A method for denoising and normalizing multi-channel mass spectrum images based on statistical method

Abstract

The application discloses a method for denoising and normalizing a multi-channel mass spectrum image based on a statistical method. The method can effectively reduce image noise, improve the accuracy of cell protein expression quantification, and reduce sample bias caused by sample preparation and preservation.

Description

Technical Field

[0001] This application relates to the field of image processing, and more particularly to a method for noise reduction and normalization on single-cell protein spectrum imaging images. Background Technology

[0002] Imaging Mass Cytometry (IMC) is a novel multiplex molecular imaging technique that combines mass spectrometry and image visualization. It uses laser ablation to generate particles, which are then carried by an inert gas stream to a mass spectrometer for detection. This allows for the simultaneous measurement of multiple protein markers within a single tissue section, providing information on the molecular structure and spatial domains of the tissue. As a hot topic in molecular imaging and mass spectrometry, IMC has received significant attention in recent years and has rapidly developed into an important tool for analyzing tissue sections. It can detect the distribution characteristics and changes in the concentration of small molecules such as genes, proteins, and drugs within organisms, providing information on molecular changes during various physiological and pathological processes. It holds significant application prospects in clinical medicine, molecular biology, and pharmacy.

[0003] With the continuous development of this technology, many technical problems have emerged that urgently need to be solved. The scale and information content of current IMC images are becoming increasingly massive, necessitating the development of corresponding efficient data analysis methods. The IMC image processing workflow mainly includes user-interactive single-cell segmentation, single-cell clustering, and cell type labeling. However, during IMC imaging, noise and batch effects are introduced into IMC images due to the different characteristics of various protein markers used, as well as various systematic and random errors introduced by laboratory mass spectrometers, imaging equipment, or their corresponding image processing software. Current IMC image processing workflows often lack effective denoising algorithms and batch effect elimination algorithms. These problems directly affect the accuracy and precision of downstream analyses such as single-cell expression quantification, clustering, and labeling, hindering the clinical application of this technology. Summary of the Invention

[0004] In IMC image processing workflows, the commonly used denoising method in this field is to select the entire image as the noise region and apply median filtering. However, in areas where real cells exist, median filtering can lead to the removal of some weakly expressed protein signals after denoising, resulting in low accuracy and precision of the denoised image. Furthermore, for many clustered noise points, simply using median filtering cannot remove those image noise points, resulting in poor denoising results.

[0005] This invention proposes a method for image denoising and inter-sample protein level normalization in imaging cellular mass spectrometry (IMC). This method can effectively reduce image noise while preserving the biological information of the image. Specifically, this invention solves the technical problem through the following technical solution.

[0006] 1. A method for denoising and protein level normalization of images obtained from imaging cellular mass spectrometry (IMC) techniques, comprising the following steps for denoising the images:

[0007] 1) Identify individual cells in the IMC image using an image single-cell segmentation method;

[0008] 2) Preprocess the IMC image;

[0009] 3) Quantify the protein expression value of the cell into the statistical value of all pixel values ​​in the segmented region of the cell;

[0010] 4) Select a certain number of random pseudo-cells on the image, and quantify the protein expression value of the random pseudo-cells according to the method in step 3) to obtain the protein expression value of the random pseudo-cells.

[0011] 5) The background noise level of the image is calculated by using the protein expression values ​​of random pseudocells. After eliminating the influence of background noise in the image, the denoised IMC image is obtained.

[0012] 2. The method described in Project 1 further includes the following step of normalizing the image:

[0013] 1) Compare the distribution of protein expression values ​​of cells before noise reduction obtained in step 3) of Project 1 with the distribution of protein expression values ​​of random pseudo cells obtained in step 4) of Project 1, and determine the threshold of protein expression values ​​of positive cells by controlling the false discovery rate (FDR).

[0014] 2) Calculate the normalization scaling parameter for each sample by using the protein expression values ​​of positive cells on multiple samples, and normalize the IMC image of each sample.

[0015] 3. The method described in Project 1, wherein the preprocessing includes converting the image pixel values ​​using the Arcsinh function to compress the scale of the pixel values ​​and make the pixel values ​​more stable; and removing "hot pixels" on the image using a filter to reduce the noise impact of outlier "hot pixels".

