AdvisorMingkai Zheng

• We propose TopoCMamba, a topology-aware framework that integrates low-frequency structural information with high-frequency boundary cues to construct topological conductance constraints. These constraints directly regulate state propagation in Mamba. The proposed method addresses cross-boundary information mixing caused by the mismatch between scan adjacency and anatomical continuity under heterogeneous clinical acquisition conditions. It consequently improves boundary delineation and cross-domain generalization in medical image segmentation.
• We propose ReBridge-Flow, a measurement-aware posterior bridge re-coupling framework for Flow Matching. It jointly re-couples the source and clean endpoints while incorporating observation constraints, thereby preserving their local compatibility with the current state. This design mitigates bridge mismatch caused by conventional local corrections, as well as the resulting structural drift, artifacts, and over-smoothing. It enables more stable image restoration with improved structural consistency.

AdvisorProf. Björn Schuller (IEEE Fellow)

• We propose PriMD, a primitive memory distillation framework that disentangles shared semantics from modality-specific information and dynamically compensates missing modalities with primitive memory, addressing the instability of missing-modality compensation caused by existing methods’ holistic generation, alignment, or distillation of missing modalities without distinguishing shared semantics from modality-specific details.
• We propose D²IB, a differentiated dual-level information bottleneck framework that models differences in modality reliability and modality-combination dominance, and performs differentiated information compression at two bottleneck levels, addressing the problems of insufficient noise suppression and inadequate learning of weak modalities caused by existing methods’ neglect of such differences.