Uterine sarcomas are a group of rare and aggressive malignant tumors originating from uterine mesenchymal tissue. Their annual incidence is approximately 1.7 per 100,000 women, and they account for 3-10% of all uterine malignancies. Despite being uncommon, they carry a disproportionately grim prognosis, driven largely by the difficulty of distinguishing them from uterine fibroids (leiomyomas), which are among the most common benign gynecologic tumors, affecting 40-80% of women over their lifetime. This clinical overlap is not minor: the two tumor types can be virtually indistinguishable on standard preoperative workup.
The morcellation crisis: The diagnostic stakes are high because uterine fibroids are frequently treated with minimally invasive laparoscopic surgery, often including power morcellation to facilitate removal through small incisions. In 2017, the U.S. Food and Drug Administration estimated that uterine sarcomas were present in approximately 1 in 225 to 1 in 580 women who underwent surgery for presumed uterine fibroids. When a sarcoma is morcellated without being recognized, tumor fragments disperse throughout the peritoneal and pelvic cavity, a process that dramatically accelerates disease progression, shortens progression-free survival, and worsens overall outcomes.
The current diagnostic gap: No reliable preoperative diagnostic criteria can definitively distinguish uterine sarcoma from other uterine tumors. The final diagnosis has historically required postoperative histopathological evaluation of resected tissue. Clinical symptoms (irregular vaginal bleeding, pelvic masses, and pelvic or abdominal pain) are shared with fibroids, and standard imaging findings also overlap substantially. The challenge of preoperative diagnosis motivates the entire body of research reviewed in this paper.
This 2022 review, published in Cancers and authored by Liu, Wang, and Rueda, synthesizes evidence from 96 articles selected from 3,480 search results in PubMed and Web of Science (spanning 2000-2022). It covers laboratory tests, imaging examinations, radiomics and machine learning methods, preoperative biopsy, integrated scoring models, and emerging molecular genetic imaging techniques, offering a comprehensive map of the current state of the field.
Uterine sarcomas are not a single disease. The three main pathological subtypes, leiomyosarcoma (LMS), endometrial stromal sarcoma (ESS), and adenosarcoma (AS), each have distinct biological behavior, prognosis, and diagnostic challenges. Carcinosarcomas were reclassified as dedifferentiated endometrial carcinomas by the International Federation of Gynaecology and Obstetrics (FIGO) in 2009 and are no longer grouped with uterine sarcomas in modern classification. Risk factors for uterine sarcoma include obesity, menopausal estrogen and progesterone use, oral contraceptives, history of pelvic radiotherapy, genetic defects, and tamoxifen use. There are also ethnic disparities: black women have a higher incidence of uterine sarcoma than white women.
Leiomyosarcoma (LMS): LMS is the most common subtype, accounting for more than 60% of uterine sarcoma cases. The average onset age is 48 years. Tumors are mostly single large masses arising from the myometrium or subserosal layer, with soft fish-like cut surfaces on gross pathology. LMS is the subtype for which preoperative differentiation from leiomyoma is most consequential, given the high overlap in imaging and clinical presentation.
Endometrial stromal sarcomas (ESS): ESS is further subdivided into low-grade (LG-ESS), high-grade (HG-ESS), and undifferentiated (UUS) variants. LG-ESS is the second most common subtype and is less malignant; more than 50% of cases occur in premenopausal women, and the ovary is the most common extrauterine site. HG-ESS was reintroduced as a distinct entity in the 2014 WHO classification, with a mean onset age of 50 years and an intermediate prognosis between LG-ESS and UUS. UUS is highly heterogeneous, occurs mainly in postmenopausal women, and lacks specific histological markers, requiring diagnosis by exclusion with extensive pathological sampling.
Adenosarcoma (AS): Adenosarcoma is a mixed tumor of low malignant potential composed of benign glandular epithelium and low-grade sarcoma. The vast majority (85%) arise from the endometrium; the remainder arise from the cervix or extrauterine sites. Tumors are typically polypoid with an average diameter of 5 cm. The wide age range of onset and the relatively indolent behavior distinguish AS from LMS and HG-ESS.
Multiple serum biomarkers have been investigated for preoperative differentiation of uterine sarcoma from uterine fibroids, with lactate dehydrogenase (LDH) emerging as the most clinically relevant despite limitations in specificity. Cancer antigen 125 (CA125) was one of the earliest markers studied; early reports suggested elevated preoperative CA125 levels in sarcoma, but subsequent work found significant overlap with early-stage disease, and CA125 is now considered a poor predictor for this indication.
