Introduction

Endometrial cancer (EC) is one of the most common types of cancer [1]. The transition from FIGO 2008 to FIGO 2023 introduces molecular classification, which included polymerase epsilon (POLE) mutation, loss of mismatch repair protein expression (MMR-D), no specific molecular profile (NSMP) and p53-abnormal (p53abn), for EC staging, aiding in prognostic risk assessment and treatment decisions [2-3]. However, limitations in economic capacity and testing technology hinder universal application. Radiogenomics, combining radiomics and genomic data, offers a noninvasive method to discern molecular characteristics [4]. This approach shows promise in characterizing tumors and genomics [5]. Our goal is to explore a novel radiogenomics approach to create a gene expression signature for EC patients from two institutions, aiming to improve prognostic and predictive information for patients.

Methods

During the period between January 2020 and August 2022, data from 254 patients diagnosed with endometrial cancer (EC) were retrospectively analyzed after histological and genetic confirmation. These patients were recruited from two medical centers: 217 individuals from the Obstetrics and Gynecology Hospital of Fudan University (Institution 1) and 37 from the Beijing Chaoyang Hospital (Institution 2). Genomic DNA was meticulously examined using a specialized endometrial cancer molecular typing test kit, covering markers such as POLE, TP53, MSH2, PMS2, MLH1, and MSH6 for detailed subtyping. The MRI examinations were conducted utilizing either a 1.5T MR scanner (MAGNETOM Avanto, Siemens Healthcare, Erlangen, Germany) or a 3T MR scanner (MAGNETOM Prisma, Siemens Healthcare, Erlangen, Germany). DICOM files of T2WI, T1WI, DWI, and T1CE were processed using ITK-SNAP. A skilled radiologist with 25 years of experience delineated the tumor regions as the region of interest (ROI), subsequently reviewed by a senior radiologist with extensive experience in pelvic female tumors. Data from Institution 1 were divided into a 7:3 ratio, facilitating the training and internal test set, while Institution 2 data served as the external test set. The research protocol was shown in Figure 1. A radiomics model was constructed based on four sequences. From each sequence, we extracted 386 radiomics features from the ROI, including both original features and those derived from the Laplacian-of-Gaussian (LoG) filter by FAE (v.0.5.8) [6]. For the training data set, we firstly normalized each feature to remove the scale effect and then incorporated the Pearson correlation coefficient for dimension reduction, recursive feature elimination for feature selection, and a logistic regression to develop the classifier. This comprehensive approach yielded the most effective radiomics model based on a 5-fold cross-validation to determine the optimal features for the further analysis. Subsequently, the final radiomics-clinical model was developed, integrating clinical features (Onset Age, CA125, and FIGO stage) with the rad-score, calculated from the linear combination of selected features derived from the four sequences. Clinical and pathological characteristics were meticulously analyzed using SPSS.

Results

The clinical and pathological characteristics of the two institutions are summarized in the results section (Table 1). Institution 1 included 19 POLE-EDM, 86 MMR-D, 69 NSMP, and 43 p53abn patients, while Institution 2 had 1 POLE-EDM, 8 MMR-D, 21 NSMP, and 7 p53abn patients. The optimal radiomics model, consisting of 19 features selected from 4 sequences, demonstrated an AUC of 0.760 (95% CI: 0.626-0.884) and 0.652 (95% CI: 0.543-0.855) in the internal and external test sets，respectively. However, the AUC value of the final model improved by 0.849 (95% CI: 0.754-0.924) and 0.673 (95% CI: 0.582-0.764) in the internal and external test set (Table 2). Notably, SHAP (SHapley Additive exPlanations) showcased the significance of the Rad-score compared with clinical features (Figure 2).

Discussion

In our study, the radiogenomics-based analysis displayed the moderate-to-high diagnostic performance in distinguishing molecular subtypes of EC. This study demonstrated how the model supports the differentiation of molecular subtypes, in line with previous research [4], but our study adds external test, combining clinical and pathological information to give a richer model. Future studies in radiogenomics should focus on evaluating different imaging modalities to directly identify the molecular subtype spectrum of TCGA from radiomic data in larger populations. The limitation of the present study is the performance on the external test set was suboptimal. Figure 3 revealed variations in the distribution of imaging features and clinical characteristics between the two institutions, which may be due to sample size inadequacies and differences in MRI scanners. Generalizability can be enhanced in future studies with larger sample sizes and standardized methods.

Conclusion

In conclusion, our results demonstrate that the predictive model derived from MRI imaging features holds significant promise in identifying molecular subtypes in EC. This model has the potential to guide clinicians in tailoring individualized treatments for EC patients.

Acknowledgements

Wenyi Yue and Ruxue Han contributed equally to this work.