A method for predicting colorectal cancer survival based on deep learning CT radiomics

Abstract

The present invention discloses a method for predicting survival of colorectal cancer based on deep learning CT imaging genomics. It belongs to the field of medical image processing technology. Specific steps: (1) acquiring data; (2) annotating colorectal tumor regions on CT imaging genomics data; (3) preprocessing the acquired data; (4) constructing a feature learning model based on a deep neural network; (5) using Lasso regression to reduce the dimension of deep high-throughput features of colorectal cancer CT imaging genomics, and establishing a risk scoring model for patients; (6) grouping according to risk scores; (7) curve and feature validity verification; (8) constructing a deep neural network multi-task logistic regression (DNN-MTLR) model for survival probability prediction; the present invention imports the system analysis after obtaining the CT images taken by the patient, and the results can provide a reference for doctors (especially young and inexperienced radiologists) to better understand the patient's condition and make the next decision.

Description

Technical Field

[0001] The present invention belongs to the field of medical image processing technology and can be used for intelligent medical disease diagnosis. It introduces a method for predicting colorectal cancer survival based on deep learning CT imaging genomics, and ultimately derives the five-year disease-free survival (DFS) probability of colorectal cancer patients. Background Art

[0002] Colorectal cancer is a common malignant tumor in the gastrointestinal tract, with high morbidity and mortality. According to the results of a global survey conducted by the International Agency for Research on Cancer in 2018, the incidence of colorectal cancer ranks third, second only to lung cancer and breast cancer. The mortality rate ranks second, second only to lung cancer. In China, the incidence and mortality rates in economically developed regions and the southeastern coastal areas are also on a significant increase.

[0003] Accurately predicting a patient's survival period has important clinical and social value. For doctors (especially young and inexperienced doctors), accurately predicting a patient's survival period can help doctors better understand the patient's condition, make a diagnosis, and make the best medical decision. For patients, accurately predicting a patient's survival period can provide patients with scientific survival expectations, so that they can better understand their own physical condition. This can guide patients to scientifically comply with treatment plans, avoid over-medicalization, reduce the family's financial burden, and also help improve the doctor-patient relationship.

[0004] With the development of imaging and artificial intelligence technologies, imaging technologies such as computed tomography (CT), positron emission tomography (PET), and magnetic resonance imaging (MR) are playing an increasingly important role in the diagnosis, treatment, and prognosis of tumors. The role of medical imaging has also gradually changed from traditional analysis methods such as disease diagnosis and screening to individualized precision diagnosis and treatment. The mainstream direction of future medical development is precision medicine, which needs to take into account individual differences in prevention and corresponding diagnosis and treatment strategies. The combination of artificial intelligence and medicine is also a necessary path for the future development of medicine. The realization of artificial intelligence is inseparable from machine learning, and deep learning is one of the technologies. Deep learning technology can be combined with CT imaging genomics features to predict the survival of colorectal cancer patients. Summary of the invention

[0005] In response to the above problems, the present invention provides a new method for predicting the survival of CRC (colorectal cancer) patients based on deep learning CT image groups; in particular, a method for predicting the survival of colorectal cancer based on deep learning CT imaging genomics is proposed.

[0006] The technical solution of the present invention is: a method for predicting colorectal cancer survival based on deep learning CT imaging genomics, and the specific steps include the following:

[0007] Step (1.1), obtaining data: the data includes clinical data and CT radiomics data;

[0008] Step (1.2), annotating the colorectal tumor region of the CT imaging data;

[0009] Step (1.3), preprocessing the acquired data;

[0010] Step (1.4), constructing a feature learning model based on a deep neural network to obtain deep high-throughput features of colorectal cancer CT imaging data;

[0011] Step (1.5), using Lasso regression to reduce the dimensionality of deep high-throughput features of colorectal cancer CT imaging data, and establish a risk scoring model for patients;

[0012] Step (1.6), according to the patient's radiomics risk score S, the cutoff value T is obtained using the median of the radiomics label score value, and the patient is divided into a high-risk survival group and a low-risk survival group;

[0013] Step (1.7), drawing KM curve for the obtained deep high-throughput features and using data analysis software to evaluate and verify the curve;

[0014] Step (1.8), construct a deep neural network multi-task logistic regression model to predict the survival probability.

