An artificial intelligence-based CT image dynamic 3D imageomics system for judging COVID-19 staging

Abstract

This invention provides an AI-based dynamic 3D radiomics system for CT images to determine the stage of COVID-19, belonging to the field of artificial intelligence. Experimental results show that the AI-based dynamic 3D radiomics system for CT images of this invention can achieve an accuracy rate of 90% in determining the stage of COVID-19 in 66 COVID-19 patients. For each stage prediction, the AUC was 0.965 (95% CI: 0.934, 0.997) for stage 1 (early stage), 0.958 (95% CI: 0.931, 0.984) for stage 2 (progression stage), 0.998 (95% CI: 0.994, 1.000) for stage 3 (peak stage), and 0.975 (95% CI: 0.956, 0.994) for stage 4 (absorption stage). Therefore, the AI-based dynamic 3D imaging omics system for CT images provided by this invention can effectively stage COVID-19 patients and can serve as a potential tool to help hospitals make reasonable resource allocations and formulate appropriate treatment plans, with broad application prospects.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence, specifically relating to an AI-based dynamic 3D radiomics system for determining the stage of COVID-19 in CT images. Background Technology

[0002] Diagnostic methods for COVID-19 include respiratory sample transcription polymerase chain reaction (RT-PCR) and chest imaging. RT-PCR has high specificity but low sensitivity, reportedly as low as 60-70%. In chest imaging, chest X-rays are of little value in early diagnosis, while CT scans can detect abnormalities before symptoms appear. Therefore, for suspected COVID-19 cases, chest CT scans are recommended for initial evaluation and follow-up.

[0003] Based on CT images, COVID-19 can be divided into four stages: early stage, progressive stage, peak stage, and absorption stage. ① Early chest manifestations are often atypical, with lesions appearing as thin, patchy ground-glass opacities (GGO), mostly localized and scattered in the middle and lower lung fields, mainly in the subpleural region. ② In the progressive stage, lesions are multiple, manifesting as GGO exudation, fusion, or consolidation, most commonly distributed in the middle and outer zones of both lung fields, and may be accompanied by a small amount of pleural effusion. ③ The peak stage (critical illness) corresponds to the late stage of the disease, with diffuse and widespread increases in lung density, known as "white lung." Lesions develop rapidly during this stage, increasing by more than 50% within 48 hours, making treatment difficult and resulting in a high mortality rate. ④ In the absorption stage, lesions shrink or are absorbed, and interstitial fibrosis may be observed in some cases.

[0004] Computer-aided diagnostic (CAD) systems are effective tools for achieving automated and rapid diagnosis. Developing an artificial intelligence system for determining the stage of COVID-19 is of great significance for the triage and staging of hospitalized COVID-19 patients in clinical practice. Summary of the Invention

[0005] The purpose of this invention is to provide an AI-based dynamic 3D radiomics system for CT images for determining the stage of COVID-19, as well as its construction method and applications.

[0006] This invention provides an artificial intelligence system for determining the stage of COVID-19, the artificial intelligence system comprising the following modules:

[0007] (a) Input Module

[0008] Enter the radiomics and clinical characteristics of COVID-19 patients in the input module;

[0009] (II) Phased Judgment Model Establishment Module

[0010] Using the radiomics and clinical features of COVID-19 patients in the training set as input, the model was trained using the machine learning classifier built into MATLAB to obtain the trained staging model.

[0011] (III) Staged Judgment and Result Output Module

[0012] Using the radiomics and clinical features of COVID-19 patients in the test set as input, the trained staging model is used to determine the stage and output the results.

[0013] Furthermore, the radiomics features are extracted using the following method: SCOAT-Net is trained and tested using CT images of COVID-19 patients in the training set, and SCOAT-Net is trained and tested using lesion segmentation models to establish a lung-lesion segmentation model; the lung-lesion segmentation model is used to segment the lungs and lesions using CT images of COVID-19 patients in the test set, and three-dimensional lung lesions are reconstructed based on the lung-lesion segmentation results; radiomics features are extracted based on the reconstructed three-dimensional lung lesions.

