A deep learning cascade CT lung nodule detection method and device
Through deep learning cascade method combined with CT imaging and cfDNA methylation characteristics, a high sensitivity and high specificity pulmonary nodule diagnostic model was constructed, solving the high sensitivity but specificity of CT screening in the prior art Low problems achieve more accurate diagnosis of pulmonary nodules and reduce detection costs.
Patent Information
- Application Number
- CN202210767083.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-07-01
- Publication Date
- 2024-12-03
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing low-dose CT screening is highly sensitive but not specific in early detection of lung cancer, resulting in many non-neoplastic pulmonary nodules being misdiagnosed as malignant, lacking diagnostic methods with high sensitivity and specificity, especially non-invasive methods. .
The deep learning cascade method was adopted, combined with YOLOX algorithm detection and vgg16-unet model to segment CT images, extract the imagingomics characteristics of lung nodules, and processed in combination with peripheral blood cfDNA methylation characteristics, and constructed a diagnostic model through a random forest algorithm.
It improves the accuracy of pulmonary nodules diagnosis and reduces the detection cost, especially in the detection rate of early lung cancer in lung nodules, which is better than the diagnostic effect of a single-computer algorithm model.
Smart Images

Figure CN115100156B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of data processing, and in particular to a deep learning cascade CT lung nodule detection method and device. Background Art
[0002] Lung cancer is the leading cause of cancer death worldwide. Most patients lack symptoms in the early stages of the disease and are diagnosed at an advanced stage, resulting in a poor prognosis. Early detection of lung cancer is crucial for reducing mortality and improving patient survival. Low-dose CT (LDCT) screening has been widely used for early detection of lung cancer, as the National Lung Screening Trial demonstrated that its use could reduce lung cancer mortality by 20%. However, due to the high sensitivity but low specificity of LDCT screening, many non-neoplastic pulmonary nodules are also detected. Although CT imaging can estimate the risk of nodule malignancy based on nodule size, lobar location, density, and margin characteristics, approximately 20% of nodules are misdiagnosed as malignant by CT screening. Therefore, new diagnostic methods with high sensitivity and specificity for distinguishing malignant tumors from benign pulmonary nodules remain a pressing need, ideally achieved through non-invasive methods.
[0003] However, CT image features have rarely been considered in previous noninvasive diagnostic models. Artificial intelligence (AI) algorithms play a crucial role in the automated detection, segmentation, and computer-aided diagnosis of malignant lesions. Based on published studies, radiomics of CT images combined with AI algorithms appears to be the most promising for differentiating benign from malignant pulmonary nodules. Radiomics is a growing field concerned with extracting a set of features from images, which allows for the automatic classification of medical images into predefined groups. Therefore, we believe that the combination of CT image-based radiomics and cfDNA methylation could further improve diagnostic accuracy. Summary of the Invention
[0004] The purpose of the present invention is to provide a deep learning cascade CT lung nodule detection method and device, aiming to improve diagnostic accuracy and reduce detection costs.
[0005] To achieve the above objectives, in a first aspect, the present invention provides a deep learning cascade CT lung nodule detection method, comprising: acquiring slice data and generating a training set, a validation set, and a test set;
[0006] The image data in the training set was used to detect nodules using the YOLOX algorithm and segmented using the vgg16-unet model to obtain the segmentation results;
[0007] Extracting radiomic features of lung nodules based on the segmentation results;
[0008] Obtain peripheral blood cfDNA methylation characteristics;
[0009] The radiomic features of lung nodules and cfDNA methylation features were combined to obtain the final features.
[0010] Based on the final features, the random forest algorithm is used to build a diagnostic model for detection.
[0011] The specific method of obtaining slice data and generating training sets and test sets is:
[0012] Get the original data;
[0013] Filter the slice data and convert it into annotated image data;
[0014] After annotating the image data, a training set and a test set are generated.
[0015] Among them, the ratio of the training set, validation set and test set is preferably 7:1.5:1.5.
[0016] Among them, the image data in the training set is detected with the YOLOX algorithm for nodules and segmented using the vgg16-unet model. The specific steps to obtain the segmentation results are:
[0017] Based on the original image of the training set, the YOLOX module is used to detect the position of the target in the original image, mark it as the ROI area, and record the first position information of the area relative to the original image;
[0018] The ROI region is sent to the vgg16-unet model for segmentation to obtain multiple segmented ROI region results, and the ROI region image is represented by a dotted line;
[0019] Acquire second position information of the dotted line relative to the first position information;
[0020] The ROI region image and the original image are fused according to the first position information and the second position information to obtain a segmentation result.