[0015] Furthermore, in step (1.1), specifically:

[0016] (1.1.1) Clinical data: including the patient's age, gender, survival status: 1 or 0 and the time from the CT image to the occurrence of interest; 1 represents death and 0 represents survival;

[0017] (1.1.2) CT radiomics data: CT image data taken of patients.

[0018] Furthermore, in step (1.2), the specific operation method of annotating the colorectal tumor area of ​​the CT imaging genomics data is as follows: the CT imaging genomics data is imported into ITK-SNAP in batches in unit order, manually annotated, the region of interest where the tumor is located is selected, and the annotated CT imaging genomics data is saved as a nii file.

[0019] Furthermore, in step (1.3), the specific steps of preprocessing the acquired data are as follows:

[0020] The data was pre-selected and deleted, and the elimination criteria were as follows:

[0021] (1.3.1) Incomplete clinical information records, where the reasons for incompleteness include: loss to follow-up, withdrawal and termination;

[0022] (1.3.2) The end of the survival observation process is caused by other reasons rather than death;

[0023] (1.3.3) According to step (1.2), the nii file of the region of interest is obtained, and the features of the region of interest are extracted by combining with the original CT imaging data. Each unit obtains a three-dimensional feature matrix f(P, P, P) containing the region of interest, where P represents the size of the matrix.

[0024] Furthermore, in step (1.4), the feature learning model based on the deep neural network is specifically described as follows: the feature matrix containing the region of interest obtained by each unit is used as the input of the network, and the size of the feature matrix is ​​[M×P×P×P], where M represents the total number of units, and P represents the dimension of the feature matrix of each unit in the total units;

[0025] Put it into the feature selector for feature selection; the feature selector consists of N 0 convolutional layers, N 0 It consists of a pooling layer, a fully connected layer, and a logistic regression output layer;

[0026] The convolutional layer includes M 1 filters, and other convolutional layers include M i filters, where the filter size is n×n×n, where n represents the filter size;

[0027] Each of the above convolutional layers is followed by a maximum pooling operation with a pool size of m×m×m. Each convolution layer has a linear rectification function. The loss function uses the mean square error, and its formula is as follows:

[0028]

[0029] Among them, y m represents the true value, Represents the predicted value.

[0030] Furthermore, in step (1.5), the specific operation method for effectively reducing the dimension of the colorectal cancer CT imaging genomics data is as follows: first, M×K node information of the fully connected layer of the feature learning model of the deep neural network is selected as the first effective feature dimensionality reduction, where M represents the total number of units and K is the number of node information; the data is standardized;

[0031] Next, the least absolute shrinkage selection operator Lasso regression is used to further effectively reduce the dimension of the features and obtain the risk coefficient score S for each person; the Lasso regression loss function is as follows:

[0032]

[0033] Among them, xi represents the feature label of each unit, yi represents the time label of each unit, and λ represents the regularization coefficient. Represents the weight coefficient.

[0034] Furthermore, in step (1.7), the specific steps for curve evaluation and verification of the selected features are as follows:

[0035] (1.7.1) Use the cutoff value T obtained in step (1.6) to draw the corresponding KM curve to visualize the results and obtain two survival probability curves;

[0036] (1.7.2) After obtaining different survival probability curves using the KM method, a chi-square test was performed using data analysis software, and finally the P value was obtained;

[0037] (1.7.3) Use the P value to determine whether there is a significant difference between the two curves.

[0038] Furthermore, in step (1.8), the specific steps of constructing a deep neural network multi-task logistic regression model for survival probability prediction are as follows:

[0039] (1.8.1) Import the final effective features obtained in step (1.5) together with the time label and survival status label into the deep neural network multi-task logistic regression model;

[0040] Among them, each layer of the deep neural network multi-task logistic regression model uses the following activation function:

[0041] Layer#1: M1 neurons, using activation function: h (1) (x) = LeakyReLu(x)

[0042] Layer#2: M2 neurons, using activation function: h (2) (x) = ReLu(x)

[0043] Layer#3: M3 neurons, using activation function: h (3) (x) = ReLu(x)

[0044] Among them, LeakyReLu represents the leaky rectified linear unit function, and ReLu represents the rectified linear unit function;

