Answers are grounded in this paper and cite their sources.
Multi-modal brain MRI segmentation faces two major challenges: scarce high-quality annotations and frequent missing modalities in clinical practice. Although multi-modal methods exploit cross-modal complementarity, their tightly coupled fusion mechanisms often trigger severe feature distribution shifts and suppress modality-specific discriminative information when sequences are absent at inference. To address these issues, we propose CPKD-Seg(Cross-Modal Prototypical Knowledge Distillation for Segmentation), a semi-supervised framework that decouples multi-modal knowledge acquisition from the inference phase. Specifically, we introduce a bidirectional knowledge distillation (BKD) strategy that facilitates a mutually regularizing interaction between the multi-modal teacher and uni-modal students. By enabling the teacher to inversely absorb localized consensus knowledge from students, BKD mitigates potential over-fitting to highly coupled multi-modal inputs, thereby preserving modality-specific features for robust uni-modal reasoning. To further alleviate the cross-modal feature discrepancy, we propose a Cross-modal Prototypical Contrast (CPC) strategy. By aligning uni-modal features with category-level prototypes, CPC imposes global semantic constraints to bridge the representation gap between heterogeneous modalities and balance feature power across different sequences. Extensive experiments demonstrate that CPKD-Seg effectively narrows the performance gap caused by limited annotations. Notably, with only 10% labeled data, our model achieves uni-modal reasoning performance comparable to its fully supervised counterpart, while significantly outperforming state-of-the-art methods in missing-modality scenarios. • Enables Robust uni-modal inference for multi-modal medical image segmentation. • Bidirectional distillation transfers semantics while keeping private features. • Cross-modal prototype comparison learning aligns pixel-level features. • CPKD-Seg achieves near full-modal performance using single-modality inference.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: