ESTRO 2024 - Abstract Book

S5146

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2024

datasets including pre-operative MRIs. In clinical practice, it may not be feasible for every case to include annotations in the format of the public dataset. In this study, we assess the robustness of radiomic risk-stratifying models within challenges and constraints inherent to clinical applications such as limited clinical labels.

Material/Methods:

We utilized two GBM datasets: the publicly available BraTS 2020 [1] with 236 patients and a local dataset (STORM_GLIO), consisting of 53 patients acquired between April 2014 - April 2018 in Wales. Both datasets incorporated overall survival (OS) information. MRI modalities included: T1, contrast-enhanced T1 (T1ce), T2, and T2 FLAIR for all cases across the two datasets. We allocated 90% of the BraTS data for training (BraTS Training) and the remaining of the BraTS (BraTS Testing) plus STORM_GLIO data for testing. The BraTS 2020 included 3 labels: enhancing tumor (ET), tumor core (TC), and whole tumor (WT). STORM_GLIO included 1 label: Gross Tumor Volume (GTV). Radiomic analysis was carried out using SPAARC software [2,3]. For each patient, 143 imaging features were extracted from each MRI modality. We employed the LASSO Cox regression and Pearson correlation for radiomic feature selection and feature weight calculation by using five-fold cross-validation after splitting the BraTS data into the training and testing data. A risk stratification model was developed using the Radiomics-based Risk Score (RRS) which is calculated using a combination of selected features and weights [4]. To evaluate the RRS, Kaplan–Meier survival analysis, log–rank test, and C-index techniques were utilized. Table 1 shows the performance of the RRS models with each modality. It can be seen that the models with ‘T2’ have significant P values (<0.05) from the log-rank test. Figure 1 presents the Kaplan-Meier survival analysis for the risk models showing that the RRS models significantly separate low and high-risk groups. In both the BraTS Testing and STORM_GLIO, the C-indices depicted in the figure consistently exceed the 0.5 C-index threshold. Following feature selection, two texture-based radiomic features, namely 'ngl_dc_entr_3D', and 'ngl_dc_energy_3D' derived from the Neighborhood Grey Level Dependence Matrix (NGLDM) [5] along with age information, were considered important for training the risk model. The weights of these three features were -176.3, -374.8, and -146.8. For each patient, the RRS score was determined by summing the multiplications of individual features and their corresponding weights. The threshold value for the RRS model was calculated at 0.078. Thus, the risk stratification based on this threshold demonstrated significant results. Results: