Model Based on Ultrasound Radiomics and Machine Learning for Differentiating Uterine Fibroids and Adenomyomas

Abstract

Objectives

To evaluate the value of radiomics based on ultrasound in differentiating uterine fibroids (UFs) and uterine adenomyomas (AMs) and construct a noninvasive tool for this differentiation.

Methods

The clinical data and ultrasound images of 659 patients diagnosed with UFs or AMs postoperatively were retrospectively analyzed at 3 institutions from January 2020 to December 2024. The cohort comprised patients with 422 UFs and 237 uterine AMs, divided into training (Institution 1 and Institution 2) and external test cohort (Institution 3). Radiomics features extracted from ultrasound images were used to develop models based on different machine learning classifiers. A combined model was developed combining radiomics features with clinical characteristics and a nomogram was depicted. The performance of the models was assessed by area under the receiver operating characteristic curve (AUC), calibration curve, and decision curve.

Results

The clinical model demonstrated moderate performance with AUCs of 0.767 (95% CI: 0.7259–0.8082) in the training cohort and 0.747 (95% CI: 0.6003–0.8933) in the test cohort. The radiomics model based on the support vector machine (SVM) showed the best performance in discriminating UFs and AMs, with AUCs of 0.899 (95% CI: 0.8721–0.9256) in the training cohort and 0.823 (95% CI: 0.7297–0.9159) in the external test cohort, respectively. The combined model presented better efficacy compared with the clinical model and the radiomics model, with AUCs of 0.923 (95% CI: 0.9015–0.9451) and 0.894 (95% CI: 0.8285–0.9596) in the training and external test cohorts, respectively. The calibration curves suggested good consistency and decision curves showed the highest overall clinical benefit for the combined model.

Conclusions

Ultrasound radiomics model based on SVM is feasible to differentiate UFs and AMs, and the combined model is a reliable and effective noninvasive tool for differentiation between UFs and AMs, which can assist clinicians in preoperative planning and fertility preservation strategies.

Abbreviations

AM

adenomyoma

AUC

area under the receiver operating characteristic curve

CEUS

contrast‐enhanced ultrasound

CI

confidence interval

GLCM

gray‐level co‐occurrence matrix

GLDM

gray‐level dependency matrix

GLRLM

gray‐level run length matrix

GLSZM

gray‐level size zone matrix

ICC

intraclass correlation coefficient

JZ

junctional zone

LASSO

least absolute shrinkage and selection operator

LR

logistic regression

MLP

multi‐layer perceptron

MRI

magnetic resonance imaging

MUSA

Morphological Uterus Sonographic Assessment

NGTDM

neighboring gray‐tone difference matrix

RF

random forest

ROI

region of interest

SE

strain elastography

SVM

support vector machine

SVM

support vector machine

TVUS

transvaginal ultrasound

UF

uterine fibroid

Uterine fibroids (UFs) are the most common benign tumors affecting women during their reproductive years, ¹ with cumulative incidence reaching approximately 70–80% by age 50. ² Adenomyosis refers to the presence of ectopic endometrial mucosa within the myometrium. Prevalence figures in the literature vary widely, from 5 to 70%, and surgical cohorts report an average of 20–30%. ³ Adenomyosis is also a prevalent gynecological disorder that significantly compromises the quality of life, typically presenting with clinical symptoms such as menorrhagia, dysmenorrhea, and infertility. ⁴ The detrimental impact of adenomyosis on fertility is multifactorial, involving impaired sperm transport, aberrant uterine peristalsis, chronic inflammation, and altered endometrial receptivity. ⁵ These factors lead to lower implantation rates and a higher likelihood of pregnancy loss during assisted reproductive technology cycles. ⁶ , ⁷ Adenomyosis is divided into 2 subtypes according to lesion distribution: focal disease, which frequently presents as adenomyomas (AMs), and diffuse disease. ⁸ Notably, although AMs account for only approximately 9.8% of all adenomyosis cases, ⁹ they are significantly associated with primary infertility. ¹⁰ While treatment for diffuse adenomyosis often focuses on symptom management and fertility preservation through medical or conservative interventional approaches, focal adenomyomas may be amenable to surgical resection. In addition, UFs can be treated with both conservative and surgical approaches based on patient conditions. ¹¹ Given the substantial differences in treatment strategies between UFs and AMs, accurate preoperative differentiation is crucial for formulating individualized treatment plans and avoiding inappropriate interventions.

