Beyond visual inspection: can a multimodal machine learning model improve the preoperative differentiation of endometrial polyps from non-polypoid endometrial lesions?

Abstract

Objective

To develop a multimodal machine learning model that integrates clinical data and ultrasound features to improve the non‑invasive preoperative differentiation between endometrial polyps (EMPs) and non‑polypoid endometrial lesions (including polypoid hyperplasia and normal endometrial tissue).

Methods

This single‑center retrospective cohort study included 913 patients with a preoperative diagnosis of endometrial polyp confirmed by postoperative pathology. Based on pathological findings, patients were categorized into a typical polyp group (n = 601) and a non‑polyp group (n = 312, consisting mainly of polypoid hyperplasia). Clinical baseline characteristics, transvaginal ultrasound (TVUS) features, and hysteroscopic observations were collected. Independent predictors were identified through multivariate analysis. Multiple prediction models—including logistic regression, decision tree, random forest, and gradient boosting decision tree (GBDT)—were constructed and evaluated using a 70:30 split into training and test sets.

Results

No significant differences were observed between the two groups in baseline characteristics. Typical EMPs more frequently presented as focal masses on ultrasound (82.4% vs. 75.3%) and exhibited detectable intralesional blood flow (51.6% vs. 38.8%). Multivariate analysis identified endometrial thickness on ultrasound, the presence of ultrasound‑detected blood flow, and hysteroscopic visualization of vegetations as independent predictors. Among the tested models, the GBDT demonstrated the best discriminative performance, with an area under the receiver operating characteristic curve (AUC) of 0.727, outperforming the traditional logistic regression model (AUC = 0.599).

Conclusion

Multimodal machine learning models, particularly GBDT, offer a potentially useful auxiliary tool for distinguishing EMPs from polypoid hyperplasia. However, their current diagnostic performance remains insufficient to support standalone clinical decision‑making. These findings highlight the limitations of existing non‑invasive methods and reinforce the necessity of histopathological confirmation. Future efforts should focus on incorporating more objective quantitative radiomic features to advance preoperative risk stratification.

Introduction

Endometrial polyps (EMPs) are common intrauterine growths that may present with abnormal uterine bleeding or infertility, but are often discovered incidentally during transvaginal ultrasound (TVUS). While most EMPs are benign, a subset may harbor premalignant or malignant changes, particularly in postmenopausal women, which necessitates accurate preoperative identification to guide appropriate surgical management. Clinically, however, the preoperative diagnosis of EMPs is frequently confounded by the presence of other focal endometrial pathologies, notably polypoid endometrial hyperplasia—a non‑neoplastic localized proliferation of the endometrium that mimics polyps on both ultrasound and hysteroscopy. The overlapping imaging features between these two entities create a persistent diagnostic dilemma, as both often appear as well‑defined, echogenic, intracavitary lesions with variable vascularity. Although 3D ultrasound with Doppler imaging and saline infusion sonography have improved endometrial assessment, none of these modalities reliably distinguishes focal hyperplasia from true polyps [1, 2]. Moreover, subjective interpretation remains highly observer-dependent, contributing to diagnostic variability [3].

The clinical significance of this diagnostic uncertainty is substantial. EMPs affect approximately 11% of premenopausal women and up to 58.9% of postmenopausal women, representing a large patient population undergoing evaluation and treatment [2, 4]. Conversely, a considerable proportion of lesions suspected as polyps preoperatively turn out to be benign polypoid hyperplasia or normal endometrium upon pathological examination—reflecting the limited specificity of current imaging criteria. This mismatch exposes a significant number of patients to unnecessary hysteroscopic surgery, with its attendant risks of bleeding, perforation, infection, and anesthesia, as well as increased healthcare costs and patient anxiety. Conversely, the misclassification of premalignant or malignant lesions as benign polyps could delay appropriate oncologic treatment. Thus, an accurate, non-invasive method to differentiate these entities before surgery would have direct clinical utility in reducing unnecessary procedures and improving risk stratification.

Existing diagnostic approaches rely heavily on subjective morphological evaluation. On TVUS, EMPs are classically described as hyperechoic, well‑defined masses, sometimes with a detectable feeding vessel, while endometrial hyperplasia typically presents as diffuse thickening [2, 5]. However, these features are neither sensitive nor specific. Notably, polypoid hyperplasia often presents as a focal lesion indistinguishable from a polyp, and conversely, patients with multiple polyps may exhibit diffuse endometrial irregularity that mimics hyperplasia. Furthermore, variability among observers and the natural biological diversity of the lesions add to the diagnostic challenges. These observations highlight that current morphological assessments—whether by ultrasound or hysteroscopy—are insufficient for reliable preoperative differentiation, and underscore the need for more quantitative, objective, and integrated diagnostic approaches.

