The overall research process, including patient screening, data collection, and model validation, is illustrated in Fig. 1. This study retrospectively reviewed the clinical records collected at the Affiliated Hospital of Putian University from August 2024 to June 2025. Sixty individuals diagnosed with uterine sarcoma and another 60 with uterine fibroids, all confirmed by pathological examination, were included. The inclusion criteria were as follows: (1) had a diagnosis of uterine sarcoma or fibroid confirmed by surgical pathology; (2) all patients underwent color Doppler ultrasound for subsequent analysis; (3) complete clinical and laboratory data were available for all patients to ensure the accuracy and reliability of the study; and (4) patients in all age groups, including both premenopausal and postmenopausal women, were included. The exclusion criteria were as follows: (1) patients with concomitant malignancies, which could interfere with the diagnosis and analysis of uterine sarcoma or fibroids; (2) long-term estrogen use, as estrogen may influence tumor development and cause bias; (3) patients diagnosed with nonprimary uterine sarcoma to ensure the specificity and consistency of the study cohort; and (4) patients who did not undergo ultrasound examination or who had significant missing clinical or laboratory data to ensure data validity and scientific rigor. The diagnoses of all the patients were independently assessed by two senior pathologists in a blinded manner and confirmed according to the 2020 World Health Organization (WHO) classification of female genital tumors [15]. For the identification of sarcomatoid changes, the following morphological criteria were strictly followed: (1) prominent nuclear atypia (increased nuclear-cytoplasmic ratio, mitotic count $\ge$10/10 high-power fields [HPF]); (2) tumor cells arranged in a spindle or epithelioid pattern with disarray; (3) presence of pathological mitoses; and (4) associated with tumor necrosis. In cases of diagnostic disagreement between the two pathologists, a third senior pathologist was consulted to make the final decision.

Fig. 1.

Flowchart of this study.

This retrospective, single-center case-control study followed the STROBE guidelines and employed a 1:1 matched design to minimize selection bias, carefully matching 60 uterine sarcoma patients and 60 uterine fibroid patients based on age ($\pm$3 years), menopausal status, and disease duration ($\pm$1 year). The primary outcome was the diagnostic performance of a nomogram integrating clinical, laboratory, and ultrasound features in differentiating uterine sarcoma from fibroids, with key endpoints including diagnostic accuracy, sensitivity, specificity, and area under the receiver operating characteristic (ROC) curve (AUC). Matching was performed by an independent researcher blinded to the final diagnosis, and a standardized data collection protocol was implemented to ensure comprehensive and consistent data acquisition across all participants. The analysis of the ultrasound images was independently completed by 2 experienced sonographers. All ultrasound features were evaluated according to the standardized terminology and definitions published by the International Endometrial Tumor Analysis (IETA) [16]. For key parameters such as tumor margins, internal echoes, and blood flow distribution, the classification criteria in the Morphological Uterine Sonographic Assessment (MUSA) consensus were adopted to ensure the consistency and accuracy of the terminology used. For the radiological identification of sarcomatoid changes, the following criteria were used: irregular margins (defined as ill-defined or infiltrative growth), highly heterogeneous internal echoes (defined as a difference in echo intensity $>$50%), and an Adler blood flow grade $\ge$3 (characterized by diffuse vascular distribution with a resistance index [RI] 0.4). A consensus was reached through discussion for all controversial cases.

The data were categorized into three groups, namely, clinical information, laboratory indicators, and ultrasound features, and analyzed via Python (version 3.12.3; Python Software Foundation, Beaverton, OR, USA). Independent predictors in each group were identified through univariate and multivariate analyses. Four predictive models were constructed: clinical, laboratory, ultrasound, and combined models. The diagnostic performance of each model was assessed via ROC curves to determine the AUC, sensitivity, and specificity. The DeLong test was used to compare the predictive ability of the models. To rigorously evaluate the models’ generalizability and mitigate overfitting, we employed two internal validation techniques. First, tenfold cross-validation was performed by randomly partitioning the dataset into 10 equal-sized folds. The models were trained on 9 folds and validated on the remaining fold; this process was repeated 10 times, and the performance metrics were averaged to yield a robust estimate. Second, bootstrap validation with 1000 repetitions was conducted. For each repetition, a bootstrap sample of the same size as the original dataset was drawn with replacement to train the model, which was then evaluated on the original dataset. The optimism (i.e., the difference between the bootstrap performance and the apparent performance) was calculated for each repetition and averaged to obtain an optimism-corrected estimate of the models’ performance.

