Magnetic resonance imaging-based radiomics model for preoperative assessment of risk stratification in endometrial cancer

Abstract

BACKGROUND

Preoperative risk stratification is significant for the management of endometrial cancer (EC) patients. Radiomics based on magnetic resonance imaging (MRI) in combination with clinical features may be useful to predict the risk grade of EC.

AIM

To construct machine learning models to predict preoperative risk stratification of patients with EC based on radiomics features extracted from MRI.

METHODS

The study comprised 112 EC patients. The participants were randomly separated into training and validation groups with a 7:3 ratio. Logistic regression analysis was applied to uncover independent clinical predictors. These predictors were then used to create a clinical nomogram. Extracted radiomics features from the T2-weighted imaging and diffusion weighted imaging sequences of MRI images, the Mann-Whitney U test, Pearson test, and least absolute shrinkage and selection operator analysis were employed to evaluate the relevant radiomic features, which were subsequently utilized to generate a radiomic signature. Seven machine learning strategies were used to construct radiomic models that relied on the screening features. The logistic regression method was used to construct a composite nomogram that incorporated both the radiomic signature and clinical independent risk indicators.

RESULTS

Having an accuracy of 0.82 along with an area under the curve (AUC) of 0.915 [95% confidence interval (CI): 0.806-0.986], the random forest method trained on radiomics characteristics performed better than expected. The predictive accuracy of radiomics prediction models surpassed that of both the clinical nomogram (AUC: 0.75, 95%CI: 0.611-0.899) and the combined nomogram (AUC: 0.869, 95%CI: 0.702-0.986) that integrated clinical parameters and radiomic signature.

CONCLUSION

The MRI-based radiomics model may be an effective tool for preoperative risk grade prediction in EC patients.

Key Words: Endometrial cancer, Risk stratification, Radiomics, Machine learning, Nomogram

Core Tip: Our research focused on the utilization of clinical features and radiomics derived from magnetic resonance imaging (MRI) in order to predict the risk grade of endometrial cancer (EC). In our studies, we constructed a predictive model capable of predicting preoperative EC risk. The MRI-based radiomics model put forth in this research showed strong predictive ability and great potential value for assessing the level of EC risk. The integration of predictive models into clinical practice would greatly enhance the preoperative selection of customized therapy.

INTRODUCTION

Endometrial cancer (EC), characterized by the abnormal growth of cells in the lining of the uterus, is a prevalent gynecologic malignancy with a significant global health impact for females[1]. The early detection rate of EC is normally high, with approximately 80% of cases being diagnosed in stage I, with almost 95% 5-year survival rates[2]. However, if a disease has spread regionally or far away, the 5-year survival rate is substantially lower (68% and 17%, respectively)[3].

According to the European Society for Medical Oncology, the European Society for Radiotherapy and Oncology, and the European Society of Gynaecological Oncology Consensus Conference non-molecular risk classification of EC from 2020[4], patients who are at a heightened risk demonstrate the presence of at least one of the subsequent histologic characteristics: Deep infiltration into the myometrium (DMI); a tumor of high grade; a histological subtype that is non-endometrioid; lymph-vascular spaces invasion (LVSI); spread beyond the uterus; or involvement of lymph nodes.

Bilateral salpingo-oophorectomy and total abdominal hysterectomy are the standard treatments for low-risk EC. Although numerous clinical trials have sought to demonstrate the benefits of adjuvant therapy and lymphadenectomy (LA) for EC in its early stages, whether it truly meets the needs of patients remains controversial[5]. Approximately 7%-10% of patients undergoing lymph node dissection have postoperative lymphatic fistula, and roughly 23% of people are said to have lower extremity lymphedema on average[6]. According to the guidelines, individuals diagnosed with low-risk endometrioid carcinoma exhibit a diminished likelihood of lymph node involvement, hence LA is not advised for these patients. However, high-risk patients who had previously had partial surgery who needed LA to complete staging may be a good candidate for this procedure. For patients with intermediate-risk and high-risk endometrial carcinoma, surgical tumor debulking including enlarged lymph nodes and adjuvant treatment is recommended.

