A method for non-invasive detection of gene mutation based on multi-view CT images

Abstract

The application discloses a kind of based on multi-view CT image's gene mutation noninvasive detection method, comprising: 1 the acquisition and pre-processing of CT data;2 construct by joint representation learning module, double confrontation representation learning module and detection module composition gene mutation state detection model;3 joint representation learning module carries out shared information extraction and mutation irrelevant information separation to CT image sample;4 using double confrontation representation learning network to supervise shared information and mutation irrelevant information;5 construct loss function and train detection model, obtain the trained detection model.The application can learn the shared information between views while reducing the mutation irrelevant information mixed in the multi-view fusion process, so as to effectively improve the noninvasive detection performance of gene mutation state.

Description

Technical Field

[0001] This invention belongs to the field of non-invasive detection technology of gene mutations in lung CT images, specifically a non-invasive detection method for gene mutations based on multi-view CT images. Background Technology

[0002] The goal of non-invasive gene mutation detection using lung CT images is to achieve non-invasive detection of gene mutation status through CT image analysis combined with artificial intelligence. Clearly identifying a patient's gene mutation status helps doctors develop targeted treatment plans, which has broad clinical application value. However, current clinical methods involve puncturing tissue samples for gene testing to determine gene mutation status, which has many drawbacks such as invasiveness, high risk, and difficulty in repeated sampling. Therefore, developing non-invasive gene mutation status detection methods is of great research significance. In recent years, researchers have proposed and successfully applied non-invasive gene mutation detection based on multi-view CT images.

[0003] However, existing non-invasive gene mutation detection methods based on multi-view CT images still have some problems that urgently need to be solved. First, studies have shown that different views of CT images can jointly reflect pathological information of abnormal lung areas in terms of texture, shape, and other aspects. This shared information between views is of great value for detecting gene mutation status. However, most existing multi-view methods adopt a strategy of learning the features of different views separately, neglecting the effective extraction of the aforementioned cross-view shared information, thus leading to a decline in the performance of the detection method. Second, CT images often contain information unrelated to the gene mutation status, such as ribs and spine appearing in the view. Existing multi-view methods directly fuse information from different views, inevitably mixing in information unrelated to the mutation status in the process, resulting in poor quality of the generated multi-view representation, which in turn affects the accuracy of detection. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention proposes a non-invasive detection method for gene mutations based on multi-view CT images. The aim is to learn shared information between different views of the combined CT images while reducing mutation-irrelevant information mixed in during the multi-view fusion process, thereby effectively improving the accuracy and efficiency of non-invasive gene mutation detection.

[0005] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0006] The present invention provides a non-invasive method for detecting gene mutations based on multi-view CT images, characterized by the following steps:

[0007] Step 1: Acquisition and preprocessing of lung CT data;

[0008] Lung CT image data of N cases with gene mutation category labels are obtained. All lung CT image data are then resampled using a trilinear interpolation algorithm to standardize voxel size, resulting in a lung CT slice set for each case and its corresponding category label, denoted as . Among them, X i This represents a lung CT scan of the i-th case, y i X represents the lung CT slice of the i-th case. i The corresponding gene mutation category label; i = 1, 2, ..., N; N represents the number of cases;

[0009] With X i Taking the center of the abnormal location as the origin, image blocks of dimension H×W pixels are cropped along the axial view direction av, the coronal view direction sv, and the sagittal view direction cv, respectively, to obtain the lung CT image I of the i-th case in the axial view direction av. i,av Lung CT images in the coronal view (SV) i,sv Lung CT images in the sagittal view CV direction i,cv ;

[0010] The lung CT image of the i-th case I i,av Lung CT images I i,sv and lung CT images I i,cv The multi-view data of the i-th case {I i,n |n=av,cv,sv};where I i,n This represents the lung CT image of the i-th case in the viewing direction n;

[0011] Step 2: Construct a non-invasive gene mutation detection model;

[0012] Step 2.1: Construct a joint representation learning module, including: a parameter-shared encoder E c And three view-specific encoders, namely axial view-specific encoders Coronal view specific encoder and sagittal view specific encoder

[0013] Step 2.2: Construct a dual-adversarial representation learning module, including: axial-coronal view discriminator D1, axial-sagittal view discriminator D2, and axial view mutation discriminator D... av Coronal View Mutation Discriminator D cv and sagittal view mutation discriminator D sv ;