[0014] The clinical features include one or more of the following: age, sex, time of onset, and time of progression.

[0015] Furthermore, the parameter settings for establishing the lung and lesion segmentation model are as follows: based on the PyTorch framework, the stochastic gradient descent method is used to optimize the dice coefficient loss function; the network parameters are initialized using the Kaiming He method; the number of iterations for model training is set to 100, the initial learning rate is 0.01, and it is multiplied by 0.1 every 10 iterations; finally, the model that has undergone 100 training cycles is selected as the segmentation model.

[0016] Furthermore, the radiomics features and clinical features are obtained by the following method: after the extracted radiomics features and clinical features are selected using a feature selection method, they are sorted in descending order of weight, and the features with the highest weight are selected; the feature selection method is selected from RF-FS, Relief-F or LLCFS, preferably RF-FS.

[0017] Furthermore, the features ranked higher by weight are one or more of the 1st to 30th features, preferably the 1st to 30th features or the 1st to 17th features.

[0018] Furthermore, the radiomics features and clinical features include one or more of the following 30 features: time of onset, ratio of lesion to lung volume, texture eighth perspective (autocorrelation coefficient), time of progression, texture fifth perspective (percentage of travel), age, texture seventh perspective (percentage of travel), intensity (median), intensity (cross-entropy), intensity (skewness), intensity (homogeneity), texture eighth perspective (correlation information measurement 2), lung region area, texture fourth perspective (cluster prominence coefficient), texture fourth perspective (maximum probability), intensity (standard deviation), texture seventh perspective (normalized inverse difference moment), intensity (mean absolute deviation), texture fourth perspective (energy), intensity (variance), texture third perspective (cluster prominence coefficient), texture fourth perspective (variance), texture fourth perspective (difference entropy), texture fourth perspective (contrast), texture fifth perspective (travel length nonuniformity), texture ninth perspective (travel length nonuniformity), texture eighth perspective (travel length nonuniformity), texture ninth perspective (cluster prominence coefficient), texture seventh perspective (travel length nonuniformity), texture eighth perspective (contrast).

[0019] And / or, the machine learning classifier provided by MATLAB is selected from RF classifier, SVM classifier, or KNN classifier;

[0020] And / or, the COVID-19 stages are early, progressive, peak, or absorption stages.

[0021] Furthermore, the machine learning classifier built into MATLAB is an RF classifier, and the radiomics features and clinical features are the following 30 features: onset time, lesion-to-lung volume ratio, texture eighth perspective (autocorrelation coefficient), progression time, texture fifth perspective (percentage of travel), age, texture seventh perspective (percentage of travel), intensity (median), intensity (cross-entropy), intensity (skewness), intensity (uniformity), texture eighth perspective (correlation information measurement 2), lung region area, texture fourth perspective (cluster prominence coefficient), texture fourth perspective (maximum probability), intensity (standard deviation), texture seventh perspective (normalized inverse difference moment), intensity (mean absolute deviation), texture fourth perspective (energy), intensity (variance), texture third perspective (cluster prominence coefficient), texture fourth perspective (variance), texture fourth perspective (difference entropy), texture fourth perspective (contrast), texture fifth perspective (travel length nonuniformity), texture ninth perspective (travel length nonuniformity), texture eighth perspective (travel length nonuniformity), texture ninth perspective (cluster prominence coefficient), texture seventh perspective (travel length nonuniformity), and texture eighth perspective (contrast).