LDH and its isoforms: LDH has more diagnostic utility than total CA125. A Chinese study found LDH 185 U/L to be an independent predictor of LMS. Goto et al. showed that combined LDH elevation plus contrast-enhanced MRI (CE-MRI) achieved high sensitivity and specificity for LMS. More precise discrimination came from LDH isoform analysis: an Italian team developed the U.M.G. risk index (named after the University of Magna Graecia), combining LDH3 and LDH1 isoenzymes into an inverse algebraic formula. A U.M.G. index 29 is considered indicative of uterine sarcoma. Validation in 179 patients with uterine fibroids achieved 91.1% specificity, with a higher false-positive rate in obese women (85.5% specificity) than in non-obese women (95.1%). LDH combined with CE-MRI and PET-CT was found to further improve both sensitivity and specificity.
Neutrophil-to-lymphocyte ratio (NLR): The NLR reflects the systemic inflammatory response to malignancy. Multiple studies have reported that elevated NLR can contribute to preoperative differentiation of uterine sarcoma, with threshold values set at 2.1, 2.12, and 2.8 across different cohorts. In one study, NLR was found to be more powerful than serum CA125 for this purpose.
MicroRNA and emerging markers: A panel of seven candidate miRNAs was identified for preoperative identification of uterine sarcomas. The optimal combination of miR-1246 and miR-191-5p achieved an AUC of 0.83 for overall sarcoma versus fibroid discrimination, and notably reached an AUC of 0.97 specifically for LMS versus leiomyoma. Growth differentiation factor-15 (GDF-15), progranulin, and osteopontin were identified through Gene Expression Omnibus and The Cancer Genome Atlas (TCGA) database analysis and measured using a novel compact chemiluminescent immunoautoanalyzer (POCube). D-dimer and C-reactive protein (CRP) combined with LDH achieved 100% specificity and 100% positive predictive value for LMS when all three markers were positive simultaneously.
Imaging is the primary clinical tool for preoperative assessment of uterine masses. Each modality has characteristic strengths and limitations for distinguishing sarcoma from leiomyoma, and multiparametric combinations are more reliable than any single technique alone.
Ultrasound: Color Doppler ultrasonography is the preferred first-line screening modality, given its cost-effectiveness and wide availability. Classic ultrasound features of uterine sarcoma include large masses with heterogeneous structure, heterogeneous echoes, irregular cystic or necrotic areas, and vascular hyperplasia, based on a multicenter analysis from 13 Italian and Spanish ultrasound centers covering 183 sarcoma patients over 10 years. An intra-mass resistance index (RI) less than or equal to 0.40 has been proposed as a diagnostic criterion for sarcoma, but RI is suboptimal for masses with large liquid-dark areas that produce no internal blood flow signal, which are then easily misclassified as degenerating fibroids. A radiomics model applied to ultrasound images demonstrated good performance in predicting mesenchymal malignant lesions, but ultrasound cannot match MRI for soft-tissue characterization and should prompt MRI evaluation when sarcoma is suspected.
MRI: MRI offers superior soft-tissue resolution and multiparametric evaluation. Sarcoma-specific MRI features include high T2 signal with areas of central necrosis, high enhancement on CE-MRI that peaks early at 60 seconds (compared to the later, lower enhancement of fibroids), high signal on diffusion-weighted imaging (DWI), and low apparent diffusion coefficient (ADC) values. Combining the T2 ratio, tumor-myometrial contrast on T2, and tumor-myometrial contrast enhancement ratio achieved 100% sensitivity. A combination of ADC and the tumor-myometrial contrast ratio (TCR) using a defined formula achieved 100% sensitivity and 100% specificity in one series of 8 sarcoma and 95 leiomyoma cases. CE-MRI had significantly higher diagnostic accuracy and specificity than DWI alone.
CT and PET-CT: CT is often overlooked but can reveal ESS as hypointense lesions on non-contrast CT with significant enhancement in the arterial phase and persistent heterogeneous enhancement in venous and delayed phases. PET-CT carries the highest diagnostic accuracy of any single imaging modality: a meta-analysis found that using an SUVmax threshold of 7.5 gave 80.8% sensitivity and 100% specificity, effectively excluding most fibroids. An SUVmax threshold of 4.4 achieved a negative predictive value (NPV) of 100%. PET-CT texture analysis combining traditional image features and SUV values achieved 100% sensitivity, 94% specificity, and 95% accuracy. The novel tracer 18F-FLT (fluorothymidine), which measures tumor proliferation, outperformed standard 18F-FDG for distinguishing uterine sarcoma from leiomyoma in early data, though it is not yet in routine clinical use.
Beyond standard MRI sequences, several emerging quantitative MRI techniques have shown high diagnostic performance for uterine sarcoma identification. These methods extract tissue properties that the naked eye cannot resolve, bridging conventional imaging and computational analysis.
