S2433
Physics - Radiomics, functional and biological imaging, and outcome prediction
ESTRO 2026
WC, Perks JR, Apte A, Aristophanous M, LoCastro E, Hsu D: Multi-institutional atlas of brain metastases informs spatial modeling for precision imaging and personalized therapy. Nature communications 2025, 16(1):4536.2. Pantelis E, Papadakis N, Verigos K, Stathochristopoulou I, Antypas C, et.al.: Integration of functional MRI and white matter tractography in stereotactic radiosurgery clinical practice. International Journal of Radiation Oncology Biology and Physics 2010, 78(1):257-267.3. Sebenius I, Seidlitz J, Warrier V, Bethlehem RA, Alexander-Bloch A, et.al : Robust estimation of cortical similarity networks from brain MRI. Nature neuroscience 2023, 26(8):1461-1471. Keywords: MRI, brain hubs, brain network Integrating Clinical, Radiomics, and Deep Learning Features for 12-Month Overall Survival Prediction in NSCLC Patients Treated with Radiotherapy Hemalatha Kanakarajan 1 , Jikai Zhou 2 , Aiara Lobo Gomes 3,2 , Wouter De Baene 1 , Fariba Tohidinezhad 4,2 , Petros Kalendralis 2 , Wenjie Liang 2 , Andre Dekker 2 , Margriet Sitskoorn 1 1 Department of Cognitive Neuropsychology, Tilburg University, Tilburg, Netherlands. 2 Department of Radiation Oncology (Maastro), GROW School for Oncology and Reproduction, Maastricht University Medical Centre, Maastricht, Netherlands. 3 Institute of Molecular Medicine, RWTH Aachen University, Aachen, Digital Poster Highlight 1930 Germany. 4 Department of Radiology & Nuclear Medicine, Erasmus Universiteit Medical Centre, Rotterdam, Netherlands Purpose/Objective: Accurate prediction of overall survival (OS) in non- small cell lung cancer (NSCLC) patients remains challenging due to wide patient heterogeneity. While radiomics and deep learning (DL) have independently shown prognostic value, their integration with clinical and dosimetric features has rarely been explored. This study aimed to develop and validate an ensemble model combining clinical, radiomic, DL features to predict 12-month OS in NSCLC patients treated with definitive radiotherapy. Material/Methods: A retrospective cohort of 225 NSCLC patients treated with definitive radiotherapy at Maastro Clinic was analyzed (training: 180; testing: 45). Radiomic features were extracted from tumor volumes using PyRadiomics, and 3D DL features from a pre-trained 3D ResNet-18 network. Clinical and dosimetric variables, including prescribed dose, were added. After feature selection via Recursive Feature Elimination (RFE) and data balancing with Synthetic Minority Oversampling Technique (SMOTE) plus
Labeling atlas, generating two types of 90 x 90 connectivity matrices: 1) DTI-based matrices derived from probabilistic tractography and quantified by streamline counts; and 2) T1 morphometric similarity matrices constructed from surface features extracted from T1-weighted MRI [3]. Morphological features including cortical thickness, surface area, curvature, gray matter volume, and sulcal depth were extracted; pairwise similarity between regions was computed using the Kullback–Leibler divergence. Six network metrics, including degree and betweenness centrality, clustering coefficient, nodal and local efficiency, and shortest path length, were derived to describe network topology. Metric stability was assessed by computing the intraclass correlation coefficient (ICC) across 88 × 90 (subjects × regions) matrices. Metrics with ICC > 0.7 were considered robust and used to identify cortical hubs. Results: DTI- and T1-based matrices showed high inter-subject consistency (ICC > 0.7). Mean ICC values and 95% confidence intervals are summarized in Table 1. DTI- based analyses identified six robust hubs (the bilateral precuneus, putamen, and thalamus), consistently ranking among top nodes across stable metrics. The convergence of DTI-derived hubs across multiple metrics highlights their robustness supporting their role as stable markers of structural brain organization. T1-surface matrices identified a smaller but consistent set of reproducible hubs, primarily in the superior frontal gyrus, posterior cingulate, and middle temporal regions. Although the hubs identified from T1-MRI were spatially distinct from those derived from DTI, their detection on conventional anatomical imaging highlights the ability of morphometric networks to provide complementary information on large-scale structural connectivity.
Conclusion: The identification of reproducible hubs from T1- derived morphometric networks suggests that conventional anatomical imaging can capture complementary features of brain organization. Although these results require validation in larger independent cohorts, they provide a promising basis for integrating morphometric connectivity analysis into clinically accessible radiotherapy workflows aimed at brain function preservation. References: 1. Barrios J, Porter E, Capaldi DP, Upadhaya T, Chen
Made with FlippingBook - Share PDF online