A comprehensive magnetic resonance imaging-based model for predicting lymphovascular space invasion in endometrial cancer: a retrospective observational study

Introduction

Endometrial cancer is a heterogeneous disease with complex and diverse molecular profiles. Comprehensive staging is essential to stratifying patients and providing tailored treatments (1). In 2023, the International Federation of Gynecology and Obstetrics (FIGO) updated the endometrial cancer staging system to incorporate pathological and molecular variables beyond anatomic findings (2,3). Lymphovascular space invasion (LVSI) at the invasive front of endometrial cancer, which is a strong independent prognostic factor for pelvic regional recurrence, distant metastasis, and overall survival, was included in the staging system as a major update, especially in early-stage disease (4,5). According to the 2023 FIGO, cases of early disease with no LVSI or only focal LVSI are categorized as IA and IB, while the presence of substantial LVSI automatically upshifts the stage to IIB. In high-grade disease, LVSI is the strongest independent risk factor for lymph node metastasis and overall survival (6,7). Therefore, knowledge of the LVSI status is crucial for aligning prognosis with treatment decision-making in both low- and high-grade endometrial cancer. Unfortunately, as a histopathological finding, LVSI status can only be assessed after surgery. Therefore, there is an urgent need for minimally invasive approaches to preoperatively identify LVSI status.

With the advantages of excellent soft-tissue contrast resolution, multiplanar imaging, and absence of ionizing radiation, magnetic resonance imaging (MRI) plays a pivotal role in the management of endometrial cancer throughout the clinical workflow, from initial staging and treatment response evaluation to preoperative planning (8). The 2023 FIGO staging system introduces refined stratification criteria, thereby amplifying the importance of MRI in endometrial cancer staging. Integration of MRI with histopathological and molecular profiling enables comprehensive alignment with the enhanced requirements of this updated classification framework (9). This evolution also poses significant challenges for endometrial cancer diagnosis based solely on conventional MR images. Studies over the past decade have demonstrated significant variability in the sensitivity, specificity, and accuracy of MRI for diagnosing deep myometrial invasion, cervical stromal invasion, and lymph node metastasis in endometrial cancer, resulting in inconsistent diagnostic performance (10-13). Substantial interobserver and intersequence variability in MRI interpretation further compounds these limitations. Radiomics and artificial intelligence algorithms address have been employed to address these challenges by converting medical images into quantitatively analyzable data, extracting high-dimensional radiomic features that capture intratumoral heterogeneity and revealing tumor phenotypic characteristics inaccessible to conventional methods. The recent integration of artificial intelligence in radiology has helped address many imaging-based diagnostic challenges and has extended the applications of radiology (14,15). This approach provides an objective foundation for risk assessment, early detection, treatment planning, and prognosis prediction in patients with endometrial cancer (10). A recent meta-analysis including 15 studies and 3,608 patients with endometrial cancer demonstrated favorable diagnostic accuracy of MRI-based radiomics for evaluating tumor grade, deep myometrial invasion, LVSI status, and lymph node metastasis, providing robust evidence supporting the role of radiomics in preoperative risk stratification for endometrial cancer (16). A radiomics signature derived from multiparametric MRI was reported to be capable of preoperatively predicting LVSI in endometrial cancer with a satisfactory performance (17). When this model was combined with clinical variables, the model yielded an area under the curve (AUC) of 0.807 [95% confidence interval (CI): 0.673–0.941] in an internal validation set, with a 77.8% sensitivity and a 78.6% specificity. However, radiomics requires domain expertise and human engineering for the design of features for pattern recognition. Therefore, radiomics may encounter a saturation curve, meaning that the model’s performance will not improve even with increased sample sizes. Deep learning, a form of representation learning in which a machine can develop its own representations needed for pattern recognition from raw data, can address these limitations by extracting complex and deep features from medical images. In one study, a fusion model combining traditional radiomics and deep learning outperformed models based solely on deep learning or radiomics in classifying early pancreatic lesions and diagnosing pancreatic cancer (18); this demonstrated the ability to improve the model performance through the integration of knowledge from human experts and machine learning-developed representations.

Therefore, the purpose of this study was to develop and validate a comprehensive model integrating clinical features, MRI-based radiomics, and deep learning features for the preoperative prediction of LVSI in a large cohort of patients with endometrial cancer. The overall aim of this work is to inform personalized and effective care in gynecologic oncology. We present this article in accordance with the TRIPOD+AI reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1031/rc).

