Training method of tumor classification prediction model based on SERS (Surface Enhanced Raman Scattering) signal diagram and application of training method

Abstract

The invention belongs to the technical field of biological information, and particularly relates to a training method of a tumor classification prediction model based on SERS (Surface Enhanced Raman Scattering) data and application thereof, and the training method of the tumor classification prediction model based on SERS signal diagrams comprises the following steps: acquiring SERS signal diagrams of different subtype tumor cells subjected to SERS test as original data; all SERS signal diagrams in the original data are sequentially subjected to preprocessing and data enhancement processing, and the original data after preprocessing and data enhancement processing and labels of corresponding subtypes of the original data are jointly used as training samples; the deep learning model is trained through the training sample, the tumor classification prediction model is obtained after training, the tumor classification prediction model after training is particularly outstanding in large-scale data, and compared with other existing image classification algorithms, the classification precision is remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the field of bioinformatics technology, and particularly relates to a training method and application of a tumor classification prediction model based on SERS signal maps. Background Art

[0002] In clinical diagnosis, tissue biopsy, i.e., pathological examination, is the gold standard for cancer diagnosis and is of great significance for prognosis evaluation and treatment guidance. Clinically, hematoxylin and eosin (H&E) staining or immunohistochemical staining are usually used for histopathological detection. However, the staining process of these methods is time-consuming and requires pathologists to have a high level of professional knowledge, which limits the application of histopathology. Therefore, there is an urgent need for advanced technologies to achieve rapid and non-invasive diagnosis of cancer types and their subtypes.

[0003] Surface-enhanced Raman scattering (SERS) technology has ultra-high sensitivity, unique molecular fingerprint spectral recognition, rapid and non-destructive analysis, and has significant advantages in disease screening and diagnosis. SERS has been explored for analyzing tissue samples because it can detect subtle molecular changes in the presence of disease. This non-invasive technology can also be applied in real time, and faster diagnostic results can be obtained compared with traditional methods. Although SERS-based tissue analysis has potential in cancer diagnosis, there are still some challenges in accurately analyzing and classifying the SERS spectra of tissues. Tumor tissues are highly heterogeneous, containing various cell components, extracellular matrix proteins, and stromal components. This complexity may lead to overlapping spectral features, making it difficult to distinguish different cancer subtypes, especially when subtle molecular changes are involved. In a clinical setting, factors such as adipose tissue, inflammation, or necrotic areas can interfere with the SERS signal, complicating data analysis. Inconsistent sample collection and preparation can further exacerbate these problems, resulting in additional noise and lower sensitivity in practical applications. Compared with direct tissue detection, cell analysis has significant advantages. The Raman spectra of cells are more uniform and stable than those of tissues, and they effectively reduce the interference of adipose tissue, thus providing more accurate molecular spectral data. So far, most existing studies are mainly based on simulated data of cell lines or laboratory environments, which are very different from the complex clinical environment and real samples. This limits the wide application of SERS technology in clinical practice. Therefore, extracting cells from clinical tissues for analysis can more accurately reflect the diversity of tumors and the real clinical situation, providing greater potential for clinical applications. SERS data exhibits high dimensionality and complexity, and its spectral signals contain rich molecular characteristic peaks. Severe overlap may occur between the spectral peaks of different substances, making it difficult to distinguish specific molecular features. CN103487425A discloses a method for discriminating cancer cells using surface-enhanced Raman spectroscopy, which mainly uses the PCA method in the analysis toolbar of SPSS software to analyze the surface-enhanced Raman spectroscopy data of cells. However, complex background noise and weak signal interference may mask valuable molecular information, rendering traditional statistical or machine learning methods ineffective in feature extraction and pattern recognition.

