204 Background: Endometrial cancer is a leading cause of cancer mortality in women worldwide. The Cancer Genome Atlas (TCGA) classifies it into four molecular subtypes—POLEmut, MMRd, NSMP, and p53abn—which are crucial for prognosis and targeted therapies like immunotherapy. Current molecular testing can be costly or subjective. While MRI, particularly T2-weighted imaging, is valuable in diagnosis, its interpretation suffers from inter-observer variability. Radiomics provides a quantitative, non-invasive approach to analyze tumor heterogeneity on MRI. This study aims to develop and integrate radiomics models based on selected MRI sequences to non-invasively predict the molecular subtypes of endometrioid adenocarcinoma. Methods: This retrospective multi-center study developed a radiomics model to predict TP53 mutation status in endometrial adenocarcinoma by extracting and selecting features from preoperative T2-weighted and ADC MRI images, following rigorous image segmentation, inter-/intra-observer consistency validation, and redundancy reduction. Results: A total of 784, 60, and 45 patients were eligible at institutions 1, 2, and 3, respectively. 73% of the participants from Institution 1 were selected as the training set (n =701), and the remaining 27% from Institution 1 were used as the internal validation set (n=194). The combined total from institution 2 and institution 3 was used as the external validation set (n =105). Institution 1 included 152 patients with p53abn. Institution 2 had patients 8 patients with p53abn. Institution 3 included 8 patients with p53abn.The best model is the XGBoost model, in the T2-OAX sequence, the internal validation set AUC reached 0.716 (95% CI 0.621–0.811), and the external validation set AUC reached 0.700 (95% CI 0.565–0.835). In the T2-SAG sequence, the internal validation set AUC reached 0.728 (95% CI 0.658–0.799), and the external validation set AUC reached 0.721 (95% CI 0.580–0.862). In the T2 fusion sequence, the internal validation set AUC reached 0.821 (95% CI 0.758–0.884), and the external validation set AUC reached 0.795 (95% CI 0.674–0.915). Conclusions: Our radiomics blend model has enabled the MRI-based EC molecular subtypes' non-invasive classification. Sub sequent exploration should be conducted with prospective validation and a larger dataset to better guide clinicians in developing personalized treatment plans for patients with EC.
Luo et al. (Tue,) studied this question.