Therefore, we merged intermediate and high-risk groups as the high-risk group. Hence, identifying an effective approach for accurately evaluating risk stratification in EC cancer before treatment is crucial to ensure precise decision-making. Precise evaluation of the primary tumor, lymph nodes, and distant metastases is crucial for predicting the risk stratification of EC, particularly in the context of DMI.

Transvaginal ultrasonography is inferior to magnetic resonance imaging (MRI), which provides higher diagnostic performance[7,8]. Preoperative evaluation of EC staging using combined MRI T2-weighted imaging (T2WI) sequences is considered the best way for accurate assessment of DMI, particularly in cases where there is disappearance of the junctional zone (commonly seen in postmenopausal patients) and poor tumor-myometrial contrast[9]. Nevertheless, there is often an overestimation of the stage in instances where the tumor affects the uterine horns, when the tumor extends extensively into the uterine wall causing thinning of the myometrium, or in the event it occurs in a leiomyomatous uterus marked due to the varied signal strength observed in the myometrium[10,11].

The utilization of radiomics within clinical decision systems has witnessed a notable surge in recent years, leading to enhanced precision in the realms of diagnosis, prognosis, and prediction[12,13]. The primary objective of radiomics is transforming medical pictures into digital information that encompasses physiological and fundamental descriptive variables, such as contrast enhancement, diffusion properties, and tracer pickup. Additionally, it encompasses fundamental describing metrics, among which are dimension, form, intensity, and texture[14]. The application of radiomics is frequently observed in the field of oncology. According to reports, medical imaging characteristics were useful in differentiating across EC risk groups[15,16].

In the assessment of the predictive capacity of three risk stratification models for lymph node transfer in endometrioid EC, Korkmaz et al[17] found that the model that exhibited the highest performance attained an area under the curve (AUC) value of 0.780. Liu et al[18] concluded that the utilization of texture analysis exhibited promising prospects in its capacity to function as imaging indicators for evaluating preoperative risk. However, additional study is required to examine the evaluation of MRI-based radiomic analysis in conjunction with clinical factors for the purpose of preoperative risk stratification in patients with EC.

The objective of the present research was to inquire into the utilization of clinical features and radiomics derived from MRI in order to predict the risk grade of EC.

MATERIALS AND METHODS

Patients

The retrospective inquiry in question was exempt from the requirement of informed consent as it had obtained approval from the institutional review board. We reviewed 315 patients with EC that had postoperative pathological results between January 2018 and January 2020 in our hospital. Based on the following criteria, 112 cases were included in the study. Patients were first stratified according to the European Society of Gynaecological Oncology, the European Society for Medical Oncology, and the European Society of Pathology 2020 guidelines and then divided into two groups: Low-risk (including patients belonging to the low-risk class according to the 2020 guidelines classification); and high-risk (including intermediate, intermediate-high, and high-risk classes).

The inclusion criteria were: (1) Patients with pathologically confirmed EC who received an MRI exam 2 wk before surgery; and (2) Individuals who had not received any therapeutic interventions, including surgical procedures, biopsies, radiation therapy, chemotherapy, or hormone therapy, prior to their MRI. The exclusion criteria were: (1) Patients with missing clinical data; (2) Poor MRI image quality or significant artifacts that did not match the requirement of analysis; and (3) Lesions with a diameter of less than 5 mm. A random assignment method was used to allocate patients in a 7:3 ratio between the training group (n = 78) and the validation cohort (n = 34). The research flowchart is portrayed in Figure 1.

Figure 1 Technology roadmap for this research. A: Flow chart of patient enrollment; B: Workflow of radiomics analysis process. AUC: Area under the curve; DT: Decision tree; EC: Endometrial cancer; KNN: K-nearest neighbor; LDA: Linear discriminant analysis; LR: Logistic regression; MRI: Magnetic resonance imaging; MSE: Mean-square error; RF: Random forest; ROC: Receiver operating characteristic; ROI: Region of interest.

Data and images collection

Clinical and histopathological traits of all selected patients, such as age, histological grade and subtype, International Federation of Gynaecology and Obstetrics staging, body mass index, hypertension, hyperlipidemia, diabetes mellitus, menstruation, menopause status, fertility history, serum cancer antigen (CA) 125 level, serum CA19-9 level, and drinking and smoking history, were collected. These data were obtained through consultation of the medical record system. MRI pictures of all patients in digital imaging and communications in the medicine type were exported.