[0014] Step 2.3: Construct a gene state detection module, including a weight generation module and a classification module;

[0015] Step 3: The joint representation learning module extracts shared information and separates mutation-irrelevant information from the lung CT images;

[0016] Step 3.1: Transfer the multi-view data {I} of the i-th case. i,n |n=av,sv,sv} Input the parameters shared by encoder E c The data is processed to obtain an invariant representation of the multi-view data. in, This shows the invariant representation of the i-th case in the view direction n;

[0017] Step 3.2: Image Ii of the lungs in the axial view direction av of the i-th case. ,av Input to the axial view specific encoder The process is performed to obtain the specific characterization of the i-th case in the axial view direction av.

[0018] The lung CT image I of the i-th case in the sagittal view direction cv i,cv Input to the coronal view specific encoder The process is performed to obtain the specific characterization of the i-th case in the sagittal view direction cv.

[0019] The lung CT image I of the i-th case in the coronal view direction sv i,sv Input to the sagittal view-specific encoder The data is processed to obtain the specific characterization of the i-th case in the coronal view direction sv.

[0020] Step 4: The dual-adversarial representation learning module supervises the shared information and mutation-independent information;

[0021] Step 4.1, with To represent the target view, with and Representation of the source view;

[0022] Will and The input is processed by the axial-coronal view discriminator D1, thereby... Determined to be true. Determined to be false;

[0023] Will and The input is processed by the axial-sagittal view discriminator D2, thereby... Determined to be true. Determined to be false;

[0024] Step 4.2, Input to the corresponding mutation discriminator D ma D mc and D ms The data is processed to obtain gene mutation category detection labels {y' based on view-specific characterization. i,n |n=av,cv,sv};where y' i,n The gene state category detection label represents the specific characterization of the i-th case based on the view direction n;

[0025] Step 5: The gene state detection module fuses and detects the characteristics;

[0026] Step 5.1: The weight generation module uses equation (7) to generate the weight w of the i-th case in the view direction n. i,n :

[0027]

[0028] In equation (7), MLP represents a multilayer perceptron;

[0029] Step 5.2: The classification module uses equation (8) to obtain the multi-view data fusion representation F of the i-th case. i :

[0030]

[0031] Step 6: Construct the total loss function and train the detection model to obtain the trained detection model;

[0032] Step 6.1: Construct the total loss function L total ;

[0033] View discrimination loss function {L} for multi-view data of N cases i,v The mean value of |i=1,2,…N} is calculated to obtain the average view discrimination loss function L. v Among them, L i,v For the i-th case, identify the view discrimination loss;

[0034] Mutation identification loss function {L} for multi-view data of N cases i,m The mean value of |i=1,2,…N} is calculated to obtain the average view discrimination loss function L. m Among them, L i,m The mutation identification loss for the i-th case;

[0035] Separation loss function {L} for multi-view data of N cases i,sep The mean value of |i=1,2,…N} is calculated to obtain the average separation loss function l.sep Among them, L i,sep Let be the separation loss function for the i-th case;

[0036] The classification loss function {L} for multi-view data of N cases i,cla The mean of |i=1,2,…N} is calculated to obtain the average classification loss function L. cla Among them, L i,cla Let be the classification loss function for the i-th case;

[0037] Construct the overall loss function L using equation (11) total ;

[0038] L total =L cla +γL sep +(αL v +βL m (11)

[0039] In equation (7), α represents L v The weighting coefficient, β represents L m The weighting coefficient, γ represents L sep Weighting coefficients;

[0040] Step 6.2: Continuously train the non-invasive gene mutation detection network using the Adam optimizer and calculate the total loss function L. total The network parameters are updated until the maximum number of training rounds is reached, thereby obtaining a trained non-invasive gene detection model based on lung CT images, which can be used to achieve non-invasive detection of gene mutation status.

[0041] The non-invasive gene mutation detection method based on multi-view CT images described in this invention is also characterized by:

[0042] The parameter-sharing encoder E c and three view-specific encoders and Each layer consists of C1 shallow convolutional layers, I1 residual blocks, and P1 max-pooling layers; each residual block consists of C2 convolutional layers and their skip connections; the activation function of all convolutional layers is set to the ReLU function;

[0043] View discriminators D1 and D2, mutation discriminator D av D cv and D sv All of them are network structures with different parameters based on generative adversarial networks;

[0044] Both the weight generation module and the classification module include C3 convolutional layers, P2 max pooling layers, and F fully connected layers; the activation function of all convolutional layers and the first F-1 fully connected layers uses the ReLU function; the activation function of the Fth fully connected layer uses the Softmax function.

