📝 Original Info
- Title: Learning Glioblastoma Tumor Heterogeneity Using Brain Inspired Topological Neural Networks
- ArXiv ID: 2602.11234
- Date: 2026-02-11
- Authors: ** 논문에 명시된 저자 정보가 제공되지 않았습니다. (저자명 및 소속은 원문을 확인해 주세요.) — **
📝 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.
💡 Deep Analysis
📄 Full Content
Glioblastoma (GBM) is a leading cause of mortality in brain cancer with limited treatment options [1,2]. Determining the aggressiveness of the cancer from MRI is critical for designating therapy roadmap by clinicians. At present radiologist determined tumor aggressiveness progression from MRI is the only clinically reliable way to inform clinicians on prognosis. Supervised Deep learning (DL) has shown promising direction to extract features from MRI for early detection and onset aggressiveness determination for cancers. Yet this roadmap of therapy for aggressive GBM is frequently unreliable and ineffective in most cases [3,4,5] because of such models are black box in nature, are heavily overfit on their training cohorts which make them unexplainable and not generalizable across cohorts. Moreover, manual feature annotation is required for reliability. The critical gap still remains in learning biologically meaningful features from tumor microenvironment of the MRI, to make clinically relevant decisions. Especially for GBM, extracting tumor periphery, tumor core, tumor holes and tumor microenvironment irregularities from MRI is very challenging with existing unsupervised or supervised approaches. Hence, clinical decision making rarely relies on artificial intelligence (AI) driven feature extraction and modeling for GBM disease. Related Works: DL has transformed neuro-oncological imaging by enabling extraction of high-dimensional embeddings from MRI scans. In [6] authors propose a Convolutional denoising autoencoder (DAE) network which is combined with a Cox proportional hazard regression loss function to predict survival in the BRATS GBM cohort. In [7] authors propose DenseNet for survival classification on the Brats 2018 challenge [8,9]. In [5], [4] and [10] authors proposed transformer, deep neural network models for survival prediction. These models based on convolutional neural nets (CNN) and Vision Transformers (ViTs) capture rich semantic and non-linear patterns; however, they are heavily blackbox in nature. Hence attributing the features used for prognostication can seldom be attributed to tumor regions in GBM. This omission is critical because GBM aggressiveness correlates strongly with topological features such as rim irregularity, tumor periphery, infiltrative outgrowths, and necrotic geometry [11,12]. Lack of attributing to these properties in tumor representation make these models clinically unreliable. The other pressing scientific challenge with these existing architectures lie with their susceptibility to scanner variability. Despite internal (training) cohort metrics the models heavily underperform to unseen validation. To address these gaps, we propose TopoGBM, a GBM tumor heterogeneity representation learning framework trained with neural networks with topology-aware regularization. Our
Fig. approach incorporates a TopoLoss term that constrains the latent embeddings to preserve persistent homology features across multiple filtrations. We explicitly penalize deviations in the Betti-number barcodes, to ensure that critical shape descriptors like connectivity of necrotic cores and the continuity of irregular tumor margins are retained within the latent manifold z. [13]. These topology-constrained latent embeddings from stage 1 are subsequently harmonized across cohorts and passed to a cross-attention survival head in stage 2, where a clinical covariate (age) acts as the query over the imaging embedding to produce discrete-time risk predictions. We re-attribute the latent embeddings to regions across the tumor and normal tissue to explain the hazard and embedding fraction. Our objective is to learn clinically relevant features from MRI using semi-supervised representation learning to make outcome predictions more interpretable and generalizable to external cohorts. In multi-institutional experiments across UPENN, RHUH, and UCSF datasets, TopoGBM improves out-of-cohort survival prediction, reduces overfitting, and enhances biological interpretability compared to conventional models. These findings highlight the value of incorporating topological priors into representation learning for robust and clinically meaningful MRI biomarkers in GBM prognosis. We make the following key contributions in this paper: 1. Brain inspired unsupervised representation learning: To enforce anatomically meaningful structure in the representation, we introduce a mammalian-brain-inspired topological regularizer (TopoLoss) on an encoder, which encourages topology-preserving organization of learned features, promoting stability to domain shift while retaining clinically relevant morphology. This topology-aware constraint biases the latent space toward robust shape-driven cues rather than spurious texture, strengthening interpretability of downstream risk signals.
We present TopoGBM as a semi-supervised representational learning framework that (i) learns a compact multimodal la-tent space representation from 3D MRI via a
Reference
This content is AI-processed based on open access ArXiv data.