Methods

Patient population

This multicenter retrospective study was approved by the Institutional Review Board (IRB) of Shenzhen People’s Hospital (IRB No. LL-KY-2024018-02), which waived the requirement for informed consent given the study’s retrospective design. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Potentially eligible patients were initially identified through a search of the electronic medical record systems and pathology databases at each participating center based on a confirmed diagnosis of endometrial endometrioid adenocarcinoma. The specific data collection period spanned January 2010 to December 2023 for Shenzhen People’s Hospital and January 2017 to December 2023 for Shenzhen Second People’s Hospital. Subsequently, patient medical records, preoperative MRI examinations, and pathology reports were systematically reviewed to apply the inclusion and exclusion criteria.

The inclusion criteria were as follows: (I) endometrial cancer with endometrioid adenocarcinoma histology; (II) MRI examination with T2-weighted imaging (T2WI), apparent diffusion coefficient (ADC), and late contrast-enhanced T1-weighted imaging (T1CE) completed within 2 weeks before treatment; and (III) availability of complete clinical data. The exclusion criteria were as follows: (I) incomplete MRI sequences, (II) lesions not visible on MRI or with a maximum diameter <1 cm, and (III) artifacts in the MRI scan. The patient flowchart is shown in Figure 1. Through application of these criteria, the final cohort had 100% multiparametric MRI data completeness. This ensured valid inputs for all planned computational image analyses, spanning both handcrafted radiomics feature extraction and deep learning model development.

Figure 1 Flowchart illustrating participant inclusion and exclusion. Center 1, Shenzhen People’s Hospital; center 2, Shenzhen Second People’s Hospital. MRI, magnetic resonance image.

This study included 475 patients from Shenzhen People’s Hospital (data collection period: January 2010 to December 2023) and 105 patients from Shenzhen Second People’s Hospital (data collection period: January 2017 to December 2023). The 475 patients from Shenzhen People’s Hospital were randomly divided into a training cohort (TC; 380 patients) and an internal validation cohort (IVC; 95 patients) at a ratio of 8:2. The 105 patients from Shenzhen Second People’s Hospital were used as the external validation cohort (EVC).

Clinical feature assessment

Clinical characteristics including age, biopsy-proven pathological grade, cancer antigen 125 (CA125) levels, and depth of myometrial invasion (DMI) were collected. The first three parameters were obtained by reviewing the electronic medical records from the two medical centers. DMI assessments were independently performed by two gynecological imaging radiologists (Y.L. and Y.W., with 6 and 9 years of specialized experience, respectively) using the picture archiving and communication system (PACS). In cases of discrepant evaluations, a final determination was adjudicated by a third senior radiologist (Q.Y., with 13 years of clinical expertise). LVSI status was determined based on the pathology report of the surgical specimen, with LVSI positivity defined as the presence of ≥5 LVSI foci in any single pathological section, corresponding precisely to the “diffuse” category in the semiquantitative classification system (19). All LVSI assessments in pathology reports were independently reviewed by two gynecologic pathologists, and any discrepancies were adjudicated by a third senior pathologist through consensus review.

MRI acquisition

All MRI examinations were performed with 3.0- or 1.5-T scanners at both participating institutions. Dynamic contrast-enhanced abdominal and pelvic MRI scans were performed at least 2 weeks after the biopsy to minimize the effects of postbiopsy inflammation and within 2 weeks before surgery. The MRI scan parameters for both institutions are shown in Table S1. Sagittal T1-weighted late contrast-enhanced images (T1CE) acquired 240 seconds after administration of a gadolinium-based contrast agent (gadopentetate meglumine) at 0.1 mmol/kg body weight, axial ADC maps, and sagittal T2-weighted images in Digital Imaging and Communications in Medicine (DICOM) format were retrieved from the PACS and transferred to a workstation equipped with an GeForce RTX 4080 SUPER GPU (Nvidia, Santa Clara, CA, USA).

Image preprocessing

Initially, N4 bias field correction was applied to the original MRI scans with the N4ITK toolkit in 3D Slicer software (v. 5.6.2), which significantly enhanced the uniformity of the image intensities across various MRI scans. Subsequently, bicubic resampling was used to standardize the image dimensions, yielding a voxel size of 1×1×1 mm³. The images were then normalized by z-scores to minimize data variability across the different centers.

Image segmentation and feature extraction

A radiologist (Y.L.) with 6 years of experience in gynecological MRI interpretation manually delineated the tumor layer by layer using ITK-SNAP (v. 3.8.0) to obtain the volumes of interest (VOIs). To assess feature stability, tumors of 50 patients, who were randomly selected from the patient cohort of Shenzhen People’s Hospital, were segmented by another radiologist (Q.Y.) with 13 years of experience in gynecological MRI interpretation and resegmented by reader (Y.L.) after a 1-month interval. Additionally, reader (Q.Y.) performed tumor resegmentation on three sequence images for a subset of 20 randomly selected patients from each MRI scanner, resulting in the creation of nine distinct datasets. These datasets were used to assess the interscanner reproducibility of the radiomics and deep learning features. A comprehensive illustration of the research design is shown in Figure 2.

