Abstract
Glioblastoma (GBM) is a highly aggressive brain tumor that exerts significant mechanical stress on surrounding healthy brain tissue, a phenomenon termed the "tumor field effect." Traditional survival prediction models often focus solely on the visible tumor, ignoring these remote biomechanical changes. This paper introduces R-DepTH, a novel MRI-based descriptor that integrates intra-tumoral texture features with biomechanical deformation measurements of the surrounding brain parenchyma. R-DepTH demonstrates superior prognostic accuracy and generalizability compared to standard radiomic and deep learning approaches across multi-institutional datasets.
Introduction
Glioblastoma (GBM) remains one of the most challenging cancers due to its aggressive growth and systemic impact on the brain. The "mass effect" or "tumor field effect"—where the expanding tumor physically distorts remote brain regions—is a primary driver of morbidity and mortality. While MRI is the standard for diagnosis, most existing survival prediction models (radiomic or deep learning-based) extract features only from the visible tumor and its immediate margins. These models often fail to capture the subtle, quantifiable biomechanical distortions in the seemingly normal "Brain Around Tumor" (BAT) regions. We hypothesize that integrating these remote structural deformations with internal tumor texture provides a more comprehensive prognostic marker.
Methods
The radiomic-Deformation and Textural Heterogeneity (r-DepTH) descriptor was developed to capture both internal morphology and external mechanical impact. The pipeline consists of three primary stages:
- Deformation Features: Patient MRIs are non-rigidly aligned to a healthy brain atlas using diffeomorphic registration (ANTs SyN). The resulting deformation field quantifies the displacement of healthy tissue. Magnitude measurements are taken in 12 concentric 5mm annular rings extending up to 60mm from the tumor boundary.
- Textural Features: 3D COLLAGE (Co-occurrence of Local Anisotropic Gradient Orientations) is used to capture internal heterogeneity by extracting higher-order statistics of local gradient tensors from the enhancing lesion and peri-tumoral regions.
- Modeling: Deformation and COLLAGE features are integrated into the R-DepTH descriptor. A LASSO + Cox regression model selects the most predictive features to calculate a survival risk score, stratifying patients into high-risk and low-risk groups.
Experimental Setup
The study utilized a multi-institutional retrospective cohort of 207 GBM patients with 1.5T MRI scans (Gd-T1w, T2w, T2w-FLAIR).
- Training Set ($S_t$): 128 cases from the Cancer Imaging Archive (TCIA) and Cleveland Clinic (CCF).
- Testing Set ($S_v$): 79 independent cases from the Medical College of Wisconsin (MCW).
Results and Discussion
R-DepTH successfully stratified patients into risk groups, achieving a Concordance-index of 0.80 on the training set ($p = 3.5 \times 10^{-7}$) and 0.75 on the independent testing set ($p = 0.0024$). Notably, the model outperformed standard radiomics (4,376 features) and CNN-based transfer learning (4,096 features), both of which failed to generalize to the test set ($p = 0.84$ and $p = 0.54$, respectively).
Key insights include:
- Remote Prognostication: Significant structural deformations were found up to 60mm from the tumor, indicating that the physical impact on distant tissue is a powerful prognostic signal.
- Feature Interpretability: The most predictive features were the skewness and kurtosis of displacement magnitudes. High-risk patients showed higher skewness, reflecting aggressive, lopsided pushing of the tumor into healthy tissue.
- Robustness: By grounding the model in biomechanical realities rather than just pixel intensities, R-DepTH proved more robust to "scanner shift" across different institutions.
Conclusion
R-DepTH provides a novel, physically-grounded approach to GBM prognosis by quantifying the tumor field effect. Future work will explore the inclusion of directional deformation attributes and validation on larger prospective cohorts, as well as application to other solid tumor types.
Summary updated