Diagnosis of Malignant Endometrial Lesions from Ultrasound Radiomics Features and Clinical Variables Using Machine Learning Methods

Abstract

Background:

The prognosis of patients with early diagnosis of malignant endometrial lesions is good. We aimed to identify benign and malignant lesions in endometrial tissue, explore effective methods for assisting diagnosis, and improve the accuracy and precision of identifying endometrial lesions.

Methods:

1142 ultrasound radiomics and 18 clinical features from 1254 patients were analyzed, from which 36 features were selected for machine learning. We sketched the region of interest (ROI) of the abnormalities on the ultrasound images. Then, the radiomics features were extracted. Six common machine learning algorithms, including Support Vector Machine (SVM), Logistic Regression, Decision Tree, Random Forest, Gradient Boosting Tree, and k-Nearest Neighbors, were employed to identify benign and malignant changes in endometrial tissue. Cross-validation and grid search techniques for hyperparameter tuning were utilized to obtain the best model performance. Accuracy, precision, sensitivity, F1-scores, area under the curve (AUC), cross-validation average score and bootstrap average accuracy were also used to evaluate algorithm performance, classification accuracy, and generalization capability.

Results:

We combined 21 ultrasound characteristics and 15 clinical characteristics to develop and validate six common machine learning algorithms. After internal validation, the best models were the Random Forest models, with accuracy of 89%, precision of 93%, sensitivity of 97%, F1-score of 95%, and AUC of 95%, as well as a 10-fold cross-validation average score of 95% and bootstrap average accuracy of 94%, implying flawless classification in the test set.

Conclusions:

We identified the clinical and ultrasound features in the early diagnosis of benign or malignant lesions in endometrial tissue. And Random Forest model algorithms have demonstrated excellent performance in identifying benign and malignant changes in endometrial tissue. This is significant for enhancing early diagnostic accuracy and improving treatment outcomes and long-term management.

Graphical abstract

Keywords

endometrial lesions / early diagnosis / ultrasound radiomics / clinical features / machine learning

1. Introduction

The rising incidence of endometrial cancer, one of the most common types of gynecological malignancies, significantly impacts women’s health [1, 2]. Unfortunately, this disease often presents with subtle or no noticeable symptoms in its early stages, leading many patients to miss the optimal window for treatment, resulting in increased treatment difficulty and decreased survival rates. Therefore, early identification of benign or malignant changes in endometrial tissue is a crucial challenge in clinical practice to reduce patient mortality and improve treatment outcomes.

The diagnostic criteria of endometrial cancer is pathologic finding, which is usually based on endometrial dilatation and curettage or hysteroscopic sampling or hysterectomy biopsy. However, this is an invasive diagnosis and does not reflect the severity of the disease. Some previous studies have identified a lot of non-invasive diagnostic factor for diagnosing suspected endometrial tumors, including demographic features, medical history, tumor biomarkers, ultrasound and magnetic resonance imaging [3, 4, 5]. Transvaginal ultrasound should be utilized as first-line imaging modality, being more available, cost-effective, and more acceptable by patients. And endometrial echogenicity, endometrial–myometrial junction, blood flow signals, ect., have been investigated on ultrasound [6]. Researchers have developed some prediction models to facilitate the early diagnosis of symptomatic patients [7], and to predict endometrial cancer only based on clinical features or ultrasound features [8, 9]. Currently, there is lack of a comprehensive, population-wide predictive model to identify benign and malignant lesions in endometrial tissue integrating of ultrasound features and clinical features.

Radiomics provided a noninvasive method to extract numerous quantitative features from the standard medical images [10]. Then, those features can then be applied in statistical analysis, machine learning, or other analysis [11]. Ultrasound radiomics has been widely used to study in endometrial cancer, including predicting high-risk endometrial cancer and predicting the prognosis of patients with endometrial cancer [12, 13, 14]. Hence, we hypothesized that ultrasound radiomics features could show endometrial cancer heterogeneity and diagnose malignant endometrial lesions.