[0022] Furthermore, the machine learning classifier built into MATLAB is an SVM classifier, and the radiomics features and clinical features are the following 17 features: onset time, ratio of lesion to lung volume, texture eighth perspective (autocorrelation coefficient), progression time, texture fifth perspective (percentage of travel), age, texture seventh perspective (percentage of travel), intensity (median), intensity (cross-entropy), intensity (skewness), intensity (uniformity), texture eighth perspective (correlation information measurement 2), lung region area, texture fourth perspective (cluster prominence coefficient), texture fourth perspective (maximum probability), intensity (standard deviation), and texture seventh perspective (normalized inverse difference moment).

[0023] The present invention also provides the use of the above-described artificial intelligence system in the preparation of a device for determining the stage of COVID-19.

[0024] The present invention also provides a computer device for determining the stage of COVID-19, the computer device including a memory, a processor and a computer program stored in the memory and executable on the processor, the computer program being the aforementioned artificial intelligence system for determining the stage of COVID-19.

[0025] Experimental results show that the AI-based dynamic 3D radiomics system for CT images of this invention can accurately determine the stage of COVID-19 in 66 COVID-19 patients (mean age 57 ± 15 years; 35 of whom were female), achieving an accuracy rate of 90%. For each stage prediction, the AUC was 0.965 (95% CI: 0.934, 0.997) for stage 1 (early stage), 0.958 (95% CI: 0.931, 0.984) for stage 2 (progression stage), 0.998 (95% CI: 0.994, 1.000) for stage 3 (peak stage), and 0.975 (95% CI: 0.956, 0.994) for stage 4 (absorption stage). Therefore, the AI-based dynamic 3D imaging omics system for CT images provided by this invention can effectively stage COVID-19 patients and can serve as a potential tool to help hospitals allocate resources rationally and develop appropriate treatment plans, with broad application prospects.

[0026] Obviously, based on the above description of the present invention, and according to common technical knowledge and conventional methods in the field, various other modifications, substitutions or alterations can be made without departing from the basic technical concept of the present invention.

[0027] The following detailed embodiments further illustrate the above-described content of the present invention. However, this should not be construed as limiting the scope of the present invention to the following examples. All technologies implemented based on the above-described content of the present invention fall within the scope of the present invention. Attached Figure Description

[0028] Figure 1 Flowchart of an AI-based dynamic 3D radiomics analysis method. (A) Use an AI model to segment the lungs and lesions of COVID-19 patients; (B) Extract 3D-dynamic radiomics features of the three-dimensional lung lesions and combine them with clinical data to assess the staging of COVID-19 patients.

[0029] Figure 2 Four stages of COVID-19 are shown in CT images of four patients. (A) Transverse unenhanced image; (B) Coronal reconstructed image. The first row represents the early stage, the second row represents the progression stage, the third row represents the peak stage, and the last row represents the absorption stage.

[0030] Figure 3 The proposed artificial intelligence model provides segmentation results of lung lesions and corresponding 3D lung lesion reconstruction results.

[0031] Figure 4 The weights of each feature in the three feature selection methods. Dynamic 3D radiomics features include structural, shape, and intensity features, while clinical features include age, sex, onset time, and progression time. (A) RF-FS selection method; (B) Relief-F feature selection method; (C) LLCFS feature selection method.

[0032] Figure 5 (A) Relationship between COVID-19 stage and onset time (in days); (B) Relationship between COVID-19 stage and lesion / lung volume ratio.

[0033] Figure 6 Normalized confusion matrix and ROC curve for four-class classification. In the ROC curve, the red dot curve represents the area under the overall mean curve (AUC). (A) RF classifier; (B) SVM classifier.

[0034] Figure 7 Flowchart of inclusion and exclusion criteria for the study population in Part (3) of the data.

[0035] Figure 8 The three parts of data used by the three AI algorithms, and their relationship with... Figure 1 Among the relevant patients, the data for each part were divided into a 4:1 ratio.

[0036] Figure 9 The SCOAT-Net architecture is used for lung and lesion segmentation in CT images of COVID-19 patients. SCOAT-Net redesigns the Unet++ architecture and introduces an attention learning mechanism consisting of channel attention modules and spatial attention modules.

