S2420
Physics - Radiomics, functional and biological imaging, and outcome prediction
ESTRO 2026
1 Radiation Physics, Department of Hematology, Oncology and Radiation Physics, Skåne University Hospital, Lund, Sweden. 2 Medical Radiation Physics, Lund University, Lund, Sweden. 3 Centre for Mathematical Sciences, Lund University, Lund, Sweden. 4 Department of Neurosurgery, Skåne University Hospital, Lund, Sweden. 5 Department of Clinical Sciences/Division of Radiology, Lund University, Lund, Sweden. 6 Department of Medical Imaging and Physiology, Skåne University Hospital, Lund, Sweden Purpose/Objective: This study investigates whether imaging parameters obtained pre-surgery, post-surgery, and pre- radiotherapy (RT) improve survival prediction for glioblastoma (GBM) patients when combined with clinical data during chemo-radiotherapy (CRT) to 60 Gy. Material/Methods: We included 238 consecutive patients with histologically confirmed IDH-wildtype GBMtreated with concurrent and adjuvant CRT at our institution (2016– 2023). Imaging data were collected at three time points (pre-surgery, post-surgery, pre-RT), alongside clinical variables (age, sex, MGMT status), radiotherapy structures and doses, and neurosurgical reports.T1- weighted MRI with and without contrast was used to quantify viable tumor at the time of RT planning by calculating the contrast-enhancing (CE) volume within the gross tumor volume (GTV). MRI preprocessing included N4 bias correction and Z-score normalization within the brain mask. The T1 image was subtracted from the T1CE image, generating an intensity histogram within the GTV (Figure 1). A histogram threshold was optimized by cross-validation(shuffle split, 100 iterations, test size=30%) to maximize the concordance index in univariate Cox modeling. Finally, CE was calculated as the volume of the voxels with an intensity above the threshold.
Patients with complete data were split into training (n=152) and validation (n=66) cohorts. Univariate Cox models were used to identify prognostic factors, and variables with p<0.005 (Bonferroni correction=10) were entered into multivariate Cox models. Elastic net regularization was applied for feature selection, and logistic regression was used to predict survival shorter than 15 months. Results: MGMT status (p<0.001), extent of surgical resection (p<0.001), GTV (p=0.003), mean brain dose (p<0.001), CTV (p<0.001), and CE (p<0.001) were significantly associated with overall survival in univariate Cox regression analysis. The CE parameter showed strong prognostic value (concordance index 0.645; HR 1.028, 95% CI 1.019–1.050). In the multivariate Cox elastic net model, non-zero variables included CE, CTV, and MGMT status. The model achieved a concordance index of 0.694 in training and 0.665 in validation data. The same variables were used in a logistic regression yielding AUCs of 0.78 (training) and 0.76 (validation) (Figure 2).
Made with FlippingBook - Share PDF online