Cancer risk assessment system based on multi-omics data

Abstract

The invention relates to the technical field of multi-omics, in particular to a cancer risk assessment system based on multi-omics data, which comprises a proteomics acquisition module, a transcriptomics acquisition module, a genomics acquisition module, a radiomics acquisition module and a central processing module, according to the method, the medical image noise layer simulates the noise according to different irradiation parameters, imaging parameters and imaging technology categories through the medical image noise parameter adjustment dictionary, so that the dynamic convolution layer and the convolution encoder learn how to extract key features of the medical image as much as possible while removing the noise; a multi-modal second-order auto-encoder is constructed through a transfer learning technology to perform multi-modal fusion and feature extraction on multi-omics data of different modals, and finally the multi-modal second-order auto-encoder is transformed into a multi-omics cancer risk assessment neural network model. According to the method, multi-omics, multi-mode and practical cancer risk assessment is realized through a transfer learning technology, a multi-mode fusion technology and a noise simulation technology.

Description

Technical Field

[0001] The present invention relates to the field of multi-omics technology, and in particular to a cancer risk assessment system based on multi-omics data. Background Art

[0002] A cancer risk assessment system is a tool that uses medical knowledge and data analysis techniques to predict an individual's risk of developing a certain cancer. The cancer risk assessment system is usually based on a large amount of epidemiological data, genetic information, lifestyle and environmental factors.

[0003] Now with the popularization of the concept of multi-omics, the industry has begun to use multi-omics technology to assess cancer risk. Cancer risk assessment based on multi-omics is mainly divided into two technical routes: statistics and deep learning.

[0004] The iClusterPlus model is a typical statistical technology route. It is a regularized regression model that establishes latent variables and observation values ​​for different data types. Its disadvantage is that it relies too much on the prior knowledge of professionals for feature selection, which leads to low universality and requires re-modeling every time a new scenario is encountered.

[0005] The autoencoder in the deep learning technology route solves this problem. As a self-supervised model, the autoencoder extracts features from the input data through the encoder and gradually passes it to the bottleneck layer. The decoder needs to reconstruct the input data as much as possible based on the feature representation compressed in scale in the bottleneck layer. Due to the size limit of the bottleneck layer, the encoder in the autoencoder must learn how to extract the features from the input data that can best help the decoder reconstruct, that is, the features that best represent the input data during the training process. The advantage of this technical route is that it is highly universal and does not need to be remodeled every time a new scene is encountered, but its problem is that the multi-omics field has its own particularity. In the multi-omics field, it is often necessary to process hundreds of different features at the same time. The burden of selecting and representing features is already too heavy, not to mention that the multicollinearity between these features is quite serious. A considerable part of the information in transcriptomics and proteomics has been represented by genomic information. Directly inputting medical images will lead to convergence difficulties and feature confusion. Summary of the invention

[0006] The present invention overcomes the shortcomings of the prior art and provides a cancer risk assessment system based on multi-omics data, including a proteomics acquisition module, a transcriptomics acquisition module, a genomics acquisition module, an imaging omics acquisition module and a central processing module. The present invention uses a medical image noise adjustment dictionary to enable the medical image noise layer to simulate noise according to different imaging technology categories, irradiation parameters and imaging parameters, so that the dynamic convolution layer and the convolution encoder learn how to extract the key features of the medical image as much as possible while removing the noise. A multimodal second-order autoencoder is constructed through transfer learning technology to perform multimodal fusion and feature extraction on multi-omics data of different modalities. Finally, the multimodal second-order autoencoder is transformed into a multi-omics cancer risk assessment neural network model. The present invention realizes multi-omics, multi-modal and practical cancer risk assessment through transfer learning technology, multimodal fusion technology and noise simulation technology.

[0007] A cancer risk assessment system based on multi-omics data, comprising: a proteomics acquisition module, a transcriptomics acquisition module, a genomics acquisition module, an imaging omics acquisition module and a central processing module;

[0008] The proteomics acquisition module, the transcriptomics acquisition module, the genomics acquisition module and the imaging omics acquisition module are electrically connected to the central processing module respectively;

[0009] The proteomics acquisition module acquires proteomics data and uploads it to the central processing module;

[0010] The transcriptomics acquisition module collects transcriptomics data and uploads it to the central processing module;

[0011] The genomics acquisition module collects genomics data and uploads it to the central processing module;

[0012] The radiomics acquisition module acquires radiomics data and uploads it to the central processing module, wherein the radiomics data includes irradiation parameters, imaging parameters, medical images, and imaging technology categories, the irradiation parameters include irradiation time and irradiation dose, and the imaging parameters include image sampling rate and pixel spacing;

[0013] The central processing module uses transfer learning technology to construct a multi-omics cancer risk assessment neural network model and uses the multi-omics cancer risk assessment neural network model to assess cancer risk.