[0037] Figure 10 The three-dimensional lesions of COVID-19 patients were decomposed using nine different viewpoints (texture features were extracted from each view).

[0038] Figure 11 The three feature selection methods correspond to each machine learning classifier in the training set. (A) RF classifier, (B) SVM classifier, (C) KNN classifier. Detailed Implementation

[0039] The raw materials and equipment used in this invention are all known products, obtained by purchasing commercially available products.

[0040] 1. Data sources and preprocessing for COVID-19 patients

[0041] The COVID-19 patient data used in this embodiment of the invention is divided into three parts, which are respectively from:

[0042] Part (1) Data: 5,750 CT images of 170 COVID-19 patients were collected from two public datasets, with lung boundaries already annotated by doctors;

[0043] Part (2) Data: A total of 19 patients and 1117 CT images were collected from Xiangya Second Hospital;

[0044] Part (3) Data: Images of chest CT scans performed continuously throughout the treatment process from January 1, 2020 to March 9, 2020 were collected from a hospital for a total of 331 patients (including 1023 CT scans and clinical data).

[0045] The data from the above three parts were randomly divided into training and test sets in a 4:1 ratio.

[0046] The CT images in the above three data sets were all labeled by two radiologists with 14 and 31 years of experience, respectively.

[0047] Data preprocessing: All raw CT images were adjusted using a fixed lung window [-1200, 0] and normalized to the range [0, 255]. No scaling techniques were used; all CT images were the same size of 512 × 512 pixels.

[0048] 2. Basic information of the 331 patients in Part (3)

[0049] 1023 CT scans from 331 COVID-19 patients were included in the staging dataset. As shown in Table 1, the most common symptoms at presentation were cough (225 out of 331 [70%]) and fever (220 out of 331 [67%]). Table 2 shows that most laboratory results were normal, with a few elevated. C-reactive protein levels, D-dimer levels, and erythrocyte sedimentation rate were elevated in all four stages, peaking in stage three. In addition, lactate dehydrogenase levels and serum creatinine levels increased only in stage three. The onset time varied significantly among the four stages, at 4.4 ± 6.1, 11.0 ± 7.1, 15.4 ± 8.3, and 26.3 ± 12.0 days, respectively.

[0050] 3. Criteria for interpreting CT images in four stages

[0051] Based on CT images, COVID-19 can be divided into four stages: early stage, progressive stage, peak stage, and absorption stage. Figure 2 ① Early chest manifestations are often atypical, with lesions appearing as thin, patchy ground-glass opacities (GGO), mostly localized and scattered in the middle and lower lung fields, primarily subpleural. ② In the progressive stage, lesions are multiple, presenting as GGO exudation, fusion, or consolidation, most commonly distributed in the middle and outer zones of both lung fields, and may be accompanied by a small amount of pleural effusion. ③ The peak stage (critical illness) corresponds to the late stage of the disease, with diffuse and widespread increases in lung density, known as "white lung." Lesions develop rapidly during this stage, increasing by more than 50% within 48 hours, making treatment difficult and resulting in a high mortality rate. ④ In the absorption stage, lesions shrink or are absorbed, and interstitial fibrosis may be observed in some cases.

[0052] Example 1: A method for constructing an AI-based dynamic 3D radiomics system for determining COVID-19 staging based on CT images.

[0053] I. Construction Method

[0054] The method for constructing an AI-based dynamic 3D radiomics system for determining COVID-19 staging in this embodiment includes the following steps:

[0055] (I) Establishing a lung-lesion segmentation model

[0056] Two segmentation models of a spatial and channel-based coarse-fine attention network (SCOAT-Net) were trained using data from parts (1) and (2). SCOAT-Net is a novel U-Net++ architecture with channel-oriented and spatial attention modules to attract the network's self-attention learning, enabling successful segmentation of target regions at both the channel and pixel levels. Figure 9 ).