Transvaginal ultrasound (TVUS) serves as the preferred initial imaging approach used to evaluate these disorders, owing to its accessibility and real‐time imaging capacity. However, the diagnostic accuracy of conventional B‐mode TVUS is highly operator‐dependent, as it requires subjective interpretation of sonographic features, contributing to considerable inter‐observer variability. Andres et al reported that 2D‐TVUS achieves reasonable sensitivity (83.8%) for detecting adenomyosis, but its specificity remains limited (63.9%), with significant variability across different sonographic features. Sensitivity was highest for heterogeneous myometrium (86.0%), whereas specificity peaked at 78.1% for globular uterus. These findings underscore the operator‐dependence and lack of standardization in conventional ultrasound assessment. ¹² Owing to its excellent soft tissue contrast, magnetic resonance imaging (MRI) is frequently used as both a problem‐solving tool and a diagnostic reference standard. Nevertheless, MRI is costly and less accessible. More importantly, its diagnostic performance can be variable and is not infallible. T2‐weighted imaging of the junctional zone (JZ) serves as the diagnostic cornerstone; however, this approach has notable limitations. The JZ thickness varies with the menstrual phase, hormonal status, and age, and can be mimicked by transient uterine contractions, potentially leading to misinterpretation. Furthermore, the sensitivity of MRI for detecting adenomyosis in the presence of coexisting fibroids can be suboptimal. ¹³ Ultrasound elastography has been explored as a potential quantitative adjunct tool, yet its application remains constrained by operator dependency, a significant learning curve, and potential artifacts. ¹⁴ , ¹⁵

Given these limitations, there is a clinical need for more objective and quantitative diagnostic approaches. Radiomics, which automatically extracts numerous quantitative features like heterogeneity and shape from medical images via high‐throughput analysis, can help to reveal crucial information of neoplastic internal physiology without being influenced by observer subjectivity. ¹⁶ , ¹⁷ Machine learning encompasses various computational techniques and algorithms that leverage informative features extracted from medical images, and this technology has seen extensive use in radiology. ¹⁸ , ¹⁹ As a non‐invasive methodology, radiomics holds significant promise for precision diagnosis in oncology. The application of ultrasound‐based radiomics in gynecological diseases has primarily centered on the benign‐malignant differentiation of ovarian masses. ²⁰ , ²¹ , ²² , ²³ However, few studies have investigated the efficacy of machine learning approaches for radiomics‐based differential diagnosis between UFs and AMs. One study has developed a preoperative ultrasound radiomics model for differentiating UFs and AMs; however, the absence of external validation, coupled with a single‐center design and modest sample size, leaves its generalizability unconfirmed. ²⁴ Therefore, we hypothesized that a radiomics model, developed from multi‐center ultrasound data and integrated with clinical risk factors, could provide a generalizable tool for preoperatively differentiating UFs from AMs.

Our study aims to explore the value of radiomics for the classification of UFs and AMs and construct a nomogram combining clinical features and radiomics scores for differentiation in individuals with UFs and AMs to guide clinical management.

Methods

Patients

This retrospective study received approval from the Ethics Committee and Institutional Review Board of the Second Affiliated Hospital of Nanchang University, which also granted a waiver of informed consent. From January 2020 to December 2024, a total of 339 patients diagnosed with AMs in the Second Affiliated Hospital of Nanchang University (Institution 1) and Jiangxi Provincial Maternal and Child Health Hospital (Institution 2) were initially included for the study and were retrieved from institutional databases. Additionally, 492 patients diagnosed with UFs postoperatively during the same period were randomly chosen for comparative analysis using the RAND function in an Excel sheet. A total of 218 patients diagnosed with UFs or AMs at Yichun Maternal and Child Health Hospital in Jiangxi Province (Institution 3) during the same period were included as external test data. For patients with multiple lesions, the largest lesion with a clear pathological diagnosis was selected as the target lesion for further analysis, ensuring 1 lesion per patient. Inclusion criteria: (1) Patients who underwent surgical excision of lesions for UFs or AMs. (2) Patients who received ultrasound evaluation for UFs or AMs within 1 month prior to surgery. Exclusion criteria: (1) Incomplete imaging or clinical data. (2) Ultrasound images of insufficient quality (eg, severe artifacts or low resolution). Due to the retrospective design, we could not consistently record the exact day of the menstrual cycle at the time of ultrasound examination and therefore could not control for it in this study. These inclusion and exclusion criteria led to the exclusion of a total of 149 UF patients and 121 AM patients from the combined initial pools of Institution 1 and 2. The primary reasons for exclusion were incomplete clinical data or suboptimal image quality precluding reliable radiomics analysis. Consequently, the final training cohort comprised 561 eligible patients (343 UFs and 218 AMs) from Institution 1 and 2. Using the same criteria, the external test cohort from Institution 3 included 98 patients (79 UFs and 19 AMs). Ultimately, 659 patients fulfilled all inclusion criteria and constituted the study sample. Of these, 561 patients (Institutions 1 and 2) were allocated to the training cohort. The remaining 98 patients (Institution 3) formed the test cohort. Figure 1 presents a flowchart summarizing the patient selection process.

Figure 1.

Flowchart of patient selection process. The flowchart illustrates the inclusion and exclusion criteria applied to patients from 3 institutions, resulting in a final training cohort of 561 patients (343 UFs and 218 AMs) from Institution 1 and Institution 2, and an external test cohort of 98 patients (79 UFs and 19 AMs) from Institution 3.