Recently, machine learning has emerged as a promising tool in medical imaging, enabling the integration of diverse data types to improve diagnostic accuracy. By combining clinical parameters (e.g., menopausal status, BMI, symptoms) with quantitative and semi-quantitative ultrasound features, multimodal machine learning models have shown superior performance compared to conventional logistic regression in various gynecologic applications, such as ovarian cancer risk prediction and endometrial cancer detection [4, 6]. However, no such model has been developed specifically to address the challenge of distinguishing EMPs from polypoid hyperplasia [7, 8]. This gap is notable, as a well‑calibrated risk prediction tool could provide clinicians with an objective, reproducible risk score to complement subjective imaging interpretation, thereby facilitating more informed shared decision‑making regarding the need for hysteroscopic intervention.

Therefore, the aim of this study is to develop and validate a multimodal machine learning model that integrates clinical baseline data, TVUS features, and hysteroscopic observations to improve the preoperative non-invasive differentiation of EMPs from non-polypoid endometrial lesions (predominantly polypoid hyperplasia). We systematically evaluate multiple machine learning algorithms and compare their performance against traditional logistic regression, with the ultimate goal of providing a clinically useful adjunctive tool to reduce unnecessary invasive procedures and enhance preoperative risk stratification.

Methods

Study design and inclusion/exclusion criteria

This research was conducted as a retrospective, observational cohort study at a single center. Its objective is to systematically analyze the variations in preoperative clinical characteristics, ultrasound imaging attributes, and hysteroscopic observations between individuals diagnosed with typical EMPs and those with non-polypoid lesions, all of which were confirmed through surgical pathology. The study aims to develop and validate a predictive model for preoperative differential diagnosis based on these identified differences. The cohort was formed from consecutive cases treated at our hospital’s gynecology department from August 2024 to July 2025, where patients were surgically managed for a preoperative diagnosis of EMP and subsequently received a conclusive postoperative pathological diagnosis.

Participants in the study had to fulfill these criteria: (1) female adults aged 18 and above; (2) hospitalized with a pre-surgical diagnosis of EMP; (3) had hysteroscopy and endometrial tissue sampling performed; (4) pathological findings could be distinctly categorized into the established study groups; (5) all necessary data was fully accessible.

In order to maintain consistency within the study group and enhance the dependability of the analytical findings, stringent exclusion criteria were established: (1) existing genital or other types of cancers; (2) being pregnant or breastfeeding; (3) significant organ failure or serious systemic illnesses; (4) ambiguous pathological diagnosis.

Utilizing the gold standard for postoperative pathological evaluation, the selected cases were categorized into two distinct groups. The first category, referred to as the EMP Group, consisted of cases where the pathological assessment confirmed the presence of EMP, characterized by the microscopic identification of typical fibrovascular cores along with associated glands and stroma. The second category, known as the Control Group, comprised non-polypoid endometrial lesions, which mainly included polypoid endometrial hyperplasia and histologically normal endometrial tissue. This composition was chosen to replicate the real‑world diagnostic challenge encountered in clinical practice, where preoperative ultrasound or hysteroscopic findings suspicious for polyps may ultimately correspond to various non‑polypoid pathologies.

Data collection

Data for this research were gathered retrospectively using the hospital’s extensive electronic health record system. The methodology adhered to established protocols to guarantee precision and thoroughness. The information obtained covered four fundamental areas.

Data regarding sociodemographic factors and initial clinical traits were gathered. This encompassed details such as the individual’s age, stature, mass (to determine body mass index), reproductive history, menopausal condition, and any clinical signs like irregular uterine bleeding.

A systematic extraction and evaluation of preoperative TVUS characteristics were conducted. To reduce the potential for observer bias, two seasoned senior sonographers, unaware of the final pathological findings, independently assessed all archived ultrasound images and initial reports. In cases where the two sonographers disagreed on any ultrasound feature, they were required to re‑examine the images together and reach a consensus through discussion. If disagreement persisted, a third senior sonographer (> 15 years of experience in gynecologic ultrasound) was consulted to provide the final adjudication. The quantitative and qualitative data extracted comprised: the detection of intrauterine fluid (indicating endometrial cavity separation), the thickness of the endometrial layer, the uniformity of the overall echogenicity of the endometrium, the identification of focal echoes (including their intensity), and the analysis of blood flow signals within any lesions present.

To ensure standardized data extraction and reproducibility, all ultrasonic characteristics were defined according to pre-specified criteria. Endometrial monolayer thickness was measured in the sagittal plane at the thickest point of the endometrial echo, from the echogenic interface between the endometrium and myometrium to the opposite endometrial surface, excluding any intracavitary fluid or lesions. Intrauterine fluid collection was defined as the presence of anechoic or hypoechoic areas within the endometrial cavity. Heterogeneous endometrial echo was defined as non‑uniform echogenicity of the endometrium on gray‑scale ultrasound, with areas of differing echotexture. An occupying signal (focal lesion) was defined as a well‑defined, echogenic structure protruding into the endometrial cavity, distinct from the surrounding endometrium; among these, hyperechoic was defined as echogenicity greater than that of the adjacent myometrium or endometrial tissue. Blood flow signal was defined as the presence of detectable color or power Doppler signals within the lesion or the adjacent endometrium. All ultrasound examinations were performed using a transvaginal prob and all measurements were performed on frozen images during real-time scanning.

