Learning Glioblastoma Tumor Heterogeneity Using Brain Inspired Topological Neural Networks

TL;DR

Glioblastoma prognosis is challenged by extreme intratumoral heterogeneity and cross-site MRI variability. The authors introduce TopoGBM, a two-stage, multimodal 3D autoencoder with a brain-inspired topology regularizer (TopoLoss) that preserves persistent homology in a latent $z \in \mathbb{R}^{d}$, and a cross-attention survival head that fuses $z$ with age to produce discrete-time risk. The model achieves a C-index of $0.67$ on internal validation and $0.58$ on external TCGA data, outperforming DL baselines that degrade under domain shift, while reconstruction residuals localize to tumor regions and roughly 50% of prognostic signal ties to tumor/peri-tumoral tissue, supporting interpretability. By integrating topology priors with multimodal MRI and clinical context, TopoGBM yields morphology-faithful, cross-site robust embeddings that have potential for clinically reliable GBM prognosis.

Abstract

Accurate prognosis for Glioblastoma (GBM) using deep learning (DL) is hindered by extreme spatial and structural heterogeneity. Moreover, inconsistent MRI acquisition protocols across institutions hinder generalizability of models. Conventional transformer and DL pipelines often fail to capture the multi-scale morphological diversity such as fragmented necrotic cores, infiltrating margins, and disjoint enhancing components leading to scanner-specific artifacts and poor cross-site prognosis. We propose TopoGBM, a learning framework designed to capture heterogeneity-preserved, scanner-robust representations from multi-parametric 3D MRI. Central to our approach is a 3D convolutional autoencoder regularized by a topological regularization that preserves the complex, non-Euclidean invariants of the tumor's manifold within a compressed latent space. By enforcing these topological priors, TopoGBM explicitly models the high-variance structural signatures characteristic of aggressive GBM. Evaluated across heterogeneous cohorts (UPENN, UCSF, RHUH) and external validation on TCGA, TopoGBM achieves better performance (C-index 0.67 test, 0.58 validation), outperforming baselines that degrade under domain shift. Mechanistic interpretability analysis reveals that reconstruction residuals are highly localized to pathologically heterogeneous zones, with tumor-restricted and healthy tissue error significantly low (Test: 0.03, Validation: 0.09). Furthermore, occlusion-based attribution localizes approximately 50% of the prognostic signal to the tumor and the diverse peritumoral microenvironment advocating clinical reliability of the unsupervised learning method. Our findings demonstrate that incorporating topological priors enables the learning of morphology-faithful embeddings that capture tumor heterogeneity while maintaining cross-institutional robustness.

Figures (4)

• Figure 1: TopoGBM Architectural Framework. The architecture consists of Stage 1: joint representational learning phase which compresses volumetric MRI data into a topologically-regularized latent representation $\mathbf{z}$ and a Stage 2 inference phase that utilizes a cross-attention head to fuse these imaging embedding features with clinical covariate. The final pipeline maps this multimodal representation to discrete-time survival bins to generate an expected risk score $r$.
• Figure 2: The regional attribution for top-performing exemplars in the expert-verified UPENN and TCGA cohorts, revealing a high degree of spatial alignment between Hazard Fractions (solid, left bars) and Embedding Fractions (hatched, right bars). Nearly 50% of the model's sensitivity is concentrated within the tumor and immediate peri-tumoral rings (0–20 voxels), demonstrating that the prognostic survival signal is intrinsically anchored to the topologically-encoded tumor morphology rather than healthy anatomical variance.
• Figure 3: Representative examples from internal UPENN (top 3 rows) and external TCGA/BraTS (bottom 3 rows) showing ground-truth FLAIR with expert tumor mask, reconstructed FLAIR, and the absolute reconstruction error over the 4 modalities). Across cohorts, the autoencoder preserves global anatomy while residual error concentrates in and around tumor regions, visually supporting tumor-enriched reconstruction discrepancies under domain shift.
• Figure 4: Distribution of per-patient reconstruction quality on Train, Test, and external Val (TCGA) using MAE, MSE, PSNR, and SSIM; each subplot shows a boxplot with points and the sample size (N) for each set