Figure 2 Study workflow. AUC, area under the curve; CI, confidence interval; CRDL, combination of clinical, radiomics, and deep learning features; DL, deep learning; MRI, magnetic resonance image; R, radiomics; SVM, support vector machine.

Radiomics features were extracted using PyRadiomics (v. 3.0.1), which included 14 shape-based, 234 first-order, 182 gray-level dependence matrices (GLDMs), 182 gray-level size-zone matrices (GLSZMs), 65 neighboring gray-tone difference matrices (NGTDMs), 208 gray-level run-length matrices (GLRLMs), and 286 gray-level co-occurrence matrices (GLCMs). A total of 3,513 (1,171×3) features were extracted from T2WI, ADC map, and T1CE.

For deep learning feature extraction, a rectangle region of interest (ROI) was segmented on a representative image with the largest tumor depicture and then augmented by an additional 10 pixels to include the peritumoral zones (20-23). A ResNet101 model with pretrained weights from the ImageNet database and modifications to accept single-channel input and remove the final fully connected layer was implemented to yield 2,048 deep learning features that were extracted from each MRI sequence, resulting in a total of 6,144 deep learning features. During the training phase, we applied real-time data augmentation methods such as random cropping and horizontal and vertical flipping. For the test images, the processing was limited to normalization. Other important hyperparameters included the use of stochastic gradient descent as the optimizer and softmax cross-entropy as the loss function. After training of the ResNet101 deep learning model, the features from the average pooling layer were extracted as deep learning features. All deep learning processing described above was performed on the OnekeyAI platform (v. 20240616).

Feature selection and model development

Intra-and interclass correlation coefficients (ICCs) were calculated to evaluate the consistency of the continuous variables extracted by various radiologists or MRI scanners. Features with ICC ≥0.9 were retained for further investigation. Least absolute shrinkage and selection operator (LASSO) regression was used to select features for signature construction.

In the training cohort, a clinical model was developed via logistic regression, while a radiomics model (R), deep learning model (DL), and a comprehensive model (CRDL) integrating clinical, radiomics, and deep learning features were developed using support vector machines (SVMs).

Statistical analysis

Categorical variables are reported as numbers and percentages, and comparisons between groups were assessed via Chi-squared or Fisher exact tests. Continuous variables are reported as means and standard deviations, and comparisons between groups were assessed with the Mann-Whitney test or Student t-test. The predictive ability of the four models was evaluated with the AUC and compared via the DeLong test. Accuracy, sensitivity, and specificity were derived from confusion matrices. Decision curve analysis (DCA) was used to evaluate the model’s clinical benefit. Data were analyzed with Python v. 3.7.12 (Python Software Foundation, Wilmington, DE, USA), and statistical analyses were performed with SPPS v. 20.0 (IBM Corp., Armonk, NY, USA) and MedCalc v. 23.0.2 (MedCalc Software, Ostend, Belgium). A two-sided P value <0.05 was considered to indicate statistical significance.

Results

Patient characteristics

The clinical pathological characteristics of each patient cohort are summarized in Table 1. Among the 580 patients, 118 (20.34%) and 462 (79.66%) were LVSI-positive and LVSI-negative, respectively. Univariate analysis of the training cohort revealed significant differences in grade (P<0.001), DMI (P<0.001), and CA125 (P=0.02) values but not age (P>0.05) between the LVSI-negative and LVSI-positive groups. In the IVC, grade and DMI also differed significantly between the groups (P<0.001), whereas CA125 values and age did not (P>0.05). In the EVC, only DMI differed significantly between the LVSI-negative and LVSI-positive groups (P<0.05).

As shown in Table 1, there were no statistically significant differences in age, grade, or DMI across the TC, IVC, and EVC (all P values >0.05). However, significant differences in CA125 levels were observed between the IVC and TC (P=0.03), whereas no such differences were present between the EV and TC (P=0.282).

Feature selection and model performance

Through backward stepwise multivariable logistic regression, grade and DMI were identified as independent predictors of LVSI. The clinical model incorporating these two variables yielded an AUC of 0.748 (95% CI: 0.672–0.823) in the TC. After feature dimensionality reduction via LASSO, the R model for the TC included 7 radiomics features and yielded an AUC of 0.810 (95% CI: 0.746–0.874). The DL model was derived from 16 DL features strongly correlated with LVSI and produced an AUC of 0.823 (95% CI: 0.759–0.887). The CRDL model, consisting of 2 clinical characteristics, 22 radiomic features, and 5 DL features, obtained an AUC of 0.924 (95% CI: 0.884–0.964). Further details of these features are shown in Figure S1 in the Supplementary Material. The performance of the CRDL model was superior to the clinical model, R model, and DL model in the TC, and these differences were statistically significant after Bonferroni correction (Table 2 and Figure 3A). The calibration and decision curves also indicated that the CRDL model can offer higher discriminatory power and greater potential for net benefit than can the other three models (Figures 3B,4A-4C).