[0014] Furthermore, the process of the central processing module using transfer learning technology to construct a multi-omics cancer risk assessment neural network model includes the following steps:

[0015] Step S1: construct a medical image denoising dataset and a medical image noise layer and use the medical image denoising dataset to obtain a medical image noise parameter dictionary;

[0016] Step S2: collecting multi-omics data sets, constructing a medical image denoising convolutional autoencoder, and training the medical image denoising convolutional autoencoder;

[0017] Step S3: construct a contraction autoencoder and train the contraction autoencoder using a multi-omics dataset;

[0018] Step S4: constructing a multimodal second-order autoencoder using a contraction autoencoder and a medical image denoising convolutional autoencoder and training the multimodal second-order autoencoder using a multi-omics dataset;

[0019] Step S5: Use a multimodal second-order autoencoder to build a multi-omics cancer risk assessment neural network model and use a multi-omics dataset to train the multi-omics cancer risk assessment neural network model.

[0020] Furthermore, step S1 specifically includes the following steps:

[0021] Step S11: collecting medical images before denoising, corresponding medical images after denoising, corresponding irradiation parameters, corresponding imaging parameters and corresponding imaging technology categories to construct a medical image denoising data set;

[0022] Step S12: construct a medical image noise layer, which adds Gaussian noise, Poisson noise and salt and pepper noise to its input. The hyperparameters of the medical image noise layer include the mean of Gaussian noise, the variance of Gaussian noise, the mean of Poisson noise, the variance of Poisson noise and the proportion of salt and pepper noise. The hyperparameters of the medical image noise layer constitute a hyperparameter group:

[0023] ;

[0024] in, represents a hyperparameter group, represents the mean of Gaussian noise, represents the variance of Gaussian noise, represents the mean of Poisson noise, represents the variance of Poisson noise, Represents the proportion of salt and pepper noise;

[0025] Step S13: Divide the medical image denoising data set according to the imaging technology category, designate the denoised medical image as the input of the medical image noise layer, designate the medical image before denoising as the target output of the medical image noise layer, construct a noise simulation symbolic regression model, use the symbolic regression method to maximize the average structural similarity index to search for the best noise simulation symbolic regression model corresponding to different imaging technology categories, extract the parameter adjustment formula set from the best noise simulation symbolic regression model, and construct a medical image noise adjustment parameter dictionary. The single key-value pair in the medical image noise adjustment parameter dictionary is as follows:

[0026] ;

[0027] ;

[0028] ;

[0029] ;

[0030] in, Represents a single key-value pair in the medical image noise parameter dictionary. Represents the imaging technology category, Represents the parameter adjustment formula set corresponding to the imaging technology category, The parameter adjustment formula representing the mean value of Gaussian noise, The parameter tuning formula representing the variance of Gaussian noise, The parameter adjustment formula representing the mean value of Poisson noise, The parameter tuning formula representing the variance of Poisson noise, The parameter adjustment formula representing the proportion of salt and pepper noise is: represents the irradiation parameter, represents the imaging parameters, represents the irradiation time, represents the radiation dose, represents the image sampling rate, Represents pixel pitch.

[0031] Furthermore, step S2 specifically includes the following steps:

[0032] Step S21: collecting multi-omics data sets;

[0033] Step S22: constructing a dynamic convolutional layer and a convolutional autoencoder, wherein the convolutional autoencoder includes a convolutional encoder, a convolutional bottleneck layer and a convolutional decoder;

[0034] Step S23: connecting the medical image noise layer, the dynamic convolution layer and the convolutional autoencoder in sequence to construct a medical image denoising convolutional autoencoder;

[0035] Step S24: extracting radiomics data from the multi-omics dataset;

[0036] Step S25: Use the medical images in the imaging omics data to train the medical image denoising convolutional autoencoder. During the training process, while inputting the medical image into the medical image noise layer, query the parameter adjustment formula set in the medical image noise adjustment parameter dictionary according to the imaging technology category corresponding to the medical image. Extract the irradiation parameters and imaging parameters from the imaging omics data and input them into the parameter adjustment formula set to obtain a hyperparameter group, and adjust the medical image noise layer according to the hyperparameter group. The dynamic convolution layer selects different convolution kernels according to the noise simulated by the medical image noise layer for different irradiation parameters, imaging parameters and imaging technology categories. The convolution encoder learns during the training process how to extract the key features of the medical image as much as possible while removing the noise, and the convolution bottleneck layer outputs the denoised key feature map.

[0037] Furthermore, step S3 specifically includes the following steps:

[0038] Step S31: extracting proteomics data, transcriptomics data and genomics data from the multi-omics data set, and aggregating the proteomics data, transcriptomics data and genomics data into a protein-transcription-gene data set;

[0039] Step S32: construct a contraction autoencoder, wherein the contraction autoencoder includes a contraction encoder, a contraction bottleneck layer and a contraction decoder;

[0040] Step S33: Use the protein-transcript-gene dataset to train the contraction autoencoder. During the training process, the contraction encoder learns how to reduce the dimensions of proteomics data, transcriptomics data, and genomics data and outputs the reduced-dimensional feature representation through the contraction bottleneck layer.