[0045] Using Equation (2), construct the view discrimination loss L for the i-th case. i,v :

[0046] L i,v =L i,f +L i,t (2)

[0047] In equation (2), L i,f Let L represent the spurious adversarial loss for the i-th case, and obtain it from equation (3). i,t Let represent the true adversarial loss for the i-th case, and obtain it from equation (4);

[0048]

[0049]

[0050] The separation loss function L for the i-th case is constructed using equation (1). i,sep :

[0051]

[0052] In equation (1), This represents the square of the Frobenius norm.

[0053] Using equation (5), construct the mutation identification loss L for the i-th case. i,m ;

[0054]

[0055] In equation (5), L i,av L represents the mutation discrimination loss of the i-th case in the axial view direction av. i,cv L represents the mutational discrimination loss of the i-th case in the sagittal view direction. i,sv denoted as mutation identification loss of the i-th case in the coronal view direction sv; CE() represents the cross-entropy loss.

[0056] The classification loss function L for the i-th case is constructed using equation (9). i,cla :

[0057] L i,cla =-y i log(MLP(F i ))-(1-y i)log(1-MLP(F i (9)

[0058] The present invention provides an electronic device, including a memory and a processor, wherein the memory is used to store a program that supports the processor in executing the non-invasive gene mutation detection method, and the processor is configured to execute the program stored in the memory.

[0059] The present invention discloses a computer-readable storage medium on which a computer program is stored, wherein the computer program, when executed by a processor, performs the steps of the non-invasive gene mutation detection method.

[0060] Compared with existing technologies, the beneficial effects of this invention are reflected in:

[0061] 1. The method of the present invention uses the central slices of abnormal locations in the axial, coronal and sagittal views of CT images as multi-view inputs and maps them to a shared subspace to obtain view-invariant representations, thereby fully extracting the shared information of multiple views; at the same time, by learning the specific representation of each view, the method also captures mutation-independent information in each view, which can enable the model to learn more discriminative representations, thereby improving the accuracy of non-invasive gene mutation detection.

[0062] 2. To enhance the learning of view-invariant and specific representations, this invention designs a dual-adversarial learning strategy that combines a view discriminator and a mutation discriminator. Specifically, by establishing an adversarial game process between the view discriminator and the shared encoder, the distribution alignment among view-invariant representations is achieved, thereby more effectively mining shared information among views. The mutation discriminator and the specific encoder are also trained adversarially to ensure that view-specific representations are independent of mutation states, thus significantly reducing mutation-irrelevant information mixed in during multi-view fusion and effectively solving the problem of poor quality of multi-view representations. This effectively improves the model's performance in non-invasive detection of gene mutations. Attached Figure Description

[0063] Figure 1 This is a schematic diagram of the process of the present invention;

[0064] Figure 2 This is a network structure diagram of the present invention. Detailed Implementation

[0065] In this embodiment, a non-invasive method for detecting gene mutations based on multi-view CT images is described, such as... Figure 1 As shown, the procedure is as follows:

[0066] Step 1: Acquisition and preprocessing of lung CT data;

[0067] Lung CT image data of N cases with gene mutation category labels are obtained. All lung CT image data are then resampled using a trilinear interpolation algorithm to standardize voxel size, resulting in a lung CT slice set for each case and its corresponding category label, denoted as . Among them, X i This represents a lung CT scan of the i-th case, y i X represents the lung CT slice of the i-th case. i The corresponding gene mutation category label; i = 1, 2, ..., N; N represents the number of cases; In addition, given the large range of grayscale variation in the CT images of each case, this invention further normalizes the grayscale of the CT images of all cases; In this embodiment, CT data of 420 cases of lung EGFR gene mutations were collected as a sample set, the CT images were resampled, and their size was adjusted to 1×1×1mm, and then grayscale normalization was performed. Specifically, since the grayscale value of air in CT images is less than -1000HU (Hounsfield Unit), while the grayscale value of the bone region is greater than +400HU, this invention first adjusts the grayscale of all CT images to the range of -1000 to 400HU, and then normalizes the grayscale value of the above CT images to the range of [0,1].

