S2414
Physics - Radiomics, functional and biological imaging, and outcome prediction
ESTRO 2026
Our multimodal radiopathomic model accurately predicts early recurrence in LARC by integrating complementary MRI, pathology, and clinical data for comprehensive tumor assessment. It achieved high predictive accuracy and is well calibrated, indicating strong clinical utility. For example, high-risk patients could be triaged to intensified therapy or closer surveillance, whereas low-risk patients might avoid overtreatment. The model relies only on routine imaging and pathology data, making it readily implementable and promising for personalized post- treatment management of LARC. References: 1. Abdollahi H, Dehesh T, Abdalvand N, Rahmim A. Radiomics and dosiomicsbased prediction of radiotherapy-induced Xerostomia in head and neck cancer patients. Int J Radiat Biol. 2023; 99(11): 1669– 83.2. Ma CY, Zhao J, Qian KY, Xu Z, Xu XT, Zhou JY. Analysis of nutritional risk, skeletal muscle depletion, and lipid metabolism phenotype in acute radiation enteritis. World J Gastrointest Surg. 2023; 15(12): 2831–43.3. Raptis S, Ilioudis C, Theodorou K. From pixels to prognosis: unveiling radiomics models with SHAP and LIME for enhanced interpretability. Biomed Phys Eng Express. 2024;10(3). Keywords: rectal cancer, TNT, early recurrence prediction PISA (Pipeline for Interpretable-by-design Survival Analysis): A novel AI pipeline applied to survival prediction in patients with bone metastases Thalea Schlender 1 , Yvette M van der Linden 1 , Peter A.N. Bosman 2,3 , Tanja Alderliesten 1 1 Radiotherapy Department, Leiden University Medical Center, Leiden, Netherlands. 2 Evolutionary Intelligence Group, Centrum Wiskunde & Informatica, Amsterdam, Netherlands. 3 Software Technology, TU Delft, Delft, Netherlands Digital Poster Highlight 513 Purpose/Objective: Traditional survival models often miss non-linear effects, while deep learning methods lack transparency. We propose PISA (Pipeline for Interpretable Survival Analysis), an automated interpretable AI pipeline that transforms survival data into feature sets for survival analysis models and generates meaningful patient stratifications. In advanced metastatic cancer, accurate and interpretable survival prediction is crucial for balancing treatments and quality of life. We focus on predicting overall survival in patients with symptomatic spinal bone metastases treated by radiotherapy, surgery or both. We evaluate PISA for both predictive accuracy and its potential to reduce
Digital Poster 440
A Multimodal Machine Learning Model to Improve Early Recurrence Prediction after Neoadjuvant Chemoradiotherapy for Rectal Cancer Chen-ying Ma, Yi Fu, Lou Liu, Jie Chen, Jing-wen Zhang, Ju-ying Zhou Radiation Oncology, The First Affiliated Hospital ofSoochow University, Suzhou, China Purpose/Objective: Locally advanced rectal cancer (LARC) patients are at high risk of early recurrence after neoadjuvant chemoradiotherapy (CRT) and surgery (about 80% of recurrences occur within 3 years). We developed a multimodal machine learning model integrating MRI radiomic features, H&E histopathology, and clinical data to predict early recurrence risk. Material/Methods: This retrospective study analyzed 88 stage II–III LARC patients who received neoadjuvant CRT and surgery (median follow-up 32 months; 2-year recurrence ~33%). Radiomic features were extracted from pre- treatment T2-weighted MRI, and pathomic features were derived from digitized H&E whole-slide images using a multiple instance learning neural network. Key clinical variables (e.g., age, carcinoembryonic antigen, tumor stage, regression grade, margin status) were also included. All features were combined in a classifier to predict early post-surgery recurrence (~18–24 months). The model was trained with cross- validation and tested on an independent cohort. Performance was evaluated by ROC AUC and other metrics (accuracy, sensitivity, specificity), along with survival analysis. The multimodal model's performance was also compared to that of single- modality models. Results: Among clinical factors, only a positive circumferential resection margin (CRM) significantly predicted early recurrence. Pathomic features alone achieved the highest single-modality performance (test AUC ~0.90), outperforming radiomics (~0.80) or clinical features. The integrated radiopathomic model achieved excellent performance (test AUC ~0.90) and significantly outperformed any individual modality (P < 0.05, DeLong's test), with better risk stratification and net benefit on decision curve analysis. SHAP interpretability analysis showed that two pathology- derived features were the most influential, followed by CRM status and intratumor heterogeneity radiomic features from all three tumor “habitats.” A fusion- model nomogram stratified patients into low- vs. high- risk groups with significantly different recurrence-free survival (P < 0.01 in both training and validation cohorts), confirming effective risk discrimination. Conclusion:
Made with FlippingBook - Share PDF online