A radiomics-based model and method for estimating the time of epidural hematoma formation

Abstract

This invention discloses a radiomics-based model and method for estimating the formation time of epidural hematoma, belonging to the fields of forensic medicine and imaging technology. By analyzing CT data of epidural hematomas within 12 hours, key features are selected using Lasso logistic regression, and an artificial intelligence algorithm is applied to construct an epidural hemorrhage time prediction model. Furthermore, to achieve more accurate prediction of epidural hematoma formation time, samples are divided into two time periods—before and after 5 hours—based on clinical research data and pathophysiological conditions of epidural hematomas. A random forest algorithm is applied to demonstrate radiomics differences between the two time periods. Separate epidural hematoma injury time estimation models are established for each of these two time periods, and the predictive performance is verified to be more accurate than that without differentiation based on pathophysiological conditions. This invention demonstrates the feasibility of radiomics technology in injury time estimation, and these results can provide a reference for forensic practice.

Description

Technical Field

[0001] This invention belongs to the fields of forensic medicine and imaging technology. It involves using CT images of human epidural hemorrhage, based on features screened by radiomics, using Lasso logistic regression to screen key features, and constructing a model to infer the formation time of epidural hematoma through weighted regression system. It has certain forensic application value. Background Technology

[0002] Injury time estimation refers to the use of natural science and technology to predict the time of formation of injuries to human tissues and organs, with its core content being the estimation of survival time after injury. Determining the injury time helps in identifying the cause of death, the manner of death, and the relationship between injury and illness, thereby narrowing down the scope of suspects, determining the time of the incident, assisting in scene reconstruction, determining the correlation between the injury and the case or event, and determining the nature of the case. Therefore, injury time estimation is an important scientific problem in forensic medicine. Blunt force causing intracranial hematoma is the most common and serious complication of brain injury, accounting for approximately 30% to 60% of traumatic brain injuries, while epidural hemorrhage is the most common type of intracranial hemorrhage. Epidural hematoma commonly occurs when blood accumulates between the skull and dura mater due to rupture of the middle meningeal artery, anterior meningeal artery, venous sinus, or rupture of the diploic vein caused by skull fractures, accounting for approximately 30% of traumatic brain hematomas. Based on the formation time of the hematoma, epidural hematomas are mainly divided into three types: acute, subacute, and chronic, with acute, subacute, and chronic hematomas accounting for approximately 86%, 10%, and 4%, respectively. In terms of location, epidural hematoma commonly occurs supratentorially, most frequently in the frontotemporal and temporoparietal regions, and relatively less frequently in the occipital region. CT is the preferred imaging modality for diagnosing epidural hematoma clinically; in the acute phase, epidural hematoma presents as a "biconvex lens" high-density shadow. Research on the estimation of the time of epidural hemorrhage injury has significant forensic value; currently, there are no reported studies on the formation time of epidural hematoma.

[0003] Previous studies have used molecular pathology techniques to infer the formation time of subdural hematomas by analyzing changes in the pathological morphology and specific cellular markers of hematomas and the dura mater. Other studies have used molecular biology techniques to predict hematoma formation time by detecting changes in the content of biomolecules within the subdural hematoma. However, these methods rely heavily on available tissue samples; furthermore, the estimated time window is large and the accuracy is poor, failing to meet the practical needs of forensic medicine. Therefore, it is particularly necessary to explore a non-invasive in vitro examination method for accurate estimation of injury time in forensic medicine. Using non-invasive imaging techniques such as X-rays, computed tomography (CT), and magnetic resonance imaging (MRI) to estimate injury time is more feasible.

[0004] Radiomics refers to the process of extracting a large number of quantitative features from digital images, then mining these features to generate or verify hypotheses. This technology can deeply mine image features and be used for precision medical diagnosis. Studies have reported that CT radiomics can effectively predict the risk of intracranial hematoma re-expansion. Therefore, we selected CT images, applied radiomics technology to screen image features, and used artificial intelligence algorithms such as deep learning to construct a model for estimating the time of forensic epidural hemorrhage, which is then applied to forensic practice.

[0005] In conclusion, the forensic inference model for the time of acute epidural hematoma hemorrhage based on CT dural radiomics technology is expected to provide new evidence for inferring the formation time of epidural hematoma and has important forensic significance. Summary of the Invention

[0006] To address the aforementioned problems, this invention provides a method for inferring the formation time of an epidural hematoma based on CT images of epidural hemorrhage, employing radiomics technology to screen omics features, and using artificial intelligence algorithms to construct an epidural hemorrhage time model, along with its forensic applications. This invention demonstrates the feasibility of radiomics technology in predicting the formation time of epidural hematoma, and uses multiple screened radiomics features to establish mathematical models and regression equations for accurate inference of the epidural hematoma formation time.

[0007] To achieve the above objectives, the present invention provides the following technical solution.

[0008] This invention provides a screening method for radiomics features used to identify the formation time of epidural hematoma, characterized by the following steps:

[0009] Step 1: Collect CT imaging data of epidural hematoma scanned within 12 hours using the same machine model, excluding CT artifacts and samples with other types of intracranial hematoma at the same location.

