ESTRO 2025 - Abstract Book

S3741

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2025

2-Tensaouti F., et al. Radiother Oncol. 2017 Mar;122(3):362-367. doi: 10.1016/j.radonc.2016.12.025. Epub 2017 Jan 12. 3- Tensaouti F., et al. Abstract 849: Radiotherapy and Oncology 194 (2024): S2030-S2033. 4-Tensaouti F et al, British Journal of Radiology , https://doi.org/10.1259/bjr.20160537 5- Nioche C., et al. LIFEx. Cancer Res 2018 ; 78 :4786–9. https://doi.org/10.1158/0008-5472.CAN-18-0125 6- https://wiki.cancerimagingarchive.net/display/Public/VASARI+Research+Project

1286

Digital Poster a preliminary study of radiation enteritis associated with temporal sequencing of total neoadjuvant therapy in locally advanced rectal cancer Chenying Ma, Yi Fu, Shuyue Li, Jie Chen, Guanghui Gan, Juying Zhou Radiation Oncology, The First Affiliated Hospital of Soochow University, Suzhou, China Purpose/Objective: To develop a delta radiomics model based on multi-temporal MRI to predict the occurrence and progression of severe acute radiation enteritis (SARE) during total neoadjuvant therapy (TNT) in patients with locally advanced rectal cancer. Material/Methods: A retrospective analysis was conducted on 92 patients with locally advanced rectal adenocarcinoma at our institution, divided into a training set of 73 cases and a validation set of 19 cases. All patients underwent baseline and post-radiotherapy pelvic contrast-enhanced MRI examinations. For each patient, the MRI T2-weighted imaging (T2WI) sequences before and after radiotherapy were registered with the planning CT sequences to obtain the clinical target volumes (CTVR0 and CTVpost-RT), which were defined as regions of interest. Radiomic features were extracted from these regions, and temporal features were obtained by calculating the differences in texture features between the two time-point T2WI sequences in the same region （ Δfeature=CTVpost-RT-CTVR0 ） . Two machine learning methods-Least Absolute Shrinkage and Selection Operator (LASSO) logistic regression and Random Forest regression-were employed to build predictive models, including a clinical factors model, a baseline MRI radiomics model, a post-radiotherapy MRI radiomics model, and a delta radiomics model. The optimal Youden index was determined, and Receiver Operating Characteristic (ROC) curves, calibration curves, and decision curves were plotted to evaluate the predictive performance of the different models. Finally, model visualization and validation were performed using an NPC-Wise model combined with the SHapley Additive exPlanations (SHAP) algorithm. Results: The predictive model based on delta features achieved an area under the ROC curve (AUC) of 0.977, with an accuracy of 91.7%, sensitivity of 87.5%, specificity of 95.0%, negative predictive value of 90.5%, and positive predictive value of 93.3%. These metrics were significantly higher than those of the baseline MRI radiomics model (AUC: 0.940), the post-radiotherapy MRI radiomics model (AUC: 0.915), and the clinical factors predictive model (AUC: 0.872). Moreover, the constructed nomogram demonstrated high predictive efficiency, with a sensitivity of 96.9% and specificity of 91.4%. The SHAP feature importance scatter plot illustrated the ranking of feature weights during the model's prediction process. Finally, we visualized at the individual level the entire process by which delta radiomic features influence the NPC-Wise model's prediction of the risk of severe acute radiation enteritis (SARE). Conclusion: The delta radiomics predictive model based on temporal features demonstrates excellent predictive performance in predicting SARE associated with TNT in patients with locally advanced rectal cancer.