[0068] With X i Taking the center of the abnormal location as the origin, image blocks of dimension H×W pixels are cropped along the axial view direction av, the coronal view direction sv, and the sagittal view direction cv, respectively, to obtain the lung CT image I of the i-th case in the axial view direction av. i,av Lung CT images in the coronal view (SV) ,sv Lung CT images in the sagittal view CV direction i,cv In this embodiment, both H and W are set to 128, and then random folding (including vertical folding and horizontal folding) and random rotation {90°, 180°, 270°, 360°} are used to augment the training set images.

[0069] The lung CT image of the i-th case I i,av Lung CT images I i,sv and lung CT images I i,cv The multi-view data of the i-th case {I i,n |n=av,cv,sv};where I i,n This represents the lung CT image of the i-th case in the viewing direction n.

[0070] Step 2, as follows Figure 2 As shown, a non-invasive detection model for gene mutations was constructed.

[0071] Step 2.1: Construct a joint representation learning module, including: a parameter-shared encoder E c And three view-specific encoders, namely axial view-specific encoders Coronal view specific encoder and sagittal view specific encoder

[0072] Parameter sharing encoder E c and three view-specific encoders and Each layer consists of c1 shallow convolutional layers, P1 max pooling layers, and I1 residual blocks; each residual block consists of C2 convolutional layers and their skip connections; in this embodiment, C1 is 1, P1 is 1, I1 is 8, and C2 is 2; the activation function of all convolutional layers is set to the ReLU function.

[0073] Step 2.2: Construct a dual-adversarial representation learning module, including: axial-coronal view discriminator D1, axial-sagittal view discriminator D2, and axial view mutation discriminator D... av Coronal View Mutation Discriminator D cv and sagittal view mutation discriminator D sv ;

[0074] Among them, the view discriminator D1 and D2, and the mutation discriminator D av D cv and D sv All of them are network structures with different parameters based on generative adversarial networks;

[0075] Step 2.3: Construct a gene state detection module, including a weight generation module and a classification module;

[0076] The weight generation module and classification module include C3 convolutional layers, P2 max pooling layers and F fully connected layers; the activation function of all convolutional layers and the first F-1 fully connected layers uses the ReLU function; the activation function of the Fth fully connected layer is the Softmax function; in this embodiment, P2 is 1 and F is 2.

[0077] Step 3: The joint representation learning module extracts shared information and separates mutation-independent information from image samples;

[0078] Step 3.1: Transfer the multi-view data {I} of the i-th case. i,n |n=av,cv,sv} Input parameter shared encoder E c The data is processed in this way, mapping different CT views to the same common subspace to obtain an invariant representation of the multi-view data. This allows for the full extraction of shared information from multiple views; among which, This shows the invariant representation of the i-th case in the view direction n;

[0079] Step 3.2: I plot the lung CT image of the i-th case in the axial view direction av. i,av Input to axial view specific encoder The process is performed to obtain the specific characterization of the i-th case in the axial view direction av. This allows information unrelated to mutations to be captured in the axial view;

[0080] The lung CT image I of the i-th case in the sagittal view direction cv i,cv Input to coronal view specific encoder The process is performed to obtain the specific characterization of the i-th case in the sagittal view direction cv. This allows for the capture of information in the coronal view that is unrelated to the mutation;

[0081] The lung CT image I of the i-th case in the coronal view direction sv i,sv Input to sagittal view specific encoder The data is processed to obtain the specific characterization of the i-th case in the coronal view direction sv. This allows information unrelated to mutations to be captured in the sagittal view.

[0082] Step 3.3: To encourage view-invariant representation and view-specific characterization It can learn different aspects of multiple views, so a separation loss is used during training to impose orthogonal constraints on the view-invariant and specific representations of each view, thus separating the two representations; the separation loss function L for the i-th case is constructed using equation (1). i,sep :

[0083]

[0084] In equation (1), This represents the square of the Frobenius norm.

[0085] Step 4: The dual-adversarial representation learning module supervises the sharing of information and mutation-irrelevant information;

[0086] Step 4.1: The process of extracting shared information is inevitably affected by differences between views. In actual clinical practice, when analyzing CT images, doctors usually use axial views as the primary method, with the other two views serving as supplementary. Therefore, it is natural that this invention applies this process to the adversarial process. To represent the target view, with and The representation of the source view is aligned with the representation distribution of the target view.