[0010] Step 2: Use radiomics software to delineate the region of interest in the epidural hematoma area, perform image filtering and transformation, and extract the corresponding radiomics features.

[0011] Step 3: Standardize the radiomics feature data of the samples obtained in Step 1, and use the Lasso package to calculate the time of sample damage formation. t With hour as the dependent variable and radiomics features as the independent variable, Lasso was applied for data dimensionality reduction and feature selection. The optimal parameter λ of the Lasso algorithm was obtained through 10-fold cross-validation, and finally, radiomics features with non-zero coefficients were selected.

[0012] Furthermore, the screening method described above identified 27 radiomics features for predicting the formation time of epidural hematoma within 0 to 12 hours, as shown in Table 1.

[0013] Furthermore, the screening method described above identified 37 radiomics features for determining the formation time of epidural hematoma in the two time periods before and after 5 hours, as shown in Table 2.

[0014] Furthermore, the screening method described above identified 13 radiomics features for predicting the formation time of epidural hematoma within 0 to 5 hours, as shown in Table 4.

[0015] Furthermore, the screening method described above identified 19 radiomics features for predicting the formation time of epidural hematoma within 6 to 12 hours, as shown in Table 5.

[0016] This invention also provides a method for establishing a radiomics-based model for predicting the formation time of epidural hematoma, characterized in that the model is used to predict the formation time of epidural hematoma within 0-12 hours, specifically including the following steps:

[0017] Step 1: Collect CT imaging data of epidural hematoma scanned within 12 hours using the same machine model, excluding CT artifacts and samples with other types of intracranial hematoma at the same location.

[0018] Step 2: Use radiomics software to delineate the region of interest in the epidural hematoma area, perform image filtering and transformation, and extract the corresponding radiomics features.

[0019] Step 3: Standardize the radiomics feature data of the samples obtained in Step 1, and use the Lasso package to calculate the time of sample damage formation. t Using (hours) as the dependent variable and radiomics features as the independent variable, Lasso was applied for data dimensionality reduction and feature selection. The optimal parameter λ for the Lasso algorithm was determined through 10-fold cross-validation, ultimately selecting 27 radiomics features with non-zero coefficients. These 27 radiomics features include:

[0020] The feature name is DISCRETIZED_HUSkewness, and the coefficient is -0.794956.

[0021] The feature name is DISCRETIZED_HISTO_Entropy_log10, and the coefficient is 0.137586;

[0022] The feature name is GLRLM_LRE, and the coefficient is 0.129044.

[0023] The feature name is GLZLM_SZE, and the coefficient is 0.547362.

[0024] The feature name is LoG-GLCM_Contrast, and the coefficient is -0.326405.

[0025] The feature name is LoG-GLCM_Correlation, and the coefficient is -0.028087.

[0026] The feature name is LoG-GLZLM_SZE, and the coefficient is -0.160379.

[0027] The feature name is SWT-LLL-GLZLM_LZE, and the coefficient is 0.528321.

[0028] The feature name is SWT-HLL-DISCRETIZED_Kurtosis, and the coefficient is -0.236168.

[0029] The feature name is SWT-HLL-GLZLM_SZE, and the coefficient is 0.226367.

[0030] The feature name is SWT-HLL-GLZLM_LZLGE, and the coefficient is 0.555236.

[0031] The feature name is SWT-LHL-CONVENTIONAL_mean, and the coefficient is 0.233665.

[0032] The feature name is SWT-LHL-DISCRETIZED_Kurtosis, and the coefficient is 0.349986.

[0033] The feature name is SWT-LHL-GLCM_Correlation, and the coefficient is 0.030686.

[0034] The feature name is SWT-LLH-GLRLM_LRHGE, and the coefficient is 0.304777.

[0035] The feature name is SWT-LLH-NGLDM_Busyness, and the coefficient is -0.408665.

[0036] The feature name is SWT-LLH-GLZLM_LZE, and the coefficient is -0.141548.

[0037] The feature name is SWT-HLH-CONVENTIONAL_Q2, and the coefficient is 0.059320.

[0038] The feature name is SWT-HLH-DISCRETIZED_Kurtosis, and the coefficient is -0.625426.

[0039] The feature name is SWT-HLH-GLCM_Contrast, and the coefficient is 0.145714.

[0040] The feature name is SWT-HLH-GLZLM_SZE, and the coefficient is 0.111274.

[0041] The feature name is SWT-LHH-CONVENTIONAL_Kurtosis, and the coefficient is -0.089323.

[0042] The feature name is SWT-LHH-GLRLM_LRLGE, and the coefficient is 0.130523.

[0043] The feature name is SWT-LHH-NGLDM_Coarseness, and the coefficient is -0.199430.

[0044] The feature name is SWT-LHH-NGLDM_Busyness, and the coefficient is 0.502334.

[0045] The feature name is SWT-HHH-CONVENTIONAL_mean, and the coefficient is 0.161991.

[0046] The feature name is SWT-HHH-GLZLM_SZLGE, and the coefficient is 0.443160.