[0004] Therefore, deep learning, with its powerful capabilities of automatic feature extraction, noise reduction, and non-linear modeling, can better discover hidden information in data, thereby accurately classifying and analyzing complex SERS signal maps. Deep learning is a popular research topic in the field of machine learning, capable of effectively analyzing and processing data such as speech, spectra, and images, and showing strong performance in data discrimination. Combining SERS cell analysis and deep learning algorithms can more accurately reconstruct tumor features and reveal the intrinsic characteristics of tumors. However, existing deep learning models based on SERS data have technical problems such as low classification prediction accuracy, difficulty in handling background signal interference, and single data augmentation methods, resulting in their inapplicability to the classification prediction of different subtypes of tumors. Summary of the Invention

[0005] In view of the problems existing in the prior art, the present invention discloses a method for training a tumor classification prediction model based on SERS signal maps and its application.

[0006] In a first aspect, the present invention provides a method for training a tumor classification prediction model based on SERS signal maps, as Figure 1 shown in Figure b of

[0007] Obtain SERS signal maps of different subtypes of tumor cells tested by SERS as original data. In the original data, different subtypes of tumor cells include several tumor cells of two or more subtypes. There are at least 2 tumor cell samples of the same subtype, and each tumor cell sample corresponds to at least 3,000 SERS signal maps obtained through continuous testing;

[0008] Successively perform preprocessing and data augmentation processing on all SERS signal maps in the original data. The steps of the preprocessing successively include baseline removal, outlier detection, and normalization; after the original data undergoes preprocessing and data augmentation processing, it is used as a training sample together with the label of its corresponding subtype;

[0009] Train a deep learning model through the training samples, and obtain a tumor classification prediction model after training;

[0010] Among them, the deep learning model is an improved AlexNet model, as Figure 2As shown, the improved AlexNet model includes a feature extraction module and a classification module. The feature extraction module includes: multiple 1D convolutional layers, ReLU, local response normalization (LRN), and a max pooling layer. It should be noted that here ReLU, namely Rectified Linear Unit, is an activation function in deep learning. The classification module includes: a flattening layer (Flatten), fully connected layers (Linear), ReLU, and Dropout. In the feature extraction module, the multiple 1D convolutional layers sequentially include a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, and a fifth convolutional layer. The first convolutional layer is used to input a 1D SERS signal map, and after the first convolutional layer, ReLU, LRN, and a max pooling layer are sequentially connected. Connecting ReLU is used to enhance the non-linear expression ability and improve the network learning ability. Connecting LRN is used to normalize adjacent neurons and enhance the generalization ability. Connecting the max pooling layer is used to reduce the dimension. Similarly, after the second convolutional layer, ReLU, LRN, and a max pooling layer are sequentially connected. Through ReLU activation for non-linear transformation, LRN normalization, and pooling to further extract high-level features. The third convolutional layer is connected to ReLU to continue extracting deeper features. The fourth convolutional layer is connected to ReLU for further learning of local patterns. The fifth convolutional layer is sequentially connected to ReLU and a max pooling layer, which is used to reduce the data dimension, reduce the computational amount, and prepare for the fully connected layer. In the classification module, it sequentially passes through the flattening layer, the first fully connected layer, ReLU, Dropout, the second fully connected layer, ReLU, and the third fully connected layer, and the third fully connected layer outputs the final class probability distribution.

[0011] Preferably, the flattening layer flattens the data into a one-dimensional vector, and the input feature dimension is 3456. The first fully connected layer reduces the data dimension from 3456 dimensions to 2048 dimensions, uses the ReLU activation function, and Dropout selects 50%, which is used to randomly discard 50% of the neurons to prevent overfitting. The second fully connected layer maintains 2048 dimensions and uses the ReLU activation function to continue enhancing the generalization ability and reducing overfitting.

[0012] Preferably, the tumor cell sample is obtained through the following operating steps: Immerse the clinical tissue sample in PBS solution, then chop each tissue sample and transfer it together with the PBS solution to a 70μm cell filter, then use a tissue grinder to grind and filter the sample, centrifuge the filtered cell solution at 1000 rpm for 5 minutes, after removing the supernatant, add PBS to redisperse the precipitate, thereby obtaining the cell suspension of each tumor cell sample.