Baseline data was retrieved from the electronic medical record system for each patient, including demographic characteristics (age), B‐mode features, and Doppler features (pseudocapsule, location, shape, outline, echogenicity, shadowing, myometrial cysts, calcification, vascularity, amount of color). The core differential sonographic features of UFs and AMs for morphological and hemodynamic comparison are presented in Figure S1.

Image Acquisition

Ultrasound Image Acquisition

This retrospective, multicenter study utilized TVUS examinations performed by experienced sonographers. Standardized high‐frequency endocavitary probes (frequency range: 5–12 MHz) from multiple manufacturers (eg, Canon, Samsung, Mindray, GE) were employed across all 3 institutions following uniform pelvic imaging protocols to ensure comparability of image quality. A physician with 10 years of experience (physician 1) selected the most representative image for each lesion. Two physicians, with 5 and 8 years of experience, respectively (physician 2, 3), evaluated the following ultrasound features of lesions, including pseudocapsule, location, number, echogenicity, shape, outline, calcification, shadowing, myometrial cysts, vascularity, and amount of color. The definitions and criteria for these features were based on the standardized terminology proposed by the Morphological Uterus Sonographic Assessment (MUSA) consensus. ²⁵ In case of any disagreements, discussion was made by the physician with 10 years of experience. Pathological outcomes were concealed from all physicians during image assessment.

Image Segmentation

The workflow of this study is shown in Figure 2. All images were preprocessed before lesion delineation and were resampled using the same voxel size (1 × 1 × 1) to reduce errors introduced by different machines. ITK‐SNAP (version 3.8, http://www.itksnap.org/) was used to process all grayscale ultrasound images. Physician 2 delineated the region of interest (ROI), which covered all the lesions, by carefully outlining the margin of lesions on ultrasound images. In cases where a pseudocapsule was present, the outlining was performed along its inner edge. Because the pseudocapsule is a reactive fibrous tissue around UFs rather than the intrinsic tissue of the lesion itself, excluding it ensures that radiomics features are extracted only from the lesion tissue, thus avoiding confounding factors and improving the accuracy of feature analysis. To assess interobserver agreement, Physician 3 independently delineated the same 50 randomly selected images. To quantify interobserver agreement, we calculated the intraclass correlation coefficient (ICC).

Figure 2.

Workflow of the study. The main steps of this study are outlined here, starting from image acquisition and segmentation, followed by feature extraction and selection, then model construction, and finally performance evaluation.

Radiomics Procedure

Feature Extraction

Radiomics features were extracted from the manually segmented ROIs using Pyradiomics (http://pyradiomics.readthedocs.io/). These features fell into 3 categories: (1) geometric features, which mainly describe the morphology, size, and location of the lesion; (2) intensity features, which focus on the strength of the internal echogenicity of the lesion; and (3) texture features, which describe the distribution and pattern of internal echogenicity within the lesion, providing detailed information on its internal structure. Texture features characterize the second‐order and higher‐order statistical relationships between pixel intensities, including gray‐level co‐occurrence matrix (GLCM), gray‐level dependency matrix (GLDM), gray‐level run length matrix (GLRLM), gray‐level size zone matrix (GLSZM), and neighboring gray‐tone difference matrix (NGTDM).

Feature Selection

Only features demonstrating excellent interobserver reproducibility (ICC >0.75) were included in subsequent analysis. Firstly, all the ultrasound radiomics features were filtered by t‐test after Z‐value normalization, retaining only those with p < .05. Second, we used Pearson correlation analysis to detect and eliminate highly correlated features. For any pair of features with a correlation coefficient exceeding 0.9, only 1 was kept to reduce multicollinearity. Thirdly, least absolute shrinkage and selection operator (LASSO) regression was employed for multivariate feature selection. LASSO shrank the coefficients of the model by introducing a penalty term, which could completely shrink the coefficients of some unimportant features to zero. Ten‐fold cross‐validation was used to adjust the parameters of the penalty to decrease redundant information and avoid overfitting. Finally, the most significant features were filtered out.

Model Construction and Evaluation

Radiomics models were constructed by various classifiers, such as logistic regression (LR), random forest (RF), support vector machine (SVM), extra trees, multi‐layer perceptron (MLP), adaptive Boosting (AdaBoost), and extreme gradient boosting (XGBoost) in the training cohort. The final radiomics models were obtained by 5‐fold cross‐validation. Subsequently, the optimal radiomics model was identified after comparison of the performance. The predictive value for each patient in the model was assigned as Radscore.

Features such as demographic characteristics, grayscale ultrasound features, and Doppler features in the training cohort were selected using univariable and multivariable logistic regression analyses with a p‐value <.05 to develop the clinical model.

A combined model incorporating both the Radscore and the independently significant clinical variables was then constructed using the SVM classifier. A nomogram was created to visually display the combined model.