[0021] The specific steps of extracting the radiomic features of pulmonary nodules based on the segmentation results are:
[0022] Based on the segmentation results, the radiomics feature extraction module is used to extract the features;
[0023] Normalize the extracted features;
[0024] After resampling the normalized extracted features, all radiomic features are extracted for each nodule;
[0025] Based on feature engineering, all the radiomic features are reduced in dimension to obtain the radiomic features of lung nodules.
[0026] Among them, the specific method of obtaining the methylation characteristics of peripheral blood cfDNA is: after obtaining all methylation characteristics by performing full methylation omics sequencing on cfDNA, the feature engineering is used to perform dimensionality reduction processing to obtain the methylation characteristics of peripheral blood cfDNA.
[0027] The feature engineering specifically includes: firstly processing by using a variance filtering method, and then processing by using an embedding method.
[0028] In a second aspect, the present invention also provides a deep learning cascade CT lung nodule detection device, comprising a data acquisition module, a segmentation module, an impactomics feature extraction module, a cfDNA methylation feature extraction module, a screening module, and a detection module;
[0029] The data acquisition module is used to acquire slice data and generate training sets, validation sets and test sets;
[0030] The segmentation module detects nodules using the YOLOX algorithm and performs segmentation using the vgg16-unet model on the image data in the training set to obtain segmentation results;
[0031] The impactomics feature extraction module is used to extract pulmonary nodule imaging features based on the segmentation results;
[0032] The cfDNA methylation feature extraction module is used to obtain the cfDNA methylation features of peripheral blood;
[0033] The screening module is used to jointly screen and process the pulmonary nodule imaging features and the cfDNA methylation features to obtain the final features;
[0034] The screening module is used to construct a diagnostic model for detection using a random forest algorithm based on the final features.
[0035] The present invention provides a deep learning cascade CT lung nodule detection method and device, comprising: acquiring slice data and generating a training set and a test set; detecting nodules using the YOLOX algorithm and segmenting the image data in the training set using the vgg16-unet model to obtain segmentation results; extracting lung nodule imaging features based on the segmentation results; acquiring peripheral blood cfDNA methylation features; jointly screening and processing the lung nodule imaging features and cfDNA methylation features to obtain final features; and constructing a diagnostic model based on the final features using a random forest algorithm for detection. The core algorithm model constructed using an artificial intelligence algorithm organically combines the high sensitivity of CT imaging features with the high specificity of cfDNA methylation features. Its diagnostic accuracy for benign and malignant lung nodules is superior to that of any single-omics algorithm model, particularly in detecting early lung cancer in lung nodules, and has great practical application value. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.
[0037] Figure 1 This is a flowchart of a deep learning cascade CT lung nodule detection method of the present invention.
[0038] Figure 2 This is a flowchart of the present invention for obtaining slice data and generating training sets and test sets.
[0039] Figure 3 This is a flowchart of the present invention that uses the YOLOX algorithm to detect nodules in the image data in the training set and uses the vgg16-unet model to segment the image data to obtain the segmentation results.
[0040] Figure 4 This is a flowchart of the present invention for extracting radiomic features of lung nodules based on the segmentation results.
[0041] Figure 5 This is a module diagram of a deep learning cascade CT lung nodule detection device of the present invention.
[0042] Figure 6 It is a cutting flow chart of the present invention.
[0043] Figure 7 This is the ROC curve result diagram of the accuracy of the random forest algorithm under the test set of the present invention.
[0044] Figure 8 This is the ROC curve result of the random forest algorithm accuracy under the independent validation set of the present invention. DETAILED DESCRIPTION
[0045] The following describes embodiments of the present invention in detail, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and are not to be construed as limiting the present invention.
[0046] A total of 127 cases of data were selected in this paper, of which 86 were malignant and 41 were benign.
[0047] See also Figures 1 to 6 In a first aspect, the present invention provides a deep learning cascade CT lung nodule detection method, comprising:
[0048] S101 obtains slice data and generates training set, validation set and test set;
[0049] The specific method is:
[0050] S201 obtains original data;
[0051] To ensure the authenticity and reliability of the training model, the dataset is the CT image diagnosis results of real patients provided by authoritative doctors.
[0052] S202 filters the slice data and converts it into annotated image data;
[0053] Medical CT images in the DICOM format contain patient information, such as name, gender, and age, as well as other image-related information, such as the equipment that captured and generated the image, and medical context. Privacy and invalid information is filtered out, retaining only the image information, and the original DICOM format is converted into annotated 2D image slices.
[0054] S203 generates a training set, a validation set, and a test set after annotating the image data.
[0055] Under the guidance of professional doctors at the hospital, we annotated all nodules involved in this project, including ground-glass nodules (malignant), solid nodules (malignant), and benign nodules, generating standard JSON files. Our artificial intelligence team then processed the data, converted the format, grayscaled it, and binarized it, generating paired PNG images of the original and training-test sets for training and testing. The preferred ratio of the training set, validation set, and test set was 7:1.5:1.5.