All the statistical calculations were conducted via Python (version 3.12.3; Python Software Foundation, Beaverton, OR, USA) and IBM SPSS Statistics (v 26.0; IBM Corp., Armonk, NY, USA). The kappa test was used to assess the diagnostic consistency among pathologists (Kappa = 0.82, p 0.001) and the evaluation consistency among sonographers (Kappa = 0.76, p 0.001), indicating a high level of consistency. Continuous variables exhibiting a normal distribution are presented as the mean $\pm$ standard deviation, whereas those not following a normal distribution are presented as the median and interquartile range. The chi-square test was applied to compare categorical data. For the three feature groups, both univariate and multivariate regression analyses were utilized. Statistical results with p values less than 0.05 were interpreted as significant.

The baseline characteristics of the study participants are summarized in Table 1. Successful matching was confirmed for age, disease duration, and menopausal status, as no significant differences were observed between the two groups (all p $>$ 0.05). However, significant differences were found in postmenopausal bleeding (p = 0.002), laboratory markers including lactate dehydrogenase (LDH) (p 0.001) and carbohydrate antigen 125 (CA125) (p 0.001), and ultrasound features such as maximal diameter (p = 0.010), margin (p 0.001), echogenicity (p = 0.003), cystic change (p = 0.003), and Adler blood flow grade (p 0.001). Calcification did not differ significantly between the groups (p = 0.462).

Univariate logistic regression analysis indicated that age (OR = 2.105, p = 0.002), menopausal status (OR = 5.222, p 0.001), and postmenopausal vaginal bleeding (OR = 14.03, p = 0.003) were significantly associated with sarcoma risk. Laboratory indices such as LDH (OR = 2.189, p 0.001) and CA125 (OR = 2.592, p 0.001) were also identified as significant predictors. Among the ultrasound features, echogenicity (OR = 0.415, p = 0.011), margin (OR = 7.094, p 0.001), sarcomatoid changes (OR = 4.239, p = 0.001), and maximal tumor diameter (OR = 1.246, p = 0.030) had a significant diagnostic impact (Table 2, Fig. 2).

Fig. 2.

Forest plot of the univariate logistic regression analysis. OR, odds ratio; CI, confidence interval; LDH, lactate dehydrogenase; CA125, carbohydrate antigen 125.

The outcomes from the multivariate logistic regression analysis are presented in Table 3 and Fig. 3. Multiple factors, including age (OR = 1.805, 95% CI: 1.225–3.749, p 0.001) and postmenopausal vaginal bleeding (OR = 8.182, 95% CI: 4.508–15.045, p = 0.002), were confirmed as independent predictors for sarcoma. In contrast, menopausal status was not an independent predictor in the multivariate model (OR = 0.565, 95% CI: 0.256–1.217, p = 0.802). Laboratory markers LDH (OR = 1.316, 95% CI: 1.107–1.606, p = 0.007) and CA125 (OR = 1.816, 95% CI: 1.153–2.394, p = 0.002) remained significant predictors. Among the ultrasound features, echogenicity (OR = 0.481, p = 0.003) and margin (OR = 6.514, p 0.001) were independently associated. In contrast, tumor morphology (OR = 1.015, p = 0.211) and maximal diameter (OR = 1.414, p = 0.574) were not significant in the multivariate analysis.

Fig. 3.

Forest plot of multivariate logistic regression analysis.

The diagnostic performance of various models is presented in Table 4 and Fig. 4. The apparent performance on the original dataset showed that the clinical model yielded an AUC of 0.764, indicating good discriminative ability; the laboratory model reached an AUC of 0.651; and the ultrasound model exhibited a higher AUC of 0.804. Importantly, the integrated clinical-laboratory-USA model achieved the highest AUC of 0.902, significantly outperforming all single models (all DeLong p 0.05).

Fig. 4.

Predictive performance of the four models. AUC, area under the curve.