The discrimination of each model was evaluated by ROC. AUC with 95% confidence intervals (CIs), accuracy, sensitivity, and specificity were calculated. The DeLong test was used to assess the significance of the AUC. Hosmer‐Lemeshow analysis was used to assess the calibration ability of the models. Finally, decision curve analysis was used to estimate the clinical utility of each model.

Statistical Analysis

Model development was performed using Python (version 3.6, https://www.python.org). R software (version 4.2.1, http://www.r-project.org) was used for statistical analysis. Normality of continuous variables was tested using the Shapiro–Wilk test. Continuous variables obeying a normal distribution were displayed as mean ± standard deviation and compared using independent samples t‐test, while continuous variables not obeying a normal distribution were presented as median (interquartile values of 25 and 75%) and compared using the Mann–Whitney U test. Categorical variables were shown as count (percentage) and compared by chi‐square test or Fisher's exact test. Univariable and multivariable logistic regression analyses were used to screen out independent variables. The optimal cutoff value was selected using the Youden's index, and the best sensitivity and specificity were obtained at the maximum value of the Youden's index. A statistically significant difference was indicated by p‐value <.05.

Results

Clinical Baseline Characteristics

The key sonographic features of uterine fibroids and adenomyomas are presented in Figure S1. Additional sonographic features and detailed image acquisition notes are provided in the Supporting Information. This study performed a systematic univariable analysis of the patients' baseline clinical characteristics, aiming to preliminarily identify differentiating factors between UFs and AMs. The results are summarized in Table 1. Variables including age, vascularity, amount of color Doppler signal, lesion location, presence of a pseudocapsule, shape, margin definition, echogenicity, shadowing, myometrial cysts, and calcification were analyzed. In the training cohort, a statistically significant difference in age was observed between the 2 groups (p = .023), with UF patients (44.84 ± 8.30 years) being slightly but significantly older than AM patients (43.54 ± 5.67 years). Vascularity demonstrated highly significant intergroup differences in both the training and external test cohorts (p < .001 for both); 47.81% of UFs showed circumferential flow while 52.75% of AMs showed translesional flow in the training cohort; 59.49% of UFs had circumferential flow and 100% of AMs had translesional flow in the test cohort, suggesting that perfusion patterns are a crucial discriminative marker. The presence or absence of a pseudocapsule was strongly associated with the diagnosis. The absence of a pseudocapsule was significantly more frequent in AM cases (training cohort: 92.20% versus 74.34%; test cohort: 100% versus 50.63%; p < .001 for both), consistent with the infiltrative growth pattern characteristic of adenomyosis. Among other sonographic morphological features, lesion location, margin definition, shadowing, and the presence of myometrial cysts all showed statistically significant differences in the training cohort (p < .05). For instance, AM lesions demonstrated a predilection for the posterior uterine wall (AM group in training cohort: 66.51% versus UF group: 34.99%) and were more frequently associated with ill‐defined margins (AM group in training cohort: 72.48% versus UF group: 46.65%); the rate of Venetian blind sign in AM group (34.86%) was significantly higher than that in UF group (11.08%), and the incidence of myometrial cysts in AM group (21.56%) was higher than that in UF group (11.37%). However, lesion shape (overwhelmingly round or oval) and the presence of calcification did not reach statistical significance in this univariate comparison between the groups.

Table 1.

Baseline Characteristics of the Study Cohorts

Variables	Training Cohort (n = 561)			Test Cohort (n = 98)
	UF (n = 343)	AM (n = 218)	p	UF (n = 79)	AM (n = 19)	p
Age (years)	44.84 ± 8.30	43.54 ± 5.67	.023*	44.15 ± 8.16	45.32 ± 6.46	.564
Vascularity			<.001*			<.001*
Circumferential flow	164 (47.81)	35 (16.06)		47 (59.49)	0
Translesional flow	43 (12.54)	115 (52.75)		23 (29.11)	19 (100.00)
None	136 (39.65)	68 (31.19)		9 (11.39)	0
Amount of color			.094			.143
No color	96 (27.99)	45 (20.64)		11 (13.92)	0
Minimal color	226 (65.89)	150 (68.81)		53 (67.09)	12 (63.16)
Moderate color	16 (4.66)	17 (7.80)		11 (13.92)	6 (31.58)
Abundant color	5 (1.46)	6 (2.75)		4 (5.06)	1 (5.26)
Location			<.001*			.049*
Ant	128 (37.32)	42 (19.27)		21 (26.58)	5 (26.32)
Fund	27 (7.87)	11 (5.05)		8 (10.13)	0
Lt Lat	30 (8.75)	11 (5.05)		5 (6.33)	0
Post	120 (34.99)	145 (66.51)		34 (43.04)	7 (36.84)
Rt Lat	15 (4.37)	8 (3.67)		3 (3.80)	0
SM	23 (6.71)	1 (0.46)		8 (10.13)	7 (36.84)
Pseudocapsule			<.001*			<.001*
None	255 (74.34)	201 (92.20)		40 (50.63)	19 (100.00)
Yes	88 (25.66)	17 (7.80)		39 (49.37)	0
Shape			1.0			1.0
Round or oval	342 (99.71)	217 (99.54)		79 (100.00)	19 (100.00)
Irregular	1 (0.29)	1 (0.46)		0	0
Outline			<.001*			.004*
Well‐defined	183 (53.35)	60 (27.52)		40 (50.63)	2 (10.53)
Ill‐defined	160 (46.65)	158 (72.48)		39 (49.37)	17 (89.47)
Echogenicity			.027*			1.0
Hypo‐echogenicity	333 (97.08)	218 (100.00)		79 (100.00)	19 (100.00)
Hyper‐echogenicity	10 (2.92)	0		0	0
Shadowing			<.001*			1.0
None	305 (88.92)	142 (65.14)		64 (81.01)	17 (89.47)
Venetian blind sign	38 (11.08)	76 (34.86)		15 (18.99)	2 (10.53)
Myometrial cysts			.002*			1.0
None	304 (88.63)	171 (78.44)		77 (97.47)	18 (94.74)
Yes	39 (11.37)	47 (21.56)		2 (2.53)	1 (5.26)
Calcification			.311			1.0
None	334 (97.38)	208 (95.41)		76 (96.20)	18 (94.74)
Yes	9 (2.62)	10 (4.59)		3 (3.80)	1 (5.26)

