Raman denoising methods and models, and raman imaging methods and devices

Abstract

The application relates to a Raman denoising method, which comprises the following steps: performing encoding processing on an initial feature map of a Raman spectrum or a Raman image to obtain an encoded feature map; wherein the encoding processing comprises the following steps: performing feature extraction processing on the initial feature map to obtain an encoded separated feature map, and performing feature fusion processing on the encoded separated feature map to obtain the encoded feature map; performing dimension integration and feature integration processing on the encoded feature map to obtain a bearing feature map; performing decoding processing on the bearing feature map to obtain a decoded feature map of the corresponding Raman spectrum or Raman image after filtering out noise; wherein the decoding processing comprises the following steps: performing expansion processing on the encoded feature map to obtain a sampling feature map, splicing the sampling feature map and the encoded separated feature map to obtain a decoded splicing feature map, and performing feature extraction processing on the decoded splicing feature map to obtain the decoded feature map. The application can simultaneously consider the spatial dimension and the spectral dimension to perform Raman spectrum or Raman image denoising, and the denoising effect is strong.

Description

Technical Field

[0001] This invention relates to the field of spectral measurement technology, and in particular to a Raman denoising method. Background Technology

[0002] Modern spectroscopic methods include infrared spectroscopy, Raman spectroscopy, fluorescence spectroscopy, and mass spectrometry. Among these, Raman spectroscopy analyzes samples at the molecular level and identifies the fingerprint spectral characteristics of their rotational or vibrational modes in a non-destructive and label-free manner. Raman spectroscopy is a highly specific optical mode. It is primarily used to study the vibrational and structural properties of molecular groups and has wide applications in life sciences, medicine, and extraterrestrial matter detection. Because each peak in a Raman spectrum corresponds one-to-one with a functional group in a chemical molecule, the types and numbers of chemical bonds contained in the molecule can be inferred by analyzing specific peaks. In biomedical research, Raman spectroscopy can non-destructively and label-free obtain the biochemical composition of specific structures within living cells and tissues, such as the content and distribution of carbohydrates, proteins, lipids, and nucleic acids. This gives Raman spectroscopy immense application potential in biomedical research.

[0003] Compared to Raman spectroscopy, Raman imaging can simultaneously capture both the chemical information and spatial distribution characteristics of a sample. Raman imaging involves acquiring Raman spectra at specific sites while simultaneously recording the spatial information of those sites, and then arranging the acquired Raman spectra according to their corresponding spatial locations to obtain a visualized three-dimensional dataset. For simple monolayer samples, Raman imaging ultimately yields three-dimensional data, including two-dimensional planar space and a third dimension of scattered light wavelength information—that is, spatial coordinate information and spectral information of the substances. Thanks to the advantages of Raman imaging, such as being non-destructive, label-free, and providing subcellular resolution molecular imaging, many biomedical applications have been realized. Among these, Raman imaging of cells can reveal the relative concentrations and spatial distribution information of substances such as nucleic acids, lipids, and proteins within cells, thereby analyzing the biochemical composition and state of cells, which is of great significance for analyzing cell differentiation.

[0004] The core principle of Raman spectroscopy and Raman imaging is Raman scattering. When incident laser light is scattered by molecules, most of the scattered light has the same wavelength as the incident light, but a very small portion of the scattered light has a different wavelength. This wavelength change is determined by the chemical structure of the scattering substance. Since the Raman scattering background from water is negligible, Raman spectroscopy or Raman imaging can be used to study biological samples, such as living cells, without interference from water.

[0005] However, the Raman scattering process is inherently weak. Typically, in... arrive Only one of the scattered photons may undergo Raman scattering. The scarcity of scattering molecules results in a very weak Raman signal, which in turn increases the proportion of background noise in the overall signal. For example, in cases involving low-intensity Raman signals, the detection of single molecules or ultra-low concentrations of analytes is susceptible to noise interference generated during the acquisition of low-intensity signals. Similarly, for the identification of deep lesions within biological tissues, deeper tissue penetration means longer photon propagation paths, and tissue heterogeneity leads to noise interference in deeper regions of the Raman signal. In ultrafast imaging, the short acquisition time at each detection point results in a low signal-to-noise ratio (SNR) for the Raman image. Due to these application scenarios, the impact of noise becomes non-negligible in current applications of Raman scattering, making it difficult to accurately obtain Raman images with high SNR under complex noise interference. Summary of the Invention

[0006] Based on this, the purpose of this invention is to provide a Raman denoising method that can simultaneously take into account the complex correlation between spatial and spectral dimensions in spectral data.