To rigorously assess the models’ generalizability and correct for overoptimism, internal validation was performed (Fig. 5). The tenfold cross-validated AUCs for the clinical, laboratory, ultrasound, and combined models were 0.748, 0.635, 0.789, and 0.885, respectively. Furthermore, bootstrap validation with 1000 repetitions yielded optimism-corrected AUCs of 0.751, 0.640, 0.791, and 0.887 for the respective models. The combined model maintained the highest performance across all validation methods, demonstrating its robust diagnostic ability.

Fig. 5.

Predictive performance and internal validation of the models. (A) ROC curves from tenfold cross-validation. The lines represent the mean ROC curve for each model averaged over the 10 folds. The combined model (AUC = 0.885) demonstrates superior discriminative ability compared to the clinical (AUC = 0.748), laboratory (AUC = 0.635), and ultrasound (AUC = 0.789) models. (B) Bootstrap-validated ROC curve of the combined model. The solid line represents the optimism-corrected mean ROC curve from 1000 bootstrap repetitions (AUC = 0.887). The shaded area represents the 95% confidence interval, illustrating the model’s stable performance. ROC, receiver operating characteristic; AUC, area under the curve.

The core objective of this study is to construct a cost-effective, easily implementable preoperative triage tool for resource-limited settings. Although MRI is highly sensitive, it has significant limitations in primary care scenarios, including high costs, poor equipment availability, and lengthy examination times. In contrast, our nomogram integrates clinical features, laboratory markers, and ultrasound parameters to provide a rapid and economical first-line screening method. This approach aims to identify high-risk patients who can then be referred for more precise MRI evaluation.

Our approach aligns with recent literature advocating the need for efficient preliminary diagnostic tools [17, 18, 19, 20]. For example, in resource-constrained areas, the low availability of MRI can lead to diagnostic delays. The nomogram, as a multidimensional integration model, can prioritize high-risk cases and optimize the allocation of medical resources. This stepwise strategy (initial screening with the nomogram followed by MRI confirmation) has been proven to enhance diagnostic efficiency and reduce unnecessary invasive procedures. This stepwise diagnostic strategy, initial risk stratification with the nomogram followed by targeted MRI confirmation in high-risk cases has the potential to enhance overall diagnostic efficiency, reduce unnecessary invasive procedures, and ensure timely referral for complex cases.

By avoiding direct reliance on MRI, our model lowers the barriers to clinical implementation. Ultrasound and clinical data are more readily accessible and easier to implement in primary healthcare settings, which aligns with the global health equity concept. Future multicenter studies can further validate the complementary value of this model when it is used in conjunction with MRI.

In summary, the exclusion of MRI is based on a comprehensive consideration of practicality, cost-effectiveness, and accessibility. Our nomogram aims to bridge the gap between primary screening and advanced imaging, providing efficient support for clinical decision-making.

Comparing our study with the literature, we observed significant parallels and distinctions [21, 22]. One study developed a similar diagnostic model with an AUC of 0.82, which is comparable to our integrated model’s performance (AUC = 0.902) [23]. Our research extends this approach by integrating ultrasound characteristics, demonstrating improved diagnostic accuracy [24, 25, 26]. Our research extends this approach not only by integrating ultrasound characteristics but also by subjecting the model to more rigorous validation, thereby providing a more reliable estimate of its future performance. Unlike previous studies that relied primarily on single-modality assessments, our nomogram’s strength lies in its multidimensional data integration, which significantly enhances diagnostic precision. These comparisons underscore the innovative nature of our approach and its potential clinical utility in differentiating uterine sarcomas from fibroids.

Despite these promising results, this study has several limitations that should be acknowledged. First, its retrospective and single-center design may introduce selection bias and limit the generalizability of our findings. Although we employed a 1:1 matched design to mitigate this risk, future prospective, multicenter studies with larger sample sizes are warranted to validate and refine our nomogram.

Second, the relatively small sample size, particularly for rare uterine sarcoma cases, may affect the stability and statistical power of the predictive model. Expanding the cohort in future research would help enhance the model’s robustness.

Third, while the nomogram demonstrates high diagnostic accuracy, it is intended as a preoperative triage tool rather than a replacement for definitive diagnostic methods such as MRI or histopathological examination. Its performance in routine clinical practice needs further evaluation.