[0013] Preferably, the operation steps of the SERS test include: dropping the extracted tumor cells on the silver nanorod array SERS substrate and naturally drying them in a refrigerator at 4°C for 8 hours; the SERS exposure time is 3 s, the laser intensity is 50%, and it is accumulated once; more preferably, the SERS substrate for the SERS test is a silver nanorod array substrate and is subjected to plasma cleaning to reduce the background signal.

[0014] As a preferred embodiment of the present application, in the improved AlexNet model, the number of input channels of the first convolutional layer is 1 for the single-channel 1D SERS signal map, and the number of output channels is 48 to increase the number of features; the number of feature map channels of the second convolutional layer is increased to 128; the number of feature map channels of the third to fifth convolutional layers is increased to 192.

[0015] As a preferred embodiment of the present application, in the preprocessing, the operation steps for baseline removal are: fitting the baseline of the SERS signal map in the original data using the adaptive iteratively reweighted penalized least squares algorithm and removing the baseline; the operation steps for outlier detection are to detect the same sample using the Isolation Forest (Isoforest) algorithm provided by the scikit-learn machine learning library. Here, it should be noted that the same sample refers to the SERS signal map from the same tumor cell, and at least 30% of the outlier data in the SERS signal map is removed; the operation steps for normalization are: after baseline removal and outlier detection, applying the min-max normalization method to normalize the SERS spectral intensity in the SERS signal map to the range of [0, 1], and taking the Raman spectrum in the range of 600-1700 cm -1 as the output result after normalization. Here, it should be noted that the SERS signal map includes Raman shift on the abscissa and intensity information on the ordinate, and normalization only adjusts the numerical value of the ordinate intensity.

[0016] The generalization ability of the model can be improved through data augmentation processing. Therefore, as a preferred embodiment of the present application, such as Figure 1As shown in Figure b in [reference], the data augmentation processing is selected from any one, two, or three of noise augmentation, offset augmentation, or scaling augmentation to generate augmented data; each data augmentation processing generates two new augmented SERS signal maps from a single SERS signal map point, and these augmented data are sequentially added to the augmented SERS signal map list; finally, the preprocessed original SERS signal map is vertically concatenated with the augmented SERS signal maps. For example, as a better implementation manner of this application, the data augmentation processing adopts the above three augmentation processing methods. At this time, the 6 new SERS signal maps added after the data augmentation processing of each preprocessed original SERS signal map and the 1 preprocessed original SERS signal map form a final SERS signal map set including the original and augmented SERS signal maps.

[0017] The noise augmentation is to add Gaussian noise to the original spectrum to generate a new spectrum. In this application, preferably, to control the noise intensity, the noise standard deviation is set to 0.01.

[0018] The offset augmentation is to apply a random shift to the variables of the original spectrum, and generate a new spectrum by horizontally shifting the variables forward or backward along the x-axis; in this application, preferably, the offset amplitude is set to 5, allowing forward or backward offset between -5 cm -1 and 5 cm -1 forward or backward.

[0019] The scaling augmentation is to generate a new spectrum by modifying the intensity of the spectrum, specifically by generating a random scaling factor between 0.9 and 1.1, and scaling the original spectrum with the random scaling factor to generate a new spectrum.

[0020] In a second aspect, the present invention provides an application of a tumor classification prediction model based on SERS signal maps trained by the training method described in the first aspect above in tumor classification prediction, that is, a tumor classification prediction method based on SERS signal maps is provided. The prediction method includes the steps of:

[0021] Obtain the SERS signal map of the target sample tested by SERS, where the target sample is a tumor cell of an unknown subtype; sequentially perform preprocessing and data augmentation processing on the SERS signal map of the target sample tested by SERS to obtain the test data of the target sample; analyze the test data of the target sample through the tumor classification prediction model, and output the corresponding type of the target sample.