Perfusion-weighted imaging (PWI) with machine learning: PWI characterizes tumor blood perfusion by tracking contrast agent distribution over time. When used alone, PWI parameters cannot reliably differentiate benign from malignant uterine lesions. However, when 21 PWI features extracted from regions of interest were fed into a machine learning classifier, the system achieved 91.7% accuracy, 100% sensitivity, and 90% specificity for distinguishing uterine sarcoma from fibroids.
Magnetic resonance spectroscopy (MRS): MRS detects tissue metabolite signatures associated with malignant transformation. The proportion of choline and lipid peaks in malignant uterine lesions is significantly higher than in benign lesions. One study combining ADC values with MRS findings achieved 98.3% accuracy (95.1-100% confidence interval) for differentiating uterine sarcomas from fibroids. A separate study confirmed that choline and lipid-positive peaks on MRS are highly suggestive of uterine malignancy, with a sensitivity of 100%, specificity of 96%, positive predictive value of 92%, and negative predictive value of 100% for the lipid peak alone.
Susceptibility-weighted sequences (SWS): SWS exploits differences in tissue magnetic susceptibility and blood oxygen level-dependent effects. Signal voids on SWS were observed in all sarcoma tissues examined but in only 4% of uterine fibroids, yielding 97% accuracy, 100% sensitivity, and 96% specificity for the differentiation. A related technique, enhanced T2 star-weighted angiography (ESWAN), enables quantitative measurement of phase, T2*, and R2* values. In a two-reader study, ESWAN T2* AUC reached 0.961 for distinguishing sarcoma from degenerative leiomyoma.
Diffusion kurtosis imaging (DKI): DKI investigates water diffusion properties using a non-Gaussian distribution model, capturing tissue microstructural complexity beyond what standard DWI-ADC provides. In 13 sarcoma cases versus 26 leiomyoma cases, mean kurtosis (MK), axial kurtosis (Ka), and radial kurtosis (Kr) all differed significantly between groups, with AUC values of 0.93, 0.99, and 0.80 respectively. Mean diffusivity (MD), axial diffusivity (Da), and radial diffusivity (Dr) AUCs were 0.94, 0.97, and 0.90. These are promising results, but studies of DKI remain very small and require multicenter validation.
Radiomics extracts high-throughput quantitative features from medical images, including shape descriptors, intensity statistics, and texture measures that are imperceptible to human readers. These features are then used to train machine learning models that can classify uterine tumors beyond what visual interpretation allows. The most productive domain for sarcoma radiomics has been T2-weighted MRI and ADC maps, given their central role in standard diagnostic workup.
Machine learning with multiparametric MRI and PET-CT integration: Nakagawa et al. compared individual MRI parameters (AUC 0.68-0.80), the SUVmax of PET (AUC 0.85), and a multivariate logistic regression model combining both, which reached an AUC of 0.92 in distinguishing uterine sarcoma from leiomyoma. This multivariate model performance was comparable to that of board-certified radiologists (AUC 0.97 and 0.89). A subsequent study by the same group applied extreme gradient boosting (XGBoost) and achieved AUC 0.93, significantly outperforming two radiologists independently (AUC 0.80 and 0.68, p = 0.03 and p less than 0.001).
Malek et al. machine learning decision trees: Analyzing 13 features from multiparametric MRI using machine learning, Malek et al. constructed two decision tree models. The simple decision tree achieved accuracy of 96.2%, sensitivity of 100%, and specificity of 95%. The complex decision tree achieved accuracy, sensitivity, and specificity all of 100%, though such results require cautious interpretation given the small dataset sizes and risk of overfitting.
Radiomics combined with clinical variables: Wang et al. compared three model types in 53 sarcoma and 81 leiomyoma cases. A T2WI-based radiomics model alone achieved AUC 0.76; a clinical variable model alone reached AUC 0.79; and a combined clinical-radiomics model achieved AUC 0.91, significantly outperforming either component (p less than 0.05). The combined model also outperformed or equaled both radiologist readers (AUC 0.78 and 0.90 respectively). ADC map-based texture analysis identified entropy as the single most discriminating feature, with AUC 0.94 in 16 sarcoma versus 31 degenerative fibroid cases. Lakhman et al. found 16 texture features that differed significantly between LMS and atypical leiomyoma (p less than 0.001 to 0.036), with unsupervised clustering achieving accuracy 0.75.
When imaging and laboratory tests are inconclusive, preoperative tissue sampling is the most direct route to histological confirmation before surgery. However, the feasibility and accuracy of biopsy depends heavily on tumor location and subtype, and no technique achieves the diagnostic certainty of postoperative pathology.