All MRI examinations were performed using 3.0 T system MRI scanners (Magnetom Aera; Siemens Healthcare, Germany). The standard scanning protocol comprised axial fast spin-echo (FSE) T1-weighted images, axial FSE T2WI, axial fat suppression FSE T2WI, and sagittal FSE T2WI. The b values utilized in diffusion weighted imaging (DWI) encompassed the values of 0 and 1000 s/mm². In the study, uterus-sagittal position T2WI were acquired for lesion segmentation. Other imaging was used as a reference for lesion segmentation. Then, the radiomics features were extracted based on T2WI and DWI sequences. Table 1 displays the MRI parameters in detail.

Table 1 Magnetic resonance imaging scanning parameters.

Sequence	TR/TE in ms	FOV in mm × mm	Matrix	Slice gap in mm	Slice thickness in mm
Axial T1WI	500/8.6	310 × 310	224 × 320	1	6
Axial T2WI	6200/95	310 × 310	307 × 384	1	6
Axial (FS)-T2WI	4000/93	310 × 310	256 × 320	1	5
Sagittal (FS)-T2WI	6000/86	240 × 240	205 × 256	1	6
Axial DWI	5000/69	327 × 245	115 × 192	1	4

DWI: Diffusion weighted imaging; FOV: Field of view; FS: Fat suppression; T1WI: T1-weighted images; T2WI: T2-weighted imaging; TE: Echo Time; TR: Repetition time.

Image segmentation and radiomic features extraction

The open-source software three-dimensional (3D) slicer (version 5.2.1), which may be accessed at https://www.slicer.org, was employed for the purpose of image segmentation. Segmentation was performed independently and manually by two experienced doctors who were unaware of the pathological result of the patient. The entire tumor was included in the 3D volume of interest (VOI), which the medical professionals defined, divided, and fused for a layer after layer screen. In the event of a disagreement, the two doctors talked it out until an agreement was reached.

The VOIs underwent resampling to a voxel size of 3 mm × 3 mm × 3 mm before extracting features to produce isotropic voxels; to ensure that the gray-level values of all photos were dispersed over the same range, image normalization was carried out. The VOI of each patient was utilized to extract radiomics features using the free and publicly accessible python program pyradiomics (https://pypi.org/project/pyradiomics/). The following categories were created from the extracted features: First-order features; two-dimensional features; gray-level cooccurrence matrix (GLCM); gray-level dependence matrix; gray-level size-zone matrix; gray-level run-length matrix; and neighboring gray tone difference matrix[19]. In total, 1037 radiomics characteristics were retrieved.

Radiomic feature picking and signature building

To mitigate interference across the feature dimensions, the features underwent standardization making use of the Z-score method, which involves deducting the average value and dividing by the standard deviation. Initially, to assess what features could individually separate the low-risk group from the high-risk group, we conducted the Mann-Whitney U test. Only features exhibiting a P value of 0.05 were retained. Subsequently, to eliminate highly correlated features, we removed the features with P > 0.9. The utilization of the least absolute shrinkage and selection operator (Lasso) in data analysis allows for the reduction of coefficients associated with variables that are not relevant to survival to zero while retaining features that have coefficients that are not equal to zero[20,21]. In this study, the Lasso method was employed to identify radiomics properties with the highest impact, and a 10-fold cross-validation approach was implemented within the training cohort. The minimum criteria (minimum lambda) were used to identify the ideal tuning parameter.

The radiomic signature (radscore) was calculated by adding up all of the filtered eigen values and multiplying the total by the appropriate coefficients. Radscore = β1X1 + β2X2 + …βnXn, with the radiomic signature denoted as radscore, βn was the coefficient, and Xn was the eigenvalue[22]. The assessment of the consistency between radscore in the training and testing sets was conducted by employing the Wilcoxon rank sum Mann-Whitney U test.

