HiGATE: hierarchical graph attention for multi-scale tissue encoder in computational pathology.

Background: Histopathological diagnosis is inherently a multi-scale reasoning process, where pathologists seamlessly integrate cellular morphology with tissue architecture. Yet, computational models remain fragmented by analyzing cells and tissues in isolation, missing the diagnostic synergy that emerges from their interplay. This disconnect limits both predictive accuracy and clinical trust. Methods: We introduce HiGATE (Hierarchical Graph Attention Tissue Encoder), a biologically-inspired framework that unifies cellular and tissue-level analysis through a novel dual-graph architecture. Unlike prior hierarchical models that rely on static, unidirectional information flow, HiGATE introduces a bidirectional Cross-Level Attention mechanism enabling dynamic, context-aware communication where cellular details inform tissue organization and architectural context refines cellular representations. The framework incorporates learnable, spatially-constrained graph construction via differentiable pooling with spatial regularization, adaptively capturing tissue heterogeneity rather than relying on fixed heuristics. Multi-modal nuclear features integrate domain-adapted visual semantics (DINOv2), morphological shape descriptors, and fine-grained morphometrics (StarDist). We validate HiGATE across four diverse datasets spanning multiple tasks: PanNuke (nuclei classification with dual hierarchical levels: 5 nuclear types and 19 tissue types), MoNuSeg (segmentation feature transfer), DigestPath (colon polyp classification), and TCGA-BRCA (whole-slide breast cancer grading). Results: On the PanNuke benchmark, HiGATE achieves state-of-the-art performance with 91.3% accuracy and an F1-score of 0.896 for nuclei classification, while also achieving 85.4% accuracy for tissue-type classification across 19 cancer types. The framework demonstrates exceptional cross-dataset generalization: MoNuSeg segmentation (Dice = 0.841), DigestPath classification (accuracy = 0.872), and TCGA-BRCA WSI-level grading (accuracy = 0.854). Similarly, at a clinically relevant high-sensitivity operating point (recall = 0.95), HiGATE maintains precision of 0.87-a 10.1% reduction in false positives over HACT-Net. A multi-reader study with five board-certified pathologists confirms the clinical relevance of HiGATE's integrated multi-scale explanations (mean diagnostic relevance = 4.1/5.0, Conclusion: HiGATE bridges the gap between high-performance AI and clinically actionable diagnostics by unifying predictive accuracy with transparent, pathologist-aligned reasoning. The bidirectional cross-scale attention mechanism constitutes a general contribution to hierarchical graph representation learning, with potential applications extending beyond computational pathology to any domain requiring sophisticated multi-scale relational reasoning. Our comprehensive validation across tasks and tissue types establishes HiGATE as a robust foundation for trustworthy diagnostic AI in personalized medicine.