Single gene mutation status prediction method, system and product based on CT images

Abstract

The present invention belongs to the technical field of medical image processing, and provides a single gene mutation state prediction method, system and product based on CT images. The technical solution is that in each feature hierarchical structure, the image block input into the layer or the feature extracted from the previous layer is subjected to multi-scale convolution and channel attention mechanism to capture the multi-scale features of the image, the multi-scale features of the image extracted by each feature hierarchical structure are subjected to the optimized SwinTransformer Block to extract the deep features of the image, and after fusing the multi-scale features and the corresponding extracted deep features, the extracted channel and space features are adaptively optimized through the CBAM module, the channel and spatial relationship of the features is optimized, the model's ability to capture detailed features in the image and its attention to key biomarkers is enhanced, the accuracy of classification is improved, and the problem of difficulty in predicting the single gene mutation state of KRAS is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image processing, and in particular relates to a method, system and product for predicting the mutation status of a single gene based on CT images. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] EGFR (epidermal growth factor receptor) and KRAS (Kirsten rat sarcoma) are the two most important mutations in patients with non-small-cell lung cancer (NSCLC). Mutations in these two genes can lead to abnormal proliferation of somatic cells, the occurrence of tumors, promote cancer progression, and have a negative impact on treatment. Therefore, for NSCLC patients, it is very important to accurately detect the mutation status of KRAS and EGFR genes.

[0004] Computed tomography (CT) is a computer-assisted X-ray procedure that combines many X-ray images to create cross-sectional views and three-dimensional images of the body's internal organs. Compared with conventional chest X-ray film methods, CT scans can observe the lungs in more detail and show the size and location of tumors. Currently, CT scans are a commonly used imaging method for tumor examinations, providing a non-invasive and easy-to-use method for lung cancer analysis. CT images can cover the entire tumor area, including the tumor microenvironment and surrounding attached tissues, allowing researchers to predict the status of related gene mutations based on the entire tumor.

[0005] The machine learning-based radiomics method is used to process CT images, but the features extracted by this method can only reflect the general characteristics of the image, and cannot accurately reflect the characteristics specifically associated with EGFR or KRAS gene mutations. In particular, the radiomics method also requires precise determination of the tumor boundary and extraction of relevant features using feature engineering methods, which is very time-consuming and requires a lot of manual labeling work and comprehensive guidance from experienced radiologists.

[0006] In recent years, deep learning methods have achieved substantial success in the field of artificial intelligence due to their powerful automated feature extraction and classification capabilities. The end-to-end nature of deep learning enables researchers to avoid tedious manual labeling and feature extraction processes, showing a strong advantage over imaging omics.

[0007] However, for NSCLC, most of the deep learning image processing methods currently used are based on the simultaneous prediction of the mutation status of at least two genes, such as EGFR and KRAS, or focus on the EGFR gene mutation status for prediction, and there are very few predictions for the single gene mutation status of KRAS. Summary of the invention

[0008] In order to solve at least one technical problem existing in the above-mentioned background technology, the present invention provides a single gene mutation status prediction method and system based on CT images, which is aimed at NSCLC, uses CT images to predict the KRAS single gene mutation status, and adaptively optimizes the multi-scale feature representation obtained by extracting each feature hierarchical structure and the corresponding extracted deep features to improve the prediction effect.

[0009] In order to achieve the above object, the present invention adopts the following technical solution:

[0010] The first aspect of the present invention provides a method for predicting the mutation status of a single gene based on CT images, comprising the following steps:

[0011] Obtain CT images of patients with non-small cell lung cancer;

[0012] Based on the CT images of patients with non-small cell lung cancer and the trained KRAS gene single mutation status prediction model, the KRAS single gene mutation status is predicted; wherein the construction process of the KRAS gene single mutation status prediction model includes:

[0013] Image feature representations of different scales are obtained through multiple feature extraction hierarchical structures;

[0014] Represent image features at different scales, and extract deep features of the image through the optimized SwinTransformer Block;

[0015] The multi-scale feature representation extracted from each feature hierarchical structure and the corresponding deep features extracted are added through residual connection. After passing through the CBAM module, the extracted channel and spatial features are adaptively optimized along the two dimensions of channel and space, and then downsampled to obtain the output features of this layer;

[0016] The output features of the previous layer are used as the input of the next layer, extracted layer by layer, and sent to the classification layer to output the final state prediction result.