[0047] Step 4: Calculate the radiomics score for each sample based on the 27 radiomics feature coefficients and intercepts selected in Step 3. ;

[0048] Step 5: Using the OLS package, apply ordinary least squares to establish a regression model on the training set between lesion formation time and the radiomics score obtained in Step 4, and derive the ordinary least squares regression equation:

[0049] ;

[0050] Step 6: Substitute the test set into the ordinary least squares regression equation from Step 5, apply the mean absolute error (MAE) to evaluate the accuracy of the model's predictions, and perform analysis.

[0051] This invention also provides a method for establishing a radiomics-based model for predicting the formation time of epidural hematoma, characterized in that the model is used to distinguish the formation time of epidural hematoma in two time periods before and after 5 hours, specifically including the following steps:

[0052] Step 1: Collect CT imaging data of epidural hematoma scanned within 12 hours using the same machine model, excluding CT artifacts and samples with other types of intracranial hematoma at the same location.

[0053] Step 2: Use radiomics software to delineate the region of interest in the epidural hematoma area, perform image filtering and transformation, and extract the corresponding radiomics features.

[0054] Step 3: Standardize the radiomics feature data of the samples obtained in Step 1, and use the Lasso package to calculate the time of sample damage formation. t Using (hours) as the dependent variable and radiomics features as the independent variable, Lasso was applied for data dimensionality reduction and feature selection. The optimal parameter λ for the Lasso algorithm was determined through 10-fold cross-validation, ultimately selecting 37 radiomics features with non-zero coefficients. These 37 radiomics features include:

[0055] Feature name: DISCRETIZED_HUStd;

[0056] Feature name: DISCRETIZED_HUSkewness;

[0057] Feature name: GLRLM_LRE;

[0058] Feature name: GLZLM_SZE;

[0059] Feature name: LoG-GLRLM_LRE;

[0060] Feature name: SWT-LLL-GLZLM_SZE;

[0061] Feature name: SWT-LLL-GLZLM_LZE;

[0062] Feature name: SWT-HLL-GLZLM_LZLGE;

[0063] Characteristic name: SWT-LHL-CONVENTIONAL_Kurtosis;

[0064] Feature name: SWT-LHL-DISCRETIZED_Kurtosis;

[0065] Feature name: SWT-LHL-GLCM_Correlation;

[0066] Feature name: SWT-LHL-GLRLM_LRE;

[0067] Feature name: SWT-LHL-GLRLM_SRLGE;

[0068] Feature name: SWT-LHL-GLRLM_SRHGE;

[0069] Feature name: SWT-HHL-CONVENTIONAL_max;

[0070] Characteristic name: SWT-HHL-CONVENTIONAL_Kurtosis;

[0071] Feature name: SWT-HHL-GLCM_Contrast;

[0072] Feature name: SWT-LLH-DISCRETIZED_HISTO_Energy;

[0073] Feature name: SWT-LLH-GLRLM_LRHGE

[0074] Feature name: SWT-LLH-NGLDM_Busyness;

[0075] Characteristic name: SWT-HLH-CONVENTIONAL_Kurtosis;

[0076] Feature name: SWT-HLH-DISCRETIZED_Kurtosis;

[0077] Feature name: SWT-HLH-GLCM_Correlation;

[0078] Feature name: SWT-HLH-GLRLM_LGRE;

[0079] Feature name: SWT-HLH-NGLDM_Busyness;

[0080] Feature name: SWT-HLH-GLZLM_SZE;

[0081] Feature name: SWT-HLH-GLZLM_LGZE;

[0082] Feature name: SWT-LHH-CONVENTIONAL_Kurtosis;

[0083] Feature name: SWT-LHH-GLCM_Energy;

[0084] Feature name: SWT-LHH-NGLDM_Coarseness;

[0085] Feature name: SWT-LHH-NGLDM_Busyness;

[0086] Feature name: SWT-LHH-GLZLM_LZE;

[0087] Feature name: SWT-HHH-CONVENTIONAL_mean;

[0088] Feature name: SWT-HHH-CONVENTIONAL_Kurtosis;

[0089] Feature name: SWT-HHH-GLRLM_RLNU;

[0090] Feature name: SWT-HHH-GLZLM_SZLGE;

[0091] Feature name: SWT-HHH-GLZLM_SZHGE;

[0092] Step 4 and Step 3 use the selected features to construct a discriminant model for the formation time of epidural hematoma within a 5-hour timeframe using the random forest algorithm.

[0093] This invention also provides a method for establishing a radiomics-based model for predicting the formation time of epidural hematoma, characterized in that the model is used to predict the formation time of epidural hematoma within 0-5 hours, specifically including the following steps:

[0094] Step 1: Collect CT imaging data of epidural hematoma scanned within 12 hours using the same machine model, excluding CT artifacts and samples with other types of intracranial hematoma at the same location.

[0095] Step 2: Use radiomics software to delineate the region of interest in the epidural hematoma area, perform image filtering and transformation, and extract the corresponding radiomics features.

