Search papers, labs, and topics across Lattice.
This paper introduces UnDA, an innovative framework for unpaired cross-modal knowledge distillation aimed at enhancing medical imaging tasks. By leveraging an anchor-guided Alignment Module and implementing Uncertainty-Weighted Optimal Transport (UCT-OT), the method effectively mitigates noise from uncertain predictions and facilitates robust knowledge transfer across modalities. Experimental results reveal significant improvements in accuracy and boundary precision for segmentation tasks, indicating that effective structural knowledge transfer is achievable without the need for paired datasets.
UnDA achieves superior cross-modal knowledge transfer in medical imaging, even in the absence of paired clinical data, by dynamically suppressing noise from uncertain predictions.
Multimodal based approaches often outperform single modality approaches in downstream tasks as the different modalities provide complementary information, yet acquiring paired clinical data remains a significant challenge in real world scenarios. While cross-modal knowledge distillation addresses this, existing methods often struggle with large modality gaps and the propagation of noise from uncertain source-domain predictions. To overcome these challenges, we propose UnDA, an anchor-guided framework for unpaired cross-modal distillation. Our approach introduces a backbone-agnostic Alignment Module that extracts semantically structured class tokens via an attention based pooling mechanism. To ensure robust knowledge transfer, we propose Uncertainty-Weighted Optimal Transport (UCT-OT), which dynamically weights feature-level alignment based on prediction confidence, effectively suppressing noisy supervision. Furthermore, a per-class ProtoNCE objective maintains stable prototype memories to enforce global discriminability across unpaired batches. Evaluations on representative segmentation tasks under strictly unpaired settings show consistent improvements in accuracy and boundary precision in the target modality, demonstrating that meaningful structural knowledge can be transferred across heterogeneous data sources without paired datasets.