[0007] A Raman denoising method includes:

[0008] The initial feature map of the Raman spectrum or Raman image is encoded to obtain an encoded feature map; wherein, the encoding process includes feature extraction processing of the initial feature map to obtain an encoded separation feature map, and feature fusion processing of the encoded separation feature map to obtain an encoded feature map;

[0009] The inherited feature map is obtained by performing dimensional integration and feature integration on the encoded feature map;

[0010] The inherited feature map is decoded to obtain the corresponding Raman spectrum or Raman image after noise filtering; wherein, the decoding process includes expanding the inherited feature map to obtain a sampled feature map, concatenating the sampled feature map with the encoded separate feature map to obtain a decoded concatenated feature map, and performing feature extraction processing on the decoded concatenated feature map to obtain a decoded feature map.

[0011] Furthermore, the feature extraction process for the initial feature map specifically includes: performing independent convolution processing on the features of the initial feature map for each channel to obtain multiple encoded convolutional feature maps, and performing point-by-point convolution processing on all encoded convolutional feature maps to obtain encoded decoupling feature maps.

[0012] Further, the feature fusion processing of the encoding separation feature map specifically includes: extracting the spatial dimension features of the encoding separation feature map to obtain a spatial feature map; extracting the spectral dimension features of the encoding separation feature map to obtain a spectral feature map; concatenating the spatial feature map and the spectral feature map to obtain an encoding concatenated feature map; performing dimensionality reduction processing on the encoding concatenated feature map to obtain a first encoding dimensionality-reduced feature map; performing pooling processing on the first encoding dimensionality-reduced feature map to obtain an encoding pooled feature map; performing attention allocation processing on the encoding pooled feature map to obtain an attention feature map; and multiplying the attention feature map with the first encoding dimensionality-reduced feature map to obtain a fused feature map with different weights.

[0013] Furthermore, the feature extraction process for the decoded spliced ​​feature map specifically includes: performing independent convolution processing on the features of the decoded spliced ​​feature map of each channel to obtain multiple decoded convolutional feature maps, and performing point-by-point convolution processing on all decoded convolutional feature maps to obtain decoded feature maps.

[0014] The present invention also provides a Raman denoising model, comprising:

[0015] At least one coding unit is used to encode an initial feature map of a Raman spectrum or Raman image to obtain a coded feature map; wherein, the coding unit includes a coding separation module and a fusion module, the coding separation module is used to perform feature extraction processing on the initial feature map to obtain a coded separation feature map, and the fusion module is used to perform feature fusion processing on the coded separation feature map to obtain a coded feature map;

[0016] A receiving unit is used to perform dimensional integration and feature integration processing on the encoded feature map to obtain the receiving feature map;

[0017] At least one decoding unit is used to decode the receiving feature map to obtain a corresponding Raman spectrum or a Raman image after noise filtering; wherein, the decoding unit includes an upsampling module, a decoding connection module, and a decoding separation module, the upsampling module is used to expand the receiving feature map to obtain a sampled feature map, the decoding connection module is used to concatenate the sampled feature map and the encoded separation feature map to obtain a decoded concatenated feature map, and the decoding separation module is used to perform feature extraction processing on the decoded concatenated feature map to obtain a decoded feature map.

[0018] Furthermore, the encoding separation module is used to perform independent convolution processing on the features of the initial feature map of each channel to obtain multiple encoded convolutional feature maps, and to perform point-by-point convolution processing on all encoded convolutional feature maps to obtain encoded separation feature maps.

[0019] Further, the fusion module is used to extract the spatial dimension features of the encoded separation feature map to obtain a spatial feature map; extract the spectral dimension features of the encoded separation feature map to obtain a spectral feature map; concatenate the spatial feature map and the spectral feature map to obtain an encoded concatenated feature map; perform dimensionality reduction processing on the encoded concatenated feature map to obtain a first encoded dimensionality-reduced feature map; perform pooling processing on the first encoded dimensionality-reduced feature map to obtain an encoded pooling feature map; perform attention allocation processing on the encoded pooling feature map to obtain an attention feature map; and multiply the attention feature map with the first encoded dimensionality-reduced feature map to obtain a fusion feature map with different weights.

[0020] Furthermore, the decoding separation module is used to perform independent convolution processing on the features of the decoded splicing feature map of each channel to obtain multiple decoded convolutional feature maps, and to perform point-by-point convolution processing on all the decoded convolutional feature maps to obtain decoded feature maps.

[0021] The present invention also provides a Raman imaging method, comprising:

[0022] Obtain the initial feature map of the Raman spectrum or Raman image;

[0023] The initial feature map of the Raman spectrum or Raman image is encoded to obtain an encoded feature map; wherein, the encoding process includes feature extraction processing of the initial feature map to obtain an encoded separation feature map, and feature fusion processing of the encoded separation feature map to obtain an encoded feature map;

[0024] The inherited feature map is obtained by performing dimensional integration and feature integration on the encoded feature map;

[0025] The inherited feature map is decoded to obtain the corresponding Raman spectrum or Raman image after noise filtering; wherein, the decoding process includes expanding the encoded feature map to obtain a sampled feature map, concatenating the sampled feature map with the encoded separated feature map to obtain a decoded concatenated feature map, and performing feature extraction processing on the decoded concatenated feature map to obtain a decoded feature map.

[0026] The present invention also provides a Raman imaging device, comprising:

[0027] A Raman spectrometer is used to acquire initial feature maps of Raman spectra or Raman images;

[0028] At least one coding unit is used to encode an initial feature map of a Raman spectrum or Raman image to obtain a coded feature map; wherein, the coding unit includes a coding separation module and a fusion module, the coding separation module is used to perform feature extraction processing on the initial feature map to obtain a coded separation feature map, and the fusion module is used to perform feature fusion processing on the coded separation feature map to obtain a coded feature map;

[0029] A receiving unit is used to perform dimensional integration and feature integration processing on the encoded feature map to obtain the receiving feature map;

[0030] At least one decoding unit is used to decode the received feature map to obtain a corresponding Raman spectrum or a Raman image after noise filtering; wherein, the decoding unit includes an upsampling module, a decoding connection module, and a decoding separation module, the upsampling module is used to expand the encoded feature map to obtain a sampled feature map, the decoding connection module is used to concatenate the sampled feature map and the encoded separation feature map to obtain a decoded concatenated feature map, and the decoding separation module is used to perform feature extraction processing on the decoded concatenated feature map to obtain a decoded feature map.

[0031] To better understand and implement this invention, the following detailed description is provided in conjunction with the accompanying drawings. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the Raman denoising model of the present invention.

[0033] Figure 2 This is a schematic diagram of the operation process of the Raman denoising model of the present invention.

[0034] Figure 3 The input image with noise is used for testing the Raman denoising model of the present invention.

[0035] Figure 4 for Figure 3 The input image is the image obtained after wavelet denoising.

[0036] Figure 5 for Figure 3 The input image is the image obtained after SG filtering.

[0037] Figure 6 for Figure 3 The input image is the image obtained after processing by a one-dimensional U-Net.

[0038] Figure 7 for Figure 3 The input image is processed by the Raman denoising model of the present invention to obtain the reconstructed image.

[0039] Figure 8 for Figure 3 The input image corresponds to the original image without noise.

[0040] Figure 9 The input spectrum with noise is used when testing the Raman denoising model of this invention.

[0041] Figure 10 for Figure 9 The spectrum obtained after wavelet denoising is applied to the input spectrum.

[0042] Figure 11 for Figure 9 The spectrum obtained after the input spectrum is filtered by SG is obtained.

[0043] Figure 12 for Figure 9 The spectrum obtained after the input spectrum is processed by a one-dimensional U-Net.

[0044] Figure 13 for Figure 9 The input spectrum is used to obtain the reconstructed spectrum after passing through the Raman denoising model of the present invention.

[0045] Figure 14 for Figure 9 The input spectrum corresponds to the noise-free original spectrum. Detailed Implementation

[0046] The applicant understands that current Raman denoising methods using deep convolutional networks can only perform denoising in a single stage, requiring continuous adjustment of denoising parameters. They only denoise the Raman spectra one by one in the Raman image or only denoise the Raman image corresponding to a certain wavenumber, without considering the complex correlation between the spatial and spectral dimensions in the Raman spectrum or Raman image stack. This leads to the potential loss of some important information during reconstruction, failing to achieve optimal denoising results, and exhibiting weak network generalization ability. Therefore, the applicant attempts to build a model using a three-dimensional convolutional network, considering the independent characteristics of Raman spectra and Raman images in both spatial and spectral dimensions. However, the increased number of parameters could lead to computational limitations or excessive computational costs. To address this, the applicant incorporates a fusion module, using pseudo-three-dimensional convolution to extract and fuse features in both the spectral and spatial dimensions, thereby reducing the number of computational parameters while maintaining the complex correlation between the spatial and spectral dimensions in the spectral data.

[0047] The Raman imaging device includes a Raman spectrometer and a Raman denoising model. The Raman spectrometer acquires Raman spectra containing chemical information of the sample and performs Raman imaging based on the Raman spectra to obtain Raman images. The Raman denoising model is connected to the Raman spectrometer, receives Raman spectra and Raman images from the Raman spectrometer, and performs denoising processing on the Raman spectra to obtain reconstructed spectra and on the Raman images to obtain reconstructed images.