[0096] Step 3: Standardize the radiomics feature data of the samples obtained in Step 1, and use the Lasso package to calculate the time of sample damage formation. t Using (hours) as the dependent variable and radiomics features as the independent variable, Lasso was applied for data dimensionality reduction and feature selection. The optimal parameter λ for the Lasso algorithm was determined through 10-fold cross-validation, ultimately selecting 13 radiomics features with non-zero coefficients. These 13 radiomics features include:

[0097] The feature name is GLRLM_LRE, and the coefficient is 0.183608.

[0098] The feature name is LoG-CONVENTIONAL_std, and the coefficient is -0.059375.

[0099] The feature name is LoG-CONVENTIONAL_Q1, and the coefficient is 0.096689.

[0100] The feature name is LoG-GLCM_Contrast, and the coefficient is -0.381450.

[0101] The feature name is SWT-HLL-GLZLM_SZLGE, and the coefficient is 0.269235.

[0102] The feature name is SWT-LHL-CONVENTIONAL_Q2, and the coefficient is 0.253596.

[0103] The feature name is SWT-LHL-CONVENTIONAL_Skewness, and the coefficient is 0.068589.

[0104] The feature name is SWT-LHL-GLRLM_RLNU, and the coefficient is -0.040628.

[0105] The feature name is SWT-HHL-GLZLM_ZLNU, and the coefficient is 0.087925.

[0106] The feature name is SWT-LLH-GLZLM_LZE, and the coefficient is -0.066777.

[0107] The feature name is SWT-LLH-GLZLM_ZLNU, and the coefficient is 0.011010.

[0108] The feature name is SWT-HLH-GLRLM_LRLGE, and the coefficient is -0.145954.

[0109] The feature name is SWT-HLH-GLZLM_LGZE, and the coefficient is -0.203099.

[0110] Step 4: Calculate the radiomics score for each sample based on the 13 radiomics feature coefficients and intercepts selected in Step 3. ;

[0111] Step 5: Using the OLS package, apply ordinary least squares to establish a regression model on the training set between lesion formation time and the radiomics score obtained in Step 4, and derive the ordinary least squares regression equation: ;

[0112] Step 6: Substitute the test set into the ordinary least squares regression equation from Step 5, apply the mean absolute error (MAE) to evaluate the accuracy of the model's predictions, and perform analysis.

[0113] This invention also provides a method for establishing a radiomics-based model for predicting the formation time of epidural hematoma, characterized in that the model is used to predict the formation time of epidural hematoma within 6 to 12 hours, and specifically includes the following steps:

[0114] Step 1: Collect CT imaging data of epidural hematoma scanned within 12 hours using the same machine model, excluding CT artifacts and samples with other types of intracranial hematoma at the same location.

[0115] Step 2: Use radiomics software to delineate the region of interest in the epidural hematoma area, perform image filtering and transformation, and extract the corresponding radiomics features.

[0116] Step 3: Standardize the radiomics feature data of the samples obtained in Step 1, and use the Lasso package to calculate the time of sample damage formation. t Using (hours) as the dependent variable and radiomics features as the independent variable, Lasso was applied for data dimensionality reduction and feature selection. The optimal parameter λ for the Lasso algorithm was determined through 10-fold cross-validation, ultimately selecting 19 radiomics features with non-zero coefficients. These 19 radiomics features include:

[0117] The feature name is GLZLM_SZE, and the coefficient is 0.110848.

[0118] The feature name is LoG-CONVENTIONAL_Q1, and the coefficient is 0.432929.

[0119] The feature name is LoG-GLCM_Contrast, and the coefficient is -0.195671.

[0120] The feature name is LoG-GLZLM_SZE, and the coefficient is -0.306248.

[0121] The feature name is SWT-HLL-DISCRETIZED_Kurtosis, and the coefficient is -0.319317.

[0122] The feature name is SWT-HLL-GLRLM_LRHGE, and the coefficient is -0.076523.

[0123] The feature name is SWT-HLL-GLZLM_SZE, and the coefficient is 0.305857.

[0124] The feature name is SWT-HLL-GLZLM_HGZE, and the coefficient is 0.190681.

[0125] The feature name is SWT-HLL-GLZLM_LZHGE, and the coefficient is 0.157684.

[0126] The feature name is SWT-LHL-CONVENTIONAL_Q2, and the coefficient is 0.016221.

[0127] The feature name is SWT-LHL-GLRLM_LRHGE, and the coefficient is -0.345238.

[0128] The feature name is SWT-LHL-GLZLM_SZLGE, and the coefficient is -0.126023.

[0129] The feature name is SWT-HHL-GLRLM_RLNU, and the coefficient is -0.326098.

[0130] The feature name is SWT-HHL-GLZLM_SZHGE, and the coefficient is -0.119980.

[0131] The feature name is SWT-LLH-CONVENTIONAL_Q1, and the coefficient is 0.147801.

[0132] The feature name is SWT-HLH-GLCM_Correlation, and the coefficient is -0.269171.

[0133] The feature name is SWT-HLH-GLZLM_GLNU, and the coefficient is 0.106151.

[0134] The feature name is SWT-LHH-GLZLM_SZE, and the coefficient is 0.468498.

[0135] The feature name is SWT-HHH-CONVENTIONAL_Skewness, and the coefficient is 0.015185.