[0022] Beneficial effects: The tumor classification prediction model training method provided by this application is mainly based on the improved AlexNet model. The original AlexNet is a deep convolutional neural network with multiple convolutional layers and pooling layers. Compared with the early LeNet, AlexNet has more layers, which enables it to better learn the abstract features of images. AlexNet uses the ReLU (Rectified Linear Unit) activation function, which helps to accelerate the training process, avoid gradient vanishing, and improve the convergence speed of the network. Dropout is introduced to reduce overfitting by randomly discarding some neurons during the training process, which helps to improve the generalization ability of the model and enhance its performance on the test set. AlexNet adopts the local response normalization (LRN) layer, which enhances the model's ability to learn local features and improves its generalization performance. Compared with the original AlexNet model, the improved AlexNet model provided by this application consists of 5 convolutional layers and 3 fully connected layers. The ReLU activation function, local response normalization, and max pooling layer are used between the convolutional layers. Dropout and the ReLU activation function are used between the fully connected layers. Dropout randomly discards 50% of the neurons to prevent model overfitting. By training with a large number of real SERS signal map samples, a tumor classification model with higher prediction accuracy is obtained, which is used for the efficient and accurate classification of different types and subtypes of tumors and auxiliary clinical diagnosis, can improve the accuracy of tumor detection and the feasibility of clinical applications, and provide technical support for personalized medicine and precision treatment.

[0023] Compared with the method of using traditional principal component analysis (PCA) in the prior art, the tumor classification prediction model trained by this application is used to analyze the SERS signal maps corresponding to different tumors to predict the tumor type. The tumor classification prediction model is trained based on the deep learning algorithm. Using the deep learning algorithm can automatically learn complex non-linear features and realize the high-dimensional data analysis of SERS signals. Moreover, the deep learning algorithm adopted by the present invention is the improved AlexNet model of the inventor of this application. By combining data preprocessing technology and data augmentation technology, the influence brought by background noise and spectral background signals is effectively reduced, ensuring that the classification data is more stable and reliable. The introduced data augmentation technologies such as noise augmentation, offset augmentation, and scaling augmentation can improve the generalization ability of the model and effectively improve the recognition accuracy of the model for unknown samples.

[0024] The test results show that the tumor classification prediction model trained by this application performs particularly outstandingly on large-scale data. Compared with other existing image classification algorithms, a significant improvement in classification accuracy is achieved, up to 99.99% at most. Brief Description of the Drawings

[0025] Figure 1Schematic diagram of the steps for obtaining the SERS signal map in Embodiment 1 of the present invention and the steps of the tumor classification prediction model training method based on the SERS signal map;

[0026] Figure 2 The structure of the improved AlexNet model of the present invention;

[0027] Figure 3 The structure of the improved AlexNet model in the specific implementation manner of the present invention;

[0028] Figure 4 The classification results of the trained model of the present invention for two subtypes of samples before and after data augmentation. Specific implementation manner

[0029] The following describes the specific embodiments of the present invention in detail with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those of ordinary skill in the art, various changes can be made without departing from the gist of the present invention, and modifications or deformations without creative labor are still within the protection scope of the present invention.

[0030] Embodiment 1

[0031] Step S1, perform plasma cleaning on the prepared SERS substrate to reduce the background signal. The SERS substrate is a silver nanorod array substrate; the prepared SERS substrate is prepared according to the method described in the public literature Chunyuan S, Boyue Y, Yu Z, et al. Ultrasensitive sliver nanorods array SERS sensor for mercury ions [J]. Biosensors and Bioelectronics, 2017, 87: 59 - 65.

[0032] Step S2, the specific steps for obtaining the SERS signal map of different subtypes of tumor cells after SERS testing, that is, the steps for obtaining the original data, are as follows: Figure 1 As shown in Figure a in, extract tumor tissue cells. The specific steps are as follows: First, immerse a 0.5×0.5 cm clinical tissue sample into a centrifuge tube containing 2 mL of 1×PBS solution. Next, use ophthalmic scissors to chop each tissue sample and transfer it together with the 1×PBS solution to a 70 μm cell filter. Then use a tissue grinder to grind and filter the sample. Centrifuge the filtered cell solution at 1000 rpm for 5 minutes. After removing the supernatant, add 1 mL of 1×PBS to redisperse the precipitate, thereby obtaining the cell suspension of each tumor sample; as Figure 1As shown in Figure b in [reference], the extracted tumor cells are dropped onto the SERS substrate of the silver nanorod array and naturally dried in a refrigerator at 4°C for 8 hours. Then, 3000 SERS signal maps of each tumor cell sample are collected. The SERS test parameters for the 3000 SERS signal maps of the tumor cell sample are: exposure time 3s, laser intensity 50%, accumulated 1 time, and 3000 SERS signal maps of each tumor cell sample are collected.