[0048] Please see Figure 1 The Raman denoising model includes at least one encoding unit, one receiving unit, and at least one decoding unit. Figure 2 This is a flowchart illustrating the operation of the Raman denoising model. The encoding unit is used to process the initial feature map. Encoding processing is performed to obtain the encoded feature map. In this context, Raman spectra or Raman images serve as the initial feature maps. The feature map in the encoding unit adopts express, This represents the sub-index of the feature map within the coding unit. The sequence number of the encoding unit where the feature map is located; the receiving unit is used for encoding the feature map. Dimensional integration and feature integration are performed to obtain the receiving feature map, wherein the feature map in the receiving unit adopts... express, The sub-index of the feature map in the receiving unit; the decoding unit is used to decode the received feature map to obtain a decoded feature map, wherein the decoded feature map is a reconstructed spectrum obtained after noise has been filtered out from the Raman spectrum, or a reconstructed image obtained after noise has been filtered out from the Raman image, and the feature map in the decoding unit adopts... express, This is the sub-index of the feature map in the decoding unit. This is the sequence number of the decoding unit where the feature map is located.

[0049] The following detailed description uses a Raman denoising model consisting of two encoding units, one receiving unit, and two decoding units as an example.

[0050] The encoding unit includes an encoding upscaling module, an encoding separation module, a fusion module, and a compression module.

[0051] Specifically, in the first encoding unit, the encoding upsizing module is used to process the initial feature map. The dimensionality is increased to obtain the encoded dimensionality-incremented feature map. Specifically, the encoding dimensionality enhancement module includes a 1×1×1 convolutional layer, and the initial feature map The input channel has 1 value, a height of 64, a width of 64, and 500 spectral bands. The encoded up-dimensional feature map is obtained. for .

[0052] The encoding separation module is used to upscale the encoded feature map. Feature extraction is performed to obtain the encoded separation feature map. Specifically, the encoding separation module includes channel-wise convolutional layers and point-wise convolutional layers. The number of convolutional kernels in the channel-wise convolutional layers is set to 16, and each convolutional kernel corresponds to one input channel for independent convolution processing, resulting in multiple encoded convolutional feature maps. The pointwise convolutional layer encodes all the convolutional feature maps. Perform pointwise convolution to obtain the encoded separation feature map. The encoded separation feature map for .

[0053] Specifically, assume the size of the input feature map is... ,in , , , These represent the number of feature channels, height, width, and number of spectral bands of the input feature map, respectively. Assume the kernel size for convolution on the input feature map is... ,in The dimension of the convolution kernel. , These represent the number of input channels and the number of output channels of the convolution kernel, respectively.

[0054] For existing modules that utilize 3D convolution, the number of computational parameters... for:

[0055]

[0056] For the encoding separation module of this application, the number of computational parameters for the channel-by-channel convolutional layer is... The number of computational parameters for pointwise convolutional layers and total calculation parameters They are respectively:

[0057]

[0058]

[0059]

[0060] Therefore, the computational costs of depthwise separable 3D convolution and standard 3D convolution can be compared:

[0061]

[0062] As can be seen from the above formulas, the number of computational parameters of the encoding separation module in this application is greatly reduced compared with the existing modules that use three-dimensional convolution, achieving the same performance with lower computational cost.

[0063] The fusion module is used to separate the encoded feature maps. Feature fusion processing is performed to obtain a fused feature map. Specifically, the fusion module includes a pseudo-3D convolutional block, a connection layer, three 1×1×1 convolutional layers, an adaptive average pooling layer, an attention layer, and a fusion layer; the pseudo-3D convolutional block extracts and encodes the feature map for separation. Spatial feature map is obtained from spatial dimensional features. Extracting and separating feature maps The spectral feature map is obtained from the spectral dimensional features. The spatial feature map for The spectral feature map for The connection layer will display the spatial feature map. and spectral feature map The coded concatenated feature map is obtained by concatenating the features. The encoded splicing feature map for The encoded splicing feature map After the first 1×1×1 convolutional layer, the dimensionality is reduced to obtain the first encoded dimensionality-reduced feature map. The first encoded dimensionality reduction feature map for The adaptive average pooling layer reduces the dimensionality of the first encoded feature map. Pooling is performed to obtain the encoded pooled feature map. The encoded pooling feature map for The encoded pooling feature map After passing through the second and third 1×1×1 convolutional layers and the attention layer, we obtain the attention feature map showing the distribution of attention weights for each channel. The attention layer employs the Sigmoid activation function, and the attention feature map... for The fusion layer will reduce the dimensionality of the first encoded feature map. With attention feature map Multiplication yields a fused feature map with different weights for each channel. The fused feature map for .