[0045] Divide the time axis into J time intervals, so that a j =[τ j-1 ,τ j ], with τ 0 = 0, and τ J =∞; as shown in the following formula:

[0046]

[0047] In each interval a j Build a logistic regression model with parameters and the response variable That is, the event occurs in interval a j is 1, otherwise it is 0;

[0048] When a unit is in interval a s When an event s∈[1,J] is experienced, the state of the remaining interval remains unchanged; therefore, the response vector is as follows:

[0049]

[0050] Among them, a j Indicates a unit time interval: one month; represents the response variable matrix; y j is the response variable: 1 represents the occurrence of the event, 0 represents the non-occurrence;

[0051] Probability density function:

[0052] Among them, exp() represents the exponential function with natural number e as the base, a s represents the sth time interval,

[0053] is the input feature vector, is a nonlinear transformation;

[0054] Survival function Representation sample In τ s-1 The probability density function after the time interval is calculated as:

[0055] Where ψ: So The eigenvector is a nonlinear transformation of the input, and its output is an R J A vector whose values ​​are mapped to the J subdivisions of the time axis as described by the following formula:

[0056]

[0057] (1.8.2) Set the ratio of the training set to the test set to 8:2 and visualize the results.

[0058] (1.8.3) Use the concordance index to evaluate the discriminative power of the deep neural network multi-task logistic regression model: The concordance index represents the overall evaluation of the discriminative power of the deep neural network multi-task logistic regression model. The value range of the concordance index is between 0 and 1. A value of 1 represents the best prediction model, a value of 0.5 represents a random prediction model, and 0 represents an inapplicable model. The calculation formula of the concordance index is as follows:

[0059]

[0060] where C-index represents the concordance index; 1 Tj<Ti Satisfy Tj < Ti = 1, otherwise 0;

[0061] (1.8.4) Use IBS to evaluate the accuracy of the deep neural network multi-task logistic regression model: Its value range is between 0 and 1. Among them, 0 is the best possible value; IBS < 0.25 represents a useful model. The calculation formula of IBS is:

[0062] where IBS represents the integrated Brier score, which is used to evaluate the accuracy of the model in predicting the survival function; N is the number of data samples, represents the actual probability of the occurrence of event t in sample i.

[0063] The beneficial effects of the present invention are as follows: The present invention combines deep learning technology with CT radiomics labels to predict the survival period of colorectal cancer patients; this technology relies on CT imaging, and CT images are relatively easy to obtain clinically; in medicine, after obtaining the CT images of patients and importing them into the system for analysis, the results can provide reference for physicians (especially young and inexperienced radiologists) to better understand the patient's condition and make the next decision; in addition, for patients, it also helps to better understand their own condition.

[0064] CT images contain rich features, but due to the large size of CT images and excessive number of slices, the data volume is large and there are many redundant features; the present invention realizes data dimensionality reduction through a DL feature selector and the least absolute shrinkage operator Lasso regression, thereby obtaining effective features with low dimensions that are beneficial for prediction.

[0065] In addition, the present invention constructs a deep neural network multi-task logistic regression (DNN-MTLR) model, which provides similar results to the CoxPH model, but does not rely on the assumptions required by the latter. The DNN-MTLR model can be used to estimate the probability of an event of interest occurring within each centerline. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 It is a schematic diagram of the process of the present invention;

[0067] Figure 2 This is a schematic diagram of manual annotation using ITK-SNAP in the present invention;

[0068] Figure 3 It is a network model diagram of the DL feature selector in the present invention;

[0069] Figure 4 It is a high-low risk grouping-KM curve diagram of the present invention;

[0070] Figure 5 It is a DNN-MTLR network model diagram in the present invention;

[0071] Figure 6 This is the result graph predicted by the DNN-MTLR model in the present invention;

[0072] Figure 7 It is a prediction diagram of a specific embodiment of the present invention. DETAILED DESCRIPTION

[0073] In order to more clearly illustrate the technical solution of the present invention, the technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings:

[0074] like Figure 1 The method for predicting colorectal cancer survival based on deep learning CT imaging genomics ultimately derives the five-year disease-free survival (DFS) probability of colorectal cancer patients. The specific steps include the following:

[0075] Step (1.1), obtaining data: the data includes clinical data and CT radiomics data;