[0017] Furthermore, image feature representations of different scales are obtained through multiple feature extraction hierarchical structures, including:

[0018] The entire CT image is divided into image blocks of fixed size. The image blocks input into the current layer or the features extracted from the previous layer are processed through multi-scale convolution and channel attention mechanism to capture the multi-scale features of the image, including:

[0019] The image block or the features extracted from the previous layer are convolved with three convolution kernels of different sizes to generate three feature maps.

[0020] The three feature maps are added and fused, and global average pooling is performed to generate a global feature vector;

[0021] The global feature vector passes through two layers of fully connected networks, and the channel weights are generated using the Sigmoid activation function;

[0022] Multiply the channel weights by the feature maps generated by the corresponding convolution kernels channel by channel, and then add the weighted feature maps to obtain the final feature representation.

[0023] Furthermore, when optimizing the SwinTransformer Block, based on the Swin Transformer V2 model, residual normalization is used to move the LN layer from the beginning of each residual unit to the back end, and the output of each residual module is normalized before being merged with the backbone network;

[0024] or,

[0025] When optimizing the SwinTransformer Block, it also includes: after the output of the first residual block, the MLP module is replaced with a convolutional feed-forward module in the ConvFormer, wherein the convolutional feed-forward module includes two 1×1 convolutions, batch normalization, and GeLU.

[0026] Furthermore, before entering the LN layer, the window attention mechanism is executed. Inside each window, a custom convolution kernel is created to build an adaptive receptive field for each pixel, establish long-distance dependencies, and calculate the self-attention inside each window.

[0027] Furthermore, the process of creating a custom convolution kernel within each window, constructing an adaptive receptive field for each pixel, establishing long-distance dependencies, and calculating self-attention within each window includes:

[0028] The features output by multi-scale convolution channel attention are used as input features;

[0029] Perform convolution operation on the input features to fuse the features of adjacent 3×3 pixels into the feature representation of the corresponding pixel position in the query, key, and value;

[0030] The scaled cosine function is used to calculate the similarity of pixel pairs by combining the dot product of the query and key vectors to obtain the relative position deviation between the pixels and;

[0031] Based on the relative position deviation between pixels and, self-attention is calculated for query, key, and value to obtain the self-attention value within each window.

[0032] Furthermore, when training the KRAS gene single mutation status prediction model, the cross entropy loss function was used for training, and the stochastic gradient descent optimization algorithm was used to optimize the model.

[0033] Furthermore, the method also includes preprocessing the CT image after acquiring the CT image of the non-small cell lung cancer patient, selecting a cubic region of interest containing the entire tumor, and then standardizing the data using Z-score standardization.

[0034] Furthermore, the accuracy and area under the ROC curve (AUC) were used to evaluate the generalization ability of the KRAS gene single mutation status prediction model.

[0035] The second aspect of the present invention provides a single gene mutation status prediction system based on CT images, comprising:

[0036] An image acquisition module, which is used to acquire CT images of patients with non-small cell lung cancer;

[0037] A gene mutation status prediction module is used to predict the KRAS single gene mutation status based on the CT images of non-small cell lung cancer patients and the trained KRAS gene single mutation status prediction model; wherein the construction process of the KRAS gene single mutation status prediction model includes:

[0038] Image feature representations of different scales are obtained through multiple feature extraction hierarchical structures;

[0039] Represent image features at different scales, and extract deep features of the image through the optimized SwinTransformer Block;

[0040] The multi-scale feature representation extracted from each feature hierarchical structure and the corresponding deep features extracted are added through residual connection. After passing through the CBAM module, the extracted channel and spatial features are adaptively optimized along the two dimensions of channel and space, and then downsampled to obtain the output features of this layer;

[0041] The output features of the previous layer are used as the input of the next layer, extracted layer by layer, and sent to the classification layer to output the final state prediction result.

[0042] A third aspect of the present invention provides a program product.

[0043] A program product, which is a computer program product, includes a computer program, and when the computer program is executed by a processor, the steps in the method for predicting the mutation status of a single gene based on CT images as described above are implemented.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] 1. The present invention captures the multi-scale features of the image through multi-scale convolution and channel attention mechanism in each feature hierarchical structure, extracts the deep features of the image through the optimized SwinTransformer Block, and after fusing the multi-scale features and the corresponding extracted deep features, adaptively optimizes the extracted channel and space features through the CBAM module, optimizes the channel and spatial relationship of the features, enhances the model's ability to capture detailed features in the image and its attention to key biomarkers, improves the accuracy of classification, and solves the problem of difficulty in predicting the single gene mutation status of KRAS.

