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.

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.

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.

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

SSL, JLW and JH designed the study and acquired the data. LZ, JLW, HW and QXA analyzed ultrasound recordings. SSL worked on clinical preprocessing and machine learning process and drafted the manuscript. XYW analyzed and interpreted the data. XYW, JH and QXA revised the manuscript. All authors revised this draft, read, and approved the final manuscript. All authors have participated sufficiently in the work and agreed to be accountable for all aspects of the work.

The study was conducted in accordance with the Declaration of Helsinki, and the protocol was approved by the Ethics Committee of Central Hospital of Enshi Tujia and Miao Autonomous Prefecture (approval number: 2023-010-02). The retrospective data was from a medical database, and the participant’s identities were kept anonymous and the information in this study was kept strictly confidential. All subjects gave their informed consent for inclusion before they participated in the study.

This research received no external funding.

The authors declare no conflict of interest.