[0064] The compression module is used to process the fused feature map. Compressed feature maps are obtained by performing compression processing. Specifically, the compression module includes a 2×2×2 max-pooling layer, and the fused feature map The compressed feature map is obtained by the size of the compressed feature map after the max pooling layer. The compressed feature map for .

[0065] The second encoding unit is connected to the first encoding unit and receives the compressed feature map from the first encoding unit. In the second encoding unit, the encoding upscaling module is used to process the compressed feature map. The dimensionality is increased to obtain the encoded dimensionality-incremented feature map. Specifically, the encoding dimensionality enhancement module includes a 1×1×1 convolutional layer to obtain the encoded dimensionality enhancement feature map. for .

[0066] The encoding separation module is used to upscale the encoded feature map. Feature extraction is performed to obtain the encoded separation feature map. Specifically, the encoding separation module includes channel-wise convolutional layers and point-wise convolutional layers. The number of convolutional kernels in the channel-wise convolutional layers is set to 16, and each convolutional kernel corresponds to one input channel for independent convolution processing, resulting in multiple encoded convolutional feature maps. The pointwise convolutional layer encodes all the convolutional feature maps. Perform pointwise convolution to obtain the encoded separation feature map. The encoded separation feature map for .

[0067] The fusion module is used to separate the encoded feature maps. Feature fusion processing is performed to obtain a fused feature map. Specifically, the fusion module includes a pseudo-3D convolutional block, a connection layer, three 1×1×1 convolutional layers, an adaptive average pooling layer, an attention layer, and a fusion layer; the pseudo-3D convolutional block extracts and encodes the feature map for separation. Spatial feature map is obtained from spatial dimensional features. Extracting and separating feature maps The spectral feature map is obtained from the spectral dimensional features. The spatial feature map for The spectral feature map for The connection layer will display the spatial feature map. and spectral feature map The coded concatenated feature map is obtained by concatenating the features. The encoded splicing feature map for The encoded splicing feature map After the first 1×1×1 convolutional layer, the dimensionality is reduced to obtain the first encoded dimensionality-reduced feature map. The first encoded dimensionality reduction feature map for The adaptive average pooling layer reduces the dimensionality of the first encoded feature map. Pooling is performed to obtain the encoded pooled feature map. The encoded pooling feature map for The encoded pooling feature map After passing through the second and third 1×1×1 convolutional layers and the attention layer, we obtain the attention feature map showing the distribution of attention weights for each channel. The attention layer employs the Sigmoid activation function, and the attention feature map... for The fusion layer will reduce the dimensionality of the first encoded feature map. With attention feature map Multiplication yields a fused feature map with different weights for each channel. The fused feature map for .

[0068] The compression module is used to process the fused feature map. Compressed feature maps are obtained by performing compression processing. Specifically, the compression module includes a 2×2×2 max-pooling layer, and the fused feature map The compressed feature map is obtained by the size of the compressed feature map after the max pooling layer. The compressed feature map for The compressed feature map serves as the final encoded feature map.

[0069] The receiving unit includes at least one receiving dimensionality-upgrading module, a receiving separation module, and at least one receiving dimensionality-reducing module, wherein the number of receiving dimensionality-upgrading modules corresponds to the number of receiving dimensionality-reducing modules. The receiving dimensionality-upgrading module includes a 1×1×1 convolutional layer for compressing the feature map. The dimensionality-upgrading process yields the dimensionality-upgrading feature map. The aforementioned upscaling feature map for The specific structure of the receiving and separating module is the same as that of the encoding and separating module, and it handles the receiving and upsizing feature map. Feature extraction is performed to obtain the receiving and separating feature map. The receiving separation feature map for The dimensionality reduction module includes a 1×1×1 convolutional layer for processing the separation feature map. Dimensionality reduction is performed to obtain the dimensionality-reduced feature map. The aforementioned dimensionality reduction feature map for .

[0070] The decoding unit includes an upsampling module, a decoding connection module, a first decoding dimensionality reduction module, a decoding separation module, and a second decoding dimensionality reduction module.