[0057] Parameter settings for the lung and lesion segmentation model: The lung and lesion segmentation model of this invention is based on the PyTorch framework and uses stochastic gradient descent (SGD) to optimize the dice coefficient loss function (DICE). Network parameters are initialized using the Kaiming He method. The number of training iterations is set to 100, with an initial learning rate of 0.01, multiplied by 0.1 every 10 iterations. The model that has undergone 100 training cycles is ultimately selected as the segmentation model.

[0058] First, the lung segmentation network SCOAT-Net was trained and tested using the data from part (1) under the above parameter settings. The lesion segmentation network SCOAT-Net was also trained and tested using the data from part (2) under the same parameter settings. Then, the trained lung and lesion segmentation networks were used to segment the lungs and lesions from part (3) of the data. Next, three-dimensional lung lesions were reconstructed based on the lung and lesion segmentation results. The segmentation code can be found at https: / / github.com / Phanzsx / SCOAT-Net.

[0059] The segmentation performance of the lung segmentation network and the lesion segmentation network on the data of parts (1) and (2) is shown in Table 7. The segmentation and reconstruction results are as follows: Figure 3 As shown.

[0060] (II) Feature Extraction and Feature Selection

[0061] 1. Feature Extraction

[0062] Radiomics feature extraction: First, common radiomics features (intensity features) are extracted based on the reconstructed 3D lung lesions; then, the 3D lung lesions are decomposed into 9 fixed-view slices. Figure 10 The common radiomics (texture features) were extracted from each slice; lung volume, lesion volume, and the ratio of lesion to lung volume were also included as shape features. In total, 314 conventional 3D radiomics features were extracted, including 3 shape features, 14 intensity features, and 297 texture features. In addition, considering that radiologists will consider lung changes when assessing staging, this invention also uses the changes in conventional 3D features from two adjacent CT scans as dynamic 3D radiomics features.

[0063] Clinical Feature Extraction: In addition to radiomics features, this invention also incorporates four clinical features, including age, gender, onset time, and progression time. Onset time (in days) represents the time since symptom onset, and progression time (in days) is the time interval between two adjacent CT scans. During the patient's first CT scan, the dynamic 3D features and progression time are set to 0.

[0064] A total of 632 features were extracted here, including 628 radiomics features and 4 clinical features (Table 6, column 1).

[0065] 2. Feature Selection

[0066] Regarding feature selection, this invention first uses three feature selection methods based on the training set: Random Forest (RF-FS), Relief-F, and Cluster Feature Selection Based on Local Learning (LLCFS). The code for these three feature selection methods can be called from FEATURE SELECTION TOOLBOX v6.2.

[0067] (III) Establishing a staged judgment model and selecting a classifier

[0068] The staging process is implemented in MATLAB software: After feature selection, the selected radiomics features and clinical features are used as inputs. Based on the training set, MATLAB's built-in machine learning classifiers are used to train the staging model. Here, MATLAB's built-in machine learning classifiers are one of the following: Random Forest (RF) classifier, Support Vector Machine (SVM) classifier, or K-Nearest Neighbors (KNN) classifier. The code for staging can be found at https: / / github.com / Phanzsx / Assess-the-COVID-19.

[0069] The effectiveness of the trained staging model in determining the stage of COVID-19 was evaluated based on the test set.

[0070] This invention employs six metrics to evaluate segmentation and staging performance. Segmentation performance is evaluated using the dice similarity coefficient (DSC) and intersection-over-union ratio (IOU). The classifier's performance is assessed using accuracy, sensitivity, specificity, and area under the receiver operating characteristic (AUC). Furthermore, t-tests are used to examine differences between independent groups, with a two-sided p-value <0.05 considered statistically significant.

[0071] II. Results Analysis

[0072] 1. Segmentation results and selected features

[0073] The segmentation performance of the lung segmentation network and the lesion segmentation network on the data of parts (1) and (2) is shown in Table 7. This invention utilizes the trained segmentation model to segment the lungs and lesions in part (3) of the data, then reconstructs the three-dimensional lung lesions, extracts radiomics features, and the segmentation and reconstruction results are shown in Table 7. Figure 3 As shown.