Information regarding the surgical intervention was gathered. This included documenting detailed observations from the hysteroscopic procedure, such as changes in endometrial thickness, the quantity and visibility of any growths, and other pertinent features.

The diagnosis from the pathological report was obtained, providing the definitive foundation for categorizing the cases and conducting further analyses.

Ethics statement

The research adhered to the ethical guidelines outlined in the Declaration of Helsinki. The Ethics Committee at Ningbo No.2 Hospital, Wenzhou Medical University approved the study protocol (Ethics number: PJ-NBEY-KY-2025-193-01). This investigation involved a retrospective analysis of medical records, ensuring that patient privacy and interests were not compromised. All clinical information was fully anonymized, and the Ethics Committee exempted the requirement for informed consent. There was no involvement from patients or the public in the design, execution, reporting, or dissemination of this study.

Statistical analysis

Data handling was conducted utilizing Microsoft Excel 2010 (Microsoft Corporation, Redmond, WA, USA) with a system of double-entry validation. For statistical evaluations, SPSS version 26.0 (IBM Corp., Armonk, NY, USA) and the SPSSPRO online service (Version 1.0.11; Beijing QingSi Data Technology Co., Ltd.) were employed. All visual data presentations were generated using GraphPad Prism version 10.1.2 (GraphPad Software, San Diego, CA, USA).

Initially, the Shapiro-Wilk test was employed to evaluate the normality of continuous variables. Those that followed a normal distribution are reported as mean ± standard deviation, with independent samples t-tests used for comparisons between groups. Conversely, data that did not meet normality assumptions are expressed as median (interquartile range, IQR), and the Mann-Whitney U test, a non-parametric method, was utilized for group comparisons. For categorical variables, results are presented as counts (percentages), and comparisons between groups were conducted using either the Chi-square test or Fisher’s exact test, depending on the situation.

The analysis of data in this research employed a systematic approach. Initially, a univariate analysis was conducted to assess the differences in all gathered clinical, ultrasound, and intraoperative observation variables between the Polyp Group and the Control Group. Subsequently, variables that demonstrated statistical significance (P < 0.05) in the univariate analysis were included in a binary logistic regression model for further multivariate analysis. A forward stepwise technique utilizing the likelihood ratio was applied to filter variables, aiming to pinpoint independent predictive factors that differentiate the two lesion types, along with their odds ratios (ORs) and 95% confidence intervals (CIs).

Drawing from the essential factors identified in the previous analysis, this research developed conventional logistic regression models alongside four machine learning approaches: the decision tree using the CART algorithm, Extremely Randomized Trees (ExtraTrees), backpropagation neural networks (BPNN), Gradient Boosting Decision Trees (GBDT), and Random Forest. A comparison was made regarding the effectiveness of these five distinct classification prediction models.

Hyperparameter tuning was performed exclusively on the training set using 5-fold stratified cross-validation. Stratified cross-validation was chosen to preserve the class distribution across each fold, ensuring that the prevalence of the outcome remained consistent in both training and validation subsets. For each model, a grid search strategy was employed to exhaustively explore pre-defined hyperparameter ranges. The following parameters and search grids were tested:

• CART: maximum depth (3, 5, 10, None), minimum samples per split (2, 5), split criterion (gini, entropy).

• ExtraTrees: number of estimators (50, 100, 200), maximum depth (5, 10, None), minimum samples per split (2, 5).

• BPNN: hidden layer size (50, 100, 150), batch size (8, 16, 32), learning rate (0.001, 0.01, 0.1).

• GBDT: number of estimators (50, 100, 200), learning rate (0.01, 0.1, 0.3), maximum depth (3, 5, 7).

• Random Forest: number of estimators (50, 100, 200), maximum depth (5, 10, None), minimum samples per split (2, 5).

Before the modeling process, the complete dataset was divided randomly into two parts: a training set and a test set, with a distribution of 70% for training and 30% for testing. Each model underwent parameter optimization through cross-validation using the training data. Their performance was then assessed on a separate test set. For each model, we calculated sensitivity, specificity, and the area under the receiver operating characteristic curve (AUC) based on predictions on the test set. For the machine learning models, additional metrics such as accuracy, recall, precision, and F1-score were also utilized. All hypothesis tests conducted in this research were two-tailed, with a significance threshold set at a P-value of less than 0.05.

Results

Baseline characteristics of the study population

This research involved 913 patients who met the eligibility criteria, as depicted in Fig. 1. According to the findings from postoperative pathology, 601 individuals (65.83%) were classified as belonging to the Typical Polyp Group, whereas 312 individuals (34.17%) were placed in the Non-polypoid Group (which included 259 cases of polypoid endometrial hyperplasia (83.01%) and 53 cases of normal endometrial tissue (16.99%)).