[0087] To reduce the impact of differences between views, this embodiment introduces an axial-coronal view discriminator D1 to reduce... and The distribution differences. Specifically, and The input is processed in the axial-coronal view discriminator D1, thereby... Determined to be true. The result is determined to be false; in this embodiment, the axial-coronal view discriminator D1 and the parameter-sharing encoder E... c Establish an adversarial game process, and distinguish as much as possible in the view discriminator D1. and At that time, parameter sharing encoder E c Make the generated and Unable to be recognized by the view discriminator D1;

[0088] Introducing the axial-sagittal view discriminator D2 reduces and The distribution differences. Specifically, and The input is processed in the axial-sagittal view discriminator D2, thereby... Determined to be true. It was determined to be false.

[0089] Using equation (2), construct the view discrimination loss L for the multi-view data of the i-th case. i,v :

[0090] L i,v =L i,f +L i,t (2)

[0091] In equation (2), L i,f Let L represent the false adversarial loss of the multi-view data for the i-th case, obtained from equation (3), and let L represent the true adversarial loss of the multi-view data for the i-th case, obtained from equation (4); minimize the view discrimination loss L of view discriminators D1 and D2. i,v It can guide a shared encoder to generate a view-invariant representation with consistent distribution. Eliminating differences between different views allows for better extraction of shared information between views;

[0092]

[0093]

[0094] During training, a gradient inversion layer (GRL) is used to train the view discrimination loss. This layer keeps the input unchanged during forward propagation but inverts the gradient sign during backpropagation.

[0095] Step 4.2: During representation learning, information leakage inevitably occurs between different representations. To address this issue, this embodiment uses the mutation discriminator D... n and specific encoder Establish a game process between them, where n = av, cv, sv; Input to the corresponding mutation discriminator D av D cv and D sv The data is processed to obtain gene mutation category detection labels {y' based on view-specific characterization. i,n |n=av,cv,sv};where y' i,n The gene state category detection label represents the specific characterization of the i-th case based on the view direction n; in this process, the mutation discriminator detects the gene mutation status of the case as accurately as possible.

[0096] In contrast, specific encoders Attempts to generate specific representations that do not contain shared information prevent the mutation discriminator from correctly detecting the gene mutation status of a sample; and the mutation discrimination loss L for the i-th case is constructed using equation (5). i,m ;

[0097]

[0098] In equation (5), L i,av L represents the mutation discrimination loss of the i-th case in the axial view direction av. i,cv L represents the mutational discrimination loss of the i-th case in the sagittal view direction. i,sv The mutation discriminant loss for the i-th case in the coronal view direction sv is represented by CE(); the cross-entropy loss is represented by CE(); the mutation discriminant D is minimized. n Mutation discrimination loss L for n = av, cv, sv i,m Can guide specific encoders Generate view-specific representations that do not contain shared information n = av, sv, sv ensures that the view-specific representation is independent of the mutation state, thereby significantly reducing the mutation-irrelevant information mixed in during multi-view fusion and effectively solving the problem of poor multi-view representation quality. During training, a gradient inversion layer is used to train the mutation discrimination loss. This layer keeps the input unchanged during forward propagation but inverts the gradient sign during backpropagation.

[0099] Step 4.3: Construct the bi-adversarial loss function L for the i-th case using equation (6). i,dadv :

[0100] L i,dadv =α i L i,v +β i L i,m (6)

[0101] In equation (6), α i L i,v The weighting coefficient, β i L i,m The weighting coefficients.

[0102] Step 5: The gene state detection module fuses and detects the characteristics;

[0103] Step 5.1: The weight generation module uses equation (7) to generate the weight w of the i-th case in the view direction n. i,n :

[0104]

[0105] In equation (7), MLP represents a multilayer perceptron.

[0106] Step 5.2: The classification module uses equation (8) to obtain the multi-view data fusion representation F of the i-th case. i :

[0107]

[0108] Step 5.3: Construct the classification loss function L for the i-th case using equation (9). i,cla :

[0109] L i,cla =-y i log(MLP(F i ))-(1-y i )log(1-MLP(F i (9).