Figure 3 Performance evaluation of predictive models for LVSI in endometrial cancer. (A) ROC curves of the clinical, DL, R, and CRDL models. (B) Calibration curves of the four models. AUC, area under the curve; CI, confidence interval; CRDL, combination of clinical, radiomics, and deep learning features; DL, deep learning; LVSI, lymphovascular space invasion; R, radiomics; ROC, receiver operating characteristic; SVM, support vector machine.

Figure 4 Generalizability and clinical utility of the CRDL model across multiple cohorts. (A-C) Decision curve comparisons of the clinical, R, DL, and CRDL models in the (A) TC, (B) IVC, and (C) EVC. (D) ROC curves of the CRDL model in the TC, IVC, and EVC. CRDL, combination of clinical, radiomics, and deep learning features; DCA, decision curve analysis; DL, deep learning; EVC, external validation cohort; IVC, internal validation cohort; R, radiomics; SVM, support vector machine; TC, training cohort.

Furthermore, the CRDL model achieved AUCs of 0.873 (95% CI: 0.779–0.967) and 0.831(95% CI: 0.707–0.956) in the IVC and EVC, respectively, which were not significantly different compared to the AUCs in the TC (IVC vs. TC: DeLong P=0.329; EVC vs. TC DeLong P=0.165), suggesting the robustness of the CRDL model (Figure 4D). Table 3 presents the CRDL model’s predictive performance metrics for LVSI, including accuracy, sensitivity, and specificity across both the IVC and EVC, along with the DeLong test results comparing AUC values between the validation cohorts and the TC.

Discussion

In this study, we developed and validated a comprehensive MRI-based model integrating clinical variables, radiomics, and deep learning features to predict LVSI in endometrial cancer preoperatively. The CRDL model, including clinical, radiomics, and deep learning features, demonstrated superior predictive performance, achieving AUCs of 0.924, 0.873, and 0.831 in the TC, IVC, and EVC, respectively. This model outperformed the clinical, R, and DL models, supporting the value of integrating multimodal data for preoperative risk stratification. Importantly, no statistically significant differences in AUC values were observed between the training and validation cohorts, indicating the robust generalizability of the model across heterogeneous populations. The CRDL model’s ability to accurately predict LVSI status has significant clinical implications, as LVSI is a critical prognostic factor in endometrial cancer, informing both staging and treatment decisions (24-26).

MRI plays an essential role in guiding the oncological treatment of patients with endometrial cancer by identifying important prognostic factors, assessing extension into adjacent organs, and evaluating their therapeutic responses. Therefore, both the American College of Radiology and the European Society of Urogenital Radiology recommend the use of pelvic MRI as a complementary preoperative assessment tool (27,28). MRI is a multiparametric imaging modality, which can probe many profiles of biocharacteristics. As a heterogeneous disease, endometrial cancer can manifest with a variety of biological behaviors. Previous research indicates that multiparametric MRI can improve the diagnosis and staging of endometrial cancer (29,30). A recent study of a nomogram incorporating conventional preoperative MRI features, including tumor morphology, maximum diameter, presence and DMI (<50% or >50%), cervical stromal invasion (yes or no), and minimal tumor-to-serosa distance, and pathological characteristics yielded an AUC of 0.834 for predicting LVSI status in endometrial cancer (31). Although its diagnostic performance was comparable to the clinical model in our study, it was significantly inferior to the CRDL model. This substantial improvement underscores the critical advantage of machine learning in capturing subtle heterogeneities beyond conventional imaging semantic features, potentially refining risk stratification for personalized surgical planning. Previous studies have reported that multisequence MRI machine learning models can predict LVSI, risk stratification, Ki-67 levels, and DNA polymerase epsilon mutation status in patients with endometrial cancer (32-34). Our multiparametric MRI protocol incorporated conventional sequences including T2WI, T1CE with gadolinium-based agents, and ADC mapping, thereby ensuring methodological consistency with established radiomics research frameworks in gynecological malignancies (35).