Values are presented as median values with interquartile values of 25 and 75% in parentheses or number (%) unless otherwise indicated. p‐Value indicated the comparison between UF group and AM group within each cohort.

None (Vascularity), other patterns (e.g., punctate flow); Ant, anterior; Fund, fundal; Lt Lat, left lateral; Post, posterior; Rt Lat, right lateral; SM, submucosal.

^* p < .05.

Radiomics Features Selection and Model Construction

A total of 1560 features were extracted, including 14 geometric features, 306 intensity features, and 1240 texture features. Firstly, following interobserver agreement analysis (ICC >0.75, p < .05), all features were normalized and subjected to a 2‐sample t‐test (p < .05), resulting in 804 features with significant differences. Secondly, to eliminate multicollinearity among these 804 features, Pearson correlation analysis was performed, which further reduced the feature set to 45. Finally, 20 features with nonzero coefficients in Figure 3 were selected by LASSO regression for modeling. The radiomics features demonstrated distinct discriminatory capabilities between UFs and AMs. Wavelet‐domain texture features, for instance, primarily captured large, homogeneous hyperechoic regions within fibroids, which correspond to their dense fibrous tissue composition. In contrast, low‐gray‐level features were more indicative of cystic or hypoechoic areas within adenomyomas, reflecting underlying pathological changes such as glandular dilation and stromal edema. Shape‐based features, including original_shape_Elongation, effectively quantified the morphological irregularity characteristic of adenomyomas, while first‐order intensity features reflected the asymmetric distribution of echogenicity within uterine lesions. Furthermore, texture features derived from GLCM and GLDM, by analyzing pixel interdependencies and spatial patterns of echogenicity, successfully differentiated the regular internal texture of fibroids from the heterogeneous texture of adenomyomas. The SVM model showed best discrimination in training cohort (AUC 0.899, 95% CI: 0.8721–0.9256). It also showed good discrimination in test cohort (AUC 0.823, 95% CI: 0.7297–0.9159). The ROC curves of all the models based on machine learning were shown in Figure 4.

Figure 3.

Coefficient distribution of the 20 features selected by LASSO regression. Using LASSO regression, we identified 20 radiomics features with non‐zero coefficients as key predictors for distinguishing UFs from AMs. Features with positive coefficients were more associated with AMs, whereas those with negative coefficients were more indicative of UFs.

Figure 4.

ROC curve analysis of different radiomics models. Panel (A) shows the ROC curves for 7 machine learning classifiers in the training cohort, and panel (B) shows the corresponding curves in the test cohort. Among the classifiers tested—including LR, SVM, RF, Extra Trees, AdaBoost, MLP, and XGBoost—SVM achieved the best discriminative performance in both cohorts.

Clinical Model Construction and Nomogram

As shown in Table 2, finally, 6 clinical and ultrasound‐based variables were retained in the final clinical model following univariate and multivariate logistic regression analyses: age, lesion location, vascularity, amount of color Doppler signal, pseudocapsule, and shadowing. Among them, age was a significant negative predictor (OR = 0.945, p < .001) (negative predictor refers to the statistical term that the increase of age is associated with the decrease of the probability of AMs), suggesting that AMs tend to occur more frequently in younger patients. Lesion location (OR = 1.239, p = .001) was also significant, with AMs showing a predilection for posterior uterine locations. In addition, vascularity (OR = 1.286, p = .002) and amount of color Doppler signal (OR = 1.892, p < .001) were positively associated with AMs, consistent with their hypervascular imaging phenotype. Notably, the absence of a pseudocapsule (OR = 0.220, p < .001) was the strongest negative predictor for AMs, reflecting its infiltrative, non‐encapsulated growth pattern. Venetian blind sign (OR = 4.233, p < .001) was a strong indicator of AMs. These features contributed significantly to the discriminative power of the clinical model and were later incorporated into the combined radiomics‐clinical nomogram. The nomogram was plotted as well in Figure 5.