[0046] 2. By introducing an improved Swin Transformer model, the window attention mechanism and multi-layer perceptron (MLP) layer in the Swin Transformer model were optimized to solve the common attention collapse problem in complex medical image analysis, thereby improving the accuracy of feature extraction and the prediction effect of the model, and significantly improving the prediction ability of the KRAS gene mutation status in CT images of NSCLC patients.

[0047] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0049] Figure 1 is a flow chart of a method for predicting a single gene mutation state based on CT images provided by an embodiment of the present invention;

[0050] Figure 2 It is a schematic diagram of the overall structure of SwinTransfomer provided by an embodiment of the present invention;

[0051] Figure 3 It is a schematic diagram of the structure of each basic layer in SwinTransfomer provided by an embodiment of the present invention;

[0052] Figure 4It is a schematic diagram of the structure of the MSCACM module provided by an embodiment of the present invention;

[0053] Figure 5 is a schematic diagram of the structure of an improved Swin Transformer block provided in an embodiment of the present invention;

[0054] Figure 6 It is a schematic diagram of the window attention mechanism optimization provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0055] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0056] It should be noted that the following detailed descriptions are all illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0057] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0058] Embodiment 1

[0059] like Figure 1 As shown, this embodiment provides a single gene mutation status prediction method based on CT images, comprising the following steps:

[0060] Step 1: Obtain CT images of patients with non-small cell lung cancer;

[0061] In this embodiment, the CT image is scanned by using Siemens Somatom Sensation AS128 or Siemens Somatom Definition Flash dual-source CT for chest CT scanning. After holding the breath, the scan is performed from the apex to the base of the lung, including both sides of the chest wall and the axilla.

[0062] CT scanning and image reconstruction parameters were as follows:

[0063] Tube voltage 120 kV, automatic tube current, pitch 0.6-1.2, conventional image layer thickness 8-10 mm, thin image layer thickness 1.0 mm, lung window width 1500-1800 HU, window level -500 600 HU, mediastinal window 250-350 HU, window level 40-50 HU, standard algorithm reconstruction. Images were acquired in DICOM format, which is a standardized format for medical images and related information designed to ensure the interoperability of medical images between different devices and systems. It is a widely used standard in the medical field for acquiring, transmitting, storing and viewing medical imaging data such as X-rays, CT scans, MRI images, etc.

[0064] Step 2: preprocessing the acquired CT images of patients with non-small cell lung cancer;

[0065] Select a cubic region of interest (ROI) that contains the entire tumor. When selecting, the ROI should include the complete tumor area, including the tumor edge. After that, adjust the pixels of the ROI in each CT slice by cubic spline interpolation;

[0066] The data was then standardized using Z-score standardization. After standardization, the data was adjusted to conform to a normal distribution with a mean of 0 and a variance of 1. The purpose of standardization is to restore the gradually deviated distribution to the standard distribution before the activation function.

[0067] Step 3: Based on the preprocessed CT images and the trained gene mutation prediction model, the KRAS single gene mutation status is predicted;

[0068] The specific steps include:

[0069] Step 301: Segment the entire CT image into image blocks of fixed size, and obtain feature representations of different scales through multiple feature extraction hierarchical structures;

[0070] The gene mutation prediction model in this embodiment is improved based on the SwinTransformer structure, and the overall structure of the model is consistent with the original SwinTransformer structure. Figure 2 As shown, it includes multiple hierarchical units, which first divide the entire CT image into small blocks of fixed size (called "patches"), and then gradually reduce the size of these blocks and increase the number of channels through each hierarchical structure to obtain hierarchical feature representations of different scales;

[0071] The structure of the basic layer in each hierarchical unit is as follows: Figure 3 As shown in Figure 2, unlike SwinTransformer, the image blocks input to this layer, or the features extracted from the previous stage, are first passed through the constructed MSCACM, as shown in Figure 2. Figure 4 The MSCACM constructed in this embodiment is shown, which is based on multi-scale convolution and channel attention mechanism to capture multi-scale information in the image, and at the same time uses the channel attention mechanism to improve the effectiveness of feature representation.