Machine learning algorithms have recently made remarkable progress in medical diagnosis, becoming powerful tools for assisting clinical decision-making. By learning patterns and features from available data, machine learning algorithms automatically build models for precise disease diagnosis and prediction [11, 15]. Compared to traditional rule-based and statistical methods, machine learning algorithms excel at discovering complex relationships and patterns in data, resulting in higher accuracy and generalization capabilities [16, 17]. Specifically, in endometrial cancer, the application of machine learning has offered notable contributions towards non-invasive diagnostics and prognostication [18]. Such as, Vipul Bhardwaj and colleagues [19] provided a sketch of endometrial cancer in risk factors and diagnostic methods, underscoring machine learning’s potential in endometrial cancer diagnosis and prognosis.

The aim of this study was (1) to identify certain ultrasound features of benign and malignant changes in endometrial tissue and explore effective methods for assisting diagnosis, and (2) to build and validate six algorithms integrated clinical and ultrasound features to identify benign and malignant endometrial lesions and provide scientific evidence for their ability to enhance early diagnostic accuracy.

2. Materials and Methods

2.1 Participants

The dataset used in this study was collected from a medical database and consists of information from 1254 patients treated between June 1, 2018 and June 1, 2023 at the Central Hospital Enshi Tujia and Miao Autonomous Prefecture, including 953 benign cases and 301 malignant cases with pathological results of intrauterine adhesions. The inclusion criteria were: (1) age >18 years, (2) pathologically confirmed diagnosis, (3) lesions located in the endometrium, (4) transrectal/transvaginal ultrasonography occurred within one month before operation, (5) tumor biomarkers detected within one month before the operation, and (6) an available full medical history. Exclusion criteria were: (1) hysterectomy, (2) receiving preoperative hormone therapy, chemotherapy, or radiation therapy, (3) tumors in other organs, (4) pregnant or lactating women, and (5) recently taking hormone drugs (Fig. 1).

2.2 Clinical Features

After collecting patients’ medical histories, tumor marker examinations and ultrasonic testing, eighteen clinical characteristics including age, body mass index (BMI), gravidity, parity, abortion, breastfeeding status, menopause status, menstrual regularity, irregular bleeding of the vagina, contact bleeding, leucorrhea with blood, hypogastralgia, hypertension, diabetes, CA125, CA15-3, CA19-9 and HE4 [3, 20] and twelve ultrasound features including uniform endometrial echogenicity, endometrial midline appearance, endometrial–myometrial junction, “bright edge”, intracavity fluid, color score, vascular pattern, arterial pulsatility index, resistance index, end-diastolic flow rate, peak flow rate, and average flow rate were included in our study [5, 20]. Especially, regarding the very differences in hormone levels and endometrium appearances of premenopause compared to postmenopause, and menopause status was also taken into consideration [3].

2.3 Ultrasonic Examination

The two-dimensional vaginal ultrasound was performed on all patients according to the international endometrial tumor analysis (IETA) expert consensus recommendation. Ultrasound features for benign and malignant lesions were defined by the IETA [21, 22]. DC-3/DC-3T model ultrasonic instruments (GE, Boston, MA, USA) with a probe frequency of 7.5 MHz were used for vaginal examinations. All ultrasound examinations and measurements were performed by a senior sonographer with 10 years of experience.

2.4 Clinical and Structural Ultrasound Features Selection

Feature selection plays a crucial role in machine learning tasks to reduce the dimensionality of the feature space, improve model generalization, and identify important features related to the target variable [23]. This study adopted feature selection methods based on domain knowledge and statistical analysis [24]. First, based on domain knowledge, features relevant to tumor diagnosis and prediction, such as age, BMI index, medical history, clinical features, and structural ultrasound features, were selected. Second, correlation analysis was performed to calculate the correlation coefficients between each feature and benign/malignant tumors. Selecting features that are significantly correlated with endometrial lesions reduce redundant information and improves model performance.

