Endometrial cancer (EC) is the sixth most common cancer in women worldwide, with 417,000 new cases and 97,000 deaths in 2020 (1). Endometrial cancer is classified into two histopathological subtypes type I- endometrioid carcinoma accounting for 80% and type II-non-endometrioid subtypes including serous, clear cell and carcinosarcoma, which confers a worse prognosis (2). Whilst traditionally, further risk stratification is achieved through histologic grade and staging, the advances in genetic profiling in EC have resulted in a shift towards a molecular classification of EC based on four distinct genomic subgroups (2, 3).

Initial disease characterisation determines appropriate therapeutic strategies such as surgery, radiation, chemotherapy, hormonal and immunotherapy. Early-stage disease is primarily treated with hysterectomy and bilateral salpingo-oophrectomy, with adjuvant therapies traditionally reserved for intermediate-high risk or advanced disease (2).

MRI and endometrial biopsy have an important role in the diagnosis and evaluation of EC.

However, there are limitations to these assessments, including the inability to radiologically assess important histopathologic features like lymphovascular space invasion (LVSI) and sampling error or bias inherent to biopsy. There remain challenges to predicting long-term outcomes and disease-free survival (DFS). Currently, the critical prognostic factors for EC include tumour grade, histological subtype, genomic subtype, deep myometrial invasion (DMI), LVSI and lymph node metastases (LNM) (2, 3). Many of the prognostic features in the International Federation of Gynecology and Obstetrics (FIGO) classification for staging endometrial cancer rely on histopathology assessment of post-hysterectomy surgical specimens (2). Optimising preoperative radiological assessment could decrease reliance on and potentially circumvent surgical staging and therapeutic lymphadenectomy.

Radiomics is the process of computational extraction of vast amounts of quantitative voxel metrics from a segmented region of interest (ROI) on diagnostic images, which undergo statistical techniques to extract spatiotemporal features, thought to reflect the underlying pathophysiology and tumour phenotype (4). Radiomics has the potential to create a non-invasive a priori approach to personalised patient risk-stratification and refine the clinical decision-making process. Despite radiomic advances in other oncological areas, there remains a relative dearth of literature with regards to the prediction of high-risk histopathologic features for tumour recurrence in EC (5). In particular, there are a limited number of publications establishing the role of MRI radiomics in endometrial cancer DFS (6–9).

Thus, we aimed to explore a multiparametric MRI radiomics model predictive for DFS and high-risk histopathologic features associated with tumour recurrence in endometrial cancer.

This was a retrospective study in a large Canadian tertiary referral institute with approval from our HIPPA-compliant institutional research and ethics board with waiver of informed consent. The initial surgical database search over a ten-year period (2010 – 2020) identified 238 consecutive patients with endometrial cancer who underwent MRI of which 167 were excluded due to occult disease on MRI, inadequate imaging, incomplete surgical data, distant metastases, concurrent neoplasm at diagnosis, fertility sparing surgery and preoperative neoadjuvant treatment ( Figure 1 ). The final study cohort comprised of 71 patients who had a preoperative MRI and underwent operative management for histopathology-proven endometrial cancer at our institution. Patients with all histological subtypes and grade of endometrial cancer were included.

Patient characteristics and data were collected from the institutional electronic patient record (EPR) system.

Our results show that multiparametric MRI radiomics could potentially be used preoperatively in patients with EC to predict DFS and important high-risk histopathologic features associated with tumour recurrence. We established a six-feature elastic net radiomic signature that was a strong predictor of DFS, while clinicopathologic features had no significant correlation with DFS. We established statistically significant individual RF for the high-risk histopathologic features and statistically significant elastic net MRI radiomic signatures specifically for DMI, hysterectomy grade, hysterectomy histology, and LVSI. This raises the possibility of inadequate clinicopathologic EC prognostication and the potential prognostic role for multiparametric MRI radiomics.

Our results demonstrating the role of MRI radiomics in predicting EC survival outcomes are in-keeping with a prospective study by Ytre-Hauge et al, in which high kurtosis in post-contrast T1WI was a good predictor of progression-free survival (HR 1.5, p<0.001) (6). In the same cohort, Jacob et al. reported an MRI radiomics model that predicted 5-year disease-specific survival (p < 0.001) (8). Whilst segmentation is a mainstay technique of radiomics literature, whole-tumor MRI RF have also been shown to determine significant predictors of progression-free survival (13). To the best of our knowledge, there is only one additional study to report MRI radiomics in predicting survival outcomes, which reports single sequence ADC metrics predicting disease recurrence and reduced survival (14). Our results contribute to the growing evidence in support of the use of MRI radiomics in the prognostication of EC survival outcomes.

In addition to DFS, we report the utility of MRI radiomics in predicting established high-risk histopathologic features for tumour recurrence in EC. LVSI represents microvessel tumour emboli and is one of the most important high-risk histopathologic features in EC, as it is considered a precursor to metastatic disease (3). In combination with similar contemporary studies, our demonstration of an MRI radiomics signature predictive for LVSI helps to establish the diagnostic predictive value of MRI radiomics for LVSI and its potential practical application in clinical decision-making (7, 15, 16). Ueno et al. and Zhang et al. also reported the predictive value of MRI radiomics in LVSI, DMI and high-grade tumours (15, 16). Long et al. in a cohort of n=184, showed T2WI and DCE T1WI-based radiomics had a strong predictive role for LVSI (AUC 0.93 95% CI: 0.875–0.991; sensitivity: 91.6%; specificity: 96.0%) (17).In contrast to these results, Bereby-Kahane et al, found a limited role for MRI radiomics in LVSI and high EC grade prediction (AUC of 0.59, sensitivity = 71%; specificity = 59%) and reported that tumour size (short axis 20mm) outperformed radiomics as their strongest predictor (18). Similar results were shown by Fasmer et al, as tumour size outperformed their MRI radiomic signatures for DMI and LMN (13). These discrepant findings may be due to their small sample sizes and, in the case of Fasmer et al, low numbers of high-risk surgicopathological features and a lack of multiparametric radiomic analysis.

Multiple studies have established robust MRI radiomic imaging biomarkers for LNM in EC (6, 9, 13, 15, 19, 20). Whilst we did not find a significant MRI radiomic signature for LNM, this is likely attributed to the low incidence of nodal positive disease in our cohort (n=14 positive pelvic nodes). The integration of clinical variables with radiomics modelling has shown MRI clinicoradiomics fusion models can also predict LNM in EC (19, 21). Xu et al. showed a combination of DCE T1WI RF, lymph node size and CA125, showed the best predictability for LNM and out-performed conventional radiologist MRI assessment (19). Yan et al. in a large (n=717) multicentre study with external validation, showed that clinicoradiomic fusion models have a strong diagnostic performance in predicting high-risk EC (AUCs of 0.75 and 0.85 in two validation groups) with applications in surgical decisions (22). Multiple additional studies, including our results, have also shown radiomics to have significant independent prediction in stratifying low or high-risk EC (6, 9, 13, 16, 18, 21–26). DMI is another important high-risk histopathologic feature for EC, with a developing role for MRI radiomics predictive models (6, 9, 13, 15, 16, 22, 25–29). Stanzione et al demonstrated radiomic-assisted MRI interpretation improved accuracy for DMI detection from 82% to 100% (p=0.48) (27). Less supportive results were reported by Otani et al, in a study of n=200 patients, their multiparametric MRI radiomics models predictive for DMI did not enhance the conventional radiologist MRI assessment (30). There are, however, multiple additional studies showing encouraging results for radiomic-assisted diagnostic performance compared with the conventional MRI evaluation in EC (19, 20, 24).

Over recent years, the advancement of molecular characterisation in EC, has resulted in a refinement of molecular prognostication based on the four distinct genomic subtypes proposed by The Cancer Genome Atlas (TCGA) (2, 3). In conjunction, radiomics has been shown to enhance EC genetic profiling in combined radiogenomic modelling that incorporates genomic tumour information (8, 9). We tested for one of the four TCGA subtypes and did not find significant individual RF or radiomic signature associations (p=0.074). Despite the increasing number of validated molecular biomarkers for EC, their clinical applications are still limited (3). This may be due in part to the limited availability for IHC MMR screening outside of tertiary centres. Similarly, there remains uncertainty regarding the transferability of MRI radiomics into clinical practice and the need for large scale validation studies is paramount. Integrating genomics, with clinicoradiomic fusion risk-stratification models could produce a more robust and clinically applicable approach to prognostication.

We must acknowledge several limitations of our study. Primarily, the retrospective single-institution study design resulting in a smaller patient cohort with the lack of validation may have also resulted in overfitting of the models. The limited sample number was too small to use model training and testing. The proportion of patients in our cohort with an IHC MMR status was low, limiting conclusions. While 3 years may be sufficient follow up time to predict an event in advanced stage (3-4) endometrial cancer, it is a relatively short time for follow up to predict DFS or survival in early-stage endometrial cancer. The analysis was initiated before the new FIGO 2023 staging system was published and therefore dichotomised into endometrioid vs non-endometrioid. Clinically, many hospitals have not yet adopted the new FIGO 2023 staging system and still categorizes endometrial cancer into endometrioid vs non-endometrioid. We utilised a tumour segmentation approach comprised of axial imaging of key pulse sequences only in majority of patients and even though all images wherein tumor was visible were contoured, theoretically in a heterogenous tumour this approach could possibly not be entirely representative. However, the literature comparing single slice and whole-tumour segmentation is currently inconclusive. Segmentations were performed by one radiologist, thus not allowing for assessment of reproducibility and stability of the radiomic features used for model development. Computational image preprocessing was not utilised, however image resampling remains a controversial process.

Additionally considering MRI data collection in this patient cohort was over a 10-year period, evolution in MRI technology could have also further contributed to data heterogeneity.

In conclusion, our results support the potential role for multiparametric MRI radiomics in EC to non-invasively predict DFS and critical high-risk histopathologic features for tumour recurrence, and thus optimise personalised patient care. However, despite the promising results, practical implementation in the clinical decision-making process has yet to come to fruition. To establish standardised non-invasive radiomics-aided preoperative prognostic modelling in EC, large-scale multi-institutional data sharing is likely necessary to facilitate comprehensive validation required for clinical translation and personalised patient management.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving humans were approved by University Health Network, Toronto, Canada. The studies were conducted in accordance with the local legislation and institutional requirements. The ethics committee/institutional review board waived the requirement of written informed consent for participation from the participants or the participants’ legal guardians/next of kin in accordance with institutional ethics review.

MRe: Formal analysis, Methodology, Writing – original draft, Writing – review & editing. MF: Conceptualization, Data curation, Investigation, Software, Writing – original draft, Writing – review & editing. JW: Data curation, Formal analysis, Investigation, Validation, Writing – original draft, Writing – review & editing. MM: Methodology, Writing – original draft, Writing – review & editing. SL: Conceptualization, Methodology, Writing – original draft, Writing – review & editing. MRo: Data curation, Formal analysis, Writing – original draft, Writing – review & editing. KH: Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Software, Supervision, Writing – original draft, Writing – review & editing. KJ: Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writing – original draft, Writing – review & editing.