Fig. 1.

1 Flow of case data collection. Abbreviations: EMP, endometrial polyps

Table 1 illustrates the differences in baseline demographic and clinical traits between the two groups. The Typical Polyp Group showed a notably reduced parity in comparison to the Non-polypoid Group (p < 0.05). However, there were no significant variations between the groups in terms of age, BMI, gravidity, menopausal status, occurrence of abnormal uterine bleeding (AUB), hypertension, or diabetes mellitus (all p > 0.05).

Table 1.

Baseline Characteristics of Patients in the EMP Group and Control Group

Characteristics	EMP (n = 601)	Control (n = 312)	Statistical Value	p value
Age (years), Mean ± SD	44.75 ± 12.11	43.95 ± 12.83	0.935	0.350
BMI (kg/m²), Mean ± SD	23.12 ± 2.39	22.90 ± 2.13	1.473	0.141
Gravidity, M (P25, P75)	2 (2, 3)	3 (2, 3)	-1.959	0.050
Parity, M (P25, P75)	2 (1, 2)	2 (2, 2)	-2.304	0.021
Menopausal status, n (%)	131 (21.8)	54 (17.3)	2.562	0.109
AUB, n (%)	241(40.1)	143 (45.8)	2.770	0.093
Hypertension, n (%)	69 (11.5)	24 (7.7)	3.222	0.073
Diabetes, n (%)	22 (3.7)	9 (2.9)	0.377	0.539

Abbreviations: AUB Abnormal uterine bleeding, BMI Body mass index, EMP Endometrial polyps

Comparison of preoperative ultrasound features

Marked variations were observed in the preoperative ultrasound characteristics between the two cohorts (refer to Table 2). The incidence of cases with a sonographically identified focal lesion was notably greater in the Typical Polyp Group (82.4%) compared to the Non-polypoid Group (75.3%, P < 0.05). For patients with confirmed lesions, the rate of those displaying hyperechoic traits was similar in both the Typical Polyp Group (88.9%, 440/495) and the Non-polypoid Group (p > 0.05). In terms of vascularity, the occurrence of detectable blood flow signals within the lesions was significantly higher in the Typical Polyp Group (51.6%) than in the Non-polypoid Group (38.8%, p < 0.001). However, the presence of intrauterine fluid and varied endometrial echogenicity did not show significant differences between the groups (both p > 0.05).

Table 2.

Ultrasonic characteristics and hysteroscopic findings of patients in the EMP group and control group

Characteristics	EMP (n = 601)	Control (n = 312)	Statistical Value	p value
Ultrasonic Characteristics
Endometrial monolayer thickness (mm), Mean ± SD	4.62 ± 2.26	4.21 ± 2.12	2.614	0.009
Intrauterine fluid collection, n (%)	61 (10.1)	33 (10.6)	0.041	0.840
Heterogeneous endometrial echo, n (%)	475 (79.0)	235 (75.3)	1.639	0.200
Occupying signal, n (%)	495 (82.4)	235 (75.3)	6.355	0.012
Hyperechoic, n (%)	440 (88.9)	211 (89.8)	0.133	0.715
Blood flow signal, n (%)	310 (51.6)	121 (38.8)	13.499	< 0.001
Hysteroscopic Findings
Endometrial thickening, n (%)	176 (29.3)	87 (27.9)	0.196	0.658
Neoplasm			10.495	0.005
Single, n (%)	281 (46.8)	145 (46.5)
Multiple, n (%)	298 (49.6)	140 (44.9)

Abbreviations: EMP Endometrial polyps

Interestingly, the ultrasound findings in the Polyp Group may reveal widespread changes in the endometrium, especially when multiple polyps are present. The average thickness of a single layer of the endometrium was higher in the Typical Polyp Group (4.62 ± 2.26 mm) compared to the Non-polypoid Group (4.21 ± 2.12 mm, p < 0.01). The occurrence of intrauterine fluid was rare in both groups, and no notable differences were found between them (p > 0.05).

Concordance between hysteroscopic and pathological diagnoses

The primary intraoperative observations for both cohorts are presented in Table 2. During direct hysteroscopic examination, surgeons noted single or multiple growths in 864 instances. Subsequent pathological evaluations revealed typical polyps in just 579 (67.1%) of these cases, while the other 285 (32.9%) were linked to conditions classified under the Non-polypoid Group (mainly polypoid hyperplasia).