We included four clinical features and conducted univariate analysis of variance on both training and validation sets. The results revealed significant differences in DMI and grading between the LVSI-negative and LVSI-positive endometrial cancer groups. After rigorous screening and analysis, two clinical features (DMI and grade) were incorporated into the CRDL model, with DMI at the highest feature proportion. This result further emphasizes the important association between DMI, grading, and LVSI status in patients with endometrial cancer. This conclusion is consistent with that of a previous study (36). It has also been reported that lymph node status is one of the most important prognostic factors in early endometrial cancer (37,38). As our study focused on preoperative prediction and all included clinical variables were obtained from preoperative curettage and conventional MRI, lymph node status was excluded. This exclusion was based on the limited accuracy of conventional preoperative MRI for detecting lymph node metastasis (39). Furthermore, this study derived both radiomics and deep learning features from multiparameter MRI to comprehensively capture the multiple biocharacteristics of endometrial cancer. The application of radiomics depends largely on the knowledge of domain experts, which limits the number and diversity of features that can be extracted. In contrast, by leveraging deep neural networks, deep learning can capture intricate patterns and subtle variations in imaging data that may not be easily discernible through traditional radiomics approaches (40). Furthermore, deep learning models can adapt to a diversity of datasets and learn hierarchical representations, offering a more flexible and scalable approach as compared to traditional feature engineering methods (41). However, deep learning often requires the collection of a large amount of data and ignores knowledge from human experts (42). In our study, we constructed a CRDL model consisting of both traditional radiomics and deep learning features, thereby establishing a robust tool for preoperative risk stratification in endometrial cancer.

A few studies have reported that a fusion model combining clinical characteristics, radiomics features, and deep learning features can provide additional information and improve diagnostic performance (43-45). To the best of our knowledge, this is the first study to describe the development of an integrated model based on the analysis of a large patient cohort and an external test dataset. Furthermore, the model also had high generalization ability, as it could obtain matched predictive performance across the TC, IVC, and EVC. The discriminatory power and net benefit support its clinical translation in real-world settings for stratifying endometrial cancer.

Despite these encouraging results, this study involved several limitations that should be addressed. First, the retrospective design was subject to selection bias, and the inability to adjust for potential confounders (e.g., molecular subtypes) due to data unavailability may impact the reliability of the conclusions. Second, intercenter heterogeneity in imaging hardware and acquisition parameters introduced information bias, while the exclusive use of data from only two Chinese institutions may compromise model generalizability. Third, the biological depth remains constrained by insufficient molecular profiling data. To address these issues, future studies could (I) implement prospective multicenter studies using standardized acquisition protocols and systematically examine confounders (e.g., molecular subtypes and treatment history) and (II) establish multimodal frameworks [for instance, by investigating emerging imaging biomarkers such as DDVD, which has demonstrated a high correlation with Ki-67 proliferation status in endometrial carcinoma (46)], clinically—by integrating MRI-derived machine learning biomarkers with sentinel lymph node biopsy to refine nodal metastasis assessment (47)—and technically, by fusing radiomics with pathomics and proteomics to elucidate tumor biology.

Conclusions

We successfully developed and tested a comprehensive MRI-based model integrating clinical variables, radiomics features, and deep learning features to predict LVSI in endometrial cancer. The excellent prediction performance of this model supports its use as a promising imaging biomarker for LVSI. The clinical translation of the model could facilitate risk stratification and inform individualized treatment decision-making for patients with endometrial cancer.

Acknowledgments

We gratefully acknowledge the Onekey AI platform development team for their technical contributions, and extend our appreciation to all study participants, researchers, and technicians whose collective efforts enabled this research.

Footnote

Reporting Checklist: The authors have completed the TRIPOD+AI reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1031/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1031/dss

