Predictive Quality Management in Nepalese Multimodal Survival Prediction in High-Grade Gliomas via Boundary-Aware Swin-TransUNet and Radiomic-Genomic Cross-Attention

Abstract

Predicting overall survival (OS) prior surgery is challenging in high-grade gliomas (HGG) because they are characterized by complex spatial tumor heterogeneity, varying microenvironmental infiltration and non-linear relationships between clinical-genomic risk factors. The paradigms that have been established so far are either based on the premise of parallel two-stage pipelines where truncation errors occur or do not support the modelling of multi-scale semantics at long range in space. To tackle these challenges, we present a Boundary-Aware Swin-TransUNet model combined with a Cross-Attention Multimodal Fusion (CAMF) module, which is integrated with an end-to-end multi-task deep framework for collaborative tumor segmentation and continuous survival hazard prediction. Surv-TransNet performs processing of 3D multiparametric Magnetic Resonance Imaging (mpMRI), in addition to clinical parameters (age, Karnofsky Performance Status, extent of resection) and key molecular biomarkers (IDH1/2 mutation status, MGMT promoter methylation). A Boundary-Preserving Contrastive Loss (BPCL) is used to regularize the spatial feature maps from a 3D Shifted-Window Transformer (SWTrans) encoder while a Transformer based Cox Proportional Hazards module models non-linear survival distributions. Surv-TransNet was tested in the BraTS 2021/2023 Survival Benchmarks (n = 2,140 patient scans across multiple institutions) and a curated clinical cohort of HGG (n = 253 patient scans). Surv-TransNet had a mean C-index of 0.812 for predicting continuous OS and a mean accuracy of 72.8% for predicting 3 classes of continuous OS (short, mid, long survival). It achieved a Whole Tumor Dice Similarity Coefficient (DSC) of 93.85% and an Enhanced Tumor DSC of 89.40% for segmentation, which is better than the existing 3D U-Net, Swin UNETR, and DeepSurv baselines. In ablation experiments, joint spatial-boundary regularization, in combination with cross-attention fusion, is responsible for a 14.2% relative increase in prognostic discrimination.