Additionally, Table 3 presents the concordance between hysteroscopic findings and final pathological diagnoses. Notably, among the 49 patients who were hysteroscopically assessed as having thickened, irregular endometrium without clear vegetations, pathological examination subsequently confirmed typical polyps in 22 cases (44.9%). Conversely, among the 864 cases in which hysteroscopy identified single or multiple intrauterine growths, only 579 (67.1%) were pathologically confirmed as true polyps, while the remaining 285 (32.9%) were ultimately diagnosed as non-polypoid lesions, predominantly polypoid hyperplasia. These findings collectively indicate that exclusive reliance on hysteroscopic morphological evaluation carries a substantial risk of both misdiagnosis (classifying non-polyps as polyps) and underdiagnosis (missing true polyps when they present with atypical appearances). This discrepancy underscores that hysteroscopy, despite being a valuable diagnostic tool, cannot substitute for histopathological confirmation, particularly when the visual findings are ambiguous or atypical. These results reinforce the necessity of integrating objective, quantitative adjunctive tools—such as the multimodal predictive model proposed in this study—to improve preoperative risk stratification and reduce diagnostic errors.

Table 3.

Concordance between hysteroscopic findings and final pathological diagnoses

	EMP	Control	Total
Neoplasm	579	285	864
No Neoplasm	22	27	49
Total	601	312	913

Abbreviations: EMP Endometrial polyps

Univariate and multivariate logistic regression analysis of risk factors

In the univariate analysis, variables that showed statistical significance (p < 0.05) included parity, endometrial thickness measured by ultrasound, the presence of a lesion identified via ultrasound, detection of blood flow signals on ultrasound, and the presence of vegetations observed during hysteroscopy. These factors were subsequently included in a binary logistic regression model. The results indicated that greater endometrial thickness (OR = 1.088, 95% CI: 1.021–1.159), the ultrasound identification of a focal lesion (OR = 1.530, 95% CI: 1.098–2.133), the presence of blood flow signals detected by ultrasound (OR = 1.682, 95% CI: 1.273–2.221), and the observation of hysteroscopic vegetations (OR = 2.493, 95% CI: 1.395–4.455) were all significantly linked to the probability of typical EMPs. However, parity did not demonstrate a significant correlation (refer to Table 4).

Table 4.

Univariate logistic regression analysis of risk factors for endometrial polyps

Variable	β	Wald χ²	OR	OR 95% CI	p value
Parity	-0.188	3.336	0.828	0.677–1.014	0.068
Endometrial monolayer thickness	0.084	6.742	1.088	1.021–1.159	0.009
Occupying signal	0.425	6.304	1.530	1.098–2.133	0.012
Blood flow signal	0.520	13.397	1.682	1.273–2.221	< 0.001
Hysteroscopic Neoplasm	0.914	9.514	2.493	1.395–4.455	0.002

Abbreviations: CI Confidence Interval, OR Odds Ratio

A subsequent analysis using multivariate logistic regression with a forward stepwise approach, which included variables with a significance level of P < 0.05 from the univariate analysis, revealed three distinct risk factors associated with typical EMPs: an increase in single-layer endometrial thickness observed via ultrasound (OR = 1.070, 95% CI: 1.003–1.141, p < 0.05), the detection of blood flow signals on ultrasound (OR = 1.577, 95% CI: 1.187–2.094, p < 0.01), and the identification of polypoid growths during hysteroscopy (OR = 2.409, 95% CI: 1.339–4.332, p < 0.01) (refer to Table 5).

Table 5.

Multivariate logistic regression analysis of risk factors for endometrial polyps

Variable	β	Wald χ²	OR	OR 95% CI	p value
Endometrial monolayer thickness	0.068	4.236	1.070	1.003–1.141	0.040
Blood flow signal	0.455	9.895	1.577	1.187–2.094	0.002
Hysteroscopic Neoplasm	0.879	8.619	2.409	1.339–4.332	0.003
Constant	-0.677	4.305	0.508	-	0.038
Occupying signal	-	-	-	-	0.143

Abbreviations: CI Confidence Interval, OR Odds Ratio

Discriminatory performance of independent risk factors

A logistic regression model for discrimination was developed based on independent risk factors identified through multivariate analysis, which included endometrial thickness measured by ultrasound, the detection of blood flow signals via ultrasound, and the visualization of vegetations through hysteroscopy. The coefficients for this model are outlined in Table 5. Table 6 provides a detailed overview of the classification performance metrics—such as positive predictive value, negative predictive value, sensitivity, and specificity—for each individual factor, with optimal cutoff points established using the Youden index. Analysis of the receiver operating characteristic (ROC) curve revealed AUC values of 0.558 (95% CI: 0.519–0.596) for ultrasound endometrial thickness, 0.564 (95% CI: 0.525–0.603) for ultrasound blood flow signals, and 0.527 (95% CI: 0.487–0.567) for hysteroscopic visualization of vegetations. (refer to Fig. 2).

Table 6.

Optimal cut-off values and diagnostic performance of key indicators

Index	Cut-off Value	Youden Index	Positive Predictive Value	Negative Predictive Value	Sensitivity (%)	Specificity (%)
Endometrial monolayer thickness	> 4.1 mm	0.1245	1.29	0.78	55.07	57.37
Blood flow signal	YES	0.1280	1.33	0.79	51.58	61.22
Hysteroscopic Neoplasm	YES	0.0499	1.05	0.42	96.34	8.65

Fig. 2.