[0076] Step (1.2), annotating the colorectal tumor region of the CT imaging data;

[0077] Step (1.3), preprocessing the acquired data;

[0078] Step (1.4), constructing a feature learning model based on a deep neural network to obtain deep high-throughput features of colorectal cancer CT imaging data;

[0079] Step (1.5), using Lasso regression to reduce the dimensionality of deep high-throughput features of colorectal cancer CT imaging data, and establish a risk scoring model for patients;

[0080] Step (1.6), according to the patient's radiomics risk score S, the cutoff value T is obtained using the median of the radiomics label score value, and the patients are divided into a high-risk survival group (S>T) and a low-risk survival group (S <T);

[0081] Step (1.7), drawing KM curve for the obtained deep high-throughput features and using data analysis software to evaluate and verify the curve;

[0082] Step (1.8), construct a deep neural network multi-task logistic regression (DNN-MTLR) model to predict the survival probability.

[0083] Furthermore, in step (1.1), specifically:

[0084] (1.1.1) Clinical data: including the patient's age, gender, survival status: 1 or 0 and the time from the CT image to the occurrence of interest; 1 represents death and 0 represents survival;

[0085] (1.1.2) CT radiomics data: CT image data taken of patients.

[0086] Furthermore, in step (1.2), the specific operation method of labeling the colorectal tumor region of the CT radiomics data is as follows: the CT radiomics data is imported into ITK-SNAP in batches according to the unit sequence, and the data is manually labeled, the region of interest where the tumor is located is selected, and the labeled CT radiomics data is saved as a nii file; the labeling result is as follows: Figure 2 shown.

[0087] Furthermore, in the step (1.3), the specific steps of preprocessing the data are as follows:

[0088] The data was pre-selected and deleted, and the elimination criteria were as follows:

[0089] (1.3.1) Incomplete clinical information records, where the reasons for incompleteness include: loss of follow-up (loss of contact), withdrawal (death from the study due to non-research factors or non-treatment factors), and termination (observation was terminated after the time specified in the design, but the subject is still alive);

[0090] (1.3.2) The end of the survival observation process is caused by other reasons rather than death;

[0091] (1.3.3) According to step (1.2), the nii file of the region of interest is obtained, and the features of the region of interest are extracted by combining with the original CT image data. Each unit obtains a three-dimensional feature matrix f(32,32,32) containing the region of interest.

[0092] Furthermore, in step (1.4), if Figure 3 As shown, the feature learning model based on deep neural network is specifically described as follows: the feature matrix containing the region of interest obtained by each unit is used as the input of the network, and the size of the feature matrix is ​​[M×P×P×P], where M represents the total number of units; P represents the dimension of the feature matrix of each unit in the total units;

[0093] Put it into the feature selector for feature selection; the feature selector consists of N 0 convolutional layers, N 0 The convolutional layer consists of M pooling layers, a fully connected layer, and a logistic regression output layer. 1 filters, and other convolutional layers include M i filters, where the filter size is n×n×n, n is the filter size;

[0094] Each of the above convolutional layers is followed by a maximum pooling operation with a pool size of m×m×m. Each convolution layer has a linear rectification function (RELU); the loss function uses the mean square error (MSE), and its formula is as follows:

[0095]

[0096] Among them, y m represents the true value, Represents the predicted value.

[0097] Furthermore, in step (1.5), the specific operation method for effectively reducing the dimension of the colorectal cancer CT imaging genomics data is as follows: first, 600×6400 node information of the fully connected layer of the feature learning model of the deep neural network is selected as the first effective feature dimensionality reduction; the data is standardized;

[0098] Next, the least absolute shrinkage selection operator Lasso regression is used to further effectively reduce the dimension of the features and obtain the risk coefficient score S for each person; the Lasso regression loss function is as follows:

[0099]

[0100] Among them, xi represents the feature label of each unit, yi represents the time label of each unit, and λ represents the regularization coefficient. Represents the weight coefficient.

[0101] Further, in step (1.6), the specific operation method for dividing patients into a high-risk survival group (S > T) and a low-risk survival group (S < T) is as follows: Locate the patient's risk coefficient score S file, and use the median of the radiomics label score value as the cut-off value T: -2.227. Taking T as the cut-off value, for S > T, it is taken as the high-risk group with a shorter survival period, and for S < T, it is taken as the low-risk group with a longer survival period.

