Shape matters: Unsupervised Exploration of Glioblastoma Imaging Survival Predictors

1458 

Abstract Submission 

Martha Foltyn-Dumitru¹, Mustafa Mahmutoglu¹, Felix Sahm¹, Wolfgang Wick¹, Martin Bendszus¹, Philipp Vollmuth¹, Marianne Schell¹

¹Heidelberg University Hospital, Heidelberg, Germany

Martha Foltyn-Dumitru  
Heidelberg University Hospital
Heidelberg, Germany

Mustafa Mahmutoglu  
Heidelberg University Hospital
Heidelberg, Germany

Felix Sahm  
Heidelberg University Hospital
Heidelberg, Germany

Wolfgang Wick  
Heidelberg University Hospital
Heidelberg, Germany

Martin Bendszus  
Heidelberg University Hospital
Heidelberg, Germany

Philipp Vollmuth  
Heidelberg University Hospital
Heidelberg, Germany

Marianne Schell  
Heidelberg University Hospital
Heidelberg, Germany

Within glioblastoma research, conventional metrics such as 2D/3D tumor size measurements have long served as reliable prognostic indicators. However, the complete prognostic potential of sophisticated structural parameters, particularly morphological radiomics, remains to be thoroughly explored. This investigation delves into a clustering technique on morphological radiomics along with tumor volume to unveil distinct glioblastoma phenotypes, assessing their prognostic impact on overall survival (OS).

A retrospective study included 436 glioblastoma patients (2009-2020) from Heidelberg University Hospital. Data were split into training and test datasets (80:20 ratio), with external validation using the UCSF glioma dataset including 397 patients (Calabrese et al., 2022). MRI acquisition involved 3D T1-weighted imaging (pre/post contrast administration), axial 2D FLAIR, and T2-weighted imaging.
Automated tumor segmentation was performed using a variant of HD-GLIO (Isensee et al., 2019; Kickingereder et al., 2019). PyRadiomics facilitated the radiomic feature extraction of nine morphological radiomic features from the tumor (van Griethuysen et al., 2017), and the total tumor volume was ascertained. Threshold determination for morphological radiomic features and tumor volume involved hierarchical Bayesian modeling within a Cox proportional hazards framework for OS (Chen et al., 2014; Fang et al., 2017). Parameters were binarized according to their respective thresholds. Subsequently, a Gower distance was computed using these binarized parameters, serving as the foundational metric for subsequent partition around medoids (PAM) clustering (Hennig, 2023; Maechler et al., 2022). Cluster robustness was quantitatively appraised using the Jaccard index across 500 bootstrap iterations.
Survival rates were visualized using Kaplan-Meier curves and tested for significance by log-rank test. Univariate and multivariate Cox regression models, adjusted for clinical covariates, explored cluster and tumor volume impact on OS. Discriminative ability was evaluated using the concordance probability (C index) and Akaike information criterion (AIC). ANOVA compared C indices for embedded models, and the evidence ratio from AIC assessed differences between non-embedded models.

PAM clustering identified two clusters with the highest silhouette coefficient (width=0.44) and high stability (Jaccard index: 0.94 for cluster 1, 0.89 for cluster 2). Cluster composition analysis showed distinct patterns: Cluster 1 (n=233) had a higher proportion of patients with higher Sphericity and Elongation, while Cluster 2 (n=115) had a higher proportion of patients with higher Maximum 3D Diameter, Surface Area, Axis Lengths, and tumor volume (p<0.001 for each).
OS differed significantly between clusters: Cluster 1 showed median OS of 18.8, 23.8, and 20.1 months in the training, test, and UCSF datasets, respectively, whereas Cluster 2 showed 11.7, 11.4, and 13.7 months (p<0.003 for all; Figure 1). Univariate Cox regression linked cluster affiliation with OS (HR=2.25, p=0.003, C index=0.625) in the test dataset. Multivariate Cox regression showed improved performance with cluster affiliation over clinical data alone (C index 0.67 vs. 0.59, p=0.003) and further enhanced predictive accuracy with preoperative ECOG status (C index=0.68, p=0.005). Cluster-based models outperformed the models with tumor volume alone (evidence ratio 5.16-5.37; Table 1).

Data-driven clustering reveals clinically relevant imaging phenotypes. This underscores the enhanced prognostic value achieved by combining morphological radiomics with tumor size, emphasizing the superiority of this integrated approach over relying solely on tumor size when predicting survival outcomes in glioblastoma patients.

Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s)

Classification and Predictive Modeling ¹

Multivariate Approaches

Multi-Modal Imaging ²

ADULTS

Data analysis

Machine Learning

Modeling

Neoplastic Disease

Neurological