Receiver operating characteristic curves of single independent risk factor for discriminating endometrial polyps

Discriminatory performance of predictive models

A total of five unique predictive models were created, and their effectiveness was assessed on the test dataset utilizing the crucial variables identified. The performance metrics for each model are thoroughly presented in Table 7. The binary logistic regression model demonstrated a sensitivity of 54.91% along with a specificity of 60.90%.

Table 7.

Evaluation results of machine learning models

Model	Accuracy	Recall	Precision	F1-score	Sensitivity (%)	Specificity (%)
BPNN	0.669	0.669	0.660	0.569	44.76	67.63
Extratrees	0.885	0.885	0.886	0.885	59.57	65.71
GBDT	0.703	0.703	0.708	0.646	57.74	73.08
Random Forest	0.945	0.945	0.945	0.945	70.88	56.09
Decision Tree	0.888	0.888	0.891	0.887	66.22	57.69

Abbreviations: BPNN Back-Propagation Neural Network, GBDT Gradient Boosting Decision Tree

The evaluation of the machine learning model’s effectiveness was conducted using several key metrics: (1) Accuracy, which measures the ratio of correctly identified samples to the overall number; a greater value signifies superior performance. (2) Recall, representing the fraction of true positive instances that were accurately detected; higher values are desirable. (3) Precision, indicating the ratio of predicted positive instances that are actually positive; again, higher values are favored. (4) F1-score, which is the harmonic average of precision and recall, offering a comprehensive assessment when these two metrics are in opposition.

The analysis of the ROC curve for differentiating EMPs yielded the following AUC results: Logistic Regression: 0.599 (95% Confidence Interval: 0.561–0.637); BP Neural Network: 0.583 (95% Confidence Interval: 0.544–0.622); ExtraTrees: 0.691 (95% Confidence Interval: 0.655–0.727); GBDT: 0.727 (95% Confidence Interval: 0.693–0.760); Random Forest: 0.680 (95% Confidence Interval: 0.644–0.717); and Decision Tree: 0.700 (95% Confidence Interval: 0.665–0.736). (refer to Fig. 3).

Fig. 3.

Receiver Operating Characteristic Curves of Multiple Models for Discriminating Endometrial Polyps. All AUC values reported in this figure are derived from predictions on the independent test set (30% of the total dataset) (Abbreviations: GBDT, Gradient Boosting Decision Tree; AUC, area under the receiver operating characteristic curve)

Supplementary notes on special cases

During the review process, several notable special cases were omitted from the main statistical analysis. These cases included: one instance of a postmenopausal endometrial polyp that showed focal carcinomatous changes, with unusual characteristics observed in both preoperative ultrasound and hysteroscopy; two instances of endometrial clear cell carcinoma, both of which exhibited AUB and showed endometrial thickening with significant vascularity on ultrasound, resembling benign conditions; and eight instances of small submucosal myomas (less than 1 cm in diameter) that were initially thought to be polyps. Sonographically, these myomas appeared as intrauterine lesions with varying echogenicity (hypoechoic, isoechoic, or heterogeneous) and were seen hysteroscopically as thickened endometrium or vegetative growths, making them challenging to differentiate from polyps. A key distinguishing characteristic was the lack of detectable intralesional blood flow on color Doppler ultrasound. These cases highlight the difficulties in preoperative differential diagnosis and emphasize the critical importance of conclusive pathological assessment.

Discussion

Complexity of differentiating typical endometrial polyps from non-polypoid lesions

This research conducts a comparative examination of clinical and imaging data from patients with typical EMPs and non-polypoid lesions, such as polypoid hyperplasia and normal endometrium, highlighting the intricacies involved in their differential diagnosis. Unlike some earlier studies that identified age and BMI as distinguishing factors [3, 6], our results indicated no significant variations in fundamental clinical attributes, including age, BMI, menopausal status, and the occurrence of AUB between the two groups. The only significant distinction observed was a reduced parity in the polyp group, aligning with previous findings that suggest nulliparity may elevate the risk of developing EMPs [9]. These outcomes imply that depending solely on standard data like age and symptoms is inadequate for effectively differentiating between these lesion types in clinical settings, emphasizing the need for enhanced imaging evaluations.

Evaluating ultrasound characteristics offers a more effective foundation for making differential diagnoses. Earlier studies have primarily concentrated on differentiating EMPs from atypical hyperplasia or endometrial carcinoma, as both conditions may manifest as thickening of the endometrium or localized intrauterine irregularities on TVUS [5, 10, 11]. Reports indicate that the agreement between observers for ultrasound identification of benign EMPs and endometrial cancer stands at 73.3% and 82.4%, respectively; however, this information does not specifically tackle the challenges in distinguishing polyps from simple hyperplasia, especially the focal type [3]. While the use of three-dimensional ultrasound along with Doppler imaging assists in differentiating typical polyps, submucosal myomas, and widespread endometrial hyperplasia [10], its effectiveness in distinguishing focal hyperplasia from polyps remains inadequately supported by existing literature.