[0071] Specifically, in the first decoding unit, the upsampling module includes a 2×2×2 upsampling layer for receiving the dimensionality-reduced feature map. The sampled feature map is obtained by performing extended processing. The sampling feature map for The decoding connection module includes a connection layer that samples the feature map. Separation feature map from the second coding unit The resulting decoded spliced ​​feature map is obtained by splicing. The decoded splicing feature map for The first dimensionality reduction module for decoding includes a 1×1×1 convolutional layer used to concatenate the decoded feature maps. The first decoded dimensionality-reduced feature map is obtained by performing dimensionality reduction processing. The first decoded dimensionality reduction feature map for The decoding separation module is used to perform dimensionality reduction on the first decoded feature map. Feature extraction is performed to obtain the decoding and separation feature map. Specifically, the encoding separation module includes a channel-wise convolutional layer and a point-wise convolutional layer. The number of convolutional kernels in the channel-wise convolutional layer is set to 16, and each convolutional kernel corresponds to one input channel for independent convolution processing, resulting in multiple decoded convolutional feature maps. The pointwise convolutional layer decodes all the convolutional feature maps. Perform pointwise convolution to obtain the decoded and separated feature map. The decoding separation feature map for The second dimensionality reduction module for decoding includes a 1×1×1 convolutional layer used to separate the feature maps during decoding. Dimensionality reduction is performed to obtain the second decoded dimensionality-reduced feature map. The second decoded dimensionality reduction feature map for .

[0072] The second decoding unit is connected to the first decoding unit and receives the second decoded dimensionality reduction feature map from the first decoding unit. The second decoding unit includes an upsampling module, a decoding connection module, a first decoding dimensionality reduction module, a decoding separation module, and a second decoding dimensionality reduction module. The upsampling module includes a 2×2×2 upsampling layer for processing the second decoded dimensionality reduction feature map. The sampled feature map is obtained by performing extended processing. The sampling feature map for The decoding connection module includes a connection layer that samples the feature map. Separation feature map from the first coding unit The resulting decoded spliced ​​feature map is obtained by splicing. The decoded splicing feature map for The first dimensionality reduction module for decoding includes a 1×1×1 convolutional layer used to concatenate the decoded feature maps. The first decoded dimensionality-reduced feature map is obtained by performing dimensionality reduction processing. The first decoded dimensionality reduction feature map for The decoding separation module is used to perform dimensionality reduction on the first decoded feature map. Feature extraction is performed to obtain the decoding and separation feature map. Specifically, the encoding separation module includes a channel-wise convolutional layer and a point-wise convolutional layer. The number of convolutional kernels in the channel-wise convolutional layer is set to 16, and each convolutional kernel corresponds to one input channel for independent convolution processing, resulting in multiple decoded convolutional feature maps. The pointwise convolutional layer decodes all the convolutional feature maps. Perform pointwise convolution to obtain the decoded and separated feature map. The decoding separation feature map for The second dimensionality reduction module for decoding includes a 1×1×1 convolutional layer used to separate the feature maps during decoding. Dimensionality reduction is performed to obtain the second decoded dimensionality-reduced feature map. The second decoded dimensionality reduction feature map for The second decoded dimensionality reduction feature map is used as the final decoded feature map.

[0073] Since the structure of each coding unit is exactly the same, with only the parameters of the coding feature map undergoing the same logical changes, it will not be described in detail here. Those skilled in the art will understand that when multiple coding units exist, the second coding unit compresses the feature map... Encoding process is performed to obtain compressed feature maps The third coding unit compresses the feature map. Encoding process is performed to obtain compressed feature maps And so on, the actual number of coding units needed depends on the initial feature map. By selecting the parameters and the level of noise reduction desired by the user, the optimal noise reduction effect can be achieved.

[0074] Similarly, the structure of each decoding unit is exactly the same, except that the parameters of the decoding feature map undergo the same logical changes, which will not be elaborated here. It is important to note that the number of decoding units should correspond to the number of encoding units. Furthermore, in the decoding connection module of the decoding unit, the decoding connection module connects the decoding feature map obtained by the upsampling module of the decoding unit and the encoding feature map obtained by the encoding separation module of the encoding unit. The encoding unit connected to this module should be symmetrical about the receiving unit's center to the decoding unit corresponding to it. Specifically, when the Raman denoising model has five encoding units, the decoding connection module of the first decoding unit should connect to the fifth encoding unit, the decoding connection module of the second decoding unit should connect to the fourth encoding unit, and so on.

[0075] After obtaining the Raman denoising model of this application, in order to verify its practical application effect, it was trained and tested using the following methods.

[0076] 169 high SNR Raman images of breast cancer cells (containing 1.4 million Raman spectra) were used as clear Raman images, with an image size of 64×64×500, where 64 represents the spatial physical size of the Raman imaging region and 500 represents the relative wavenumber of the Raman image. The spectral range of the Raman images is... arrive The spatial resolution is 0.5 μm.