Model construction

By applying univariate and multivariate regression analysis, the clinical independent predictors were discovered. Applying the independent predictors, a clinical nomogram was then produced. To categorize the EC based on their radiomics features because we needed a model to solve a binary classification problem and we did not have a very large sample size, seven popular binary machine learning algorithms [logistic regression (LR), K-nearest neighbor (KNN), decision tree (DT), random forest (RF), gradient boosting (GB), eXtreme GB (XGBoost), and GBDT] with good explanations were used. A combined prediction multivariate logistic regression model was developed by including the radiomic signature (radscore) and clinically significant predictors that demonstrated statistical significance in a multivariate regression analysis.

Model performance and comparison and statistical analysis

The building of the ideal radiomics model was accomplished by employing a machine learning technique that demonstrated the maximum AUC value in the validation cohort[23]. AUCs of the models were compared via the Delong test[24]. The calibration was verified using calibration curves. The method of decision curve analysis was employed in order to assess and quantify the clinical value associated with each model[25].

Statistical tests suitable for the data type were employed to compare the baseline data between the training and validation cohorts. Specifically, the χ² test or the Fisher exact test were used for categorical variables, while the continuous variables were assessed using either the two-sample t-test or the Mann-Whitney U test. Statistical significance was attributed to data sets with P values lower than 0.05. Statistical analyses were performed utilizing the statistical product and service solutions (version 26.0, international business machines corporation, https://www.ibm.com/spss), R software (version 3.5.1, https://www.r-project.org/), and Python (version 3.5.6, https: //www.python.org/).

RESULTS

Demographic characteristics

A sample size of 112 individuals, aged between 31 years and 84 years (average age of 58.35 ± 11.75 years), was partitioned into two separate sections: Low-risk (n = 56) and high-risk (n = 56; including 42 intermediate risk cases and 14 high-risk cases). The participants were randomized to the training and validation subgroups in a random manner, following a 7:3 ratio. In the training subgroup, there were 39 patients classified as low-risk and an equal number of 39 patients classified as high-risk. Similarly, in the validation cohort, there were 17 patients classified as low-risk and an equal number of 17 patients classified as high-risk. The training and validation cohorts did not significantly differ in either clinical or histological characteristics. Table 2 displays the fundamental characteristics of the individuals.

Table 2 Characteristics of endometrial cancer patients in the training and validation cohorts.

Characteristics	Training cohort	Validation cohort	P value
Number	78	34
Age in yr	58.04 ± 11.54	59.06 ± 12.04	0.367
Histological grade1			0.465
G1/G2	58	23
G3	20	11
Risk classes			1.000
Low-risk	39	17
High-risk	39	17
Histological subtype1			0.018
Endometrioid	66	22
Serous	6	5
Clear cell	4	4
Carcinosarcoma	2	3
FIGO staging			0.765
I	50	21
II	11	7
III	9	4
IV	8	2
BMI
> 24	44	19	0.959
≤ 24	34	15
Hypertension			0.812
Positive	20	8
Negative	58	26
DM			0.605
Positive	15	8
Negative	63	26
Hyperlipidemia			0.338
Positive	5	4
Negative	73	30
Menstruation			0.395
Regular	48	18
Irregular	30	16
Menopausal status
Premenopausal	57	27	0.477
Postmenopausal	21	7
Fertility			0.669
Fertility	69	31
Nonfertility	9	3
CA125			0.617
< 35 U/mL	61	28
≥ 35 U/mL	17	6
CA19-9			0.631
< 27 U/mL	67	28
≥ 27 U/mL	11	6
Family history of cancer			0.292
Positive	10	7
Negative	68	27
Drinking history			0.915
Positive	5	2
Negative	73	32
Smoking history			0.606
Positive	4	1
Negative	74	33

¹Histological grade and subtype of endometrioid adenocarcinoma from surgical specimen.

Data are n or mean ± standard deviation. FIGO: The International Federation of Gynaecology and Obstetrics; BMI: Body mass index; CA125: Cancer antigen 125; CA19-9: Cancer antigen 19-9; DM: Diabetes mellitus.

Radiomic feature selection and radiomic signature construction

The 3D Slicer software was utilized to extract a total of 1037 quantitative imaging feature parameters for each patient. Following deredundancy processing, 788 features were kept, and 15 feature parameters were chosen using the Lasso dimensionality reduction approach (Figure 2). The radiomics score for each patient was computed via a linear equation using the following formula:

Figure 2 Least absolute shrinkage and selection operator regression model for screening the radiomics characteristics of the training group. A: Screening of the radiomics features was performed through least absolute shrinkage and selection operator (Lasso) regression. The cross validation for Lasso regression, where the parameter λ was adjusted to find the best function set, is shown. The vertical dotted line on the left panel represents the log(λ) corresponding to the optimal λ; B: Screening of the radiomics features was performed through Lasso regression. The coefficients of texture parameters changed with λ. The vertical line corresponds to the 10 features selected with non-zero Lasso cross-validation coefficients. MSE: Mean-square error.

Radscore = (0.085 × original_firstorder_Minimum) + (0.046 × original_glcm_ldmn) + [(-0.039) × log-sigma-4-4-mm-3D_firstorder_Energy] + [(-0.075) × log-sigma-4-4-mm-3D_firstorder_Maximum] + [(-0.133) × wavelet-LLH_firstorder_Median] + (0.072 × wavelet-LLH_firstorder_Skewness) + [(-0.057) × wavelet-LLH_glszm_LowGrayLevelZoneEmphasis] + [(-0.077) × wavelet-LHL_firstorder_Median] + (0.048 × wavelet-LHL_glcm_ClusterProminence) + (0.145 × wavelet-LHH_glcm_ClusterProminence) + (0.046 × wavelet-HHL_glcm_JointEnergy) + [(-0.088) × wavelet-HHH_glcm_MCC] + [(-0.075) × wavelet-LHH_gldm_LargeDependenceLowGrayLevelEmphasis] + [(-0.038) × wavelet-LLL_glrlm_LongRunHighGrayLevelEmphasis] + (0.046 × wavelet-LLL_glszm_LargeAreaLowGrayLevelEmphasis).

The study observed that the radscore exhibited greater values in the high-risk subgroup in comparison to the low-risk subgroup. Additionally, statistical analysis demonstrated that the radscore had a strong capability to differentiate between EC patients with low risk and those with high risk, as observed in both the training and validation cohorts. Low-risk EC patients had a radscore of 0.25 ± 0.20 in the training cohort compared to high-risk patient radscores of 0.75 ± 0.15 (P = 0.000); in the validation cohort, low-risk EC patients had a radscore of 0.24 ± 0.39 compared to high-risk patient radscores of 0.67 ± 0.21 (P = 0.000) (Figure 3).

Figure 3 Comparison of low-risk and high-risk endometrial cancer radscores. A: Training groups; B: Testing groups. 1: Low-risk endometrial cancer; 0: High-risk endometrial cancer.

Development and validation of the clinical model

Logistic regression analysis, both univariate and multivariate, was used to filter out all of the clinically suspicious risk factors (Table 3). The findings revealed that age, CA125, and CA19-9 were clinical independent predictors. A clinical model was constructed with the predictors. The AUC was 0.751 [95% confidence interval (CI): 0.611-0.899] in validation set, and accuracy was 0.706. Figure 4A depicts the clinical prediction model nomogram.

Figure 4 Nomogram to predict endometrial cancer risk. Cancer antigen (CA) 125 label 1 corresponds to serum levels below 35 U/mL, and label 0 corresponds to serum levels higher than 35 U/mL. CA19-9 label 1 corresponds to serum levels below 27 U/mL, and label 0 corresponds to serum levels higher than 27 U/mL. A: Nomogram developed by clinical predictors; B: Radiomics-clinical combined nomogram.

Table 3 Univariate regression analysis and multivariate regression analysis.