2.5 Lesions Segmentation

The ITK-SNAP software (http://www.itksnap.org) was utilized to segment the region of interest (ROI) of lesions. Two sonographers with 10 years of experience manually segmented the ROIs by selecting the most representative the ultrasound image (DICOM format) with the largest solid components. Discrepancies were solved by being re-segmented by a senior sonographer with 20 years of experience. All sonographers were blinded to the specific histopathological type and other futures.

2.6 Features Extraction and Selection

The original ultrasonic images and sketched ROIs images were imported into the A.K. software version 3.0.0 (GE Healthcare, Waukesha, Wisconsin, USA). Then, the original image and ROI images of each patient were automatically matched. We adopted nonlinear intensity transformation on image voxels, Gaussian Laplace filter and eight wavelet transform to obtain high-throughput features. A total of 1130 characteristics of seven categories were extracted: (1) histogram parameters; (2) morphology; (3) gray level co-occurrence matrix (GLCM); (4) gray-level run-length matrix (GLRLM); (5) gray-level size zone matrix (GLSZM); (6) neighboring gray-tone difference matrix (NGTDM), and (7) gray-level dependence matrix (GLDM).

Subsequently, univariate logistic regression analysis was used to select candidate features, and 46 features were included. Then, Person’s rank correlation coefficient was also used to calculate the correlation between features, and one of the features with correlation coefficient greater than 0.9 between any two features was retained, and 27 features were retained. Finally, the least absolute shrinkage and selection operator (LASSO) regression analysis were applied for selection of final features in the training set, and retaining only those with a p value of less than 0.05. Ultimately, a total of 11 ultrasound radiomics features were utilized to establish the models (Fig. 2).

2.7 Algorithm Selection and Validation

From the large pool of supervised or unsupervised learning-based classification algorithms available, this study employed six classic machine learning models for training and testing. These models included the Support Vector Machine (SVM), Logistic Regression, Decision Tree, Random Forest, Gradient Boosting Tree, and the k-nearest neighbor (k-NN), all of which are widely used in modeling and classification.

During the model training process, 80% of the benign and malignant groups were randomly selected to make up the training sets, and ten-fold cross-validation was performed. The development sets were randomly divided into ten equal parts, where the nine parts was used to training sets and one part was iteratively excluded for test purposes. The remaining 20% of the groups comprised the validation sets and used grid search to select the best hyperparameter settings [25]. The training dataset was applied to fit the model, while selecting and tuning using the validation set obtained the models’ predictive ability in identifying benign and malignant tumors.

Bootstrap is a resampling method with replacement used to estimate the distribution of sample statistics. Bootstrap average accuracy is the average of accuracies obtained through bootstrap resampling, providing a more robust evaluation of model performance. And cross-validation average score, accuracy, precision, F1-score, sensitivity and area under the curve (AUC) were also used to evaluate performance of our models.

2.8 Statistical Analysis

GraphPad Prism software version 9.5 (GraphPad Software Inc., San Diego, CA, USA) was used for all statistical analyses in this study. Normality was first tested using the Shapiro-Wilk test. Subsequently, quantitative data was expressed as mean and interquartile range (IQR). An independent sample t-test compared normal distributions, and the Mann-Whitney U test was suitable for data with abnormal distributions. Nevertheless, categorical variables were expressed as percentages (%). The Chi-square test or Fisher’s exact test was utilized to compare classification variable data. p 0.05 was considered as statistically significant.

3. Results

3.1 Demographic Characteristic

This study included 1254 patients, and the mean age at ultrasonic imaging examination in benign lesions group and malignant lesions group was 39 years (IQR: 38–40 years), 56 (IQR: 54–58 years). Table 1 details clinical variables and demographics. There were 953 (76%) patients with benign lesions, including 451 cases of endometrial polyps, 165 cases of intrauterine adhesions, 145 cases of endometrial simple hyperplasia, 103 cases of endometrial polypoid hyperplasia, 46 cases of endometrial complex hyperplasia, 35 submucous myomas, and 8 cases of endometritis. Additionally, there were 301 (24%) malignant lesions, including 169 cases of endometrioid adenocarcinoma, 63 cases of endometrial atypical hyperplasia, 30 cases of endometrial carcinoma, 14 cases of endometrial infiltrating adenocarcinoma, 10 cases of uterine carcinosarcoma, 10 cases of serous adenocarcinoma of endometrium, 3 cases of endometrial low-grade squamous intraepithelial lesion, and 2 cases of giant cell type high-grade undifferentiated sarcoma of uterus.

