriminative training results;

[0125] S2-1-8. Calculate the loss function based on the discriminative training results, the generated training samples, and the fluorescence noise spectral data; the loss function includes a reconstruction loss function, an adversarial loss function, and a fluorescence supervision loss function;

[0126] Reconstruction loss function Adopt the L1 norm loss function, adversarial loss function Adopt the WGAN-GP loss function, fluorescence supervision loss function Adopt the MSE loss function, thus, the loss function The corresponding formula is:

[0127] ;

[0128] Among them, , , represent coefficients, , , The sum of is 1.

[0129] S2-1-9. Adjust the weight parameters of the adversarial generative network based on the loss function.

[0130] In actual use, only the generator performs fluorescence removal on the input original Raman spectral data to obtain the processed Raman spectral data.

[0131] This method constructs an adversarial generation network by using TransU-Net and Fourier transform, designs a dual decoder, extracts the frequency-domain and time-domain features of Raman spectral data, which is beneficial to separating Raman scattering signals and fluorescence noise, highly preserves deep features, and can extract key characteristic peaks.

[0132] S2-2. Perform data enhancement on the processed Raman spectral data through operations such as baseline correction and smoothing filtering to obtain Raman spectral data.

[0133] S3. Extract features from the Raman spectral data to obtain Raman data; the feature extraction method used in S3 can adopt the wavelet transform method or use the CNN model for feature extraction, which can effectively extract the frequency-domain / spatial-domain features in the Raman spectrum and is beneficial for subsequent multi-level classification.

[0134] S4. Input the Raman data and the original Raman spectral data into the tumor recognition and classification model to obtain the tumor classification result; the classification result includes tumor and non-tumor, and can also include the status of a certain molecular subtype.

[0135] As Figure 5 shown, the tumor recognition and classification model includes an initial classification module and a final classification module; the initial classification module includes a parallel XGBoost model and a 1D-ResNet model; the final classification module includes a three-layer graph convolutional network (GCN network), a self-attention layer (Self-attention), and a classification layer connected in series in sequence; the classification layer includes a pooling layer and a fully connected classification layer connected in series.

[0136] S4 includes:

[0137] S4-1. Input the Raman data into the XGBoost model and output the first initial classification and recognition result.

[0138] S4-2. Input the Raman data and the original Raman spectral data into the 1D-ResNet model and output the second initial classification and recognition result.

[0139] S4-3. Based on the first initial classification and recognition result and the second initial classification and recognition result, calculate the initial tumor classification result through a weighted voting mechanism.

[0140] The training process of the initial classification module is as follows:

[0141] S4-3-1. Obtain Raman training data and original Raman spectral training data.

[0142] S4-3-2. Input the Raman training data into the XGBoost model and output the first initial recognition and classification training result.

[0143] S4-3-3. Input the Raman training data and the original Raman spectrum data into the 1D-ResNet model, and output the second initial recognition and classification training result;

[0144] S4-3-4. Based on the first initial recognition and classification training result and the second initial recognition and classification training result, calculate the initial classification training result through a weighted voting mechanism;

[0145] S4-3-5. Based on the initial classification training result, adjust the weight parameters of the initial classification module.

[0146] S4-4. Based on the initial tumor classification result, obtain the corresponding metabolites and construct the corresponding relationship graph;

[0147] The S4-4 includes:

[0148] S4-4-1. Based on the initial tumor classification result, obtain the corresponding metabolites;

[0149] S4-4-2. Perform peak position matching on the metabolites and the initial tumor classification result to obtain the corresponding peak position matching result; Use the Bio-Rad KnowItAll database to map the Raman peaks corresponding to the initial tumor classification result to the metabolites, such as 1660 cm -1 →β-sheet protein;

[0150] S4-4-3. Based on the peak position matching result, use the metabolite interaction network (such as the STRING database) to screen relevant pathways; For example, when detecting breast cancer, the relevant pathways are the HER2 signaling pathway, glycolysis, etc.; The relevant pathways refer to the interaction relationships between each pair of metabolites;

[0151] S4-4-4. Use the metabolites and the Raman peaks corresponding to the initial tumor classification result as nodes, and use the relevant pathways and the mapping relationship between the Raman peaks corresponding to the initial tumor classification result and the metabolites as edges to construct metabolite nodes, Raman peak nodes, metabolite-Raman peak edges, and metabolite-metabolite edges to obtain the corresponding relationship graph.

[0152] S4-5. Input the Raman data and the relationship graph corresponding to the initial tumor classification result into the final classification module, and output the tumor classification result.

