Hypergraph facilitated feature fusion of multi-prototype for BRAF mutation prediction of thyroid cancer using cytology slides

Predicting BRAF mutation status from fine-needle aspiration cytology (FNAC) slides is crucial for personalized treatment of thyroid cancer, but remains challenging due to background noise, cellular sparsity and annotation scarcity. We propose PHICS , a P henotypic H ypergraph framework with I terative C ontrast S trategy, designed to address these challenges. A key innovation is its integration of single-cell morphological quantification with dynamic hypergraph feature fusion, which uncovers higher-order interaction patterns among cellular prototypes for robust slide representation. The iterative contrast strategy employs a reinforcement learning-guided cell-sampler to dynamically construct hypergraphs as augmented views for contrastive learning-based pretraining, further promoting hypergraph representational ability. Evaluated on cross-center cohorts, PHICS achieves state-of-the-art performance. As the first hypergraph-based application for FNAC data, PHICS establishes a traceable paradigm for genotype-phenotype correlation analysis, offering a clinically promising tool for preoperative molecular profiling.