Characteristics	Univariate regression OR (95%CI)	P value	Multivariate regression OR (95%CI)	P value
Age	Reference 1.091 (1.038-1.147)	0.000	Reference 1.10 (1.000-1.019)	0.048
BMI
> 24	Reference
≤ 24	2.333 (0.933-5.833)	0.070
Hypertension
Positive	Reference
Negative	0.438 (0.152-1.256)	0.124
Diabetes mellitus
Positive	Reference
Negative	0.848 (0.274-2.619)	0.774
Hyperlipidemia
Positive	Reference
Negative	0.649 (0.102-4.113)	0.646
Menstruation
Regular	Reference
Irregular	1.591 (0.615-4.114)	0.338
Menopausal status
Premenopausal	Reference
Postmenopausal	2.560 (0.898-7.296)	0.079
Reproductive history
Positive	Reference
Negative	0.777 (0.192-3.142)	0.724
CA125
< 35 U/mL	Reference		Reference
≥ 35 U/mL	26.435 (3.284-29.776)	0.002	1.515 (1.163-1.973)	0.003
CA19-9			Reference
< 27 U/mL	Reference		1.355 (1.030-1.783)	0.033
≥ 27 U/mL	5.550 (1.114-27.648)	0.036
Family history of cancer
Negative	Reference
Positive	1.000 (0.265-3.772)	1.000
Drinking history
Negative	Reference
Positive	4.343 (0.463-40.749)	0.199
Smoke history
Negative	Reference
Positive	0.316 (0.031-3.177)	0.328
Radscore	1.027 (1.002-1.054)	0.038	1.025 (0.990-1.060)	0.016

BMI: Body mass index; CA125: Cancer antigen 125; CA19-9: Cancer antigen 19-9; CI; Confidence interval; OR: Odds ratio.

Development and validation of the radiomics models

Radiomics models were created using LR, KNN, DT, RF, GB, XGBoost, and GBDT. The model performance information is displayed in Table 4. The discrimination of the RF algorithm model outperformed the others.

Table 4 Outcomes of radiomics models in validation set.

Algorithms	AUC (95%CI)	Accuracy	Sensitivity	Specificity
LR	0.794 (0.706-0.976)	0.794	0.765	0.824
KNN	0.893 (0.766-0.971)	0.765	0.823	0.705
DT	0.676 (0.529-0.824)	0.676	0.529	0.824
RF	0.915 (0.806-0.986)	0.824	0.824	0.824
GB	0.907 (0.792-0.979)	0.794	0.706	0.882
XGBoost	0.869 (0.739-0.965)	0.765	0.706	0.824
GBDT	0.872 (0.742-0.958)	0.765	0.647	0.882

AUC: Area under the curve; CI: Confidence interval; DT: Decision tree; GB: Gradient boosting; GBDT: Gradient boosting decision tree; KNN: K-nearest neighbor; LR: Logistic regression; RF: Random forest; XGBoost: eXtreme gradient boosting.

Development and validation of the combined model

The LR technique was employed to generate a nomogram that integrated the variables of age, CA125, CA19-9, and radscore into a combined model. The AUC in the test collection demonstrated outstanding discrimination, with a value of 0.869 (95%CI: 0.702-0.986). Figure 4B displays the nomogram representing the combined prediction models.

Comparison of models

Figure 5, including the receiver operating characteristic curves, calibration curves, and decision curve analysis, presents a comparative analysis of the performance between a clinical model, the best radiomics model, and the combined model. According to the Delong test (P < 0.05), the radiomics model exhibited superior discriminatory performance compared to both the clinical and combined models.

Figure 5 Evaluation of the radiomics model, clinical model, and combined model for predicating endometrial cancer risk grading. A: Receiver operating characteristic curve; B: Calibration curve; C: Decision curve analysis. AUC: Area under the curve.

DISCUSSION

This study involved the development of three models that were based on clinical independent risk factors, radiomics features, and the combination of clinical risk factors with radscore. The objective was to predict risk stratification for EC. This study demonstrated that the MRI-based radiomics models and the clinical models have good efficacy when assessing the risk grade of EC. The diagnostic efficacy of the combined model (AUC: 0.869, 95%CI: 0.702-0.986) was higher than the clinical model (AUC: 0.75, 95%CI: 0.611-0.899), and the radiomics model (AUC: 0.915, 95%CI: 0.806-0.986) developed based on the imaging features extracted from the T2WI and DWI sequence had the highest efficacy. The RF model performed best in this investigation out of various classifiers utilized to build the radiomics predictive model. This study showcased the efficacy of utilizing a comprehensive tumor radiomics-based analysis to evaluate the risk classification of EC before surgical intervention. The radiomics model may be clinically effective in the individualized surgical care of EC patients since radiomics may extract relevant data about high-risk indicators prior to surgery.