[0041] Furthermore, step S4 specifically includes the following steps:

[0042] Step S41: extracting a contraction encoder and a contraction bottleneck layer from the contraction autoencoder, and extracting a dynamic convolution layer, a convolution encoder, and a convolution bottleneck layer from the medical image denoising convolution autoencoder;

[0043] Step S42: constructing a concatenated fusion layer and a sparse autoencoder, wherein the sparse autoencoder includes a sparse encoder, a sparse bottleneck layer and a sparse decoder;

[0044] Step S43: connecting the shrinkage bottleneck layer and the convolution bottleneck layer to the splicing fusion layer, connecting the splicing fusion layer to the sparse autoencoder, and constructing a multimodal second-order autoencoder;

[0045] Step S44: Freeze the weights of the contraction encoder, the contraction bottleneck layer, the dynamic convolution layer, the convolution encoder, and the convolution bottleneck layer, specify the protein-transcription-gene dataset as the input of the contraction encoder, specify the medical images in the imaging omics data as the input of the dynamic convolution layer, extract clinical data from the multi-omics dataset as the target output of the sparse autoencoder, the clinical data includes the main site of tumor occurrence, AJCC pathological stage and the specific type of cancer, train the multimodal second-order autoencoder, the sparse encoder learns during the training process how to extract the key features in the input as much as possible while fusing the inputs of different modalities, and the sparse bottleneck layer outputs the multimodal key feature representation.

[0046] Furthermore, step S5 specifically includes the following steps:

[0047] Step S51: remove the sparse decoder in the multimodal second-order autoencoder, build a multi-layer perceptron and connect the sparse bottleneck layer to the multi-layer perceptron, specify the loss function of the multi-layer perceptron as categorical_crossentropy, and build a multi-omics cancer risk assessment neural network model;

[0048] Step S52: One-hot encode the specific types of cancer in the multi-omics dataset to generate a one-hot code for the cancer type, specify the protein-transcription-gene dataset as the input of the contraction encoder, specify the medical images in the imaging omics data as the input of the dynamic convolution layer, specify the one-hot code for the cancer type as the target output of the multi-layer perceptron, freeze the weights of each layer except the multi-layer perceptron in the multi-omics cancer risk assessment neural network model, train the multi-omics cancer risk assessment neural network model, and the multi-layer perceptron learns how to map the multimodal key feature representation into the probability distribution of the cancer type during the training process.

[0049] Furthermore, the process of assessing cancer risk using the multi-omics cancer risk assessment neural network model specifically includes the following steps:

[0050] Step P1: collecting the target user's proteomics data, the target user's transcriptomics data, the target user's genomics data, and the target user's imaging data;

[0051] Step P2: Aggregate the target user's proteomics data, the target user's transcriptomics data, and the target user's genomics data into a protein-transcriptome-gene data instance, and extract the target user's medical image from the target user's radiomics data;

[0052] Step P3: Input the protein-transcription-gene data instance into the contraction encoder of the multi-omics cancer risk assessment neural network model, input the target user's medical image into the dynamic convolution layer of the multi-omics cancer risk assessment neural network model, and the multi-omics cancer risk assessment neural network model outputs the cancer type probability distribution of the target user;

[0053] Step P4: Extract the cancer type with the highest probability from the target user's cancer type probability distribution as the highest risk cancer type, and extract the probability corresponding to the highest risk cancer type.

[0054] The beneficial effects of the present invention are as follows:

[0055] (1) The present invention provides a cancer risk assessment system based on multi-omics data, including a proteomics acquisition module, a transcriptomics acquisition module, a genomics acquisition module, an imaging omics acquisition module and a central processing module. The present invention uses a medical image noise adjustment dictionary to enable the medical image noise layer to simulate noise according to different imaging technology categories, irradiation parameters and imaging parameters, so that the dynamic convolution layer and the convolution encoder learn how to extract the key features of the medical image as much as possible while removing the noise. A multimodal second-order autoencoder is constructed through transfer learning technology to perform multimodal fusion and feature extraction on multi-omics data of different modalities. Finally, the multimodal second-order autoencoder is transformed into a multi-omics cancer risk assessment neural network model. The present invention realizes multi-omics, multi-modal, and practical cancer risk assessment through transfer learning technology, multimodal fusion technology and noise simulation technology.

[0056] (2) The medical image noise layer proposed in the present invention is designed using prior knowledge in the field of medical imaging, and Gaussian noise, Poisson noise and salt and pepper noise are added to the input. Gaussian noise is used to simulate structural noise in medical images. Structural noise is caused by the working conditions of medical imaging equipment and the characteristics of imaging technology. For example, in MRI imaging, the noise caused by the uneven magnetic field conforms to the Gaussian distribution. Poisson noise is used to simulate the quantum noise caused by the quantum fluctuations of imaging rays. The quantum noise conforms to the Poisson distribution. Salt and pepper noise is caused by the transmission process of the channel and the working conditions of the image sensor, and manifests as randomly appearing isolated black and white noise points. The medical image noise parameter adjustment dictionary proposed in the present invention enables the medical image noise layer to dynamically adjust the noise simulation parameters according to different imaging technology categories, irradiation parameters and imaging parameters, helping the dynamic convolution layer and the convolution encoder to learn how to extract the key features of the medical image as much as possible while removing the noise, thereby realizing automatic denoising of the medical image.