3.2 Clinical and Ultrasound Features

The database comprises 1160 features belonging to four different categories: 18 features from demographic features, 12 features from structural ultrasound features, and 1130 ultrasound radiomics features extracted from A.K. software. Finally, 36 features were selected to build six models, including 15 clinical variables and demographics features including age, BMI, gravidity, parity, abortion, menopause status, menstrual regularity, irregular bleeding of the vagina, contact bleeding, leucorrhea with blood, diabetes, CA125, CA15-3, CA19-9, and HE4 (Table 1), and 10 structural ultrasound features including uniform endometrial echogenicity, endometrial–myometrial junction, intracavity fluid, color score, vascular pattern, arterial pulsatility index, resistance index, end diastolic flow rate, peak flow rate, and average flow rate (Table 2), and 11 ultrasound radiomics features of large dependence low gray level emphasis, non-uniformity of run length, size zone matrix gray level non-uniformity, statistical kurtosis of voxel intensities, statistical uniformity of voxel intensities, run length matrix gray level non-uniformity hara entroy, high gray level run emphasis, long run high gray level emphasis, size zone matrix zone size non-uniformity and short run high gray level emphasis after the univariate logistic regression, Person’s rank correlation coefficient and LASSO regression analysis (Table 3 and Fig. 3). Table 2 details the structural ultrasound features, and the details of ultrasonographic radiomics features were shown in Table 3.

3.3 Machine Learning Models

Six different machine learning models were established to identify benign and malignant changes in endometrial tissue, including SVM, Logistic Regression, Decision Tree, Random Forest, Gradient Boosting Tree, and k-NN. Our results showed that the accuracy of these six models were 86%, 80%, 83%, 89%, 83% and 78%, respectively (Table 4), and the Random Forest model perform best. Besides, the Random Forest models also achieved perfect compared to other six models with precision of 93%, sensitivity of 97%, F1-score of 95%, and AUC of 95%, as well as a 10-fold cross-validation average score of 95% and bootstrap average accuracy of 94%, implying flawless classification in the test set, which means Random Forest classifier can be comprehensive and applicable to identify malignant changes in endometrial tissue. Table 4 and Fig. 4 showcase the results of the performance metrics.

4. Discussion

In this study, we first extracted ultrasound radiomics features, and found models that cooperated radiomics features with clinical parameters improved the accuracy and precision of identifying endometrial lesions. Then, six common machine learning algorithms were built and validated, including SVM, Logistic Regression, Decision Tree, Random Forest, Gradient Boosting Tree, and k-NN, to identify benign and malignant changes in endometrial tissue. And Random Forest model algorithms demonstrated excellent performance in identifying benign and malignant changes in endometrial tissue. The model provided an important reference for improving accuracy of early diagnosis and treatment effects of endometrial cancer.