[0033] Step S3, the obtained SERS signal maps are further preprocessed and data enhanced. The preprocessing steps are as follows: The adaptive iteratively reweighted penalized least squares algorithm is used for baseline fitting of the SERS signal maps to infer the baseline; The Isolation Forest (Isoforest) algorithm provided by the scikit-learn machine learning library is used to detect outliers in the SERS signal maps of the same sample, and 30% of the outlier data is removed, leaving 2000 SERS signal maps; After baseline removal and outlier exclusion, the min-max normalization method is applied to normalize the spectral intensity data to the range [0,1], and the Raman spectra in the range of 600 - 1700 cm -1 are used as data; Three data enhancement methods are applied to the preprocessed SERS signal map data: noise, shift, and scale enhancement, to generate enhanced data. Each method generates two new enhanced spectra from the single SERS signal map data respectively, and these enhanced data are sequentially added to the enhanced spectrum list; Finally, the preprocessed spectra and the enhanced spectra are vertically concatenated to form the final SERS signal map set containing 798000 SERS signal maps. Among them, the strategy of noise enhancement is to add Gaussian noise to the original spectrum to generate a new spectrum. In this study, to control the noise intensity, the noise standard deviation is set to 0.01; Shift enhancement is to apply a random shift to the variables of the original spectrum, and a new spectrum is generated by horizontally moving the variables forward or backward along the x-axis. In this study, the shift amplitude is set to 5, -5 cm - -1 and 5 cm - -1 for forward or backward shift; The strategy of scale enhancement is to modify the intensity of the spectrum to generate a new spectrum. A random scaling factor between 0.9 and 1.1 is generated, and the original spectrum is scaled with this factor to produce a new spectrum.

[0034] Step S4, construct a deep neural network model, namely the improved AlexNet model. Taking AlexNet as the backbone network, structurally, the AlexNet model includes five convolutional layers and three fully connected layers. Between the convolutional layers, the ReLU activation function, local response normalization, and max pooling layer are adopted. Between the fully connected layers, the Dropout and ReLU activation functions are used, and Dropout randomly discards 50% of the neurons to prevent model overfitting.

[0035] Figure 3 This is the structure of the improved AlexNet model of the convolutional neural network according to the embodiments of the present invention. The original AlexNet model of the convolutional neural network was proposed in 2019. Reference: Lu S, Lu Z, Zhang Y D. Pathological brain detection based on AlexNet and transfer learning[J]. Journal of computational science, 2019, 30: 41-47., and this model was developed for two-dimensional image analysis.

[0036] According to the embodiments of the present invention, the original convolutional neural model is revised as shown in Figure 3 to analyze the SERS signal map as one-dimensional data.

[0037] The original convolutional neural network model consists of five convolutional layers and three fully connected layers, where the convolutional layers are equipped with ReLU activation functions, local response normalization, and max pooling. The last three layers (FC8, Softmax, classification layer) are replaced with a binary classification structure. The parameters of the first few layers are retained from the ImageNet pre-trained model, and only the new fully connected layers are trained to improve the adaptability to small-sample data. The training uses SGDM optimization, and the final model achieves a 100% classification accuracy in the experiment;