[0153] Therefore, the training process of the graph convolutional network is:

[0154] S4-5-1. Obtain the Raman training data and the relationship training graph corresponding to the initial classification training result;

[0155] S4-5-2. Input the Raman training data and the relationship training graph corresponding to the initial classification training results into the first-layer graph convolutional network, and output the first-order adjacency training information;

[0156] S4-5-3. Input the first-order adjacency training information into the second-layer graph convolutional network, and output the second-order adjacency training information;

[0157] S4-5-4. Input the second-order adjacency training information into the third-layer graph convolutional network, and output the third-order adjacency training information;

[0158] S4-5-5. Input the third-order adjacency training information into the self-attention layer, and use the attention mechanism to extract key features to obtain the key adjacency training features;

[0159] S4-5-6. Input the key adjacency training features into the classification layer, and output the final classification training results;

[0160] S4-5-7. Based on the final classification training results, adjust the weight parameters of the target tumor recognition classification model.

[0161] This method realizes the rapid localization of breast cancer-related components from complex Raman spectra through the efficient rough screening of the first-level classification model and the accurate recognition of the second-level classification model.

[0162] Taking breast cancer as an example, select the serum sample data of a certain hospital for experiments, classify the Raman peaks corresponding to HER-2 in the serum samples of patients, and distinguish the Raman spectra corresponding to the two categories of IHC0 and IHC3.

[0163] The Raman data corresponding to IHC0 and IHC3 are as Figure 6 shown. The Raman peaks of the two groups basically coincide, but there are inter-group differences in the intensity of the Raman scattering signals. The Raman peaks of the two groups are related to the glucose component in the serum sample at the wavenumbers of 824 cm -1 , 849 cm -1 and 869 cm -1 ; they are related to the lipid component in the serum sample at the wavenumbers of 1102 cm -1 , 1095 cm -1 and 1097 cm -1 ; they are related to the lipid component and protein component in the serum sample at the wavenumbers of 1451 cm -1 , 1461 cm -1 , 1463 cm -1 , 1465 cm -1 , 1468 cm -1 and 1479 cm -1 ; they are related to the lipid component and protein component in the serum sample at the wavenumbers of 1308 cm -1 and 1305 cm-1 The Raman peaks at [position] are related to the carotenoid components in the serum sample; at wavenumbers of 1655 cm -1 , 1661 cm -1 , 1665 cm -1 , 1672 cm -1 , 1677 cm -1 and 1679 cm -1 the Raman peaks at [position] are related to the phospholipid components and fatty acid components in the serum sample. The differences in Raman spectra between the two categories of IHC0 and IHC3 reflect the changes in biochemical components, providing a basis for distinguishing different types of IHC classification.

[0164] Calculate the evaluation metrics of this method (MSA-FFT-U Net-XGBoost-ResNet-GCN), Method 1 (MSA-U Net-XGBoost-ResNet-GCN) without introducing the multi-head attention mechanism, and Method 2 (FFT-U Net-XGBoost-ResNet-GCN) without introducing the frequency-domain attention mechanism.

[0165] The evaluation metrics selected are accuracy, AUC score, and F1 score, and the corresponding values are shown in Table 1.

[0166] Table 1

[0167]

[0168] As can be seen from Table 1, the accuracy, AUC score, and F1 score of this method are 0.967, 0.970, and 0.969 respectively; the accuracy, AUC score, and F1 score of Method 1 are 0.905, 0.908, and 0.893 respectively; the accuracy, AUC score, and F1 score of Method 2 are 0.882, 0.895, and 0.888 respectively. Therefore, the accuracy, AUC score, and F1 score of this method are all higher than those of Method 1 and Method 2, further proving that introducing the attention mechanism in this method can effectively improve the feature processing ability and classification accuracy of the model.

[0169] As Figure 7 shown, a serum data classification system based on Raman spectroscopy includes:

[0170] A Raman spectroscopy acquisition and processing module, which is used to obtain the serum sample of the patient and perform sample preprocessing and Raman spectroscopy measurement to obtain the corresponding original Raman spectroscopy data;

[0171] A Raman spectroscopy enhancement module, which is used to perform fluorescence removal processing and data enhancement on the original Raman spectroscopy data to obtain Raman spectroscopy data;

[0172] A Raman feature extraction module, which is used to extract features from Raman spectroscopy data to obtain Raman data;

[0173] A tumor classification module, which is used to input the Raman data and the original Raman spectroscopy data into a tumor recognition and classification model to obtain a tumor classification result.