[0016] 4. The method described in Project 1, wherein step 4) selecting random pseudo-cells further includes selecting random pseudo-cells on the image, and the method for selecting random pseudo-cells is as follows:

[0017] 1) Calculate the major axis, minor axis, orientation, and area of ​​all cells obtained by cell segmentation in step 1) of Project 1, and fit the major axis, minor axis, and orientation separately using individual Gaussian models;

[0018] 2) Draw random numbers from the three distributions mentioned above, and randomly select a coordinate point on the image. Using this point as the center, form an elliptical random pseudo-cell on the image. This random pseudo-cell must satisfy the following two conditions:

[0019] Condition 1: The area of ​​the random pseudo-cell is within the range of all cell areas;

[0020] Condition 2: When part of the random pseudo-cell extends beyond the image boundary, the area of ​​the random pseudo-cell within the image must be greater than 5 pixels.

[0021] 5. The method described in Project 4, wherein all pixels of the random pseudo-cells are within the noise region of the image, and the noise region is defined as follows:

[0022] 1) X ori <X θ The pixel is set to 0, where X ori It is the pixel value of the original IMC image, X θ It is the set minimum expression threshold;

[0023] 2) Apply a median filter to the image after the previous step.

[0024] Then, the area with a pixel value of 0 in the image obtained in step 2) above is defined as the noise area.

[0025] 6. The method described in Project 1, wherein the background noise level X of the image noise Calculate as follows:

[0026] X noise =mean(X) decoy,i )

[0027] Among them, X decoy,i The average protein expression value of the pixel on random pseudocell i in the IMC image, i.e.

[0028] X decoy,i =C i / N i

[0029] Among them, C i N represents the protein expression level of random pseudocell i. i denoted as the total number of pixels in random pseudo-cell i.

[0030] 7. The method described in Project 6, wherein the pixel value X of the denoised IMC image... denoise Calculate as follows:

[0031] X denoise =max(0,X) origin -X noise )

[0032] Among them, X denoise X represents the pixel value after noise reduction. origin These are the pixel values ​​before noise reduction obtained after step 2) of Project 1.

[0033] X noise This represents the background noise level of the image.

[0034] 8. The method described in Project 2, wherein the False Discovery Rate (FDR) is calculated as follows:

[0035]

[0036] Wherein, FP is the number of false positive cells, that is, the number of cells with protein expression values ​​greater than the threshold extracted from random pseudo cells in step 4) of Project 1, and TP is the number of true positive cells, that is, the number of cells with protein expression values ​​greater than the threshold before noise reduction in step 3) of Project 1.

[0037] 9. The method described in Project 2, wherein the normalization of cellular protein expression values ​​includes:

[0038] For each protein biomarker, the protein expression value of all positive cells in each sample is calculated on the denoised IMC image. Then, using the protein expression values ​​of all positive cells in each sample, the normalization scaling parameter r for each sample is obtained as follows. k :

[0039]

[0040] in, r is the average of the protein expression values ​​of all positive cells in sample k; k Then, the IMC image is normalized as follows:

[0041] X scaled,k =r k ·X denoise,k

[0042] Among them, X scaled,k X represents the normalized pixel value of sample k. denoise,k Let K be the pixel value after noise reduction for sample k, where K is the total number of samples.

[0043] 10. The method described in Project 2 is to pass the normalized image through a filter, such as a median filter, preferably a 5×5 median filter, and finally output the noise-reduced and normalized IMC image. The protein expression value of the cell is obtained by quantifying the noise-reduced and normalized IMC image according to the method in step 3) of Project 1.

[0044] 11. The method described in any one of items 1-10 further includes visualization of the denoising result, including setting all pixels outside the range of positive cells to 0, then performing Gaussian filtering on the resulting image, and visualizing the image of the filtering result. Attached Figure Description

[0045] Figure 1 This is a flowchart of image noise reduction and protein expression value normalization between samples for IMC images.

[0046] Figure 2 This is the single-cell segmentation result obtained from step 1) of Project 1 (in this example, we use the Dice-XMBD method for single-cell segmentation of the image). The left image shows the cell nucleus channel of the original image, and the right image shows the single-cell segmentation image.