This innovative research addresses the need for early diagnosis and the application of machine learning algorithms to distinguish benign and malignant endometrial lesions. The potential of these algorithms to enhance clinical diagnosis and treatment decisions is significant, offering new opportunities to revolutionize medical practice and improve women’s health outcomes. By exploring and comparing the performance of various machine learning models, this study seeks to provide valuable insight for the medical community, ultimately paving the way for more precise and effective management of endometrial cancer. Previous studies have explored ultrasound diagnosis of endometrial malignant lesions, but failed to extract radiomics features [3, 20]. And another study developed and evaluated various machine learning model utilizing non-invasive clinical parameters for the classification of endometrial non-benign lesions in postmenopausal women, and the Random Forest demonstrated superior recognition capabilities compared to other models [26]. However, the radiomics features were not included. Clinical and demographic characteristics can also be used in assisting diagnosis. The clinical risk factors for malignancy in our study were retrospective. A systematic review of published literature associated these clinical risk factors with endometrial malignancy lesions based on different study populations, sample sizes, and demographic characteristics [9, 27]. For example, Friberg et al. [28] conducted a meta-analysis including sixteen studies and found that diabetes was associated with an increased risk of endometrial cancer (relative risk [RR] 2.10, 95% confidence interval [CI] 1.75–2.53). Additionally, recent study suggests obesity is an independent predictor of endometrial cancer. The occurrence of endometrial carcinoma in women with a BMI greater than 30 kg/m² was four times higher than in normal women [9].

Ultrasound plays a significant role in the field of obstetrics and gynecology, especially in endometrial tissues. The transvaginal ultrasound examination is noninvasive, cheap, effective, well-tolerated, and can be widely used in clinical practice, and the anatomical characteristics and blood flow signals can be comprehensively scanned [29]. Our study extracted the structural and blood flow signal ultrasound features and ultrasoundbased radiomics features. Structural ultrasound features (e.g., endometrial midline appearance: non-linear and no “bright edge” sign) reflect the lesion structure well, while blood flow parameters assess the myometrial invasion [3, 30, 31]. Ultrasound radiomics features reflecting the texture information and intra-lesion homogeneity marker of the lesions, have been demonstrated to be closely related to genetic and biological features of the lesions [32]. Ultrasound radiomics has been widely used in the field of cancer, including endometrial cancer [12, 13, 32]. An European multicenter study [12] shown that radiomics have some ability to predict high-risk endometrial cancer, and developed and external validated a clinical-ultrasound radiomics model for discriminating high-risk and low-risk endometrial cancers from other endometrial cancers. Huang et al. [13] established a nomogram integrating ultrasound radiomics features and clinical parameters for predicting the prognosis of patients with endometrial cancer. Therefore, our study tentatively explored building a comprehensive prediction model by integrating multi-modal, multi-dimensional, and comprehensive data.

In clinical applications, machine-learning algorithms offer vast potential for identifying benign and malignant lesions of endometrial cancer. They can assist clinicians in early diagnosis and improve accuracy and precision while reducing misdiagnoses and missed diagnoses [33]. Particularly, the Random Forest model may show superior performance in distinguishing complex benign and malignant lesions. Furthermore, machine-learning algorithms have broad applications in the diagnosis, prediction, and treatment decisions associated with other gynecological diseases [34]. For instance, they can be used in the early detection and prediction of gynecological malignancies such as breast and ovarian cancers, providing personalized treatment plans [35].

Although this study demonstrated excellent performance in benign and malignant lesion identification in endometrial cancer, there are potential biases and limitations. First, this is a large sample, single-center study, thus, multicenter and external validation is needed. Second, pathological diagnosis is still the gold standard, however, more features need to be incorporated to build a comprehensive model in early stage, especially using MRI. Third, data quality and sample size are crucial for the diagnosability of machine learning algorithms. Additionally, the interpretability of the algorithms remains a concern, as the black-box models’ (e.g., SVM) decision-making process is difficult to explain, potentially limiting their use in clinical applications [36].

Future improvement methods and research directions should include a multicenter study, optimizing data preprocessing, a larger sample size and external validation to ensure the quality and balance of data. Second, further feature selection and extraction methods can be explored, and domain knowledge should be incorporated to enhance feature representation. Finally, deep-learning methods to discover more complex features and patterns from large-scale data, research on model interpretability, and reliable support for algorithmic clinical decision-making should be considered.

5. Conclusions

In conclusion, our study identified clinical features, structural ultrasound features and ultrasound radiomics features from malignant endometrial lesions, and built multiple machine learning algorithms based on this data. Based on our internal validation result, the best models were the Random Forest models. Thus, we suggest using the Random Forest models that combine clinical and ultrasound features, which could significantly improve early diagnosis accuracy and assist clinical decision-making in clinical practice.

