ESTRO 2025 - Abstract Book

S3734

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2025

Purpose/Objective: The use of hybrid magnetic resonance linear accelerator (MR-linac) in radiotherapy enables longitudinal analysis of tumor imaging features during fractionated radiotherapy. This study aimed to investigate radiomic features extracted from MR scans of liver cancer patients treated with fractionated stereotactic body radiotherapy (SBRT) and describe their temporal changes over the course of treatment. Material/Methods: The study is based on 276 MR scans of 46 liver cancer patients, acquired with the 0.35T MRIdian system at the time of pre-treatment simulation and at each of the five SBRT fractions. The images were processed using bias field correction and z-score normalization, followed by resampling to 2 mm cubic voxels. Radiomic features were extracted from the gross tumor volume (GTV) at all timepoints using our in-house developed software, Z-Rad [1]. Highly cross correlated features were eliminated (Spearman’s | r |>0.90). To account for all timepoints, for each feature we calculated a time-dependent slope via linear fitting that describes its individual temporal evolution. Using these temporal trends, uniform manifold approximation and projection (UMAP) was applied for dimensionality reduction, followed by hierarchical density-based spatial clustering of applications with noise (HDBSCAN) to identify patient groups. A Kruskal-Wallis test was performed to identify distinguishing features among these groups, followed by false discovery rate (FDR) correction for multiple testing (p-value<0.05). Finally, we investigated whether the features are a surrogate of GTV volume changes or if they carry additional independent information. Results: After pre-processing, 56 radiomic features were used in the analysis from the initial 160 features extracted from the GTV. Three patient groups were identified: group 1 (N=6 patients), group 2 (N=22 patients) and group 3 (N=18 patients). The Kruskal-Wallis test revealed 28 features that significantly distinguished the patient groups, consisting of 16 texture features and 12 intensity features. Among these, the three most significant features (p-value<0.001) included a neighbouring grey level dependence-based feature (high dependence low grey level emphasis), an intensity-based statistical feature (coefficient of variation) and an intensity histogram feature (histogram mode) (Figure 1). These features also showed low correlation with the GTV volume (| r |<0.50) (Figure 2).