[0057] (3) The design of the medical image noise parameter adjustment dictionary of the present invention is based on the prior knowledge in the field of medical imaging. The symbolic regression technology is used to fit the relationship between the irradiation parameters and the imaging parameters and the medical image noise layer. Since the cause of quantum noise is the quantum fluctuation of the imaging ray, the input variables of the parameter adjustment formula of the mean of Poisson noise and the parameter adjustment formula of the variance of Poisson noise are irradiation parameters. Compared with the fixed mean and variance, the parameter adjustment formula used to determine the mean and variance takes into account the accumulation of statistical fluctuations and physical factors caused by changes in irradiation time and irradiation dose, which is closer to the characteristics of quantum noise in reality. Since the increase of irradiation parameters will increase the intensity of Gaussian noise while increasing the signal strength, the imaging parameters will affect the output image. Sensitivity to noise, so Gaussian noise is affected by both Gaussian noise and illumination parameters. Since salt and pepper noise is not generated in the imaging process but in the transmission and processing process, it is not affected by illumination parameters. The noise simulation scheme combining the medical image noise parameter adjustment dictionary with the medical image noise layer of the present invention introduces the idea of ​​physical simulation, and fits the relationship between illumination parameters, imaging parameters and hyperparameter groups through symbolic regression, which is closer to the real noise characteristics. Compared with black box models such as adversarial generative models, it has better interpretability and is closer to the real noise generation process than simple data-driven models, realizing realistic, transparent and interpretable medical imaging noise simulation.

[0058] (3) Existing medical image denoising technology treats different types of medical images equally. Although this improves the efficiency of medical image processing, it leads to the problem of feature loss caused by over-denoising, and ultimately results in a medical image that is too smooth and lacks reference value. In other words, existing medical image denoising technology often identifies overly subtle features as noise points and removes them. The dynamic convolution layer selects different convolution kernels based on the features in the medical image. By selecting different convolution kernels, it implicitly injects the recognition of noise features caused by different irradiation parameters, imaging parameters and imaging technology categories into the medical image, thereby assisting the convolution encoder to remove noise more specifically, solving the problem of over-denoising in existing medical image denoising technology.

[0059] (4) Limited by experimental conditions and sequencing processes, there are always some subtle errors and disturbances in proteomics data, transcriptomics data and genomics data, and the regularization technology of the shrinkage autoencoder enables it to resist extremely small input disturbances. The present invention pre-trains the shrinkage autoencoder and the medical image denoising convolutional autoencoder respectively and introduces the shrinkage autoencoder and the medical image denoising convolutional autoencoder into the multimodal second-order autoencoder using transfer learning technology, making the system as a whole more interpretable, facilitating subsequent expansion, and improving users' confidence in the system. In addition, the transfer learning technology can also enable the encoder to learn more general feature representations, thereby improving the universality of the system.

[0060] (5) The existing multimodal autoencoding schemes mainly have the following three architectures:

[0061] Data-level fusion: First, the data is fused and then processed uniformly by an autoencoder. This solution is obviously not suitable for cross-dimensional multimodal fusion, but only for multimodal fusion of the same dimension such as audio and text, or images and matrices.

[0062] Multimodal autoencoder: Instead of fusing data, it processes data by feeding different modal inputs into different parts of the multimodal autoencoder. However, the problem is that multi-omics tasks are more specific than conventional multimodal tasks. In the field of multi-omics, hundreds of different features often need to be processed at the same time. For a single encoder, the burden of selecting features is already too heavy. Moreover, the multicollinearity between these features is quite serious. A considerable part of the information in transcriptomics and proteomics has been represented by genomic information. Directly inputting the information of another modality will lead to convergence difficulties and feature confusion.

[0063] Output vector concatenation: The outputs of different modal autoencoders are connected to obtain a multimodal joint representation. However, directly concatenating vectors obviously cannot represent the relationship between different modal information.

[0064] The multimodal second-order autoencoder proposed in the present invention realizes feature extraction of information between different modalities by first dividing and conquering and then comprehensively analyzing the method by introducing the concept of second-order autoencoder, first eliminating multicollinearity and then performing multimodal fusion, effectively avoiding the convergence difficulty and feature confusion problems of a single multimodal autoencoder, and the re-representation function of the second-order autoencoder further explores the relationship between different modal information. In summary, the multimodal second-order autoencoder proposed in the present invention has significant improvements over the existing multimodal autoencoder scheme in preventing the aggravation of multicollinearity caused by multimodal fusion and exploring the relationship between different modal information, thereby indirectly improving the rationality and accuracy of cancer risk assessment.