[0056] S102 uses the YOLOX algorithm to detect nodules in the image data in the training set and uses the vgg16-unet model to segment the image data to obtain the segmentation results;
[0057] The specific steps are:
[0058] S301 detects the position of the target in the original image using the YOLOX module based on the original image of the training set, marks it as a ROI region, and records the first position information of the region relative to the original image;
[0059] The YOLOX module detects the location of the target (i.e., the nodule) in the original image, marks it as a ROI, records its initial position relative to the original image, and crops the ROI image as input to the vgg16-unet module. The resulting ROI image is significantly smaller than the original image, and the target object's proportion of the entire image is significantly increased, facilitating further processing.
[0060] S302: sending the ROI region into the vgg16-unet model for segmentation to obtain multiple segmented ROI region results, wherein the ROI region image is represented by a dotted line;
[0061] Since the original image may contain multiple targets, each image may generate multiple ROI area images. After obtaining these images, they are sent to the vgg16-unet model for segmentation. Multiple segmented ROI area results can be obtained. The segmented edges are represented by dotted lines, and the second position information of the dotted line position relative to the ROI area is recorded.
[0062] S303: Acquire second position information of the dotted line relative to the first position information;
[0063] S304: The ROI region image and the original image are fused according to the first position information and the second position information to obtain a segmentation result.
[0064] Since this problem is actually a binary classification problem of image pixels, a black and white binary image is output as the segmentation result. Finally, the positions of all ROI region rectangles and all nodule boundary dashed lines in the original image are determined based on the first and second position information. That is, all segmented ROI region images and the original image are fed into the image fusion module for fusion to obtain the final segmentation result.
[0065] S103 extracts radiomic features of the lung nodules based on the segmentation results;
[0066] The specific steps are:
[0067] S401 uses an imaging omics feature extraction module to extract features based on the segmentation results to obtain extracted features;
[0068] The radiomics feature extraction module is pyradiomics, which can extract radiomics features from medical images. It supports feature extraction in both 2D and 3D. Feature extraction can be performed directly from the image obtained from the segmentation result.
[0069] S402 normalizes the extracted features;
[0070] S403 extracts all radiomics features for each nodule after resampling the normalized extracted features;
[0071] S404 performs dimensionality reduction on all the radiomic features based on feature engineering to obtain radiomic features of pulmonary nodules. The feature engineering specifically includes: firstly performing processing using a variance filtering method, and then performing processing using an embedding method.
[0072] The specific steps of the variance filtering method are: setting a variance threshold for a specified feature, calculating the variance value of the specified feature, and identifying a feature with a variance value greater than the variance threshold as a radiomics feature. This allows for a first dimensionality reduction process to reduce the number of features, and then further processing using the embedding method to further reduce the number of features.
[0073] The specific steps are: A regression model with mean error as the loss function is used to train the features after the initial dimensionality reduction process to obtain weight coefficients for each feature. A weight threshold is set, and when the weight coefficient exceeds the weight threshold, it is considered an influencing omics feature. This can reduce the originally acquired 1,800-dimensional data to 100 dimensions, thereby improving the subsequent running speed.
[0074] S104 obtains the methylation characteristics of peripheral blood cfDNA;
[0075] Specifically, after obtaining all methylation features through full methylome sequencing of cfDNA, feature engineering was used to reduce dimensionality to obtain peripheral blood cfDNA methylation features. Specifically, feature engineering involved first using variance filtering and then embedding to reduce the data from 698 dimensions to 93 dimensions.
[0076] S105 combines the radiomics features of lung nodules with the cfDNA methylation features to obtain the final features;
[0077] The cfDNA methylation signatures from peripheral blood and the radiomics features from lung nodules were combined to create a 193-dimensional feature set. The Boruta feature screening method was then used to filter these 193-dimensional features, ultimately reducing them to 15 dimensions. This method then selected highly representative features, providing a more effective basis for subsequent classification and discrimination. This method ultimately screened 15 combined methylation and imaging features for inclusion in the diagnostic model.
[0078] S106 uses the random forest algorithm to build a diagnostic model for detection based on the final features.
[0079] In order to build a diagnostic model for benign and malignant lung nodules, the data of the 127 patients after the above screening were divided into training set, test set and independent validation set in a ratio of 7:2:1. The following are the results of the test set and independent validation set respectively. A total of 4 classification algorithm models were constructed here. By using the training set to train each model, and then the trained models were used to classify the test set and independent validation set respectively, the results showed that the application of random forest algorithm (Random Forest) modeling combined with imaging genomics features and cfDNA methylation features for the diagnosis of benign and malignant lung nodules was the best. The final ROC curve results are shown in the figure below. Figure 7 、 Figure 8 shown.