Some studies have explored preoperative prediction of EC risk stratification. Yan et al[26] found that variables for high-risk EC included metabolic syndrome, CA125, age, and tumor grade after curettage. Saarelainen et al[27] also concluded that CA125 paired with other serum indicators was superior to CA125 alone in predicting the recurrence of disease in EC patients. CA125 is a factor that relates to high-risk EC. In line with this, age, CA125, and CA19-9 were independent risk factors associated with EC risk stratification in this study.

Previous studies have noted that patients with EC, particularly those exhibiting unfavorable prognostic factors such as tumor recurrences, grade 3 tumors, DMI, lymph node metastasis, and extrauterine disease, utilize CA125 as a biomarker for tumor detection[28]. CA125 has been integrated into many pre-surgery prediction models, with varying levels of effectiveness. In clinical settings, the utilization of a threshold value of 35 U/mL for CA125 can prove to be a valuable diagnostic tool due to its high sensitivity. This is evidenced by the fact that a mere 1% of females who are in good health exhibit CA125 levels surpassing this cutoff point[29]. Therefore, this widely used cutoff was also used in this study.

Furthermore, there is research that has demonstrated notable correlations between International Federation of Gynaecology and Obstetrics staging and several other characteristics. These factors encompass inadequate histological classification, progressed clinical stage, existence of metastatic lymph nodes, and heightened blood CA19-9 levels[30]. Elderly females are more likely to have metabolic syndrome, have worse basic health status, and are more likely to develop high-risk EC. A study showed that reproductive history was a high-risk indicator for the emergence of EC[31]. In the current study, however, this conclusion was not found, suggesting that reproductive history may be more related to the incidence of EC instead of risk grade.

Previous studies have found a connection between the characteristics of tumor texture and diagnosis and grading[32,33]. Celli et al[34] found that the differentiation of low-risk EC from other risk categories, as well as the identification of the existence of LVSI, can be achieved with moderate to high accuracy by the utilization of MRI-based whole-tumor radiomics and radio-genomic investigations. Using MRI texture features, Ueno et al[35] established preliminary mathematical modelling that showed a correlation with DMI, LVSI, and high tumor grade. The AUC of the model in predicting high-risk EC was 0.83, but they did not perform the volumetric analysis. High-throughput parameters were extracted for the current investigation, and 15 statistically significant features were then found to create a radiomics model. The model performed well in identifying low-risk and high-risk EC; the sensitivity and specificity of the validation cohort were 82.4% and 82.4%, and the AUC was 0.915.

The GLCM feature characterizes the spatial arrangement of grey intensity in a two-dimensional manner by quantifying the probability of observing specific pixel pairs with particular grey values[36,37]. The chosen feature in this study comprised six GLCM features, indicating a close association between the distribution of texture and spatial heterogeneity within tumors and tumor differentiation. Moreover, alterations in the internal texture of tumors play a crucial role in determining tumor risk grading.

Among the seven machine learning algorithms we used to develop a radiomics model, the RF model performed best. It is based on DT analysis; the algorithm builds DTs by employing a random selection process to determine a subset of the samples and a subset of the features to be used for the split at each node. This greatly reduces the variance term in the diagnostic error. This is an advantage over bagging of the DTs[38]. In the area of biomedical research, the characteristics of RF make it the perfect material for creating diagnostic models.

Our study had some limitations. First, our model might have been overfitted because we only included 112 patients in this early investigation. Second, even though the MRI was performed according to a set protocol, the retrospective nature of the study could result in inhomogeneity un the imaging data. Although normalization was used in the picture analysis, image standardization and normalization techniques still require research. Third, we included only T2WI map images in our model, which may potentially enable missing significant data. To limit the potential of biases from only one sequence, future studies should use multiparametric techniques.

CONCLUSION

The MRI-based radiomics model put forth in this research showed strong predictive ability and great potential value for assessing the level of EC risk. The integration of predictive models into clinical practice would greatly enhance the preoperative selection of customized therapy.

ACKNOWLEDGEMENTS

The authors express their heartfelt thanks to the staff of the Department of Obstetrics and Gynecology of our hospital.