Funding: This work was supported by Shenzhen Natural Science Foundation in Basic Research Fund (No. JCYJ20230807112009019).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1031/coif). J.G. serves as an unpaid editorial board member of Quantitative Imaging in Medicine and Surgery. All authors report that this work was supported by Shenzhen Natural Science Foundation in Basic Research Fund (No. JCYJ20230807112009019). The authors have no other conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This multicenter retrospective study was approved by the review board of Shenzhen People’s Hospital (IRB No. LL-KY-2024018-02), which waived the need for obtaining informed consent given the study’s retrospective design. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Levine DAThe Cancer Genome Atlas Research Network. Integrated genomic characterization of endometrial carcinoma. Nature 2013;497:67-73.
2. Berek JS, Matias-Guiu X, Creutzberg C, Fotopoulou C, Gaffney D, Kehoe S, Lindemann K, Mutch D, Concin NEndometrial Cancer Staging Subcommittee, FIGO Women's Cancer Committee. FIGO staging of endometrial cancer: 2023. Int J Gynaecol Obstet 2023;162:383-94.
3. Schwameis R, Fanfani F, Ebner C, Zimmermann N, Peters I, Nero C, Marth C, Ristl R, Leitner K, Grimm C, Oberndorfer F, Capasso I, Zeimet AG, Polterauer S, Scambia G, Fagotti A, Concin N. Verification of the prognostic precision of the new 2023 FIGO staging system in endometrial cancer patients - An international pooled analysis of three ESGO accredited centres. Eur J Cancer 2023;193:113317.
4. Peters EEM, Bartosch C, McCluggage WG, Genestie C, Lax SF, Nout R, Oosting J, Singh N, Smit HCSH, Smit VTHBM, Van de Vijver KK, Bosse T. Reproducibility of lymphovascular space invasion (LVSI) assessment in endometrial cancer. Histopathology 2019;75:128-36.
5. Pifer PM, Bhargava R, Patel AK, Ling DC, Vargo JA, Orr BC, Sukumvanich P, Courtney-B MB, Boisen MM, Berger JL, Taylor S, Olawaiye AB, Comerci JT, Lesnock JL, Edwards RP, Beriwal S. Is the risk of substantial LVSI in stage I endometrial cancer similar to PORTEC in the North American population? - A single-institution study. Gynecol Oncol 2020;159:23-9.
6. Stalberg K, Bjurberg M, Borgfeldt C, Carlson J, Dahm KP, Flöter RA, Hellman K, Hjerpe E, Holmberg E, Kjølhede P, Marcickiewicz J, Rosenberg P, Tholander B, Åvall-Lundqvist E, Högberg T. Lymphovascular space invasion as a predictive factor for lymph node metastases and survival in endometrioid endometrial cancer - a Swedish Gynecologic Cancer Group (SweGCG) study. Acta Oncol 2019;58:1628-33.
7. Leon-Castillo A, Horeweg N, Peters EEM, Rutten T, Ter HN, Smit VTHBM, Kroon CD, Boennelycke M, Hogdall E, Hogdall C, Nout RRA, Creutzberg CL, Ortoft G, Bosse T. Prognostic relevance of the molecular classification in high-grade endometrial cancer for patients staged by lymphadenectomy and without adjuvant treatment. Gynecol Oncol 2022;164:577-86.
8. Maheshwari E, Nougaret S, Stein EB, Rauch GM, Hwang KP, Stafford RJ, Klopp AH, Soliman PT, Maturen K E, Rockall AG, Lee S I, Sadowski EA, Venkatesan AM. Update on MRI in Evaluation and Treatment of Endometrial Cancer. Radiographics 2022;42:2112-30.
9. Avesani G, Bonatti M, Venkatesan AM, Nougaret S, Sala E. RadioGraphics Update: 2023 FIGO Staging System for Endometrial Cancer. Radiographics 2024;44:e240084.
10. Leo E, Stanzione A, Miele M, Cuocolo R, Sica G, Scaglione M, Camera L, Maurea S, Mainenti PP. Artificial Intelligence and Radiomics for Endometrial Cancer MRI: Exploring the Whats, Whys and Hows. J Clin Med 2023;13:226.
11. Antonsen SL, Jensen LN, Loft A, Berthelsen AK, Costa J, Tabor A, Qvist I, Hansen MR, Fisker R, Søgaard-Andersen E, Sperling L, Nielsen AL, Asmussen J, Høgdall E, Fagö-Olsen CL, Christensen IJ, Nedergaard L, Jochumsen K, Høgdall C. MRI, PET/CT and ultrasound in the preoperative staging of endometrial cancer - a multicenter prospective comparative study. Gynecologic Oncology 2013;128:300-8.
12. Haldorsen IS, Salvesen HB. Staging of endometrial carcinomas with MRI using traditional and novel MRI techniques. Clin Radiol 2012;67:2-12.