[0074] The weights of the 632 features are assigned as shown in Table 6 when using different feature selection methods, and then these features are sorted in descending order. For each classifier, the highest accuracy (n) is selected, where n is the number of features, and the accuracy is the result of ten-fold cross-validation on the training set. Figure 11 Finally, for each classifier performing separate classifications over four epochs, the optimal feature selection method is determined based on the highest accuracy, yielding the optimal number of features. This part regarding feature ranking and feature selection is implemented only on the training set.

[0075] Table 6 shows that each of the three feature selection methods can rank these features differently. Furthermore, this invention calculates the sum of weights for different types of features (…). Figure 4 The study found that dynamic 3D radiomics features and clinical features were important factors in the evaluation of the training set phase. The invention also listed the top 30 features for each feature selection method based on their weights, finding that some features (onset time, age, intensity features) were equally significant (Tables 3, 8-9). After comparing the three feature selection methods, the Random Forest (RF-FS) feature selection method was chosen.

[0076] After sorting these features, each classifier selects the top n (1≤n≤632) features and calculates the accuracy using tenfold cross-validation on the training set. Finally, as shown in Table 10, the optimal feature selection method with the optimal number of features is obtained for each classifier. The top 30 features of the RF-FS feature selection method are listed in Table 3. We found that the top two features are the onset time (Table 3) and the ratio of lesion volume to lung volume (…). Figure 5 The ratios of lesion volume to lung volume in the four stages were 2.2%±3.6%, 14.8%±14.1%, 46.0%±16.5%, and 7.8%±9.5%, respectively.

[0077] 2. COVID-19 staging results

[0078] Using selected radiomics and clinical features from two adjacent CT scans as input, a staging model was trained on the training set. The trained staging model was then used to evaluate its performance on the test set. The results are shown in Table 4. Furthermore, considering the poor accuracy of the KNN classifier, this invention only used RF and SVM classifiers for detailed analysis. More specifically, this invention calculated the accuracy, sensitivity, and specificity of four binary classification methods (Stage 1 / Stage 2-3-4, Stage 2 / Stage 1-3-4, Stage 3 / Stage 1-2-4, Stage 4 / Stage 1-2-3) on the test set, and also calculated the confusion matrix and ROC curve on the test set. The results are shown in Table 5. Figure 6 As shown.

[0079] From Table 5 and Figure 6 As can be seen, the RF and SVM classifiers achieved comparable performance in terms of accuracy, specificity, and AUC. However, the RF and SVM classifiers differed in sensitivity. Specifically, the RF classifier outperformed the SVM classifier during the absorption phase, while the SVM classifier outperformed the RF classifier in the early stages. Furthermore, this invention found that both classifiers achieved 100% sensitivity during the peak phase. Based on these results, the diagnostic efficiency of all classifiers was quite considerable. More specifically, the RF classifier was the most effective in terms of overall accuracy (90%). Compared to the RF classifier, the SVM classifier showed the highest sensitivity in early diagnosis.

[0080] Example 2: An AI-based dynamic 3D radiomics system for determining COVID-19 staging (using an RF classifier to establish a staging model).

[0081] The system used in this embodiment to determine the stage of COVID-19 includes the following three modules:

[0082] (a) Input Module

[0083] Input the radiomics features and clinical features of COVID-19 patients in the training set into the input module. The radiomics features and clinical features are the 30 features shown in Table 3.

[0084] The feature extraction method is the same as in Example 1.

[0085] (II) Phased Judgment Model Establishment Module

[0086] Using the radiomics features and clinical features from module (I) as input, the model is trained using the RF classifier, a machine learning classifier built into MATLAB, to obtain the trained staging model.