[0072] like Figure 4 As shown in Figure 2, the specific feature extraction process includes:

[0073] Step 3011: The input feature map is convolved with three convolution kernels of different sizes (1x1, 3x3, 5x5) to generate three feature maps;

[0074] Step 3012: Add and fuse the three feature maps, and perform global average pooling to generate a 1x1xC global feature vector;

[0075] Step 3013: The global feature vector passes through two layers of fully connected networks (with ReLU activation in the middle), the output dimension becomes 1x1x(3C), and the Sigmoid activation function is used to generate channel weights;

[0076] Step 3014: Multiply the channel weights by the feature maps generated by the corresponding convolution kernels channel by channel, and then add the weighted feature maps to obtain the final feature representation.

[0077] Through the MSCACM module, feature information of different scales can be captured through convolution kernels of different sizes. Combining these multi-scale convolution operations, image features can be fully extracted, and the weight of each channel can be dynamically adjusted to enhance important features and reduce interference from irrelevant information.

[0078] Step 302: extracting multi-scale features of the image from each feature hierarchical structure, and extracting deep features of the image through the optimized SwinTransformer Block;

[0079] After extracting shallow features, the SwinTransformer Block is used to extract deep features and optimize the SwinTransformer Block. For stable training of deep network models, the original SwinTransformer and ViT both use a pre-normalization method, but the activation values ​​are accumulated layer by layer, and the amplitude of the deep layer is significantly larger than that of the early layer. Figure 5 As shown in the figure, to solve this problem, in the Swin Transformer V2 model, residual normalization is used to move the LN layer from the beginning of each residual unit to the back end. The output of each residual module will be normalized before being merged with the backbone network. Therefore, the amplitude of the main branch will not accumulate as the number of layers increases.

[0080] In addition, after the output of the first residual block, this study borrows the CFFN module in ConvFormer and replaces the MLP with a convolution feed-forward network (CFFN). The CFFN consists of two 1×1 convolutions, batch normalization, and GeLU.

[0081] A Swin Transformer Block process can be expressed by the following formula (1) and formula (2):

[0082] (1),

[0083] (2),

[0084] in, is the feature output by multi-scale convolution channel attention, To perform window multi-head self-attention processing, is the feature output by window self-attention, is the feature output by the convolutional feedforward module.

[0085] The CFFN module in formula (2) can be expressed by the following formula (3):

[0086] (3)

[0087] In formula (3) Is Convolution, batch normalization and GeLU activation function together can be expressed by the following formula (4):

[0088] (4)

[0089] Figure 3 In the Swin Transformer Block in , the window attention mechanism (windowattention) process is specifically performed, that is, Figure 5 In each window, we optimize the self-attention process by referring to Swin Transformer V2 and ConvFormer. The optimization process is as follows: Figure 6 As shown in the figure, in the specific window attention mechanism, a custom convolution kernel is created to build an adaptive receptive field for each pixel and establish long-distance dependencies.

[0090] Take the features output by multi-scale convolution channel attention as input , first by the size of The convolution kernel is obtained by convolution operation. That is, query, key, value, which can be specifically expressed by the following formulas (5)-(7):

[0091] (5)

[0092] (6)

[0093] (7)

[0094] In formulas (5)-(7), , is the learnable weight in the 2D convolution kernel, The dimension of and the internal attention in Vision Transformer The dimension corresponds to l , g , i , k Represents pixel coordinates.

[0095] For input features Perform convolution operation to merge the features of adjacent 3×3 pixels into The feature representation of the corresponding pixel position in .

[0096] Then, the calculated Perform self-attention calculation. In Swin Transformer, self-attention is calculated by formula (8):

[0097] (8)

[0098] in, is the dimension of the vector, is the bias term used to fit the data;

[0099] In formula (8), query and key Vector dot product is used to calculate the similarity of pixel pairs. However, the attention maps learned in some blocks and heads are often dominated by a small number of pixel pairs. Therefore, in this embodiment, the scaled cosine function proposed by Swin TransformerV2 is used to calculate the similarity of pixel pairs. The calculation formula is:

[0100] (9)

[0101] in, For the i A query vector (Query), For the j A key vector (Key), is the temperature parameter, which is used to scale the value of cosine similarity to control the smoothness of softmax. To represent the query vector q The bias term between the key vector k and the key vector k is used to adjust the calculated similarity value.

[0102] The relative position deviation between pixels and can be calculated by formula (9).

[0103] (10)

[0104] The self-attention within each window can be calculated by formula (10).

[0105] Step 303: The features extracted by SwinTransformer Block and the multi-scale features initially extracted by MSCACM are added through residual connection, and the attention map is inferred along the two dimensions of channel and space through the CBAM module. The channel and space features are adaptively optimized, and finally down-sampling is performed, that is, the image blocks are fused as the output of this layer.