[0065] (6) The multimodal second-order autoencoder proposed in the present invention does not use the default self-supervised learning method, but uses clinical data as the target output of the sparse autoencoder. This is because the sparse decoder is not suitable for directly predicting the probability distribution of different cancer types, and cannot output medical images, proteomics data, transcriptomics data and genomics data at the same time. If only proteomics data, transcriptomics data and genomics data are output, the importance of medical images is ignored. The selection of clinical data as the target output of the sparse autoencoder is actually a compromise. The reason for selecting clinical data is that clinical data in the multi-omics data set can best characterize the specific characteristics of cancer. The process of forcing the multimodal second-order autoencoder to establish the association between multi-omics data and clinical data forces the sparse encoder to extract the most critical part of the multimodal key feature representation, further reducing the multicollinearity of the multimodal key feature representation, facilitating the multi-layer perceptron to establish mapping, and ultimately improving the prediction accuracy of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 A module diagram of a cancer risk assessment system based on multi-omics data proposed by the present invention;

[0067] Figure 2 This is a schematic diagram of the model of the medical image denoising convolutional autoencoder proposed in the present invention;

[0068] Figure 3 This is a simplified model diagram of the contraction autoencoder;

[0069] Figure 4 A schematic diagram of the model of the multimodal second-order autoencoder proposed in the present invention;

[0070] Figure 5 This is a schematic diagram of the multi-omics cancer risk assessment neural network model proposed in the present invention.

[0071] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention. DETAILED DESCRIPTION

[0072] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the examples given are only used to explain the present invention, and are not used to limit the scope of the present invention.

[0073] Example 1, see Figure 1 , a cancer risk assessment system based on multi-omics data, comprising: a proteomics acquisition module, a transcriptomics acquisition module, a genomics acquisition module, an imaging omics acquisition module and a central processing module;

[0074] The proteomics acquisition module, the transcriptomics acquisition module, the genomics acquisition module and the imaging omics acquisition module are electrically connected to the central processing module respectively;

[0075] The proteomics acquisition module acquires proteomics data and uploads it to the central processing module;

[0076] The transcriptomics acquisition module collects transcriptomics data and uploads it to the central processing module;

[0077] The genomics acquisition module collects genomics data and uploads it to the central processing module;

[0078] The radiomics acquisition module acquires radiomics data and uploads it to the central processing module, wherein the radiomics data includes irradiation parameters, imaging parameters, medical images, and imaging technology categories, the irradiation parameters include irradiation time and irradiation dose, and the imaging parameters include image sampling rate and pixel spacing;

[0079] The central processing module uses transfer learning technology to construct a multi-omics cancer risk assessment neural network model and uses the multi-omics cancer risk assessment neural network model to assess cancer risk. In this embodiment, the imaging omics data uses the DICOM format, and the imaging technology category, irradiation parameters and imaging parameters come from the DICOM header file.

[0080] Embodiment 2: Based on the above embodiment, the process of the central processing module using transfer learning technology to construct a multi-omics cancer risk assessment neural network model includes the following steps:

[0081] Step S1: construct a medical image denoising dataset and a medical image noise layer and use the medical image denoising dataset to obtain a medical image noise parameter dictionary;

[0082] Step S2: collecting multi-omics data sets, constructing a medical image denoising convolutional autoencoder, and training the medical image denoising convolutional autoencoder;

[0083] Step S3: construct a contraction autoencoder and train the contraction autoencoder using a multi-omics dataset;

[0084] Step S4: constructing a multimodal second-order autoencoder using a contraction autoencoder and a medical image denoising convolutional autoencoder and training the multimodal second-order autoencoder using a multi-omics dataset;

[0085] Step S5: Use a multimodal second-order autoencoder to build a multi-omics cancer risk assessment neural network model and use a multi-omics dataset to train the multi-omics cancer risk assessment neural network model.

[0086] Embodiment 3: This embodiment is based on the above embodiment, and step S1 specifically includes the following steps:

[0087] Step S11: collecting medical images before denoising, corresponding medical images after denoising, corresponding irradiation parameters, corresponding imaging parameters and corresponding imaging technology categories to construct a medical image denoising data set;

[0088] Step S12: construct a medical image noise layer, which adds Gaussian noise, Poisson noise and salt and pepper noise to its input. The hyperparameters of the medical image noise layer include the mean of Gaussian noise, the variance of Gaussian noise, the mean of Poisson noise, the variance of Poisson noise and the proportion of salt and pepper noise. The hyperparameters of the medical image noise layer constitute a hyperparameter group:

[0089] ;

[0090] in, represents a hyperparameter group, represents the mean of Gaussian noise, represents the variance of Gaussian noise, represents the mean of Poisson noise, represents the variance of Poisson noise, Represents the proportion of salt and pepper noise;

[0091] Step S13: Divide the medical image denoising data set according to the imaging technology category, designate the denoised medical image as the input of the medical image noise layer, designate the medical image before denoising as the target output of the medical image noise layer, construct a noise simulation symbolic regression model, use the symbolic regression method to maximize the average structural similarity index to search for the best noise simulation symbolic regression model corresponding to different imaging technology categories, extract the parameter adjustment formula set from the best noise simulation symbolic regression model, and construct a medical image noise adjustment parameter dictionary. The single key-value pair in the medical image noise adjustment parameter dictionary is as follows:

[0092] ;

[0093] ;

[0094] ;

[0095] ;

[0096] in, Represents a single key-value pair in the medical image noise parameter dictionary. Represents the imaging technology category, Represents the parameter adjustment formula set corresponding to the imaging technology category, The parameter adjustment formula representing the mean value of Gaussian noise, The parameter tuning formula representing the variance of Gaussian noise, The parameter adjustment formula representing the mean value of Poisson noise, The parameter tuning formula representing the variance of Poisson noise, The parameter adjustment formula representing the proportion of salt and pepper noise is: represents the irradiation parameter, represents the imaging parameters, represents the irradiation time, represents the radiation dose, represents the image sampling rate, Represents the pixel spacing. In this embodiment, the medical image noise layer is implemented by using the scikit-image library with layers.core.Lambda in Keras, and the symbolic regression method is implemented by using the PySR library.

[0097] Embodiment 4: This embodiment is based on the above embodiment. In step S13, the process of constructing the noise simulation symbol regression model specifically includes the following steps:

[0098] Step S131: establishing a Gaussian noise sign regression sub-model, specifying the input variables of the Gaussian noise sign regression sub-model as the irradiation parameter and the imaging parameter, and the output variables as the variance of the Gaussian noise and the mean of the Gaussian noise;

[0099] Step S132: establishing a Poisson noise sign regression sub-model, specifying the input variables of the Poisson noise sign regression sub-model as the irradiation parameters, and the output variables as the variance of the Poisson noise and the mean of the Poisson noise;

[0100] Step S133: establishing a salt and pepper noise sign regression sub-model, specifying the input variable of the salt and pepper noise sign regression sub-model as the imaging parameter, and the output variable as the proportion of salt and pepper noise;

[0101] Step S134: rewrite the loss function of PySR into an average structural similarity loss function. The calculation process of the average structural similarity loss function is to first extract a parameter adjustment formula set from a Gaussian noise symbol regression submodel, a Poisson noise symbol regression submodel, and a salt and pepper noise symbol regression submodel. In the process of adding noise to the medical image in the medical image noise layer, the irradiation parameters and the imaging parameters are input into the parameter adjustment formula set to obtain a hyperparameter group, and the medical image noise layer is adjusted according to the hyperparameter group. The average structural similarity index between the image output by the medical image noise layer and the medical image before denoising is calculated, and the loss value is obtained by inverting the average structural similarity index.

[0102] Step S135: Specify the number of iterations to be 100 and construct a noise simulation symbolic regression model.

[0103] Example 5, see Figure 2 , this embodiment is based on the above embodiment, and step S2 specifically includes the following steps:

[0104] Step S21: Collecting multi-omics data sets. In this embodiment, the multi-omics data sets are from The Cancer Genome Atlas and The Cancer Imaging Archive;

[0105] Step S22: constructing a dynamic convolutional layer and a convolutional autoencoder, wherein the convolutional autoencoder includes a convolutional encoder, a convolutional bottleneck layer and a convolutional decoder;

[0106] Step S23: connecting the medical image noise layer, the dynamic convolution layer and the convolutional autoencoder in sequence to construct a medical image denoising convolutional autoencoder;

[0107] Step S24: extracting radiomics data from the multi-omics dataset;

[0108] Step S25: Use the medical images in the imaging omics data to train the medical image denoising convolutional autoencoder. During the training process, while inputting the medical image into the medical image noise layer, query the parameter adjustment formula set in the medical image noise adjustment parameter dictionary according to the imaging technology category corresponding to the medical image. Extract the irradiation parameters and imaging parameters from the imaging omics data and input them into the parameter adjustment formula set to obtain a hyperparameter group, and adjust the medical image noise layer according to the hyperparameter group. The dynamic convolution layer selects different convolution kernels according to the noise simulated by the medical image noise layer for different irradiation parameters, imaging parameters and imaging technology categories. The convolution encoder learns during the training process how to extract the key features of the medical image as much as possible while removing the noise, and the convolution bottleneck layer outputs the denoised key feature map.

[0109] Example 6, see Figure 3 , this embodiment is based on the above embodiment, and step S3 specifically includes the following steps:

[0110] Step S31: extracting proteomics data, transcriptomics data and genomics data from the multi-omics data set, and aggregating the proteomics data, transcriptomics data and genomics data into a protein-transcription-gene data set;

[0111] Step S32: construct a contraction autoencoder, wherein the contraction autoencoder includes a contraction encoder, a contraction bottleneck layer and a contraction decoder;

[0112] Step S33: Use the protein-transcript-gene dataset to train the contraction autoencoder. During the training process, the contraction encoder learns how to reduce the dimensions of proteomics data, transcriptomics data, and genomics data and outputs the reduced-dimensional feature representation through the contraction bottleneck layer.