[0136] Step 4: Calculate the radiomics score for each sample based on the 19 radiomics feature coefficients and intercepts selected in Step 3. ;

[0137] Step 5: Using the OLS package, apply ordinary least squares to establish a regression model on the training set between lesion formation time and the radiomics score obtained in Step 4, and derive the ordinary least squares regression equation: ;

[0138] Step 6: Substitute the test set into the ordinary least squares regression equation from Step 5, apply the mean absolute error (MAE) to evaluate the accuracy of the model's predictions, and perform analysis.

[0139] The present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that the program is executed by a processor to implement a model constructed by the screening method for identifying radiomics features for epidural hematoma formation time as described in any one of claims 1-5 or the method for establishing a radiomics-based epidural hematoma formation time prediction model as described in any one of claims 6-9.

[0140] The beneficial effects of the present invention compared with the prior art.

[0141] (1) This invention provides a method for inferring the formation time of epidural hematoma based on radiomics features by collecting CT images of patients with epidural hematoma within 12 hours of injury, which is conducive to the transformation of research results.

[0142] (2) The present invention provides radiomics features for judging CT images of epidural hematoma within 5 hours and 6-12 hours. The 37 radiomics features have the characteristic of high model prediction accuracy on the random forest machine learning algorithm.

[0143] (3) This invention provides two linear regression equations based on radiomics features to infer the formation time of epidural hematoma in two time periods 5 hours before and after the formation time of epidural hematoma, which can realize the calculation of the formation time of epidural hematoma and is expected to provide more favorable assistance for the inference of injury time in forensic practice. Attached Figure Description

[0144] Figure 1 Radiomics workflow diagram.

[0145] Figure 2 Lasso regression with 10-fold cross-validation was used to screen features from the training set samples 0-12 hours post-injury.

[0146] Figure 3 Linear fitting of scatter plots of training set samples from 0 to 12 hours post-injury.

[0147] Figure 4 Validation results of the test set samples from 0 to 12 hours post-injury.

[0148] Figure 5 Radiomic features used to distinguish between the 0-5h and 6-12h time periods were selected by Lasso regression with 10-fold cross-validation.

[0149] Figure 6 The ROC curve of a random forest classification model, constructed using 37 radiomics features selected by Lasso regression, was validated on the test set.

[0150] Figure 7 Features were selected from the training set samples 0-5 hours after injury using Lasso regression with 10-fold cross-validation.

[0151] Figure 8 Validation results of the test set samples from 0 to 5 hours post-injury.

[0152] Figure 9 Features were selected from the training set samples 6-12 hours after injury using Lasso regression with 10-fold cross-validation.

[0153] Figure 10 Validation results of the test set samples 6-12 hours after injury. Detailed Implementation

[0154] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. The following examples illustrate the present invention in detail, but are not intended to limit the scope of the invention. These embodiments employ conventional experimental techniques, which are familiar to those skilled in the art and can be performed according to the instructions provided by the material manufacturers.

[0155] 1. Research subjects and grouping.

[0156] Study Subjects: To ensure no machine-related interference, this invention selected CT data of patients with epidural hematoma scanned using a Philips Brilliance ICT (256) scanner. Specific scanning parameters were: matrix 512*512, slice thickness 5mm, voltage 120KV, window level 40, and window width 80. A standard head scan was used, from the base of the skull to the top. CT artifacts and samples with other types of intracranial hematoma besides epidural hematoma were excluded. A total of 95 CT images of patients were collected from 0 to 12 hours post-injury, along with corresponding clinical baseline information. The data were randomly divided into a training set and a test set at a 7:3 ratio. X-rays were then used to analyze the data. 2 Fisher's test was used to analyze the clinical baseline information of patients in the training and test sets. The results showed that there was no statistically significant difference in clinical information between the two groups, indicating that the subsequent analysis would not be affected by the imbalance of clinical baseline information.

[0157] 2. Radiomics feature extraction and preliminary screening.

[0158] LIFEx was used to delineate regions of interest (ROIs) in the collected CT data for epidural hematoma areas. CT HU values ​​and intra-hematoma imaging CT values ​​showed no correlation with the time of epidural hematoma formation. Radiomic features of the ROI regions were extracted and categorized into five main types: statistical features, first-order histogram features, shape features, second-order texture features, and filtering features. The filtering features included Laplacian of Gaussian (LOG) filtering and Stationary Wavelet Transform (SWT). A total of 535 features were collected. The intraclass correlation coefficient (ICC) is commonly used to evaluate the results of different observers on the same quantitative observation. In this invention, 30 images were randomly selected, and another physician with radiological experience delineated the ROI of the epidural hematoma and extracted features. The interstitial coefficient (ICC) was calculated using the features extracted by the two observers for the same hematoma. Generally, an ICC above 0.75 is considered to indicate good reliability of the feature. Therefore, this invention initially included 503 features after removing features with an ICC below 0.75. The Pearson coefficient was used to assess the correlation between different features. Redundant features with a coefficient greater than 0.9 were removed to further optimize the number of features. Finally, 269 features were included in subsequent studies.