The research indicated that standard EMPs are more commonly identified as distinct focal masses (82.4% compared to 75.3%) and are more prone to show internal blood flow signals (51.6% versus 38.8%), which corresponds with the understanding of polyps as genuine neoplastic entities featuring a fibrovascular structure. Interestingly, the hyperechoic characteristic, typically regarded as a common ultrasound trait of polyps [1], did not provide significant differentiation between the two categories in our analysis. Additionally, a substantial number of cases in the non-polypoid category (75.3%) also appeared as focal masses on ultrasound, while the polyp category could display diffuse thickening of the endometrium akin to hyperplastic conditions.

The overlapping characteristics suggest that identifying a mass on ultrasound should not be viewed as a straightforward diagnostic measure. A thorough assessment of the lesion’s border clarity, its base appearance (whether narrow-based or pedunculated), and the blood flow pattern (such as a single dominant vessel) may provide more valuable distinctions than merely labeling it as a mass or hyperechoic. This highlights the need for enhanced technical and reporting standards in ultrasound evaluations. Additionally, our study found no significant differences between groups regarding the presence of intrauterine fluid and varied endometrial echogenicity, indicating their limited effectiveness as distinguishing factors.

Limitations of hysteroscopist’s view diagnosis

This research offers significant revelations regarding the use of hysteroscopy for identifying intrauterine abnormalities. Our analysis indicated that of the cases identified as vegetations during the procedure, merely 67.1% were confirmed as typical polyps upon pathological examination, with almost a third being non-neoplastic alterations such as polypoid hyperplasia. In contrast, in instances where hysteroscopy did not reveal any clear vegetations (only showing endometrial irregularities or thickening), 44.9% still contained pathologically verified polyps. These results underscore the challenges of depending exclusively on hysteroscopic visual evaluations.

It is important to emphasize that the designation of hysteroscopy as the gold standard for diagnosis arises primarily from its capacity to facilitate precise biopsy or resection while directly visualizing the area, which allows for a conclusive histopathological assessment, rather than solely from the reliability of its visual morphological evaluation [12]. Research indicates that although hysteroscopy greatly enhances the assessment of intrauterine conditions, confirmation through biopsy remains essential [13]. Notably, around 30.6% of cases suspected to be polyps during hysteroscopy were overlooked by dilation and curettage (D&C) biopsy, and all submucosal myomas detected via hysteroscopy were not identified through D&C [14], highlighting the critical need for direct visual sampling to achieve a definitive diagnosis.

In situations where preoperative evaluations are ambiguous or there is clinical concern, hysteroscopy’s primary benefit is its ability to direct accurate histopathological sampling. It is essential to perform systematic or targeted biopsies of any suspicious regions, even if their visual characteristics are unusual, to ensure that small, flat polyps or localized malignant lesions are not overlooked. Our results advocate for a reclassification of hysteroscopy as a platform for precision biopsy guidance and sampling, with its clinical effectiveness heavily reliant on its connection to pathological diagnosis.

Predictive model performance evaluation and challenges in clinical translation

Utilizing the independent risk factors revealed through multivariate analysis—such as endometrial thickness, ultrasound blood flow signals, and hysteroscopic findings—this research developed and evaluated various predictive models. Nonetheless, the ability of these models to differentiate outcomes, including the traditional Logistic Regression model (AUC = 0.599) and more advanced machine learning approaches (like the GBDT model with an AUC of 0.727 and the Decision Tree model at AUC = 0.700), fell short of an excellent standard (AUC > 0.9). This finding has important consequences: given the current state of technology, no individual marker or combination of models can accurately and non-invasively differentiate between typical polyps and polypoid hyperplasia before a pathological assessment.

The investigation of these models is still significant. For instance, the GBDT model exhibited a relatively high level of specificity, whereas the Random Forest model was more sensitive. This indicates a promising avenue for future research into combining models to mitigate the risks associated with misdiagnosis and overlooked diagnoses. While these models offer an initial quantitative framework for assessing preoperative risks, their existing efficacy does not suffice for independent clinical decision-making. Nevertheless, we emphasize that the primary value of this model lies not in replacing pathological examination, but in providing a quantitative, standardized risk estimate that can complement clinical judgment. In practice, such a model could serve as a triage tool to reduce unnecessary hysteroscopies and to support more informed shared decision-making, particularly in settings where patient preference or comorbidities weigh against immediate invasive evaluation. As the model’s performance improves with future data integration, it may gradually decrease the number of benign biopsies performed, aligning with the broader trend toward precision medicine.

Caution from special cases

The retrospective examination of this research reveals that certain unique instances, while excluded from the main statistical evaluation, hold considerable clinical importance.