[0113] Embodiment 7, see Figure 4 , this embodiment is based on the above embodiment, and step S4 specifically includes the following steps:

[0114] Step S41: extracting a contraction encoder and a contraction bottleneck layer from the contraction autoencoder, and extracting a dynamic convolution layer, a convolution encoder, and a convolution bottleneck layer from the medical image denoising convolution autoencoder;

[0115] Step S42: constructing a concatenated fusion layer and a sparse autoencoder, wherein the sparse autoencoder includes a sparse encoder, a sparse bottleneck layer and a sparse decoder;

[0116] Step S43: connecting the shrinkage bottleneck layer and the convolution bottleneck layer to the splicing fusion layer, connecting the splicing fusion layer to the sparse autoencoder, and constructing a multimodal second-order autoencoder;

[0117] Step S44: Freeze the weights of the contraction encoder, the contraction bottleneck layer, the dynamic convolution layer, the convolution encoder, and the convolution bottleneck layer, specify the protein-transcription-gene dataset as the input of the contraction encoder, specify the medical images in the imaging omics data as the input of the dynamic convolution layer, extract clinical data from the multi-omics dataset as the target output of the sparse autoencoder, the clinical data includes the main site of tumor occurrence, AJCC pathological stage and the specific type of cancer, train the multimodal second-order autoencoder, the sparse encoder learns during the training process how to extract the key features in the input as much as possible while fusing the inputs of different modalities, and the sparse bottleneck layer outputs the multimodal key feature representation.

[0118] Embodiment 8, see Figure 5 , this embodiment is based on the above embodiment, and step S5 specifically includes the following steps:

[0119] Step S51: remove the sparse decoder in the multimodal second-order autoencoder, build a multi-layer perceptron and connect the sparse bottleneck layer to the multi-layer perceptron, specify the loss function of the multi-layer perceptron as categorical_crossentropy, and build a multi-omics cancer risk assessment neural network model;

[0120] Step S52: One-hot encode the specific types of cancer in the multi-omics dataset to generate a one-hot code for the cancer type, specify the protein-transcription-gene dataset as the input of the contraction encoder, specify the medical images in the imaging omics data as the input of the dynamic convolution layer, specify the one-hot code for the cancer type as the target output of the multi-layer perceptron, freeze the weights of each layer except the multi-layer perceptron in the multi-omics cancer risk assessment neural network model, train the multi-omics cancer risk assessment neural network model, and the multi-layer perceptron learns how to map the multimodal key feature representation into the probability distribution of the cancer type during the training process. In this embodiment, the one-hot encoding is completed using the scikit-learn library.

[0121] Embodiment 9: Based on the above embodiment, this embodiment uses a multi-omics cancer risk assessment neural network model to assess cancer risk, and the process specifically includes the following steps:

[0122] Step P1: collecting the target user's proteomics data, the target user's transcriptomics data, the target user's genomics data, and the target user's imaging data;

[0123] Step P2: Aggregate the target user's proteomics data, the target user's transcriptomics data, and the target user's genomics data into a protein-transcriptome-gene data instance, and extract the target user's medical image from the target user's radiomics data;

[0124] Step P3: Input the protein-transcription-gene data instance into the contraction encoder of the multi-omics cancer risk assessment neural network model, input the target user's medical image into the dynamic convolution layer of the multi-omics cancer risk assessment neural network model, and the multi-omics cancer risk assessment neural network model outputs the cancer type probability distribution of the target user;

[0125] Step P4: Extract the cancer type with the highest probability from the target user's cancer type probability distribution as the highest risk cancer type, and extract the probability corresponding to the highest risk cancer type.

[0126] Example 10. This example runs in the Windows operating system environment, relies on the Anacond3 environment, uses Keras as the algorithm framework of the multi-omics cancer risk assessment neural network model, the genomics acquisition module and the transcriptomics acquisition module are implemented using HiSeq2500, the proteomics acquisition module is implemented using TripleTOF 5600, and the imaging omics acquisition module and the central processing module are implemented using A320 157128-48TA.

[0127] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

[0128] The present invention and its implementation methods are described above. Such description is not restrictive. What is shown in the accompanying drawings is only one of the implementation methods of the present invention. The actual structure is not limited thereto. In short, if ordinary technicians in this field are inspired by the present invention and do not deviate from the purpose of the invention, they can creatively design structural methods and embodiments similar to the technical scheme, which should all fall within the scope of protection of the present invention.

Claims

1. A cancer risk assessment system based on multi-omics data, characterized by: include: Proteomics acquisition module, transcriptomics acquisition module, genomics acquisition module, imaging omics acquisition module and central processing module; The proteomics acquisition module, the transcriptomics acquisition module, the genomics acquisition module and the imaging omics acquisition module are electrically connected to the central processing module respectively; The central processing module uses transfer learning technology to construct a multi-omics cancer risk assessment neural network model and uses the multi-omics cancer risk assessment neural network model to assess cancer risk; The process of the central processing module using transfer learning technology to build a multi-omics cancer risk assessment neural network model includes the following steps: Step S1: construct a medical image denoising dataset and a medical image noise layer and use the medical image denoising dataset to obtain a medical image noise parameter dictionary; Step S2: collecting multi-omics data sets, constructing a medical image denoising convolutional autoencoder, and training the medical image denoising convolutional autoencoder; Step S3: construct a contraction autoencoder and train the contraction autoencoder using a multi-omics dataset; Step S4: constructing a multimodal second-order autoencoder using a contraction autoencoder and a medical image denoising convolutional autoencoder and training the multimodal second-order autoencoder using a multi-omics dataset; Step S5: Use a multimodal second-order autoencoder to build a multi-omics cancer risk assessment neural network model and use a multi-omics dataset to train the multi-omics cancer risk assessment neural network model.