[0174] In summary, the present invention combines time-domain and frequency-domain features through an adversarial generative network (TransU-Net + frequency-domain attention mechanism), significantly improving the fluorescence noise separation ability. At the same time, a dual-decoder design (Raman signal decoding + fluorescence noise decoding) is adopted to retain key Raman peak information. In terms of classification, parallel XGBoost (feature importance analysis) and 1D-ResNet (deep feature extraction) improve the initial classification accuracy through weighted voting. At the same time, a "metabolite-Raman peak" relationship graph is constructed, combined with a graph convolutional network (GCN) and a self-attention mechanism, to accurately identify target tumor-related components. In addition, a secondary classification model (initial classification + target tumor recognition) realizes progressive analysis from rough screening to fine classification, and outputs probabilities and key metabolite-Raman peak pairs. The system adopts a modular design and step-by-step processing (data acquisition → fluorescence removal and enhancement → feature extraction → classification) to improve scalability and maintainability. At the same time, an adversarial generative network and a multi-level classification model are integrated to reduce manual intervention and are applicable to the rapid analysis of large-scale clinical data.

[0175] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, and all of them should be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A method for classifying serum data based on Raman spectroscopy, characterized in that, Including: Obtain the serum sample of the patient, perform sample preprocessing and Raman spectroscopy measurement to obtain the corresponding original Raman spectroscopy data; Perform fluorescence removal processing and data enhancement on the original Raman spectroscopy data to obtain Raman spectroscopy data; Extract features from the Raman spectroscopy data to obtain Raman data; Input the Raman data and the original Raman spectroscopy data into the tumor recognition and classification model to obtain the tumor classification result; the tumor classification result includes tumors and non-tumors; the tumor recognition and classification model includes an initial classification module and a final classification module; the initial classification module includes a parallel XGBoost model and a 1D-ResNet model; the final classification module includes a three-layer graph convolutional network, a self-attention layer, and a classification layer connected in series in sequence; the classification layer includes a pooling layer and a fully connected classification layer connected in series.

2. The method for classifying serum data based on Raman spectroscopy according to claim 1, wherein The fluorescence removal processing uses an adversarial generation network; the adversarial generation network includes a generator and a discriminator; the generator uses a TransU-Net network; the TransU-Net network includes an encoder and a decoder connected in series; the encoder includes a first convolutional block, a second convolutional block, a third convolutional block, a Transformer module, and a fourth convolutional block connected in series in sequence; the structures of the first convolutional block to the third convolutional block are exactly the same; the fourth convolutional block includes two convolutional layers; the decoder includes a Raman decoding module and a fluorescence noise decoding module connected in parallel; the Raman decoding module includes a first transposed convolutional block, a second transposed convolutional block, a third transposed convolutional block, and a fourth transposed convolutional block connected in series in sequence and with exactly the same structure; the fluorescence noise decoding module includes a DWConv layer, a fifth convolutional block, and a Conv5 layer connected in series in sequence; the fifth convolutional block includes two convolutional layers; The first convolutional block includes a Conv1 layer, a Conv2 layer, a Maxpool layer, and an FFT layer connected in series in sequence; the first transposed convolutional block includes a Conv3 layer, a Conv4 layer, an Upsampling layer, and an IFFT layer connected in series in sequence; The Transformer module includes N Transformer layers connected in series; each Transformer layer includes a Layer Norm1 layer, an MSA layer, a Layer Norm2 layer, and an MLP layer connected in series in sequence.

3. A method for classifying serum data based on Raman spectroscopy according to claim 2, characterized in that, The training process of the adversarial generation network includes: Obtain the original Raman spectroscopy training data and its labels; Input the original Raman spectroscopy training data into the encoder, and process it through the first convolutional block to the third convolutional block in sequence, and respectively output the first spectral encoded data to the third spectral encoded data at the corresponding levels; Input the third spectral encoded data into the Transformer module, and output the spectral encoded key data; Input the spectral encoded key data into the fourth convolutional block, and output the fourth spectral encoded data; Input the first spectral encoded data to the fourth spectral encoded data into the Raman decoding module, and output the Raman spectroscopy training decoded data, that is, the generated training samples of the generator; Input the first spectral encoded data to the fourth spectral encoded data into the fluorescence noise decoding module to obtain the fluorescence noise spectral data; Input the generated training samples and their labels into the discriminator to obtain the corresponding discriminative training results; Calculate the loss function based on the discriminative training results, the generated training samples, and the fluorescence noise spectral data; the loss function includes a reconstruction loss function, an adversarial loss function, and a fluorescence supervision loss function; Adjust the weight parameters of the adversarial generation network based on the loss function.

