HINTS: Hierarchically Disentangling Subregional Heterogeneity With Structural Priors for Multi-Modal Survival Analysis

Radiological tumor subregions can reflect intra-tumor heterogeneity (ITH), which is critical for accurate survival prediction. Existing deep survival methods primarily rely on voxel- or patch-level feature extraction to characterize tumor representations. However, such regular grid-based partitioning strategies fail to preserve the intrinsic morphology and spatial topology of tumor subregions, limiting their ability to capture intra-tumor heterogeneity. In this paper, we propose HINTS, a structure-aware framework for multi-modal survival analysis. HINTS leverages superpixels to construct structural priors for tumor subregions, enabling fine-grained heterogeneity modeling while preserving their natural morphology. We design Disentangling Heterogeneity Block that hierarchically aggregates subregional information and employs JSD loss to disentangle shared and modality-specific representations, thereby capturing complementary heterogeneity across modalities. In addition, we propose SQC Block that incorporates Quaternion Singular Value Decomposition to model high-order interactions among multi-modal subregional features. Experimental results on our NPC-1352 dataset and two public datasets demonstrate that structure-aware subregion modeling substantially improves survival prediction performance. The source code is available at https://github.com/LWX-Research/HINTS. Note to Practitioners - Accurately predicting the survival of cancer patients from MRI scans is crucial for personalized treatment planning. Tumors are often heterogeneous, with different subregions showing distinct structural and pathological characteristics. Existing automated approaches typically divide tumors into regular grids, which can distort natural boundaries and overlook important local details, limiting prediction accuracy. Our approach, HINTS, overcomes these limitations by adaptively segmenting tumors into irregular subregions that reflect their natural morphology and by integrating information from multiple MRI sequences to capture complementary signals from each modality. By modeling both the structure and the relationships among subregions, HINTS provides more reliable survival predictions. Through extensive experiments, we demonstrated that it improves prognostic accuracy compared to conventional grid-based methods. In practice, this means clinicians or AI systems can obtain more accurate survival estimates while preserving critical intra-tumor details, potentially supporting more informed treatment decisions.