2. A cancer risk assessment system based on multi-omics data according to claim 1, characterized in that: The proteomics acquisition module acquires proteomics data and uploads it to the central processing module; The transcriptomics acquisition module collects transcriptomics data and uploads it to the central processing module; The genomics acquisition module collects genomics data and uploads it to the central processing module; The radiomics acquisition module acquires radiomics data and uploads it to the central processing module. The radiomics data includes irradiation parameters, imaging parameters, medical images and imaging technology categories. The irradiation parameters include irradiation time and irradiation dose, and the imaging parameters include image sampling rate and pixel spacing.

3. The cancer risk assessment system based on multi-omics data according to claim 2, characterized in that: Step S1 specifically includes the following steps: Step S11: collecting medical images before denoising, corresponding medical images after denoising, corresponding irradiation parameters, corresponding imaging parameters and corresponding imaging technology categories to construct a medical image denoising data set; Step S12: constructing a medical image noise layer, wherein the medical image noise layer adds Gaussian noise, Poisson noise and salt and pepper noise to its input, and the hyperparameters of the medical image noise layer include the mean of Gaussian noise, the variance of Gaussian noise, the mean of Poisson noise, the variance of Poisson noise and the proportion of salt and pepper noise, and the hyperparameters of the medical image noise layer constitute a hyperparameter group; Step S13: Divide the medical image denoising data set according to the imaging technology category, designate the denoised medical image as the input of the medical image noise layer, designate the medical image before denoising as the target output of the medical image noise layer, construct a noise simulation symbolic regression model, use the symbolic regression method to maximize the average structural similarity index to search for the best noise simulation symbolic regression model corresponding to different imaging technology categories, extract the parameter adjustment formula set from the best noise simulation symbolic regression model, and construct a medical image noise parameter adjustment dictionary.

4. The cancer risk assessment system based on multi-omics data according to claim 3, characterized in that: Step S2 specifically includes the following steps: Step S21: collecting multi-omics data sets; Step S22: constructing a dynamic convolutional layer and a convolutional autoencoder, wherein the convolutional autoencoder includes a convolutional encoder, a convolutional bottleneck layer and a convolutional decoder; Step S23: connecting the medical image noise layer, the dynamic convolution layer and the convolutional autoencoder in sequence to construct a medical image denoising convolutional autoencoder; Step S24: extracting radiomics data from the multi-omics dataset; Step S25: Use the medical images in the imaging omics data to train the medical image denoising convolutional autoencoder. During the training process, while inputting the medical image into the medical image noise layer, query the parameter adjustment formula set in the medical image noise adjustment parameter dictionary according to the imaging technology category corresponding to the medical image. Extract the irradiation parameters and imaging parameters from the imaging omics data and input them into the parameter adjustment formula set to obtain a hyperparameter group, and adjust the medical image noise layer according to the hyperparameter group. The dynamic convolution layer selects different convolution kernels according to the noise simulated by the medical image noise layer for different irradiation parameters, imaging parameters and imaging technology categories. The convolution encoder learns during the training process how to extract the key features of the medical image as much as possible while removing the noise, and the convolution bottleneck layer outputs the denoised key feature map.

5. The cancer risk assessment system based on multi-omics data according to claim 4, characterized in that: The contraction autoencoder includes a contraction encoder, a contraction bottleneck layer and a contraction decoder.

6. The cancer risk assessment system based on multi-omics data according to claim 5, characterized in that: Step S4 specifically includes the following steps: Step S41: extracting a contraction encoder and a contraction bottleneck layer from the contraction autoencoder, and extracting a dynamic convolution layer, a convolution encoder, and a convolution bottleneck layer from the medical image denoising convolution autoencoder; Step S42: constructing a concatenated fusion layer and a sparse autoencoder, wherein the sparse autoencoder includes a sparse encoder, a sparse bottleneck layer and a sparse decoder; Step S43: connecting the shrinkage bottleneck layer and the convolution bottleneck layer to the splicing fusion layer, connecting the splicing fusion layer to the sparse autoencoder, and constructing a multimodal second-order autoencoder; Step S44: Freeze the weights of the contraction encoder, the contraction bottleneck layer, the dynamic convolution layer, the convolution encoder, and the convolution bottleneck layer, specify the protein-transcription-gene dataset as the input of the contraction encoder, specify the medical images in the imaging omics data as the input of the dynamic convolution layer, extract clinical data from the multi-omics dataset as the target output of the sparse autoencoder, and train the multimodal second-order autoencoder.

Images