The identification of postmenopausal polyps exhibiting localized carcinomatous changes, along with atypical clear cell carcinoma of the endometrium, suggests that cancers can resemble the characteristics of benign polyps or hyperplastic conditions during imaging and hysteroscopic evaluations. This resemblance complicates the diagnostic process, particularly when ultrasound or hysteroscopic findings are unusual. This observation is consistent with earlier research indicating that tumors like clear cell carcinoma [15] and serous carcinoma [16, 17] can also manifest in a polyp-like form. Consequently, it is crucial to conduct thorough pathological assessments on all surgically removed intrauterine specimens, particularly those from postmenopausal individuals.

In this investigation, the differential diagnosis revealed that eight small submucosal myomas, initially thought to be polyps, exhibited no clear intralesional blood flow on color Doppler ultrasound prior to surgery. This observation indicates that the lack of blood flow signals could be a significant indicator for identifying submucosal myomas, particularly the smaller ones, as opposed to vascularized polyps in imaging studies. However, further research with larger sample sizes is needed to confirm its diagnostic relevance. This aligns with previous studies that have indicated early-stage myomas may be incorrectly identified as cystic lesions due to the absence of typical blood flow signals [18]. These insights expand the differential diagnostic framework for focal intrauterine lesions, highlighting the importance of distinguishing not only between polyps and various hyperplastic conditions but also considering other possibilities like submucosal myomas.

Limitations

This research has a number of constraints. To begin with, being a retrospective study conducted at a single center, it is prone to selection bias. Additionally, while experienced physicians assessed the ultrasound characteristics, some level of subjectivity remains. Furthermore, although we implemented a consensus protocol for discrepant readings, we did not formally calculate interobserver agreement (e.g., Cohen’s kappa) for the ultrasound features, which may affect the reproducibility of our findings and should be addressed in future prospective studies. Lastly, the instances of unique cases were few, and any associated findings should be approached with caution as they are merely observational indicators.

Future studies should focus on creating extensive prospective cohorts through collaboration across multiple centers. Utilizing quantitative analysis techniques like radiomics to extract hidden insights from ultrasound and hysteroscopic images, along with incorporating potential molecular pathological indicators, can lead to the development of advanced diagnostic and risk assessment systems. The primary objective is to combine various types of information for more accurate preoperative evaluation of lesions and to refine surgical criteria. This strategy aims to prevent unwarranted invasive procedures while facilitating the prompt detection and intervention of malignant lesions, thus progressing the diagnosis and treatment of endometrial conditions towards a more personalized and precise approach.

Conclusion

To summarize, this research thoroughly explores the intricate issues and real-world difficulties involved in distinguishing standard EMPs from non-polypoid endometrial lesions (predominantly polypoid hyperplasia). The findings reveal that the effectiveness of common clinical characteristics is restricted, ultrasound imaging features exhibit significant overlap, and visual assessments through hysteroscopy cannot be fully trusted. While predictive models that incorporate clinical and imaging characteristics provide some additional diagnostic assistance, there remains a disparity between their current ability to differentiate and the objective of achieving a highly accurate non-invasive diagnosis. These results highlight the necessity to reassess existing diagnostic procedures: there should be initiatives aimed at enhancing the standardization and precision of ultrasound descriptions and reports, as well as emphasizing that the primary function of hysteroscopy in diagnosis is to collect tissue samples rather than relying exclusively on visual morphological evaluations.

Abbreviation

AUB

Abnormal uterine bleeding

AUC

Area under the receiver operating characteristic curve

BPNN

Backpropagation neural network

BMI

Body mass index

CI

Confidence interval

D&C

Dilation and curettage

EMPs

Endometrial polyps

ExtraTrees

Extremely Randomized Trees

GBDT

Gradient boosting decision tree

IQR

Interquartile range

OR

Odds ratio

ROC

Receiver operating characteristic

TVUS

Transvaginal ultrasound

Authors’ contributions

All authors contributed to the study conception and design. Material preparation, data collection and analysis were performed by D.M, Z.Y., C.W., H.L. and H.Y. The first draft of the manuscript was written by D.M, Z.Y. and C.W. All authors read and approved the final manuscript. Y.Y. contributed to the study as supervision and critical revision of the manuscript for important intellectual content.

Funding

This study was not funded.

Data availability

The data analyzed in this study is subject to the following licenses/restrictions: All data generated or analyzed during this study are included in this article. Further enquiries can be directed to the corresponding author. Requests to access these datasets should be directed to Yingsha Yao, yaoyingsha@zju.edu.cn.

Declarations

Ethics approval and consent to participate

No animal studies are presented in this manuscript. Approval for the study protocol was granted by the Ethics Committee at Ningbo No.2 Hospital, Wenzhou Medical University (Ethics number: PJ-NBEY-KY-2025-193-01). The consent to participate was waived by the Institutional Review Board. All procedures performed in studies involving human participants were in accordance with the ethical standards of the Declaration of Helsinki.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.