[0159] 3. Establish a regression model for 0-12 hours.

[0160] Damage formation time of the training set ( t With hours as the dependent variable and 269 radiomics features as independent variables, Lasso was used for data dimensionality reduction and feature selection. During feature selection, 10-fold cross-validation was selected to determine the optimal λ of the model. Figure 2 To plot the relationship between mean squared error and λ, a vertical dashed line is drawn using the minimum standard deviation and one standard deviation of the minimum standard to represent the optimal value. The optimal parameter λ is 0.281. Based on the optimal λ, Lasso coefficient curves for 269 features are plotted. Finally, 27 features with non-zero coefficients are selected. The specific feature names and corresponding coefficients are shown in Table 1.

[0161] Table 1. Names and coefficients of 27 radiomics features

[0162] .

[0163] The Radscore for each sample is calculated using the following formula:

[0164] .

[0165] Figure 3The Radscore scatter plot for the training set samples shows the solid line as the fitted curve and the shaded area as the 95% confidence interval. A significant linear correlation can be observed between radiomics scores and damage formation time (in hours). Using the ols package, an ordinary least squares regression equation is established for the radiomics scores and damage formation time of the training set samples. The specific formula is as follows:

[0166] .

[0167] Substituting the test set samples into the above regression equation, the MAE was applied to evaluate the accuracy of the model prediction, and the MAE of the test set was found to be 2.43h. Figure 4 This is a scatter plot of the validation results for the test set samples, with the black line representing the diagonal of y = x.

[0168] 4. Distinguish between radiomics differences in the 0-5h and 6-12h time periods.

[0169] According to previous literature, the volume of epidural hematoma no longer changes significantly 5 hours after trauma, and the aforementioned mean expiratory epidural hematoma (MAE) is not accurate enough. To achieve more accurate prediction of epidural hematoma travel time, based on this pathophysiological phenomenon, we divided the training and test sets into two groups according to two time periods: 0–5 h and 6–12 h. In the training set, we used the 0–5 h and 6–12 h time periods as dependent variables and various radiomics features as independent variables, applying Lasso for data dimensionality reduction and feature selection. For feature selection, we used 10-fold cross-validation to determine the optimal parameter λ for the Lasso model. Figure 5 The relationship between mean squared error and λ, and the coefficient curves are shown in Table 2. The optimal λ is 0.036. Thirty-seven features with non-zero coefficients were selected. The specific feature names are listed in Table 2. Using the 37 features selected by Lasso, a discriminant model for the formation time of epidural hematoma within a 5-hour time interval was built using the random forest algorithm. The model was then validated using a test set. To avoid the influence of different parameter settings, the model was built and validated using default parameters. The results show that the model has an AUC of 0.793 on the test set, and 22 out of 29 samples on the validation set can be accurately divided into two time intervals. The prediction accuracy of this classifier model can reach 75.86%. Figure 6 As shown in Table 2, this indicates that there are radiomics differences between the two time periods.

[0170] Table 2. Names of 37 radiomics features

[0171] .

[0172] Table 3. Prediction results of the classification model based on 37 image features for grouping test set samples 5 hours before and after.

[0173] .

[0174] 5. Establish regression models for the two time periods before and after 5 hours.

[0175] The above results indicate that there are differences in radiomics characteristics of epidural hematoma before and after 5 hours, suggesting that the different phases of epidural hematoma expansion and stabilization should be considered when establishing an injury time inference model. Therefore, this invention will further screen features that play an important role in inferring the formation time of epidural hematoma at different time periods and construct a mathematical regression model based on these features for practical application.

[0176] For the 0-5 hour timeframe, each time point in the training set was used as the dependent variable, and various radiomics features were used as independent variables. Lasso was applied for data dimensionality reduction and feature selection. During feature selection, 10-fold cross-validation was chosen to determine the optimal parameter λ for the Lasso model. Figure 7 To plot the relationship between mean squared error and λ, a vertical dashed line is drawn using the minimum standard and one standard deviation of the minimum standard to obtain the optimal value. The optimal parameter λ is 0.193. Lasso coefficient curves for 269 features are plotted. Finally, 13 non-zero coefficient features are selected based on the λ value, and the radiomics score (Radscore) for each sample is calculated. The specific feature names and coefficients are shown in Table 4. Figure 8 To validate the results on the test set, MAE=0.96h, proving that the regression model has good predictive performance. The regression formula obtained from the training set is as follows:

[0177] .

[0178] Table 4. Names, coefficients, and intercepts of the 13 non-zero features used to calculate radiomics scores.

[0179] .

[0180] For the 6-12 hour time period, the method was the same as above, and a total of 19 features with non-zero coefficients were selected. Table 5 shows the specific feature names and coefficients. The model validation results are shown in [Table 5]. Figure 10 The MAE is 1.18h. The regression formula is as follows:

[0181] .

[0182] Table 5. Names, coefficients, and intercepts of the 19 non-zero features used to calculate radiomics scores.

[0183] .