[0102] Further, in step (1.7), the specific operation steps for curve evaluation and verification of the selected features are as follows:

[0103] (1.7.1) Plot the corresponding KM curve using the cut-off value T obtained in step (1.6) to visualize the results and obtain two survival probability curves;

[0104] (1.7.2) After obtaining different survival probability curves using the KM method, it is insufficient to directly determine whether there are significant differences between multiple curves through direct observation. Use IBM SPSS Statistics 26 to perform a log-rank test and finally obtain the P value;

[0105] (1.7.3) Judge whether there are significant differences between the two curves according to the P value; generally, it is considered that P < 0.05 indicates a statistical difference; the obtained result P < 0.01 shows a statistical difference.

[0106] Further, in step (1.8), the specific operation steps for constructing a deep neural network multi-task logistic regression (DNN-MTLR) model for survival probability prediction are as follows:

[0107] (1.8.1) Import the final effective features obtained in step (5), along with the time label and survival status label, into the deep neural network multi-task logistic regression (DNN-MTLR) model; among them, the DNN-MTLR model is as Figure 5 shown;

[0108] Each layer uses the following activation functions:

[0109] Layer#1: 326 neurons, using the activation function: h (1) (x) = LeakyReLu(x)

[0110] Layer#2: 652 neurons, using the activation function: h (2) (x) = ReLu(x)

[0111] Layer#3: 1304 neurons, using the activation function: h (3) (x) = ReLu(x)

[0112] Among them, LeakyReLu is the leaky rectified linear unit function, and ReLu is the rectified linear unit function;

[0113] Divide the time axis into J time intervals, so that a j =[τ j-1 ,τ j ], with τ 0 = 0, and τ J =∞; as shown in the following figure:

[0114]

[0115] In each interval a j Build a logistic regression model with parameters and the response variable That is, the event occurs in interval a j is 1, otherwise it is 0;

[0116] However, since the effect of repeated events is not analyzed, it is necessary to ensure that when a cell is in interval a s When an event s∈[1,J] is experienced, the state of the remaining interval remains unchanged; therefore, the response vector is as follows:

[0117]

[0118] Among them, a j Indicates a unit time interval: one month; represents the response variable matrix; y j is the response variable: 1 represents the occurrence of the event, 0 represents the non-occurrence;

[0119] Probability density function:

[0120] Among them, exp() represents the exponential function with natural number e as the base, a s represents the sth time interval,

[0121] is the input feature vector, is a nonlinear transformation;

[0122] Survival function Representation sample In τ s-1 The probability density function after the time interval is calculated as:

[0123] Where ψ: So The feature vector is a non-linear transformation of the input, and its output is an R J vector, the value of which is mapped to the J subdivision of the time axis, as described by the following formula:

[0124]

[0125] (1.8.2), where the ratio of the training set to the test set is set to 8:2, and the results are visualized;

[0126] (1.8.3), Use the concordance index (C-index) DNN-MTLR model to evaluate the discrimination ability of the deep neural network multi-task logistic regression model: where the C-index represents the overall evaluation of the discrimination ability of the deep neural network multi-task logistic regression model, and the C-index is obtained: (0.82). 1 is the best prediction model, the value 0.5 is the random prediction model, and 0 is the inapplicable model. The formula for calculating its C-index is as follows:

[0127]

[0128] Among them, the C-index represents the concordance index, and ηi represents the risk score of a unit i; 1 Tj<Ti satisfies Tj < Ti = 1, otherwise it is 0;

[0129] (1.8.4), Use the integrated Brier score (IBS) to evaluate the accuracy of the deep neural network multi-task logistic regression (DNN-MTLR) model: where IBS represents the accuracy of evaluating the model's predicted survival function, and the IBS value: (0.06), where 0 is the best possible value; IBS < 0.25 represents a useful model; among them, the IBS calculation formula is:

[0130] Among them, IBS represents the integrated Brier score, which is used to evaluate the accuracy of the model's predicted survival function. N is the number of data samples, represents the actual probability of the occurrence of event t for sample i. Specific embodiments:

[0132] (1). Obtain data: Obtain the CT imaging omics data of patient A.

[0133] (2). Perform colorectal tumor region annotation on the CT imaging omics data.