13. Haldorsen IS, Husby JA, Werner HMJ, Magnussen IJ, Rørvik J, Helland H, Trovik J, Salvesen Øyvind O, Espeland A, Salvesen HB. Standard 1.5-T MRI of endometrial carcinomas: modest agreement between radiologists. Eur Radiol 2012;22:1601-11.
14. Rajpurkar P, Lungren MP. The Current and Future State of AI Interpretation of Medical Images. N Engl J Med 2023;388:1981-90.
15. Esteva A, Robicquet A, Ramsundar B, Kuleshov V, DePristo M, Chou K, Cui C, Corrado G, Thrun S, Dean J. A guide to deep learning in healthcare. Nat Med 2019;25:24-9.
16. Di Donato V, Kontopantelis E, Cuccu I, Sgamba L, Golia D'Augè T, Pernazza A, Della Rocca C, Manganaro L, Catalano C, Perniola G, Palaia I, Tomao F, Giannini A, Muzii L, Bogani G. Magnetic resonance imaging-radiomics in endometrial cancer: a systematic review and meta-analysis. Int J Gynecol Cancer 2023;33:1070-6.
17. Luo Y, Mei D, Gong J, Zuo M, Guo X. Multiparametric MRI-Based Radiomics Nomogram for Predicting Lymphovascular Space Invasion in Endometrial Carcinoma. J Magn Reson Imaging 2020;52:1257-62.
18. Preuss K, Thach N, Liang XY, Baine M, Chen J, Zhang C, Du HF, Yu HF, Lin C, Hollingsworth M, Zheng DD. Using Quantitative Imaging for Personalized Medicine in Pancreatic Cancer: A Review of Radiomics and Deep Learning Applications. Cancers 2022;14:1654.
19. Restaino S, Tortorella L, Dinoi G, Zannoni GF, Baroni A, Capasso I, Distefano E, Sozzi G, Chiantera V, Scambia G, Fanfani F. Semiquantitative evaluation of lymph-vascular space invasion in patients affected by endometrial cancer: Prognostic and clinical implications. Eur J Cancer 2021;142:29-37.
20. Yu Q, Ning Y, Wang A, Li S, Gu J, Li Q, Chen X, Lv F, Zhang X, Yue Q, Peng J. Deep learning-assisted diagnosis of benign and malignant parotid tumors based on contrast-enhanced CT: a multicenter study. Eur Radiol 2023;33:6054-65.
21. Li S, Deng YQ, Hua HL, Li SL, Chen XX, Xie BJ, Zhu ZL, Liu RY, Huang J, Tao ZZ. Deep learning for locally advanced nasopharyngeal carcinoma prognostication based on pre- and post-treatment MRI. Comput Methods Programs Biomed 2022;219:106785.
22. Braman NM, Etesami M, Prasanna P, Dubchuk C, Gilmore H, Tiwari P, Plecha D, Madabhushi A. Intratumoral and peritumoral radiomics for the pretreatment prediction of pathological complete response to neoadjuvant chemotherapy based on breast DCE-MRI. Breast Cancer Res 2017;19:57.
23. Qiu Y, Liu YF, Shu X, Qiao XF, Ai GY, He XJ. Peritumoral Radiomics Strategy Based on Ensemble Learning for the Prediction of Gleason Grade Group of Prostate Cancer. Acad Radiol 2023;30:S1-3.
24. Gaffney D, Matias GX, Mutch D, Scambia G, Creutzberg C, Fotopoulou C, Berek J S, Concin N. 2023 FIGO staging system for endometrial cancer: The evolution of the revolution. Gynecol Oncol 2024;184:245-53.
25. Harris KL, Maurer KA, Jarboe E, Werner TL, Gaffney D. LVSI positive and NX in early endometrial cancer: Surgical restaging (and no further treatment if N0), or adjuvant ERT? Gynecol Oncol 2020;156:243-50.
26. Pifer PM, Jaishankar S, Bhargava R, Schad MD, Keller A, Musunuru HB, Cohen M, Sukumvanich P, Courtney-B M, Boisen M, Berger J L, Olawaiye A, Lesnock J, Edwards R, Taylor S, Vargo J A, Beriwal S. Is Substantial Lymphovascular Space Invasion Prognostic in Patients With Pathologically Lymph Node-Negative Endometrial Cancer?. Int J Radiat Oncol Biol Phys 2023;117:148-53.
27. Expert Panel on GYN and OB Imaging. Reinhold C, Ueno Y, Akin EA, Bhosale PR, Dudiak KM, Jhingran A, Kang SK, Kilcoyne A, Lakhman Y, Nicola R, Pandharipande PV, Paspulati R, Shinagare AB, Small W, Vargas HA, Whitcomb BP, Glanc P. ACR Appropriateness Criteria® Pretreatment Evaluation and Follow-Up of Endometrial Cancer. J Am Coll Radiol 2020;17:S472-86.
28. Nougaret S, Horta M, Sala E, Lakhman Y, Thomassin NI, Kido A, Masselli G, Bharwani N, Sadowski E, Ertmer A, Otero-G M, Kubik-H RA, Cunha T M, Rockall A, Forstner R. Endometrial Cancer MRI staging: Updated Guidelines of the European Society of Urogenital Radiology. Eur Radiol 2019;29:792-805.
29. Pati SK, Mondal K, Bodhey NK, Bagde N, Gupta RK, Shukla A. Role of Multiparametric MRI in the Preoperative Evaluation of Endometrial Carcinoma: A Cross-Sectional Study. Cureus 2024;16:e65058.
30. Lefebvre TL, Ueno Y, Dohan A, Chatterjee A, Vallières M, Winter-Reinhold E, Saif S, Levesque IR, Zeng XZ, Forghani R, Seuntjens J, Soyer P, Savadjiev P, Reinhold C. Development and Validation of Multiparametric MRI-based Radiomics Models for Preoperative Risk Stratification of Endometrial Cancer. Radiology 2022;305:375-86.
31. Valletta R, Avesani G, Vingiani V, Proner B, Steinkasserer M, Notaro S, Vanzo F, Negri G, Vercelli C, Bonatti M. A Nomogram for Preoperative Prediction of Tumor Aggressiveness and Lymphovascular Space Involvement in Patients with Endometrial Cancer. J Clin Med 2025;14:3914.
32. Liu D, Huang J, Zhang Y, Shen H, Wang X, Huang Z, Chen X, Qiao Z, Hu C. Multimodal MRI-based radiomics models for the preoperative prediction of lymphovascular space invasion of endometrial carcinoma. BMC Med Imaging 2024;24:252.
33. Ding SX, Sun YF, Meng H, Wang JN, Xue LY, Gao BL, Yin XP. Radiomics model based on multi-sequence MRI for preoperative prediction of ki-67 expression levels in early endometrial cancer. Sci Rep 2023;13:22052.
34. Lin Z, Gu W, Guo Q, Xiao M, Li R, Deng L, Li Y, Cui Y, Li H, Qiang J. Multisequence MRI-based radiomics model for predicting POLE mutation status in patients with endometrial cancer. Br J Radiol 2023;96:20221063.
35. Jia Y, Hou L, Zhao J, Ren J, Li D, Li H, Cui Y. Radiomics analysis of multiparametric MRI for preoperative prediction of microsatellite instability status in endometrial cancer: a dual-center study. Front Oncol 2024;14:1333020.
36. Kim SI, Yoon JH, Lee SJ, Song MJ, Kim JH, Lee HN, Jung G, Yoo JG. Prediction of lymphovascular space invasion in patients with endometrial cancer. Int J Med Sci 2021;18:2828-34.
37. Cheng F, Yang L, Wen Q, Xu J, Shao F. Risk factors and prognostic analysis of endometrial cancer with para-aortic lymph node metastasis. Discov Oncol 2025;16:1344.
38. Hong JH, Kang J, Lee SJ, Lee KH, Hur SY, Kim YS. High-Risk Early-Stage Endometrial Cancer: Role of Adjuvant Therapy and Prognostic Factors Affecting Survival. Cancers (Basel) 2025;17:2056.
39. Yan BC, Li Y, Ma FH, Zhang GF, Feng F, Sun MH, Lin GW, Qiang JW. Radiologists with MRI-based radiomics aids to predict the pelvic lymph node metastasis in endometrial cancer: a multicenter study. Eur Radiol 2021;31:411-22.
40. Scapicchio C, Gabelloni M, Barucci A, Cioni D, Saba L, Neri E. A deep look into radiomics. Radiol Med 2021;126:1296-311.
41. Moulaei K, Afshari L, Moulaei R, Sabet B, Mousavi SM, Afrash MR. Explainable artificial intelligence for stroke prediction through comparison of deep learning and machine learning models. Sci Rep 2024;14:31392.
42. Alzubaidi L, Bai J, Al-Sabaawi A, Santamaria J, Albahri AS. AI-dabbagh BSN, Fadhel MA, Manoufali M, Zhang J, Al Timemy AH, Duan Y, Abdullah A, Farhan L, Lu Y, Gupta A, Albu F, Abbosh A, Gu Y. A survey on deep learning tools dealing with data scarcity: definitions, challenges, solutions, tips, and applications. J Big Data 2023;10:46.
43. Tian WW, Yan QQ, Huang XY, Feng R, Shan F, Geng DY, Zhang ZY. Predicting occult lymph node metastasis in solid-predominantly invasive lung adenocarcinoma across multiple centers using radiomics-deep learning fusion model. Cancer Imaging 2024;24:8.
44. Liang X, Tang K, Ke X, Jiang J, Li S, Xue C, Deng J, Liu X, Yan C, Gao M, Zhou J, Zhao L. Development of an MRI-Based Comprehensive Model Fusing Clinical, Radiomics and Deep Learning Models for Preoperative Histological Stratification in Intracranial Solitary Fibrous Tumor. J Magn Reson Imaging 2024;60:523-33.
45. Li X, Yang L, Jiao X. Comparison of Traditional Radiomics, Deep Learning Radiomics and Fusion Methods for Axillary Lymph Node Metastasis Prediction in Breast Cancer. Acad Radiol 2023;30:1281-87.
46. Wang F, Wang Y, Qi L, Liang J, Xiao BH, Zhang C, Wáng YXJ, Ye Z. High correlation between Ki-67 expression and a novel perfusion MRI biomarker diffusion-derived vessel density (DDVD) in endometrial carcinoma. Magn Reson Imaging 2025;117:110324.
47. Bogani G, Giannini A, Vizza E, Di Donato V, Raspagliesi F. Sentinel node mapping in endometrial cancer. J Gynecol Oncol 2024;35:e29.

(English Language Editor: J. Gray)