[0184] By performing radiomics analysis on CT data of epidural hematoma within 12 hours, this invention demonstrates the feasibility of radiomics technology in injury time estimation, further screening important imaging features and verifying their accuracy. Simultaneously, this invention considers the pathophysiology of hematoma to establish injury time estimation models for different stages, proving that the models have more accurate predictive capabilities. These results can provide a reference for forensic practice and lay the foundation for further translation.

Claims

1. A method for constructing a predictive model for epidural hematoma formation time based on radiomics, characterized in that, Includes the following steps: Step 1: Collect epidural hematoma images scanned by the same type of CT scanner within 12 hours after the injury, and exclude samples with artifacts or accompanied by other types of intracranial hematoma. Step 2: Use radiomics software to delineate the region of interest in the epidural hematoma area, perform image filtering and transformation, and extract the corresponding radiomics features. Step 3: Standardize the radiomics feature data of the samples obtained in Step 1, and use the Lasso package with the time t (hours) of sample damage formation as the dependent variable and the radiomics feature as the independent variable. Apply Lasso to perform data dimensionality reduction and feature selection. The optimal parameter λ of the Lasso algorithm is obtained through 10-fold cross-validation. Finally, radiomics features with non-zero coefficients are selected. Step 4: Calculate the radiomics score for each sample based on the radiomics feature coefficients and intercepts selected in Step 3. Step 5: Using the OLS package, apply ordinary least squares to build a predictive model of damage formation time and the radiomics score described in Step 4 on the training set: The prediction model for the 0-12 hour all-time period is based on Lasso regression and 10-fold cross-validation. 27 key features are selected, and after calculating the Radscore, the first linear prediction model is established using ordinary least squares: Y_Time = 1.1511 * Radscore - 2.8189; Predictive model for subdivided time periods of injury ≤ 5 hours: Based on Lasso regression and 10-fold cross-validation, 13 key features were selected. After calculating the Radscore, a second linear prediction model was established using ordinary least squares: Y_Time = 1.4617 * Radscore - 1.1542; Predictive models for the 6-12 hour sub-period of injury time: Based on Lasso regression and 10-fold cross-validation, 19 key features were selected. After calculating the Radscore, a third linear prediction model was established using ordinary least squares: Y_Time = 1.4543 * Radscore - 3.7944; Y_Time is the predicted time of epidural hematoma formation, and Radscore is the radiomics score. Step 6: Input the test set into the prediction model from Step 5, apply the Mean Absolute Error (MAE) to evaluate the accuracy of the model's predictions, and perform analysis.

2. The construction method according to claim 1, characterized in that, The steps also include: establishing a discriminant model for the formation time of epidural hematoma within a 5-hour timeframe based on 37 radiomics features using a random forest algorithm.

3. The construction method according to claim 2, characterized in that, The 37 radiomics features include: Feature name DISCRETIZED_HUStd; Feature name DISCRETIZED_HUSkewness; Feature name GLRLM_LRE; Feature name GLZLM_SZE; Feature name LoG-GLRLM_LRE; Feature name SWT-LLL-GLZLM_SZE; Feature name SWT-LLL-GLZLM_LZE; Feature name SWT-HLL-GLZLM_LZLGE; Feature name SWT-LHL-CONVENTIONAL_Kurtosis; Feature name SWT-LHL-DISCRETIZED_Kurtosis; Feature name SWT-LHL -GLCM_Correlation; Feature name SWT-LHL-GLRLM_LRE; Feature name SWT-LHL-GLRLM_SRLGE; Feature name SWT-LHL-GLRLM_SRHGE; Feature name SWT-HHL-CONVENTIONAL_max; Feature name SWT-HHL-CONVENTIONAL_Kurtosis; Feature name SWT-HHL-GLCM_Contrast; Feature name SWT-LLH-DISCRETIZED_HISTO_Energy; Feature name SWT-LLH-GLRLM_LRHGE; Feature name S SWT-HLH-NGLDM_Busyness; SWT-HLH-CONVENTIONAL_Kurtosis; SWT-HLH-DISCRETIZED_Kurtosis; SWT-HLH-GLCM_Correlation; SWT-HLH-GLRLM_LGRE; SWT-HLH-NGLDM_Busyness; SWT-HLH-GLZLM_SZE; SWT-HLH-GLZLM_LGZE; SWT-LHH-CONVENTIONAL_Kurt osis; Feature name SWT-LHH-GLCM_Energy; Feature name SWT-LHH-NGLDM_Coarseness; Feature name SWT-LHH-NGLDM_Busyness; Feature name SWT-LHH-GLZLM_LZE; Feature name SWT-HHH-CONVENTIONAL_mean; Feature name SWT-HHH-CONVENTIONAL_Kurtosis; Feature name SWT-HHH-GLRLM_RLNU; Feature name SWT-HHH-GLZLM_SZLGE; Feature name SWT-HHH-GLZLM_SZHGE.

