HiCAF-Net: A Hierarchical Cross-Attention Fusion framework for cross-cancer subtype classification using...

Accurate classification of cancer subtypes is fundamental to personalized medicine, yet it remains hindered by pronounced tumor heterogeneity and the inherent limitations of unimodal analysis in capturing complex biological interactions. Existing multimodal fusion frameworks often struggle to effectively integrate multi-scale morphological features with high-dimensional genomic data, while their focus on single-cancer cohorts overlooks the predictive power of pan-cancer invariant hallmarks. In this study, we propose HiCAF-Net, a novel Hierarchical Cross-Attention Fusion and multitask learning framework designed for actionable knowledge discovery from fragmented oncology data. HiCAF-Net implements a coarse-to-fine visual extraction strategy, capturing macroscopic tissue architectures via attention-based multiple instance learning and microscopic nuclear topologies through graph convolutional networks. To bridge the semantic gap between modalities, we introduce a dual-layer stepwise cross-attention mechanism that progressively aligns these multi-scale pathological primitives with transcriptomic profiles. Furthermore, adversarial domain adaptation is integrated to synchronize optimization across heterogeneous cancer types, mitigating task conflicts and uncovering shared cross-cancer commonalities. Extensive experiments on datasets encompassing eight distinct cancer types (including TCGA cohorts and a private PTMC dataset) demonstrate that HiCAF-Net significantly outperforms state-of-the-art single-modal and single-task baselines. Our results underscore the framework's robust generalization and its potential as an efficient, deployable solution for joint pan-cancer analysis in clinical settings.