Latent knowledge-driven unified representation learning for brain tumor segmentation with missing modalities.

Abstract

Gliomas are the most common primary malignant brain tumors, characterized by high heterogeneity, diffuse infiltration, and poor prognosis. Accurate tumor segmentation plays a fundamental role in diagnosis, treatment planning, and quantitative analysis. Although multimodal MRI provides rich anatomical and functional information, missing imaging modalities are frequently encountered in real-world clinical practice. Here, we propose a Latent Knowledge-driven Unified Representation Learning (LK-URL) framework for robust glioma segmentation under missing-modality conditions. LK-URL incorporates a Dynamic Adaptive Unified Interaction Module (DAUIM) to model intra- and inter-modality interactions, enabling enhanced tumor-aware feature representation and cross-modal compensation. Furthermore, a Latent Knowledge Collaborative Learning (LKCL) strategy is introduced to transfer structural knowledge from complete to incomplete modality settings via internal full-modality supervision, thereby improving the consistency of feature representations across diverse modality combinations and enhancing segmentation robustness. Extensive experiments on three datasets demonstrate that the proposed method consistently outperforms state-of-the-art approaches under missing-modality scenarios, while exhibiting strong generalization ability in real-world clinical settings. The code is available at https://github.com/joey-AI-medical-learning/LK-URL.