Availability of Data and Materials

The datasets analyzed during the current study are available from the corresponding author on reasonable request.

References

[1] Novac L, Grigore T, Cernea N, Niculescu M, Cotarcea S. Incidence of endometrial carcinoma in patients with endometrial hyperplasia. European Journal of Gynaecological Oncology. 2005; 26: 561–563.

[2] Brooks RA, Fleming GF, Lastra RR, Lee NK, Moroney JW, Son CH, et al. Current recommendations and recent progress in endometrial cancer. CA: a Cancer Journal for Clinicians. 2019; 69: 258–279. https://doi.org/10.3322/caac.21561.

[3] Lin D, Wang H, Liu L, Zhao L, Chen J, Tian H, et al. IETA Ultrasonic Features Combined with GI-RADS Classification System and Tumor Biomarkers for Surveillance of Endometrial Carcinoma: An Innovative Study. Cancers. 2022; 14: 5631. https://doi.org/10.3390/cancers14225631.

[4] Nah EH, Cho S, Park H, Kim S, Kwon E, Cho HI. Establishment and validation of reference intervals for tumor markers (AFP, CEA, CA19-9, CA15-3, CA125, PSA, HE4, Cyfra 21-1, and ProGRP) in primary care centers in Korea: A cross-sectional retrospective study. Health Science Reports. 2023; 6: e1107. https://doi.org/10.1002/hsr2.1107.

[5] Cerovac A, Habek D, Hrgović Z. Ultrasound Characteristics of Myometrial Invasion in Endometrial Carcinoma: A Prospective Cohort Study. Clinical and Experimental Obstetrics & Gynecology. 2024; 51: 50. https://doi.org/10.31083/j.ceog5102050.

[6] Capozzi VA, Merisio C, Rolla M, Pugliese M, Morganelli G, Cianciolo A, et al. Confounding factors of transvaginal ultrasound accuracy in endometrial cancer. Journal of Obstetrics and Gynaecology: the Journal of the Institute of Obstetrics and Gynaecology. 2021; 41: 779–784. https://doi.org/10.1080/01443615.2020.1799342.

[7] Jacobs I, Gentry-Maharaj A, Burnell M, Manchanda R, Singh N, Sharma A, et al. Sensitivity of transvaginal ultrasound screening for endometrial cancer in postmenopausal women: a case-control study within the UKCTOCS cohort. The Lancet. Oncology. 2011; 12: 38–48. https://doi.org/10.1016/S1470-2045(10)70268-0.

[8] Ambrosio M, Raffone A, Alletto A, Cini C, Filipponi F, Neola D, et al. Is preoperative ultrasound tumor size a prognostic factor in endometrial carcinoma patients? Frontiers in Oncology. 2022; 12: 993629. https://doi.org/10.3389/fonc.2022.993629.

[9] Wise MR, Gill P, Lensen S, Thompson JM, Farquhar CM. Body mass index trumps age in decision for endometrial biopsy: cohort study of symptomatic premenopausal women. American Journal of Obstetrics and Gynecology. 2016; 215: 598. e1–598. e8. https://doi.org/10.1016/j.ajog.2016.06.006..

[10] Kumar V, Gu Y, Basu S, Berglund A, Eschrich SA, Schabath MB, et al. Radiomics: the process and the challenges. Magnetic Resonance Imaging. 2012; 30: 1234–1248. https://doi.org/10.1016/j.mri.2012.06.010.

[11] Han H. Diagnostic value of dynamic enhanced multi-slice spiral CT in lymph node metastasis of cervical cancer and analysis of the causes of missed diagnosis. European Journal of Gynaecological Oncology. 2024; 45: 35–40.

[12] Moro F, Albanese M, Boldrini L, Chiappa V, Lenkowicz J, Bertolina F, et al. Developing and validating ultrasound-based radiomics models for predicting high-risk endometrial cancer. Ultrasound in Obstetrics & Gynecology: the Official Journal of the International Society of Ultrasound in Obstetrics and Gynecology. 2022; 60: 256–268. https://doi.org/10.1002/uog.24805.