[0038] In contrast, Figure 3The deep learning model of the embodiment of the present invention simplifies the convolutional layer. Specifically, the AlexNet model consists of 5 convolutional layers and 3 fully connected layers. The ReLU activation function, local response normalization, and max pooling layer are used between the convolutional layers. AlexNet adopts the local response normalization (LRN) layer, which enhances the model's ability to learn local features and improves its generalization performance. Dropout and the ReLU activation function are used between the fully connected layers. Dropout randomly discards 50% of the neurons to prevent the model from overfitting. Specifically, the structure of the improved AlexNet model in Embodiment 1 is as follows: features are extracted through multiple 1D convolutional layers, ReLU, local response normalization (LRN), and max pooling layers. The input channel number of the first convolutional layer is 1 for the single-channel 1D SERS signal map, and the output channel number is 48 to increase the number of features. The convolutional kernel size is 11, the stride is 4, and the padding is 2. The ReLU activation function is used to improve the feature expression ability. LRN is used to normalize the neurons in the neighborhood to improve the generalization ability. Max pooling is used with a pooling kernel of 3 and a stride of 2 for dimensionality reduction. The second convolutional layer increases the number of feature map channels to 128, the convolutional kernel size is 5, and the padding is 2 to keep the size consistent. ReLU+LRN+pooling is repeated; for the third to fifth convolutional layers, the number of feature map channels is increased to 192, and the convolutional kernel size is 3 with a padding of 1. Max pooling is used with a pooling kernel of 3 and a stride of 2 for dimensionality reduction. Classification is performed through flattening (Flatten), fully connected layer (Linear), ReLU activation function, and Dropout. The first fully connected layer reduces from 3456 dimensions to 2048 dimensions and uses ReLU activation. 50% of the neurons are randomly discarded to prevent overfitting. The second fully connected layer maintains 2048 dimensions and uses ReLU activation. The third fully connected layer outputs the final number of categories as 3.

[0039] Step S5: Input the preprocessed and data-augmented SERS signal maps into the deep learning model. For each sample, there are 14,000 corresponding SERS signal maps, and an appropriate amount of spectra are randomly selected from the total SERS signal maps of all samples, and then divided into a training set, a validation set, and a test set in a ratio of 6:2:2. The training set is used to train the deep learning model to obtain a tumor classification prediction model after training.

[0040] Test Example 1 The improved AlexNet model of this application is used for the prediction of tumor subtype classification

[0041] The original data of Test Example 1 are SERS signal maps obtained by SERS testing of cells extracted from two breast cancer subtypes, Luminal B and HER2+. The samples include 17 cases of Luminal B and 2 cases of HER2+. One tumor cell was extracted from each sample. The SERS testing was the same as that in Example 1. After continuous testing, 3000 original SERS signal maps were obtained. The obtained original data were subjected to the same preprocessing and data augmentation as in Example 1: 2000 SERS signal maps were retained for each tumor cell sample through preprocessing. Subsequently, six new SERS signal maps were generated for each SERS signal map by three data augmentation methods and vertically connected to the corresponding original SERS signal map. That is, for each preprocessed original SERS signal map, six new SERS signal maps were added after data augmentation, together with one original SERS signal map. That is, seven SERS signal maps were retained for each original SERS signal map. Therefore, in this test example, 7×2000 = 14000 SERS signal maps corresponded to each tumor cell sample after data augmentation. For example, there were 17 tumor cell samples of the Luminal B subtype, and 14000×17 = 238000 SERS signal maps corresponded to each tumor cell sample after processing. There were 2 tumor cell samples of the HER2+ subtype, and 14000×2 = 28000 SERS signal maps corresponded to each tumor cell sample after processing. 28000 spectra were randomly selected from the 238000 SERS signal maps of the Luminal B subtype and divided into a training set, a validation set, and a test set in a ratio of 6:2:2. All 28000 spectra of the HER2+ subtype were divided into a training set, a validation set, and a test set in a ratio of 6:2:2. The training sets divided from the above two subtypes were respectively input into the improved AlexNet model of the present application for training to obtain a trained tumor classification prediction model, and the test set was input into the trained tumor classification prediction model for testing.