[0080] In a second aspect, the present invention also provides a deep learning cascade CT lung nodule detection device, comprising a data acquisition module, a segmentation module, an impactomics feature extraction module, a cfDNA methylation feature extraction module, a screening module, and a detection module;
[0081] The data acquisition module is used to acquire slice data and generate training sets and test sets;
[0082] The segmentation module segments the image data in the training set using the YOLOX algorithm and the vgg16-unet model to obtain segmentation results;
[0083] The impactomics feature extraction module is used to extract pulmonary nodule imaging features based on the segmentation results;
[0084] The cfDNA methylation feature extraction module is used to obtain the cfDNA methylation features of peripheral blood;
[0085] The screening module is used to jointly screen and process the pulmonary nodule imaging features and the cfDNA methylation features to obtain the final features;
[0086] The screening module is used to construct a diagnostic model for detection using a random forest algorithm based on the final features.
[0087] In this multi-omics combined AI-powered lung nodule diagnosis system, CT radiomics automatically segments lung nodule margins, ensuring the accuracy of subsequent radiomics data extraction. This ensures the convenience of clinical application and the consistency and objectivity of large-scale data analysis. Four cfDNA methylation signatures and 11 CT imaging feature parameters were identified through the screening of cfDNA methylation and CT radiomics features. This feature screening method utilizes a minimal number of cfDNA methylation feature variables and combines them with CT imaging features for modeling, ensuring the accuracy of the diagnostic system while significantly reducing the cost of cfDNA methylation testing, facilitating clinical application and alleviating the medical burden.
[0088] This multi-omics combined artificial intelligence diagnosis system for lung nodules uses a unique artificial intelligence algorithm to construct a core algorithm model that organically combines the high sensitivity of CT imaging omics with the high specificity of cfDNA methylation characteristics. Its diagnostic accuracy for benign and malignant lung nodules is better than any single-omics algorithm model, especially in the detection rate of early lung cancer in lung nodules. It has unique advantages and great practical application value.
[0089] The above disclosure is only a preferred embodiment of the present invention, and certainly cannot be used to limit the scope of the rights of the present invention. Ordinary technicians in this field can understand that all or part of the processes of the above embodiment and equivalent changes made in accordance with the claims of the present invention are still within the scope of the invention.
Claims
1. A deep learning cascade CT lung nodule detection method, It is characterized in that Including: obtaining slice data and generating training sets, validation sets and test sets; The image data in the training set is used to detect nodules using the YOLOX algorithm and segmented using the vgg16-unet model to obtain the segmentation results; Extracting radiomic features of pulmonary nodules based on the segmentation results; Obtain peripheral blood cfDNA methylation characteristics; The radiomic features of lung nodules and cfDNA methylation features were jointly screened and processed to obtain the final features; Based on the final features, a random forest algorithm is used to build a diagnostic model for detection; The specific steps of using the YOLOX algorithm to detect nodules in the image data in the training set and using the vgg16-unet model to segment the image data to obtain the segmentation results are: Based on the original image of the training set, the YOLOX module is used to detect the position of the target in the original image, mark it as the ROI area, and record the first position information of the area relative to the original image; The ROI region is sent to the vgg16-unet model for segmentation to obtain multiple segmented ROI region results, and the ROI region image is represented by a dotted line; Acquire second position information of the dotted line relative to the first position information; According to the first position information and the second position information, the ROI area image and the original image are merged to obtain a segmentation result; The specific steps of extracting the radiomics features of pulmonary nodules based on the segmentation results are: Based on the segmentation results, the radiomics feature extraction module is used to extract the extracted features; Normalize the extracted features; After resampling the normalized extracted features, all radiomic features are extracted for each nodule; Based on feature engineering, all the radiomic features are reduced in dimension to obtain the radiomic features of lung nodules.
2. A deep learning cascade CT lung nodule detection method as claimed in claim 1, characterized in that: The specific method of obtaining slice data and generating training sets and test sets is: Get the original data; Filter the slice data and convert it into annotated image data; After annotating the image data, training sets and test sets are generated.
3. A deep learning cascade CT lung nodule detection method as claimed in claim 2, characterized in that: The ratio of the training set, validation set and test set is 7:1.5:1.
5.
4. A deep learning cascade CT lung nodule detection method as claimed in claim 3, characterized in that: The specific method of obtaining the methylation characteristics of peripheral blood cfDNA is: after obtaining all methylation characteristics by performing full methylome sequencing on cfDNA, the feature engineering is used to perform dimensionality reduction processing to obtain the methylation characteristics of peripheral blood cfDNA.
5. A deep learning cascade CT lung nodule detection method as claimed in claim 4, characterized in that: The feature engineering is specifically as follows: firstly, a variance filtering method is used for processing, and then an embedding method is used for processing.