[0047] Figure 3 This is the result of preprocessing step 2) of project 1. Figure 3 In the image 'a': the left image is the original image of the CD3 protein channel, and the right image is the image after Arcsinh conversion. Figure 3 In the image b: the left image is the image after Arcsinh transformation, and the right image is the image after removing "hot pixels".

[0048] Figure 4 This is a diagram showing the noise region identified in Project 5 when the minimum expression threshold parameter is set to 0.05 times the 99th percentile pixel value of the original IMC image (default parameter value), and a diagram showing the extraction of random pseudo-cells in Project 4. The left image shows the distinction between the real cell expression region and the noise region, with green representing the cell protein expression region and black representing the noise region; the right image is a diagram showing the extraction of random pseudo-cells.

[0049] Figure 5 This is a schematic diagram illustrating the calculation of FDR and determination of the protein expression threshold for positive cells in step 1) of project 2. The blue curve represents the distribution curve of protein expression levels in random pseudo-cells, and the blue shaded area represents the number of false positive cells; the red curve represents the distribution curve of cell protein expression levels within the image, and the red shaded area represents the number of true positive cells.

[0050] Figure 6 The image shows the noise reduction effect after setting the number of random pseudo-cells to 1000, the false detection rate (FDR) to 0.01, setting all pixels outside the range of positive cells to 0, and applying a 5×5 Gaussian filter. The left image is the image before noise reduction, and the right image is the image after noise reduction.

[0051] Figure 7 This is the result of a manual noise reduction method for CD3 protein channel images.

[0052] Figure 8 This is a diagram showing the effect of applying a mean filtering noise reduction method to CD3 protein channel images. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0054] The following steps can be performed on IMC images to reduce noise and normalize protein levels between samples (the flowchart is shown below). Figure 1 (As shown).

[0055] Step 1: Cell segmentation;

[0056] We perform single-cell segmentation on the image using existing single-cell segmentation methods. In one example, we use a pre-trained Dice-XMBD model for single-cell segmentation. After cell segmentation, we obtain cell masks with pixel-level precision and the same size as the original image. We discard segmented regions with cell areas smaller than 5 pixels to obtain the single-cell segmentation result.

[0057] Step 2: Image preprocessing;

[0058] (1) Use the Arcsinh function to transform the image pixel values, compress the scale of the pixel values, and make the pixel values ​​more stable;

[0059] (2) Use a 5×5 image filter to remove "hot pixels" from the image. When the center pixel value of the median filter window is in the top 2% of all pixel values ​​and is at least 4 times higher than the median of all pixel values ​​in the window, the center pixel is identified as a "hot pixel," and its pixel value is adjusted to the median of all pixels in the window. This step reduces the noise impact of outlier "hot pixels."

[0060] Step 3: Quantify single-cell protein expression values;

[0061] After cell segmentation (step 1) and image preprocessing (step 2), the cell protein expression value before noise reduction is quantified into the statistical value of all pixel values ​​in the cell in the preprocessed image. In one example, we calculate the average value of all pixel values ​​in the cell to quantify the cell protein expression value before noise reduction.

[0062] Step 4: Extract a certain number of random pseudo-cells and calculate their protein expression values;

[0063] 4.1 Before extracting random pseudo-cells, distinguish between regions of high probability of real cell expression and noise regions in the image. Regions with high pixel values ​​are defined as regions of high confidence in real cell expression, while regions with very low pixel values ​​are defined as noise regions. The pixel values ​​of noise regions are set to zero, i.e., the pixel values ​​of the image are adjusted to:

[0064]

[0065] Among them, X ori It is the pixel value of the original IMC image, X θ It is the set minimum expression threshold, X adj These are the adjusted pixel values.

[0066] Next, a 5×5 median filter is applied to the adjusted image.

[0067] In the image obtained after the above processing steps, the pixel value of the noise region is 0, and the pixel value of the real cell expression region with high confidence is greater than 0, thus completing the distinction between the real cell expression region with high confidence and the noise region.