[13] Huang XW, Ding J, Zheng RR, Ma JY, Cai MT, Powell M, et al. An ultrasound-based radiomics model for survival prediction in patients with endometrial cancer. Journal of Medical Ultrasonics (2001). 2023; 50: 501–510. https://doi.org/10.1007/s10396-023-01331-w.

[14] Yao F, Ding J, Hu Z, Cai M, Liu J, Huang X, et al. Ultrasound-based radiomics score: a potential biomarker for the prediction of progression-free survival in ovarian epithelial cancer. Abdominal Radiology (New York). 2021; 46: 4936–4945. https://doi.org/10.1007/s00261-021-03163-z.

[15] Feng Y, Wang Z, Xiao M, Li J, Su Y, Delvoux B, et al. An Applicable Machine Learning Model Based on Preoperative Examinations Predicts Histology, Stage, and Grade for Endometrial Cancer. Frontiers in Oncology. 2022; 12: 904597. https://doi.org/10.3389/fonc.2022.904597.

[16] Onan A. Biomedical Text Categorization Based on Ensemble Pruning and Optimized Topic Modelling. Computational and Mathematical Methods in Medicine. 2018; 2018: 2497471. https://doi.org/10.1155/2018/2497471.

[17] Binson VA, Thomas S, Subramoniam M, Arun J, Naveen S, Madhu S. A Review of Machine Learning Algorithms for Biomedical Applications. Annals of Biomedical Engineering. 2024; 52: 1159–1183. https://doi.org/10.1007/s10439-024-03459-3.

[18] Tonni G, Grisolia G. Simulator, machine learning, and artificial intelligence: Time has come to assist prenatal ultrasound diagnosis. Journal of Clinical Ultrasound: JCU. 2023; 51: 1164–1165. https://doi.org/10.1002/jcu.23512.

[19] Bhardwaj V, Sharma A, Parambath SV, Gul I, Zhang X, Lobie PE, et al. Machine Learning for Endometrial Cancer Prediction and Prognostication. Frontiers in Oncology. 2022; 12: 852746. https://doi.org/10.3389/fonc.2022.852746.

[20] Chen B, Wang P, He W, Yang P, Kong Z, Wang D, et al. Standardized IETA criteria enhance accuracy of junior and intermediate ultrasound radiologists in diagnosing malignant endometrial and intrauterine lesions. Ultrasound in Obstetrics & Gynecology: the Official Journal of the International Society of Ultrasound in Obstetrics and Gynecology. 2024; 64: 528–537. https://doi.org/10.1002/uog.29102.

[21] Leone FPG, Timmerman D, Bourne T, Valentin L, Epstein E, Goldstein SR, et al. Terms, definitions and measurements to describe the sonographic features of the endometrium and intrauterine lesions: a consensus opinion from the International Endometrial Tumor Analysis (IETA) group. Ultrasound in Obstetrics & Gynecology: the Official Journal of the International Society of Ultrasound in Obstetrics and Gynecology. 2010; 35: 103–112. https://doi.org/10.1002/uog.7487.

[22] Epstein E, Fischerova D, Valentin L, Testa AC, Franchi D, Sladkevicius P, et al. Ultrasound characteristics of endometrial cancer as defined by International Endometrial Tumor Analysis (IETA) consensus nomenclature: prospective multicenter study. Ultrasound in Obstetrics & Gynecology: the Official Journal of the International Society of Ultrasound in Obstetrics and Gynecology. 2018; 51: 818–828. https://doi.org/10.1002/uog.18909.

[23] Li S, Oh S. Improving feature selection performance using pairwise pre-evaluation. BMC Bioinformatics. 2016; 17: 312. https://doi.org/10.1186/s12859-016-1178-3.