4. A method for classifying serum data based on Raman spectroscopy according to claim 3, characterized in that The processing process of the first convolutional block includes: Input the original Raman spectral training data into the Conv1 layer, and output the initial convolutional spectral training data; Input the initial convolutional spectral training data into the Conv2 layer, and output the convolutional spectral training data; Input the convolutional spectral training data into the Maxpool layer, and output the spectral training pooled data; Input the spectral training pooled data into the FFT layer, and output the spectral training frequency-domain data; Concatenate the spectral training pooled data and the spectral training frequency-domain data to obtain the first spectral encoded data.

5. A method for classifying serum data based on Raman spectroscopy according to claim 3, characterized in that, The processing process of the first deconvolutional block of the Raman decoding module is: Concatenate the third spectral encoded data and the fourth spectral encoded data to obtain the spectral encoded concatenated data; Input the spectral encoded concatenated data into the Conv3 layer, and output the first convolutional spectral decoded data; Input the first convolutional spectral decoded data into the Conv4 layer, and output the second convolutional spectral decoded data; Input the second convolutional spectral decoded data into the Upsampling layer, and output the spectral sampled decoded data; Input the spectral sampled decoded data into the IFFT layer, and output the spectral time-domain decoded data; Concatenate the spectral time-domain decoded data and the spectral sampled decoded data, and output the first spectral decoded data.

6. A method for classifying serum data based on Raman spectroscopy according to claim 3, characterized in that, The processing process of the first Transformer layer in the Transformer module includes: Input the third spectral encoded data into the Layer Norm1 layer, and output the normalized third spectral encoded data; Input the normalized third spectral encoded data into the MSA layer, and output the spectral encoded tensor data; Concatenate the spectral encoded tensor data and the third spectral encoded data to obtain the spectral encoded tensor concatenated data; Input the spectral encoded tensor concatenated data into the Layer Norm2 layer, and output the normalized spectral encoded tensor concatenated data; Input the normalized spectral encoded tensor concatenated data into the MLP layer, and output the first spectral encoded initial key data; Concatenate the spectral encoded tensor concatenated data and the first spectral encoded initial key data to obtain the first spectral encoded key data.

7. A method for classifying serum data based on Raman spectroscopy according to claim 1, characterized in that The process of inputting Raman data and original Raman spectral data into the tumor recognition and classification model to obtain the tumor classification result includes: Input the Raman data into the XGBoost model, and output the first initial classification and recognition result; Input the Raman data and the original Raman spectral data into the 1D-ResNet model, and output the second initial classification and recognition result; Calculate the initial tumor classification result through a weighted voting mechanism based on the first initial classification and recognition result and the second initial classification and recognition result; Based on the initial tumor classification result, obtain the corresponding metabolites and construct the corresponding relationship graph; Input the Raman data and the relationship graph corresponding to the initial tumor classification result into the final classification module, and output the tumor classification result.

8. A method for classifying serum data based on Raman spectroscopy according to claim 7, characterized in that, The step of "Based on the initial tumor classification result, obtain the corresponding metabolites and construct the corresponding relationship graph" includes: Based on the initial tumor classification result, obtain the corresponding metabolites; Perform peak position matching on the metabolites and the initial tumor classification result to obtain the corresponding peak position matching result; Based on the peak position matching result, use the metabolite interaction network to screen relevant pathways; Take the metabolites and the Raman peaks corresponding to the initial tumor classification result as nodes, and take the relevant pathways and the mapping relationship between the Raman peaks corresponding to the initial tumor classification result and the metabolites as edges to construct metabolite nodes, Raman peak nodes, metabolite-Raman peak edges and metabolite-metabolite edges, and obtain the corresponding relationship graph.

9. A serum data classification system based on Raman spectroscopy, which is used to implement a serum data classification method based on Raman spectroscopy according to any one of claims 1 to 8, characterized in that, It includes: A Raman spectroscopy acquisition and processing module, which is used to obtain the serum sample of the patient and perform sample pretreatment and Raman spectroscopy measurement to obtain the corresponding original Raman spectroscopy data; A Raman spectroscopy enhancement module, which is used to perform fluorescence removal processing and data enhancement on the original Raman spectroscopy data to obtain Raman spectroscopy data; A Raman feature extraction module, which is used to extract features from the Raman spectroscopy data to obtain Raman data; A tumor classification module, which is used to input the Raman data and the original Raman spectroscopy data into the tumor recognition and classification model to obtain the tumor classification result; the tumor recognition and classification model includes an initial classification module and a final classification module; the initial classification module includes a parallel XGBoost model and a 1D-ResNet model; the final classification module includes a three-layer graph convolutional network, a self-attention layer and a classification layer connected in series in sequence; the classification layer includes a pooling layer and a fully connected classification layer connected in series.

Images