[0087] (III) Staged Judgment and Result Output Module

[0088] Using the radiomics and clinical features (30 features as shown in Table 3) of COVID-19 patients in the test set as input, the staging judgment model trained by module (II) is used to make staging judgment and output the judgment result.

[0089] Example 3: An AI-based dynamic 3D radiomics system for determining COVID-19 staging (using an SVM classifier to establish a staging model).

[0090] The system used in this embodiment to determine the stage of COVID-19 includes the following three modules:

[0091] (a) Input Module

[0092] Input the radiomics and clinical features of COVID-19 patients in the test set into the input module. The radiomics and clinical features are features 1 to 17 shown in Table 3.

[0093] The feature extraction method is the same as in Example 1.

[0094] (II) Phased Judgment Model Establishment Module

[0095] Using the radiomics features and clinical features from module (I) as input, the model is trained using the SVM classifier, a machine learning classifier built into MATLAB, to obtain the trained staging judgment model.

[0096] (III) Staged Judgment and Result Output Module

[0097] Using the radiomics and clinical features (features 1-17 in Table 3) of COVID-19 patients in the test set as input, the staging judgment model trained by module (II) is used to make staging judgment and output the judgment result.

[0098] Tables 1 through 10 are shown below:

[0099] Table 1: Clinical characteristics of patients

[0100]

[0101] Table 2: Patient Laboratory Results

[0102]

[0103] Note: The units for white blood cells, neutrophils, and lymphocytes are G / L; the units for C-reactive protein and D-dimer are mg / L; the units for alanine aminotransferase, aspartate aminotransferase, and lactate dehydrogenase are U / L; the units for blood urea nitrogen are mmol / L; the units for serum creatinine and serum uric acid are μmol / L; the units for erythrocyte sedimentation rate are mm / h; and the time from onset of illness is in days.

[0104] Table 3: Top 30 features ranked by weight using the Random Forest Feature Selection (RF-FS) method.

[0105]

[0106] Table 4: Determining the optimal feature selection method and number of features based on the accuracy of three machine learning classifiers on the training and test sets.

[0107]

[0108] Table 5: Performance metrics of RF and SVM classifiers on the test set

[0109]

[0110] Table 6: Detailed weights of dynamic 3D radiomics features and four clinical features in the three feature selection methods.

[0111]

[0112] Table 7: Segmentation performance of the lung and lesion segmentation dataset

[0113]

[0114] Note: DSC represents the dice similarity coefficient, and IOU represents the intersection-union ratio.

[0115] Table 8: Top 30 features by weight using the Relief-F feature selection method

[0116]

[0117] Table 9: Top 30 features by weight using the LLCFS feature selection method

[0118]

[0119] Table 10: The accuracy of three classifiers using 10x cross-validation on the training set determines the optimal feature selection method and number of features.

[0120]

[0121] In summary, this invention provides an artificial intelligence system for determining the stage of COVID-19, belonging to the field of artificial intelligence. Experimental results show that the AI-based dynamic 3D radiomics system for CT images of this invention achieves a 90% accuracy rate in determining the stage of COVID-19 in 66 COVID-19 patients. For each stage prediction, the AUC was 0.965 (95% CI: 0.934, 0.997) for stage 1 (early stage), 0.958 (95% CI: 0.931, 0.984) for stage 2 (progression stage), 0.998 (95% CI: 0.994, 1.000) for stage 3 (peak stage), and 0.975 (95% CI: 0.956, 0.994) for stage 4 (absorption stage). Therefore, the AI-based dynamic 3D imaging omics system for CT images provided by this invention can effectively stage COVID-19 patients and can serve as a potential tool to help hospitals allocate resources rationally and develop appropriate treatment plans, with broad application prospects.