Preoperative biopsy approaches: Diagnostic curettage, hysteroscopic endometrial biopsy, and ultrasound-guided puncture biopsy are the available routes. Because LMS arises predominantly in the myometrium, diagnostic curettage reaches the lesion in only 42.9% of cases. ESS, which can present as polypoid masses in the uterine cavity, is accessible in 83.3% of curettage cases. Hysteroscopy improves yield by allowing direct visualization and targeted biopsy of suspicious lesions, but still cannot reach deep myometrial tumors. Transvaginal or transabdominal fine-needle aspiration biopsy, in which 4-5 passes are made per tumor to reduce sampling error, was evaluated in 10 patients with imaging-suggestive malignancy without intraoperative evidence of seeding. A 3-year follow-up study of preoperative puncture biopsy confirmed the feasibility of reaching atypical myometrial tumors by vaginal ultrasound-guided biopsy.
AI-assisted pathology from biopsy images: For patients at risk of LG-ESS, tissue biopsy images were preprocessed with segmentation and staining normalization algorithms, and multiple classical machine learning and deep learning models were applied to classify tissue images as benign or malignant. The optimal model achieved AUC 0.87, suggesting that trained machine learning algorithms can assist pathological reading of difficult biopsy specimens.
Integrated scoring models: Clinical scoring systems integrate age, serum markers, imaging, and endometrial cytology into composite risk scores. The revised PREoperative sarcoma score (rPRESS), an update to the original Nagai et al. system, combines serum LDH, MRI findings, endometrial cytology, and patient age to improve diagnostic accuracy. The pLMS score (Koehler et al., 2019) for LMS is based on clinical features including abnormal uterine bleeding, menorrhagia, dysmenorrhea, suspicious ultrasound presentation, and tumor diameter, with a score greater than 1 confirming sarcoma and a score between minus 3 and 1 triggering additional investigations. However, an independent validation study concluded that pLMS was not reliable and did not recommend it for clinical use. Risk prediction models incorporating age, race, BMI, number of myomas, uterine size, and pelvic pain level have also been proposed, with uniformly unsatisfactory results across validations.
Molecular genetic imaging: A survivin promoter-based genetic imaging approach was demonstrated to accurately differentiate sarcomas from benign tumors in human cell lines and mouse LMS models. This technology pairs an imaging reporter gene with a complementary imaging agent to measure gene expression or protein interactions in vivo, and represents a candidate for future clinical development.
Despite impressive results from individual studies, the preoperative identification of uterine sarcoma remains an unsolved clinical problem. Several structural limitations prevent the existing body of research from generating the evidence base needed for clinical implementation.
Small single-center datasets and retrospective design: Almost all reviewed studies are retrospective single-center analyses with small sample sizes. Uterine sarcoma is rare, making large prospective cohorts difficult to assemble. Single-center models suffer from institutional bias in imaging protocols, staining procedures, scanner hardware, and patient demographics. Retrospective designs also prevent standardization of data acquisition, and follow-up times are often insufficient to capture long-term outcomes. The authors cite Pergialiotis's argument that broader inter-institutional and international collaboration will be required to overcome these structural limitations.
Lack of external validation: Most AI and radiomics studies in this domain report only internal cross-validation results. External validation on independent test sets is the minimum standard for clinical credibility, yet it is consistently absent from the literature in this area. Performance metrics from internal validation systematically overestimate real-world accuracy, and the gap between internal and external performance is expected to be significant given the small training set sizes involved.
Technical and clinical integration challenges: Even when individual diagnostic tools show high accuracy, clinical adoption requires integration across modalities. The current evidence base supports a layered diagnostic approach: ultrasound for initial screening, MRI (with multiparametric sequences where available) for suspicious masses, LDH and its isoforms as a routine adjunct laboratory test, and PET-CT when prior investigations remain inconclusive. Preoperative biopsy is reserved for selected cases with high clinical suspicion and imaging findings that remain ambiguous. Multidisciplinary collaboration, including gynecologists, radiologists, pathologists, and clinical data scientists, is essential for advancing this field.
Future research priorities: The authors emphasize three forward directions. First, the creation of a gynecologic cancer database to enable large-scale multi-institutional AI model training. Second, increased recruitment of talent in computer science and basic medicine to support hybrid clinical-computational research programs. Third, continued development of noninvasive tissue characterization tools, particularly multiparametric MRI radiomics combined with molecular biomarkers (miRNA, LDH isoforms, GDF-15), as an integrated diagnostic panel that does not require preoperative tissue sampling. Newer imaging agents such as 18F-FLT PET and survivin promoter-based molecular imaging represent longer-term translational opportunities.