[0110] Step 6: Construct the total loss function and train the detection model to obtain the trained detection model;

[0111] Step 6.1: Construct the total loss function L total ;

[0112] View discrimination loss function {L} for multi-view data of N cases i,v The mean value of |i=1,2,…N} is calculated to obtain the average view discrimination loss function L. v ;

[0113] Mutation identification loss function {L} for multi-view data of N cases i,m The mean value of |i=1,2,…N} is calculated to obtain the average view discrimination loss function L. m ;

[0114] Separation loss function {L} for multi-view data of N cases i,sep The mean value of |i=1,2,…N} is calculated to obtain the average separation loss function L. sep ;

[0115] The classification loss function {L} for multi-view data of N cases i,cla The mean of |i=1,2,…N} is calculated to obtain the average classification loss function L. cla .

[0116] Construct the overall loss function L using equation (11) total ;

[0117] L total =L cla +γL sep +(αL v +βL m (11)

[0118] In equation (11), γ represents L sep The weighting coefficient, α represents L v The weighting coefficient, β represents L m The weighting coefficients are: in this embodiment, γ is 0.5, and α and β are 0.7 and 0.1, respectively.

[0119] Step 6.2: Continuously train the non-invasive gene mutation detection network using the Adam optimizer and calculate the total loss function L. total The network parameters are updated until the maximum number of training rounds is reached, thereby obtaining a trained non-invasive gene detection model based on lung CT images, which can be used to achieve non-invasive detection of gene mutation status.

[0120] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the above-described method, and the processor is configured to execute the program stored in the memory.

[0121] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

[0122] To quantitatively evaluate the performance of the proposed method, this embodiment presents the performance comparison results of a five-fold cross-validation experiment on a dataset of 420 lung EGFR gene mutation CT images, with the multi-view classification method of Su et al., Feng et al., and the deep learning method in the field of multi-view medical imaging of Wu et al., Li et al., as shown in Table 1.

[0123] Table 1 Comparison with other methods

[0124]

[0125] In the experimental process, this invention first learns view-invariant representations by combining different views of CT images, fully extracting shared information from multiple views and thus making full use of the shared information between views. Simultaneously, by learning the specific representations of each view, this method also captures mutation-independent information in each view, significantly reducing mutation-independent information mixed in during multi-view fusion. Furthermore, a dual adversarial representation learning strategy is used to strengthen the learning of invariant and specific representations, effectively improving the model's detection performance.