[0106] Step 304: Finally, the features extracted by multiple Swin Transformer based basic layers are sent to the final classification layer to predict the mutation status of the KRAS single gene.

[0107] During model training, this embodiment uses the cross entropy loss function for training, and uses the stochastic gradient descent optimization algorithm to optimize the model. In addition, techniques such as weight decay regularization, batch normalization, layer normalization, and discarding are also applied to optimize the model. Specifically, the "Kaiming uniform" method is used to initialize the network weights in the model, and the AdamW optimizer is used for 300 training cycles, using a cosine decay learning rate scheduler and a linear warm-up of 20 cycles. Other training settings include a learning rate of 0.001, a batch size of 128, a weight decay of 0.0001, and a discard coefficient of 0.2.

[0108] Taking into account the imbalance of samples, accuracy (ACC) and area under the ROC curve AUC were selected to evaluate the generalization ability of the model.

[0109] The accuracy is defined by the following formula:

[0110] (11)

[0111] in, Represents the number of samples that the model correctly predicts as positive examples, Indicates the number of samples that the model correctly predicts as negative examples, Indicates the number of samples where the model incorrectly predicts negative examples as positive examples, Indicates the number of examples where the model incorrectly predicted positive examples as negative examples.

[0112] The vertical axis of the ROC curve represents the true positive rate (TPR), and the horizontal axis represents the false positive rate (FPR), which are expressed by the following formulas:

[0113] (12)

[0114] (13)

[0115] AUC is the area under the ROC curve, which can be obtained by adding the areas of each part under the ROC curve.

[0116] Assume that the ROC curve has coordinates by connecting The point formation, where ,and Represents the number of coordinate points. AUC can be calculated by the following formula:

[0117] (14)

[0118] After comparing with the existing classification model, the following comparison results are obtained, as shown in Table 1.

[0119] Table 1 Comparison of prediction results between the method of the present invention and the existing classification model

[0120]

[0121] By comparison, it can be seen that the research results of the method of the present invention show obvious advantages over the existing research based on machine learning and traditional imaging omics methods, especially in dealing with specific challenges of medical images, such as high heterogeneity and sample inhomogeneity. The model can provide more stable and reliable output. By in-depth analysis of the model's performance on NSCLC lung cancer images of different types and stages, its application potential in early diagnosis and personalized treatment strategies is further confirmed.

[0122] Embodiment 2

[0123] This embodiment provides a single gene mutation status prediction system based on CT images, including:

[0124] An image acquisition module, which is used to acquire CT images of patients with non-small cell lung cancer;

[0125] A gene mutation status prediction module is used to predict the KRAS single gene mutation status based on the CT images of non-small cell lung cancer patients and the trained KRAS gene single mutation status prediction model; wherein the construction process of the KRAS gene single mutation status prediction model includes:

[0126] Image feature representations of different scales are obtained through multiple feature extraction hierarchical structures;

[0127] Represent image features at different scales, and extract deep features of the image through the optimized SwinTransformer Block;

[0128] The multi-scale feature representation extracted from each feature hierarchical structure and the corresponding deep features extracted are added through residual connection. After passing through the CBAM module, the extracted channel and spatial features are adaptively optimized along the two dimensions of channel and space, and then downsampled to obtain the output features of this layer;

[0129] The output features of the previous layer are used as the input of the next layer, extracted layer by layer, and sent to the classification layer to output the final state prediction result.

[0130] In this embodiment, the implementation process of each module is consistent with the content of the first embodiment, and is implemented with reference to the process of the first embodiment.

[0131] Embodiment 3

[0132] This embodiment provides a program product, which is a computer program product, including a computer program. When the computer program is executed by a processor, the steps in the method for predicting the mutation status of a single gene based on CT images as described in Embodiment 1 are implemented.