[24] Huang Y, McCullagh P, Black N, Harper R. Feature selection and classification model construction on type 2 diabetic patients’ data. Artificial Intelligence in Medicine. 2007; 41: 251–262. https://doi.org/10.1016/j.artmed.2007.07.002.

[25] Al-Abdaly NM, Al-Taai SR, Imran H, Ibrahim M. Development of prediction model of steel fiber-reinforced concrete compressive strength using random forest algorithm combined with hyperparameter tuning and k-fold cross-validation. Eastern-European Journal of Enterprise Technologies. 2021; 5: 59–65.

[26] Lai J, Rao B, Tian Z, Zhai QJ, Wang YL, Chen SK, et al. Postmenopausal endometrial non-benign lesion risk classification through a clinical parameter-based machine learning model. Computers in Biology and Medicine. 2024; 172: 108243. https://doi.org/10.1016/j.compbiomed.2024.108243.

[27] Brunelli AC, Brito LGO, Moro FAS, Jales RM, Yela DA, Benetti-Pinto CL. Ultrasound Elastography for the Diagnosis of Endometriosis and Adenomyosis: A Systematic Review with Meta-analysis. Ultrasound in Medicine & Biology. 2023; 49: 699–709. https://doi.org/10.1016/j.ultrasmedbio.2022.11.006.

[28] Friberg E, Orsini N, Mantzoros CS, Wolk A. Diabetes mellitus and risk of endometrial cancer: a meta-analysis. Diabetologia. 2007; 50: 1365–1374. https://doi.org/10.1007/s00125-007-0681-5.

[29] Xu J, Rao X, Lu W, Xie X, Wang X, Li X. Noninvasive Predictor for Premalignant and Cancerous Lesions in Endometrial Polyps Diagnosed by Ultrasound. Frontiers in Oncology. 2022; 11: 812033. https://doi.org/10.3389/fonc.2021.812033.

[30] Liao YM, Li Y, Yu HX, Li YK, Du JH, Chen H. Diagnostic value of endometrial volume and flow parameters under 3D ultrasound acquisition in combination with serum CA125 in endometrial lesions. Taiwanese Journal of Obstetrics & Gynecology. 2021; 60: 492–497. https://doi.org/10.1016/j.tjog.2021.03.018.

[31] Huang ZR, Li L, Huang H, Cheng MQ, De Li M, Guo HL, et al. Value of Multimodal Data From Clinical and Sonographic Parameters in Predicting Recurrence of Hepatocellular Carcinoma After Curative Treatment. Ultrasound in Medicine & Biology. 2023; 49: 1789–1797. https://doi.org/10.1016/j.ultrasmedbio.2023.04.001.

[32] Guo Y, Hu Y, Qiao M, Wang Y, Yu J, Li J, et al. Radiomics Analysis on Ultrasound for Prediction of Biologic Behavior in Breast Invasive Ductal Carcinoma. Clinical Breast Cancer. 2018; 18: e335–e344. https://doi.org/10.1016/j.clbc.2017.08.002.

[33] Zhang Y, Wang Z, Zhang J, Wang C, Wang Y, Chen H, et al. Deep learning model for classifying endometrial lesions. Journal of Translational Medicine. 2021; 19: 10. https://doi.org/10.1186/s12967-020-02660-x.

[34] Iqbal MJ, Javed Z, Sadia H, Qureshi IA, Irshad A, Ahmed R, et al. Clinical applications of artificial intelligence and machine learning in cancer diagnosis: looking into the future. Cancer Cell International. 2021; 21: 270. https://doi.org/10.1186/s12935-021-01981-1.

[35] Zhang PY, Yu Y. Precise Personalized Medicine in Gynecology Cancer and Infertility. Frontiers in Cell and Developmental Biology. 2020; 7: 382. https://doi.org/10.3389/fcell.2019.00382.

[36] Rudin C. Stop Explaining Black Box Machine Learning Models for High Stakes Decisions and Use Interpretable Models Instead. Nature Machine Intelligence. 2019; 1: 206–215. https://doi.org/10.1038/s42256-019-0048-x.