[0068] 4.2 Extracting Random Pseudo-Cells that Meet the Rules. In one example, we extract 1000 random pseudo-cells. Calculate the major axis, minor axis, orientation, and area of ​​all cells obtained from cell segmentation in step 1, and fit the major axis, minor axis, and orientation using separate Gaussian models. Random numbers are drawn from the above three distributions, and a coordinate point is randomly selected on the image. An elliptical random pseudo-cell is formed on the image centered at this point. This random pseudo-cell needs to meet the following three conditions:

[0069] a) Condition 1: The area of ​​the random pseudo-cell is within the range of the area of ​​all cells;

[0070] b) Condition 2: When part of the random pseudo-cell extends beyond the image boundary, ensure that the area of ​​the random pseudo-cell within the image is greater than 5 pixels.

[0071] c) Condition 3: All pixels of the random pseudo-cell are within the noise region mentioned above.

[0072] 4.3 Quantify the protein expression values ​​of random pseudocells using the method in step 3.

[0073] 4.4 The average protein expression value of the extracted random pseudo-cells is defined as the background noise level of the image, and its calculation method is as follows:

[0074] X noise =mean(X) decoy,i )

[0075] Among them, X decoy,iThe average protein expression value of the pixel on random pseudocell i in the IMC image, i.e.

[0076] X decoy,i =C i / N i

[0077] Among them, C i N represents the protein expression level of random pseudocell i. i denoted as the total number of pixels in random pseudo-cell i.

[0078] Step 5: Eliminate the influence of background noise;

[0079] To eliminate the influence of background noise in the image, the pixel value of the image obtained in step 2 before noise reduction is subtracted from the background noise level in step 4 (step 4.4), which is the pixel value X of the image after noise reduction. denoise for:

[0080] X denoise =max(0,X) origin -X denoise )

[0081] Among them, X denoise X represents the pixel value after noise reduction. origin X represents the pixel value before noise reduction. noise This represents the background noise level of the image.

[0082] Step 6: Determine the protein expression threshold for positive cells;

[0083] In this invention, the error detection rate (FDR) is calculated as follows:

[0084]

[0085] Wherein, FP (False Positive) is the number of false positive cells, that is, the number of cells with protein expression values ​​greater than the threshold among the 1000 random pseudo cells extracted in step 4, and TP (True Positive) is the number of true positive cells, that is, the number of cells with protein expression values ​​greater than the threshold before noise reduction in step 3.

[0086] In one example, we set the false detection rate (FDR) to 0.01, which controls the proportion of false positive cells in positive cells to 0.01, thereby determining the protein expression value threshold for positive cells. Cells with protein expression values ​​greater than or equal to the threshold before noise reduction in step 3 are defined as positive cells in the image, and cells with values ​​less than the threshold are defined as negative cells.

[0087] Step 7: Normalize cellular protein expression values;

[0088] The expression intensity of the same protein biomarker varies among different samples. Therefore, it is necessary to normalize the cellular protein expression values ​​of all samples for the same protein biomarker in order to conduct reasonable subsequent biological data analysis, such as unsupervised cell clustering.

[0089] (1) For each protein marker, calculate the noise-reduced protein expression value of the positive cells identified in step 6 in each sample. Then, using the protein expression value of the positive cells in each sample, obtain the normalization scaling parameter r for each sample according to the following method. k :

[0090]

[0091] in, is the average protein expression value of positive cells in sample k, where K is the total number of samples. (2)

[0092] Obtain the normalized scaling parameter r k Then, the IMC image is normalized as follows:

[0093] X scaled,k =r k ·X denoise,k

[0094] Among them, X scaled,k X represents the normalized pixel value of sample k. denoise,k The pixel value after noise reduction for sample k.

[0095] (3) In order to further remove salt and pepper noise, the image after background noise elimination and normalization is filtered, preferably a 5×5 median filter, and the final denoised and normalized IMC image is output. The method in step 3 is used to quantify protein in single cells and the final denoised and normalized cell protein expression value is output.

[0096] Step 8: Interactive parameter tuning and visualization of denoising results (optional).

[0097] Building upon step 6, all pixels outside the range of positive cells are set to 0. Then, the resulting image is Gaussian filtered and visualized, presenting the denoised image to researchers for observation, analysis, and parameter adjustment. User-adjustable parameters include: minimum expression threshold (default value is 0.05 times the 99th percentile pixel value of the preprocessed image), number of false cells (default value is 1000), and false detection rate (FDR) (default value is 0.01).