[0042] Figure 4Shows the classification results of the cell SERS signal maps extracted from two cancer subtypes using the improved AlexNet model of this application before and after data augmentation. Without data augmentation, the classification results for the Luminal B subtype were: accuracy 96.30%, precision 97.56%, recall 94.75%, and F1 score 96.14%. For HER2+, the results were: accuracy 96.30%, precision 95.17%, recall 97.76%, and F1 score 96.45%. After applying data augmentation, all evaluation metrics (accuracy, precision, recall, and F1 score) were significantly improved. Specifically, for the Luminal B subtype, the classification results were: accuracy 99.68%, precision 99.88%, recall 99.48%, and F1 score 99.68%. For HER2+, the results were: accuracy 99.68%, precision 99.48%, recall 99.87%, and F1 score 99.68%. These results indicate that data augmentation plays an important role in improving the classification ability of the model and effectively enhancing its performance in identifying breast cancer subtype samples. Through data augmentation, the generalization ability of the model is improved, enabling it to operate more stably when dealing with complex and highly diverse data. This allows the model to better capture the subtle differences between subtypes, achieving higher classification accuracy and consistency.

[0043] Comparison of the prediction effects of the test case 2 model for tumor subtype classification

[0044] In addition, using the same original data as in Test Example 1, similarly, the SERS signal graphs before and after data augmentation are respectively input into seven existing deep learning models (VGG19, LeNet, AlexNet, LSTM, ResNet18, GoogleNet, and EfficientNet), and divided into training set, validation set, and test set in the same ratio as in Test Example 1, that is, 6:2:2, for comparison, and trained, validated, and tested for comparison. The test results are shown in Table 2 below. The AlexNet model showed the best performance in distinguishing Luminal B and HER2+ subtypes, obtaining classification results with an accuracy of 96.30%, a precision of 96.37%, a recall of 96.26%, and an F1 value of 96.29%. In addition, the recognition accuracies of VGG19, LeNet, LSTM, ResNet18, GoogleNet, and EfficientNet all exceeded 90%. After applying data augmentation, overfitting of the model was effectively controlled by enhancing data diversity, and significant improvements were achieved in the accuracy, precision, recall, and F1 value of all models. Among them, the AlexNet model showed the best performance, achieving classification results with an accuracy of 99.68%, a precision of 99.68%, a recall of 99.68%, and an F1 value of 99.68%. In addition, except for the AlexNet model, the recognition accuracies of other models also increased to approximately 99%. These results indicate that data augmentation significantly improves the classification ability of the model, and the enhanced model has a high recognition accuracy for breast cancer subtype tumor samples.

[0045] Table 2: Results of Comparative Experiments

[0046]

Claims

1. A tumor classification prediction model training method based on SERS signal graphs, the training method comprising the steps of: obtaining SERS signal graphs of different subtypes of tumor cells tested by SERS as raw data, wherein the different subtypes of tumor cells in the raw data include several tumor cells of two or more subtypes, the same subtype has at least 2 tumor cell samples, and each tumor cell sample corresponds to at least 3000 SERS signal graphs obtained through continuous testing; performing preprocessing and data enhancement processing on all SERS signal graphs in the raw data in sequence, wherein the preprocessing steps include baseline removal, outlier detection and normalization in sequence; The original data, after preprocessing and data enhancement, are used together with the labels of the corresponding subtypes as training samples; The deep learning model is trained using the training samples, and a tumor classification prediction model is obtained after training.

2. A tumor classification prediction model training method based on SERS signal graph according to claim 1, characterized in that: The deep learning model is an improved AlexNet model, which includes a feature extraction module and a classification module. The feature extraction module includes: multiple 1D convolutional layers, ReLU, local response normalization (LRN), and a maximum pooling layer; the classification module includes: a flattening layer, a fully connected layer, ReLU, and Dropout; in the feature extraction module, the multiple 1D convolutional layers include a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, and a fifth convolutional layer in sequence, the first convolutional layer is used to input a 1D SERS signal graph, and the first convolutional layer is sequentially connected to ReLU, LRN, and a maximum pooling layer; the second convolutional layer is sequentially connected to ReLU, LRN, and a maximum pooling layer, and the third convolutional layer is connected to ReLU; the fourth convolutional layer is connected to ReLU; the fifth convolutional layer is sequentially connected to ReLU and a maximum pooling layer; in the classification module, the flattening layer, the first fully connected layer, ReLU, Dropout, the second fully connected layer, ReLU, and the third fully connected layer are sequentially passed through, and the third fully connected layer outputs the final category probability distribution.

