S429
Clinical - Gynaecological
ESTRO 2026
LASSO regularization. Logistic regression models—clinical-only, radiomic-only, and combined—were validated using Leave-One- Out Cross-Validation (LOOCV), and model performance was compared using Receiver Operating Characteristic (ROC) curve analysis with Area Under the Curve (AUC) values as the primary metric (Figure 1).
Conclusion: This study demonstrates that radiomic features derived from pretreatment MRI can noninvasively predict early chemoradiotherapy response in locally advanced cervical cancer, supporting individualized treatment adaptation. Early response assessment may refine brachytherapy planning and provide insights into tumor chemoradioresistance and long- term outcomes. The lower performance of the combined model likely reflects the limited sample size; ongoing data collection is expected to strengthen model generalizability and predictive accuracy. References: Gui B, Autorino R, Miccò M, et al. Pretreatment MRI Radiomics Based Response Prediction Model in Locally Advanced Cervical Cancer. Diagnostics (Basel). 2021;11(4):631. Published 2021 Mar 31. doi:10.3390/diagnostics11040631Autorino R, Gui B, Panza G, et al. Radiomics-based prediction of two-year clinical outcome in locally advanced cervical cancer patients undergoing neoadjuvant chemoradiotherapy. Radiol Med. 2022;127(5):498-506. doi:10.1007/s11547- 022-01482-9Wang W, Yang G, Liu Y, et al. Multimodal deep learning model for prognostic prediction in cervical cancer receiving definitive radiotherapy: a multi- center study. NPJ Digit Med. 2025;8(1):503. Published 2025 Aug 4. doi:10.1038/s41746- 025-01903-9 Keywords: Radiomics, cervix cancer, artificial intelligence
Results: The median age was 54 years. Of 68 patients, 43 (63.2%) achieved CR and 25 (36.8%) had non-CR on MRI. Most patients had squamous cell carcinoma (86.8%) and were HPV positive (75.0%). The predominant FIGO stage was IIIC1 (48.5%). The feature selection pipeline reduced the dataset from 113 variables to seven predictors (four radiomic and three clinical). The selected radiomic features—
original_shape_LeastAxisLength, original_shape_MinorAxisLength,
original_gldm_DependenceEntropy, and original_glszm_ZoneEntropy—reflected tumor shape and texture heterogeneity.The radiomic-only model achieved the best performance (AUC = 0.773, sensitivity = 87.2%, accuracy = 69.4%), outperforming the clinical-only (AUC = 0.613) and combined (AUC = 0.717) models (p = 0.005) (Figure 2).
Made with FlippingBook - Share PDF online