[0098] Example

[0099] The method of this invention was applied to IMC images of melanoma patients. The dataset was acquired as follows:

[0100] First, target proteins were screened to obtain an antibody panel (containing 25 pre-labeled antibodies, 10 third-party antibodies, and DNA). Different antibody concentration gradients were designed for preliminary experiments to obtain the optimal antibody concentration ratio. Then, paraffin sections of the sample tissue were stained. After antibody incubation, a panoramic scan of the tissue sections was performed using the Hyperion tissue mass spectrometry imaging system (Fluidigm), and suitable regions of interest (ROIs) were selected for imaging.

[0101] We used this method to denoise and normalize the original IMC image. The experimental parameters and results for each step are as follows:

[0102] The single-cell segmentation results obtained in step 1 using Dice-XMBD (code address: https: / / github.com / xmuyulab / Dice-XMBD) are shown in the image. Figure 2 );

[0103] Step 2 involves image preprocessing, including pixel value transformation using the Arcsinh function (the code is the asinh function from the math module of the Python standard library), and applying a 5×5 median filter to adjust outlier hot pixels located in the top 2% that are higher than the median pixel value within a 4x window to the median pixel value within the window. The results are shown in [link to image processing steps]. Figure 3 ;

[0104] Step 3: After cell segmentation (Step 1) and image preprocessing (Step 2), the cell protein expression value before noise reduction is quantified into the average value of all pixels in the cell in the preprocessed image;

[0105] Step 4: Extract a certain number of random pseudo-cells and calculate their protein expression values. First, set the minimum expression threshold parameter to 0.05 times the 99th percentile pixel value of the original IMC image. Set pixel values ​​below the minimum expression threshold to 0. Then, pass the resulting image through a 5×5 median filter to distinguish between real cell expression regions with high confidence and noise regions. Next, based on the single-cell segmentation results obtained in Step 1, obtain the area, major axis, minor axis, and orientation of all cells in the IMC image. Use separate Gaussian models to fit the major axis, minor axis, and orientation respectively. Finally, extract random numbers for the major axis, minor axis, and orientation from the above three distributions and form pseudo-random cells with an area within the range of all cell areas in the noise region (see schematic diagram). Figure 4 The number of random pseudo-cells was set to 1000. After extracting 1000 pseudo-cells, the protein expression value of the random pseudo-cells was quantified using the method in step 3. The average protein expression value of the extracted random pseudo-cells was defined as the background noise level of the image.

[0106] Step 5: Subtract the background noise level from the image in Step 4 from the pixel values ​​of the image before noise reduction after preprocessing in Step 2, thereby eliminating the influence of background noise in the image.

[0107] Step 6: Generate the distribution of protein expression values ​​from the 1000 random pseudo-cells extracted in Step 4 and the undenoised protein expression values ​​obtained in Step 3. The number of false positive cells is defined as the number of cells with protein expression values ​​greater than a threshold among the 1000 random pseudo-cells. The number of true positive cells is defined as the number of cells with protein expression values ​​greater than a threshold before denoising. Calculate the FDR (Fake Response Rate), i.e., the proportion of false positive cells among positive cells, and determine the protein expression value threshold for positive cells to distinguish between positive (true) and negative cells. See [link to detailed process] for further information. Figure 5 ;

[0108] Step 7: For each protein marker, calculate the denoised protein expression value of the positive cells identified in Step 6 in each sample, and calculate the mean value of the positive cell protein expression value in each sample. The normalization scaling parameter for each sample is defined as the ratio of the mean value of the largest positive cell protein expression value in all samples to the mean value of the positive cell protein expression value in this sample. Multiply the pixel value of the denoised image by the normalization scaling parameter and pass it through a 5×5 median filter to obtain the denoised and normalized IMC image. Quantify the denoised and normalized cell protein expression value on the denoised and normalized IMC image according to the method in Step 3.

[0109] Step 8: Set the number of random fake cells to 1000, the false detection rate (FDR) to 0.01, set all pixels outside the range of positive cells to 0, apply a 5×5 Gaussian filter, and see the result image after noise reduction. Figure 6 .

