2578 - Non-Invasive Molecular Risk Stratification in Endometrial Cancer Using Phenotypes Extracted from Routine MRI Reports

Purpose/Objective(s): Molecular classification increasingly guides adjuvant radiotherapy and systemic therapy decisions in endometrial cancer, yet access to molecular testing remains variable. We hypothesized that routinely documented MRI phenotypes could predict molecular subtype. The primary objective was to evaluate whether text-extracted MRI features could classify dMMR, p53-abnormal, POLE-mutated, and NSMP tumors, measured by area under the receiver operating characteristic curve (AUC) and overall classification accuracy.

Materials/Methods: We retrospectively studied 63 consecutive patients with biopsy-proven endometrial carcinoma who underwent preoperative MRI at a tertiary center. Narrative reports were analyzed using a rule-based natural language processing approach to extract predefined descriptors, including depth of myometrial invasion, cervical involvement, T2 heterogeneity, diffusion restriction, enhancement pattern, and nodal features. Final surgical pathology provided molecular classification (dMMR, p53-abnormal, POLE-mutated, NSMP). Supervised machine-learning models were trained on text-derived features with 80:20 internal validation. Discrimination was assessed using AUC and overall accuracy.

Results: The cohort comprised dMMR (n = 19), p53-abnormal (n = 17), POLE-mutated (n = 5), and NSMP tumors (n = 22). p53-abnormal cancers commonly showed irregular margins and marked diffusion restriction, whereas dMMR tumors more often demonstrated heterogeneous T2 signal. Models achieved AUCs of 0.78 for dMMR, 0.81 for p53-abnormal, and 0.70 for POLE, with 74% overall accuracy. Incorporation of age, grade, and BMI increased accuracy to 82%. In roughly 20% of cases, imaging-based predictions suggested high-risk molecular features not initially suspected clinically.

Conclusion: Structured analysis of routine MRI report descriptors demonstrates meaningful association with molecular subtype in endometrial cancer. This pragmatic, low-cost approach requires no additional imaging or specialized software and may help prioritize molecular testing and inform treatment planning, particularly in resource-constrained settings. Validation in larger, multi-institutional cohorts is needed.