[0134] (3). Preprocess the obtained data to obtain time and status labels.

[0135] (4). Construct a feature learning model based on a deep neural network to obtain the CT imaging omics deep high-throughput features of patient A.

[0136] (5) The lasso regression operator is used to reduce the dimensionality of the deep high-throughput features of CT imaging data, and a risk scoring model for patient A is established.

[0137] (6) Based on the radiomics risk score of patient A, patient A was divided into a high-risk group.

[0138] (7) Evaluate and verify the obtained deep high-throughput features.

[0139] (8) The features obtained by dimensionality reduction are put into the deep neural network multi-task logistic regression model to predict the survival probability and finally obtain the prediction results.

[0140] The results are as follows Figure 7 The probability results by year are shown in the following table:

[0141] month 12 24 36 48 60 Survival probability 97.649882％ 93.804818％ 85.995414％ 66.933918％ 48.437748％

[0142] The results showed that this type of method can be used to predict the survival of patients with colorectal cancer.

Claims

1. A method for predicting colorectal cancer survival based on deep learning CT radiomics, characterized in that: The specific steps include the following: Step (1.1), obtaining data: the data includes clinical data and CT radiomics data; Step (1.2), annotating the colorectal tumor region of the CT imaging data; Step (1.3), preprocessing the acquired data; Step (1.4), constructing a feature learning model based on a deep neural network to obtain deep high-throughput features of colorectal cancer CT imaging data; Step (1.5), using Lasso regression to reduce the dimensionality of deep high-throughput features of colorectal cancer CT imaging data, and establish a risk scoring model for patients; Step (1.6), according to the patient's radiomics risk score S, the cutoff value T is obtained using the median of the radiomics label score value, and the patient is divided into a high-risk survival group and a low-risk survival group; Step (1.7), using a drawing KM curve for the obtained deep high-throughput features and using data analysis software to evaluate and verify the curve; Step (1.8), construct a deep neural network multi-task logistic regression model to predict the survival probability.

2. The method for predicting colorectal cancer survival based on deep learning CT radiomics according to claim 1, characterized in that: In step (1.1), specifically: (1.1.1) Clinical data: including the patient's age, gender, survival status: 1 or 0 and the time from the CT image to the time of interest; 1 represents death and 0 represents survival; (1.1.2) CT radiomics data: CT image data taken of patients.

3. The method for predicting colorectal cancer survival based on deep learning CT radiomics according to claim 1, characterized in that: In step (1.2), the specific operation method of annotating the colorectal tumor area of ​​the CT imaging genomics data is as follows: the CT imaging genomics data is imported into ITK-SNAP in batches in unit order, manually annotated, the region of interest where the tumor is located is selected, and the annotated CT imaging genomics data is saved as a nii file.

4. The method for predicting colorectal cancer survival based on deep learning CT radiomics according to claim 1, characterized in that: In the step (1.3), the specific steps of preprocessing the acquired data are as follows: The data was pre-selected and deleted, and the elimination criteria were as follows: (1.3.1) Incomplete clinical information records, where the reasons for incompleteness include: loss to follow-up, withdrawal and termination; (1.3.2) The end of the survival observation process is caused by other reasons rather than death; (1.3.3) According to step (1.2), the nii file of the region of interest is obtained, and the features of the region of interest are extracted by combining with the original CT imaging data. Each unit obtains a three-dimensional feature matrix f(P, P, P) containing the region of interest, where P represents the size of the matrix.

5. The method for predicting colorectal cancer survival based on deep learning CT radiomics according to claim 1, characterized in that: In step (1.4), the feature learning model based on deep neural network is specifically described as follows: the feature matrix containing the region of interest obtained by each unit is used as the input of the network, The size of its feature matrix is ​​[M×P×P×P], where M represents the total number of units and P represents the dimension of the feature matrix of each unit in the total unit; Put it into the feature selector for feature selection; the feature selector consists of N0 convolutional layers, N0 pooling layers, a fully connected layer and a logistic regression output layer; The convolutional layer includes M1 filters, and the other convolutional layers include M i filters, where the filter size is n×n×n, where n represents the filter size; After each of the above convolutional layers, there is a max pooling operation with a pool size of m×m×m. Each convolutional layer has a rectified linear unit function; the loss function uses the mean squared error, and its formula is as follows: Among them, y m represents the true value, Represents the predicted value.