Table 2.

Univariable and Multivariable Logistic Regression Analysis of Clinical and Ultrasound Features

Variables	Univariate Analysis		Multivariate Analysis
	OR (95% CI)	p	OR (95% CI)	p
Age(years)	0.989 (0.986–0.993)	<.001*	0.945 (0.931–0.959)	<.001*
Pseudocapsule	0.193 (0.125–0.299)	<.001*	0.220 (0.134–0.360)	<.001*
Vascularity	0.875 (0.811–0.943)	.003*	1.286 (1.122–1.474)	.002*
Amount of color	0.829 (0.770–0.891)	<.001*	1.892 (1.468–2.440)	<.001*
Location	0.936 (0.885–0.989)	.050	1.239 (1.116–1.376)	.001*
Shadowing	2.000 (1.442–2.773)	<.001*	4.233 (2.861–6.265)	<.001*
Outline	0.987 (0.821–1.188)	.911	—	—
Myometrial cyst	1.205 (0.844–1.721)	.389	—	—
Echogenicity	0.000 (<0.001–>999)	.998	—	—
Shape	1.000 (0.098–10.237)	1.000	—	—
Calcification	1.111 (0.522–2.366)	.819	—	—

OR, odds ratio; CI, confidence interval.

^* p < .05.

Figure 5.

Nomogram of the combined model. This nomogram integrates 6 clinical and ultrasound features—age, lesion location, vascularity, amount of color, pseudocapsule, and shadowing—along with the Radscore. To use it, draw a vertical line from each variable to the Points axis, sum the points, and then draw a vertical line from the Total Points axis to obtain the predicted probability of AMs versus UFs.

Model Evaluation

As shown in Table 3, the clinical model demonstrated moderate predictive performance in the training cohort (AUC 0.767, 95% CI: 0.7259–0.8082) and test cohort (AUC 0.747, 95% CI: 0.6003–0.8933). The radiomics model exhibited improved predictive efficacy, achieving an AUC of 0.899 (95% CI: 0.8721–0.9256) in the training cohort and 0.823 (95% CI: 0.7297–0.9159) in the test cohort. The combined model further enhanced predictive performance, with an AUC of 0.923 (95% CI: 0.9015–0.9451) in the training cohort and 0.894 (95% CI: 0.8285–0.9596) in the test cohort in Figure 6. Delong's test revealed that the combined model showed statistically significant improvement compared to the clinical model in both the training cohort (p < .001) and the test cohort (p = .039). When compared with the radiomics model, the combined model also demonstrated a significant advantage in the training cohort (p = .004), and showed statistically significant improvement in the test cohort (p = .029). The calibration curves indicated good agreement between predicted probabilities and actual outcomes for all models in both cohorts in Figure 7. Decision curve analysis confirmed the superior clinical utility of the combined model across most threshold probabilities in Figure 8.

Table 3.

Diagnostic Performance of Clinical, Radiomics, and Combined Models

Model	Accuracy	AUC	95% CI	Sensitivity	Specificity	PPV	NPV	Cohort
Clinical	0.749	0.767	0.7259–0.8082	0.528	0.889	0.752	0.748	Train
Radiomics	0.813	0.899	0.8721–0.9256	0.908	0.752	0.700	0.928	Train
Combined	0.870	0.923	0.9015–0.9451	0.844	0.886	0.825	0.899	Train
Clinical	0.847	0.747	0.6003–0.8933	0.684	0.886	0.591	0.921	Test
Radiomics	0.673	0.823	0.7297–0.9159	0.947	0.608	0.367	0.980	Test
Combined	0.867	0.894	0.8285–0.9596	0.789	0.886	0.625	0.946	Test

AUC, area under the receiver operating characteristic curve; CI, confidence interval; PPV, positive predictive value; NPV, negative predictive value.

Figure 6.

ROC curves of the clinical, radiomics, and combined models. In the training cohort (A) and test cohort (B), the combined model—which integrated clinical characteristics with the Radscore—demonstrated superior discriminative ability compared to either the clinical model or the radiomics model alone.

Figure 7.

Calibration curves of the clinical, radiomics, and combined models. The curves illustrate the agreement between predicted probabilities and observed outcomes for the 3 models in the training (A) and test (B) cohorts. The dashed diagonal line indicates ideal calibration. Among all models, the combined model demonstrated the best calibration, with its predicted probabilities most closely aligning with the actual outcomes.

Figure 8.