Claims

1. An artificial intelligence system for determining the stage of COVID-19, characterized in that: The artificial intelligence system includes the following modules: (a) Input Module Enter the radiomics and clinical characteristics of COVID-19 patients in the input module; (II) Phased Judgment Model Establishment Module Using the radiomics and clinical features of COVID-19 patients in the training set as input, the model was trained using the machine learning classifier built into MATLAB to obtain the trained staging model. (III) Staged Judgment and Result Output Module Using the radiomics and clinical features of COVID-19 patients in the test set as input, the trained staging model is used to determine the stage and output the results. The radiomics features were extracted using the following method: SCOAT-Net was trained and tested using CT images of COVID-19 patients in the training set, and SCOAT-Net was also trained and tested to establish a lung-lesion segmentation model; the lung-lesion segmentation model was used to segment the lungs and lesions using CT images of COVID-19 patients in the test set, and three-dimensional lung lesions were reconstructed based on the lung-lesion segmentation results; radiomics features were extracted based on the reconstructed three-dimensional lung lesions. The clinical features include one or more of the following: age, sex, time of onset, and time of progression; The parameter settings for establishing the lung and lesion segmentation model are as follows: Based on the PyTorch framework, the dice coefficient loss function is optimized using the stochastic gradient descent method. The radiomics features and clinical features were obtained by the following method: the extracted radiomics features and clinical features were selected using a feature selection method, and then sorted in descending order of weight, selecting the features with the highest weight; the feature selection method was RF-FS.

2. The artificial intelligence system according to claim 1, characterized in that: The network parameters were initialized using the Kaiming initialization method; the number of iterations for model training was set to 100, the initial learning rate was 0.01, and it was multiplied by 0.1 every 10 iterations; finally, the model that had undergone 100 training cycles was selected as the segmentation model.

3. The artificial intelligence system according to claim 1, characterized in that: The features ranked higher by weight are one or more of the 1st to 30th features.

4. The artificial intelligence system according to claim 1, characterized in that: The COVID-19 stages are defined as early stage, progressive stage, peak stage, or absorption stage; The built-in machine learning classifier in MATLAB is the RF classifier. The radiomics features and clinical features are the following 30 features: onset time, lesion-to-lung volume ratio, texture eighth view - autocorrelation coefficient, progression time, texture fifth view - percentage of travel, age, texture seventh view - percentage of travel, intensity - median, intensity - cross-entropy, intensity - skewness, intensity - homogeneity, texture eighth view - correlation information measurement 2, lung region area, texture fourth view - cluster prominence coefficient, texture fourth view - maximum probability, intensity - standard deviation, texture seventh view - normalized inverse difference moment, intensity - mean absolute deviation, texture fourth view - energy, intensity - variance, texture third view - cluster prominence coefficient, texture fourth view - variance, texture fourth view - difference entropy, texture fourth view - contrast, texture fifth view - travel length non-uniformity, texture ninth view - travel length non-uniformity, texture eighth view - travel length non-uniformity, texture ninth view - cluster prominence coefficient, texture seventh view - travel length non-uniformity, and texture eighth view - contrast.

5. The artificial intelligence system according to claim 1, characterized in that: The COVID-19 stages are defined as early stage, progressive stage, peak stage, or absorption stage; The machine learning classifier built into MATLAB is an SVM classifier. The radiomics features and clinical features are the following 17 features: onset time, ratio of lesion to lung volume, texture eighth view - autocorrelation coefficient, progression time, texture fifth view - percentage of travel, age, texture seventh view - percentage of travel, intensity - median, intensity - cross-entropy, intensity - skewness, intensity - uniformity, texture eighth view - information measurement of correlation 2, lung region area, texture fourth view - cluster prominence coefficient, texture fourth view - maximum probability, intensity - standard deviation, and texture seventh view - normalized inverse difference moment.

6. Use of the artificial intelligence system according to any one of claims 1 to 5 in the preparation of a device for determining the stage of COVID-19.

7. A computer device for determining the stage of COVID-19, characterized in that: The computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is the artificial intelligence system for determining the stage of COVID-19 as described in any one of claims 1 to 5.

Images