[0110] Comparative Example

[0111] Manually determine signal threshold for noise reduction

[0112] Using image visualization software (Fluidigm's MCD Viewer software, version v1.0.560.6), each channel of the IMC image was observed. Different minimum and maximum protein expression thresholds could be manually set in the channel settings. The software identified signals below the minimum protein expression threshold as noise and set their values ​​to zero, while signals above the maximum protein expression threshold were set to the maximum threshold, thus eliminating the influence of hot pixel values ​​on subsequent single-cell protein quantification analysis. In the example of this invention, in the CD3 protein channel image, the minimum protein expression threshold was manually set to 3 and the maximum protein expression threshold to 4.8 in the software, resulting in the software-denoised image. Figure 7 ).

[0113] Because of the differences between different samples and different proteins, the noise threshold of different protein channels in a sample is not the same. The method of manually determining the threshold to reduce noise requires a lot of manpower and time, and the noise reduction result is also highly subjective.

[0114] Application of medium-range filtering for noise reduction

[0115] Median filtering is used to filter the IMC image to eliminate noise similar to salt-and-pepper noise. In the example of this invention, the `medianBlur` function provided by the OpenCV-python software library (version 4.2.0.34) is used, with the window size parameter set to 5, to perform a 5×5 median filtering operation on the CD3 protein channel image. 5×5 median filtering sets the value of each pixel in the image to the median of all pixel values ​​within a 5×5 neighborhood window of that point, thereby performing local averaging on the image (see...). Figure 8 ).

[0116] pass Figure 6-8 The comparison shows that, compared to manually determining the threshold for noise reduction, the median filtering method is faster, but it may not completely eliminate clustered noise and may remove some weakly expressed protein signals, resulting in poor noise reduction results. The noise reduction method of this invention achieves the best effect (see...). Figure 6 ).

[0117] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1.A method for denoising and protein quantity normalization of images of imaging mass cytometry (IMC), comprising the following steps of denoising the images: 1) identifying single cells on the IMC images by an image single cell segmentation method; 2) pre-processing the IMC images; 3) quantifying the protein expression value of a cell as a statistical value of all pixel values in the segmented region of the cell; 4) selecting a certain number of random pseudo-cells on the images, and quantifying the protein expression value of the random pseudo-cells according to the method of step 3) to obtain the protein expression value of the random pseudo-cells; 5) calculating the background noise level of the images by the protein expression value of the random pseudo-cells, and obtaining the denoised IMC images after eliminating the influence of the background noise in the images, wherein the selection of the random pseudo-cells in step 4) further comprises selecting random pseudo-cells on the images, and the method of selecting the random pseudo-cells is as follows: 4.1) calculating the long axis, short axis, direction and area size of all cells obtained by cell segmentation in step 1), and fitting the long axis, short axis and direction using separate Gaussian models; 4.2) extracting random numbers from the fitting distribution of the long axis, short axis and direction, and randomly selecting a coordinate point on the images, and forming an elliptical random pseudo-cell centered at the point on the images, which needs to meet the following two conditions: condition 1: the area size of the random pseudo-cell is within the range of the area sizes of all cells; condition 2: when part of the position of the random pseudo-cell exceeds the boundary of the images, the area of the random pseudo-cell within the images is greater than 5 pixel points. 2.The method of claim 1, further comprising the following steps of normalizing the images: comparing the protein expression value distribution of the cells before denoising obtained in step 3) with the protein expression value distribution of the random pseudo-cells obtained in step 4), and determining the protein expression value threshold of positive cells by controlling the false discovery rate (FDR) ; and calculating the normalization scaling parameter of each sample using the protein expression values of the positive cells on multiple samples, and normalizing the IMC images of each sample. 3.The method of claim 1, wherein the pre-processing comprises converting the image pixel values by an Arcsinh function to compress the scale of the pixel values; and removing "hot pixel points" on the images by a filter, wherein the "hot pixel point" is a pixel point whose central pixel value in the median filter window is located in the top 2% of all pixel expression values, and is at least 4 times higher than the median value of all pixel values in the window. ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​

Images