Decision curve analysis of the clinical, radiomics, and combined models. The net clinical benefit of each model is shown across different threshold probabilities. In both the training cohort (A) and test cohort (B), the combined model provided the highest overall clinical utility.

Discussion

This study successfully developed and validated a differentiation model based on ultrasound radiomics features combined with multiple machine learning algorithms to differentiate UFs from AMs. Furthermore, a clinically practical nomogram was constructed by integrating the Radscore with clinical characteristics and ultrasound features including B‐mode ultrasound and Doppler parameters, providing an intuitive and non‐invasive tool for the differentiation between UFs and AMs.

To our knowledge, this is the first multicenter study to apply multiple machine learning classifiers for distinguishing UFs and AMs. Few studies have developed preoperative ultrasound radiomics models based on different machine learning algorithms to differentiate UFs and AMs. One such study, while reporting high AUC, was based on a single‐center cohort with a limited sample size, highlighting the need for multicenter validation such as ours. ²⁴ This methodological approach effectively addresses some limitations in existing research. Previous studies such as TVUS combined with strain elastography (SE), have reported favorable diagnostic performance but remain constrained by operator dependence and lack of standardization, essentially not transcending the realm of subjective visual assessment. ²⁶ Similarly, although single‐center studies have reported high efficacy for SE in differentiating UFs and AMs (with sensitivity of 100% and specificity of 96.23%), ¹⁵ or suggested the utility of shear‐wave elastography for distinguishing normal myometrium from adenomyosis, ²⁷ their findings require multicenter validation. Furthermore, these techniques face challenges in clinical adoption due to operator variability and artifacts. In recent years, while the application of artificial intelligence in gynecological ultrasound has increased, most research has focused on UFs detection or lesion segmentation, ²⁸ with intelligent studies specifically addressing the clinical challenge of differentiating UFs from AMs remaining relatively scarce. Research utilizing other modalities like contrast‐enhanced ultrasound (CEUS) has achieved a relatively high AUC (0.908) but with limited specificity (70%). ²⁹ Studies based on traditional morphological features (eg, border, lacunae) reported diagnostic sensitivity of approximately 82.0% and specificity of 88.3%, ³⁰ and those based on Doppler parameters (eg, pulsatility index) reported sensitivity of about 82% and specificity of approximately 84%. ³¹ While these provide valuable references, they fail to capture the deeper quantitative image information imperceptible to the naked eye. Although consensus guidelines like MUSA outline key sonographic differential points ²⁵ and reviews such as O'Shea's systematically summarize typical imaging differences, ³² these methods heavily rely on operator expertise, and diagnostic performance remains suboptimal. Compared to prior work, our radiomics model performs comparably to Jing et al's radiomics model (AUC 0.823 versus 0.854) in the test cohort ²⁴ and CEUS (AUC 0.899 versus 0.908) in the training cohort, and offers objective quantification lacking in elastography and Doppler‐based methods, with the added advantage of multi‐center external validation. Moreover, our study, based on multi‐center data and employing 7 machine learning algorithms, established a combined model achieving AUCs of 0.923 and 0.894 in the training and external test cohorts, respectively. This performance is significantly superior to models based on traditional ultrasound features, Doppler parameters, and CEUS models, demonstrating better generalizability and clinical value for differentiation.

The SVM‐based model performed optimally in this study. As a classical algorithm robust for high‐dimensional and sparse data, SVM effectively captures complex feature interactions through its kernel functions, offering strong generalization capabilities. ³³ This makes it particularly suitable for our radiomics analysis. In our study, the SVM‐based radiomics model achieved an AUC of 0.899 in the training cohort and 0.823 in the external test cohort, indicating excellent and stable diagnostic efficacy. In contrast, ensemble methods such as RF and XGBoost exhibited greater performance variability in external test, which may stem from their sensitivity to feature collinearity and dependence on hyperparameter tuning. Although LR offers strong interpretability, its linear limitations resulted in lower AUCs (0.767). As 2 other ensemble learning methods, AdaBoost and Extra Trees did not surpass SVM's performance in this study, possibly because their feature selection mechanisms did not achieve optimal discriminatory efficiency under the high‐dimensional and sparse characteristics of our dataset. Although MLP, as a neural network model, possesses the ability to fit complex patterns, its performance in this study also failed to reach optimal levels, which may be related to its high sensitivity to network architecture design and hyperparameter tuning, as well as the tendency to produce unstable training results with limited samples.