[0077] The intensity of each Raman image depends on the sample characteristics, the acquisition environment, and the Raman spectrometer. Therefore, to remove the intensity non-uniformity between Raman images, the Raman image data is normalized by minima.

[0078]

[0079] in, and These represent the maximum and minimum intensity values ​​in the Raman image, respectively. The normalized Raman image has a size of H×W×B, where H, W, and B represent its height, width, and number of spectral bands, respectively.

[0080] Raman spectroscopy noise mainly includes shot noise, dark noise, and readout noise. Shot noise is related to intrinsic quantum processes during photon generation, arising from quantum fluctuations when photons are converted into electrons via the photoelectric effect, and follows a Poisson distribution. Dark noise, as a signal-dependent noise, is proportional to signal intensity and is also called multiplicative noise. When electrons are converted into electrical signals, readout noise and dark noise are introduced, following a Gaussian distribution. Gaussian noise, as additive noise, is independent of signal intensity. To better reflect the noise distribution of real data, a mixed Poisson-Gaussian noise model was selected to generate a noise dataset based on high SNR Raman spectral images of breast cancer cells. The mathematical expression of the mixed Poisson-Gaussian noise model is:

[0081]

[0082] in, It is Poisson noise, determined by signal strength weighting; , is a constant that depends on the sensor and analog gain; It is Gaussian noise, which is the remaining noise that is unrelated to the signal; The noise is a mixed random noise. Mixed Poisson-Gaussian noise is added to 169 clear images to generate 169 noisy Raman images with low SNR, forming the dataset. The linear model of the noisy Raman images is then performed as follows:

[0083]

[0084] in, , These represent the noisy Raman image, the clean Raman spectrum image, and the image with added mixed random noise, respectively.

[0085] According to the ratio of 85:10:5 The dataset is divided into training, validation, and test sets. 144 hyperspectral Raman images are used as the training set, 17 as the validation set, and 8 as the test set. These images are then input into the Raman denoising model to obtain the corresponding output results.

[0086] To better quantify the reconstruction effect and compare it with other reconstruction methods, structural similarity is introduced. SSIM ) and peak signal-to-noise ratio ( PSNR The reconstructed image is evaluated.

[0087] Two images x and y Between SSIM Defined as:

[0088]

[0089] image y Relative to label image x The PSNR is defined as:

[0090]

[0091] Please see Figures 3 to 8 As can be seen, the Raman image with noise undergoes different denoising processes and is compared with the original image without noise. Specifically, the table headers are the input Raman image with noise, the Raman image after wavelet denoising, the Raman image after SG filtering, the Raman image after denoising by a one-dimensional U-Net network, and the reconstructed image after Raman denoising of the present invention. Based on this, a quantitative evaluation comparison table of different reconstruction methods in the spatial dimension is established to more intuitively show the effect.

[0092] Table 1

[0093]

[0094] Please refer to Table 1, which is a quantitative evaluation comparison table of different reconstruction methods in the spatial dimension. It can be seen that the reconstructed image of the Raman denoising model of the present invention has an SSIM of 0.9489 and a PSNR of 40.58dB. Compared with other denoising methods, it is closer to the original image and has a significant outstanding effect.

[0095] In addition, mean square error (MSE) is also introduced. MSE ) and signal-to-noise ratio ( SNR Evaluation of the reconstructed spectrum a and b of MSE Represented as:

[0096]

[0097] spectrum a and b of SNR Represented as:

[0098]

[0099] Please see Figures 9 to 14 As can be seen, the noisy Raman spectrum undergoes different denoising processes and is compared with the original noise-free spectrum. Specifically, the table headers are the noisy input Raman spectrum, the Raman spectrum after wavelet denoising, the Raman spectrum after SG filtering, the Raman spectrum after denoising by a one-dimensional U-Net network, and the reconstructed spectrum after Raman denoising of the present invention. Based on this, a quantitative evaluation comparison table of different reconstruction methods in the spectral dimension is established to more intuitively show the effect.

[0100] Table 2

[0101]

[0102] Please refer to Table 2, which is a quantitative evaluation comparison table of different reconstruction methods in the spectral dimension. It can be seen that the MSE of the reconstructed spectrum of the Raman denoising model of the present invention reaches 0.000019 and the SNR reaches 18.71dB, which is also better than other denoising methods.

[0103] The Raman denoising method of this invention can simultaneously achieve overall denoising of the spectral and spatial axes from both spectral and spatial dimensions. Compared with the denoising results of single-stage methods, the proposed method can achieve better denoising effects in both spectral and spatial dimensions. Moreover, it does not require the introduction of any additional hardware and can reconstruct low SNR cell Raman images into high SNR cell Raman images simply through post-processing at the data end.