Claims

1. A non-invasive method for detecting gene mutations based on multi-view CT images, characterized in that, Includes the following steps: Step 1: Acquisition and preprocessing of lung CT data; Get those with gene mutation category tags The lung CT image data of each case were collected, and all lung CT image data were resampled using a trilinear interpolation algorithm to standardize the voxel size, thus obtaining a set of lung CT slices for each case and its corresponding category label, denoted as . ,in, Indicates the first A lung CT scan of a case. Indicates the first Lung CT slices of the case Corresponding gene mutation category labels; ; Indicates the number of cases; by The abnormal location center is taken as the origin, and along the axial view direction respectively. Coronal view direction Sagittal view direction Cropping an image patch with dimensions H×W pixels yields the first... One case in axial view direction CT images of the lungs Coronal view direction CT images of the lungs Sagittal view direction CT images of the lungs ; The first Lung CT images of the case Lung CT images and lung CT images Composition of the first Multi-view data for individual cases ;in, Indicates the first Individual cases in view direction CT images of the lungs; Step 2: Construct a non-invasive gene mutation detection model; Step 2.1: Construct a joint representation learning module, including: a parameter-shared encoder. And three view-specific encoders, namely axial view-specific encoders Coronal view specific encoder and sagittal view specific encoder ; Step 2.2: Construct a dual adversarial representation learning module, including: an axial-coronal view discriminator. Axial-Sagittal View Discriminator Axial view mutation discriminator Coronal view mutation discriminator and sagittal view mutation discriminator ; Step 2.3: Construct a gene state detection module, including a weight generation module and a classification module; Step 3: The joint representation learning module extracts shared information and separates mutation-irrelevant information from the lung CT images; Step 3.1, place the first Multi-view data for individual cases Input the parameters shared encoder The data is processed to obtain an invariant representation of the multi-view data. ,in, Show the first Individual cases in view direction Invariant representations on; Step 3.2, place the first One case in axial view direction CT images of the lungs Input to the axial view specific encoder Processing is performed to obtain the first... One case in axial view direction Specific characterization on ; The first One case in the sagittal view direction CT images of the lungs Input to the coronal view specific encoder Processing is performed to obtain the first... One case in the sagittal view direction Specific characterization on ; The first One case in the coronal view direction CT images of the lungs Input to the sagittal view-specific encoder Processing is performed to obtain the first... One case in the coronal view direction Specific characterization on ; Step 4: The dual-adversarial representation learning module supervises the shared information and mutation-independent information; Step 4.1, with To represent the target view, with and Representation of the source view; Will and Input to the axial-coronal view discriminator Processing is performed in order to... Determined to be true. Determined to be false; Will and Input to the axial-sagittal view discriminator Processing is performed in order to... Determined to be true. Determined to be false; Step 4.2, Input to the corresponding axial view mutation discriminator Coronal view mutation discriminator and sagittal view mutation discriminator The data is processed to obtain gene mutation category detection labels based on view-specific characterization. ;in, Indicates the first Cases based on view orientation Specific characterization of gene state category detection labels; Step 5: The gene state detection module fuses and detects the characteristics; Step 5.1: The weight generation module uses equation (7) to generate the first weight. Individual cases in view direction Weight : (7) In equation (7), This represents a multilayer perceptron; Step 5.2: The classification module uses equation (8) to obtain the first... Multi-view data fusion characterization of individual cases : (8) Step 6: Construct the total loss function and train the detection model to obtain the trained detection model; Step 6.1: Construct the total loss function ; right View discrimination loss function for multi-view data of a case Calculate the mean to obtain the average view discrimination loss function. ;in, For the first The visual identification loss of individual cases; right Mutation identification loss function for multi-view data of individual cases Calculate the mean to obtain the average view discrimination loss function. ;in, For the first Loss of mutation identification in individual cases; right Separation loss function for multi-view data of individual cases Calculate the mean to obtain the average separation loss function. ;in, For the first The separation loss function for each case; right Classification loss function for multi-view data of individual cases Calculate the mean to obtain the average classification loss function. ;in, For the first The classification loss function for each case; Construct the overall loss function using equation (11) ; (11) In equation (7), represent The weighting coefficients, represent The weighting coefficients, represent Weighting coefficients; Step 6.2: Continuously train the non-invasive gene mutation detection network using the Adam optimizer and calculate the total loss function. The network parameters are updated until the maximum number of training rounds is reached, thereby obtaining a trained non-invasive gene detection model based on lung CT images, which can be used to achieve non-invasive detection of gene mutation status.

2. The non-invasive gene mutation detection method based on multi-view CT images according to claim 1, characterized in that: The parameter-sharing encoder and three view-specific encoders , and All by A shallow convolutional layer Each residual block and It consists of a maximum pooling layer; each residual block is composed of... The system consists of several convolutional layers and their skip connections; the activation function of all convolutional layers is set to the ReLU function. View discriminator and Mutation discriminator , and All of them are network structures with different parameters based on generative adversarial networks; Both the weight generation module and the classification module include Each convolutional layer There are 1 max-pooling layer and F fully connected layers; the activation function of all convolutional layers and the first F-1 fully connected layers is the ReLU function; the activation function of the Fth fully connected layer is the Softmax function.

3. The non-invasive gene mutation detection method based on multi-view CT images according to claim 2, characterized in that: Construct the first using equation (2) Individual case view identification loss : (2) In equation (2), Indicates the first The false adversarial loss for each case is obtained from equation (3). Indicates the first The true adversarial loss for each case is obtained from equation (4); (3) (4)。 4. The non-invasive gene mutation detection method based on multi-view CT images according to claim 3, characterized in that: Construct the first using equation (1) Separation loss function for each case : (1) In equation (1), This represents the square of the Frobenius norm.

5. The non-invasive gene mutation detection method based on multi-view CT images according to claim 4, characterized in that: Construct the first using equation (5) Mutation identification loss in individual cases ; (5) In equation (5), Indicates the first One case in axial view direction mutation identification loss, Indicates the first One case in the sagittal view direction Mutation identification loss, Indicates the first One case in the coronal view direction Mutation identification loss; This represents the cross-entropy loss.

6. The non-invasive gene mutation detection method based on multi-view CT images according to claim 5, characterized in that: Construct the first using equation (9) Classification loss function for each case : (9)。 7. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store programs that support the processor in executing any of the non-invasive gene mutation detection methods according to claims 1-6, and the processor is configured to execute the programs stored in the memory.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is run by the processor, it performs the steps of any of the non-invasive gene mutation detection methods described in claims 1-6.

Images