6. The method for predicting colorectal cancer survival based on deep learning CT radiomics according to claim 1, characterized in that: In step (1.5), the specific operation method for effective dimensionality reduction of the colorectal cancer CT imaging omics data is as follows: First, select the M×K node information of the fully connected layer of the feature learning model of the deep neural network as the first effective feature dimensionality reduction, where M represents the total number of units and K is the number of node information; standardize the data; Next, use the least absolute shrinkage and selection operator Lasso regression to further effectively reduce the dimensions of the features and obtain the risk coefficient score S for each person; the Lasso regression loss function is as follows: Among them, xi represents the feature label of each unit, yi represents the time label of each unit, and λ represents the regularization coefficient. Represents the weight coefficient.

7. The method for predicting colorectal cancer survival based on deep learning CT radiomics according to claim 1, characterized in that: In step (1.7), the specific operation steps for curve evaluation and verification of the selected features are as follows: (1.7.1) Plot the corresponding KM curve with the cut-off value T obtained in step (1.6) to visualize the results and obtain two survival probability curves; (1.7.2) After obtaining different survival probability curves using the KM method, perform a chi-square test through data analysis software and finally obtain the P value; (1.7.3) Judge whether there is a significant difference between the two curves according to the P value.

8. The method for predicting colorectal cancer survival based on deep learning CT radiomics according to claim 1, characterized in that: In step (1.8), the specific operation steps for constructing a deep neural network multi-task logistic regression model for survival probability prediction are as follows: (1.8.1) Import the final effective features obtained in step (1.5), together with the time label and survival status label, into the deep neural network multi-task logistic regression model; Among them, each layer of the deep neural network multi-task logistic regression model uses the following activation functions: Layer#1: M1 neurons, using activation function: h (1) (x) = LeakyReLu(x) Layer#2: M2 neurons, using activation function: h (2) (x) = ReLu(x) Layer#3: M3 neurons, using activation function: h (3) (x) = ReLu(x) Among them, LeakyReLu represents the leaky rectified linear unit function, and ReLu represents the rectified linear unit function; Divide the time axis into J time intervals, so that a j =[τ j-1 ,τ j ], with τ0 = 0, and τ J =∞; as shown in the following formula: In each interval a j Build a logistic regression model with parameters and the response variable That is, the event occurs in interval a j is 1, otherwise it is 0; When a unit is in interval a s When an event s∈[1,J] is experienced, the state of the remaining interval remains unchanged; therefore, the response vector is as follows: Among them, a j Indicates a unit time interval: one month; represents the response variable matrix; y j is the response variable: 1 represents the occurrence of the event, 0 represents the non-occurrence; Probability density function: Among them, exp() represents the exponential function with natural number e as the base, a s represents the sth time interval, is the input feature vector, is a nonlinear transformation; Survival function Representation sample In τ s-1 The probability density function after the time interval is calculated as: in, So The eigenvector is a nonlinear transformation of the input, and its output is an R J A vector whose values ​​are mapped to the J subdivisions of the time axis as described by the following formula: (1.8.2) Among them, the ratio of the training set to the test set is set to 8:2, and the results are visualized; (1.8.3) Use the concordance index to evaluate the discriminative ability of the deep neural network multi-task logistic regression model: Among them, the concordance index represents the overall evaluation of the discriminative ability of the deep neural network multi-task logistic regression model. The value range of the concordance index is between 0 and 1. The value 1 is the best prediction model, the value 0.5 is the random prediction model, and 0 is an inapplicable model; the calculation formula of the concordance index is as follows: Among them, C-index represents the concordance index, ηi represents the risk score of a unit i; 1Tj<Ti satisfies Tj<Ti = 1, otherwise it is 0; (1.8.4) Use IBS to evaluate the accuracy of the deep neural network multi-task logistic regression model: Its value range is between 0 and 1, where 0 is the best possible value; IBS < 0.25 indicates a useful model; among them, the IBS calculation formula is: Among them, IBS represents the comprehensive Brinell score, which is used to evaluate the accuracy of the model in predicting the survival function; N is the number of data samples, Represents the actual probability of event t occurring in sample i.

Images