[0133] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A single gene mutation status prediction method based on CT images, characterized in that: The steps include: Obtain CT images of patients with non-small cell lung cancer; Based on the CT images of patients with non-small cell lung cancer and the trained KRAS gene single mutation status prediction model, the KRAS single gene mutation status is predicted; wherein the construction process of the KRAS gene single mutation status prediction model includes: Image feature representations of different scales are obtained through multiple feature extraction hierarchical structures; Represent image features at different scales, and extract deep features of the image through the optimized SwinTransformer Block; When optimizing the SwinTransformer Block, based on the Swin Transformer V2 model, residual normalization is used to move the LN layer from the beginning of each residual unit to the back end, and the output of each residual module is normalized before being merged with the backbone network; After the output of the first residual block, the MLP module is replaced with a convolutional feed-forward module in ConvFormer, which includes two 1×1 convolutions, batch normalization, and GeLU; The multi-scale feature representation extracted from each feature hierarchical structure and the corresponding deep features extracted are added through residual connection. After passing through the CBAM module, the extracted channel and spatial features are adaptively optimized along the two dimensions of channel and space, and then downsampled to obtain the output features of this layer; The output features of the previous layer are used as the input of the next layer, extracted layer by layer, and sent to the classification layer to output the final state prediction result.

2. The method for predicting single gene mutation status based on CT images according to claim 1, characterized in that: Image feature representations of different scales are obtained through multiple feature extraction hierarchical structures, including: The entire CT image is divided into image blocks of fixed size. The image blocks input into the current layer or the features extracted from the previous layer are processed through multi-scale convolution and channel attention mechanism to capture the multi-scale features of the image, including: The image block or the features extracted from the previous layer are convolved with three convolution kernels of different sizes to generate three feature maps. The three feature maps are added and fused, and global average pooling is performed to generate a global feature vector; The global feature vector passes through two layers of fully connected networks, and the channel weights are generated using the Sigmoid activation function; Multiply the channel weights by the feature maps generated by the corresponding convolution kernels channel by channel, and then add the weighted feature maps to obtain the final feature representation.

3. The method for predicting single gene mutation status based on CT images according to claim 1, characterized in that: Before entering the LN layer, the window attention mechanism is executed. Inside each window, a custom convolution kernel is created to build an adaptive receptive field for each pixel, establish long-distance dependencies, and calculate the self-attention inside each window.

4. The method for predicting single gene mutation status based on CT images according to claim 3, characterized in that: The process of creating a custom convolution kernel inside each window, building an adaptive receptive field for each pixel, establishing long-distance dependencies, and calculating the self-attention inside each window includes: The features output by multi-scale convolution channel attention are used as input features; Perform convolution operation on the input features to fuse the features of adjacent 3×3 pixels into the feature representation of the corresponding pixel position in the query, key, and value; The scaled cosine function is used to calculate the similarity of pixel pairs by combining the dot product of the query and key vectors to obtain the relative position deviation between the pixels and; Based on the relative position deviation between pixels and, self-attention is calculated for query, key, and value to obtain the self-attention value within each window.

5. The method for predicting single gene mutation status based on CT images according to claim 1, characterized in that: When training the KRAS gene single mutation status prediction model, the cross entropy loss function was used for training, and the stochastic gradient descent optimization algorithm was used to optimize the model.

6. The method for predicting single gene mutation status based on CT images according to claim 1, characterized in that: The method further comprises, after acquiring a CT image of a patient with non-small cell lung cancer, preprocessing the CT image, selecting a cubic region of interest containing the entire tumor, and then using Z Score normalization normalizes the data.

7. The method for predicting single gene mutation status based on CT images according to claim 1, characterized in that: The accuracy and area under the ROC curve (AUC) were used to evaluate the generalization ability of the KRAS gene single mutation status prediction model.

8. A single gene mutation status prediction system based on CT images, using a single gene mutation status prediction method based on CT images as claimed in any one of claims 1 to 7, characterized in that: include: An image acquisition module, which is used to acquire CT images of patients with non-small cell lung cancer; A gene mutation status prediction module is used to predict the KRAS single gene mutation status based on the CT images of non-small cell lung cancer patients and the trained KRAS gene single mutation status prediction model; wherein the construction process of the KRAS gene single mutation status prediction model includes: Image feature representations of different scales are obtained through multiple feature extraction hierarchical structures; Represent image features at different scales, and extract deep features of the image through the optimized SwinTransformer Block; The multi-scale feature representation extracted from each feature hierarchical structure and the corresponding deep features extracted are added through residual connection. After passing through the CBAM module, the extracted channel and spatial features are adaptively optimized along the two dimensions of channel and space, and then downsampled to obtain the output features of this layer; The output features of the previous layer are used as the input of the next layer, extracted layer by layer, and sent to the classification layer to output the final state prediction result.

9. A program product, the program product being a computer program product, comprising a computer program, characterized in that: When the computer program is executed by a processor, the steps in the method for predicting a single gene mutation status based on CT images as described in any one of claims 1 to 7 are implemented.

Images