4. The construction method according to claim 1, characterized in that, The 27 radiomics features used in step 5 to predict the time of epidural hematoma formation throughout the 0-12 hour timeframe include: Feature name DISCRETIZED_HUSkewness, coefficient -0.794956; Feature name DISCRETIZED_HISTO_Entropy_log10, coefficient 0.137586; Feature name GLRLM_LRE, coefficient 0.129044; Feature name GLZLM_SZE, coefficient 0.547362; Feature name LoG-GLCM_Contrast, coefficient -0.326405; Feature name LoG-GLCM_Correlation, coefficient -0.028087; Feature name LoG-GLZLM_SZE, coefficient -0.160379; The following features are listed: SWT-LLL-GLZLM_LZE (coefficient 0.528321); SWT-HLL-DISCRETIZED_Kurtosis (coefficient -0.236168); SWT-HLL-GLZLM_SZE (coefficient 0.226367); SWT-HLL-GLZLM_LZLGE (coefficient 0.555236); SWT-LHL-CONVENTIONAL_mean (coefficient 0.233665); SWT-LHL-DISCRETIZED_Kurtosis (coefficient 0.349986); and SWT-LHL-GL... CM_Correlation, coefficient 0.030686; Feature name SWT-LLH-GLRLM_LRHGE, coefficient 0.304777; Feature name SWT-LLH-NGLDM_Busyness, coefficient -0.408665; Feature name SWT-LLH-GLZLM_LZE, coefficient -0.141548; Feature name SWT-HLH-CONVENTIONAL_Q2, coefficient 0.059320; Feature name SWT-HLH-DISCRETIZED_Kurtosis, coefficient -0.625426; Feature name SWT-HLH-GLCM_Contra st, coefficient is 0.145714; feature name SWT-HLH-GLZLM_SZE, coefficient is 0.111274; feature name SWT-LHH-CONVENTIONAL_Kurtosis, coefficient is -0.089323; feature name SWT-LHH-GLRLM_LRLGE, coefficient is 0.130523; feature name SWT-LHH-NGLDM_Coarseness, coefficient is -0.199430; feature name SWT-LHH-NGLDM_Busyness, coefficient is 0.502334; feature name SWT-HHH-CONVENTIONAL_mean, coefficient is 0.161991; Feature name: SWT-HHH-GLZLM_SZLGE; Coefficient: 0.443160; Intercept 5.515152.

5. The construction method according to claim 1, characterized in that, Step 5 uses 13 radiomics features to predict the time of epidural hematoma formation in the subdivision of injury time ≤ 5 hours, including: Feature name GLRLM_LRE, coefficient 0.183608; Feature name LoG-CONVENTIONAL_std, coefficient -0.059375; Feature name LoG-CONVENTIONAL_Q1, coefficient 0.096689; Feature name LoG-GLCM_Contrast, coefficient -0.381450; Feature name SWT-HLL-GLZLM_SZLGE, coefficient 0.26923; Feature name SWT-LHL-CONVENTIONAL_Q2, coefficient 0.253596; Feature name SWT-LHL-CONVENTIONAL_ Skewness, coefficient 0.068589; Feature name SWT-LHL-GLRLM_RLNU, coefficient -0.040628; Feature name SWT-HHL-GLZLM_ZLNU, coefficient 0.087925; Feature name SWT-LLH-GLZLM_LZE, coefficient -0.066777; Feature name SWT-LLH-GLZLM_ZLNU, coefficient 0.011010; Feature name SWT-HLH-GLRLM_LRLGE, coefficient -0.145954; Feature name SWT-HLH-GLZLM_LGZE, coefficient -0.203099; Intercept 2.500000.

6. The construction method according to claim 1, characterized in that, In step 5, 19 radiomics features were used to predict the time of epidural hematoma formation when the injury time was subdivided into 6–12 hours, including: Feature name GLZLM_SZE, coefficient 0.110848; Feature name LoG-CONVENTIONAL_Q1, coefficient 0.432929; Feature name LoG-GLCM_Contrast, coefficient -0.195671; Feature name LoG-GLZLM_SZE, coefficient -0.306248; Feature name SWT-HLL-DISCRETIZED_Kurtosis, coefficient -0.319317; Feature name SWT-HLL-GLRLM_LRHGE, coefficient -0.076523; Feature name SWT-HLL-GLZLM_SZE, coefficient 0.305857; Feature name SWT-HLL-GLZLM_HGZE, coefficient 0.190681; Feature name SWT-HLL-GLZLM_LZHGE, coefficient 0.157684; Feature name SWT-LHL-CONVENTIONAL_Q2, coefficient 0.0 16221; Feature name SWT-LHL-GLRLM_LRHGE, coefficient -0.345238; Feature name SWT-LHL-GLZLM_SZLGE, coefficient -0.126023; Feature name SWT-HHL-GLRLM_RLNU, coefficient -0.326098; Feature name SWT-HHL-GLZLM_SZHGE, coefficient -0.119980; Feature name SWT-LLH-CONVENTION AL_Q1, coefficient is 0.147801; feature name SWT-HLH-GLCM_Correlation, coefficient is -0.269171; feature name SWT-HLH-GLZLM_GLNU, coefficient is 0.106151; feature name SWT-LHH-GLZLM_SZE, coefficient is 0.468498; feature name SWT-HHH-CONVENTIONAL_Skewness, coefficient is 0.015185; Intercept 8.352941.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the construction method as described in any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the construction method as described in any one of claims 1 to 6.

Images