Radiomics features can be regarded as quantitative characterizations of tumor heterogeneity and can be instrumental in revealing the internal composition of lesions. ¹⁶ Our study demonstrated that ultrasound‐based radiomics features can effectively differentiate uterine UFs from AMs. Texture features, which quantify the spatial distribution and interdependencies of echogenicity within lesions, can serve as sensitive indicators of the internal structural organization of uterine myometrial lesions. ³⁴ , ³⁵ Among the 20 features used to develop the radiomics model, texture features accounted for a large proportion (18 features, 90%), which may be related to the tissue heterogeneity of uterine myometrial lesions, particularly the disordered internal architecture of adenomyomas. Meanwhile, the majority of the selected features were wavelet‐transformed texture features (9 features, 45%). This highlights the critical role of multi‐scale analysis in revealing the complex spatial heterogeneity of uterine lesions. ³⁵ , ³⁶ , ³⁷

In our study, we plotted a nomogram integrating statistically significant clinical features (eg, age, pseudocapsule) and Radscore for clinical use. We found age was inversely correlated with the likelihood of adenomyomas, aligning with its estrogen‐responsive nature in younger reproductive‐age women. This observation is further reinforced by evidence suggesting that adenomyosis can originate earlier in life, as seen in studies of adolescents. ³⁸ In our study, AMs were more commonly situated in posterior uterine segments, aligning with prior observations on their anatomical distribution. ²⁸ Hemodynamic parameters—vascularity and amount of color Doppler signal—were positively associated with AMs, reflecting increased vascular density and perfusion, which are often observed in AMs due to proliferative stromal components. These findings are consistent with prior studies using Doppler indices to assess vascular patterns in uterine pathology. ³¹ The absence of a pseudocapsule was a defining negative feature for AMs. This feature echoes the MUSA consensus criteria, where fibroids are typically encapsulated, well‐circumscribed, whereas AMs lack a discrete capsule and exhibit a more poorly defined appearance. ³⁹ Previous studies have suggested that myometrial cysts and Venetian blind sign may be a prominent feature in AMs, ⁴⁰ , ⁴¹ but myometrial cysts were not included in our model. The myometrial cysts, while more prevalent in AMs on univariable analysis (21.56% versus 11.37%, p = .002), were not retained as an independent predictor in the multivariate clinical model (p = .389). This likely reflects their statistical collinearity with more robust discriminators—such as the absence of a pseudocapsule and ill‐defined margins—which more directly capture the infiltrative growth pattern of AMs. Their moderate sensitivity and non‐specificity further limit their standalone discriminatory power in a multivariate context. The Venetian blind sign demonstrated a significantly higher prevalence in the AM group on univariable analysis (34.86% versus 11.08%, p < .001), and its independent predictive value for AMs was further confirmed in multivariable analysis (OR = 4.233, p < .001). This feature exhibited consistent results across both univariable and multivariable analyses, collectively indicating its significant association with AMs. Although the high PI value reported by Chiang et al was not directly included in our model, ³¹ the hemodynamic impedance differences it reflects might be intrinsically linked to the heterogeneity captured by radiomic texture features. The combined model achieved AUCs of 0.923 and 0.894 in the training and test cohorts, respectively, significantly outperforming single models and also exceeding the performance (AUC 0.908) reported by Zhang et al, ²⁹ indicating that multi‐feature integration effectively enhances discriminatory power.

This study has several limitations. First, the retrospective design may introduce selection bias, and the significant difference in prevalence between AMs and UFs might affect model fitting due to imbalanced sample distribution. Second, despite standardized image preprocessing, machine‐dependent bias from different ultrasound devices cannot be completely eliminated. Future analysis stratified by device model is warranted. Third, this study did not explore the potential of deep learning methods; subsequent work could attempt to integrate deep learning features upon expanding the sample size. Finally, this study was based solely on conventional grayscale and Doppler ultrasound; multimodal information such as CEUS or elastography was not included. Future research could aim to build multi‐modal radiomics models to further improve diagnostic performance. ⁴²

Conclusion

The quantitative ultrasound radiomics model developed based on SVM has the ability to accurately distinguish between UFs and AMs. In addition, the nomogram that integrated B‐mode features, Doppler features, and ultrasound radiomics features serves as a convenient visualization tool for differentiating UFs and AMs, which can assist surgeons in making correct decisions.

Supporting information

Data S1. Supplementary materials.

Figure S1. Representative ultrasound images showing key sonographic features of UFs and AMs. (A, B) Images from a 40‐year‐old woman with a uterine fibroid. (A) A well‐defined lesion with a distinct pseudocapsule is visible on grayscale imaging (white arrows). (B) Power Doppler reveals circumferential vascularity around the lesion (white arrows). (C) Grayscale image from a 33‐year‐old woman with an adenomyoma shows posterior acoustic shadowing, a finding consistent with the Venetian blind sign (white arrows). (D, E) Images from a 40‐year‐old woman with an adenomyoma. (D) The lesion appears ill‐defined, lacking a pseudocapsule and showing indistinct margins (white arrows). (E) Power Doppler demonstrates translesional vascularity within the lesion (white arrows). (F) Grayscale image from a 44‐year‐old woman with an adenomyoma reveals small myometrial cysts (white arrows). All images were acquired using transvaginal ultrasound probes (5–12 MHz). Patient identifiers and manufacturer logos have been removed.

Data Availability Statement

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.