[0104] The above-described embodiments are merely one implementation of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and the present invention also intends to include these modifications and variations.

Claims

1. A Raman denoising method, characterized in that: include: The initial feature map of the Raman spectrum or Raman image is encoded to obtain an encoded feature map; wherein, the encoding process includes feature extraction processing of the initial feature map to obtain an encoded separation feature map, and feature fusion processing of the encoded separation feature map to obtain an encoded feature map; The inherited feature map is obtained by performing dimensional integration and feature integration on the encoded feature map; The inherited feature map is decoded to obtain the corresponding Raman spectrum or Raman image after noise filtering; wherein, the decoding process includes expanding the inherited feature map to obtain a sampled feature map, concatenating the sampled feature map with the encoded separate feature map to obtain a decoded concatenated feature map, and performing feature extraction processing on the decoded concatenated feature map to obtain a decoded feature map; The feature fusion processing of the encoded separate feature map specifically includes: extracting the spatial dimension features of the encoded separate feature map to obtain a spatial feature map; extracting the spectral dimension features of the encoded separate feature map to obtain a spectral feature map; concatenating the spatial feature map and the spectral feature map to obtain an encoded concatenated feature map; performing dimensionality reduction processing on the encoded concatenated feature map to obtain a first encoded dimensionality-reduced feature map; performing pooling processing on the first encoded dimensionality-reduced feature map to obtain an encoded pooled feature map; performing attention allocation processing on the encoded pooled feature map to obtain an attention feature map; and multiplying the attention feature map with the first encoded dimensionality-reduced feature map to obtain a fused feature map with different weights.

2. The Raman denoising method according to claim 1, characterized in that: The feature extraction process for the initial feature map specifically includes: performing independent convolution processing on the features of the initial feature map for each channel to obtain multiple encoded convolutional feature maps, and performing point-by-point convolution processing on all encoded convolutional feature maps to obtain encoded decoupling feature maps.

3. The Raman denoising method according to claim 2, characterized in that: The specific features extraction process for the decoded spliced ​​feature map includes: performing independent convolution processing on the features of the decoded spliced ​​feature map of each channel to obtain multiple decoded convolutional feature maps, and performing point-by-point convolution processing on all decoded convolutional feature maps to obtain the decoded feature map.

4. A Raman imaging device, characterized in that: include: At least one coding unit is used to encode an initial feature map of a Raman spectrum or Raman image to obtain a coded feature map; wherein, the coding unit includes a coding separation module and a fusion module, the coding separation module is used to perform feature extraction processing on the initial feature map to obtain a coded separation feature map, and the fusion module is used to perform feature fusion processing on the coded separation feature map to obtain a coded feature map; A receiving unit is used to perform dimensional integration and feature integration processing on the encoded feature map to obtain the receiving feature map; At least one decoding unit is used to decode the received feature map to obtain a corresponding Raman spectrum or a Raman image after noise filtering; wherein, the decoding unit includes an upsampling module, a decoding connection module, and a decoding separation module, the upsampling module is used to expand the received feature map to obtain a sampled feature map, the decoding connection module is used to concatenate the sampled feature map with the encoded separation feature map to obtain a decoded concatenated feature map, and the decoding separation module is used to perform feature extraction processing on the decoded concatenated feature map to obtain a decoded feature map; The fusion module is used to perform feature fusion processing on the code-separated feature map, specifically including: extracting the spatial dimension features of the code-separated feature map to obtain a spatial feature map; extracting the spectral dimension features of the code-separated feature map to obtain a spectral feature map; concatenating the spatial feature map and the spectral feature map to obtain a code-concatenated feature map; performing dimensionality reduction processing on the code-concatenated feature map to obtain a first code-reduced feature map; performing pooling processing on the first code-reduced feature map to obtain a code-pooled feature map; performing attention allocation processing on the code-pooled feature map to obtain an attention feature map; and multiplying the attention feature map with the first code-reduced feature map to obtain a fused feature map with different weights.

5. The Raman imaging device according to claim 4, characterized in that: The encoding separation module is used to perform independent convolution processing on the features of the initial feature map of each channel to obtain multiple encoded convolutional feature maps, and to perform point-by-point convolution processing on all encoded convolutional feature maps to obtain encoded separation feature maps.

6. The Raman imaging device according to claim 5, characterized in that: The decoding separation module is used to perform independent convolution processing on the features of the decoded splicing feature map of each channel to obtain multiple decoded convolutional feature maps, and to perform point-by-point convolution processing on all decoded convolutional feature maps to obtain decoded feature maps.

Images