3. A tumor classification prediction model training method based on SERS signal graph according to claim 2, characterized in that: The flattening layer flattens the data into a one-dimensional vector, and the input feature dimension is 3456; the first fully connected layer reduces the data dimension from 3456 to 2048, uses the ReLU activation function, and selects 50% Dropout to randomly discard 50% of the neurons; the second fully connected layer maintains 2048 dimensions.

4. The tumor classification prediction model training method based on SERS signal graph according to claim 2, characterized in that: In the improved AlexNet model, the first convolutional layer has an input channel number of 1 for a single-channel 1D SERS signal map and an output channel number of 48; the second convolutional layer increases the number of feature map channels to 128; and the third to fifth convolutional layers increase the number of feature map channels to 192.

5. The tumor classification prediction model training method based on SERS signal graph according to claim 1, characterized in that: The tumor cell samples are obtained by the following steps: immersing clinical tissue samples in a PBS solution, then mincing each tissue sample and transferring it together with the PBS solution to a 70 μm cell strainer, then grinding and filtering the sample using a tissue grinder, centrifuging the filtered cell solution at 1000 rpm for 5 minutes, removing the supernatant, and adding PBS to re-disperse the precipitate, thereby obtaining a cell suspension of each tumor cell sample.

6. The tumor classification prediction model training method based on SERS signal graph according to claim 1, characterized in that: The operation steps of the SERS test include: dropping the extracted tumor cells on the silver nanorod array SERS substrate and drying them naturally in a refrigerator at 4° C. for 8 hours; the SERS exposure time is 3 seconds, the laser intensity is 50%, and the total is 1 time.

7. The tumor classification prediction model training method based on SERS signal graph according to claim 1, characterized in that: In the preprocessing, the operation step of removing the baseline is: using an adaptive iterative reweighted penalty least squares algorithm to fit the baseline of the SERS signal map in the original data, and removing the baseline; the operation step of outlier detection is using an isolation forest (Isoforest) algorithm provided by a scikit-learn machine learning library to detect the same sample, and remove at least 30% of the outlier data in the SERS signal map; the operation step of normalization is: after baseline removal and outlier detection, applying a minimum-maximum normalization method to normalize the SERS spectrum intensity in the SERS signal map to the range of [0,1], and converting 600-1700cm -1 The Raman spectrum of the range is taken as the output result after normalization.

8. The tumor classification prediction model training method based on SERS signal graph according to claim 1, characterized in that: The data enhancement processing is selected from any one, two or three of noise enhancement, offset enhancement or scaling enhancement to generate enhanced data; each data enhancement processing generates two new enhanced SERS signal graphs from a single SERS signal graph point, and these enhanced data are sequentially added to the enhanced SERS signal graph list.

9. A tumor classification prediction model training method based on SERS signal graph according to claim 8, characterized in that: The data enhancement process is selected from three of noise enhancement, offset enhancement or scaling enhancement. The noise standard deviation in noise enhancement is set to 0.01; the offset amplitude in offset enhancement is set to 5, allowing -5cm -1 and 5cm -1 Shift forward or backward between Scaling enhancement is performed by generating a random scaling factor between 0.9 and 1.1 and scaling the original spectrum with the random scaling factor to produce a new spectrum.

10. Application of a tumor classification prediction model based on a SERS signal map trained by the training method according to any one of claims 1 to 9 in tumor classification prediction, i.e., a tumor classification prediction method based on a SERS signal map, the prediction method comprising the steps of: Acquiring a SERS signal graph of a target sample after SERS testing, wherein the target sample is a tumor cell of unknown subtype; The SERS signal graph of the target sample tested by SERS is subjected to preprocessing and data enhancement processing in sequence to obtain test data of the target sample; The test data of the target sample is analyzed by the tumor classification prediction model, and the corresponding type of the target sample is output.

Images